A smart fire-fighting remote monitoring system based on Internet of Things
By deeply integrating multimodal sensing terminals and edge intelligent analysis units and using uncertainty modeling, the problem of high false alarm rate in traditional fire monitoring systems in complex environments has been solved, achieving highly reliable fire early warning and intelligent management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- YUXIN IND GRP CO LTD
- Filing Date
- 2026-03-19
- Publication Date
- 2026-06-19
AI Technical Summary
Traditional fire monitoring systems are prone to false alarms in complex industrial environments. They lack deep integration of multi-source data and uncertainty modeling, resulting in poor early warning robustness and difficulty in meeting the demand for highly reliable early fire warning.
Employing multimodal sensing terminals, edge intelligent analysis units, uncertainty modeling engines, and remote monitoring centers, the system uses deep neural networks and Bayesian deep learning frameworks for data fusion and uncertainty quantification. Combined with dynamic thresholds and hierarchical alarm strategies, it achieves multi-dimensional collaborative discrimination and reliable early warning of fire characteristics.
It improves the accuracy and reliability of fire early warning, reduces false alarm rate, reduces unnecessary evacuation costs and maintenance burden, and enhances the intelligence and scalability of fire monitoring system.
Smart Images
Figure CN122245060A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of Internet of Things and intelligent fire protection technology, specifically relating to an intelligent fire protection remote monitoring system based on the Internet of Things. Background Technology
[0002] With the deep integration of IoT and AI technologies, smart fire protection remote monitoring systems have become a core infrastructure for ensuring urban public safety and complex industrial production environments. By deploying high-density sensor networks and intelligent analysis terminals, smart fire protection achieves real-time monitoring and closed-loop management of fire hazards, shortening the fire response cycle.
[0003] The IoT-based smart fire remote monitoring system, through the collaboration of visual perception modules and environmental parameter sensors, is committed to achieving in-depth analysis and multi-dimensional judgment of early fire characteristics. Especially in monitoring scenarios with high dynamic environmental interference, such as tunnels and factory areas, higher requirements are placed on the system's data processing accuracy, feature fusion depth, and recognition robustness under complex working conditions.
[0004] Traditional fire monitoring systems typically rely on single physical parameter thresholds for fire detection. When faced with interference from water mist, dust, welding sparks, or momentary intense light in industrial environments, they are prone to frequent false alarms, leading to complacency among on-duty personnel or incurring significant evacuation costs. Existing image recognition modules and environmental sensor arrays often operate in isolated analytical states, lacking deep fusion mechanisms at the feature level. This makes it difficult to establish a spatiotemporal correlation mapping between visual textures and physical indicators, resulting in insufficient ability to capture the nonlinear evolution logic of fires.
[0005] Traditional early warning models generally lack the ability to model the uncertainty of the perception process and cannot quantify the decision confidence when there are conflicts in multi-source data. This results in poor early warning robustness in extreme or edge application scenarios, making it difficult to meet the practical application requirements of high-reliability early fire warning. Summary of the Invention
[0006] The purpose of this invention is to provide an intelligent fire protection remote monitoring system based on the Internet of Things, which can effectively solve the problems mentioned in the background art.
[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A smart fire protection remote monitoring system based on the Internet of Things includes a multimodal sensing terminal, an edge intelligent analysis unit, an uncertainty modeling engine, a remote monitoring center, and a communication network module. The multimodal sensing terminal is used to simultaneously collect visual image data and environmental physical parameters on site. It includes a high-definition video acquisition device and an environmental sensor array. The high-definition video acquisition device is used to acquire dynamic images within the monitoring area in real time, and the environmental sensor array is used to continuously monitor temperature change curves, smoke concentration, carbon monoxide and carbon dioxide gas concentrations. The edge intelligent analysis unit is connected to the multimodal sensing terminal and is used to extract features from the collected visual image data and perform time-series analysis on environmental physical parameters. The edge intelligent analysis unit has a built-in deep neural network model to identify the dynamic texture features of flames and the diffusion morphology features of smoke, and generate corresponding visual confidence indices. The uncertainty modeling engine is connected to the edge intelligent analysis unit and is used to receive the temporal characteristics of the visual confidence index and environmental physical parameters. Based on the Bayesian deep learning framework, it quantifies the uncertainty of the current environmental state. When there is a conflict in the spatiotemporal dimension of multi-source data or the confidence is lower than the preset threshold, the event is marked as a low-confidence suspected event. A fire warning signal is generated only when the multimodal data is highly consistent and the comprehensive confidence exceeds the dynamic threshold. The remote monitoring center is connected to the uncertainty modeling engine through the communication network module. It is used to receive and store the fire early warning signal and the corresponding raw sensing data, push hierarchical alarm information to the user terminal, and support remote access to historical event records and system operation status. The communication network module is used to establish a stable, low-latency data transmission channel between the multimodal sensing terminal, the edge intelligent analysis unit, the uncertainty modeling engine, and the remote monitoring center, ensuring the synchronous uploading of multi-source heterogeneous data and the real-time issuance of instructions.
