Construction safety monitoring method and system for small-scale engineering

Abstract

This invention discloses a construction safety monitoring method and system for small-scale engineering projects, relating to the field of construction safety supervision technology. The method includes: acquiring the reporting information of small-scale engineering projects; determining the terminal type based on the reporting information; calculating the number of terminals based on the construction area, risk level, and construction stage in the reporting information; generating a visual installation point map based on the terminal type and number; binding terminal devices according to the visual installation point map; synchronously collecting terminal data through the terminal devices; inputting the terminal data into a pre-constructed multimodal deep learning model to output standardized risk event objects; and inputting the standardized risk event objects into a rule engine to match preset logical rules and generate early warning instructions. This invention achieves scene-adaptive deployment of monitoring terminals, improves the accuracy of risk identification through multimodal data fusion, and realizes an automated closed-loop early warning handling system based on a rule engine.

Description

Technical Field

[0001] This invention belongs to the field of construction risk monitoring technology, specifically relating to a construction safety monitoring method and system for small-scale projects. Background Technology

[0002] With the advancement of urbanization, the number of small-scale construction projects, such as home renovations and minor municipal renovations, is increasing, highlighting the growing need for construction safety supervision. Currently, construction safety monitoring for these small-scale projects mainly faces the following challenges: Firstly, in terms of the deployment and configuration of monitoring terminals, existing professional engineering monitoring systems are usually designed for large construction sites, based on wired networks, and are complex to deploy with high hardware costs. Their terminal configuration methods are rigid and cannot adaptively determine the type and number of terminals and generate reasonable installation point schemes according to the specific characteristics of small and scattered projects, resulting in a serious mismatch between deployment costs and actual needs. On the other hand, civilian security equipment lacks engineering configuration guidance and is difficult to meet professional supervision needs.

[0003] Secondly, in terms of risk identification, most existing civilian security general-purpose smart cameras rely on a single video modality for identification, lacking specialized analysis capabilities for construction scenarios. They also cannot perform multi-modal fusion analysis with data collected by IoT sensors such as noise, dust, and vibration, resulting in insufficient business identification accuracy, high false alarm rate, and difficulty in adapting to the complex and ever-changing construction environment of small-scale projects.

[0004] Finally, regarding the regulatory process, existing technologies generally suffer from an overemphasis on monitoring and a neglect of management. After detecting a risk event, the system often only provides front-end alerts or simple message pushes, lacking a rule-based automated handling mechanism. This results in a disconnect between early warning information and offline handling, failing to match appropriate logical rules based on the specific circumstances of the risk event and automatically generate warning instructions to drive subsequent handling processes. Consequently, the regulatory process cannot form an effective closed loop.

[0005] In summary, existing technologies in the field of safety monitoring for small-scale construction projects suffer from technical problems such as a lack of scenario adaptability in deployment and configuration, low accuracy in risk identification due to single-modal data, and a lack of a rule-based closed-loop mechanism for handling risk events. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to overcome the technical problems of existing small-scale engineering construction safety monitoring schemes, such as lack of scenario adaptability in deployment and configuration, low accuracy of risk identification due to single modal data, and lack of rule-based closed-loop mechanism for handling risk events. In order to provide a construction safety monitoring method and system for small-scale engineering projects.

[0007] A construction safety monitoring method for small-scale projects includes the following steps: Obtain the reporting information of small-scale projects, determine the terminal type based on the reporting information, calculate the number of terminals based on the construction area, risk level and construction stage in the reporting information; generate a visual installation point map based on the terminal type and the number of terminals, and bind the terminal equipment according to the visual installation point map; Terminal data is collected synchronously through the terminal device; The terminal data is input into a pre-built multimodal deep learning model, which outputs standardized risk event objects. The standardized risk event object is input into the rule engine, matched with preset logical rules, and an early warning instruction is generated.

[0008] Furthermore, the number of terminals is calculated based on the construction area, risk level, and construction stage in the reported information, including: Based on the project type preset in the reported information, the basic quantity N is... base The area coefficient α is calculated based on the construction area. area The risk coefficient β is determined based on the aforementioned risk level. risk The stage coefficient γ is determined based on the construction stage. stage ; The method for calculating the number of terminals is expressed as follows: N=N base ×α area ×β risk ×γ stage The number of terminals is N rounded up.

