System and method for industrial risk assessment via computer vision

Abstract

A device, system and method comprising computer vision techniques for fire prevention / detection and risk assessment, as well as for determining deviations from an ideal operational state. The present invention includes for example systems and methods which leverage data collected by camera systems composed of infrared and visible light sensors to detect and / or prevent a fire from starting, and additionally, use this data to determine a risk assessment for the building. The present invention also provides for example a system and method for monitoring and controlling safety risks in indoor industrial environments by determining deviations from an ideal operational state using computer vision techniques and game-theoretic competitive ranking frameworks.

Description

[0001] Title: SYSTEM AND METHOD FOR INDUSTRIAL RISK ASSESSMENT VIA

[0002] COMPUTER VISION

[0003] Inventors: Drew Hanover, Thomas Laengle, Florian Trautweiler

[0004] Field of the Invention

[0005] The present invention relates to the field of automated risk assessment in industrial environments and also risk reduction, including computer vision techniques for fire prevent! on / detecti on and risk assessment.

[0006] Background

[0007] In industrial environments, maintaining optimal operating conditions is crucial for both safety and operational efficiency. Traditional risk assessment methods typically involve periodic inspections by qualified risk engineers who evaluate various factors including environmental state, cleanliness, occupancy, electrical installations, and other safety -related parameters. These assessments form the basis for insurance premium pricing and risk management protocols.

[0008] However, conventional periodic assessments suffer from significant limitations. First, they provide only a snapshot of conditions at a specific moment in time, rather than a continuous representation of day-to-day operations. Second, facilities are often notified in advance of inspections, allowing them to temporarily optimize conditions that may not reflect normal operations. This advance notification may lead to inaccurate risk assessments and potentially dangerous operating conditions between inspections. Furthermore, risk assessments have a direct impact on underwriting, and therefore act as a pricing mechanism for insurances. If the facility is arbitrarily different (read cleaner) on the day of an inspection, it can positively impact the risk assessment report resulting in lower premiums for the insured, and non-representative risk conditions for the insurer.

[0009] Traditional fire detection technologies also have limitations, for example in high- ceiling industrial buildings where ceiling-mounted smoke detectors may only detect hazards after substantial development. These limitations create a need for more comprehensive, continuous monitoring systems that can detect risks at their source and provide more accurate real-time assessment of safety conditions. As a specific example of a type of risk, fire risk assessment is a procedure mandated by many insurance providers to understand the risk of catastrophic loss due to fire damage. Typically, these assessments are conducted once or twice per year, with qualified risk engineers representing an underwriter. The risk engineer’s job is to determine the likelihood and potential scale of a fire based on many factors. These factors may include, but are not limited to, environmental state, cleanliness, occupancy, electrical installations, sprinklers, alarming systems, fire doors, geographical location and of course the intended purpose of the building, i.e. manufacturing, residential, occupational and so on.

[0010] The risk engineer takes these factors into consideration and provides risk scores to the underwriter. The underwriter then uses these risk scores as a primary factor in insurance premium prices.

[0011] One challenge with these risk assessments is their periodic nature. Insurance clients are often notified in advance when a risk assessment is to occur. The risk engineers may encourage clients to make the building as prepared as possible before, or on the day of inspection. This means that when the risk engineer arrives to conduct the assessment, the actual condition of the building does not accurately reflect the normal day-to-day operating characteristics.

[0012] Taking a sawmill or pellet manufacturing plant for example, the owner of the mill may make significant efforts to clean the facility, removing highly flammable sawdust from dangerous areas which would otherwise exist during normal operations. They may also repair dangerous, yet still functional equipment to avoid any demerits to their risk score in order to ultimately receive the lowest possible insurance premiums.

[0013] Furthermore, traditional smoke and fire detection technologies have significant limitations, particularly in industrial buildings with high ceilings. Conventional smoke detectors and sprinklers are typically ceiling-mounted, which can delay detection in large- volume spaces. In such environments, smoke or heat may take considerable time to reach the sensor level, and by the time an alarm is triggered, the fire may already be well developed. This delay can result in extensive damage and contributes to higher insurance premiums across the entire customer segment. Brief Summary of the Invention

[0014] This summary is provided to introduce a selection of concepts in a simplified form that are further described below in the detailed description. This summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used as an aid in determining the scope of the claimed subject matter.

[0015] The present invention relates to a device, system and method comprising computer vision techniques for fire prevention / detection and risk assessment, as well as for determining deviations from an ideal operational state. The present invention includes for example systems and methods which leverage data collected by camera systems composed of infrared and visible light sensors to detect and / or prevent a fire from starting, and additionally, use this data to determine a risk assessment for the building. The present invention also provides for example a system and method for monitoring and controlling safety risks in indoor industrial environments by determining deviations from an ideal operational state using computer vision techniques and game-theoretic competitive ranking frameworks.

[0016] For example, a system is provided comprising a plurality of cameras physically positioned throughout an indoor industrial environment, a processor, and a memory storing instructions. When executed by the processor, these instructions establish a game-theoretic competitive ranking framework where each camera functions as a player and each captured image represents a move made by the respective camera.

[0017] The system performs head-to-head matches between images from different cameras and between current and past images from the same camera, evaluating which image in each match is closer to an ideal operational state using a trained artificial intelligence model. This ideal operational state comprises proper arrangement of operational materials, machines, and humans; correct spatial organization; and appropriate environmental parameters including temperature.

[0018] Based on these evaluations, the system updates ratings for each camera according to a competitive ranking algorithm, determines if physical deviations from the ideal operational state exceed predetermined thresholds, and automatically triggers physical alarm devices or activates safety mitigation systems when necessary.

[0019] Without wishing to be limited by a closed list, this approach provides continuous monitoring and assessment of industrial environments, enabling early detection of safety risks and more accurate representation of day-to-day operating conditions compared to conventional periodic assessments. This technology creates opportunities to shift the risk assessment paradigm, moving from periodic, snapshot assessment to 24 / 7 monitoring of high fire-risk manufacturers such as wood products, steel, chemical, battery, or food / agri cultural manufacturers using advanced computer vision techniques across the infrared and visible light spectrums.

[0020] Without wishing to be limited by a closed list, visible light sensors, or Standard Cameras (SCs), have many advantages in fire risk assessment.

[0021] Fire, smoke, and sparks have unique visual characteristics which may be exploited to dramatically improve the usefulness of SCs. For example, motion, color or contrast analysis algorithms from classical computer vision may be used to detect smoke without the need of human supervision. In the same vein, millions of images containing smoke, fire, and sparks exist across the internet. Modem machine learning algorithms such as classifiers or singleshot detection (SSD) networks may be trained to recognize the visual characteristics of fire, smoke, or sparks with unprecedented accuracy. Once recognition has occurred, alarm systems may be triggered to provide an advantage over typical fire alarm systems such as smoke detectors. This is because the visual signature of smoke, fire, or sparks may be recognized at the source, rather than waiting for smoke or heat to rise to the ceiling where a smoke detector is typically mounted.

[0022] Machine learning approaches may also provide certain advantages to their classical computer vision counterparts such as early detection, confidence measurements, and continuous improvement with more data.

[0023] When images are captured by the SC, the data may be transmitted to a computational unit which may be an onboard computer (inside the camera housing), an on-site network computer, or a cloud computer. The data is then processed by an algorithm with the primary purpose of detecting fire, smoke, or sparks.

[0024] Detections can then be counted and analyzed to determine their source, frequency, confidence, or other patterns which may be useful in quantifying the risk of fire to a facility. Humans in the loop may be necessary to improve machine learning based approaches. For example, a human may be tasked with manually labeling detections as true or false positives. With the newly labeled data, the machine learning algorithm can then be retrained to improve the performance and reduce false positives. Non-limiting Examples of Advantages of the Invention

[0025] The invention provides several significant advantages over conventional risk assessment methods (without wishing to be limited by a closed list):

[0026] 1. Continuous Monitoring: Unlike periodic inspections, the system provides continuous monitoring of industrial environments, enabling detection of risks as they develop rather than after they have become significant hazards.

[0027] 2. Objective Assessment: The game-theoretic framework provides an objective method for assessing conditions, reducing the subjectivity inherent in human evaluations.

[0028] 3. Early Detection: By continuously monitoring conditions and detecting deviations from the ideal operational state, the system enables early intervention before conditions deteriorate to the point of creating serious hazards.

[0029] 4. Accurate Representation: The system provides a more accurate representation of day-to-day operating conditions compared to periodic inspections, which may not reflect normal operations.

[0030] 5. Dynamic Risk Assessment: The continuous updating of ratings and risk scores enables a dynamic assessment of safety risks that adapts to changing conditions in the industrial environment.

[0031] 6. Incentivized Safety Improvements: By potentially linking risk scores to insurance premiums, the system creates a financial incentive for facilities to maintain safer operating conditions.

[0032] Implementation of the method and system of the present invention involves performing or completing certain selected tasks or steps manually, automatically, or a combination thereof. Moreover, according to actual instrumentation and equipment of preferred embodiments of the method and system of the present invention, several selected steps could be implemented by hardware or by software on any operating system of any firmware or a combination thereof. For example, as hardware, selected steps of the invention could be implemented as a chip or a circuit. As software, selected steps of the invention could be implemented as a plurality of software instructions being executed by a computer using any suitable operating system. In any case, selected steps of the method and system of the invention could be described as being performed by a data processor, such as a computing platform for executing a plurality of instructions. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. The materials, methods, and examples provided herein are illustrative only and not intended to be limiting.

[0033] An algorithm as described herein may refer to any series of functions, steps, one or more methods or one or more processes, for example for performing data analysis.

[0034] Implementation of the apparatuses, devices, methods and systems of the present invention involve performing or completing certain selected tasks or steps manually, automatically, or a combination thereof. Specifically, several selected steps may be implemented by hardware or by software on an operating system, of a firmware, and / or a combination thereof. For example, as hardware, selected steps of at least some embodiments of the disclosure may be implemented as a chip or circuit (e.g., ASIC). As software, selected steps of at least some embodiments of the disclosure may be implemented as a number of software instructions being executed by a computer (e.g., a processor of the computer) using an operating system. In any case, selected steps of methods of at least some embodiments of the disclosure may be described as being performed by a processor, such as a computing platform for executing a plurality of instructions. The processor is configured to execute a predefined set of operations in response to receiving a corresponding instruction selected from a predefined native instruction set of codes.

[0035] Software (e.g., an application, computer instructions) which is configured to perform (or cause to be performed) certain functionality may also be referred to as a “module” for performing that functionality, and also may be referred to a “processor” for performing such functionality. Thus, a processor, according to some embodiments, may be a hardware component, or, according to some embodiments, a software component.

[0036] Further to this end, in some embodiments: a processor may also be referred to as a module; in some embodiments, a processor may comprise one or more modules; in some embodiments, a module may comprise computer instructions - which may be a set of instructions, an application, software - which are operable on a computational device (e.g., a processor) to cause the computational device to conduct and / or achieve one or more specific functionality.

[0037] Some embodiments are described with regard to a "computer," a "computer network," and / or a “computer operational on a computer network.” It is noted that any device featuring a processor (which may be referred to as “data processor”; “pre-processor” may also be referred to as “processor”) and the ability to execute one or more instructions may be described as a computer, a computational device, and a processor (e.g., see above), including but not limited to a personal computer (PC), a server, a cellular telephone, an IP telephone, a smart phone, a PDA (personal digital assistant), a thin client, a mobile communication device, a smart watch, head mounted display or other wearable that is able to communicate externally, a virtual or cloud based processor, a pager, and / or a similar device. Two or more such devices in communication with each other may be a "computer network."

[0038] Brief Description of the Drawings

[0039] The invention is herein described, by way of example only, with reference to the accompanying drawings. With specific reference now to the drawings in detail, it is stressed that the particulars shown are by way of example and for purposes of illustrative discussion of the preferred embodiments of the present invention only, and are presented in order to provide what is believed to be the most useful and readily understood description of the principles and conceptual aspects of the invention. In this regard, no attempt is made to show structural details of the invention in more detail than is necessary for a fundamental understanding of the invention, the description taken with the drawings making apparent to those skilled in the art how the several forms of the invention may be embodied in practice. In the drawings:

[0040] Figures 1A-1D provide an illustrative overview of how a network of cameras capable of sensing various light spectrums may be used to provide a quantitative risk assessment score based on day-to-day operations within a facility;

[0041] Figures 2A, 2B and 3 relate to non-limiting, illustrative examples of networks, sensors and systems for receiving data for risk determination;

[0042] Figures 4-13 relate to non-limiting examples of devices, systems and methods that use computer vision to obtain quantitative data, in order to determine quantifiable risk;

[0043] Figures 14-17 relate to the cleanliness of the environment as a risk factor;

[0044] Figure 18 illustrates a block diagram of a monitoring system for intruder detection and risk assessment;

[0045] Figure 19 shows a flowchart for calculating a hazardous materials risk score calculation;

[0046] Figure 20 demonstrates an exemplary program flow of how to use the long term history of these alerts as a means to aggregate information; Figure 21 illustrates a block diagram of a monitoring system for integrating computer vision risk assessment with classical risk assessment to generate underwriting scores, pricing, and recommendations;

[0047] Figures 22 and 23 show non-limiting examples of monitoring performed by exemplary systems as shown herein;

[0048] Figure 24 illustrates an exemplary dashboard display with a line graph showing temperature data tracked over time in an industrial environment;

[0049] Figures 25 A and 25B show a system for monitoring cleanliness in industrial environments which may track cleanliness ratings for multiple cameras over time;

[0050] Figure 26 illustrates a thermal image view of a monitoring system showing multiple temperature measurement regions; and

[0051] Figure 27 displays an exemplary dashboard visualization of the comprehensive risk scoring system used for fire risk assessment.

[0052] Detailed Description

[0053] The following description sets forth exemplary aspects of the present invention. It should be recognized, however, that such description is not intended as a limitation on the scope of the present invention. Rather, the description also encompasses combinations and modifications to those exemplary aspects described herein.

[0054] Although reference is made herein to “computer vision” and to “cameras”, it is understood that the present invention may be extended to include other types of sensors, and / or to replace one or more cameras with one or more other types of sensors.

[0055] The present invention relates to a system and method for monitoring and controlling safety risks in indoor industrial environments by determining deviations from an ideal operational state using computer vision techniques and optionally game-theoretic competitive ranking frameworks. Computer vision systems, optionally using infrared and visible light sensors, enhance various types of risk assessments, including fire risk assessments, by monitoring facilities. The system preferably comprises a plurality of cameras physically positioned throughout an indoor industrial environment, a processor, and a memory storing instructions.

[0056] The camera network optionally includes multiple types of sensors to capture different aspects of the industrial environment. An infrared sensor (Thermal Camera) is optionally used to monitor critical and potentially high-risk areas in a facility. Infrared sensors are optionally selected due to their ability to sense heat without making direct contact like traditional point detectors such as thermocouples or thermistors. The sensor preferably monitors the facility continuously, and preferably provides temperature data for every pixel in the thermal image multiple times per second. Infrared cameras measure temperature which, when combined with statistical processing algorithms, may be used to detect anomalies. The thermal camera provides temperature data that may be used to calculate metrics such as minimum, maximum, and average temperatures for various regions within an image.

[0057] A visible light sensor (Standard Camera) is preferably used to monitor operations within the facility. On the visible light spectrum, it is possible to visualize elements including but not limited to dust, smoke, fire, occupancy, cleanliness, safety equipment, hazardous materials, or potential intruders. Visible light sensors capture images which may contain signs of smoke, fire, and various status of operation such as maintenance, housekeeping, or production. Computer vision algorithms may be employed to process these images to determine if a sufficient risk is present.

[0058] The system may evaluate images to detect the presence of smoke, which may indicate an incipient fire. Smoke detection may leverage energy-based image analysis that identifies regions where texture and contrast have been reduced — a common visual signature of smoke presence. These candidate regions undergo classification through trained neural networks with temporal consistency checks to reduce false positives. For fire detection, the system employs a sequential process that includes motion detection via background subtraction, color filtering for fire-typical hues, region proposal for areas of interest, and neural network classification.

[0059] The system preferably monitors various aspects of the industrial environment, including machines, processes, and specific areas. The cameras may be strategically positioned to observe critical operational zones, production equipment, storage areas, and other locations where safety risks may arise. When executed by the processor, the stored instructions may operate a game-theoretic competitive ranking framework wherein each camera functions as a player and each captured image represents a move made by the respective camera, to analyze risks related to quality, and to preferably convert them to quantifiable risks. This framework is configured to identify physical deviations from predetermined operational safety parameters. The data from these sensors are analyzed using computer systems. These analyses, which may be augmented with artificial intelligence, inform a unified risk score for insurance underwriting, distinguishing between low-risk (clean and well-maintained) and high-risk (dirty with maintenance issues) facilities. The system incorporates various computer vision modules to analyze different aspects of risk.

[0060] The system preferably assesses cleanliness levels within the facility using a game- theoretic competitive ranking framework. This approach models cleanliness assessment as a competitive ranking system inspired by the Elo rating methodology. In this framework, each camera in the facility functions as a "player" in the system, with captured images representing "moves" or "strategies" that document the visual state of monitored areas. The system orchestrates head-to-head "matches" between images, comparing both current conditions across different facility locations and historical conditions at the same location over time. Additional modules may include temperature analysis, intruder detection, and hazardous materials identification, all contributing to the overall risk assessment.

[0061] The system may integrate traditional risk assessment approaches to provide a comprehensive evaluation. Classical risk matrices may be incorporated to evaluate the severity and likelihood of potential hazards. The system supports risk management protocols by identifying areas requiring attention and suggesting mitigation strategies. Data from previous incidents and losses may be incorporated into the risk assessment framework. The system may consider compliance with relevant building codes and safety regulations.

[0062] The system preferably generates scores that quantify the level of risk. Computer Vision (CV) Scores are derived from the analysis of sensor data using the various computer vision modules. The system integrates multiple computer vision scores, including fire detection, smoke detection, and cleanliness assessment, into a unified CV Score. The Classical Risk Assessment (CRA) Score represents the baseline risk evaluation based on traditional risk assessment methods. Traditional risk assessment models may be used as a baseline score, which is then adjusted based on the day-to-day conditions and events detected by the system.

[0063] The unified risk score may be used to inform insurance underwriting decisions, potentially enabling more dynamic and responsive insurance pricing models. The fire risk scoring mechanism may be configured to dynamically adjust insurance premiums based on day-to-day operations, rewarding clients who maintain clean and well-maintained facilities and penalizing those who fail to take adequate prevention measures. By providing continuous monitoring rather than periodic inspections, the system offers insurers a more accurate representation of day-to-day operational conditions that influence safety risks.

[0064] The present invention may function as a telematics device focused on risk assessment, prevention of catastrophic losses, worker safety, and ultimately facility management for industrial manufacturers. The camera network includes multiple modalities for comprehensive monitoring. Infrared / Thermal Cameras capture temperature data and are particularly effective at identifying hotspots or temperature anomalies that may indicate potential fire risks. Visible Light / RGB Cameras capture standard visual data, allowing for the detection of smoke, fire, cleanliness conditions, and other visual indicators of risk. The Computer Processing System (CPS) processes information from the camera sensors and relays this information to a higher-level data storage system such as a cloud or on-premise server for later analysis.

[0065] Using the temperature data from the Thermal Camera, the CPS may calculate metrics such as min, max, average temperatures for various regions of interest within an image. Metrics from the Thermal Camera are periodically uploaded to a cloud, or on-premise storage system for long-term trend analysis. Comparisons of these metrics may be made against predefined or automatically learned thresholds. In the event of temperature threshold exceedance, alarms may be triggered and forwarded to a response team. Additionally, the alarm conditions are sent to a cloud, or on-premise storage system for later inclusion in a risk assessment score.

[0066] Images from the Standard Camera may be processed to determine if a sufficient fire risk is present. Computer vision algorithms may be used to determine if smoke, fire, or unwanted personnel are present. In order to determine cleanliness levels within a facility, learning-based algorithms may be used to determine acceptable levels of cleanliness.

[0067] Infrared sensors may be used to monitor temperature variations in industrial environments. These thermal imaging systems capture thermal image views that display temperature measurements across different regions of industrial equipment or machinery. The system calculates various temperature metrics including minimum, maximum, and average temperatures for different regions, temperature trends over time, and comparison against thresholds for anomaly detection. Temperature monitoring over time allows for trend analysis, early detection of temperature anomalies, and proactive maintenance scheduling based on temperature patterns. The clear visualization of temperature thresholds in relation to actual measurements facilitates quick identification of potential overheating risks or process deviations.

[0068] The system preferably employs sophisticated algorithms to detect fire and smoke in various environmental conditions. For fire detection, the system optionally employs a sequential process that includes motion detection via background subtraction, color filtering for fire-typical hues, region proposal for areas of interest, and neural network classification. The fire detection process begins with background subtraction to detect motion, followed by color filtering to identify pixels or regions with colors characteristic of fire. The system then extracts regions of interest from the filtered frames, which are processed through neural network analysis. The neural network is preferably trained to detect visual patterns associated with fire or other hazardous conditions.

[0069] Similarly, smoke detection preferably leverages energy-based image analysis including image energy calculation using Sobel filtering, energy drop detection compared to reference background, bounding box generation around regions of interest, neural network classification, and temporal consistency checking. The smoke detection process begins with calculating image energy using Sobel filtering, which computes the spatial gradient intensity of an image. The current frame's energy map is compared to a reference background frame, with regions showing significant energy drops considered candidates for smoke detection. Bounding boxes are created around these regions, which are then analyzed by a neural network classifier. The results are tracked over time to identify persistent patterns, with alerts generated for sustained detections.

[0070] The system optionally uses convolutional neural networks (CNNs) for various detection tasks. The fire detection system uses a neural network architecture such as EfficientNet for classification. EfficientNet is a powerful and computationally efficient family of convolutional neural networks that employs a compound scaling method to balance network depth, width, and input resolution. The classification process is optimized through compound scaling that balances network depth, width, and input resolution to achieve high accuracy while maintaining computational efficiency. In the non-limiting example described herein, the network was trained on millions of images scraped from the internet and synthetic images containing scenes of fire and smoke. A random set of augmentations may optionally be applied to each data point to reduce overfitting on the training dataset, upsample minority classes, and reduce the transfer gap to real cameras. These augmentations included artificial noise, geometric transformations, color and exposure adjustments, compression artifact simulation, and Mixup / Cutmix augmentations. The network then outputs a confidence score indicating the likelihood that the region contains fire. Each region's score is stored in a temporal buffer to track confidence scores over time, enabling analysis of temporal patterns or persistence of detected features.

[0071] The smoke detection system optionally similarly uses advanced neural networks, but with specific preprocessing steps optimized for smoke characteristics including energy calculation via Sobel filtering, energy drop detection to identify regions where smoke may be present, neural network classification of candidate regions, and temporal consistency checking to reduce false positives. The smoke detection algorithm analyzes visual characteristics such as opacity, color, and movement patterns to differentiate smoke from other visual phenomena. The confidence scores reflect the algorithm's assessment of these characteristics in determining the likelihood of smoke presence.

[0072] The system optionally further features the incorporation of a game theory -based competitive ranking system for assessing facility cleanliness, for example as part of overall fire risk evaluation and / or other types of risk evaluation (including but not limited to risk of accidents, injuries and the like). This approach models cleanliness assessment as a competitive ranking system inspired by the Elo rating methodology. In this framework, each camera in the facility functions as a "player" in the system, with captured images representing "moves" or "strategies" that document the visual state of monitored areas. The system orchestrates head-to-head "matches" between images, comparing both current conditions across different facility locations and historical conditions at the same location over time. These comparisons yield outcomes (win, loss, or draw) that update each camera's cleanliness rating accordingly.

[0073] The classic Elo system is optionally used to quantify the relative strength of players (in this case, cameras) based on the outcome of a comparison. Each camera is assigned a rating (e.g., 1500 baseline). When two sets of images are compared, the expected outcome is calculated using a logistic function. For each match, the outcome is determined through evaluation by either an advanced computer vision model or a trained human expert. The evaluator assesses which image appears cleaner based on visual indicators such as dust accumulation, waste materials, spills, or general tidiness.

[0074] Areas consistently determined to be clean receive higher ratings, while declining ratings may signal deteriorating conditions that correlate with increased fire risk. The game- theoretic framework avoids fixed thresholds or manual scoring systems, which may struggle with consistency and objectivity. Instead, cleanliness is assessed through relative performance over time via structured competitive evaluation. This enables more sophisticated analyses, including trend identification, cross-area benchmarking, and the establishment of facility-wide cleanliness standards.

[0075] The system may also monitor for unauthorized personnel or intruders, which can represent a significant safety risk. This includes motion detection to identify movement in restricted areas, human detection and recognition algorithms, alert generation for unauthorized presence, and integration with access control systems. Employee personnel or intruders may play a substantial role in fire risk. Arson is a serious concern across the entire insurance industry — be it from insurance fraud or vandalism. With a combination of infrared and standard camera sensors, it is possible to detect individuals lighting a fire, smoking, or entering forbidden areas.

[0076] Motion and / or human recognition algorithms may be used to identify individuals and alert necessary personnel. When individuals enter areas where, or when they are not supposed to, these detections may be included in an Intruder Score component of the overall risk assessment. The system can identify potentially hazardous materials that may increase fire risk through object detection algorithms to identify gas cylinders, fuel containers, etc., classification of detected objects based on risk level, and integration of hazardous material detections into the overall risk score.

[0077] Hazardous materials such as gas cylinders for welding, gas canisters, open chemicals, unattended batteries, running or broken vehicles, or oils may be a major cause of accidents and fire. Information from the standard cameras may be used to capture images of these items and alert necessary personnel. Machine learning algorithms like classifiers or single-shot detectors may be trained to recognize these objects using openly available datasets or data from individual facilities. Leaking gas cylinders, such as those used in welding, can also be detected using infrared sensors. Detections may be logged and assessed by facility personnel. These detections may be included in a Hazardous Materials Score component of the overall risk assessment.

[0078] The system preferably integrates multiple computer vision scores into a unified CV Score. These components are weighted according to industry-specific risk factors, with weights determined based on historical loss data and insurance underwriting insights. The scoring framework uses parameters (sigma), weights (alpha), and bias terms (beta) to calculate component scores and the overall risk score. This includes temperature scoring based on temperature anomalies and threshold exceedances, fire scoring based on the number and confidence of fire detections, smoke scoring based on smoke detection events, and cleanliness scoring based on the game-theoretic cleanliness assessment. Each of these factors (sigma) may be combined in aggregate via a weighting function composed of a weight (alpha) and a bias term (beta). The result of this aggregation is a comprehensive risk score.

[0079] The Classical Risk Assessment (CRA) provides a baseline risk score based on traditional risk evaluation methods, which may include periodic inspections by qualified risk engineers, evaluation of safety systems (sprinklers, alarms, etc.), building code compliance assessment, and historical incident data analysis. The CV Scores are then integrated with the CRA Score to generate a comprehensive risk assessment.

[0080] Traditional risk assessment typically involves periodic inspections by qualified risk engineers who evaluate various factors including environmental state, cleanliness, occupancy, electrical installations, and other safety-related parameters. However, these periodic assessments provide only a snapshot of conditions at a specific moment in time, rather than a continuous representation of day-to-day operations. Facilities are often notified in advance of inspections, allowing them to temporarily optimize conditions that may not reflect normal operations. In contrast, the present invention is able to provide continuous 24 / 7 monitoring, enabling real-time assessment of safety conditions and early detection of risks. This ongoing assessment allows for more accurate representations of day-to-day operating conditions and fire risks.

[0081] The risk scores derived from these monitoring systems may be seamlessly incorporated into existing insurance pricing frameworks, creating a more dynamic and responsive risk assessment model. While the specific algorithms for translating these scores into premium adjustments may remain the responsibility of individual insurers, the availability of consistent, quantifiable data addresses a critical gap in traditional risk evaluations. By implementing these advanced monitoring solutions, insurance companies may establish a more collaborative relationship with policyholders centered around proactive risk management rather than reactive claims processing. The systems provide facility managers with actionable insights to improve safety protocols, creating a positive feedback loop where improved conditions lead to lower risk scores, which may then translate to reduced premiums through the insurer's pricing model.

[0082] The present invention may be implemented through systems that serve as a plug-in upgrade for existing IP camera networks. This on-premise server equipped with specialized GPUs processes video feeds from existing cameras, applying the fire and smoke detection algorithms while periodically uploading snapshot images for cleanliness assessment. When potential incidents are detected, the system records video of the event and securely uploads it to cloud storage for later review and analysis. This architecture balances edge computing for real-time detection with cloud-based storage and processing for long-term analysis and trend identification.