[0008] Preferably, the high-definition video acquisition device employs an imaging unit with wide dynamic range and low-light enhancement capabilities, which can maintain effective capture of smoke and flame characteristics in environments with drastic changes in light or dim lighting.
[0009] Furthermore, the environmental sensor array integrates a temperature sensor, a photoelectric smoke sensor, an electrochemical carbon monoxide sensor, and an infrared carbon dioxide sensor. Each sensor has a self-calibration function to eliminate drift errors caused by long-term operation.
[0010] Furthermore, the deep neural network model used by the edge intelligent analysis unit is a visual Transformer architecture based on the attention mechanism or an improved convolutional neural network structure. Its training process incorporates a large number of interference samples from complex industrial scenarios to improve the ability to identify non-fire interference sources such as water mist, dust, welding sparks and strong light reflection.
[0011] Preferably, the uncertainty modeling engine introduces a Monte Carlo Dropout mechanism or variational inference method to sample the probability distribution of the model output multiple times without increasing computational overhead, thereby obtaining the uncertainty range of the current judgment result and dynamically adjusting the triggering conditions of the early warning decision accordingly.
[0012] Furthermore, the dynamic threshold is adaptively adjusted based on historical environmental baseline data and current operating conditions. When the system detects that the environment is in a high-interference state, it automatically increases the confidence threshold to suppress false alarms, while during periods of stable environment, it appropriately decreases the threshold to ensure a sensitive response to early fires.
[0013] Furthermore, the remote monitoring center is equipped with an event classification management module, which can classify alarm information into three levels—low confidence suspected, medium confidence attention, and high confidence confirmation—based on the confidence level output by the uncertainty modeling engine, and assign different handling strategies and notification priorities to each level, thereby avoiding a decrease in operator vigilance due to frequent low-level alarms.
[0014] Compared with the prior art, the present invention has the following beneficial effects: 1. The IoT-based smart fire remote monitoring system provided by this invention breaks through the limitation of the separation between vision and perception in traditional systems by deeply integrating visual images and environmental sensing data at the feature level, and realizes multi-dimensional collaborative discrimination of early fire characteristics.
[0015] 2. This invention introduces a Bayesian deep learning framework to quantitatively model the uncertainties in the perception process, which can effectively distinguish between real fires and common interference sources such as water mist, dust, and welding sparks in complex industrial environments, reduce false alarm rates, and avoid emergency response failures caused by the "boy who cried wolf" effect.
[0016] 3. This invention adopts a dynamic threshold mechanism and an event-level alarm strategy, which takes into account both the sensitivity and reliability of the early warning, and reduces unnecessary evacuation costs and maintenance burdens while ensuring the safety of high-risk locations.
[0017] 4. The collaborative architecture of this invention not only ensures local real-time response capabilities but also supports centralized cloud management and big data backtracking analysis, thereby improving the intelligence level and scalability of the entire fire monitoring system. Attached Figure Description
[0018] Figure 1 This is a schematic diagram of the overall technical solution architecture according to the present invention; Figure 2 This is a schematic diagram of the core principle framework of multimodal data fusion and fire early warning based on an uncertainty modeling engine according to the present invention; Figure 3This is a flowchart illustrating the logical flow of the edge intelligent analysis unit in this invention for feature extraction and temporal analysis of visual images and environmental parameters. Figure 4 This is a flowchart illustrating the logical process of quantifying environmental state uncertainty and determining dynamic thresholds based on Bayesian deep learning according to the present invention. Figure 5 This is a schematic diagram of the multi-level interaction relationship and hierarchical alarm data flow between the remote monitoring center and each module according to the present invention. Detailed Implementation
[0019] Example 1: Please refer to the appendix Figure 1 To be continued Figure 5 To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.
[0020] A smart fire protection remote monitoring system based on the Internet of Things includes a multimodal sensing terminal, an edge intelligent analysis unit, an uncertainty modeling engine, a remote monitoring center, and a communication network module. The multimodal sensing terminal is used to simultaneously collect visual image data and environmental physical parameters from the site. The edge intelligent analysis unit is used to extract features from the collected visual image data and perform time-series analysis on the environmental physical parameters to generate corresponding visual confidence indices.
[0021] The uncertainty modeling engine is used to quantify the uncertainty of the current environmental state based on the Bayesian deep learning framework and generate fire early warning signals; the remote monitoring center is used to receive and store fire early warning signals and raw sensing data, and push graded alarm information; the communication network module is used to establish data transmission channels between the modules.