[0009] Furthermore, a visual installation point map is generated based on the terminal type and number, including: The construction site plan is divided into several functional areas and assigned attribute labels, and the plan is discretized into a grid. Based on the attribute labels of the functional areas and the risk factors mapped according to the terminal type, calculate the risk weight W for each grid. grid =Σ(α i ×β i ×γ i ), generate a risk heatmap, where α i For the regional basic weights, β i γ is the risk factor coefficient. i This is the distance attenuation factor; Based on the number of terminals and the risk heat map, a set coverage algorithm is used to greedily select the set of points that maximizes the coverage risk weight from the candidate points, and the set of points is modified in combination with preset installation constraint rules to output a visual installation point map.

[0010] Furthermore, the terminal data includes video streams, audio streams, sensor timing data, and engineering metadata; The terminal data is input into a pre-built multimodal deep learning model, which outputs standardized risk event objects, including: The video stream, audio stream, sensor time-series data, and engineering metadata are respectively input into the feature extraction branch corresponding to the multimodal deep learning model to obtain multimodal feature vectors; The multimodal feature vectors are concatenated and input into a fusion network to output a standardized risk event object containing risk category, risk level, confidence level, and spatiotemporal location.

[0011] Furthermore, the terminal data is input into a pre-built multimodal deep learning model, which outputs standardized risk event objects, and also includes cross-validation of the multimodal data: When features extracted from video streams identify violations of personnel safety protection, and features extracted from sensor time-series data detect abnormal physical environment parameters in the corresponding area, the risk level of the standardized risk event object is increased. When hot work features are identified based on features extracted from video streams, and abnormal temperature and humidity are detected based on features extracted from sensor time-series data but no fire-fighting equipment is identified, a standardized risk event object indicating non-compliance of hot work operations is output. When mechanical operation sounds are identified based on features extracted from the audio stream, and the current time is determined to be outside the permitted construction period based on project metadata, a standardized risk event object for construction during the violation period is output.

[0012] Furthermore, it also includes privacy-aware dynamic video processing steps: Real-time blurring of video streams that have been identified as non-public areas; When a preset type of emergency event is detected in the non-public area based on the standardized risk event object, temporary permission is requested to remove the obfuscation. Real-time obfuscation processing is restored after the event ends.

[0013] Furthermore, the standardized risk event object is input into the rule engine, matched with preset logical rules, and an early warning instruction is generated, including: The standardized risk event object, combined with the current time and the permitted work period and area permissions in the reporting information, is input into the rule engine; When a scenario that meets the preset exemption conditions is detected, the risk level of the standardized risk event object is reduced; When the risk level of a standardized risk event object exceeds a preset threshold and alarms are triggered continuously in the same area, the weight coefficient of the corresponding logical rule is dynamically increased and an early warning instruction is generated.

[0014] Furthermore, it also includes closed-loop optimization steps: Collect misjudgment cases, new risk events, and handling results arising from the flow of the aforementioned early warning instructions; The misjudged cases are periodically extracted as incremental data to fine-tune the multimodal deep learning model, and the logical rule thresholds and weights in the rule engine are optimized simultaneously.

[0015] A construction safety monitoring system for small-scale engineering projects, used to implement the aforementioned construction safety monitoring method for small-scale engineering projects, the system comprising: The terminal deployment configuration module is used to obtain the reporting information of small-scale projects, determine the terminal type based on the reporting information, calculate the number of terminals based on the construction area, risk level and construction stage in the reporting information, generate a visual installation point map based on the terminal type and the number of terminals, and bind the terminal devices according to the visual installation point map. A multimodal data acquisition module is used to synchronously acquire terminal data through the terminal device; The intelligent analysis engine module is used to input the terminal data into a pre-built multimodal deep learning model and output standardized risk event objects; The rules engine and the handling and scheduling module are used to input the standardized risk event objects into the rules engine, match them with preset logical rules, and generate early warning instructions.

[0016] Furthermore, the multimodal data acquisition module includes a smart video terminal that supports a unified interface and plug-and-play design, as well as various IoT sensors; the smart analysis engine module communicates with the rule engine and the handling and scheduling module through a standardized data format for risk event objects.

[0017] Beneficial Effects: This invention provides a construction safety monitoring method for small-scale, scattered engineering projects. By acquiring the reported information of these projects and automatically calculating the number of terminals and determining their types based on the construction area, risk level, and construction stage, a visual installation point map is generated to guide the binding and deployment of terminal equipment. This method can accurately match monitoring hardware resources according to the actual characteristics of different small-scale, scattered projects, avoiding redundant deployment or insufficient coverage of monitoring equipment, effectively reducing deployment costs, and improving the efficiency and scientific nature of equipment configuration.