[0083] The system connects to a video management system (VMS) that displays the processed video data on a computer monitor, allowing for remote monitoring and quick response to potential safety risks detected by the system. In addition to the wired components, the system may include mobile device connectivity that receives alerts and notifications, allowing for quick response to potential fire risks. The system may also incorporate additional monitoring capabilities beyond video analysis, such as dedicated smoke and fire alert sensors that work in conjunction with the video-based detection to provide a more comprehensive fire risk assessment.

[0084] The approaches described herein may be applicable across a range of industries and facility types. In sawmills, accumulated sawdust can create a significant fire hazard. The system can continuously monitor cleanliness levels and detect areas where sawdust is accumulating, enabling timely cleaning before conditions become hazardous. In recycling facilities, which often handle combustible materials and may have machinery that generates heat, the system can monitor for temperature anomalies, accumulated debris, and other fire risk factors. Food processing facilities often generate combustible dust as a byproduct of production. The system can monitor dust levels and identify areas requiring additional cleaning efforts. The system may be particularly beneficial for environments with elevated fire risks or where traditional fire detection methods face limitations, such as chemical processing plants, battery manufacturing facilities, steel production plants, and agricultural processing facilities.

[0085] The system provides comprehensive risk assessment capabilities, including temperature monitoring and anomaly detection, fire and smoke detection at early stages, cleanliness assessment using the game-theoretic framework, identification of hazardous materials and unauthorized personnel, and long-term trend analysis for predictive risk assessment. By providing more comprehensive and timely fire risk data, the disclosed techniques may enable improved safety protocols, more informed insurance underwriting, and potentially reduced premiums for facilities that maintain lower risk profiles. The systems and methods may support a shift towards more proactive and data-driven fire prevention strategies aligned with insurance industry requirements, such as cleaning plans and monitoring that can directly influence premium calculations.

[0086] Companies are required to have a cleaning plan that defines the cleaning intervals necessary for risk reduction. This plan must be monitored, and a designated person must ensure that the cleaning is carried out properly. The system provides objective data to verify compliance with these requirements, which can have a direct influence on insurance premiums.

[0087] Turning now to the drawings, Figures 1A-1D provide an illustrative overview of how a network of cameras capable of sensing various light spectrums may be used to provide a quantitative risk assessment score based on day-to-day operations within a facility.

[0088] Without wishing to be limited by a single description, the overall system and method of Figures 1A-1D may be described as providing a monitoring system as follows. The monitoring system preferably employs a sophisticated network of cameras capable of sensing various light spectrums to provide a quantitative risk assessment based on day-to-day operations within a facility. This system represents an integrated approach to risk monitoring and assessment, combining advanced technological capabilities with comprehensive analytical methodologies.

[0089] The system preferably uses a camera network to capture data from different areas of observation throughout a facility. For example and without limitation, the cameras may look at machines, processes, loading zones, or various other areas / equipment common in manufacturing or industrial facilities. The cameras capture information based on various light spectrums. This information is then processed by a series of various computer vision algorithms which can calculate independent forms of risk, as described in greater detail below.

[0090] These cameras optionally and preferably incorporate both infrared and visible light sensors, enabling the collection of multi-spectral data that provides a more complete picture of conditions within the monitored environment. However alternatively only infrared or visible light sensors are used. The data captured by these sensors is transmitted to a computer vision processor for detailed analysis.

[0091] The computer vision processor optionally and preferably contains multiple specialized modules designed to process different aspects of the captured data. These include modules for detecting fire, smoke, and sparks; monitoring cleanliness levels; identifying hazardous elements or conditions; and detecting unauthorized personnel or intruders. Additionally, a temperature monitoring module processes thermal imagery to identify temperature anomalies or hotspots that might indicate potential issues.

[0092] Working in parallel with the computer vision analysis, a comprehensive risk assessment module preferably evaluates various risk factors related to the facility. This module considers environmental factors associated with the geographic location of the facility, analyzes historical events or incidents, assesses fire safety measures and protocols, evaluates cleanliness standards and maintenance practices, and examines compliance with relevant building codes and regulations. This thorough approach to risk assessment incorporates multiple methodologies to provide a comprehensive evaluation of potential risks.

[0093] The outputs from both the computer vision processor and the risk assessment module preferably feed into a scoring mechanism. The computer vision score is derived from individual component scores related to temperature, fire / smoke / spark detection, cleanliness, hazardous conditions, and intruder detection. Similarly, the risk assessment module produces its own score based on its evaluation of risk factors.

[0094] These scores are then preferably integrated by a combined scoring module, which aggregates the data to produce a comprehensive evaluation of risk. This combined score is transmitted to an underwriting module, which can use this information to make informed decisions about risk management and insurance underwriting.

[0095] Without wishing to be limited by a closed list, by combining advanced computer vision techniques with traditional risk assessment methodologies, this integrated system provides a more comprehensive and accurate evaluation of potential risks in various industrial and commercial settings. The ability to leverage both real-time monitoring through the camera network and historical data analysis offers a robust solution for risk management and loss prevention.

[0096] Turning back to the drawings, Figure 1A illustrates a monitoring system 100. The monitoring system 100 may use a network of cameras capable of sensing various light spectrums. In some cases, the monitoring system 100 includes a camera network 102. The camera network 102 may capture data from different areas of observation, including a machine 104, a process 106, and an area 108.

[0097] The camera network 102 may include infrared and visible light sensors. In some cases, the camera network 102 provides outputs to capture both infrared light 110 and visible light 112. These outputs may be transmitted to a computer vision processor 114, which analyzes the captured data.

[0098] The computer vision processor 114 may generate a computer vision score 116. In parallel, a risk assessment module 118 may produce a classical risk assessment score 120. Both the computer vision score 116 and the classical risk assessment score 120 may be input into a combined scoring module 122. Combined scoring module 122 may aggregate these CV Scores using a weighted formula. Optionally this task is performed by computer vision processor 114, risk assessment module 118 and / or risk assessment score 120. The weighted formula may include weights (alpha), factors (sigma), and a bias term (beta). The weights may be assigned based on the relative impact of each risk factor on the overall fire risk. The factors may represent the quantified measures of the risk factors, and the bias term may be used to adjust the final score. The result of this aggregation may be a unified risk score, which may provide a comprehensive measure of the fire risk and / or other risks based on the detected risk factors.

[0099] For example, the weighted formula for aggregating risk factors may employ a mathematical framework that combines multiple component scores into a unified risk assessment. This aggregation model follows the general form R = S(ai x oi) + . where R represents the final risk score, oi denotes the individual risk factors (such as temperature anomalies, fire detection events, cleanliness ratings, hazardous material detections, and intruder identifications), ai represents the corresponding weight coefficients that reflect the relative importance of each factor, and is a bias term that calibrates the baseline risk level according to industry-specific standards. The weight coefficients are preferably determined through a combination of historical loss data analysis, domain expert input, and machine learning techniques that identify correlations between observed risk factors and actual incident occurrences. For industrial environments with extensive historical data, these weights may be dynamically adjusted using Bayesian updating mechanisms that refine the relative importance of different risk factors based on emerging patterns specific to the facility, its operational characteristics, and its industry classification.

[0100] The implementation of this weighted formula optionally and preferably incorporates adaptive normalization to ensure consistent interpretation of diverse risk factors across different scales and distributions. Each raw risk factor oi undergoes preprocessing through a scaling function that maps its values to a standardized range (typically 0-100), accounting for the specific statistical properties of that factor's distribution. For example, temperature data may be normalized using a sigmoid function that emphasizes deviations from expected operational ranges, while cleanliness scores derived from the game-theoretic framework maintain their inherent 0-100 scale. The bias term P serves multiple purposes within the formula: it establishes a minimum risk baseline appropriate for the facility type, adjusts for known but unquantifiable risk elements not directly captured by the monitoring systems, and compensates for regional variations in risk profiles.

[0101] The combined scoring module 122 may process these inputs and transmit the results to an underwriting module 124. The underwriting module 124 may represent the final component in the monitoring system 100's processing chain.

[0102] In some cases, the monitoring system 100 includes an AVIAN Vision plug-in upgrade for existing IP camera networks. This upgrade may enhance the capabilities of the camera network 102 for specific applications such as fire and smoke detection.

[0103] The system components may be arranged in a sequential flow, with data moving from the initial capture by the camera network 102 through various processing stages to the final underwriting module 124. The connections between components may indicate the flow of information through the monitoring system 100.

[0104] FIG. IB illustrates a block diagram of a computer vision processor 114. The computer vision processor 114 may include multiple monitoring modules that process different aspects of sensor data.

[0105] A temperature monitoring module 150 may process temperature-related data. In some cases, the temperature monitoring module 150 analyzes thermal imagery to detect temperature anomalies or hotspots within monitored areas.

[0106] A fire / smoke / sparks (FSS) detection module 152 may analyze data to detect fire, smoke or spark events. The FSS detection module 152 may use a multi-stage fire detection pipeline including motion detection, color filtering, and region proposal. In some cases, motion detection via background subtraction may be used to identify areas of activity. Color filtering may then be applied to isolate pixels matching fire or smoke characteristics. Region proposal techniques may generate candidate areas for further analysis.

[0107] A cleanliness monitoring module 154 may process data to assess cleanliness levels. In some cases, the cleanliness monitoring module 154 analyzes visual imagery to detect dust, debris, or other indicators of cleanliness status.

[0108] A hazardous elements monitoring module 156 may analyze data to detect hazardous materials or conditions. The hazardous elements monitoring module 156 may use computer vision techniques to identify potentially dangerous objects or situations within monitored areas.

[0109] An intruder detection module 158 may process data to detect unauthorized personnel or intruders. In some cases, the intruder detection module 158 uses motion detection and object classification to identify human presence in restricted areas.

[0110] Employee personnel or intruders may also play a substantial role in fire risk. Arson is a serious concern across the entire insurance industry - be it from an insurance fraud or vandalism perspective. With a combination of infrared and standard camera sensors, it is possible to detect individuals lighting a fire, smoking, or entering forbidden areas. Motion and or human recognition algorithms may be used to identify individuals and alert necessary personnel.

[0111] The computer vision processor 114 may use various algorithms to analyze sensor data, including convolutional neural networks, vision transformer networks, isolation forest, semantic segmentation, generative adversarial networks, reinforcement learning with human feedback, autoencoders, or K-Nearest Neighbors. These algorithms may be applied across the different monitoring modules to extract relevant features and make detections.

[0112] In some cases, synthetic data generation techniques may be used to increase dataset size and variety for training the detection algorithms. Generative adversarial networks may be employed to create artificial but realistic training examples, augmenting the available real- world data.

[0113] The modules within the computer vision processor 114 may work together to provide comprehensive analysis of sensor inputs. For example, temperature data from the temperature monitoring module 150 may be correlated with visual detections from the FSS detection module 152 to improve fire detection accuracy. The outputs from these modules may be combined to generate overall assessments of monitored conditions.

[0114] Figure 1C illustrates a block diagram showing the components of a computer vision scoring system. A computer vision score 116 may be derived from multiple individual scoring components. These components may include a temperature score 160, which evaluates temperature-related data, and a fire / smoke / spark (FSS) score 162, which assesses detection of smoke, fire and spark events. The system may also incorporate a cleanliness score 164 that quantifies cleanliness levels, a hazardous score 166 that evaluates hazardous conditions, and an intruders score 168 that tracks unauthorized access detection.

[0115] In some cases, the computer vision score 116 may be calculated using a weighted formula. The weighted formula may combine the individual scores, with each score potentially having a different weight based on its relative importance or relevance to the overall assessment. For example, in a manufacturing environment where fire risk is a primary concern, the FSS score 162 may be given a higher weight compared to other scores.

[0116] The temperature score 160 may contribute to the computer vision score 116 by providing information about thermal conditions in monitored areas. Abnormal temperature patterns detected and quantified by the temperature score 160 may indicate potential issues or risks.

[0117] The FSS score 162 may play a role in fire safety assessment. This score may reflect the system's ability to detect early signs of fire, smoke, or sparks, potentially allowing for rapid response to fire-related incidents.

[0118] The cleanliness score 164 may provide insights into the overall cleanliness and maintenance of the monitored environment. Optionally, the cleanliness monitoring module may use a game-based assessment approach with Elo-based scoring, as described in greater detail below. This approach may involve comparing cleanliness levels between different areas or time periods, with each comparison treated as a "game" that affects the overall cleanliness rating.

[0119] The hazardous score 166 may contribute to the computer vision score 116 by quantifying the presence or likelihood of hazardous conditions. This may include detection of dangerous materials, unsafe practices, or environmental hazards.

[0120] The intruders score 168 may add a security dimension to the overall assessment. By tracking and quantifying unauthorized access or presence, this score may help in evaluating security risks within the monitored area.

[0121] By combining these diverse scoring components, the computer vision score 116 may provide a comprehensive assessment of various aspects of safety, security, and operational conditions within the monitored environment. The weighted formula approach may allow for flexibility in adjusting the relative importance of different factors based on specific needs or risk profiles of different environments or industries. Figure ID illustrates a risk assessment module 118. The risk assessment module 118 may include multiple components for evaluating different aspects of risk.

[0122] A geographic location 170 component may consider the physical location and environmental factors of a facility. In some cases, the geographic location 170 component may take into account factors such as proximity to natural hazards, climate conditions, or accessibility for emergency services.

[0123] A past incidents 172 component may analyze previous events or incidents that have occurred at the facility or similar facilities in the industry. The past incidents 172 component may help identify patterns or recurring issues that could impact risk assessment.

[0124] A fire safety 174 component may evaluate fire prevention and protection measures in place at the facility. In some cases, the fire safety 174 component may assess factors such as the presence and condition of fire suppression systems, evacuation procedures, or fire- resistant building materials.

[0125] A cleanliness 176 component may assess the cleanliness conditions and maintenance standards of the facility. The cleanliness 176 component may contribute to risk assessment by considering how well the facility is maintained and whether proper housekeeping practices are followed.

[0126] A building code 178 component may examine compliance with relevant building regulations and standards. In some cases, the building code 178 component may evaluate whether the facility meets current safety and construction requirements.

[0127] The risk assessment module 118 may take into account long term temperature trend analysis across multiple thermal cameras. This analysis may provide insights into temperature patterns and anomalies over time, which may be used to assess potential risks related to equipment performance or environmental conditions.

[0128] The risk assessment module 118 may adapt to different industries through specific mechanisms. For example, in wood processing industries, the risk assessment module 118 may place greater emphasis on factors related to combustible dust accumulation and ignition sources. In food manufacturing industries, the risk assessment module 118 may focus more on hygiene standards and temperature control for perishable goods. These industry-specific adaptations may allow the risk assessment module 118 to provide more relevant and accurate risk evaluations for different types of facilities.

[0129] Optionally, Underwriting Module 124 may adjust premium prices based on the assessed risk levels. For instance, facilities that maintain clean and well-maintained environments and exhibit fewer risk factors may be rewarded with lower insurance premiums. Conversely, facilities that fail to take adequate prevention measures and exhibit a higher number of risk factors may be penalized with higher insurance premiums. This dynamic adjustment of insurance premiums may provide a more accurate reflection of the actual fire risk within a facility.

[0130] The dynamic adjustment of insurance premiums based on quantified fire risk is preferably implemented through a multi-tiered algorithmic framework within the underwriting module. This framework establishes a mathematical correlation between the unified risk score and premium calculations using a sensitivity matrix that maps score ranges to premium modification factors.

[0131] For example, the underwriting module may maintain a sliding window of historical risk scores, typically spanning 30-90 days depending on the facility type, which are processed through a weighted averaging function that prioritizes recent data while accounting for longer-term trends. This temporal analysis enables the system to differentiate between transient risk fluctuations and persistent operational changes. Premium adjustments are preferably triggered when either: (1) the rolling average risk score crosses predefined threshold boundaries (typically set at 5-point intervals on the 0-100 scale), or (2) when consistent directional movement in risk scores is detected over three consecutive assessment periods. The magnitude of premium adjustments optionally and preferably follows a nonlinear function where each 10-point improvement in risk score below the baseline typically correlates to a 5-15% premium reduction, while deterioration in scores triggers proportionally larger increases to incentivize risk mitigation.

[0132] The system preferably further incorporates facility-specific calibration factors to normalize premium adjustments across different industrial environments. For example, in sawmills where baseline fire risk is inherently higher, the premium sensitivity to cleanliness scores is amplified compared to other risk factors, with a 20-point improvement in cleanliness potentially yielding up to a 25% premium reduction. Conversely, for chemical processing facilities, temperature anomaly detection carries greater weight, where maintaining temperatures within optimal ranges for three consecutive months might result in a 10-12% premium discount.

[0133] The underwriting module also optionally and preferably features an exception handling process for sudden significant risk changes, such as the detection of critical hazardous materials or repeated high-temperature events, which can trigger immediate premium reassessment rather than waiting for the next scheduled adjustment period. This approach supports insurance pricing responsiveness to both gradual operational improvements and acute risk developments, creating a financially motivated feedback loop that encourages proactive safety measures and continuous improvement in fire risk management practices.

[0134] Figures 2A, 2B and 3 relate to non-limiting, illustrative examples of networks, sensors and systems for receiving data for risk determination.

[0135] Figure 2A illustrates an exemplary network system for processing and managing camera data. The network system 200 includes a camera network 102 that connects to a network switch 204. The network switch 204 distributes data to multiple components including a compute server 202 and a video management system 208.

[0136] The compute server 202 may process data received from the camera network 102 via the network switch 204. The network switch 204 may also connect to a cloud network 206, allowing for remote data access and storage capabilities.

[0137] The video management system 208 may receive camera data through the network switch 204 and may interface with recording storage 210 for storing video data. The video management system 208 may also connect to display devices 212, enabling visualization of camera feeds and recorded content.

[0138] In some cases, the network system 200 may follow a centralized topology with the network switch 204 acting as a distribution point for data flow between components. The camera network 102 may serve as the input source, while the compute server 202, cloud network 206, and video management system 208 may handle different aspects of data processing, storage, and display functionality.

[0139] The camera network 102 may capture both visible and infrared light data, which may be transmitted through the network switch 204 to the compute server 202 for analysis. The compute server 202 may run computer vision algorithms to process the camera data and generate insights.

[0140] In some implementations, the cloud network 206 may provide additional computational resources or serve as a backup storage solution. The video management system 208 may allow operators to view live feeds, replay recorded footage, and manage camera settings through the display devices 212.

[0141] The recording storage 210 may utilize various storage technologies such as hard disk drives, solid-state drives, or network-attached storage to maintain a historical record of camera data. This stored data may be accessed for post-event analysis or used to train and improve the system's detection algorithms over time. Figure 2B illustrates an exemplary, non-limiting network for video management and fire risk assessment. A network 250 includes multiple interconnected components that work together to capture, process, and analyze video data for fire detection and risk assessment purposes.

[0142] The network 250 flow preferably begins with a standard IP camera 252 that captures visual data from the monitored environment. This camera connects to a Power over Ethernet (PoE) switch 254, which provides both power and network connectivity to the camera.

[0143] From the PoE switch, the video data is routed to a server 256, optionally located onsite, which may for example comprise an Al server as shown. Server 256 optionally serves as the primary processing hub for the network, analyzing the incoming video feeds using artificial intelligence algorithms designed for fire detection and risk assessment.

[0144] Server 256 preferably connects to a video management system (VMS) 258 that displays the processed video data on a computer monitor.

[0145] In addition to the wired components, network 250 preferably includes a mobile device 260 that receives alerts and notifications. This allows for remote monitoring and quick response to potential fire risks detected by the system.

[0146] Network 250 may also incorporate a smoke and fire alert sensor 262, which provides additional monitoring capabilities beyond video analysis. This sensor may work in conjunction with the video-based detection to provide a more comprehensive fire risk assessment.

[0147] The network architecture of network 250 preferably follows a centralized topology, with server 256 acting as the core processing unit. This design enables efficient video data collection, processing, and alert distribution across the connected devices.

[0148] The system utilizes both wired connections for high-bandwidth video feeds and wireless communication for mobile alerts. This hybrid approach allows for real-time monitoring of the environment while providing flexibility in how alerts and notifications are delivered to relevant personnel.

[0149] In operation, network 250 preferably continuously monitors the environment through the IP camera 252. The video feed is processed by server 256, which analyzes the imagery for signs of fire or potential fire risks. When a risk is detected, the system may generate alerts that are sent to the mobile device 260 and displayed on the VMS interface.

[0150] The optional integration of the smoke and fire alert sensor 262 provides an additional layer of detection capability. This sensor may work independently or in conjunction with the video analysis to trigger alerts when smoke or fire is detected. By combining video surveillance, artificial intelligence processing, and traditional sensor technology, network 250 creates a comprehensive system for fire risk assessment and management. The centralized processing and distributed alert capabilities allow for rapid response to potential fire hazards in the monitored environment.

[0151] Figure 3 illustrates a block diagram of an exemplary, optionally implementation of a camera for use with the systems and methods as described herein. The camera 300 may include multiple components.

[0152] The camera 300 may contain an infrared light sensor 302 for capturing thermal information and a visible light sensor 304 for capturing visible spectrum data. These sensors may work in tandem to provide comprehensive visual information across different light spectrums.

[0153] A microphone 306 may be included for audio capture capabilities. The microphone 306 may allow the camera to detect and record sound, which may be useful for identifying certain events or conditions.

[0154] The camera 300 may incorporate a compute module 308 for processing the captured data. This module may analyze inputs from the sensors and microphone, potentially running algorithms for detection and classification tasks.

[0155] A network interface 310 may enable communication with external systems and networks. This interface may allow the camera to transmit processed data, receive updates, or integrate with larger monitoring systems.

[0156] The camera 300 may include an output 312 for transmitting processed data. This output may send information to other components of the monitoring system or to display devices for visualization.

[0157] An alarm 314 may be provided for generating alerts or notifications based on detected conditions. The alarm may be triggered by the compute module based on analysis of sensor data or received instructions from the network.

[0158] The components within the camera 300 may interact in various ways. For example, data from the infrared light sensor 302 and visible light sensor 304 may be combined in the compute module 308 for enhanced event detection. The compute module 308 may process audio from the microphone 306 in conjunction with visual data for more comprehensive monitoring.

[0159] When the compute module 308 detects a condition of interest, it may send an alert through the network interface 310, activate the alarm 314, and transmit relevant data through the output 312. This integrated approach may allow the camera 300 to function as a self- contained monitoring unit within a larger surveillance system.

[0160] As noted above, the camera system may include a combination of infrared and visible light sensors to capture different types of data, whether these sensors are combined in a single device such as a camera for example, or are separated. The infrared sensor captures temperature data, while the visible light sensor captures images of the facility. This combination of sensors may provide a more comprehensive assessment of the fire risk within a facility.

[0161] A camera equipped with an infrared sensor may capture information from light with wavelengths of 780 nm to beyond 14 pm. Infrared sensors are commonly used in thermographic inspection (on the longwave infrared spectrum, 8-14 pm) in order to measure temperatures of a scene without contact. This is achieved by a sensor which can measure the infrared radiation of an object, and then correlate this measurement with a temperature. Infrared images can contain hundreds of thousands of individual temperature measurements.

[0162] Contactless temperature measurements may be preferred for fire prevention, as infrared cameras can oversee large areas. Infrared cameras may be used to detect temperature anomalies before ignition occurs. This enables alerts to be sent long before a traditional smoke alarm system, however this capacity comes with the requirement that the hot object must be within the camera’s field of view. Over time, data from infrared sensors may be used to detect when machines begin to break down such as motors, gearings, belts, etc. This type of long term historical analysis can enable predictive, or preventative maintenance where maintenance is completed before a catastrophic failure occurs which may lead to a higher risk of fire.

[0163] Figures 4-13 relate to non-limiting examples of devices, systems and methods that use computer vision to obtain quantitative data, in order to determine quantifiable risk.

[0164] Turning now to Figure 4, a non-limiting, exemplary monitoring method is shown as a block diagram. The monitoring method 400 may include several interconnected components for processing and analyzing data. The process preferably begins with monitoring parameters, which may include various factors such as the temperature range, the frequency of temperature measurements, and the specific regions within the facility to be monitored. These parameters provide input to data analysis, which may involve processing the temperature data to calculate metrics such as minimum, maximum, and average temperatures for various regions within an image. For example, the flow may begin with an infrared sensor 402, which may capture infrared light information from the monitored environment. This sensor may provide thermal data that can be used for temperature analysis.

[0165] Connected to the infrared sensor 402 is a temperature monitor 404. This component may perform various temperature monitoring functions, which may include measuring minimum, maximum, and average temperatures, analyzing temperature distributions, conducting trend analysis, triggering alarms, and detecting anomalies based on the infrared data received. For example, the data analysis may also involve comparing the calculated metrics against predefined or automatically learned thresholds to identify temperature anomalies.

[0166] For example, temperature monitor 404 may determine temperature anomalies over time (by comparing information from the same sensor or group of sensors) or over space (by comparing information provided from different sensors or sensor groups). For example, the temperature monitoring system may capture essential thermal metrics by measuring minimum, maximum, and average temperatures across various regions within a facility and / or over time. These fundamental measurements provide critical baseline data about the thermal environment, allowing for immediate identification of potential hotspots and cold zones, whether spatially or over time. The minimum temperature identifies the coolest areas within the monitored space, while maximum temperature highlights potential overheating concerns that might indicate equipment malfunction or fire risk. Average temperature calculations offer a broader overview of general thermal conditions, establishing normal operating parameters for different facility zones and equipment types, and / or over time.

[0167] Temperature distribution analysis may be performed by temperature monitor 404 to examine how thermal readings are spread across a statistical range, revealing important patterns beyond simple minimum, maximum, and average values. Narrow distributions with few outliers may indicate stable, predictable thermal environments where temperatures remain consistently within expected parameters. In contrast, wide or long-tailed distributions may signal concerning variability, where unexpected temperature spikes occur with greater frequency. By analyzing these distribution patterns over time, the system can identify subtle shifts in thermal behavior that might escape detection through conventional threshold monitoring, allowing for more nuanced risk assessment that considers not just absolute temperatures but their statistical behavior.

[0168] Temperature monitor 404 may perform trend analysis. Trend analysis examines temperature data over extended time periods to identify gradual shifts, recurring patterns, and emerging anomalies that might not be apparent in snapshot measurements. This longitudinal perspective allows the system to detect slowly developing issues, such as progressive equipment deterioration, seasonal variations, or cyclical operational effects on thermal conditions. By establishing baseline temperature profiles for different operational states and tracking deviations from these expected patterns, trend analysis can provide early warning of potential problems long before they reach critical thresholds. This predictive capability enables proactive maintenance interventions, reducing the risk of catastrophic failures and enhancing overall operational reliability.

[0169] The temperature monitor 404 may feed into an event processor 406. This processor may receive events, metrics, and anomaly data from the temperature monitor. The event processor 406 may analyze and categorize the information, potentially identifying patterns or significant occurrences in the temperature data. The data obtained by event processor 406 may then be fed into a data collection (not shown), which may involve storing the analyzed temperature data for later use. Data aggregation may be performed, as the process of collecting temperature data and channeling it to the analysis step. This may involve aggregating the temperature data over a specified time period or across multiple regions within the facility. The aggregated temperature data may provide a more comprehensive view of the temperature conditions within the facility, enabling more accurate risk assessment.

[0170] Following the event processor 406 is a risk scoring processor 408. This component may implement a temperature risk scoring algorithm, using the processed event data to evaluate potential risks associated with the detected temperature patterns and anomalies.

[0171] The final component in the system is the risk score output 410. This output may present the results of the risk scoring processor, providing a quantified assessment of temperature-related risks based on the analyzed data.

[0172] In this monitoring system 400, data may flow sequentially from the infrared sensor 402 through various stages of processing and analysis, culminating in a risk score that may be used for decision-making or further action.

[0173] Optionally, a risk bias term may be applied to the analyzed temperature data, before or after data aggregation. The risk bias term may be a constant value that is added to the aggregated temperature data to adjust the final risk score. The use of a risk bias term may help to account for inherent biases in the temperature data, such as systematic errors in the temperature measurements or variations in the temperature conditions across different regions within the facility. Risk score output 410 may be calculated based on the aggregated temperature data and the risk bias term. In some aspects, the risk score output may be a numerical value on a scale from 0 to 100, with higher values indicating higher levels of fire risk. The risk score output may be used to inform insurance underwriting decisions, potentially adjusting premium prices based on the assessed risk levels. Minimum, maximum, and average temperatures of a scene may be used as an infrared risk factor. Higher temperatures may indicate a higher level of risk relative to lower temperature environments. Optionally, distributional analysis of the temperature data may be conducted to understand how controlled the temperatures may be on a day-to-day basis. Scenes which exhibit wide, or long-tailed distributions may indicate higher risk when compared to those with narrower distributions with fewer outliers as the temperatures in these cases may be considered to be largely predictable and constant.