[0022] The multimodal sensing terminal is deployed in a controlled industrial environment or inside a public building. It integrates a high-definition video acquisition device and an environmental sensor array, achieving tight hardware-level coupling through a physical interface. The high-definition video acquisition device uses an imaging unit with wide dynamic range and low-light enhancement capabilities. Its dynamic range parameter is configured to be no less than 120 dB, ensuring that even in extreme environments with strong backlight or extremely low light, it can still clearly capture the translucent edge features of smoke and the core texture information of flames.
[0023] The high-definition video acquisition device is configured to support high-definition image output of no less than 30 frames per second, and pushes the raw video stream to the edge intelligent analysis unit in real time via a MIPI interface or an Ethernet interface. The environmental sensor array includes multiple complementary physical sensing dimensions, specifically a temperature sensor, a photoelectric smoke sensor, an electrochemical carbon monoxide sensor, and an infrared carbon dioxide sensor. The temperature sensor uses a highly sensitive thermistor element, with a sampling frequency set to 10 Hz, capable of capturing minute temperature fluctuations on the order of 0.1 degrees Celsius per second, generating a continuous temperature change curve.
[0024] The photoelectric smoke sensor utilizes the Mie scattering principle and, through the spatial arrangement of an infrared LED and a photodetector, achieves real-time monitoring of the concentration of fine particulate matter in the air, converting it into a standard voltage signal. Both the electrochemical carbon monoxide sensor and the infrared carbon dioxide sensor are equipped with self-calibration circuits. These circuits automatically compensate for zero-point drift errors caused by environmental humidity, pressure fluctuations, or sensor aging by periodically comparing the signal to a built-in reference voltage standard.
[0025] The edge intelligent analysis unit employs a high-performance heterogeneous computing architecture, integrating a high-performance graphics processor and a neural network acceleration module. This unit is configured to perform real-time preprocessing on the received visual image stream, including image denoising, histogram equalization, and multi-scale spatial domain transformation, to eliminate interference from water mist and dust scattering in the environment. The edge intelligent analysis unit internally deploys an improved deep neural network model, preferably a visual Transformer architecture based on an attention mechanism.
[0026] The visual architecture, by introducing a global attention module, is configured to evaluate the importance of each pixel region in the image, accurately locating suspected ignition points or smoke diffusion areas in complex backgrounds. The deep neural network model is configured to recognize dynamic texture features of the flame, including but not limited to the flickering frequency of the flame edges, color gradient distribution, and irregular displacement of the flame's center.
[0027] The deep neural network model is also configured to identify the diffusion morphology of smoke, analyzing the smoke's expansion rate, drift direction, and transmittance attenuation coefficient across consecutive frames using spatiotemporal convolutional layers. During the identification process, the edge intelligent analysis unit calculates a visual confidence index for each identification result based on the probability distribution function of the model's output layer. This visual confidence index reflects the degree of certainty the visual model has regarding the presence of fire characteristics in the current area.
[0028] The edge intelligent analysis unit is also configured to perform time-series analysis on the physical parameters input from the environmental sensor array. This time-series analysis includes calculating the first and second derivatives of the environmental physical parameters to assess their growth slope and acceleration. For example, when the system detects that the slope of temperature increase exceeds a preset threshold for increase per unit time, and the carbon monoxide concentration exhibits a synchronous linear upward trend, the edge intelligent analysis unit maps this physical law into a feature vector and sends it, along with a visual confidence index, to the uncertainty modeling engine.
[0029] The uncertainty modeling engine is the core of the system's logical decision-making. It runs on a processing kernel with floating-point operation advantages and performs deep probabilistic inference on the perceived data based on a Bayesian deep learning framework. The uncertainty modeling engine is configured to perform multiple forward propagations during the inference phase by introducing a Monte Carlo Dropout mechanism.
[0030] During model inference, the system disconnects some neurons in the neural network with a preset random inactivation probability. By repeating this independent inference operation 50 to 100 times, a series of independent probability predictions are obtained. The uncertainty modeling engine is configured to calculate the mean and variance of this series of predictions. The mean of the predictions is defined as the overall confidence level, while the variance is defined as the cognitive uncertainty of the current decision. This mechanism effectively distinguishes between the agnostic nature caused by the model's lack of experience with such samples and the randomness caused by sensor noise.
[0031] The uncertainty modeling engine performs uncertainty quantification based on the following logic: when the visual module has low visual confidence due to dim lighting and water mist, if the temperature change curve and gas concentration curve provided by the environmental sensor array do not show a high correlation with fire characteristics, the posterior probability distribution output by the Bayesian deep learning framework will show extremely high variance.
[0032] The uncertainty modeling engine is configured to classify this situation as a high-uncertainty interference event and mark it as a low-confidence suspected event, suppressing the system from triggering fire alarms and filtering out false alarms caused by sudden changes in light and shadow, water vapor evaporation, or dust. Only when the multimodal data exhibits high consistency in the spatiotemporal dimensions—that is, when the visually identified flame features, drastic temperature changes, and sharp increases in smoke concentration coincide in time stamps, and the uncertainty index is below a preset stability threshold—will the uncertainty modeling engine determine that the overall confidence level exceeds the dynamic threshold and generate a fire warning signal.