[0018] This invention synchronously collects terminal data through terminal devices and inputs it into a pre-built multimodal deep learning model. It then uses multimodal data fusion for comprehensive analysis and outputs standardized risk event objects. This approach overcomes the limitations and high false alarm rates of single-modal data recognition, enabling more accurate capture of complex risks in construction scenarios and improving the accuracy and reliability of construction safety risk identification. Simultaneously, the output of standardized risk event objects provides unified and standardized data support for subsequent processes.

[0019] This invention inputs standardized risk event objects into a rule engine, matching them with preset logical rules to generate early warning instructions. This mechanism effectively connects risk identification with early warning response logic. Through the rule engine, early warning strategies become configurable and interpretable, solving the problem of disconnect between monitoring and response in existing technologies. It achieves automated closed-loop management from risk perception to early warning triggering, improving the response speed and process standardization of safety supervision for small-scale projects. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the main method steps of the present invention; Figure 2 This is a schematic block diagram of the system structure of the present invention. Detailed Implementation

[0022] To make the above-mentioned objectives, features, and advantages of this application more apparent and understandable, the specific embodiments of this application are described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0023] In the description of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0024] In this application, unless otherwise expressly specified and limited, the terms "installation," "connection," "joining," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise expressly limited. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0025] Example 1: This embodiment provides a construction safety monitoring method for small-scale, scattered engineering projects. This method achieves efficient and intelligent supervision of safety in small-scale, scattered engineering projects through adaptive deployment of lightweight terminals, fusion analysis of multimodal data, and rule-driven closed-loop processing. (Refer to...) Figure 1 As shown, the method includes the following steps: Step S1: Obtain the reporting information of small-scale projects, determine the terminal type based on the reporting information, calculate the number of terminals based on the construction area, risk level and construction stage in the reporting information; generate a visual installation point map based on the terminal type and the number of terminals, and bind the terminal equipment according to the visual installation point map.

[0026] Specifically, the platform has a built-in knowledge base and decision tree model. Users fill in project registration information on the platform's App or Web, including the following key fields: project type (interior decoration, exterior wall construction, road excavation, temporary power supply, small-scale demolition, etc.), project area (㎡), risk level prediction (low, medium, high, which can be manually selected or automatically assessed based on historical data), construction stage (civil engineering, water and electricity, masonry, painting, etc.), allowed work hours, area permissions, and optional floor plan / plan (upload CAD drawings or hand-drawn sketches).

[0027] Step S1.1: Determine the terminal type based on risk factor mapping The platform has a built-in project type-risk factor mapping table, which automatically determines the risk items that need to be monitored based on the project type, and then matches the corresponding sensing terminal type. In some implementations of this embodiment, the mapping is as follows: Interior decoration → Risk factors: dust, noise, fire, personnel violations → Terminals: dust sensors, noise sensors, smart cameras (smoke and fire detection + safety helmet detection); Exterior wall construction → Risk factors: falling objects from heights, safety belt wearing, wind speed → Terminals: vibration sensors, smart cameras (high-altitude recognition / safety rope recognition), wind speed sensors; Road excavation → Risk factors: underground pipelines, landslides, noise → Terminals: displacement sensors, noise sensors, smart cameras (area intrusion detection); Temporary power supply → Risk factors: leakage current, fire → Terminals: leakage current sensor, temperature and humidity sensor, smart camera (smoke and fire detection).

[0028] Each risk factor corresponds to one or more sensor types, ultimately generating a list of terminal types.

[0029] Step S1.2: Calculate the number of terminals Based on the project area, risk level coefficient, and construction stage, the calculation is performed using an empirical formula and rounded up: N=N base ×α area ×β risk ×γ stage ; Where, N base This indicates the basic quantity, preset according to the project type, such as 2 basic camera foundations for indoor decoration and 3 for exterior wall construction; α area The area coefficient is represented as max(1, area / 50㎡), with 1 additional coefficient for every 50㎡ of interior decoration, 1 additional coefficient for every 30㎡ of exterior wall construction, 1 additional coefficient for every 20m of road excavation, and a fixed 1 for temporary power supply; β risk This indicates a risk coefficient of 1 (low), 1.2 (medium), or 1.5 (high); γ stage This indicates a stage coefficient; during critical stages such as hot work, the number of cameras is increased.