[0174] Figures 5 and 6 show flowcharts of non-limiting, illustrative, exemplary algorithms related to temperature monitoring and anomaly detection in a fire risk assessment system. Figure 5 illustrates the Temperature Sensing Algorithm, while Figure 6 depicts the Temperature Anomaly Detection process.

[0175] Figure 5 illustrates a flowchart depicting a method for processing infrared data and detecting temperature anomalies. Such an illustrative flowchart provides an example program logic that takes an infrared image, calculates metrics, and determines if an anomaly is present. Additionally, this program may alert factory personnel for maintenance or emergency purposes. The data is then stored for long term trend analysis, as discussed below.

[0176] A method 500 begins with obtaining infrared data step 502. In this initial stage, the system may acquire infrared data from one or more sensors or cameras monitoring the environment. The infrared data may be captured by an infrared sensor or thermal camera.

[0177] Following the data acquisition, the method proceeds to the convert to temperature step 504. During this phase, the raw infrared data may be transformed into temperature measurements, allowing for more meaningful analysis. For example, this infrared data may then be converted to degrees Celsius, providing a temperature reading for each pixel in the thermal image.

[0178] The next step in the process is to obtain regions of interest and temperature limits 506. This stage may involve defining specific areas within the monitored space that require particular attention, as well as establishing temperature thresholds for these regions.

[0179] Once the regions and limits are established, the method moves to the calculate metrics step 508. In this phase, various temperature-related metrics may be computed based on the converted temperature data and the defined regions of interest, and optionally according to temperature limits. These metrics may include, but are not limited to, minimum, maximum, and average temperatures for various regions within an image.

[0180] The method then enters a decision point, to determine whether a temperature anomaly is detected in step 510. This stage may involve analyzing the calculated metrics for each monitored region to determine if any temperature readings fall outside the expected or acceptable range. Concurrently, the data may be stored for future use, enabling long-term trend analysis and facilitating the detection of temperature anomalies over time (see step 514).

[0181] If a temperature anomaly is detected, the method branches into two parallel paths. One path leads to the send alarm step 512, where the system may generate and transmit alerts or notifications about the detected anomaly.

[0182] Simultaneously, the method proceeds to the store data step 514. This step may involve recording the temperature data, calculated metrics, and any detected anomalies for future reference or analysis. Optionally, the data is stored even if no temperature anomalies are present.

[0183] If no temperature anomaly is detected at the decision point 510, the method bypasses the alarm step and proceeds directly to the store data step 514.

[0184] Trend and historical analysis may be conducted to evaluate how temperatures within a scene change over time. Temperature changes may be due to seasonal or ambient effects, or something more dangerous such as machine degradation. With long term analysis of temperature data, it is possible to understand how an individual motor or gearing has changed over time. For example, if a motor is running 5 degrees Celsius warmer than normal, this may indicate an upcoming bearing or armature failure which can lead to fires. Normal operating characteristics may be defined by looking at a long term time series analysis with the temperature information captured by the infrared sensor.

[0185] An optional parameter that may be obtained with infrared sensors is the determination of the amount of high-temperature events - i.e. an event where a temperature exceeds a threshold. The threshold may be set manually, or learned from historical data. Facilities where there are frequent high-temperature events may be considered to have much higher risk than those which stay within the normal, and safe operating bounds.

[0186] Figure 6 illustrates a flowchart depicting an exemplary method for analyzing temperature data and updating risk scores. The method provides an example program flow of how to use long term infrared data to detect anomalies based on data variance. When large changes in the data variance are detected, this method preferably triggers a series of updates such as updating the temperature thresholds, calculating a new risk score, and creating a report for transparency. Long-term infrared data from a database may be analyzed statistically to detect changes in variance and identify outliers or trends that deviate from the norm. This analysis may involve comparing the current temperature data against historical temperature data to identify any deviations or anomalies. If anomalies are present, an alarm may be sent, and the temperature risk score may be updated, which may also trigger the creation of a report. Additionally, temperature limits may be updated based on the findings, allowing for dynamic adjustment of the temperature thresholds based on the observed temperature conditions within the facility.

[0187] The method begins by obtaining long term infrared data from the database in step 602. In this initial stage, the system may retrieve historical infrared data from a storage database, which may contain temperature readings collected over an extended period.

[0188] Following the data retrieval, the method proceeds to the run statistical analysis step 604. During this phase, various statistical techniques may be applied to the long-term data to identify trends, patterns, and potential anomalies.

[0189] The next step in the process is to evaluate the delta variance in step 606. This stage may involve calculating and assessing the variance in temperature readings over time, which may help identify significant deviations from normal operating conditions.

[0190] Based on the delta variance evaluation, the method may branch into two potential paths. One path leads to the run outlier and trend deviation detection step 608, where advanced algorithms may be employed to identify data points that significantly deviate from expected values or trends.

[0191] The method then reaches a decision point, represented by the anomalies present step 610. This stage may involve analyzing the results of the outlier and trend deviation detection to determine if any significant anomalies have been identified.

[0192] If anomalies are detected, the method proceeds to the send alarm step 612. In this phase, the system may generate and transmit alerts or notifications about the detected anomalies to relevant personnel or systems.

[0193] Following the alarm step, or if no anomalies are detected, the method moves to the update temperature limits step 614. This stage may involve adjusting the temperature thresholds based on the analysis of long-term data and any detected anomalies. The method then proceeds to the create report step 616. In this phase, a comprehensive report may be generated, summarizing the analysis results, detected anomalies, and any updates to temperature limits.

[0194] The final step in the method is to update temperature risk 618. This stage may involve recalculating and updating the overall temperature risk score based on the analysis results and any changes made to temperature limits.

[0195] Optionally when considering the various factors mentioned above, it is possible to calculate a weighted risk score based on the infrared data. Inputs into an infrared or temperature risk score may include the number of high temperature events, the number of anomalies detected relative to long term historical analysis, the scale of the standard deviation of the temperature distribution, and or the average temperature within the scene of the infrared camera. Each of these factors (sigma) may be combined in aggregate via a weighting function composed of a weight (alpha) and a bias term (beta). The result of this aggregation is a risk component based only on the temperatures within a scene. The score ranges from 0 - 100, with 0 being a facility which exhibits minimal temperature anomalies, low average temperatures, and zero threshold exceedances and vice-versa. This component may be combined with other risk components to provide a comprehensive assessment of the fire risk.

[0196] Figure 7 illustrates a non-limiting, exemplary method for fire / smoke / spark (FSS) risk assessment. A method 700 begins with visible light input 702, which serves as the primary data source for the monitoring process.

[0197] The visible light input 702 feeds into a monitoring module 704. This module performs comprehensive FSS monitoring functions. The monitoring module 704 conducts detections, identifying potential fire, smoke, or spark events within the visible light data. It also measures the confidence level of each detection, providing an assessment of the reliability of the identified events. The module analyzes the type of detection, distinguishing between fire, smoke, and spark occurrences. Additionally, it performs frequency analysis, tracking how often these events occur over time. The monitoring module 704 also measures the time and duration of detected events, providing crucial information about the persistence and evolution of potential hazards. Monitoring module 704 preferably processes input data to determine the occurrence, confidence, and duration of fire, smoke, or spark events. In some aspects, the FSS Detection module may use computer vision algorithms to analyze images captured by the camera system. These algorithms may be capable of identifying visual characteristics associated with fire, smoke, or sparks, and may generate output data indicating the occurrence of such events. Such an implementation is described in greater detail below. The monitoring module 704 generates monitoring data 706, which includes events, metrics, and anomalies identified during the analysis process. This comprehensive set of data serves as the input for the subsequent risk assessment stage.

[0198] The monitoring data 706 is then processed by a risk scoring module 708. This module implements an FSS risk scoring algorithm, which evaluates the monitoring data to assess the overall risk level associated with the detected fire, smoke, and spark events. Optionally, risk score module 708 calculates a score based on weighted factors such as the number of events (ol), the confidence of each event (o2), and the duration of each event (o3). Each of these factors may be multiplied by their respective weights (al, a2, a3) and summed with a bias term (P). The weights and bias term may be predetermined or dynamically adjusted based on various factors, such as the specific characteristics of the facility or the historical data of fire, smoke, or spark events.

[0199] With the analyzed detections, non-limiting examples of the aggregate metrics or factors, applied as sigma, include but are not limited to the number of detections, their frequency, confidence levels, etc may be fed into a risk scoring mechanism which weight each factor by a weight alpha and a bias beta for a component risk score.

[0200] Finally, the risk scoring module 708 produces a risk score output 710. This output represents a quantified assessment of the FSS-related risks based on the analyzed data, providing a clear indication of the potential hazards present in the monitored environment. The output of the FSS Risk Scoring module is the FSS Risk Score, which quantifies the level of fire, smoke, or spark risk within a facility. In some aspects, the FSS Risk Score may be a numerical value on a scale from 0 to 100, with higher values indicating higher levels of fire risk. The FSS Risk Score may be used to inform insurance underwriting decisions, potentially adjusting premium prices based on the assessed risk levels. In some cases, the FSS Risk Score may be combined with other risk scores, such as those derived from temperature data or cleanliness levels, to provide a more comprehensive assessment of the fire risk within a facility.

[0201] Figures 8A-8D relate to an exemplary process and implementation for training a neural network for risk detection, for example for fire risk detection. Various implementations of such neural networks are possible and are included within the scope of the present invention. The images so obtained, and / or image data, are preferably analyzed by a computer vision technology as previously described. Some examples of computer vision algorithms which may be used to quantify risk include, but are not limited to, convolutional neural networks, vision transformer networks, motion detection via background subtraction, isolation forest, semantic segmentation, generative adversarial networks, reinforcement learning with human feedback, autoencoders, or K-Nearest Neighbors.

[0202] These algorithms may run directly onboard the camera, on a local computing server, or in the cloud, or a combination of a plurality of these options.

[0203] Some examples of these algorithms may include temperature analysis from the infrared light spectrum, smoke, fire, or spark detection from the visible light spectrum.

[0204] One challenge is acquiring representative and meaningful data which may be used to train the algorithms above. For example, an adaptive temperature thresholding algorithm applied to an infrared camera may require several days, weeks, or even months of data in order to learn the normal operation patterns of a facility. The same may be said regarding a cleanliness assessment, as the day-to-day variations in facility cleanliness must be observed over time in order to establish a baseline level from which deviations may be quantified.

[0205] In the example of smoke, fire, or spark detection, millions of images exist across the internet which may be used to train a baseline convolutional neural network for classification or detection such as MobileNet, VGG, EfficientNet, EfficientDet, YOLO, ResNet, Inception, and so on.

[0206] The baseline model represents a general network that is non-specific to any particular camera sensor or environment, and therefore may struggle when deploying the network in the target environment.

[0207] For example, if a deployed camera sensor has a substantially different light or exposure response than the camera sensors used to collect the original data scraped from the internet, the subsequent network performance may degrade.

[0208] To combat this, scenarios may be created to generate specific data to the environment or sensor. One may wish to gather data from real fires, sparks, or smoke events in order to fully capture the real response of the sensor. The ISO / TS 7240-29 standard for video fire detectors defines typical test fires which may be used to collect real world data.

[0209] Furthermore, synthetic data generation may be used to increase the size and variety of the dataset. To achieve this, masks of real world fires are collected. The masks of the fire shape, size, color, orientation, and location are then applied to images similar to the target deployment environment. Next, a generative adversarial network (GAN) may be used to synthesize new images which have visually similar content to that of the real world data.

[0210] Detection of hazardous objects in the scene may require locating exactly where in the image an object lies. Common single or multi-shot detection networks such as MobileDet, YOLO, R-CNN, RetinaNet, or SSD may be used to recognize and locate certain objects within an image. Training data may be acquired in a similar fashion, using open source data on the internet.

[0211] Once camera systems have been installed, data from these systems may be used to fine-tune the baseline model for environmental or customer specific deployments. As more systems are installed, availability of real world data becomes trivial, enabling frequent finetuning events for marginal model improvement in real world applications.

[0212] A larger challenge is ensuring the quality of data remains high across both the internet sourced data and the captured data from real world deployments. A multi-stage approach may be used which combines human labeling and classification with auto labeling systems. For example, an initial model may be trained using manually labeled data. This model can then be used to begin auto-labeling similar cases on which it has been trained on. Predictions from the network which are strong or weak may be flagged for human review, manually labeled, then the cycle may repeat.

[0213] Turning back to the drawings, Figure 8A illustrates an exemplary neural network training process for developing a fire detection model. The neural network training process 800 begins with a data collection step 802, which incorporates multiple sources of image data. In some cases, the data collection step 802 may include an internet scraped image step 804, where images are gathered from online sources. A synthetic image step 806 may generate artificial images to supplement the dataset. Additionally, a visual FX (effects) overlay step 808 may apply visual effects to existing images to create more diverse training samples.

[0214] The outputs from these data collection methods converge into a training dataset step 810, which compiles all the gathered images into a unified dataset. From the training dataset step 810, the neural network training process 800 proceeds to a data augmentation step 812. The data augmentation step 812 connects to a data augmentation type step 814, where various augmentation techniques may be applied. These techniques may include artificial noise, geometric transformations, and color adjustments to increase the diversity and robustness of the training data.

[0215] The data augmentation type step 814 leads to an augmented dataset step 816, which represents the expanded and enhanced training data. The augmented dataset step 816 feeds into two parallel paths: a hyperparameter tuning step 818 and a training configuration step 820.

[0216] In some cases, the hyperparameter tuning step 818 may utilize Bayesian optimization to determine optimal model parameters. The training configuration step 820 may set up the training environment, including specifying the batch size. In some cases, the batch size may be set to 64 based on the available GPU memory for training.

[0217] Both the hyperparameter tuning step 818 and the training configuration step 820 inform a model training step 822. The model training step 822 may utilize the EfficientNet family of models for fire detection. In some cases, an Adam optimizer with decoupled weight regularization may be employed during training. Additionally, a cosine annealing learning rate scheduler with a warmup period may be used to adjust the learning rate throughout the training process.

[0218] The neural network training process 800 concludes with a trained fire detection model step 824, which represents the final output of the training pipeline. This trained fire detection model may be capable of identifying fire in various scenarios based on the diverse dataset and optimization techniques applied during the training process.

[0219] To increase the efficiency and accuracy of the operations of the network, preferably each data point has one or more augmentations applied. Data augmentation is preferably used for building resilient, high-performing models. Such data augmentation may overcome problems where training datasets are imbalanced, or potentially different in distribution from real-world deployment environments. A comprehensive augmentation strategy may simultaneously reduce overfitting, balance class distributions, and close the sim-to-real domain gap. Such pipelines may integrate stochastic, domain-aware augmentations tailored to each dataset and training objective, combining geometric, photometric, and structural transformations with more advanced simulation techniques.

[0220] One goal of data augmentation is to mitigate overfitting, especially in deep learning models that tend to memorize high-frequency patterns in limited or biased datasets. To achieve this, random augmentations may be applied independently to each data point during training, introducing enough variability to encourage the model to leam invariant, generalizable features. Geometric augmentations such as random rotations, flips, translations, and elastic deformations of the data expose the model to variations in object orientation and spatial configuration, forcing it to develop robustness to viewpoint shifts. These transformations simulate different camera perspectives or physical poses, helping models generalize beyond fixed spatial biases present in the training data.

[0221] Alongside geometric changes, photometric augmentations like color jittering, gamma correction, and exposure adjustments may simulate variations in illumination conditions. These are especially effective when training for outdoor or mobile environments where lighting is inconsistent or when synthetic images need to match the noise patterns and color responses of real sensors. Additionally, the injection of artificial noise, such as Gaussian, Poisson, or camera-specific sensor noise, may be used to introduce real-world imperfections. Compression artifacts (e.g., JPEG distortion or quantization errors) may also be added to emulate the effects of video compression, image streaming, or storage degradation, as such factors may otherwise create discrepancies between training and deployment environments.

[0222] Another objective of augmentation pipelines is to address class imbalance. In real- world datasets, certain classes may be vastly underrepresented. This imbalance may lead to biased predictions and low recall for minority classes. To correct this, techniques such as SMOTE (Synthetic Minority Oversampling Technique) may be used to synthesize new training examples by interpolating between existing minority samples in feature space. In vision tasks, this technique may be adapted with targeted augmentations: specific transformations (e.g., rotations, brightness shifts, blur) are applied selectively to minorityclass images to effectively amplify their representation in the dataset without introducing redundancy. Moreover, advanced blending techniques like Mixup and Cutmix further enhance the diversity of training signals by creating composite images that combine pixels and labels from multiple classes. These strategies help regularize decision boundaries and improve generalization, particularly in low-data or multi-class scenarios. Class imbalance may also be addressed with modifications to the loss function which biases loss to be higher on images which prove to be difficult to classify during training. This can shift the network behavior from having a polarized output (i.e. confidence scores always 0 or 1.0) to a more nuanced prediction reflecting its uncertainty on challenging cases. Further examples of addressing class imbalance may come from smarter sampling at the mini-batch level. Rather than randomly sampling images from the dataset, a difficulty heuristic may be used to construct more challenging distributions of sample images for a mini-batch.

[0223] Another aspect of modem augmentation frameworks is the need to bridge the sim-to- real gap: the discrepancy between clean, synthetic training data and noisy, real-world inputs. Bridging this gap is preferred where models are trained in simulation but then deployed on real-world sensors. To improve domain transfer, sensor-specific noise modeling may be introduced during augmentation. This may include but is not limited to simulating shot noise, read noise, and motion blur that match the physical properties of the target camera system. In addition, domain randomization may be employed to expose the model to a wide range of environmental conditions during training, including random lighting spectra, weather perturbations, and imaging artifacts. This approach, originally popularized in robotics, has proven highly effective in video surveillance and drone-based applications. Studies indicate that models trained with these augmentations demonstrate up to 15% higher real-world detection accuracy compared to models trained on clean, synthetic data alone.

[0224] To implement these strategies in practice, augmentation pipelines may be designed with a modular and stochastic architecture. Transformations may be applied with randomized parameters, such as ±15° for rotation, ±0.2 for brightness delta, ensuring that each data point follows a unique path through the augmentation space. The order of operations may also be important: geometric transformations are typically applied first to maintain consistent spatial references, followed by photometric adjustments and then noise layers. In sensor-aligned augmentation, it is common to profile the target deployment environment first (e.g., collecting sensor noise histograms or analyzing image compression ratios) and then tune augmentation parameters accordingly.

[0225] Frameworks like PyTorch, Albumentations, and TensorFlow Image offer powerful interfaces for implementing complex augmentation chains. A representative PyTorch pipeline might include a combination of horizontal flipping, color jittering, Gaussian blur, random rotation, and a custom noise layer that injects sensor-like distortion.

[0226] The performance impact of advanced augmentation may be substantial. Models trained with randomized, class-aware, and sensor-specific augmentation pipelines are expected to show at least a 20-30% reduction in generalization error over unaugmented baselines. In imbalanced datasets, these strategies have been shown to improve minority-class Fl -scores by up to 18%, for example when Mixup and Cutmix are employed. Sim-to-real transfer learning also benefits significantly: benchmarks show that models trained with compression and noise simulation achieve 90-95% mAP on real-world test sets, compared to <75% when trained on clean, unaugmented synthetic data.

[0227] Such data augmentation is a composite system of geometric diversity, photometric realism, statistical balancing, and domain simulation. When these components are integrated thoughtfully, data augmentation becomes not only a defense against overfitting but also a vehicle for robust, deployable performance across complex and shifting environments. These different augmentation methods may be applied to the present invention.

[0228] To ensure robust model performance under varying real-world conditions, data augmentation may be employed as a critical preprocessing step to simulate different lighting conditions and camera angles. This process involves intentionally altering the visual characteristics of input images before training, thereby increasing the diversity of the dataset and enhancing the model’s ability to generalize. For example, lighting variations may be simulated through brightness, contrast, and saturation adjustments, which alter the image’s overall luminance, dynamic range, and color intensity. These transformations prepare the model to handle overexposed, underexposed, or unevenly lit environments: conditions common in outdoor scenes, surveillance footage, and industrial inspection tasks. Additionally, hue shifts and color channel swapping may be applied to simulate discrepancies between different camera sensors or color correction settings, further improving model robustness to variations in device hardware or calibration.

[0229] In scenarios involving multiple viewpoints or inconsistent angles, which may be seen for example in robotics, video streams, or mobile vision systems, geometric augmentations like affine transformations (rotation, perspective shift, shear) may be used to complement these lighting-based augmentations. Lighting augmentation may be further executed in a stochastic, pixel-level fashion to simulate localized shadows, lens flares, or color casting that arise in real-world settings.

[0230] Beyond these traditional augmentations, advanced techniques such as Fourier domain adaptation may be applied to simulate complex domain shifts, such as those caused by compression artifacts or transmission noise for example, as seen in RTSP streams encoded with H.264 or H.265 codecs. In this approach, noise patterns characteristic of a target domain may be transferred onto source images by manipulating the amplitude or phase components in the frequency domain, thereby imprinting artifacts like block noise, temporal aliasing, or bitrate-induced distortion. This technique allows models trained on clean, high-resolution images to better handle real-world deployment conditions, such as low-latency video streams from security cameras or embedded systems for example.

[0231] Taken together, these augmentation strategies may be applied to transform a standard image into one that appears substantially different, and which mimics the unpredictable and noisy environments encountered in practice. By exposing the model to such variations during training, this preprocessing pipeline may be applied to improve the model’s ability to generalize, adapt to new domains, and maintain high performance under suboptimal imaging conditions.

[0232] For the present implementation, as a non-limiting example, a random set of augmentations was applied to each data point to reduce overfitting on the training dataset, upsample minority classes as well as reduce the transfer gap to the real cameras. A wide range of augmentations including artificial noise, geometric transformations, color and exposure adjustments, compression artifact simulation and Mixup / Cutmix augmentations were used (as non-limiting examples only). In operation, the network then outputs a confidence score indicating the likelihood that the region contains fire (or some other parameter of interest, such as an obstruction, flammable object etc). Each region’s score is then optionally stored in a temporal buffer.

[0233] Figure 8B illustrates an exemplary inference process for analyzing high-resolution region crops of a plurality of images. The inference process 830 may begin with high- resolution region crops 832 being input into the system. These high-resolution region crops 832 may be processed to generate a confidence score 834. In some cases, the confidence score 834 may represent the likelihood that the analyzed region contains a feature of interest, such as fire or smoke.

[0234] The confidence score 834 may then be stored in a temporal buffer 836. The temporal buffer 836 may allow the system to track confidence scores over time, enabling analysis of temporal patterns or persistence of detected features.

[0235] From the temporal buffer 836, the inference process 830 may proceed to a threshold decision 838. The threshold decision 838 may evaluate if the confidence score exceeds a predetermined threshold for a required time period. This temporal evaluation may help reduce false positives by ensuring that detected features persist over time.

[0236] Based on the outcome of the threshold decision 838, the inference process 830 may branch into two paths. If the confidence score exceeds the threshold for the required time period, the process may proceed to alert generation 840. Alert generation 840 may involve creating and sending notifications to relevant parties or systems about the detected feature of interest.

[0237] Alternatively, if the confidence score does not exceed the threshold for the required time period, the inference process 830 may proceed to monitoring continuation 842. Monitoring continuation 842 may involve ongoing analysis of new high-resolution region crops, maintaining the system's vigilance without triggering unnecessary alerts.

[0238] In some cases, the inference process 830 may operate continuously, processing new high-resolution region crops as they become available. This continuous operation may enable real-time monitoring and rapid response to detected features of interest.

[0239] Figure 8C illustrates a non-limiting example of a neural network architecture called EfficientNet-BO. The EfficientNet-BO architecture may be used as a classifier for fire and / or smoke detection in images in any of the devices, systems or methods described herein. It is a non-limiting example of such a neural network, and was used to perform the fire and smoke detection examples, described in greater detail below. The architecture begins with an initial convolution block 1 that processes input images at a resolution of 224x224 using a 3x3 kernel. This initial block may extract low-level features from the input image.

[0240] Following the initial convolution, a mobile block one 2 operates at 112x112 resolution with two layers using 3x3 kernels. This block may further refine the extracted features while reducing the spatial dimensions.

[0241] The network continues with a mobile block six first 3 maintaining the 112x112 resolution across two layers with 3x3 kernels. This block may allow for additional feature extraction without changing the spatial resolution.

[0242] A mobile block six second 4 reduces the resolution to 56x56 using two layers with 5x5 kernels. The larger kernel size may enable the network to capture broader spatial relationships in the image.

[0243] A mobile block six third 5 further reduces resolution to 28x28 across three layers using 3x3 kernels. This progressive reduction in resolution may help the network focus on more abstract features relevant to fire detection.

[0244] A mobile block six fourth 6 operates at 14x14 resolution using three layers with 5x5 kernels. The combination of reduced resolution and larger kernels may allow for capturing complex patterns across a wider receptive field.

[0245] A mobile block six fifth 7 processes data at 7x7 resolution across four layers using 5x5 kernels. This block may further refine high-level features useful for fire classification.

[0246] A mobile block six sixth 8 maintains the 7x7 resolution with one layer using a 3x3 kernel. This final mobile block may perform additional feature refinement before the final stages of the network.

[0247] The architecture includes a final convolution block 9 using a 1x1 kernel at 7x7 resolution. This block may serve to combine and compress the learned features.

[0248] Following the final convolution, a pooling layer 10 may reduce the spatial dimensions and aggregate feature information.

[0249] The architecture concludes with a dense layer 11 that maintains the 7x7 resolution. This layer may perform the final classification task, determining the likelihood of fire presence in the input image.

[0250] Each block in the EfficientNet-BO architecture processes the data at specific resolutions using defined kernel sizes and numbers of layers. This structure may allow the network to efficiently transform the input image through multiple stages of feature extraction and refinement, ultimately producing a classification output for fire detection. Figure 8D illustrates a flowchart of an exemplary, non-limiting method for processing video input and generating alerts. The method 850 begins with a step 851 of receiving video input. From the video input, the method 850 proceeds to a step 852 of performing motion detection on the video frames.

[0251] Following motion detection, the method 850 advances to a step 854 of color filtering. The color filtering step 854 may identify pixels or regions with colors characteristic of fire or smoke. After color filtering, the method 850 moves to a step 856 of extracting regions of interest from the filtered video frames.

[0252] The extracted regions are then processed through a step 858 of neural network analysis. The neural network may be trained to detect visual patterns associated with fire, smoke, or other hazardous conditions.

[0253] In some cases, the method 850 incorporates temporal consistency checking. A step 860 determines if a time threshold is met. This step 860 may evaluate whether the neural network's confidence score for detecting a hazardous condition remains above a specified threshold for a certain duration. If the time threshold is not met, the method 850 proceeds to a step 862 of continuing monitoring.

[0254] If the time threshold is met, indicating a persistent detection of a potential hazard, the method 850 advances to a step 864 of sending an alert to a cloud system. The cloud system may provide additional processing capabilities and storage for alert data.

[0255] After sending the alert to the cloud, the method 850 reaches a decision point at step 866 regarding the use of a vision model for further validation. The vision model for example may comprise a larger model and / or a plurality of such models. If the vision model is not to be used, the method 850 proceeds directly to a step 868 of escalating the alert to a customer or relevant personnel.

[0256] In some cases, when the vision model is to be used, the method 850 moves to a step 870 of processing the alert data with the vision model. The vision model may be a more sophisticated or specialized algorithm designed to reduce false positives and provide a final validation of the detected hazard.

[0257] Following the vision model processing, the method 850 reaches a decision step 872. Based on the output of the vision model, if the potential hazard is confirmed, the method 850 proceeds to the step 874 of escalating the alert to the customer. If the vision model rejects the detection, the method 850 moves to a step 876 of discarding the alert.

[0258] This approach of cloud alert and validation with a foundation model as a final decision step may enhance the accuracy of the alert system and reduce false alarms. The method 850 provides a multi-stage process for video analysis and alert generation, incorporating both edge processing and cloud-based validation to improve reliability in hazard detection.

[0259] Figures 9A-9D illustrate flowcharts related to exemplary methods for detecting fire in video streams. The method 900 comprises at least three stages, as shown in Figure 9A: a stage to receive video images 901, a stage motion detection 902, a stage color filtering 904, and a stage region proposal 906. The output of this method may comprise output regions 908 as described in greater detail below.