[0033] The dynamic threshold is not fixed, but is adaptively adjusted by the uncertainty modeling engine based on historical environmental baseline data. The historical environmental baseline data is stored in local non-volatile memory and records the mean and variance distribution of physical parameters of the monitored area over the past 24 hours or longer.
[0034] When the system detects that the current environment is in a state of high interference, the uncertainty modeling engine will automatically increase the confidence threshold requirement for triggering fire warnings to enhance the ability to suppress false alarms; while during periods of stable environment and low background noise, the system will appropriately reduce the dynamic threshold to achieve agile response to very early faint smoke or hidden fires.
[0035] The communication network module is responsible for the data backbone transmission of the entire system, supporting industrial Ethernet, 5G ultra-reliable low-latency communication protocols, and proprietary wireless sensor network protocols. The module is configured to perform protocol encapsulation and time synchronization marking on the collected multi-source heterogeneous data, ensuring strict clock alignment between video data and environmental sensor data at different sampling frequencies within the uncertainty modeling engine. The module also includes a link quality monitoring unit, which automatically switches to redundant communication links when the main transmission link experiences increased packet loss or latency, ensuring that the transmission reliability of fire warning signals meets industrial-grade standards.
[0036] The remote monitoring center is configured as the system's centralized management platform, possessing a multi-layered data architecture. The remote monitoring center includes a real-time data receiving layer, a distributed storage layer, and a hierarchical alarm management module. The hierarchical alarm management module can classify collected early warning information into three independent levels based on the precise confidence level output by the uncertainty modeling engine: low-confidence suspected level, medium-confidence attention level, and high-confidence confirmation level.
[0037] For low-confidence suspected alarms, the remote monitoring center only logs the event in the background without triggering an alarm sound, but a silent notification appears in the taskbar of the operator's terminal. For medium-confidence attention alarms, the system automatically displays the monitoring screen of the relevant area, prompting the operator to confirm remotely. For high-confidence confirmation alarms, the remote monitoring center immediately triggers a network-wide alarm, pushes an instant message with location information to the designated fire safety officer's mobile phone, and activates the building's automatic fire alarm linkage logic. This tiered strategy effectively avoids operator fatigue caused by frequent low-level alarms, ensuring the system's credibility in critical moments.
[0038] Example 2: This example describes a variant system architecture based on edge collaboration and a distributed consensus mechanism, building upon Example 1. In this architecture, the system is configured to include multiple distributed multimodal sensing nodes, each independently integrating a miniaturized edge processing module.
[0039] The distributed sensing nodes are connected via a peer-to-peer network, forming a collaborative sensing grid system. In this architecture, the functionality of the edge intelligent analysis unit is distributed across each sensing node. Each sensing node is configured to extract visual and environmental features of its local monitoring area in real time. The uncertainty modeling engine is configured to use a distributed computing mode; that is, when a node detects a suspected fire, it sends its identification results, confidence index, and corresponding raw data summary to its neighboring nodes.
[0040] Upon receiving a suspected signal, neighboring nodes automatically adjust their visual acquisition parameters, such as increasing the magnification in a specific direction or adjusting the photosensitive gain, to attempt lateral verification of the ignition point from different spatial angles. Verification data between nodes is exchanged via a communication network module, and the uncertainty modeling engine is configured to execute a weighted fusion algorithm based on spatial correlation. If multiple sensing nodes simultaneously observe the same physical evolution trend at different physical locations, the uncertainty modeling engine exponentially increases the overall confidence level and reduces cognitive uncertainty.
[0041] At the hardware implementation level, the multimodal sensing terminal in this embodiment further integrates an infrared thermal imaging sensor array. The infrared thermal imaging sensor array is configured to acquire images of the absolute temperature field distribution of the monitored area. The edge intelligent analysis unit is configured to extract targets from the infrared thermal imaging images and identify pixel clusters with high-temperature anomaly characteristics. This fusion of infrared visual features and visible light visual features enhances the system's penetration detection capability in extremely dark or smoky environments.
[0042] The uncertainty modeling engine uses infrared temperature measurement data as an important supplement to physical parameters. When the visible light visual sensor fails due to smoke obstruction, if the infrared thermal imaging sensor detects an obvious abnormal temperature field distribution and the evolution logic of the abnormal distribution conforms to the fire physics model, the system can still maintain a high early warning accuracy.
[0043] The communication network module in this embodiment combines LoRa wireless communication technology with narrowband IoT technology, which has strong anti-interference capabilities. This configuration is particularly suitable for environments such as large petrochemical storage areas where large-scale cabling is inconvenient. Each sensing node is equipped with a high-capacity lithium battery pack and a solar power system. Through a low-power management mechanism, it ensures that the system can operate continuously for no less than 30 days without an external power source. The edge intelligent analysis unit is configured to have a sleep-wake function. It operates in an extremely low-power mode in standby mode when environmental parameters are stable. Only when the sensor triggers an over-limit interrupt or the vision module detects a pixel change will the high-computing core be activated to perform deep neural network inference.