[0030] For example: The interior decoration area is 120㎡, and the risk level is medium → the number of cameras = 2×(120 / 50≈2.4)×1.2=5.76 → rounded up to 6.

[0031] In this embodiment, an example representation of the project type-risk factor mapping table is as follows: Table 1: Project Type-Risk Factor Mapping Table

[0032] Step S1.3: Generate a visual installation point map Based on project registration information, this algorithm automatically calculates the optimal installation location using a rule engine. The algorithm integrates region segmentation, risk heatmaps, coverage optimization, and rule constraints. Step S1.3.1: Area Segmentation: Divide the construction site plan into several functional areas, assigning attribute labels to each area: Entrance / Exit (high-traffic area, cameras required), High-Risk Work Area (hot work area, high-altitude work area, dust-generating area, requiring key monitoring), Passage / Corridor (essential routes for personnel movement), Material Storage Area (potential fire or collapse area), and General Area (Rest area, completed area, lower monitoring priority). If CAD drawings are uploaded, the system automatically identifies walls, doors, windows, and functional labels; if not uploaded, a standard layout is preset according to the project type (e.g., interior decoration defaults to living room, kitchen, bedroom, and balcony).

[0033] Step S1.3.2: Risk Heatmap Generation: Discretize the planar map into a grid (e.g., 1m×1m), and calculate the risk weight W for each grid according to the regional attributes and risk factors. grid =Σ(α i ×β i ×γ i ), where α i The basic weights for each region are γ (2.0 for high-risk areas and 0.5 for ordinary areas). i For risk factor coefficients (hot work 1.5, dust 1.2, noise 0.8, etc.), γ i This is a distance attenuation factor relative to critical equipment (the closer the equipment, the higher the weight). A higher weight indicates that the location requires more monitoring coverage.

[0034] Step S1.3.3: Candidate Point Generation: Generate a set of candidate installation points based on the terminal type. Camera candidate points include wall corners, ceiling centers, above entrances / exits, and near critical equipment, ensuring each candidate point has a reasonable field of view (e.g., 2.5m above the ground, unobstructed). Sensor candidate points are concentrated in high-risk areas (noise sensors near equipment, dust sensors near the source of dust, vibration sensors fixed to walls or equipment). Supplementary candidate points are generated at the grid center and key nodes to cover blind spots. All candidate points have their coordinates, type, and coverage area (fan-shaped or rectangular field of view) recorded.

[0035] Step S1.3.4: Coverage Optimization (Greedy Selection): Using an improved set coverage algorithm, under the constraint of a given number of terminals, select the set of points that maximizes the coverage risk weight. (1) Initialize the covered mesh set to empty; (2) For each terminal type, execute the following loop: while the number of selected points < the target number: select new points from the candidate points that can cover the most uncovered risk weights and add them to the set, and update the covered grid; (3) If there are still remaining terminals, they will be added to the suboptimal area to ensure basic coverage.

[0036] Step S1.3.5: Rule Constraint Correction: The rule engine implements this in configurable IF-THEN form to adjust the initial points. Avoidance rule: IF candidate point is located in obstacle area THEN offset to the nearest valid coordinate; Redundancy rule: IF area type = "hot zone" AND number of selected cameras < 2 THEN, force the addition of one camera location; Viewpoint rule: IF If there is a pillar obstructing the camera's coverage area, THEN Adjust the candidate point position or add an auxiliary camera; Convenient rule: IF the point is more than 10m away from the existing power source AND there are other candidate points less than 5m away THEN select the closer point.

[0037] The final output includes a labeled floor plan, featuring equipment icons and types, equipment numbers, recommended installation height / angle (e.g., "CAM1 - 2.5m from the ground, facing southeast"), an overlaid risk heat map, and suggested installation sequence. The construction team installs the equipment according to the plan and scans the code to bind it.

[0038] Step S1.4: Automatic distribution of algorithm packages enables scene-adaptive configuration. After the device is bound, the platform automatically sends optimized AI recognition model packages to the corresponding terminals according to the project type and construction stage: for indoor decoration scenarios, models for safety helmet recognition, smoke and fire recognition, and personnel intrusion recognition are sent; for exterior wall construction scenarios, models for safety rope recognition and falling object warning are sent; and for road excavation scenarios, models for area intrusion recognition and machinery status monitoring are sent.