[0260] The stage motion detection 902 begins with a background subtraction at 920 from the received images (stage 901), as shown in Figure 9B. The method 900 then proceeds to detect motion 922 and identify foreground regions 924 where activity may be occurring. These steps lead to generating a motion mask 926.

[0261] The stage color filtering 904 processes an input frame 930, as shown in Figure 9C. The process may filter fire colors 932 by applying HSV / RGB thresholds 934 to the input frame. In some cases, HSV (Hue, Saturation, Value) thresholds may be used to isolate fire- like colors. In other cases, RGB (Red, Green, Blue) thresholds may be applied. The motion mask (from stage 926) may then be combined at 936 with the color filtering results to retain candidate fire pixels 938.

[0262] The stage region proposal 906 involves extracting connected components 940 from the combined masks. Next candidate regions 942 may be identified, as well as crop high resolution patches 944 from these regions. These patches may become candidate inference regions 946. The method 900 may conclude with output regions 908 that are prepared for further analysis.

[0263] Figure 10 illustrates a series of surveillance camera images demonstrating fire detection capabilities in various environmental conditions. The figure comprises four separate camera views arranged in a 2x2 grid layout, each showing different instances of fire detection.

[0264] In each image of Figure 10, a fire region is identified and marked with a red rectangular box. The fire region designation provides a clear visual indicator of where the system has detected potential fire hazards within the monitored areas. The fire region is detected with a probability from 0 to 1, in which 1 is 100%.

[0265] The top left image of Figure 10 presents an elevated view of an industrial or commercial space. A fire region of probability 0.97 is detected and highlighted in the middle of the frame, demonstrating the system's ability to identify fire incidents from a distance. In the top right image of Figure 10, a closer view of a fire region of probability 0.98 is displayed. This image shows more intense flames visible within the detection box, illustrating how the system may mark fire regions with varying levels of fire intensity.

[0266] The bottom left image of Figure 10 captures a fire region of probability 1.00 in a darker indoor environment. The image was taken in a dimly lit corridor or passageway, showcasing the system's capability to identify fire hazards in low-light conditions.

[0267] In the bottom right image of Figure 10, a fire region of probability 1.00 is detected at ground level in an industrial setting. This demonstrates how the system may identify fire incidents in different spatial orientations and contexts within a facility.

[0268] The consistent application of the fire region 1.00 marker across diverse environmental conditions and lighting scenarios illustrates the robustness of the fire detection system. In some cases, the system may use computer vision algorithms to analyze the visual characteristics of potential fire incidents and determine when to apply the fire region 1.00 designation.

[0269] The red rectangular boxes used to highlight the fire regions may be used to provide a standardized visual cue for rapid identification of fire hazards across different camera views. In some implementations, these visual markers may be integrated with alert systems to notify personnel or trigger automated responses when fire regions are detected.

[0270] Figures 11-13 relate to non-limiting exemplary methods for smoke detection, for example in images or video streams.

[0271] Figure 11 illustrates a flowchart of an exemplary, non-limiting method for smoke detection. The method 1100 may be used for detecting smoke in video streams for example.

[0272] The method 1100 begins with a step 1101 where an input image is received. This input image may be a frame from a video stream captured by a camera monitoring an area of interest.

[0273] In step 1102, image energy is calculated. In some cases, this energy calculation may be performed using Sobel filtering. Sobel filtering is a technique that computes the spatial gradient intensity of an image, providing a measure of how rapidly intensity changes across the image. This calculation results in an energy map of the current frame.

[0274] Following the energy calculation, the method 1100 proceeds to step 1104 for energy drop detection. In this step, the current frame's energy map may be compared to a reference background frame or a previous frame. Regions where energy has dropped significantly are considered candidates for smoke detection, as smoke tends to obscure edges and reduce contrast in an image. After energy drop detection, the method 1100 advances to step 1106, where bounding boxes are generated. These bounding boxes may be created around regions where significant energy drops have been detected, identifying areas of interest for further analysis.

[0275] The method 1100 then proceeds to step 1108, where a neural network classifier is applied. This classifier may analyze the regions within the bounding boxes to determine the likelihood of smoke presence based on learned features and patterns.

[0276] Following the classification, the method 1100 moves to step 1110, where tracking occurs in a temporal buffer. This step may involve storing and analyzing the classification results over time to identify persistent patterns or trends.

[0277] From step 1110, the method 1100 reaches a decision point where it checks if energy remains low. If the energy remains low, the method 1100 proceeds to step 1112, where a prealarm candidate is identified. This pre-alarm candidate may indicate a high likelihood of smoke presence that warrants further attention or action.

[0278] If the energy does not remain low, the method 1100 moves to step 1114 to continue monitoring. This step ensures ongoing surveillance of the area, allowing the method to detect any future energy drops that may indicate smoke presence.

[0279] Figures 12A and 12B illustrate a flowchart for an exemplary, non-limiting smoke detection process. The process 1200 begins with a video frame input 1201, which provides image data for analysis. The process then moves to a background model computation 1202, where a reference background model may be established.

[0280] Following the background model computation 1202, an energy calculation 1204 may be performed on the image data. The process then proceeds to an energy drop detection 1206, which analyzes changes in the energy levels of the scene. Based on the detected energy drops, a region proposal 1208 may identify areas of interest within the frame.

[0281] The identified regions may undergo neural network processing 1210, which analyzes the visual characteristics of these areas. The results from this processing may be stored in a temporal buffer 1212. A duration decision 1214 may evaluate whether the detected conditions persist for a specified time period.

[0282] If the duration threshold is met, the process may move to an alert transmission 1216. If the duration threshold is not met, the process may continue with monitoring continuation 1218.

[0283] After the alert transmission 1216, the process may advance to a foundation model 1220, which performs additional analysis of the detected event. The results of this analysis may lead to a smoke decision 1222, which determines whether smoke is present in the scene. Based on the smoke decision 1222, the process may branch into two paths: if smoke is detected, the process may move to a customer alert 1224, where notifications may be sent to relevant parties. If smoke is not detected, the process may move to an alert discard 1226, where the alert may be dismissed.

[0284] In some cases, the video frame input 1201 may be provided by a surveillance camera or other imaging device monitoring an industrial or commercial space. The background model computation 1202 may involve creating a reference image of the scene under normal conditions, which can be used for comparison in subsequent frames.

[0285] The energy calculation 1204 may involve applying image processing techniques, such as edge detection or texture analysis, to quantify the visual information content of each frame. The energy drop detection 1206 may then compare the current frame's energy to the background model, identifying areas where visual information has been obscured or altered.

[0286] In some implementations, the region proposal 1208 may use the results of the energy drop detection 1206 to define bounding boxes around areas of interest. These regions may then be processed by a neural network in the neural network processing 1210 step. The neural network may be trained to recognize visual patterns associated with smoke.

[0287] The temporal buffer 1212 may store the results of the neural network processing 1210 over multiple frames. This allows the duration decision 1214 to assess whether potential smoke detections persist over time, reducing false positives from transient visual effects.

[0288] In some cases, the foundation model 1220 may be a large, pre-trained vision model capable of more sophisticated image analysis. This model may provide an additional layer of verification before triggering a customer alert 1224.

[0289] The customer alert 1224 may involve sending notifications through various channels, such as email, SMS, or a dedicated monitoring interface. The alert discard 1226 may log the event for future analysis but not escalate it to human operators.

[0290] Figure 13 illustrates a series of images showing smoke detection capabilities in industrial environments. The figure contains four panels arranged in a 2x2 grid, each depicting different instances of smoke detection analysis in industrial settings.

[0291] In some cases, the smoke detection system may utilize rectangular detection regions to identify areas where smoke is present. These detection regions may be visually represented by red rectangular boxes overlaid on the image.

[0292] The top two panels of Figure 13 show similar views of an industrial interior space with overhead lighting. In these panels, red rectangular boxes indicate regions where smoke has been detected. Associated with each detection region, a smoke detection score may be displayed, representing the system's confidence in detecting smoke within that region.

[0293] In one panel, a smoke detection score 0.95 is shown, while another panel displays a smoke confidence threshold 0.99. These numerical values may represent different levels of confidence in smoke detection, with higher numbers potentially indicating greater certainty in the presence of smoke.

[0294] The bottom two panels present additional views of industrial spaces where smoke detection is being performed. The bottom left panel shows a wider view of an industrial area with piping or ductwork, while the bottom right panel depicts a more enclosed space. In one of these panels, a red rectangular box indicates a smoke detection region with a smoke alert threshold 0.90.

[0295] The smoke detection scores and thresholds shown in the panels (0.95, 0.99, and 0.90) may be used to determine when to trigger alerts or initiate automated actions based on the detection of smoke in the monitored areas. In some cases, the system may use these scores to classify detections as low, medium, or high confidence, allowing for more nuanced responses to potential smoke events.

[0296] The smoke detection system may be capable of identifying smoke in various industrial environments, including open areas with complex machinery and more confined spaces. This versatility may allow for comprehensive smoke monitoring across different types of industrial facilities.

[0297] In some implementations, the smoke detection algorithm may analyze visual characteristics such as opacity, color, contrast, and movement patterns to differentiate smoke from other visual phenomena. The confidence scores may reflect the algorithm's assessment of these characteristics in determining the likelihood of smoke presence.

[0298] The use of multiple detection regions within a single frame, as shown in Figure 13, may be used to enable the system to localize smoke sources and track the spread of smoke across an industrial space. This localization capability may be valuable for pinpointing the origin of a fire or identifying areas of concern for further investigation.

[0299] The above detailed examples relate to various potential implementations of the present invention and are not intended to be limiting.

[0300] A wide range of algorithms and modeling strategies may be used for image classification in conjunction with the present invention, depending on task complexity, available data, and deployment constraints. Convolutional Neural Networks (CNNs) remain the backbone of modem image classification, offering layered spatial feature extraction with minimal preprocessing. Notable variants such as AlexNet, VGG, ResNet, Inception-v3, and MobileNet offer trade-offs in depth, efficiency, and adaptability. More recently, Vision Transformers (ViTs) have emerged as powerful alternatives, using self-attention to capture long-range spatial dependencies and delivering state-of-the-art performance when trained at scale. Models from the YOLO family, while originally intended for object detection, offer lightweight and real-time classification capabilities, making them well-suited for edge deployments. The Segment Anything Model (SAM), with its promptable segmentation capabilities, enhances classification by isolating relevant image regions, particularly in complex or cluttered scenes. Traditional machine learning techniques such as Support Vector Machines (SVMs), K-Nearest Neighbors (KNN), and Random Forests remain useful for structured or low-data scenarios, often operating on extracted features. Finally, transfer learning enables rapid model adaptation by leveraging pretrained architectures like EfficientNet or ResNet, significantly reducing training time and boosting accuracy across diverse domains. These approaches may be applied independently or in combination as part of the classification framework for the present invention.

[0301] Convolutional Neural Networks (CNNs)

[0302] Convolutional Neural Networks (CNNs) represent a foundational architecture for modem image classification tasks. CNNs are specifically designed to handle grid-like data such as images by applying convolutional layers that extract spatial hierarchies of features. Each convolutional layer applies filters to detect patterns like edges, textures, and eventually, complex shapes. Pooling layers downsample feature maps to reduce computational complexity, while fully connected layers at the end of the network make predictions based on the learned features. One of the most notable advantages of CNNs is their ability to learn directly from pixel data with minimal preprocessing, making them extremely effective across a wide range of computer vision tasks. Various types of CNNs are suitable for use with the present invention, including variations thereof such as for EfficientNet as described above.

[0303] Among the early CNN architectures, AlexNet marked a significant turning point in 2012 by winning the ImageNet competition with a dramatic leap in accuracy over traditional methods. It introduced the use of ReLU activations, dropout regularization, and GPU training to accelerate learning. Following AlexNet, VGG16 and VGG19 emphasized architectural simplicity by stacking multiple convolutional layers with small 3x3 kernels, resulting in deep but manageable networks. These models became popular benchmarks due to their straightforward design, although they are relatively resource-intensive.

[0304] To tackle deeper architectures more efficiently, ResNet introduced the concept of residual learning via skip connections. These connections allow gradients to bypass certain layers, effectively mitigating the vanishing gradient problem that had limited earlier deep networks. ResNef s success lies in its ability to train extremely deep networks — ResNet-50, ResNet-101, and ResNet- 152 are common variants — while maintaining stable optimization and superior accuracy.

[0305] Inception-v3, another influential CNN, adopts a modular approach using Inception modules that apply multiple convolution filters of varying sizes in parallel. This design captures both fine-grained and coarse information while reducing redundant computations. Inception-v3 also uses factorized convolutions and auxiliary classifiers for regularization, resulting in a model that is both efficient and highly accurate.

[0306] MobileNet, developed for deployment on mobile and edge devices, introduces depthwise separable convolutions to drastically reduce computation and memory footprint without sacrificing performance. Instead of applying full convolutions across all input channels, MobileNet separates the operation into depth-wise and point-wise convolutions, achieving lightweight architectures ideal for real-time applications.

[0307] Vision Transformers (ViTs)

[0308] Vision Transformers (ViTs) represent a paradigm shift in image classification by replacing convolutions with self-attention mechanisms. Inspired by the success of transformers in natural language processing, ViTs divide an image into fixed-size patches (e.g., 16x16), flatten each patch, and encode it into a vector. These vectors are then processed as a sequence (similar to word embeddings in NLP) through a stack of transformer layers that use multi-head self-attention to model long-range dependencies across the image.

[0309] One of the key advantages of ViTs is their scalability and ability to learn global contextual relationships more effectively than CNNs, which are inherently local in their operations. ViTs have demonstrated strong performance on large-scale datasets like ImageNet and in some cases have outperformed CNNs when pre-trained with sufficient data. Moreover, ViTs are architecture-agnostic: they avoid task-specific inductive biases like translation invariance, which can be both a strength (more flexibility) and a weakness (requires more data to train effectively). The introduction of hybrid models like DeiT (Data- efficient Image Transformers) and Swin Transformers further enhances ViTs’ practicality by incorporating hierarchical features and local attention mechanisms, making them viable alternatives to CNNs for a wide array of vision tasks. Various types of ViTs are suitable for use with the present invention.

[0310] YOLO (You Only Look Once)

[0311] Although initially developed for object detection, the YOLO (You Only Look Once) family of models has also proven effective for classification tasks, especially in scenarios requiring high-speed inference. YOLO models frame image processing as a single regression problem, simultaneously predicting class labels and bounding boxes from entire images in one pass through the network. This end-to-end pipeline significantly improves inference speed compared to traditional multi-stage approaches.

[0312] Y0L0v7, Y0L0v8, and Y0L0v9 continue to refine the balance between speed and accuracy. For classification-specific tasks, these models can be adapted by removing or simplifying the bounding box regression head, focusing solely on class prediction. Their lightweight and optimized architecture makes them ideal for edge computing, autonomous vehicles, and surveillance applications where both rapid detection and classification are critical. YOLO’s various backbone architectures are often used in custom classification pipelines for tasks requiring low latency and robust generalization under resource constraints. Various types of YOLOs are suitable for use with the present invention.

[0313] Segment Anything Model (SAM)

[0314] The Segment Anything Model (SAM), developed by Meta Al, is primarily designed for instance segmentation — automatically identifying and masking all objects in an image. However, SAM can also be adapted for image classification tasks. By generating masks corresponding to distinct objects within an image, SAM enables downstream classification models to focus only on relevant regions, improving both interpretability and accuracy.

[0315] SAM uses a promptable architecture, where users can guide the segmentation process with points, boxes, or text prompts. When paired with a classification head, SAM facilitates fine-grained object categorization by providing clean object boundaries and region-of-interest (ROI) masks. This is particularly useful in applications such as medical imaging, remote sensing, and robotics, where object boundaries carry semantic significance. Its zero-shot capability makes SAM an appealing tool for rapid deployment in unfamiliar environments with minimal retraining. SAM may also be used with the present invention.

[0316] Contrastive Learning Image Preprocessing

[0317] Contrastive Learning Image Preprocessing (CLIP) was developed by OpenAI as a means to improve generalization for image classification tasks. Using rich text descriptions paired with images, a visual encoder model is trained alongside a text encoder. The contrastive loss between the latent spaces (image and text) is minimized such that there is a high overlap between the two modalities. In doing so, CLIP showed substantial benefits over existing methods like VGG and ResNet across a variety of datasets, including strong zeroshot performance on various image classification tasks.

[0318] The dataset that was created covers a huge corpus of information. It is possible to use foundation models like CLIP for transfer learning. What this means in practice is that the larger model capabilities may be distilled into much smaller models that are domain specific. The distillation process is like a student-teacher mimicry where the large model is the teacher and the student's job is to minimize the loss function between what it outputs and what the teacher says the answer should be. Typical loss functions may include but are not limited to cosine similarity, mean squared error, and many others.

[0319] The CLIP model is a multi-modal model that is able to understand textual context when paired with images. After distillation, the models are able to run on computationally limited hardware such as cameras or other edge devices.

[0320] Support Vector Machines (SVMs)

[0321] Support Vector Machines (SVMs) are classical machine learning models well-suited to both binary and multi-class image classification tasks, especially when the feature space is structured and high-dimensional. SVMs operate by identifying the optimal hyperplane that separates different classes with the maximum margin. When data is not linearly separable, SVMs can use kernel tricks — such as the radial basis function (RBF) kernel — to map input data to a higher-dimensional space where a linear separator becomes feasible.

[0322] SVMs are particularly useful when training data is limited, and overfitting is a concern. Although they lack the representational power of deep networks, SVMs perform well with handcrafted features like Histogram of Oriented Gradients (HOG), SIFT, or PCA- reduced embeddings. Their deterministic nature and robustness to small datasets make them ideal for tasks like face recognition, document classification, and industrial inspection.

[0323] Various types of SVMs are suitable for use with the present invention.

[0324] K-Nearest Neighbors (KNN)

[0325] K-Nearest Neighbors (KNN) is a simple, non-parametric algorithm that classifies images based on feature similarity. For a given test image, KNN identifies the 'k' most similar instances in the training dataset (usually based on Euclidean distance) and assigns the most common label among these neighbors to the test image. KNN does not require a training phase in the traditional sense — it merely stores all labeled instances and performs computations at inference time.

[0326] While KNN is computationally expensive for large datasets and not scalable for realtime applications, it remains a strong baseline and a powerful tool in low-data or exploratory settings. Feature extraction is often decoupled from the classifier: CNNs or other models can generate embeddings, and KNN then operates in this reduced feature space. Its transparency and ease of interpretation make it a valuable educational and analytical tool. It may also be used with the present invention.

[0327] Random Forest

[0328] Random Forest is an ensemble learning method that constructs multiple decision trees during training and outputs the mode of their predictions. Each tree is trained on a random subset of the data and features, introducing diversity that improves generalization and reduces overfitting. For image classification, features are typically extracted beforehand using methods like HOG or CNN embeddings, and the Random Forest algorithm handles classification based on these inputs.

[0329] Random Forest is particularly effective for imbalanced datasets or situations where interpretability and robustness are crucial. It offers insights into feature importance and decision paths, which can be invaluable in domains like bioinformatics or industrial diagnostics. While not as powerful as CNNs for raw pixel input, it remains a strong model when paired with effective feature engineering or transfer learning. It may also be used with the present invention. Transfer Learning

[0330] Transfer learning is a powerful technique that leverages pre-trained models — typically trained on massive datasets like ImageNet — to jumpstart training on a new task with a smaller dataset. The core idea is that early layers of CNNs learn general-purpose features (e.g., edges, textures) that are transferable across domains. By fine-tuning only the later layers or adding a new classification head, practitioners can quickly adapt a model to specific tasks such as medical diagnosis, wildlife monitoring, or defect detection.

[0331] Transfer learning dramatically reduces training time and data requirements, while still achieving high accuracy. Models like ResNet, Inception, and EfficientNet are frequently used as backbones for transfer learning. It is especially effective in fields where labeled data is scarce or expensive to obtain, and it has become a cornerstone technique for computer vision pipelines in both industry and academia. Transfer learning may be used with various types of pretrained models as described herein for the present invention.

[0332] INDUSTRY ENVIRONMENT CLEANLINESS

[0333] Cleanliness may be conceptualized as the absence of deviations from an ideal operational state where elements exist in their proper order and location, and where foreign objects are absent. This ideal operational state represents a baseline condition characterized by surfaces free of foreign objects, unmarred by substances that alter their expected appearance, and exhibiting an overall spatial organization where objects maintain their intended positions. When these deviations appear, whether as discrete items out of place, residual traces of activity, or overall disorganization, they constitute measurable departures from the clean state. This approach allows cleanliness to be quantified along a spectrum rather than as a binary condition, where the magnitude, type, and distribution of deviations collectively determine how far a space has moved from its ideal clean condition.

[0334] For example, the system as described herein for detecting cleanliness (or the lack thereof) is preferably able to recognize industrial environment safety conditions, by being able to recognize images of both ideal and non-ideal industrial environments. The ideal operational state preferably relates to a baseline condition characterized by surfaces free of foreign objects, unmarred by substances that alter their expected appearance (which may be included in the definition of foreign objects), and exhibiting an overall spatial organization where objects maintain their intended positions, as noted above. For example, the ideal operational state may comprise a physical arrangement showing the presence of only material, machines, and humans directly involved in operational processes of the industrial environment; proper spatial organization of materials and equipment according to predefined safety parameters; and / or the absence of foreign objects (again noting that foreign substances or materials, or substances or materials that alter their expected appearance may be included within the definition of foreign objects).

[0335] Predefined safety parameters for proper spatial organization of materials and equipment in industrial environments may comprise standardized arrangements, clearances, and positioning protocols designed to minimize risk while maximizing operational efficiency. These parameters establish spatial relationships that maintain accessibility, visibility, ergonomic function, and emergency response capability throughout a facility. At their foundation, these parameters include maintaining designated clearances around machinery, equipment, and electrical installations according to manufacturer specifications and relevant regulatory standards including but not limited to OSHA, NFPA, or ISO requirements, and / or other requirements or regulations as appropriate. These clearances typically range from 0.6 to 1.2 meters depending on the equipment type, creating safety zones that prevent accidental contact, provide service access, and reduce the risk of entrapment or crushing injuries.

[0336] Proper spatial organization further features the establishment and maintenance of unobstructed emergency egress paths with minimum widths of 0.7 to 1.1 meters, depending on occupancy and building codes. These pathways are to remain free of any temporary storage, equipment, or materials, and maintain direct line-of-sight visibility to emergency exits and equipment. Exit routes are configured to ensure that no point within the facility exceeds maximum permitted travel distances to exits, typically 60 to 90 meters in industrial settings, with shorter distances maintained for high-hazard operations. The positioning of fire suppression equipment follows similar logic, featuring strategic placement of fire extinguishers at maximum distances of 23 meters in low-hazard areas and 15 meters in high- hazard zones, with unobstructed access.

[0337] Material storage zones constitute another element of proper spatial organization, featuring demarcation through floor markings, signage, or physical barriers that separate storage from operational and transit areas. Storage configurations may adhere to maximum height restrictions based on stability calculations, typically limited to 1.8 meters for manual handling or higher limits with mechanical handling equipment when properly secured. Storage arrangements may maintain minimum clearances of 0.5 meters from sprinkler heads, 1 meter from heating elements, and appropriate distances from light fixtures to prevent obstruction or fire hazards. For hazardous materials, additional segregation parameters specify minimum separation distances between incompatible substance classes, with specific clearances ranging from 2 to 5 meters depending on material reactivity and quantity.

[0338] Workflow organization parameters establish provisions for material staging zones positioned to minimize transport distances while preventing cross-contamination between process phases. These parameters typically specify maximum buffer volumes at workstations, limited to quantities that can be processed within a defined timeframe, usually 1-2 production hours, to prevent excessive material accumulation. Traffic lanes for personnel and material handling equipment may require minimum widths of 0.8 meters for pedestrian-only paths and 1.8 to 3.5 meters for vehicular routes, with wider dimensions needed at intersections and turns. These lanes are clearly delineated through floor markings, maintain appropriate sightlines at blind comers, and incorporate designated crossing points where pedestrian and vehicular paths intersect.

[0339] Equipment positioning follows ergonomic parameters that place frequently accessed controls, displays, and workpoints within optimal reach envelopes, typically between 0.6 and 1.7 meters from the floor depending on the operation. Heavy machinery may be positioned according to floor loading calculations, with weight distributed appropriately across structural support points and vibration isolation systems deployed as needed. Mobile equipment preferably has designated parking or storage positions when not in active use, with these locations selected to prevent obstruction of workflows, emergency systems, or utility access points. Infrastructure access parameters may stipulate minimum clearances around electrical panels (0.9 meters), utility shutoffs, and maintenance points to ensure immediate accessibility during both routine operations and emergency situations.

[0340] The system as described herein is preferably calibrated to recognize these spatial organization parameters across diverse industrial environments, identifying deviations such as encroachment on clearance zones, obstructed pathways, or improperly positioned equipment. By monitoring adherence to these predefined safety parameters over time, the system may detect deterioration in spatial organization before it reaches critical thresholds, enabling proactive intervention to maintain optimal safety conditions. This continuous assessment transforms spatial organization from a static compliance requirement to a dynamic component of operational safety management, where proper positioning and organization of materials and equipment becomes an integral element of the overall cleanliness and safety profile.

[0341] Foreign objects in industrial environments may be defined as objects, materials, or substances that are not intended to be present in a specific industrial area according to operational protocols, safety guidelines, or standard workspace organization. These include items that are either completely extraneous to industrial processes or operational items that have been displaced from their designated locations. Foreign objects represent potential hazards by creating obstructions, contamination risks, fire dangers, or workflow impediments.

[0342] Within industrial settings, foreign objects typically fall into several distinct categories that the system as described herein may be capable of identifying. Process-unrelated items comprise a significant category, including personal belongings such as food containers, beverages, and clothing items, as well as packaging materials like strapping, plastic wrap, and cardboard, or tools from unrelated work processes that have no functional purpose in the current operational area. The presence of such items not only indicates a deviation from cleanliness standards but may introduce contamination or fire risks in sensitive manufacturing environments.

[0343] Displaced operational materials constitute another category of foreign objects, comprising raw materials, components, or finished products that have fallen, spilled, or been improperly stored outside their designated containers, conveyors, or storage areas. While these materials are integral to operations, their incorrect placement transforms them into hazards and indicators of suboptimal cleanliness conditions. The system as described herein may be able to distinguish between properly positioned operational materials and those that have become displaced, potentially creating slip hazards or obstructions.

[0344] Maintenance debris represents a common foreign object category in industrial environments, including used parts, packaging from replacement components, cleaning materials, or maintenance tools left behind after repair or service activities. This category highlights the temporal aspect of foreign object detection, where items may be appropriately present during maintenance operations but become foreign objects when not removed upon completion of these activities. The persistence of such debris indicates incomplete maintenance protocols and compromises workplace cleanliness standards.

[0345] Production waste comprises scrap materials, byproducts, shavings, dust, or other process residues that have accumulated beyond acceptable levels or outside designated collection areas. Unlike displaced materials, these are expected byproducts of industrial processes but become foreign objects when they exceed threshold accumulation levels or appear outside designated containment systems. This category may represent a substantial fire risk in many industrial settings, particularly in wood processing, metal fabrication, or textile manufacturing, where combustible particulates may accumulate. Transient equipment constitutes another category, including temporary machinery, dollies, carts, or auxiliary equipment that should be returned to storage areas when not actively in use. These items, while necessary for operational flexibility, become foreign objects when left in walkways, emergency exit paths, or production areas after their immediate purpose has been fulfilled. The system as described herein is preferably able to assess their contextual appropriateness based on ongoing operations in the monitored area.

[0346] By comprehensively identifying these categories of foreign objects, the system may distinguish between items that legitimately belong in an industrial space versus those that constitute deviations from the ideal clean state. This granular understanding of foreign objects enables more precise cleanliness assessments, particularly when analyzing images for safety compliance in varied industrial contexts. The detection of such objects becomes a component in quantifying cleanliness as a spectrum rather than a binary condition, where the number, type, and distribution of foreign objects collectively determine how far a space has deviated from its ideal clean condition.

[0347] In regard to the above, features of the industrial environment preferably relate to such an ideal operational state with regard to each of the above categories that define the ideal vs the non-ideal industrial environment. An image of the industrial environment may be analyzed to determine such features. A trained Al model which is capable of recognizing such features may comprise an Al model that is specifically trained on images of the industrial environment, whether labeled or unlabeled, such that supervised or unsupervised learning may be employed. Alternatively, such a trained Al model may comprise such a model that is trained on many different types of images, such as a general or foundational Al model. As described in greater detail below, preferably deviations from the ideal operational state are analyzed by comparing images, whether of different parts of the same industrial environment, or the same or at least similar parts over time.