[0044] The uncertainty modeling engine in this embodiment also introduces a variational inference method as an optimized alternative or supplement to the Monte Carlo Dropout mechanism. This variational inference method approximates a complex posterior probability distribution by optimizing the variational distribution, transforming the probability deduction problem into an optimization problem and improving computational efficiency under large-scale node collaboration. The system obtains the optimal approximation of model parameters by minimizing the Cramer-Rhodes lower bound or the variational evidence lower bound, extracting the most robust fire feature representation from complex background noise.
[0045] Example 3: This example further details the specific application configuration of the present invention in a large underground tunnel environment and the in-depth optimization details of its internal system logic. In underground tunnel application scenarios, due to the characteristics of long-distance enclosed spaces, smoke generated in the early stages of a fire often forms a long-distance wall-following flow along the tunnel ceiling, and the high-speed ventilation facilities in the tunnel can cause traditional point-type smoke sensors to fail.
[0046] For these special working conditions, the multimodal sensing terminals are configured to be arranged at preset intervals along the tunnel length. Each high-definition video acquisition device in the multimodal sensing terminal is assigned specific geometric topological coordinates. The edge intelligent analysis unit is configured to have background modeling and dynamic suppression capabilities, specifically for filtering interference from continuously flowing vehicle lights, exhaust emissions, and water mist sprayed by tunnel cleaning machines within the tunnel.
[0047] The neural network model integrated within the edge intelligent analysis unit employs a fusion structure of a long short-term memory network and a three-dimensional convolutional neural network. This fusion structure is configured to analyze not only the visual features of a single frame image but also the temporal correlation of smoke movement within a video sequence. Since the movement of smoke within a tunnel exhibits typical hydrodynamic characteristics, the edge intelligent analysis unit calculates the optical flow field of the image to assess whether the motion vector of a suspected target conforms to hydrodynamic laws, effectively distinguishing slowly drifting smoke from the background of rapidly passing vehicles.
[0048] Regarding uncertainty quantification, the uncertainty modeling engine in this embodiment is configured to use tunnel ventilation parameters as important input conditions. These ventilation parameters include the on / off status of the tunnel axial flow fan, real-time readings from the wind speed sensor, and wind direction information. The uncertainty modeling engine establishes an uncertainty correction logic based on physical feedback: when the vision module detects a suspected smoke signal, and the drift direction of this suspected smoke signal is opposite to or completely unrelated to the current tunnel wind direction, the uncertainty modeling engine increases the variance of the identification result, determining that it is highly likely to be heat generated by vehicle braking or a false positive error in the image algorithm, thus avoiding frequent false alarms caused by vehicle emissions.
[0049] In this embodiment, the remote monitoring center integrates a 3D digital twin module. This module is configured to create a virtual physical model of the tunnel and synchronously map real-time data collected by various multimodal sensing terminals into the virtual model. When a fire warning signal is generated, the remote monitoring center not only displays the confidence level but also demonstrates the possible spread trend and affected area of the smoke through 3D simulation. Simultaneously, the system, through a communication network module, links with the tunnel's broadcasting system, indicator lights, and ventilation control system to automatically issue fire emergency plans and guide the orderly evacuation of vehicles within the tunnel.
[0050] In this embodiment, the environmental sensor array of the multimodal sensing terminal also integrates an acoustic anomaly detection sensor. This sensor is configured to capture potential explosions, impacts, or sharp friction sounds within the tunnel. These acoustic features are converted into time-frequency data by the edge intelligent analysis unit and used as input for uncertainty modeling as another modality. When the acoustic sensor detects a suspected explosion, and the vision module observes a sudden change in image brightness and smoke generation almost simultaneously, the system's uncertainty modeling engine quickly outputs a fire confirmation signal with a confidence level approaching 100%, achieving earlier fire warnings than traditional physical sensors.
[0051] The uncertainty modeling engine also possesses continuous learning and self-evolution capabilities. The remote monitoring center periodically transmits manually verified fire cases back to the cloud server for incremental learning. The cloud server uses these high-value samples to fine-tune the neural network model, optimizing the weight coefficients for uncertainty judgments. Subsequently, the updated model parameters are asynchronously distributed to each edge intelligent analysis unit via the communication network module, enabling the system to continuously accumulate industry knowledge over time and achieve ultimate suppression of specific environmental interference sources.