[0039] Step S2: Collect terminal data synchronously through the terminal device.

[0040] Terminal data includes video streams, audio streams, sensor timing data, and engineering metadata. The intelligent sensing terminal layer consists of flexibly combinable intelligent video terminals (with built-in basic AI computing power) and various IoT sensors (noise, particulate matter, vibration, temperature, and humidity). It adopts a unified interface and plug-and-play design, supporting solar power supply and wireless transmission. The multimodal data fusion center receives and synchronizes the aforementioned terminal data in time and space.

[0041] Step S3: Input the terminal data into the pre-built multimodal deep learning model and output standardized risk event objects.

[0042] The core of the business intelligence analysis engine is a multimodal deep learning model, and the specific processing procedure is as follows: The structure of a multimodal deep learning model includes: Video stream processing branch: Lightweight convolutional neural networks (such as MobileNetV3 or EfficientNet-Lite) are used to extract spatial features of video frames, and temporal convolutional networks (TCN) or LSTM are used to capture action timing and output action category and confidence. Audio stream processing branch: Converts audio signals into Mel spectrograms and uses a CNN classifier to identify specific mechanical sounds and abnormal sounds; Sensor data branch: Employs a fully connected network (MLP) to process time-series characteristics such as noise, dust, vibration, temperature, and humidity, and outputs environmental status and anomaly level; Project metadata branch: Encodes structured data such as project type, construction stage, permitted working time, and area permissions into feature vectors, and maps them to a unified feature space through a fully connected layer; Fusion layer: The feature vectors output from each branch are concatenated and input into a multilayer perceptron (MLP) for multimodal fusion.

[0043] The model input and output are defined as follows: The input dimensions are defined as follows: video stream (16×224×224×3 RGB frame sequence), audio stream (10-second window Mel spectrogram 128×128), sensor timing (60×4 matrix of readings in the last 60 seconds), and engineering metadata (one-hot encoded + scalar value).

[0044] The output is a standardized risk event object, including: risk category (12 preset categories such as not wearing a safety helmet, unauthorized hot work, etc.), risk level (low 0-0.3, medium 0.3-0.7, high 0.7-1.0), confidence level (0~1), spatiotemporal location (timestamp, camera / sensor ID, coordinates), and evidence fragments (alarm screenshot URL, waveform fragment URL).

[0045] In some implementations of this embodiment, the output data structure is exemplified as follows: { "risk_type":"Not wearing a helmet", "risk_level":0.86, "confidence":0.94, "timestamp":"2025-03-26T10:23:45Z", "location":{"camera_id":"CAM_01","bbox":[120,80,200,160]}, "evidence_url":"https: / / xxx / screenshot.jpg" } Multimodal fusion enhancement is achieved through a cross-validation mechanism. In some specific implementations of this embodiment, the cross-validation mechanism includes the following rules: When a person is identified as not wearing a safety helmet based on the video stream, and the vibration sensor in the area detects high-frequency micro-vibrations (indicating falling objects from a height), the alarm level will be raised from "normal" to "severe". When an open flame or smoke is detected based on the video stream, but no fire extinguisher is detected, and the temperature and humidity sensor data spikes abnormally, a high-risk alarm for "non-compliant hot work" is generated. When a specific mechanical operation sound is detected based on the audio stream, and the current time is determined to be outside the permitted construction time (such as at night) based on the project metadata, a "construction during illegal hours" alarm is triggered.

[0046] In this embodiment, the model training method includes: Dataset construction: More than 5,000 hours of real data from over 100 small-scale engineering sites were collected and labeled by safety engineers; data augmentation methods such as random brightness adjustment, horizontal flipping, adding background noise, and adding Gaussian noise were used; the dataset was divided into a 70% training set, a 15% validation set, and a 15% test set.

[0047] Pre-training and fine-tuning: The video branch was fine-tuned after pre-training on ImageNet and Kinetics-400; the audio branch was fine-tuned after pre-training on AudioSet; the sensor branch was fine-tuned after reconstructing normal working condition features using an autoencoder in unsupervised pre-training; and the fusion layer was jointly optimized end-to-end.

[0048] Loss function: The multi-task learning loss is L=Lcls+λ1·Lrank+λ2·Lcontrast, where Lcls is the cross-entropy loss (risk category), Lrank is the ranking loss (risk level ranking), and Lcontrast is the contrast loss (to maintain semantic consistency of different modal features before fusion).