[0348] Figures 14-17 relate to the cleanliness of the environment as a risk factor. Cleanliness, or lack thereof, may contribute significantly to fire risk. For example, in facilities like sawmills, there is often a large amount of flammable sawdust as a byproduct of manufacturing. Sawdust when exposed to a sufficient heat source or spark is not only extremely flammable, but may be explosive under certain conditions. Therefore, an essential component in fire risk assessment is the day to day cleanliness factor. In traditional risk assessment, a risk engineer may visit the facility to determine for him / herself how clean the facility is. This is unfortunately only an instantaneous snapshot of the facility. With a network of cameras, however, cleanliness may be monitored 24 / 7 to dynamically adjust the risk levels based on the true conditions of the factory on a daily basis.

[0349] Images from the facility may be used to train a machine learning model which leams a baseline level of cleanliness and continuously assesses the cleanliness level. A preference model or human feedback model may be used to improve the performance of the algorithm over time, resulting in highly accurate metrics for determining the state of cleanliness within the facility. These metrics may include, but are not limited to, day-to-day fluctuations, dirt / dust content, or byproduct recognition. These metrics can then be used to determine a component risk score which we denote as the Cleanliness Score. The Cleanliness Score considers the aggregated metrics sigma and is calculated by summing a weight alpha and a bias term beta.

[0350] Furthermore, direct preference optimization (DPO) may be leveraged to learn how humans tend to label certain cases. This may be used to assess cleanliness levels, as an absolute metric of cleanliness may not be feasible based solely on the characteristics within the image. DPO leams how humans make subjective decisions, such as clean, cleaner, cleanest, dirty, dirtier, or dirtiest.

[0351] Direct Preference Optimization (DPO) is a streamlined alignment method for large language models (LLMs) that directly incorporates human preferences into training without requiring reinforcement learning or separate reward models. This approach supports alignment of Al systems with human values and expectations.

[0352] DPO operates through a core mechanism that uses binary preference data to adjust model behavior through a classification-based loss function. Unlike Reinforcement Learning from Human Feedback (RLHF), which requires training a separate reward model and applying complex reinforcement learning techniques, DPO bypasses these steps by reparameterizing the optimization problem. This allows the LLM's policy to be aligned directly with human judgments, creating a more straightforward path to model improvement.

[0353] The methodology relies on several components, beginning with preference data consisting of response pairs for given inputs. These pairs include a preferred response and a dispreferred response, typically derived from user feedback, comparative testing, or manual annotations by human evaluators. This preference data forms the foundation upon which the model leams to distinguish between desirable and undesirable outputs.

[0354] At the heart of DPO is a specialized loss function that maximizes the likelihood of preferred responses while simultaneously minimizing dispreferred ones. This function is derived from the Bradley-Terry preference model but reparameterized to avoid explicit reward modeling. The loss function incorporates both the policy being optimized and a reference model, typically the initial fine-tuned model, which serves to prevent drastic deviations from established behavior. A hyperparameter controls the degree of adherence to this reference policy, allowing for calibrated adjustments.

[0355] DPO incorporates regularization through an implicit KL-divergence term that constrains the policy to remain reasonably close to the reference model. This feature helps avoid unstable or nonsensical outputs that might otherwise emerge during optimization, ensuring the resulting model maintains coherence while improving along desired dimensions.

[0356] In practical implementation, DPO follows a straightforward workflow beginning with the collection of preference pairs for diverse prompts. Starting with a pre-trained or finetuned base LLM as the reference model, the system then updates the LLM's parameters using the specialized DPO loss function applied to the preference data. This process effectively steers the model toward generating responses more aligned with human preferences.

[0357] DPO has already demonstrated its effectiveness in models like Zephyr and NeuralChat, showing promise in tasks requiring nuanced alignment such as tone adjustment and summarization. Its reliance on high-quality preference data makes it helpful in scenarios where human judgment is critical but traditional reward modeling approaches would be impractical or overly complex.

[0358] Turning back to the Figures, Figure 14 shows an exemplary, non-limiting flowchart for a cleanliness monitoring process.

[0359] The method begins with visible light input 1404, which serves as the primary data source for the cleanliness assessment process. This input may be received from one or more cameras or imaging devices monitoring the environment.

[0360] The visible light input 1404 feeds into a cleanliness monitoring module 1406. This module performs comprehensive cleanliness monitoring functions. The cleanliness monitoring module 1406 may establish baseline measurements from Classical Risk Assessment CRA data, providing a reference point for cleanliness levels. It may track day-to- day changes in cleanliness, allowing for the identification of trends or sudden shifts in cleanliness conditions. The module may also perform outlier detection, identifying areas or instances where cleanliness levels deviate significantly from the norm. Additionally, the cleanliness monitoring module 1406 may monitor specific aspects of cleanliness, including the presence of dust, dirt, waste products, and byproducts in the monitored environment.

[0361] In some aspects, the Cleanliness Monitoring module may use images captured by the standard camerato assess the cleanliness of the facility. The images may be processed using computer vision algorithms to identify elements such as dust, debris, or other byproducts that may indicate a lack of cleanliness. The module may establish a baseline level of cleanliness based on the initial images captured by the camera, and may continuously monitor the facility to detect deviations from this baseline. The module may also adjust the parameters used for cleanliness assessment based on the detected byproducts, allowing for dynamic adjustment of the cleanliness standards.

[0362] The cleanliness monitoring module 1406 generates monitoring data 1408, which includes events, metrics, and anomalies identified during the analysis process. This comprehensive set of data serves as the input for the subsequent scoring stage.

[0363] The monitoring data 1408 is then processed by a clean scoring algorithm 1410. This algorithm evaluates the monitoring data to assess the overall cleanliness level and may take into account various factors such as the severity and frequency of cleanliness issues, the types of contaminants detected, and historical cleanliness trends. The algorithm may use weighted factors, such as the amount of dust or debris detected in the images, the frequency of cleanliness deviations, and the severity of the deviations. Each of these factors may be multiplied by their respective weights and summed with a bias term to calculate the cleanliness score. The weights and bias term may be predetermined or dynamically adjusted based on various factors, such as the specific characteristics of the facility or the historical data of cleanliness deviations.

[0364] Finally, the clean scoring algorithm 1410 produces a clean score output 1412. This output represents a quantified assessment of the cleanliness level based on the analyzed data, providing a clear indication of the overall cleanliness status in the monitored environment. The Clean Score may be a numerical value on a scale from 0 to 100. The Clean Score may be used to inform insurance underwriting decisions, potentially adjusting premium prices based on the assessed cleanliness levels. In some cases, the Clean Score may be combined with other risk scores, such as those derived from temperature data or fire, smoke, or spark events, to provide a more comprehensive assessment of the fire risk within a facility.

[0365] Various methods may be used to determine cleanliness. Figures 15A and 15B relate to a non-limiting example for determining the state of the facility, and the extent to which it deviates from the ideal operational state, by using game theory and the Elo system to determine cleanliness.

[0366] Without wishing to be limited by a closed list, an innovation of the present invention lies in its application of game theory principles to the assessment of industrial environments. Rather than using fixed thresholds, the system employs a competitive ranking approach inspired by methods used in competitive game rankings. This next section relates to this innovation and other innovations that the system uses when employing game theory principles. Of course, these innovations may also be used separately and / or in various combinations, including without limitation combinations with any of the previously described systems, methods and / or technologies provided herein.

[0367] Game Theory Based Competitive Ranking for Sensor-Driven Systems

[0368] Game theory provides a mathematically rigorous foundation for designing competitive ranking systems that dynamically assess the performance, reliability, or utility of agents: whether human players, Al agents, or, in this case, sensor components. By modeling sensors as “players” and their data outputs as “moves,” it is possible to build an algorithmic framework that continuously evaluates system components through structured, game-like interactions.

[0369] This approach allows for real-time, data-driven ranking and validation of hardware and software subsystems, particularly in large-scale environments like factories, warehouses, data centers, transportation hubs, or robotic fleets.

[0370] In systems where sensors output clearly defined, directly comparable values, such as temperature, voltage, or vibration RMS (Root Mean Square), traditional analytical or statistical methods may suffice. As a note, RMS is a common quantitative measure used to assess the intensity of vibrations, for example in rotating machinery like motors, pumps, and industrial equipment. In these cases, a simple magnitude comparison may provide a reliable answer to questions like “Which sensor runs hotter?” or “Is this motor vibrating more than usual?”

[0371] This game-theoretic ranking approach may be applied for example in domains where quantification is difficult, incomplete, or ambiguous. In these situations, the “ground truth” may be unknown or dynamic, and comparing raw values may not produce meaningful insights.

[0372] In this sense, the approach mirrors systems used in image-based ranking or preference-based platforms, such as dating apps or content recommendation engines. A historical example is Facemash, Mark Zuckerberg’s early web application, which used pairwise comparisons between photos of individuals to estimate "attractiveness", a metric nearly impossible to quantify objectively. While certain criteria may be able to be defined objectively such as facial symmetry, width, ocular spacing, etc. these features alone are not sufficient to make a decision on objective attractiveness. Rather than define attractiveness directly, the system let users decide which image was more appealing, then applied Elo-style updates to converge on a stable ranking for each user. Similarly, modem platforms like Tinder and Bumble use engagement-based ranking systems to match users with similar interaction profiles, not based on hard attributes but on dynamic, evolving signals of "success" (e.g., swipes, messages, time spent).

[0373] By reframing sensor evaluation as a similar pairwise preference task, the system is able to assess otherwise unquantifiable characteristics, such as which sensor is “behaving more normally,” or which one aligns better with a set of trusted reference behaviors under ambiguous conditions. The system does not assume that "normality" or any other quantitative characteristic may be defined up front. Instead, it lets the competitive interactions between sensors surface latent qualities that correlate with overall system health, redundancy, or trustworthiness. This may be valuable for example in multi-agent, multi-sensor environments, where some sensors may fail silently, drift slowly, or degrade under specific environmental conditions. It may also be valuable for computer vision, in which sensors capture images containing semantic information or meaning beyond the raw pixel values.

[0374] Framework Overview: Sensors as Strategic Players

[0375] In this framework, each sensor (or in broader terms, each subsystem within a network) is conceptualized as a strategic player in a repeated game. This approach borrows from classical game theory, in which players interact over time, making decisions (or moves) that impact not only their own standing but also the broader ecosystem in which they operate. Here, a "move" corresponds to a data point or a set of data captured by a sensor during a specific time interval. These moves are not isolated; rather, they are part of an ongoing competitive process through which the relative performance, reliability, or utility of each sensor is continuously assessed.

[0376] The system orchestrates structured head-to-head matches to evaluate and rank the sensors. One form of these matches involves direct comparisons between image sequences generated by cameras across a facility to obtain a cleanliness metric over time relative to other areas in the facility. In the comparison, the image is directly inspected for aspects of cleanliness or tidiness such as dust / dirt present in the image, waste, or hazardous materials that could be dangerous for humans. Over time, a competitive ranking can be established which directly measures how clean an area of the facility is relative to itself, as well as other areas in the facility. Furthermore, these competitions can be extended to other facilities to establish industry standard levels of cleanliness purely from historical camera images taking snapshots of the operational condition.

[0377] Each of these match types feeds into a competitive ranking algorithm, based on the Elo rating system and its modem extensions. These scoring systems translate the outcomes of sensor matches into dynamic numerical ratings that reflect the sensor’s comparative trustworthiness, accuracy, or consistency. As new data arrives and more comparisons are made, these ratings are automatically updated, creating a self-correcting and self-adaptive model of system-wide sensor quality. The result is a robust, real-time framework for assessing distributed components, capable of identifying underperforming sensors, highlighting anomalies, and improving the overall fidelity of sensor-driven decision-making across complex environments.

[0378] Elo Rating System: Core Mechanics

[0379] The classic Elo system may be used to quantify the relative strength of players (in this case, sensors) based on the outcome of a comparison. Each sensor is assigned a rating (e.g., 1500 baseline). When two data sets compete, the expected outcome is calculated using the logistic function:

[0380] EA-

[0381] Where RA and RB are the ratings of sensors A and B; and EA is the expected score for sensor A.

[0382] This logistic function outputs a value between 0 and 1, representing the probability that Player A will win the match. A rating difference of 0 gives an expected score of 0.5 (equal chance), while a larger difference pushes the probability closer to 1 (if RA>RB) or 0 (if RA< RB).

[0383] After comparison, actual outcomes (e.g., accuracy, agreement with ground truth, reliability) are used to update scores. A K-factor controls how responsive the system is to new data (e.g., higher K for early calibration, lower K for stable operation).

[0384] In a sensor network, a sensor that consistently aligns with peers and historical expectations will gain rating. However, a sensor that begins to deviate will lose rating, prompting potential alerts or recalibration.

[0385] Turning back to the drawings, Figure 15 A illustrates an exemplary, non-limiting method for evaluating cleanliness using multiple cameras. A method 1500 utilizes a game theory based competitive ranking system for cleanliness assessment.

[0386] The method 1500 begins with cameras 1502, labeled as Camera A, Camera B, and Camera C, capturing images 1504 at a given time. These images 1504 serve as inputs for the cleanliness assessment process.

[0387] In a pairwise comparisons step 1506, the method 1500 compares the images from different cameras against each other. For example, Image Al may be compared with Image Bl, Image Al with Image Cl, and Image Bl with Image Cl. This pairwise comparison approach allows for a relative assessment of cleanliness between different monitored areas.

[0388] Following the pairwise comparisons, the method 1500 proceeds to a vision evaluations step 1508. In this step, each comparison is assessed to determine which image appears cleaner, dirtier, or if there is no discernible difference between the images. This evaluation may be performed by a computer vision model or a trained human evaluator.

[0389] The method 1500 then advances to a match result step 1510. In this step, the outcomes of the vision evaluations are recorded as Win / Loss / Draw (W / L / D). These results form the basis for updating the cleanliness rankings of the monitored areas.

[0390] Based on the match results, the method 1500 proceeds to an Elo rating update step 1512. In this step, the cleanliness scores for each camera (representing different monitored areas) are updated using an Elo-based rating system. The Elo system, originally developed for chess rankings, is adapted here for cleanliness assessment. In this context, a "win" may result in an increase in the cleanliness score, while a "loss" may lead to a decrease.

[0391] Finally, the method 1500 concludes with an updated cleanliness scores step 1514. In this step, the system maintains a historical record of cleanliness levels for each monitored area based on the accumulated Elo rating updates. This allows for tracking cleanliness trends over time and comparing the relative cleanliness of different areas within a facility.

[0392] Figure 15B illustrates an exemplary flowchart for the assessment process within the game theory-based competitive ranking system for cleanliness evaluation, providing a detailed, exemplary implementation of the method described in Figure 15 A.

[0393] The process begins at step 1550 with "Start Assessment," which initiates the cleanliness evaluation workflow.

[0394] From there, it proceeds to step 1552, "Initialize System," where the system parameters and evaluation framework are set up for the cleanliness assessment process.

[0395] Next, at step 1554, "Add Cameras with Initial Elo=1500," the system incorporates cameras that will capture images for evaluation, assigning each camera an initial Elo rating of 1500. This establishes a baseline cleanliness score for all monitored areas, following the standard Elo rating system convention where 1500 represents an average or neutral rating.

[0396] The process then moves to step 1556, "Image Processing," where the images captured by the cameras undergo preparation and analysis to extract relevant cleanliness features.

[0397] Following image processing, the workflow advances to step 1558, "Elo Rating Update," where the system adjusts the cleanliness scores for each camera based on the pairwise comparison outcomes (wins, losses, or draws) as determined by the vision evaluation.

[0398] At decision point 1560, "More Camera Pairs?", the system checks whether additional camera comparisons need to be processed. If the answer is "Yes," the process loops back to step 1556 to process additional image pairs. This creates an iterative evaluation cycle that continues until all camera pairs have been compared.

[0399] When there are no more camera pairs to evaluate (the "No" path from decision point 1560), the process reaches step 1562, "End Assessment," which concludes the current cleanliness assessment cycle.

[0400] The method of Figure 15B shows a non-limiting exemplary implementation of the competitive ranking approach described in Figure 15 A, showing how the system systematically processes camera images, conducts pairwise comparisons, and updates Elo ratings to generate objective cleanliness assessments across different monitored areas within a facility.

[0401] Figures 15 A and 15B relate to the Elo system for determining ratings of cleanliness within a game theoretical framework. There are many variations on the Elo system, as well as non-Elo systems, which may be suitable for application within the devices, systems and methods of the present invention.

[0402] To make competitive ranking systems more adaptable and accurate in complex, real- world environments, several modem extensions have evolved from the original Elo framework. These enhancements are for example valuable in multi-sensor systems, where factors such as uncertainty, volatility, and non-linear relationships between data sources frequently occur. By incorporating advanced statistical methods and dynamic modeling, these extensions provide deeper insight into sensor performance and reliability. Some of these improvements and extensions are listed below, without wishing to be limited by a closed list, as any such variant, extension or related algorithm to Elo or any of the below extensions may be employed within the implementation of the present invention.

[0403] One such improvement is the Glicko and Glicko-2 system, which builds on Elo by introducing two critical components: rating deviation (RD) and volatility. Rating deviation reflects the system’s confidence in a rating for a given sensor. Sensors that are newly introduced or that provide inconsistent data will have a higher RD, signaling greater uncertainty in their assessed skill or reliability. This metric allows the system to weigh new or erratic sensor readings with appropriate caution. The Glicko-2 model adds an additional parameter: volatility. Volatility tracks how much a sensor’s performance (or rating) varies over time. For example, if a vibration sensor installed on a factory floor begins to produce erratic data after a period of stability, its volatility score will increase. Even if the average rating remains moderate, this rising volatility serves as a warning signal — prompting inspections or maintenance before a full failure occurs.

[0404] Another exemplary method is a-Rank, designed to handle ranking scenarios where outcomes are non-transitive. Traditional Elo assumes transitive skill relationships: if Sensor A outperforms Sensor B, and Sensor B outperforms Sensor C, then Sensor A should also outperform Sensor C. However, in many real-world sensor systems, especially under shifting environmental conditions, this assumption doesn’t hold. a-Rank addresses this complexity by modeling the set of observed outcomes as an irreducible Markov chain over the possible data strategies or outputs. It then analyzes these chains to identify stable ranking equilibria, even in the absence of a clear linear hierarchy. This is particularly useful in scenarios such as a network of humidity sensors responding differently under varying conditions. One sensor may perform best in cold conditions, another in high humidity, and a third in fluctuating temperatures. a-Rank captures these dominance cycles, allowing for contextual superiority rankings based on performance niches. A further enhancement comes from Bayesian Elo, which incorporates probabilistic reasoning to better manage uncertainty. Unlike the deterministic updates in standard Elo, Bayesian Elo treats each sensor’s skill estimate as a probability distribution rather than a single point estimate. This method uses Bayesian inference to combine prior expectations about a sensor’s reliability with newly observed data, updating the rating along with a credibility interval that represents confidence in that rating. This is especially useful in environments where data may be sparse, noisy, or occasionally missing. For example, in a remote warehouse with intermittent connectivity, the system may only receive limited updates from certain sensors. Bayesian Elo can still generate meaningful skill estimations, though with wider confidence intervals to reflect the lack of consistent data — giving operators a nuanced view of which ratings are solid and which are speculative.

[0405] Finally, dynamic Elo and online learning extensions provide tools for continuously updating sensor ratings in real-time, adaptive systems. These approaches are designed for non-stationary environments where conditions change rapidly, and sensor performance evolves in response. They often integrate with reinforcement learning models or edge-based Al systems, where trust scores and reliability metrics must be recalibrated frequently as new data flows in. Consider a fleet of autonomous delivery robots equipped with various environmental and positional sensors. As the robots traverse new terrain — be it uneven surfaces, inclement weather, or crowded urban areas — sensor modules can display different reliability patterns. A dynamic Elo model allows these reliability scores to adjust on the fly, ensuring that real-time decision-making is based on the most current and contextually relevant evaluations.

[0406] Collectively, these extensions transform Elo-based ranking from a basic competitive scoring tool into a robust, context-aware evaluation engine suitable for complex sensor- driven systems. They make it possible to not only compare hardware and software components in isolation but also to understand their behavior over time, under uncertainty, and in multi-dimensional competitive environments. These variations and extensions are also encompassed within the present invention.

[0407] Game Theory Beyond Elo for Ranking Systems

[0408] Beyond Elo and related systems, there are game-theoretic foundational systems that not only define how agents interact, but also govern how systems adapt over time through equilibrium dynamics, probabilistic inference, and stochastic processes. Such systems may also be implemented with the present invention in regard to the determination of deviations from an ideal operational state of an industrial environment, which is also referred to herein as “cleanliness”.

[0409] One non-limiting example of a framework within this foundation is Nash Ranking, which formulates the problem of determining ranks as a search for Nash equilibria in symmetric games. In this setting, each player (or agent) selects strategies such that no individual has anything to gain by deviating from their choice, assuming others maintain theirs. This concept becomes especially powerful in multi-agent systems, where the objective is not just to determine who is best, but to understand which strategy profiles are stable over repeated interactions. From a computational perspective, the problem becomes significantly complex when more than two players are involved, as it falls within the class of PPAD- complete problems: a class that includes problems guaranteed to have solutions but for which no efficient general algorithm is known. This complexity underscores the need for specialized algorithms and heuristics when implementing Nash-based ranking in practical systems involving three or more interacting components, such as distributed sensor arrays or robot swarms.

[0410] Another example of a suitable approach involves Markov chain methods, which model rating transitions as probabilistic state changes over time. In these systems, sensor or agent ratings are encoded within a transition matrix P(t), which evolves based on observed performance and competitive outcomes. Over time, this matrix converges to a stationary distribution, denoted as 71, such that:

[0411] This convergence represents a stable equilibrium of skill levels, in which each agent’s rating reflects its long-run performance relative to others. These models are particularly useful in systems where the environment is dynamic or outcomes are inherently stochastic, allowing for a smooth, time-evolving ranking process that reflects not just instantaneous performance, but broader performance trajectories.

[0412] A particularly advanced method that builds on Markov dynamics is a-Rank, which is specifically designed to handle non-transitive dominance cycles. Traditional ranking systems assume that if Sensor A outperforms Sensor B, and B outperforms C, then A should outperform C. However, in real-world systems, especially those involving autonomous agents or sensors in fluctuating environments, such transitivity often does not hold. a-Rank addresses this by constructing irreducible Markov chains over all strategy profiles and analyzing their stationary distributions to determine dominant behavioral patterns. It is especially effective in contexts like multi-agent reinforcement learning and evolutionary game theory, where agents constantly adapt to one another and no single optimal strategy exists. For sensor networks, a-Rank allows the system to capture niche-specific performance, for example by recognizing that one sensor is more accurate in high humidity, another in dry heat, and a third under varying loads, thereby allowing context-aware trust modeling.

[0413] Another noteworthy extension grounded in Bayesian game theory is TrueSkill™, a ranking system developed by Microsoft for multiplayer environments. It moves beyond Elo’s deterministic point estimates by modeling each agent's skill as a Gaussian belief distribution, characterized by a mean and variance. This probabilistic representation allows for more flexible and informative ranking updates, particularly in settings with partial observability or highly variable performance. One of TrueSkill's innovations is team skill aggregation, where the system combines the ratings of individual players (or components) to generate a composite score for the team as a whole. This is especially relevant in industrial systems where multiple subsystems or sensors contribute to a unified output, such as robots collaborating in assembly tasks or sensors monitoring different aspects of climate control in a data center. Additionally, TrueSkill includes a model for draw probability, which accounts for the likelihood of near-equal performance between agents. The draw margin is calculated using the inverse cumulative distribution function (<!> ') and incorporates both the uncertainty of individual players and a predefined sensitivity parameter:

[0414] Draw Margin

[0415] This allows the system to treat ambiguous or closely matched outcomes with nuance, rather than forcing a binary win / loss decision.

[0416] Finally, TrueSkill supports real-time updates through Bayesian approximation, enabling on-the-fly recalibration of agent skill levels as new data arrives. This is particularly advantageous in sensor networks or robotic platforms operating under live conditions, where decisions must be based on the most recent and reliable information. The combination of probabilistic reasoning, team-level modeling, and dynamic updates makes TrueSkill an effective tool for maintaining robustness and adaptability in complex, data-rich systems. One or more of these game-theoretic approaches (Nash equilibrium-based models, Markov processes, a-Rank, and Bayesian systems like TrueSkill) may form the mathematical backbone of competitive ranking for multi-agent systems. When applied to environments consisting of sensors, robots, software agents, or hybrid architectures, they provide a principled and flexible means of evaluating performance, detecting failure, and driving adaptive optimization, all within a dynamic and strategically structured ecosystem.

[0417] Operational Mechanics: How Matches May Work

[0418] As a nonlimiting example, the core of this game-theoretic sensor evaluation framework lies in its match protocols, which define how comparisons are made between data sets. These matches are the fundamental units of interaction in the system, akin to games in a tournament, where each outcome contributes to the evolving performance profile of the players; in this case, the sensors. Two primary types of match structures are employed to assess performance: cross-sensor matches and historical self-matches.

[0419] A cross-sensor match involves comparing data sets from different sensors that are measuring the same parameter at the same moment in time. For instance, multiple cameras may capture images of a facility from different vantage points over time, which may be compared. Comparisons across these images may highlight areas in the facility which fall below a certain cleaning or maintenance frequency. Similarly, these comparisons may also highlight areas in the facility which are well maintained and kept clean according to a regular interval.

[0420] In contrast, a historical self-match focuses on temporal consistency. It compares a sensor’s current reading to its own historical norms, such as long-term averages, trends, or acceptable variance ranges. This allows the system to track performance over time for a given sensor and enables the ability to answer the question “how consistent is this area being maintained or cleaned?”. For example, take a series of images from a single camera sensor mounted at a fixed location. The images may span various durations of time, but for the sake of the example assume we have 3 months of images of a facility. If a match were to take place between the image at day 90 versus the image at day 1, the system could make a decision on the cleanliness between these two images - i.e. the image taken at day 90 appears to be cleaner than the image taken at day 1. If these comparisons across the permutation of all possible image pairs were to be continued, a ranked preference which sorts images from cleanest to dirtiest would be obtained. Using the ELO mechanism, not only is a sorted preference over time obtained, but the cleanliness score for any given day, on any given camera, may be directly quantified. This enables the system to track housekeeping performance over time, and alert maintenance personnel when cleanliness deviations start to occur from a baseline norm such as a running mean or a variance based prediction model.

[0421] Each match yields either a binary outcome (e.g., win / loss based on closer alignment with expected values) or a probabilistic result (e.g., a likelihood score reflecting degrees of alignment or error margins). For example, the system may determine that Sensor A's reading was closer to a known ground truth or average cluster value than Sensor B’s, and assign the "win" accordingly. In anon-limiting example, an image from Sensor A may indicate a thin layer of dust over electrical equipment or machinery. An image from Sensor B may show a storage facility which has just undergone its weekly housekeeping. When prompted, the system can take the pixel values and semantic understanding of the image content into account and make a preference decision on which image it believes to be cleaner.

[0422] Alternatively, it may assign both sensors a partial score, or non-preferential “tie” if both deviated within acceptable limits. These outcomes are then used to update each sensor’s standing in the competitive ranking system, whether using Elo, Glicko, a-Rank, or Bayesian methods. Over time, this iterative process builds a high-resolution performance profile for every sensor in the system. Cleanliness assessment has similar challenges to defining “attractiveness”. For example, there may be some quantifiable or detectable signals within an image such as edge content (Laplacian), color segmentation, clutter recognition, these features alone are not sufficient to make an objective decision on which image represents a cleaner state. Only through relative comparisons to other images (historical snapshots or pairwise comparisons to other images across the facility) is the system able to establish a ranking which may be used to track housekeeping performance over time.

[0423] Rating Update Process

[0424] After each match, the system preferably updates the ratings for each camera based on the match results according to a competitive ranking algorithm. This algorithm may implement various rating systems including but not limited to the Elo rating system, Glicko-2 rating system, TrueSkill™ system, or Bayesian rating system.

[0425] In an implementation using the Elo rating system, for example, each camera is initially assigned a baseline rating (typically 1500). When two cameras are compared in a match, the expected outcome is calculated using a logistic function based on the current ratings of the cameras. After the match, the ratings are updated based on the difference between the expected outcome and the actual outcome. The updated ratings quantify the degree of physical deviation from safety parameters in each camera's monitored area. These ratings provide a continuous, objective measure of how closely each area maintains the ideal operational state, allowing for tracking of trends over time and comparison between different areas of the facility.