[0052] In this embodiment, the hierarchical alarm module of the remote monitoring center is also deeply integrated with a geographic information system. When a fire warning signal is confirmed to be of high confidence level, the system automatically plans the optimal entry route for external fire rescue vehicles and pushes the locations of fire hydrants inside the tunnel, the opening status of refuge passages, and the possible concentration areas of trapped personnel to the mobile terminals of rescue personnel in real time. Through this fully closed-loop intelligent monitoring, the present invention realizes a complete automated process from environmental monitoring, risk identification, uncertainty filtering to decision execution, thereby improving the level of tunnel fire safety.
[0053] The communication network module also features protocol conversion capabilities, enabling it to connect to existing PLC control systems and industrial ring networks within the tunnel. This protocol conversion function uses a preset mapping table to convert the system's fire warning signals into standard data message formats, ensuring seamless compatibility with existing fire protection facilities. Regarding data security, all transmitted sensing data and control commands are encrypted and decrypted in real-time using a hardware encryption chip to prevent system failures or data leaks caused by unauthorized access.
[0054] Example 4: This example describes a system optimization scheme for large chemical plant production workshops and other scenarios with high dynamics, high dust levels, and strong chemical electromagnetic interference. In this scenario, the hardware shell of the multimodal sensing terminal uses a special material with national explosion-proof certification and integrates an ultrasonic self-cleaning component.
[0055] The ultrasonic self-cleaning component is configured to monitor the lens transmittance of the high-definition video acquisition device. When dust accumulation causes the transmittance to fall below a preset normal operating threshold, a high-frequency vibration mechanism is automatically activated to remove the accumulated dust from the lens surface, ensuring long-term visual clarity. The environmental sensor array employs a redundant deployment strategy. For core parameters such as temperature and carbon monoxide, each node is equipped with three independent sensor probes. A hardware-level three-to-two logical voting mechanism eliminates false alarms or missed alarms caused by a single sensor hardware failure.
[0056] In this embodiment, the edge intelligent analysis unit employs neural network compression technology based on weight pruning and quantization acceleration, aiming to efficiently deploy complex visual Transformer models on embedded computing platforms. The quantized model maintains recognition accuracy while reducing inference power consumption, thus mitigating heat buildup within the explosion-proof housing.
[0057] In this embodiment, the uncertainty modeling engine is configured to focus on handling the evolution logic of chemical fires. Fires in chemical plant areas are often accompanied by the leakage and combustion of specific chemical substances, and the resulting smoke color has material identification characteristics. The edge intelligent analysis unit is configured to extract the color moment features and spectral distribution features of the smoke, while the uncertainty modeling engine introduces expert knowledge base association logic to cross-compare the visually captured smoke color with the specific gas components detected by sensors.
[0058] If the visual characteristics show dense, dark black smoke, and the carbon monoxide sensor reading is at an extremely high level, and the joint probability distribution output by the Bayesian framework exhibits a narrow peak shape (low variance), then it is judged as a fire alarm with extremely high confidence. Conversely, if the visual system detects a large amount of white water vapor-like signal, but the humidity sensor reading in the environment momentarily saturates, and the gas concentration remains unchanged, the uncertainty modeling engine will reduce the confidence of this identification result to extremely low, marking it as normal steam emissions during the production process.
[0059] The remote monitoring center in this embodiment also has full lifecycle management capabilities, enabling real-time monitoring of the hardware health status of each sensing node, including indicators such as battery level, sensor operating current, and communication error rate. The system uses a predictive maintenance-based algorithm to assess the failure probability of sensor probes and issue maintenance recommendations before the probes reach the end of their lifespan, ensuring the long-term stable operation of the smart fire protection network.
[0060] In this embodiment, the communication network module also employs time-sensitive networking technology. By assigning the highest priority to the bandwidth reservation for fire warning messages, it ensures that the transmission latency of warning information is strictly locked within 10 milliseconds even during peak factory network bandwidth periods, thus winning extremely valuable golden time for emergency shutdown and personnel evacuation in the factory area.
[0061] In summary, the IoT-based smart fire remote monitoring system provided by this invention achieves high-fidelity acquisition of environmental information through multimodal sensing terminals, completes in-depth analysis at the feature level using edge intelligent analysis units, and realizes quantitative and accurate decision-making for fire assessment in complex environments through an uncertainty modeling engine. Furthermore, it is supplemented by a hierarchical alarm mechanism in the remote monitoring center and reliable transmission through the communication network module. The system completely solves the core pain points of traditional fire early warning systems, such as high false alarm rates, poor environmental adaptability, and limited sensing dimensions, and can be widely applied to various fire prevention and control scenarios, including industrial, transportation, and public buildings.
[0062] The scope of protection of this invention is not limited to the specific forms described in the above embodiments. For those skilled in the art, any equivalent substitutions or partial improvements made to the system's module composition, neural network architecture, sensor combination methods, and communication protocols without departing from the core technical concept of this invention should fall within the scope of protection of this invention. All parameters, thresholds, and calculation logic used in the system can be adaptively adjusted and optimized according to the needs of the actual deployment environment.
[0063] The modules in this invention can be implemented using hardware circuits, software modules running on a general-purpose processor, or a combination of hardware and software. For example, the uncertainty modeling engine can be embedded in a specific integrated circuit chip or deployed as a set of pre-compiled executable code on an edge server. The logical connections between the modules should be understood as functional associations, and their physical implementation can be through bus connections, wireless connections, or network link connections. The terms "first," "second," etc., used in this invention are only used to distinguish different functional components and do not imply a specific order or weight relationship between them. Without causing contradictions, the technical features in the various embodiments of this invention can be freely combined in any way to form customized solutions for specific application scenarios.
Claims
1. A smart fire protection remote monitoring system based on the Internet of Things, characterized in that, It includes a multimodal sensing terminal, an edge intelligent analysis unit, an uncertainty modeling engine, a remote monitoring center, and a communication network module; The multimodal sensing terminal is used to simultaneously collect visual image data and environmental physical parameters on site, and includes a high-definition video acquisition device and an environmental sensor array. The edge intelligent analysis unit is connected to the multimodal sensing terminal and is used to extract features from the collected visual image data and perform time-series analysis on environmental physical parameters. It identifies the dynamic texture features of flames and the diffusion morphology features of smoke through a built-in deep neural network model and generates a visual confidence index. The uncertainty modeling engine is connected to the edge intelligent analysis unit and is used to receive the temporal features of the visual confidence index and environmental physical parameters, quantify the uncertainty of the current environmental state based on the Bayesian deep learning framework, and generate a fire warning signal when the multimodal data are consistent and the comprehensive confidence exceeds the dynamic threshold. The remote monitoring center is connected to the uncertainty modeling engine through the communication network module, and is used to store the fire early warning signal and the corresponding raw sensing data, and push graded alarm information. The communication network module is used to establish a data transmission channel between the multimodal sensing terminal, the edge intelligent analysis unit, the uncertainty modeling engine, and the remote monitoring center.
2. The smart fire protection remote monitoring system based on the Internet of Things according to claim 1, characterized in that, The high-definition video acquisition device in the multimodal sensing terminal includes an imaging unit with wide dynamic range and low-light enhancement capability. The dynamic range parameter of the imaging unit is configured to be no less than 120 dB, which is used to capture the translucent edge features of smoke and the texture information of flames in strong backlight or low-light environments, and supports high-definition image output of no less than 30 frames per second. The environmental sensor array integrates a temperature sensor, a photoelectric smoke sensor, an electrochemical carbon monoxide sensor, and an infrared carbon dioxide sensor. The temperature sensor uses a thermistor element, and its sampling frequency is set to 10 Hz to capture temperature fluctuations on the order of 0.1 degrees Celsius per second. The photoelectric smoke sensor uses the Mie scattering principle to monitor the concentration of particulate matter in the air; Both the electrochemical carbon monoxide sensor and the infrared carbon dioxide sensor are equipped with a self-calibration circuit. The self-calibration circuit automatically compensates for zero-point drift errors caused by environmental humidity, pressure fluctuations, or sensor aging by periodically comparing with the built-in reference voltage reference.
3. The smart fire protection remote monitoring system based on the Internet of Things according to claim 1, characterized in that, The edge intelligent analysis unit is configured with a heterogeneous computing architecture, including a graphics processor and a neural network acceleration module, for preprocessing operations such as image denoising, histogram equalization and multi-scale spatial domain transformation on the visual image data. The deep neural network model is a visual Transformer architecture based on an attention mechanism. The visual Transformer architecture evaluates the importance of each pixel region in the image through a global attention module, and locates suspected ignition points or smoke diffusion areas in the background. The edge intelligent analysis unit is used to identify the dynamic texture features, which include the flashing frequency of the flame edge, the color gradient distribution, and the irregular displacement of the shape center. The edge intelligent analysis unit is also used to identify the diffusion morphology characteristics of the smoke, analyze the expansion rate, drift direction and transmittance attenuation coefficient of the smoke between consecutive frames through spatiotemporal convolutional layers, and calculate the visual confidence index based on the probability distribution function of the model output layer.
4. The smart fire protection remote monitoring system based on the Internet of Things according to claim 1, characterized in that, The edge intelligent analysis unit is also used to perform the time series analysis on the environmental physical parameters. The time series analysis includes calculating the first and second derivatives of the environmental physical parameters to evaluate the growth slope and acceleration of the parameters. When the edge intelligent analysis unit detects that the slope of the temperature increase in the environmental physical parameters exceeds the preset unit time growth threshold, and the carbon monoxide concentration shows a synchronous linear upward trend, the edge intelligent analysis unit maps the physical law of the parameters into a feature vector, and sends the feature vector and the visual confidence index to the uncertainty modeling engine.