[0426] Threshold Determination and Alarm Triggering

[0427] Based on the updated ratings, the system may determine if the physical deviations from the ideal operational state exceed predetermined thresholds. These thresholds may be set based on industry standards, historical data, or specific requirements of the facility.

[0428] If the deviations exceed the threshold, the system preferably automatically triggers a physical alarm device, activates a safety mitigation system, or both. The safety mitigation system may comprise various components such as automated ventilation systems, emergency shutdown sequences, fire suppression systems, or access control systems. The specific action taken depends on the type and severity of the detected deviation.

[0429] By continuously monitoring conditions and automatically responding to deviations that exceed thresholds, the system provides an effective means of implementing automated control of safety conditions in the industrial environment, identifying and responding to conditions that increase risk before they develop into serious hazards.

[0430] Figure 16 illustrates a flowchart for an exemplary assessment method 1600. This method uses the previously described Elo method; however, as described above, a variety of different methods may be used for assessing the relative differences between images, whether from the same and / or different cameras. As noted above, such differences relate to deviations from the ideal operational state of the industrial environment as defined above. The method 1600 begins with a step 1602, where the system is initialized. Following initialization, the method 1600 proceeds to a step 1604, where a plurality of cameras present in the system are added with an initial Elo rating of 1500, as a camera initialization step. A different Elo rating may be used; furthermore, if the Elo system is not used, then a different type of camera initialization may be used.

[0431] The method 1600 then moves to a step 1606, where images from the cameras are stored. At step 1608, the system processes image comparisons. The method 1600 continues to step 1610, where it iterates through camera pairs.

[0432] At step 1612, the method 1600 checks if images are available. If images are available, the method 1600 proceeds to a step 1614, where images are compared with a Vision Model (VM). The method 1600 then moves to a step 1616 to update Elo ratings, followed by a step 1618 to save the updated ratings.

[0433] If images are not available at the step 1612, or after completing the step 1618, the method 1600 proceeds to a step 1620 to check if more pairs are available. If more pairs are available, the method 1600 returns to a step 1624 to move to the next pair. If no more pairs are available, the method 1600 proceeds to a step 1622, where the assessment ends.

[0434] In some cases, the Vision Model used in step 1614 may be a machine learning algorithm trained to evaluate image cleanliness. The Vision Model may compare two images and determine which one appears cleaner based on various visual features.

[0435] The Elo rating update in step 1616 may involve adjusting the ratings of the cameras based on the outcome of the image comparison. For example, if the Vision Model determines that an image from Camera A is cleaner than an image from Camera B, Camera A's Elo rating may be increased while Camera B's rating may be decreased.

[0436] The iterative process through camera pairs in the step 1610 may ensure that each camera is compared against multiple other cameras, providing a more robust assessment of relative cleanliness across the monitored areas.

[0437] In some cases, the method 1600 may be executed periodically, such as daily or weekly, to maintain up-to-date cleanliness ratings for all cameras in the system. The resulting Elo ratings may be used to identify areas that consistently maintain high cleanliness levels and those that may require additional attention or cleaning efforts.

[0438] Figure 17 illustrates an example of cleanliness ratings for five cameras tracked over a period of approximately one year. The graph in Figure 17 displays cleanliness ratings on the vertical axis, ranging from approximately 1200 to 1700, with time shown on the horizontal axis spanning from early 2024 to late 2024. Each camera is represented by a distinct line on the graph, labeled as camera l through camera_5.

[0439] The cleanliness ratings for different cameras may exhibit varying trends over time. For example, camera_2 and camera_5 demonstrate increasing cleanliness ratings throughout the monitored period. The ratings for these cameras start around 1400 and rise to approximately 1600-1700 by the end of the period. This upward trend may indicate improving cleanliness conditions in the areas monitored by these cameras.

[0440] In contrast, camera_3 shows a declining trend in cleanliness rating over the time period. The rating for camera_3 begins around 1400 and decreases to approximately 1200 by the end of the monitored period. This downward trend may suggest deteriorating cleanliness conditions in the area monitored by camera_3. Camera_4 and camera_l maintain relatively stable cleanliness ratings around 1400 throughout the monitored period. The consistent ratings for these cameras may indicate steady cleanliness conditions in their respective monitored areas.

[0441] The variations in cleanliness ratings between cameras and over time may provide insights into the effectiveness of cleaning procedures, environmental factors affecting cleanliness, or changes in operational conditions in different areas of an industrial facility. By tracking these ratings over extended periods, facility managers may identify areas requiring additional attention or evaluate the impact of implemented cleanliness measures.

[0442] In some cases, the cleanliness rating system may use a scale where higher values indicate cleaner conditions. The range of ratings from 1200 to 1700 may represent different levels of cleanliness, with 1700 potentially indicating very clean conditions and 1200 suggesting less clean environments.

[0443] The long-term tracking of cleanliness ratings, as illustrated in Figure 17A, may enable the identification of seasonal patterns, gradual changes in cleanliness conditions, or the effects of specific events or interventions on cleanliness levels in different areas of a monitored facility.

[0444] In some cases, a thermal image may include multiple detection regions with associated confidence scores. The thermal image may display eight detection regions marked by rectangular boxes, with seven regions having corresponding confidence scores.

[0445] Figure 17B relates to non-limiting examples of different images and how they may be assessed for cleanliness. These images show an industrial environment with significant amounts of wood, as well as machinery. In this example, at the far left, the image shows that the environment is clean. For example, in a reverse 0-100 score, in which lower scores indicate a cleaner environment, the score for the far left image could be 9 in this example. The middle image shows the same environment, but less clean. More wood dust is present, as well as pieces of wood that are not currently required for the industrial process, such that they represent clutter, which are examples of foreign objects. The score for the central image could be 47 in this example. The far right image shows the same environment in a dirty state, with a score of 89 for example, featuring many more such foreign objects. Thus, the methods and systems described herein are able to convert a more quality related characteristic (cleanliness) into a quantifiable characteristic, with a specific numeric value assigned. Game-Based Cleanliness Assessment

[0446] The game-theoretic framework offers an innovative and effective approach when applied to visual cleanliness monitoring in industrial and commercial facilities. This methodology models cleanliness assessment as a competitive ranking system, transforming what is traditionally a subjective evaluation into a quantifiable, consistent, and dynamic measurement process.

[0447] In this implementation, each camera within the facility functions as a strategic player in the competitive system. Unlike conventional approaches that might focus solely on the images themselves, this framework recognizes the camera as the persistent agent whose performance is evaluated continuously through the visual data it captures. The actual images represent discrete moves or strategies deployed by the player (camera) at specific moments in time, effectively documenting the visual state of the monitored area.

[0448] The system orchestrates head-to-head matches between periodically captured images, which occur along two primary dimensions of comparison:

[0449] • Between images from different cameras throughout the facility, which enables relative cleanliness assessment across distinct operational areas

[0450] • Against historical images from the same camera, allowing for absolute cleanliness tracking over extended time periods at a single location

[0451] For each match, the outcome is determined through evaluation by either an advanced computer vision model or a trained human expert. The evaluator assesses which image appears cleaner based on visual indicators such as dust accumulation, waste materials, spills, or general tidiness. This comparative judgment yields straightforward match results:

[0452] • A win when one camera's area is judged to be demonstrably cleaner

[0453] • A loss when another camera's area is assessed as less clean

[0454] • A draw when no significant difference in cleanliness is detected

[0455] Following each match outcome, the system updates each camera's cleanliness rating using the Elo algorithm's mathematical framework, or another framework as described above. In this context, higher ratings represent consistently cleaner operational areas, while declining ratings may signal deteriorating cleanliness conditions. These ratings evolve continuously as new "moves" (images) are evaluated, creating a dynamic measurement system that adapts to changing facility conditions.

[0456] Without wishing to be limited by a closed list, one aspect that distinguishes this approach from conventional cleanliness monitoring is its recognition that the camera, not the individual image, is the agent being evaluated. The images merely represent temporal snapshots of the area's cleanliness state (that is, of the output of the sensor, which in this case is a camera). By focusing on the camera as the persistent entity, the system can track longterm cleanliness trends while accommodating natural variations in lighting, positioning, and environmental factors.

[0457] Additionally, the game-theoretic framework avoids fixed thresholds or manual scoring systems, which may struggle with consistency and objectivity. Instead, cleanliness is assessed through relative performance over time via structured competitive evaluation. Without wishing to be limited by a closed list, this enables more sophisticated analyses, including trend identification, cross-area benchmarking, and the establishment of facilitywide cleanliness standards.

[0458] The competitive ranking approach may be extended beyond a single facility to establish industry-standard levels of cleanliness based purely on historical camera images documenting operational conditions. This creates the potential for standardized cleanliness metrics that may be applied consistently across multiple locations, providing comparative insights for operational excellence and compliance purposes.

[0459] By applying game theory to what has traditionally been a subjective visual assessment task, this framework delivers a robust, self-calibrating system that transforms cleanliness monitoring from periodic manual inspection to continuous, objective evaluation, thereby providing facility managers with unprecedented visibility into cleanliness conditions and trends across their operations.

[0460] Federated Learning Approach for Industrial Risk Assessment

[0461] The present invention may be implemented through a federated learning approach, for example for the previously described cleanliness assessment (assessment of deviations from the ideal operational state of the industrial environment) which enables collaborative model training across multiple industrial environments without transferring raw image data, thereby preserving privacy while improving the artificial intelligence model for risk assessment. This method is particularly valuable in scenarios where data cannot be centralized due to regulatory, logistical, or proprietary concerns, such as when monitoring multiple manufacturing facilities operated by different entities or in jurisdictions with strict data privacy regulations.

[0462] Federated Learning Implementation

[0463] In the context of the game-theoretic competitive ranking framework for industrial risk assessment, the federated learning process typically involves several iterative stages. First, a central server initializes a baseline global model for industrial environment feature recognition, potentially using publicly available datasets of industrial environments or pretrained weights. This model is distributed to participating client systems installed at various industrial facilities being monitored.

[0464] Each client system then trains the model using its local image data captured from the cameras physically positioned throughout its respective industrial environment. Importantly, the raw image data never leaves the client system, maintaining privacy and security of potentially sensitive industrial operations. The training adapts to local data patterns specific to each facility, such as particular machinery arrangements, operational workflows, or environmental conditions unique to that location.

[0465] The client systems subsequently send encrypted model updates to the central server, which aggregates these updates using methods such as Federated Averaging to create an improved global artificial intelligence model. This updated model, with enhanced capabilities to recognize industrial environment features and evaluate deviations from ideal operational states, is then redistributed to all client systems, and the cycle repeats at predetermined intervals to continuously refine the model's performance.

[0466] Agent-Based Learning Integration

[0467] The federated learning approach may be combined with agent-based learning techniques, wherein each industrial environment monitoring system functions as an autonomous agent. Each agent independently implements the game-theoretic competitive ranking framework described herein, with cameras functioning as players and captured images as moves, while simultaneously contributing to the collective improvement of the global artificial intelligence model.

[0468] This combined approach enables each monitoring system to maintain local optimization for its specific industrial environment while benefiting from the collective learning across all participating environments. The agent-based components allow for adaptive local responses to specific risk factors prevalent in a particular facility, while the federated learning component ensures that insights gained from diverse industrial scenarios improve the overall performance of the risk assessment system across all installations.

[0469] By implementing this federated learning approach, the present invention achieves significant advantages in detection accuracy and adaptability. Risk patterns identified in one industrial environment can inform assessment in others without compromising operational security. For example, a specific arrangement of materials identified as hazardous in one facility can be recognized in others before it results in an incident, even if that exact arrangement has never been observed in those environments previously.

[0470] Temporal Model Enhancement

[0471] The federated learning system may additionally incorporate temporal aspects, allowing the global model to improve through aggregated learning across multiple industrial environments over time periods with varying operational conditions. This enables the system to recognize seasonal patterns, cyclical risk factors, and gradual deviations that might otherwise go undetected in isolated monitoring systems.

[0472] Through this federated, agent-based approach, the artificial intelligence model evolves continuously, developing increasingly sophisticated capabilities to identify deviations from the ideal operational state in industrial environments, thereby enhancing safety risk assessment and mitigation across distributed manufacturing and processing facilities.

[0473] Further Risk and Hazard Assessment

[0474] Figure 18 illustrates a block diagram of a monitoring system for intruder detection and risk assessment. When individuals enter areas where, or when they are not supposed to, these detections may be included in another risk component which is shown as the Intruder Score. The system 1800 begins with two primary inputs: visible light input 1802 and infrared light input 1804, which serve as the data sources for the monitoring process.

[0475] The visible light input 1802 and infrared light input 1804 feed into an intruder monitoring module 1806. This module performs comprehensive intruder detection and analysis functions. The intruder monitoring module 1806 may conduct various types of detection:

[0476] Motion Detection: The module may analyze changes in pixel values or patterns across consecutive frames to identify movement within the monitored area. This may help detect potential intruders or unauthorized activity. Life Detection: By processing both visible and infrared data, the module may identify signs of living entities, such as heat signatures or characteristic movements associated with humans or animals.

[0477] Vehicle Detection: The intruder monitoring module 1806 may employ computer vision algorithms to recognize and track vehicles entering or moving within the monitored space.

[0478] Time / Duration / Frequency Analysis: The module may record and analyze temporal aspects of detected events. This may include logging the time of occurrence for each detection, measuring the duration of events e.g., how long an intruder remains in a specific area, and calculating the frequency of detections over time.

[0479] The intruder monitoring module 1806 generates monitoring data 1808, which includes events, metrics, and anomalies identified during the analysis process. This comprehensive set of data may encompass details about detected motions, life forms, vehicles, and temporal patterns of intrusions or suspicious activities.

[0480] The monitoring data 1808 is then processed by a risk scoring module 1810. This module implements an intruder risk scoring algorithm, which evaluates the monitoring data to assess the overall risk level associated with the detected intruder-related events.

[0481] Finally, the risk scoring module 1810 produces a risk score output 1812. This output represents a quantified assessment of the intruder-related risks based on the analyzed data, providing a clear indication of the potential security threats present in the monitored environment.

[0482] Figure 19 shows a flowchart for calculating a hazardous materials risk score calculation. Hazardous materials may include but are not limited to gas cylinders for welding, gas canisters, open chemicals, unattended batteries, running or broken vehicles, or oils. Such materials may be a major cause of accidents and fire. Information from the cameras may be used to capture images of the aforementioned items and alert necessary personnel. Machine learning algorithms like classifiers or SSDs may be trained to recognize these objects using openly available datasets, or data from individual facilities. Leaking gas cylinders, such as those used in welding, can also be detected using infrared sensors. Detections may be logged and assessed by facility personnel. These detections may be included in a risk component designated as the Hazardous Materials Score.

[0483] The system 1900 optionally begins with two primary inputs: visible light input 1902 and infrared light input 1904, which serve as the data sources for the monitoring process. Of course , other monitoring devices may be employed. The collected data may include temperature readings, images of the facility, and other relevant information that may be used to assess the risk associated with hazardous materials.

[0484] The visible light input 1902 and infrared light input 1904 feed into a hazardous materials monitor 1906. This module performs comprehensive detection and analysis functions for various hazardous materials and potentially dangerous objects. These analysis functions include monitoring parameters, which may include various factors such as the presence and quantity of hazardous materials, the storage conditions of these materials, and the status of vehicles that may pose a fire risk. The hazardous materials monitor 1906 may conduct detection and analysis of several key elements, which may include but are not limited to the following:

[0485] Oils: The module may detect the presence of oil spills or leaks, which may pose slip hazards or fire risks in certain environments.

[0486] Fuel canisters: The system may identify and track the location of fuel canisters, which may contain flammable or combustible materials.

[0487] Charging stations: The hazardous materials monitor 1906 may monitor electrical charging stations for signs of overheating or malfunction, which could lead to fire hazards.

[0488] Batteries: The module may detect and analyze the condition of batteries, particularly large industrial batteries, for signs of damage, leakage, or overheating.

[0489] Forklifts: The system may track the movement and operation of forklifts, which may carry hazardous materials or pose collision risks.

[0490] Chemicals: The hazardous materials monitor 1906 may identify and track the presence of various chemicals, including those that may be corrosive, toxic, or reactive.

[0491] Out of place objects: The module may detect objects that are not in their designated locations, which may indicate potential safety hazards or security risks.

[0492] The hazardous materials monitor 1906 generates monitoring data 1908, which includes events, metrics, and anomalies identified during the analysis process. This comprehensive set of data may encompass details about detected hazardous materials, their locations, quantities, and any unusual patterns or incidents related to these materials. Optionally, the data analysis may involve processing the collected data to extract relevant features and metrics for monitoring data 1908. For example, the data analysis may calculate the number of hazardous materials detected in the images, identify potential fire hazards associated with these materials, and determine the frequency and duration of detected events.

[0493] The monitoring data 1908 is then processed by a hazardous materials processor 1910. This module may perform more in-depth analysis of the detected hazardous materials, potentially correlating data from multiple sensors, applying material-specific risk models, or comparing current conditions to historical baselines.

[0494] Finally, the output from the hazardous materials processor 1910 is fed into a risk score processor 1912. This module implements a risk scoring algorithm that evaluates the processed hazardous materials data to assess the overall risk level associated with the detected hazardous materials and conditions.

[0495] The risk score processor 1912 produces a final risk assessment output, which represents a quantified evaluation of the hazardous material-related risks based on the analyzed data. This output may provide a clear indication of the potential dangers present in the monitored environment due to the presence, condition, or handling of hazardous materials. Such a risk assessment output may also relate to data aggregation. The data aggregation may involve combining the analyzed data over a specified time period or across multiple regions within the facility. The aggregated data may provide a more comprehensive view of the hazardous materials conditions and fire risk within the facility.

[0496] Risk weight factor and risk bias term may optionally then be applied to the aggregated data to produce the final risk score output. The risk weight factor may be a constant value that is multiplied with the aggregated data to adjust the final risk score. The risk bias term may be a constant value that is added to the weighted data to further adjust the final risk score.

[0497] The culmination of this process, as noted above, is the risk score output, which quantifies the overall fire risk associated with hazardous materials within the facility. The risk score output may be a numerical value on a scale from 0 to 100, with higher values indicating higher levels of fire risk. The risk score output may be used to inform insurance underwriting decisions, potentially adjusting premium prices based on the assessed risk levels. In some cases, the hazardous materials risk score calculation may be combined with other risk scores, such as those derived from temperature data or fire, smoke, or spark events, to provide a more comprehensive assessment of the fire risk within a facility.

[0498] Figure 20 demonstrates an exemplary program flow of how to use the long term history of these alerts as a means to aggregate information. As shown, the process may involve getting data from alarm tables, calculating metrics such as frequency, duration, confidence, and other metrics, and then updating risk scores. The alarm tables may store data related to previous alarms, including the time and date of each alarm, the location of the detected anomaly, and the confidence score associated with the alarm.

[0499] In some cases, the system may use a variety of computer vision algorithms to process the data from the cameras. These algorithms can detect temperature anomalies, smoke, fire, sparks, cleanliness levels, intruders, and hazardous materials. The use of these algorithms may enhance the accuracy and reliability of the anomaly detection process, enabling the system to detect a wide range of fire risks and adjust the risk scores accordingly. The system may employ any suitable devices and / or methods as described herein.

[0500] Turning now to the figure, Figure 20 shows a non-limiting exemplary method for processing alarm data and updating risk scores.

[0501] The method 2000 begins with getting data from alarm tables 2002. In this initial stage, the system may retrieve historical alarm data from a database or storage system, which may contain records of previous alerts, their types, and associated timestamps.

[0502] Following the data retrieval, the process moves to calculate frequency, duration, confidence, and metrics 2004 based on the retrieved alarm data. This step may involve analyzing the alarm data to determine how often alarms occur, how long they last, the system's confidence in each alarm's validity, and other relevant metrics that may provide insights into the nature and patterns of the alarms.

[0503] The metrics may be calculated based on the data from the alarm tables. These metrics may include the frequency of alarms, the average duration of detected anomalies, the average confidence score of the alarms, and other metrics that may be relevant to assessing the fire risk. These metrics may be used to update the risk scores, which may be used to inform insurance underwriting decisions and adjust premium prices based on the assessed risk levels.

[0504] After calculating these metrics, the method proceeds to aggregate metrics 2006. In this phase, the system may combine and summarize the calculated metrics to provide a comprehensive overview of the alarm data. This aggregation may involve statistical analysis, trend identification, or other data consolidation techniques.

[0505] The process then reaches a decision point where it evaluates changes in aggregate 2007. This step may involve comparing the newly aggregated metrics to previous data or predefined thresholds to identify significant changes or anomalies in the alarm patterns.

[0506] Based on this evaluation, the process branches into two paths. If changes are detected in the aggregate data that meet certain criteria, the method proceeds to send alarm 2008. This step may involve notifying relevant personnel or systems about the detected changes in alarm patterns, potentially indicating a shift in risk levels or the emergence of new threats.

[0507] After sending the alarm, or if no significant changes are detected, the process moves to create report 2010. This step may involve generating a comprehensive document or data summary that outlines the findings from the alarm data analysis, including trends, anomalies, and potential areas of concern. The report preferably includes the associated risk scores. The report may be used for auditing purposes, to inform safety measures, or to provide evidence of compliance with safety regulations.

[0508] Following the report creation, the method proceeds to update risk scores 2012. In this final stage, the system may adjust risk assessment values based on the insights gained from the alarm data analysis. This update may involve modifying risk scores for specific areas, processes, or overall facility safety, reflecting the most recent alarm data and analysis results. For example, the metrics may also be aggregated to provide a comprehensive view of the fire risk within the facility. The aggregation may involve combining the metrics over a specified time period or across multiple regions within the facility. If changes in the aggregate are detected, an alarm may be sent. The alarm may notify personnel of a potential increase in fire risk, prompting further investigation or action.

[0509] Figure 21 illustrates a block diagram of a monitoring system for integrating computer vision risk assessment with classical risk assessment to generate underwriting scores, pricing, and recommendations. The system comprises several interconnected modules that process and analyze risk-related data.

[0510] The monitoring system 2100 receives inputs from two primary sources: computer vision risk scores 2102 and classical risk assessment score 2104. The computer vision risk scores 2102 may be derived from the analysis of visual data captured by cameras or other imaging devices, potentially incorporating assessments related to fire risks, cleanliness levels, intruder detection, and hazardous material identification. The classical risk assessment score 2104 may be based on traditional risk evaluation methods, potentially including factors such as historical incident data, industry standards, and expert assessments.

[0511] These inputs feed into an underwriting score module 2106, which serves as the central processing unit for risk evaluation. CRA Score and the CV Risk Scores in order to dynamically re-assess premium prices for the insured client. The objective of underwriting score module 2106 is to provide a tailored pricing and recommendation experience based on the day-to-day operational status of the insured facility. The underwriting score module 2106 may provide recommendations to the insured on how to reduce their cumulative risk score, and subsequently, the prices they pay for the premium.

[0512] A unified risk score may be calculated from the above information which combines the CRA with the individual risk scores from each computer vision algorithm. Risk scores from the aforementioned algorithms may for example be pooled together as Computer Vision (CV) Risk Scores. Optionally, the system may use a fire risk scoring mechanism to quantify the level of risk, for example as a baseline score. This score may take into account long-term temperature trend analysis across a multitude of thermal cameras, events from standard cameras, and data from in-person risk assessments. The risk score may be used to dynamically adjust insurance premiums based on day-to-day operations, rewarding clients who maintain clean and well- maintained facilities and penalizing those who fail to take adequate prevention measures.

[0513] Also optionally, the system may use a classical risk assessment as a baseline score, which is then adjusted based on the day-to-day conditions and events detected by the system. The Classical Risk Assessment Score may consider factors such as the condition of fire alarm systems, or sprinklers, the building to determine if it is up to code, the nature of the facility (i.e. manufacturing, high fire risk activities, etc), or geographic location. This baseline score may then be modified using the individual CV Scores in order to provide a more accurate assessment of the true risk condition for the individual client.

[0514] The underwriting score module 2106 may perform a number of different types of calculations, including without limitation the following:

[0515] 1. Apply Baseline from CRA Score: The module may use the classical risk assessment score 2104 as a starting point or baseline for risk evaluation. This approach may leverage established risk assessment methodologies while allowing for enhancement through computer vision data.

[0516] 2. Periodically modify score based on CV scores: The underwriting score module 2106 may regularly update the risk assessment by incorporating the computer vision risk scores 2102. This periodic modification may allow the system to adapt to changing conditions detected through visual monitoring, potentially providing a more dynamic and responsive risk assessment.

[0517] 3. Adaptively change pricing based on new unified score: As the underwriting score evolves through the integration of classical and computer vision risk assessments, the module may adjust pricing recommendations to reflect the most current risk evaluation.

[0518] The underwriting score module 2106 provides outputs to two separate modules: a price module 2108 and a recommendations module 2110.

[0519] The price module 2108 may process the underwriting score data to determine appropriate pricing for insurance policies or risk management services. This module may take into account the unified risk score, potentially along with other factors such as market conditions or company policies, to generate pricing recommendations that accurately reflect the assessed risk levels. The recommendations module 2110 may analyze the processed data to generate actionable insights and suggestions for risk mitigation. These recommendations may include suggested improvements to safety protocols, changes in facility layout, or adjustments to monitoring practices based on the integrated risk assessment.

[0520] For example, if the suggestions provided by the recommendations module 2110 are followed, the system may adjust premium prices based on the assessed risk levels. For instance, facilities that maintain clean and well-maintained environments and exhibit fewer risk factors may be rewarded with lower insurance premiums. Conversely, facilities that fail to take adequate prevention measures and exhibit a higher number of risk factors may be penalized with higher insurance premiums. This dynamic adjustment of insurance premiums may provide a more accurate reflection of the actual fire risk within a facility.

[0521] Figures 22 and 23 show non-limiting examples of monitoring performed by exemplary systems as shown herein.

[0522] In some cases, as shown with regard to Figures 22A-C, thermal imaging systems may be used to monitor temperature variations in industrial environments. These thermal imaging systems may capture thermal image views that display temperature measurements across different regions of industrial equipment or machinery. In this example, a pellet press is monitored from a distance. The system picks up a hotspot in the ducting system (22C) and immediately alerts the maintenance personnel. Up to 7 minutes later, smoke appears in the visible light image (22B)

[0523] A thermal image view may include a temperature indicator 76.1 that provides a reference temperature reading. The temperature indicator 76.1 may be positioned in an upper portion of the thermal image to allow for easy visibility.

[0524] In some implementations, the thermal image view may display specific measurement regions corresponding to different parts of industrial equipment. For example, a press left 45.6 region may be defined to monitor temperatures in a left press area. Similarly, a right press 77.6 region may be established to track temperatures in a right press area. These measurement regions may be denoted by rectangular boxes overlaid on the thermal image.

[0525] The thermal image view may utilize a color gradient to represent temperature variations across the monitored area. In some cases, darker colors in the thermal image may indicate lower temperatures, while brighter colors may represent higher temperatures. This color-based visualization may allow for quick identification of temperature anomalies or hotspots. Temperature readings for the defined measurement regions may be displayed numerically within or near the corresponding rectangular boxes. For instance, the press left 45.6 region may show a temperature reading of 45.6 degrees, while the right press 77.6 region may display a temperature of 77.6 degrees. These numerical indicators may provide precise temperature data for specific areas of interest.

[0526] In some implementations, the thermal imaging system may combine thermal views with visible light views of the same area. This combination may allow for comparison between thermal patterns and standard visual information, enhancing the overall monitoring capabilities.

[0527] Multiple monitoring regions may be defined within a single thermal image view to track temperatures across various components or zones of industrial equipment. These monitoring regions may be strategically placed to measure temperatures at specific points or areas of interest within the field of view.

[0528] The thermal imaging system may include timestamp information and temperature scales to provide context for when measurements were taken and the range of temperatures being monitored. This contextual information may be useful for tracking temperature trends over time and identifying potential issues in industrial processes.

[0529] In some cases, the thermal image views may be part of a larger monitoring system that includes additional features such as alarm thresholds, data logging, and trend analysis. These features may work in conjunction with the thermal imaging capabilities to provide comprehensive temperature monitoring and risk assessment in industrial settings.

[0530] Figures 23 A and 23B show similar views. In these scenarios, a high-speed lumber planing machine is observed. The purpose of the machine is to trim the lumber to specific dimensions and surface tolerance at high-speed for maximum output. Due to the speed and friction induced by the cutting heads on the planing machine, sudden rises in temperature are not uncommon. In Figure 23A, we see a small puff of smoke in the image outlined with the red bounding box. The corresponding hotspot in the thermal image can be seen indicated by the orange bounding box.

[0531] Similarly in 23B, an off-cut of lumber is stuck in the outlet roller causing increased friction between the wood and the roller. This results in an increased temperature (indicated 103.7 C) of the roller and a substantial risk of fire.

[0532] Figure 24 illustrates an exemplary dashboard display with a line graph showing temperature data tracked over time in an industrial environment. The graph displays temperature measurements from August 24 to September 17, spanning approximately one month. Multiple colored lines represent different temperature metrics, with two horizontal threshold lines visible on the graph.

[0533] A red horizontal line indicates a threshold temperature value near 120 degrees Celsius. Such a threshold temperature may relate to an emergency or critical temperature limit for the monitored equipment or process. Below this, a yellow horizontal line shows another threshold value near 80 degrees Celsius, which may serve as a warning level temperature for example.

[0534] The primary temperature measurements, depicted by purple and turquoise lines, fluctuate between approximately 20 and 60 degrees Celsius throughout the recorded period. These temperature variations exhibit regular periodic patterns, suggesting cyclical processes or operations in the monitored environment.

[0535] In some cases, the temperature measurements approach but do not exceed the yellow threshold line, indicating that the system may operate near but within acceptable temperature ranges. In this non-limiting example, the graph shows no instances where the measured temperatures cross the red threshold line during the monitored period. If the red threshold line were to be crossed, preferably an alert would be sent, for example to a mobile device as previously described.

[0536] The periodic nature of the temperature fluctuations may correspond to daily operational cycles, batch processes, or other recurring activities in the industrial setting. The consistency of these patterns over the month-long period suggests stable and repeatable processes. Such stable and repeatable processes may be much easier to attain with such monitoring, as fluctuations outside of the desired band may be quickly detected.

[0537] A timestamp indicator on the graph shows "Sep 16, 12:00", providing a reference point for recent temperature data. The vertical axis of the graph displays temperature values ranging from approximately 10 to 120 degrees Celsius, encompassing the full range of observed and threshold temperatures.

[0538] This temperature monitoring over time may allow for trend analysis, early detection of temperature anomalies, and proactive maintenance scheduling based on temperature patterns.

[0539] In this example, the minimum temperature (cyan) decreased between the end of August and the middle of September which aligns with seasonal relationships based on the system’s geographic location. Alternatively, the maximum temperature (purple) increased during this same time period. Based on the anomaly conditions described in Figure 6, this condition was flagged. The clear visualization of temperature thresholds in relation to actual measurements may facilitate quick identification of potential overheating risks or process deviations.

[0540] Figures 25 A and 25B show a system for monitoring cleanliness in industrial environments which may track cleanliness ratings for multiple cameras over time.

[0541] Figure 25 A shows similarity metrics to determine if a camera viewpoint has been tampered with. The cameras as described herein may have a built-in tamper detection algorithm which can detect when a camera has been physically moved, or disturbed (lens covering). When global changes to the scene occur, a tamper alert may be sent out. In some cases where large motion is present in the scene, this may lead to false alerts. An example of this may be when a crane moves in front of the camera for an extended period of time and then moves out of the field of view as shown in the right hand image. The tamper detection alerts may be adjusted for such occurrences.

[0542] Figure 25B shows potential causes for alarms induced by solar radiation. Green bounding boxes indicate areas that are being analyzed in the left hand image. The red number “66” indicates a high temperature alert, which may however be due to solar radiation. The right hand image shows a photograph of the actual environment, demonstrating that solar radiation is the case. East and western facing outdoor cameras and / or cameras that face an exterior window may experience challenges with false alarms due to the presence of sunlight. Optionally, the cameras are angled downward and the camera shield is oriented forward to reduce the possibility of false alarms, and to avoid possible sensor damage. The alerts may be adjusted to account for such anomalies.

[0543] Figure 26 illustrates a thermal image view of a monitoring system showing multiple temperature measurement regions. The thermal image displays temperature variations across different areas, with warmer regions appearing brighter against a darker background representing cooler temperatures.

[0544] In some cases, a temperature measurement zone 33.7 may be defined within the thermal image to monitor a specific area of interest. The temperature measurement zone 33.7 may provide a reference temperature for comparison with other regions in the image.

[0545] The thermal image may include multiple monitoring regions overlaid on the thermal visualization. A first monitoring region 23.1 and a second monitoring region 27.6 may be positioned in the upper portion of the image. The first monitoring region 23.1 may display a temperature measurement of 23.1°C, while the second monitoring region 27.6 may show a temperature of 27.6°C. In some cases, a third monitoring region 24.5 and a fourth monitoring region 25.3 may be located in the middle section of the thermal view. The third monitoring region 24.5 may indicate a temperature of 24.5°C, and the fourth monitoring region 25.3 may display a temperature of 25.3°C.

[0546] The lower portion of the image may contain a fifth monitoring region 27.5 and a sixth monitoring region 27.0. The fifth monitoring region 27.5 may show a temperature measurement of 27.5°C, while the sixth monitoring region 27.0 may indicate a temperature of 27.0°C.

[0547] Each monitoring region may be depicted as a rectangular box, polygon, line, or singular point overlaid on the thermal image, with temperature measurements displayed within or near each region. The monitoring regions may be strategically placed to indicate the measurement of temperatures at specific points or areas of interest within the field of view.

[0548] In some cases, the thermal visualization may use a color gradient to represent temperature variations. Brighter areas in the thermal image may indicate higher temperatures, while darker areas may represent lower temperatures. This color-based representation may allow for quick visual identification of temperature patterns across the monitored area.

[0549] The thermal image may include additional information such as a timestamp and temperature scale. These elements may provide context for when the measurements were taken and the temperature range being monitored. In some cases, a menu or information panel may be displayed alongside the thermal image, showing temperature readings for various monitored components or regions.

[0550] Figure 27 displays an exemplary dashboard visualization of the comprehensive risk scoring system used for fire risk assessment. The interface presents the calculated risk score along with its contributing components in a user-friendly format. The dashboard shows the overall risk score prominently, providing facility managers and risk assessors with an at-a- glance understanding of the current fire risk level.

[0551] Below this aggregate score, the dashboard breaks down a plurality of contributing factors, including but not limited to: Temperature Scoring, which displays the current score related to high temperature incidents detected within the facility; Fire Scoring, which shows the score derived from the fire detection algorithm based on the frequency of fire events detected; Smoke Scoring, which presents the score generated by the smoke detection network's identification of smoke presence; and Cleanliness Scoring, which exhibits the score produced by the game theory -based competitive ranking system that evaluates facility cleanliness. Each component is displayed with its individual score and relative contribution to the overall CV (Computer Vision) Score, making it easy for users to identify which factors are most significantly affecting the current risk assessment. The dashboard may include visualization elements such as gauges, charts, or color-coding to provide intuitive understanding of risk levels. This allows facility managers to quickly identify areas of concern and take appropriate preventive measures. Furthermore, recommendations from the system may be provided to users in order to improve specific scores such as reducing their Response Time, improving their housekeeping frequency, or otherwise.

[0552] The dashboard may have a repository for long term temperature trend analysis, anomalies detected by the system, annual thermographic scans inside electrical cabinets, or annual risk reports provided by the user’s insurance.

[0553] CASE STUDIES AND FURTHER IMPLEMENTATION EXAMPLES

[0554] Case Study 1: Swiss Sawmill Fire Prevention System

[0555] This case study relates to exemplary installations, demonstrating at least some embodiments of the present invention in a real life situation.

[0556] A large Swiss sawmill company implemented experimental camera systems for fire prevention and predictive maintenance to demonstrate to their insurance provider that adequate measures were being taken to prevent fires. Ten cameras were initially installed as part of this demonstration.

[0557] The initial installation utilized wifi-based cameras. In some aspects, the system included a network of cameras strategically placed to monitor various areas within the facility. The cameras were configured to operate continuously, providing 24 / 7 monitoring of the facility. This continuous monitoring offered an advantage over periodic assessments, particularly in high fire-risk environments such as wood products, steel, chemical, battery, or food / agri cultural manufacturing facilities. The continuous monitoring provided a more accurate and dynamic assessment of the fire risk within the facility, enabling the system to detect temperature anomalies and adjust the fire risk score in real-time.

[0558] In some cases, the cameras were connected via a wireless network. This wireless connection allowed for flexible placement of the cameras and easy adjustment of the camera positions as the facility layout or operations changed. However, in other cases, the cameras were connected via a wired network to ensure a stable and reliable data transmission, especially in environments with high levels of electromagnetic interference. The cameras were equipped with various sensors to capture different types of data. For example, the cameras included infrared sensors to capture temperature data, visible light sensors to capture images of the facility, and other sensors to detect smoke, fire, sparks, cleanliness levels, intruders, and hazardous materials. This combination of sensors provided a more comprehensive assessment of the fire risk within the facility.

[0559] The system also included a data storage system, such as a cloud or on-premise server, to store the data collected by the cameras for later analysis. This data storage system allowed for long-term trend analysis and facilitated the detection of temperature anomalies over time. The data storage system also stored manufacturer-specific data, such as the location of the facility, the ambient temperatures, and other factors that affected the fire risk.

[0560] The system used a variety of computer vision algorithms to process the data from the cameras. These algorithms could detect temperature anomalies, smoke, fire, sparks, cleanliness levels, intruders, and hazardous materials. The use of these algorithms enhanced the accuracy and reliability of the anomaly detection process, enabling the system to detect a wide range of fire risks and adjust the risk scores accordingly.

[0561] After the initial wireless implementation, prototypes of a wired camera design were developed. The design was modified to use wired network connections to eliminate problems with magnetic interference from the large machines often present in manufacturing facilities. These wired network connections provided a stable and reliable data transmission, which was particularly beneficial in environments with high levels of electromagnetic interference.

[0562] The camera system included various components such as an infrared sensor, a visible light sensor, a microphone, a compute module, a network interface, an output, and an alarm. Each of these components contributed to the overall functionality of the camera system. For example, the infrared sensor captured temperature data, the visible light sensor captured images of the facility, the microphone captured audio data, the compute module processed the captured data, the network interface facilitated data transmission, the output displayed or transmitted information, and the alarm provided notifications or alerts.

[0563] Figures 24-26 show images from such sawmills and demonstrate how well the system worked, in the exemplary implementation.

[0564] Case Study 2: Austrian CNC Machine Monitoring

[0565] A subsequent installation monitored a CNC machine for window sill manufacturing in Austria. The camera system was configured to operate continuously, providing 24 / 7 monitoring of the specific area or process. This continuous monitoring offered an advantage over periodic assessments, particularly in high fire-risk environments such as wood products manufacturing.

[0566] The continuous monitoring provided a more accurate and dynamic assessment of the fire risk within the specific area or process, enabling the system to detect temperature anomalies and adjust the fire risk score in real-time. The camera system included various components such as an infrared sensor, a visible light sensor, a microphone, a compute module, a network interface, an output, and an alarm. Each of these components contributed to the overall functionality of the camera system.

[0567] The infrared sensor captured temperature data, the visible light sensor captured images of the facility, the microphone captured audio data, the compute module processed the captured data, the network interface facilitated data transmission, the output displayed or transmitted information, and the alarm provided notifications or alerts. The camera system used a variety of computer vision algorithms to process the data from the cameras. These algorithms could detect temperature anomalies, smoke, fire, sparks, cleanliness levels, intruders, and hazardous materials. The use of these algorithms enhanced the accuracy and reliability of the anomaly detection process, enabling the system to detect a wide range of fire risks and adjust the risk scores accordingly.

[0568] Case Study 3: Multi-Country Testing Results

[0569] This case study relates to test results from tests of experimental installations, according to aspects of the present invention. Experimental installations with over 250 individual cameras were deployed across Switzerland, France, Austria, Germany, Spain, Portugal, United States, and Canada Novel smoke, fire, and spark detection methods using machine learning methods, along with novel and improved computer vision methods, were developed.

[0570] One significant finding involved an infrared camera capturing a high temperature reading and a cool temperature reading, indicating the presence of a substantial heat source. In fact, the algorithms had detected individuals smoking within a customer facility from an infrared camera. This suggested that the infrared camera was capable of detecting temperature variations within a scene, which proved pivotal for early fire detection.

[0571] The infrared camera was configured to capture temperature data for every pixel in the thermal image. This data was used to calculate metrics such as minimum, maximum, and average temperatures for various regions within an image. The calculated metrics were compared against predefined or automatically learned thresholds to identify temperature anomalies. If a temperature threshold was exceeded, an alarm was triggered, potentially enabling early detection and prevention of fires.

[0572] The infrared camera was used to monitor high-risk areas in a facility. The camera was strategically placed to oversee areas where high temperatures were likely to occur, such as near machinery or equipment that generated heat during operation. By continuously monitoring these high-risk areas, the system was able to detect temperature anomalies and adjust the fire risk score in real-time, providing a more accurate and dynamic assessment of the fire risk within the facility.

[0573] Furthermore, over 100 standard network cameras were upgraded with smoke and fire detection capabilities.

[0574] Testing included ISO test fire TF2 specified by the ISO-7240 standard. In this test, an RGB camera monitored a hotplate which was smoking. The exemplary smoke detection algorithm being tested was able to recognize the visible characteristics of smoke and define a confidence level for the prediction. The infrared imaging showed the center of the hot plate dominating with over 200 degrees Celsius measured.

[0575] Additional testing included ISO test fire TF3 specified by the ISO-7240 standard. In this test, an RGB camera monitored a polyurethane pad which had ignited. The exemplary fire detection algorithm being tested was able to recognize the visible characteristics of fire and define a confidence level for the prediction. The infrared imaging clearly showed the fire with temperatures exceeding 210 degrees Celsius.

[0576] Additional Case Studies

[0577] Customer A experienced an issue where a saw-blade overheated and shattered. The detection system of the present invention was able to recognize the heat signature from the saw and alert the personnel that there was something wrong. Within 30 minutes, the issue was repaired. The information was sent to their insurance provider, both to demonstrate the efficacy of the system and to indicate their care in regard to prevention of fire risk.

[0578] Customer B experienced a situation where an old motor had an electrical fault. The fuse on the motor was overloaded, leading to a rapid rise in temperature. Before the fuse even blew, the customer had already received an emergency alert from our system, and they were able to repair the device immediately. An insurer may use the information from this alert to better price the premium.

[0579] Customer C was undergoing routine maintenance on a saw line. Maintenance in this case required a welding repair. Welding in a sawmill may be extraordinarily dangerous due to the large amount of combustible material present. The system of the present invention was able to detect when the repair person was conducting the repair, and ensured that no substantial temperature rise occurred after the repair was complete.

[0580] Customer D installed the system of the present invention, and after a few weeks the thresholds for a certain camera were adapting to the environment. During this adaptation period, an overhead crane was in the field of view of the camera. The crane did not move throughout the period, therefore the system was able to leam what the appropriate thresholds should be given the historical data. Then, suddenly the crane moved, revealing a very hot halogen light which is highly discouraged by risk engineers due to the large current draw and heat dissipation. The system of the present invention alerted the customer, and they were able to make adjustments.

[0581] Customer E installed the system of the present invention. The system detected irregular temperatures on a conveyor belt, which is an early warning signal for overstretching or a misaligned belt. Due to the timely intervention because of the system alerts, damage, production downtime and repair costs were prevented. Within the first 6 months alone, two such incidents were avoided.

[0582] Customer F installed the system of the present invention. The system detected an unusually high temperature of 76.1 °C at a pellet press and triggered an alarm at an early stage. Staff were on site within a few minutes and were able to eliminate the danger. This quick alert prevented a potential fire, as well as significant potential damages.

[0583] Customer G installed the system of the present invention. A defective bearing on a roller conveyor overheated to 140 °C within a very short time and was already starting to smoke. The system triggered an early alarm at 53 °C, allowing the danger to be quickly mitigated, despite the rapid overheating. Due to the rapid response, major damage and possible downtime were avoided.

[0584] Non-limiting, Illustrative Implementation Examples

[0585] The invention may be beneficial in industries with elevated fire risks, such as wood products manufacturing, food processing, and other industrial environments where combustible materials are present, without wishing to be limited by a closed list. As nonlimiting examples:

[0586] Sawmills face significant fire hazards due to accumulated sawdust throughout their operations. The system may continuously monitor cleanliness levels across the facility, detecting specific areas where sawdust is accumulating to dangerous levels. This enables maintenance teams to implement timely cleaning protocols before conditions become hazardous, particularly around cutting equipment, conveyor systems, and dust collection points where friction and heat may ignite particulates. Case Study 1 and Figures 24-26 relate to implementations at such sawmills.

[0587] Food processing facilities often generate combustible dust as a byproduct of production, especially in operations involving grain, sugar, flour, and other powdered ingredients. The system may monitor dust levels throughout production lines, storage areas, and ventilation systems, identifying areas requiring additional cleaning efforts before dust concentrations reach combustible thresholds. This proactive approach helps maintain compliance with safety regulations while preventing potentially catastrophic dust explosions.

[0588] Battery manufacturing involves volatile chemicals that may be extremely reactive, causing devastating fires that are difficult to extinguish. Lithium and other metal compounds used in modem battery production present unique fire risks that traditional systems may not adequately address. The system may monitor for signs of irregularities in lithium cell production, including temperature variations, gas emissions, and other precursors to thermal runaway events, enabling immediate intervention before conditions become critical.

[0589] Recycling and waste processing facilities face constant risk of fire due to an influx of battery cells being improperly disposed of in general waste streams. Material recovery facilities (MRFs) must also contend with other potentially dangerous items such as propane tanks, natural gas cylinders, and fuel containers that may be extremely hazardous during processing. The system may monitor for hotspots and detect dangerous items prior to their breakdown, allowing for safe removal before they enter shredders or compactors where impact could trigger explosions or fires.

[0590] Other manufacturing facilities may have multiple high-risk areas, such as those with electrical equipment or hot processes. The system may monitor these areas for temperature anomalies, smoke, or other signs of potential fire hazards. This includes environments like metal fabrication where welding operations create sparks, plastic manufacturing where static electricity may accumulate, or chemical processing where flammable materials are present. The continuous monitoring capabilities allow for customized detection parameters based on the specific risks present in each manufacturing context.

[0591] It is appreciated that certain features of the invention, which are, for clarity, described in the context of separate embodiments, may also be provided in combination in a single embodiment. Conversely, various features of the invention, which are, for brevity, described in the context of a single embodiment, may also be provided separately or in any suitable sub-combination.

[0592] Although the invention has been described in conjunction with specific embodiments thereof, it is evident that many alternatives, modifications and variations will be apparent to those skilled in the art. Accordingly, it is intended to embrace all such alternatives, modifications and variations that fall within the spirit and broad scope of the appended claims. All publications, patents and patent applications mentioned in this specification are herein incorporated in their entirety by reference into the specification, to the same extent as if each individual publication, patent or patent application was specifically and individually indicated to be incorporated herein by reference. In addition, citation or identification of any reference in this application shall not be construed as an admission that such reference is available as prior art to the present invention.

Claims

AMENDED CLAIMS received by the International Bureau on 01 October 2025 (01.10.2025)WHAT IS CLAIMED:

1. A system for monitoring an indoor industrial environment by determining deviations from an ideal operational state, the system comprising: a plurality of cameras physically positioned throughout the indoor industrial environment for capturing real-time image data; a processor; and a memory storing instructions that, when executed by the processor, cause the processor to process the image data from the plurality of cameras; characterized in that the memory storing instructions, when executed by the processor, cause the processor to:(a) establish a game-theoretic competitive ranking framework wherein each camera functions as a player and each captured image represents a move made by the respective camera, the framework configured to identify physical deviations from the ideal operational state;(b) perform head-to-head matches between: (i) images from different cameras for assessing relative physical deviations across different zones of the industrial environment, and (ii) current and past images from the same camera fortracking absolute physical deviations overtime;(c) evaluate, using a trained artificial intelligence model capable of recognizing industrial environment features, which image in each match is closer to the ideal operational state, wherein the ideal operational state comprises one or more of: (i) physical arrangement showing a presence of only material, machines, and humans directly involved in operational processes of the industrial environment, (ii) spatial organization of materials and equipment according to predefined safety parameters, and / or (iii) an absence of foreign objects; the evaluation yielding a match result comprising one of: a win when one camera's image is judged closer to the ideal operational state, a loss when the other camera's image is judged farther from the ideal operational state, or a draw when no clear difference is detected;(d) update a rating for each camera based on the match results according to a competitive ranking algorithm that adjusts each camera's rating based on the expected and actual outcomes of matches, wherein the rating quantifies the degree of physical deviation from the ideal operational states in each camera's monitored area;(e) determine if the physical deviations from the ideal operational state exceed a predetermined threshold based on the updated ratings; and(f) automatically trigger a physical alarm device or activate a safety mitigation system, or both, if the physical deviations exceed the threshold, thereby implementing automated control of safety conditions in the industrial environment by identifying and responding to conditions that increase risk.

2. The system of claim 1, wherein the ideal operational state further comprises:(a) predefined safety parameters for proper spatial organization of materials and equipment, wherein the predefined safety parameters include:(i) maintaining designated clearances around machinery, equipment, and electrical installations according to manufacturer specifications and regulatory standards;(ii) unobstructed emergency egress paths;(iii) designated material storage zones with maximum height restrictions and minimum clearances of from sprinkler heads and heating elements; and(iv) demarcated traffic lanes for personnel and material handling equipment; and(b) wherein the trained artificial intelligence model is configured to identify deviations from these predefined safety parameters by detecting encroachment on clearance zones, obstructed pathways, or improperly positioned equipment.

3. The system of claims 1 or 2, wherein the absence of foreign objects comprises the absence of:(a) process-unrelated items;(b) displaced operational materials;(c) maintenance debris;(d) production waste; and(e) transient equipment; wherein the trained artificial intelligence model is configured to distinguish between items that belong in the industrial environment versus those that constitute deviations from the ideal operational state.

4. The system of claim 3, wherein the absence of foreign objects comprises the absence of:(a) process-unrelated items including personal belongings, packaging materials, and tools from unrelated work processes;(b) displaced operational materials comprising raw materials, components, or finished products that have fallen, spilled, or been improperly stored outside their designated containers, conveyors, or storage areas;(c) maintenance debris including used parts, packaging from replacement components, cleaning materials, or maintenance tools left behind after repair or service activities;(d) production waste comprising scrap materials, byproducts, shavings, dust, or other process residues that have accumulated beyond acceptable levels or outside designated collection areas; and(e) transient equipment including temporary machinery, dollies, carts, or auxiliary equipment left in walkways, emergency exit paths, or production areas after their immediate purpose has been fulfilled; wherein the trained artificial intelligence model is configured to distinguish between items that belong in the industrial environment versus those that constitute deviations from the ideal operational state.

5. The system of any of claims 1-4, wherein updating the rating for each camera comprises: calculating an expected outcome for each match based on current ratings; comparing the expected outcome to the actual match result; and adjusting the ratings based on the difference between expected and actual outcomes.

6. The system of any of the above claims, wherein the plurality of cameras comprises at least one of visible light cameras, infrared cameras, thermal cameras, or multispectral cameras, or a combination thereof.

7. The system of any of the above claims, wherein the instructions further cause the processor to: generate a cleanliness score for each camera based on its updated rating; and display the cleanliness scores on a user interface.

8. The system of claim 7, wherein the instructions further cause the processor to: track changes in the cleanliness scores over time; and generate alerts if the cleanliness scores fall below a predetermined threshold.

9. The system of any of the above claims, wherein the instructions further cause the processor to: identify specific objects or conditions contributing to deviations from the ideal operational state; and include information about the identified objects or conditions in the triggered alarm.

10. The system of any of the above claims, wherein the instructions further cause the processor to: adjust the threshold for triggering the alarm based on historical data and patterns of deviations.

11. The system of any of the above claims, wherein the instructions further cause the processor to: perform image segmentation on the received image data to isolate different elements within the industrial environment.

12. The system of claim 11, wherein the instructions further cause the processor to: classify the isolated elements as either contributing to or deviating from the ideal operational state.

13. The system of any of the above claims, wherein the instructions further cause the processor to: apply different weightings to different types of physical deviations from the ideal operational state when updating the ratings.

14. The system of any of the above claims, wherein the instructions further cause the processor to: generate heatmaps visualizing the spatial distribution of deviations from the ideal operational state across at least a portion of the industrial environment.

15. The system of claim 14, wherein said portion of the industrial environment is defined according to a visual field of at least one camera in the system, andwherein said ideal operational state is defined according to data learned by an Al model according to a plurality of images received by said at least one camera over a period of time.

16. The system of any of the above claims, wherein the instructions further cause the processor to: implement a temporal analysis to detect gradual changes in the industrial environment over time that contribute to deviations from the ideal operational state.

17. The system of any of the above claims, wherein the instructions further cause the processor to: generate recommended actions to address identified deviations from the ideal operational state; and display the recommended actions on a user interface.

18. The system of any of the above claims, wherein the instructions further cause the processor to: adjust the frequency of head-to-head matches based on the rate of change in deviations from the ideal operational state.

19. The system of any of the above claims, wherein the safety mitigation system comprises at least one of an automated ventilation system, an emergency shutdown sequence, a fire suppression system, or an access control system; and wherein activation of the safety mitigation system comprises automatically controlling the operation of the corresponding system based on the type and severity of the detected physical deviations.

20. The system of any of the above claims, wherein the artificial intelligence model comprises a convolutional neural network, a visual transformer model or a variant thereof trained on images of ideal and non-ideal industrial environments.

21. The system of any of the above claims, wherein the game-theoretic competitive ranking framework implements an Elo rating system, Glicko-2 rating system, TrueSkill™ system, or Bayesian rating system.

22. The system of any of the above claims, wherein the instructions further cause the system to: implement a federated learning approach, an agent based learning approach or a combined approach thereof, to improve the artificial intelligence model using data from multiple industrial environments, from one industrial environment over a plurality of time periods or a combination thereof.

23. The system of any of the above claims, wherein the instructions further cause the processor to: generate a risk assessment score based on the updated ratings of the cameras; and incorporate the risk assessment score into an overall risk determination for the industrial environment.

24. A method for monitoring an indoor industrial environment by determining deviations from an ideal operational state by operating the system according to any of the above claims, the method comprising:(a) capturing real-time image data using the plurality of cameras physically positioned throughout the indoor industrial environment;(b) processing the image data to establish a game -theoretic competitive ranking framework wherein each camera functions as a player and each captured image represents a move made by the respective camera;(c) performing head-to-head matches between: (i) images from different cameras for assessing relative physical deviations across different zones of the industrial environment, and (ii) current and past images from the same camera for tracking absolute physical deviations overtime;(d) evaluating, using the trained artificial intelligence model, which image in each match is closer to the ideal operational state;(e) updating a rating for each camera based on the match results according to the competitive ranking algorithm;(f) determining if the physical deviations from the ideal operational state exceed the predetermined threshold based on the updated ratings; and(g) automatically triggering the physical alarm device or activating the safety mitigation system, or both, if the physical deviations exceed the threshold..

25. A system for monitoring and controlling safety risks in an indoor industrial environment by determining deviations from an ideal operational state, the system comprising: a plurality of cameras physically positioned throughout the indoor industrial environment for capturing real-time image data; a processor; and a memory storing instructions that, when executed by the processor, cause the processor to process the image data from the plurality of cameras; characterized in that the memory storing instructions, when executed by the processor, cause the processor to: (a) establish a competitive technical ranking framework wherein each camera functions as a monitoring node and each captured image represents a data point for spatial-temporal evaluation, the framework configured to identify physical deviations from the ideal operational state; (b) perform comparative evaluations between: (i) images from different cameras for assessing relative physical deviations across different zones of the industrial environment, and (ii) current and past images from the same camera for tracking absolute physical deviations over time; (c) evaluate, using a trained convolutional neural network implemented on the processor and configured to process the image data for recognizing industrial environment features, which image in each comparative evaluation is closer to the ideal operational state, wherein the ideal operational state comprises one or more of: (i) physical arrangement showing a presence of only material, machines, and humans directly involved in operational processes of the industrial environment, (ii) spatial organization of materials and equipment according to predefined safety parameters, and / or (iii) an absence of foreign objects; the evaluation yielding a comparison result comprising one of: a positive status when one camera's image is judged closer to the ideal operational state, a negative status when the other camera's image is judged farther from the ideal operational state, or a neutral status when no clear difference is detected; (d) update a technical safety rating for each camera based on the comparison results according to a competitive ranking algorithm that adjusts each camera's rating based on the expected and actual outcomes of comparisons, wherein the rating quantifies the degree of physical deviation from the ideal operational states in each camera's monitored area; (e) determine if the physical deviations from the ideal operational state exceed a predetermined safety threshold based on the updated technical safety ratings; and (f) automatically trigger a physical alarm device or activate a safety mitigation system, or both, if the physical deviations exceed the threshold, thereby implementing automated control of safety conditions in theindustrial environment by identifying and responding to conditions that increase risk.