5. The smart fire protection remote monitoring system based on the Internet of Things according to claim 1, characterized in that, The uncertainty modeling engine runs on a processing kernel with floating-point operation advantages. The uncertainty modeling engine is configured with a Monte Carlo Dropout mechanism to perform multiple forward propagations during the inference phase. When performing model inference, the uncertainty modeling engine disconnects some neurons in the neural network with a preset random inactivation probability. By repeatedly performing independent inference operations 50 to 100 times, a series of mutually independent probability prediction results are obtained. The uncertainty modeling engine is also used to calculate the mean and variance of the probability prediction results, wherein the mean of the probability prediction results is defined as the comprehensive confidence level, and the variance of the probability prediction results is defined as the cognitive uncertainty of the current decision, used to distinguish between the unknowability caused by the lack of samples in the model and the randomness caused by sensor noise.
6. The smart fire protection remote monitoring system based on the Internet of Things according to claim 1, characterized in that, The uncertainty modeling engine performs uncertainty quantification logic as follows: when the visual confidence level generated by the edge intelligent analysis unit is lower than the preset reliability threshold, and the temperature change curve and gas concentration curve provided by the environmental sensor array do not show a high correlation with fire characteristics, the posterior probability distribution output by the Bayesian deep learning framework shows a high variance state higher than the preset stable value. The uncertainty modeling engine judges this state as a high uncertainty interference event and marks the corresponding event as a low confidence suspected event to suppress fire alarm blasting. When the flame features identified by the visual image data, the temperature jumps in the environmental physical parameters, and the increase in smoke concentration coincide on the timestamp, and the cognitive uncertainty is lower than a preset stability threshold, the uncertainty modeling engine determines that the overall confidence level exceeds the dynamic threshold and generates the fire warning signal.
7. The IoT-based smart fire remote monitoring system according to claim 1, characterized in that, The dynamic threshold is adaptively adjusted by the uncertainty modeling engine based on historical environmental baseline data. The historical environmental baseline data is stored in local non-volatile memory and is used to record the mean and variance distribution of physical parameters of the monitoring area within a preset period. When the uncertainty modeling engine detects that the current environment is in a high-interference state, it automatically increases the confidence threshold requirement for triggering the fire warning signal. When the environment is detected to be stable and the background noise is below a preset noise threshold, the uncertainty modeling engine lowers the dynamic threshold to improve the response sensitivity to early faint smoke or hidden fires.
8. The smart fire protection remote monitoring system based on the Internet of Things according to claim 1, characterized in that, The remote monitoring center includes a real-time data receiving layer, a distributed storage layer, and a hierarchical alarm management module. The hierarchical alarm management module is used to classify the collected early warning information into low-confidence suspected level, medium-confidence attention level and high-confidence confirmation level according to the confidence level output by the uncertainty modeling engine. For the low confidence level of suspicion, the remote monitoring center records the event log in the background and silently prompts the operator in the taskbar of the duty officer's terminal. For the medium confidence level of concern, the remote monitoring center automatically pops up the monitoring screen of the associated area and prompts the on-duty personnel to remotely confirm. In response to the high confidence level of the confirmation, the remote monitoring center triggers a network-wide alarm, pushes an instant message with location information to the designated fire safety officer, and activates the automatic fire-fighting linkage logic within the building.
9. The smart fire protection remote monitoring system based on the Internet of Things according to claim 1, characterized in that, The communication network module supports industrial Ethernet, fifth-generation mobile communication protocol and wireless sensor network protocol, and is used to encapsulate and time-synchronize the collected multi-source heterogeneous data to ensure that the visual image data and the environmental physical parameters at different sampling frequencies are clock-aligned in the uncertainty modeling engine. The communication network module also includes a link quality monitoring unit, which is used to automatically switch to a redundant communication link when the packet loss rate of the main transmission link increases or the delay increases. The communication network module also has a protocol conversion function, which is used to connect to external control systems and industrial ring networks and convert the fire warning signal into a standard data message format.
10. The smart fire protection remote monitoring system based on the Internet of Things according to claim 1, characterized in that, The system comprises multiple distributed multimodal sensing nodes, and each of the multimodal sensing nodes is connected to form a collaborative sensing grid system through a peer-to-peer network. The multimodal sensing terminal also integrates an infrared thermal imaging sensor array for acquiring absolute temperature field distribution images of the monitored area. The edge intelligent analysis unit is used to extract targets from the absolute temperature field distribution images to identify high-temperature anomaly feature pixel clusters. The multimodal sensing terminal also integrates an acoustic anomaly detection sensor, which is used to capture explosion sounds, impact sounds, or friction sounds within the monitoring area, and convert the acoustic features into time-frequency map data and input them into the edge intelligent analysis unit. The remote monitoring center integrates a three-dimensional digital twin module to establish a virtual physical model of the monitoring area and synchronously maps the real-time data collected by various multimodal sensing terminals into the virtual physical model to display the smoke diffusion trend and the affected area.