26. The system of claim 25, wherein the ideal operational state further comprises: (a) predefined safety parameters for proper spatial organization of materials and equipment, wherein the predefined safety parameters include: (i) maintaining designated clearances around machinery, equipment, and electrical installations according to manufacturer specifications and regulatory standards, wherein the clearances range from 0.6 to 1.2 meters depending on equipment type; (ii) unobstructed emergency egress paths with minimum widths of 0.7 to 1.1 meters; (iii) designated material storage zones with maximum height restrictions and minimum clearances of 0.5 meters from sprinkler heads and 1 meter from heating elements; and (iv) demarcated traffic lanes for personnel and material handling equipment with minimum widths of 0.8 meters for pedestrian-only paths and 1.8 to 3.5 meters for vehicular routes; and (b) wherein the trained convolutional neural network is configured to identify deviations from these predefined safety parameters by detecting encroachment on clearance zones, obstructed pathways, or improperly positioned equipment.

27. The system of claim 25, wherein the absence of foreign objects comprises the absence of: (a) process-unrelated items including personal belongings, packaging materials, and tools from unrelated work processes; (b) displaced operational materials comprising raw materials, components, or finished products that have fallen, spilled, or been improperly stored outside their designated containers, conveyors, or storage areas; (c) maintenance debris including used parts, packaging from replacement components, cleaning materials, or maintenance tools left behind after repair or service activities; (d) production waste comprising scrap materials, byproducts, shavings, dust, or other process residues that have accumulated beyond acceptable levels or outside designated collection areas; and (e) transient equipment including temporary machinery, dollies, carts, or auxiliary equipment left in walkways, emergency exit paths, or production areas after their immediate purpose has been fulfilled; wherein the trained convolutional neural network is configured to distinguish between items that legitimately belong in the industrial environment versus those that constitute deviations from the ideal operational state.

28. The system of claim 25, wherein processing the visible light image data to detect smoke comprises: calculating image energy using Sobel filtering to identify regions in the visible light image data where texture and contrast have been reduced; comparing a current frame's energy map to a reference background frame to identify regions showing significant energy drops; generating bounding boxes around the identified regions; and classifying the identified regions using a trained neural network model and performing temporal consistency checks to reduce false positives.

29. The system of claim 25, wherein processing the visible light image data to detect fire comprises: performing motion detection via background subtraction on the visible light image data; applying color filtering to identify regions with fire-characteristic color patterns; extracting regions of interest from the filtered frames; and classifying the extracted regions using a neural network trained on fire and non-fire images.

30. The system of claim 25, wherein the instructions further configure the system to assess cleanliness levels within the industrial environment by: comparing current visible light images with historical images from the same locations; quantifying differences in visual characteristics between the current and historical images; and identifying areas requiring maintenance based on the quantified differences, wherein the identified areas are incorporated into the physical deviations for safety risk assessment.

31. The system of claim 25, wherein the instructions further configure the system to: detect unauthorized personnel in the industrial environment based on motion detection and human detection algorithms applied to the visible light image data; and include the detection of unauthorized personnel in the physical deviations for safety risk assessment.

32. The system of claim 25, wherein the instructions further configure the system to: store the thermal image data and the visible light image data in a database; analyze historical trends in the temperature metrics; and automatically adjust the predefined temperature thresholds based on the historical trends.

33. The system of claim 25, wherein the control signals include: alarm signals transmitted to response personnel when temperature anomalies exceed critical thresholds; automatic shutdown signals for equipment in areas where fire or smoke is detected; and notification signals indicating areas requiring maintenance based on the identified physical deviations.

34. A computer-implemented method for monitoring and controlling safety risks in an industrial environment comprising a plurality of cameras including at least one infrared thermal camera and at least one visible light camera, the method comprising: acquiring thermal image data from the at least one infrared thermal camera; calculating temperature metrics based on the thermal image data, wherein the temperature metrics comprise minimum temperature, maximum temperature, and average temperature values for defined regions within the industrial environment; detecting temperature anomalies by comparing the calculated temperature metrics with predefined temperature thresholds for the defined regions; acquiring visible light image data from the at least one visible light camera; processing the visible light image data using computer vision algorithms to detect at least one physical condition selected from: smoke, fire, occupancy, and presence of hazardous materials; combining the detected temperature anomalies and the detected physical conditions to identify physical deviations from predetermined operational safety parameters; and generating control signals based on the identified physical deviations to mitigate detected safety risks.

35. The method of claim 34, wherein processing the visible light image data to detect smoke comprises: calculating image energy using Sobel filtering to identify regions in the visible light image data where texture and contrast have been reduced; comparing a current frame's energy map to a reference background frame to identify regions showing significant energy drops; generating bounding boxes around the identified regions; and classifying the identified regions using a trained neural network model and performing temporal consistency checks to reduce false positives.

36. The method of claim 34, wherein processing the visible light image data to detect fire comprises: performing motion detection via background subtraction on the visible light image data; applying color filtering to identify regions with fire-characteristiccolor patterns; extracting regions of interest from the filtered frames; and classifying the extracted regions using a neural network trained on fire and non-fire images.

37. The method of claim 34, further comprising: storing the thermal image data and the visible light image data in a database; analyzing historical trends in the temperature metrics; and automatically adjusting the predefined temperature thresholds based on the historical trends.

38. The method of claim 34, further comprising: implementing a competitive ranking system to assess cleanliness levels within the industrial environment, wherein: a) current images of different facility locations are compared to historical images of the same locations; b) each location is assigned a cleanliness rating; and c) areas with declining cleanliness ratings are identified and incorporated into the physical deviations for safety risk assessment.

39. The method of claim 34, wherein generating control signals comprises at least one of: transmitting alarm signals to response personnel when temperature anomalies exceed critical thresholds; automatically shutting down equipment in areas where fire or smoke is detected; and generating notification signals indicating areas requiring maintenance based on the identified physical deviations.

40. A non-transitory computer-readable storage medium storing instructions that, when executed by a processor, cause the processor to perform the method of any one of claims 34 to 39.41 . A method for monitoring and controlling safety risks in an indoor industrial environment by determining deviations from an ideal operational state, the method comprising: capturing real-time image data using a plurality of cameras physically positioned throughout the indoor industrial environment; processing the captured image data using a processor executing instructions stored in a memory; establishing a game-theoretic competitive ranking framework wherein each camera functions as a player and each captured image represents a move made by the respective camera, the framework configured to identify physical deviations from predetermined operational safety parameters; performing head-to-head matches between: (i) images fromdifferent cameras for assessing relative physical deviations across different zones of the industrial environment, and (ii) current and past images from the same camera for tracking absolute physical deviations over time; evaluating, using a trained artificial intelligence model specifically adapted to recognize industrial environment safety conditions, which image in each match is closer to the ideal operational state, wherein the ideal operational state comprises one or more of: (i) physical arrangement showing only material, machines, and humans directly involved in operational processes of the industrial environment, (ii) proper spatial organization of materials and equipment according to predefined safety parameters, and (iii) an environmental parameter comprising at least temperature within process-specific optimal ranges; generating a match result comprising one of: a win when one camera's image is judged closer to the ideal operational state, a loss when the other camera's image is judged farther from the ideal operational state, or a draw when no clear difference is detected; updating a rating for each camera based on the match results according to a competitive ranking algorithm that adjusts each camera's rating based on the expected and actual outcomes of matches, wherein the rating quantifies the degree of physical deviation from safety parameters in each camera's monitored area; determining if the physical deviations from the ideal operational state exceed a predetermined threshold based on the updated ratings; and automatically triggering a physical alarm device or activating a safety mitigation system, or both, if the physical deviations exceed the threshold, thereby implementing automated control of safety conditions in the industrial environment by identifying and responding to conditions that increase risk.

42. The method of claim 41, wherein the trained artificial intelligence model comprises a convolutional neural network, a visual transformer model, or a variant thereof trained on a dataset comprising labeled examples of ideal and non-ideal industrial environments.

43. The method of claim 41, wherein the game-theoretic competitive ranking framework implements one or more of: an Elo rating system, a Glicko-2 rating system, a TrueSkill™ system, or a Bayesian rating system.

44. The method of any of claims 41 to 43, wherein updating the rating for each camera comprises: calculating an expected outcome for each match based on current ratings; comparing the expected outcome to the actual match result; and adjusting the ratings based on the difference between expected and actual outcomes.

45. The method of any of the above claims, wherein the plurality of cameras comprises at least one of visible light cameras, infrared cameras, thermal cameras, or multispectral cameras, or a combination thereof.

46. The method of any of the above claims, further comprising: measuring one or more environmental parameters using at least one sensor; incorporating the measured environmental parameters into the evaluation of the ideal operational state; and adjusting the rating updates based on the measured environmental parameters.

47. The method of any of the above claims, further comprising: generating a cleanliness score for each camera based on its updated rating; and displaying the cleanliness scores on a user interface.

48. The method of claim 47, further comprising: tracking changes in the cleanliness scores overtime; and generating alerts if the cleanliness scores fall below a predetermined threshold.

49. The method of any of the above claims, further comprising: identifying specific objects or conditions contributing to deviations from the ideal operational state; and including information about the identified objects or conditions in the triggered alarm.

50. The method of any of the above claims, further comprising: dynamically adjusting the threshold for triggering the alarm based on historical data and patterns of deviations.51 . The method of any of the above claims, further comprising: performing image segmentation on the received image data to isolate different elements within the industrial environment.

52. The method of any of the above claims, further comprising: classifying the isolated elements as either contributing to or deviating from the ideal operational state.

53. The method of any of the above claims, further comprising: applying different weightings to different types of physical deviations from the ideal operational state when updating the ratings.

54. The method of any of the above claims, further comprising: generating heatmaps visualizing the spatial distribution of deviations from the ideal operational state across at least a portion of the industrial environment.

55. The method of claim 54, wherein said portion of the industrial environment is defined according to a visual field of at least one camera in the system, and wherein said ideal operational state is defined according to data learned by the artificial intelligence model from a plurality of images received by said at least one camera over a period of time.

56. The method of any of the above claims, further comprising: implementing a temporal analysis to detect gradual changes in the industrial environment over time that contribute to deviations from the ideal operational state.

57. The method of any of the above claims, further comprising: generating recommended actions to address identified deviations from the ideal operational state; and displaying the recommended actions on a user interface.

58. The method of any of the above claims, further comprising: adjusting the frequency of head-to-head matches based on the rate of change in deviations from the ideal operational state.

59. The method of any of the above claims, wherein the safety mitigation system comprises at least one of an automated ventilation system, an emergency shutdown sequence, a fire suppression system, or an access control system; and wherein activation of the safety mitigation system comprises automatically controlling the operation of the corresponding system based on the type and severity of the detected physical deviations.

60. The method of any of the above claims, further comprising: implementing a federated learning approach, an agent-based learning approach, or a combined approach thereof, to improve the artificial intelligence model using data from multiple industrial environments, from one industrial environment over a plurality of time periods, or a combination thereof.61 . The method of any of the above claims, further comprising: generating a risk assessment score based on the updated ratings of the cameras; and incorporating the risk assessment score into an overall risk determination for the industrial environment.

62. The method of claim 61, further comprising: communicating the risk assessment score to an insurance underwriting system for dynamic adjustment of insurance premiums based on real-time safety conditions.

63. The method of any of the above claims, further comprising training the artificial intelligence model using a dataset comprising: images scraped from internet sources showing various industrial environments; synthetic images generated using generative models with prompting; and images with visual effects overlaid to simulate smoke, fire, or other hazardous conditions.

64. The method of claim 63, wherein training the artificial intelligence model further comprises applying data augmentation techniques including one or more of: artificial noise, geometric transformations, color adjustments, and exposure adjustments.

65. The method of any of the above claims, wherein evaluating which image is closer to the ideal operational state comprises: for fire detection, performing background subtraction to detect motion, applying color filtering for fire-typical hues, and extracting regions of interest for neural network analysis; and for smoke detection, calculating image energy using Sobel filtering, detecting regions with energy drops, and classifying these regions using a neural network.

66. The method of any of the above claims, wherein the memory further comprises instructions that, when executed by the at least one processor, cause the system to verify potential hazard detections using a foundation model that provides secondary analysis before triggering alerts.

67. The method of any of the above claims, further comprising storing confidence scores for detected hazards in a temporal buffer to track persistence of detections over time and reduce false positives.

68. The method of any of the above claims, wherein the competitive ranking algorithm incorporates volatility tracking to identify cameras showing increased variation in performance, which may indicate degrading conditions or sensor malfunction.

69. The method of any of the above claims, wherein performing head-to-head matches comprises conducting both cross-sensor matches between different cameras at the same moment in time and historical self-matches comparing a camera's current reading to its historical norms.

70. The method of any of the above claims, wherein the game-theoretic competitive ranking framework is configured to handle non-transitive outcomes where if Camera A outperforms Camera B, and Camera B outperforms Camera C, Camera A might not necessarily outperform Camera C.71 . A computer-implemented system for fire risk assessment in an industrial environment, the system comprising: at least one processor; and a memory coupled to the at least one processor, the memory comprising instructions that, when executed by the at least one processor, cause the system to: receive image data from a plurality of cameras monitoring the industrial environment; implement a game-theoretic ranking framework wherein cameras function as players and images function as moves in a competitive assessment of operational safety; generate head-to-head match results by comparing pairs of images to determine which image is closer to an ideal operational state representing safe operational conditions; update competitive rankings for each camera based on the match results using a rating algorithm that quantifies deviationfrom safety parameters; calculate a fire risk score based on the updated competitive rankings; and initiate preventive actions when the fire risk score exceeds a predetermined threshold.

72. The system of claim 71, wherein the memory further comprises instructions that, when executed by the at least one processor, cause the system to display a dashboard visualization presenting the calculated fire risk score along with contributing components including temperature scoring, fire scoring, smoke scoring, and cleanliness scoring.

73. The system of claims 71 or 72, wherein the method is implemented on an edge computing system comprising an on-premise server equipped with specialized GPUs that processes video feeds from the cameras in real-time and periodically uploads snapshot images for longer-term analysis.

74. The system of any of claims 71 to 73, wherein the system is configured to integrate with existing IP camera networks in the industrial environment, providing a plug-in upgrade that adds risk assessment capabilities without requiring replacement of existing camera infrastructure.

75. The system of any of claims 71 to 74, wherein the rating algorithm implements a competitive ranking method that adjusts ratings according to a mathematical formula that accounts for relative performance in pairwise comparisons.

76. The system of any of claims 71 to 75, wherein the memory further comprises instructions that, when executed by the at least one processor, cause the system to detect fire or smoke by analyzing the image data using a multi-stage detection pipeline comprising motion detection, color filtering, and neural network classification.

77. The system of any of claims 71 to 76, wherein the memory further comprises instructions that, when executed by the at least one processor, cause the system to detect smoke by performing energy-based image analysis to identify regions where texture and contrast have been reduced, followed by verification using a neural network classifier.

78. The system of any of claims 71 to 77, wherein the memory further comprises instructions that, when executed by the at least one processor, cause the system to: store a historical record of competitive rankings; and analyze trends in the competitive rankings to identify areas with deteriorating safety conditions.

79. The system of any of claims 71 to 78, wherein the memory further comprises instructions that, when executed by the at least one processor, cause the system to calculate the fire risk score by weighting different components according to industryspecific risk factors determined based on historical loss data.

80. A non-transitory computer-readable storage medium storing instructions that, when executed by a computing system, cause the computing system to: implement a game- theoretic competitive ranking framework for evaluating safety conditions in an industrial environment using image data from multiple cameras; perform pairwise comparisons between images to determine relative deviations from an ideal operational state; update competitive rankings for monitored areas based on the pairwise comparison results; identify potential safety risks based on changes in the competitive rankings; and initiate automated risk mitigation measures when identified risks exceed predetermined thresholds.81 . A system for automatic fire risk assessment, comprising: an infrared sensor configured to monitor high-risk areas within a facility and to provide temperature data for every pixel in a thermal image; a visible light sensor configured to oversee day-to- day operations within the facility and to visualize elements including at least one of dust, smoke, fire, occupancy, cleanliness, safety equipment, hazardous materials, or potential intruders; a computer processing system operatively coupled to the infrared sensor and the visible light sensor, configured to process information from the sensors and to relay this information to a data storage system; a data storage system configured to store long-term data of the facility, including temperature data from the infrared sensor and events detected by the visible light sensor; and a fire risk scoring mechanism configured to quantify the level of fire risk based on a long-term temperature trend analysis across a multitude of thermal cameras, events from standard cameras, and data from in-person risk assessments, wherein the fire risk scoring mechanism is further configured to dynamically adjust insurance premiums based on the quantified level of fire risk.

82. The system of claim 81, wherein the infrared sensor is further configured to calculate minimum, maximum, and average temperatures for various regions within an image and to trigger alarms in case of temperature threshold exceedance.

83. The system of claim 81, wherein the visible light sensor is further configured to process images using computer vision algorithms to determine the presence of a sufficient fire risk.

84. The system of claim 81, wherein the data storage system is further configured to store manufacturer-specific data affecting fire risk, including the location of the facility and ambient temperatures.

85. The system of claim 81, wherein the fire risk scoring mechanism is further configured to take into account events detected from images processed by the visible light sensor for calculating the fire risk score.

86. A method for automatic fire risk assessment, comprising the steps of: monitoring a facility using an infrared sensor and a visible light sensor; processing data from the sensors using a computer processing system; storing the processed data in a data storage system; and calculating a fire risk score using a fire risk scoring mechanism based on long-term temperature trend analysis, events detected from images, and data from in-person risk assessments, wherein the calculated fire risk score is used to dynamically adjust insurance premiums.

87. The method of claim 86, wherein the step of monitoring the facility includes capturing temperature data and images of the facility.

88. The method of claims 86 or 87, wherein the step of processing the data includes calculating metrics from the temperature data and analyzing the images to detect fire risks.

89. The method of any of the above claims, wherein the step of storing the processed data includes storing the data in a cloud or on-premise data storage system for later analysis.

90. The method of any of the above claims, wherein the step of calculating the fire risk score includes taking into account events detected from the images processed by the visible light sensor.91 . The method of any of the above claims, further comprising the step of configuring the infrared sensor to calculate minimum, maximum, and average temperatures for various regions within an image and to compare these metrics against predefined or automatically learned thresholds to trigger alarms in case of temperature threshold exceedance.

92. The method of any of the above claims, further comprising the step of configuring the visible light sensor to utilize computer vision algorithms to process images for the detection of elements indicative of fire risk, including dust, smoke, fire, and hazardous materials.

93. The method of any of the above claims, further comprising the step of utilizing the computer processing system to compare calculated temperature metrics against historical temperature data stored in the data storage system to identify temperature anomalies indicative of fire risk.

94. The method of any of the above claims, further comprising the step of utilizing the data storage system to store long-term facility data, including temperature trends and events detected by the visible light sensor, for use in subsequent fire risk score calculations.

95. The method of any of the above claims, further comprising the step of dynamically adjusting insurance premiums based on the fire risk score calculated by the fire risk scoring mechanism, wherein the adjustment rewards facilities with lower risk scores and penalizes those with higher risk scores.

96. The method of any of the above claims, further comprising the step of utilizing the fire risk scoring mechanism to integrate data from in-person risk assessments with sensor data to enhance the accuracy of the fire risk score.

97. The method of any of the above claims, further comprising the step of configuring the fire risk scoring mechanism to adjust the fire risk score based on manufacturerspecific data, including the location of the facility and ambient temperatures, which may affect the fire risk.

98. The method of any of the above claims, further comprising the step of employing the computer processing system to relay processed sensor data to a higher-level data storage system, wherein the data is analyzed for long-term trend analysis contributing to the fire risk score.

99. The method of any of the above claims, further comprising the step of configuring the data storage system to store data from a multitude of thermal cameras and standard cameras, wherein the stored data is analyzed for detecting trends and events that contribute to the overall fire risk assessment.

100. The method of any of the above claims, further comprising the step of implementing the fire risk scoring mechanism to utilize a weighted formula that aggregates quantified measures of detected risk factors, including temperature anomalies and events from standard cameras, to produce a unified fire risk score.

101. The method of any of the above claims, wherein the facility monitored is within the wood products industry, and the monitoring includes capturing temperaturedata and images specifically for the detection of dust accumulation that may pose a fire risk.

102. The method of any of the above claims, further comprising the step of analyzing the images captured by the visible light sensor to detect the presence of fine-grained wood by-products, wherein the detection of an accumulation beyond a predetermined threshold triggers an alert.

103. The method of any of the above claims, wherein the facility monitored is within the food processing industry, and the monitoring includes the use of the infrared sensor to detect temperature anomalies associated with processing equipment that may result in dust ignition.

104. The method of any of the above claims, wherein the facility monitored includes grain storage facilities, and the monitoring includes the use of the visible light sensor to detect changes in dust levels that may indicate a risk of combustion.

105. The method of any of the above claims, wherein the facility monitored includes flour storage facilities, and the monitoring includes the use of the computer processing system to analyze temperature and visual data for the early detection of smoldering or hot spots indicative of fire risk.

106. The method of any of the above claims, wherein the facility monitored involves storage of hazardous materials, and the monitoring includes the use of the infrared sensor to detect abnormal temperature readings indicative of chemical reactions that may pose a fire risk.

107. The method of any of the above claims, wherein the facility monitored is susceptible to electrical faults, and the monitoring includes the use of the infrared sensor to detect overheating electrical components that may lead to fires.

108. The method of any of the above claims, wherein the facility monitored is susceptible to arson, and the monitoring includes the use of the visible light sensor to detect unauthorized presence or activity within the facility during non-operational hours.

109. The method of any of the above claims, wherein the facility monitored involves hot work operations, and the monitoring includes the use of the infrared sensor to detect and track hot spots created during welding or cutting processes that may pose a fire risk.1 10. The method of any of the above claims, wherein the facility monitored includes areas where hazardous materials are handled or stored, and the monitoringincludes the use of the visible light sensor to detect spills or leaks of flammable or combustible materials.

111. A system for monitoring an industrial environment by detecting deviations from an ideal operational state, the system comprising: a camera network comprising a plurality of cameras positioned throughout the industrial environment for capturing image data; a computer vision processor configured to process the image data from the camera network; and a memory storing instructions that, when executed by the computer vision processor, cause the computer vision processor to: (a) analyze the image data using a plurality of monitoring modules to generate individual risk scores; (b) combine the individual risk scores using a weighted formula to generate a computer vision score; (c) receive a classical risk assessment score from a risk assessment module; and (d) provide the computer vision score and the classical risk assessment score to a combined scoring module that generates a unified risk score.

112. The system of claim 111, wherein the plurality of monitoring modules comprises at least one of: a temperature monitoring module configured to process temperature -related data; a fire / smoke / sparks detection module configured to detect fire, smoke, or spark events; a cleanliness monitoring module configured to assess cleanliness levels; a hazardous elements monitoring module configured to detect hazardous materials or conditions; and an intruder detection module configured to detect unauthorized personnel.

113. The system of claim 111 or 112, wherein the plurality of cameras comprises at least one infrared light sensor for capturing thermal information and at least one visible light sensor for capturing visible spectrum data.

114. The system of any one of claims 111 to 113, wherein the computer vision processor is configured to implement algorithms comprising at least one of: convolutional neural networks, vision transformer networks, isolation forest, semantic segmentation, generative adversarial networks, reinforcement learning with human feedback, autoencoders, or K-Nearest Neighbors.

115. The system of any one of claims 111 to 114, wherein the weighted formula for generating the computer vision score comprises weights (alpha), factors (sigma), and a bias term (beta).

116. The system of any one of claims 111 to 115, wherein the system further comprises an underwriting module configured to receive the unified risk score and adjust insurance premium prices based on the unified risk score.

117. The system of any one of claims 111 to 116, wherein the risk assessment module comprises components for evaluating different aspects of risk including at least one of: a geographic location component; a past incidents component; a fire safety component; a cleanliness component; and a building code component.

118. The system of claim 112, wherein the temperature monitoring module is configured to: measure minimum, maximum, and average temperatures; analyze temperature distributions; conduct trend analysis; trigger alarms; and detect anomalies based on the infrared data received.

119. The system of claim 118, wherein the temperature monitoring module is further configured to determine temperature anomalies over time by comparing information from the same sensor or group of sensors, or over space by comparing information provided from different sensors or sensor groups.

120. The system of claim 112, wherein the fire / smoke / sparks detection module implements a multi-stage fire detection pipeline including motion detection, color filtering, and region proposal.

121. The system of claim 120, wherein: the motion detection is implemented via background subtraction to identify areas of activity; the color filtering is applied to isolate pixels matching fire or smoke characteristics; and the region proposal techniques generate candidate areas for further analysis.

122. The system of claim 112, wherein the cleanliness monitoring module is configured to analyze visual imagery to detect dust, debris, or other indicators of cleanliness status.

123. The system of claim 112, wherein the hazardous elements monitoring module is configured to use computer vision techniques to identify potentially dangerous objects or situations within monitored areas.

124. The system of claim 112, wherein the intruder detection module is configured to use motion detection and object classification to identify human presence in restricted areas.

125. The system of any one of claims 112 to 124, wherein the system is configured to generate individual risk scores for each monitoring module, and wherein each riskscore comprises a numerical value on a scale from 0 to 100, with higher values indicating higher levels of risk.

126. The system of any one of claims 111 to 125, wherein the camera network is connected to a network switch that distributes data to a compute server and a video management system.

127. The system of claim 126, wherein the video management system is connected to recording storage for storing video data and to display devices for visualization of camera feeds and recorded content.

128. The system of any one of claims 111 to 127, wherein the system further comprises a microphone for audio capture capabilities.

129. The system of any one of claims 111 to 128, wherein the system further comprises an alarm for generating alerts or notifications based on detected conditions.

130. The system of claim 129, wherein the alarm is triggered by the computer vision processor based on analysis of sensor data or received instructions from a network.

131. A method for fire detection in an industrial environment using the system of any one of claims 111 to 130, the method comprising: receiving video images; performing motion detection on the video images via background subtraction; performing color filtering by applying HSV / RGB thresholds to identify fire-like colors; generating a region proposal by extracting connected components from combined masks; and applying a neural network to analyze the regions to detect fire.

132. A method for smoke detection in an industrial environment using the system of any one of claims 111 to 130, the method comprising: receiving an input image; calculating image energy using Sobel filtering; detecting energy drops by comparing the current frame's energy map to a reference background frame; generating bounding boxes around regions where significant energy drops have been detected; applying a neural network classifier to analyze the regions; and tracking classification results in a temporal buffer.

133. The method of claim 132, further comprising: determining if energy remains low; identifying a pre-alarm candidate if the energy remains low; and continuing monitoring if the energy does not remain low.

134. A method for analyzing temperature data in an industrial environment using the system of any one of claims 111 to 130, the method comprising: obtaining infrared data; converting the infrared data to temperature measurements; obtainingregions of interest and temperature limits; calculating metrics based on the temperature measurements and regions of interest; determining whether a temperature anomaly is detected; and sending an alarm if a temperature anomaly is detected.

135. The method of claim 134, further comprising storing the temperature data, calculated metrics, and any detected anomalies for future reference or analysis.

136. A method for analyzing long-term temperature data in an industrial environment using the system of any one of claims 111 to 130, the method comprising: obtaining long-term infrared data from a database; running statistical analysis on the long-term data; evaluating delta variance in temperature readings over time; running outlier and trend deviation detection; determining if anomalies are present; sending an alarm if anomalies are detected; updating temperature limits; and updating a temperature risk score.

137. A method for training a neural network for fire detection using the system of any one of claims 111 to 130, the method comprising: collecting image data from multiple sources including internet scraped images, synthetic images, and visual effects overlays; compiling the images into a unified training dataset; applying data augmentation techniques to the training dataset; tuning hyperparameters; setting training configuration; and training the neural network using the augmented dataset.

138. The method of claim 137, wherein the data augmentation techniques include at least one of: artificial noise, geometric transformations, color adjustments, compression artifact simulation, and Mixup / Cutmix augmentations.

139. The method of claim 137 or 138, wherein training the neural network comprises using an EfficientNet family model with an Adam optimizer and a cosine annealing learning rate scheduler.

140. A computer vision scoring system implemented on the system of any one of claims 111 to 130, the scoring system deriving a computer vision score from multiple individual scoring components comprising: a temperature score evaluating temperature-related data; a fire / smoke / spark score assessing detection of smoke, fire, and spark events; a cleanliness score quantifying cleanliness levels; a hazardous score evaluating hazardous conditions; and an intruders score tracking unauthorized access detection.

Images