[0049] Optimizer and Strategy: The AdamW optimizer is used with an initial learning rate of 1e-3, cosine annealing scheduling, and early stopping (the validation set loss stops if it does not decrease for 5 consecutive epochs). Convergence occurs in approximately 50-80 epochs.

[0050] Privacy-aware dynamic video processing steps: In non-public areas (such as marked bedrooms and bathrooms), the video stream is blurred in real time by default. When the business intelligence analysis engine detects an emergency event in the area (such as fire smoke, a person falling, or an unauthorized intrusion), it automatically requests temporary permission from the administrator to remove the blurring for emergency confirmation. The blurring is automatically restored after the event ends.

[0051] Step S4: Input the standardized risk event object into the rule engine, match it with the preset logical rules, and generate an early warning instruction.

[0052] The business intelligence analysis engine and the rules engine are the two core decision-making modules, which work closely together through standardized interfaces (fixed JSON format for risk event objects) and event-driven mechanisms (such as Kafka message middleware).

[0053] Step S4.1: Event-driven and rule-matching In some implementations of this embodiment, the rules built into the rule engine include: The rules engine has built-in configurable IF-THEN rules that match in parallel when a risk event is received: Rule 1: IF(risk_type="not wearing a safety helmet" AND confidence>0.8)THEN generate a "general warning" → push to the on-site supervisor's App; Rule 2: IF(risk_type="hot work" AND confidence>0.9 AND no fire extinguisher in the video)THEN generate an "emergency warning" → simultaneously push to the construction party and the regulatory party; Rule 3: IF(risk_type="Noise Exceeding Standard" AND Current Time NOTIN Permitted Construction Period) THEN Generate "Illegal Construction Warning" → Activate On-Site Audible and Visual Removal; Rule 4: IF (risk_level>0.7 AND consecutive alarms in the same area within 30 seconds) THEN automatically escalate the notification level.

[0054] Step S4.2: Dynamic Interaction and Mutual Correction AI output for rule correction: When the rule engine detects an exempt scenario (such as "managers temporarily entering without wearing safety helmets"), it can reduce the risk level of the AI ​​output through an indicator function to avoid false alarms; AI enhances rule weights: When AI identifies high-risk behaviors (such as "working at height without a safety rope"), it dynamically increases the weight coefficient of the corresponding rule, so that similar events can trigger high-level warnings more quickly. Dynamic threshold adjustment: The rule engine automatically lowers the sensor anomaly threshold based on environmental changes (such as heavy rain), and the AI ​​model adjusts the anomaly detection sensitivity accordingly.

[0055] Step S4.3: Closed-loop optimization mechanism: The handling and scheduling module stores user-reviewed misjudged cases, new risk events, and handling results into the data storage module. Incremental data is extracted periodically (e.g., weekly) to fine-tune the AI ​​model, while simultaneously optimizing thresholds, weights, or adding new rules to the rule base. New models and rule bases seamlessly replace older versions through hot deployment. When performance metrics (accuracy ≥ 95%, false negative rate ≤ 1%, false positive rate ≤ 3%) meet the standards after multiple iterations without significant improvement, iteration is paused, and the system enters a stable operation phase.

[0056] Example 2: Reference Figure 2 As shown, this embodiment provides a construction safety monitoring system for small-scale projects, which implements the method of Embodiment 1. The system adopts a lightweight "cloud-edge-device" collaborative architecture, including: an intelligent sensing terminal layer, an edge computing layer, an IoT cloud platform layer, and a user interaction layer.

[0057] The specific modules include: Terminal deployment configuration module: used to obtain reporting information, map risk factors, calculate the number of terminals, generate a visual installation point map and bind terminal devices, and automatically distribute scene recognition algorithm packages; Multimodal data acquisition module: namely the intelligent sensing terminal layer, including intelligent video terminals that support unified interfaces and plug-and-play design and various IoT sensors, used to synchronously acquire terminal data; Intelligent Analysis Engine Module: This is the business intelligent analysis engine, used to extract features from collected multimodal data, perform fusion and cross-validation, and output standardized risk event objects. Privacy-aware processing module: used to blur video streams that have been identified as non-public areas, and to request deblurring when an emergency event is detected; The rules engine and handling scheduling module are used to generate early warning instructions and transfer handling tasks based on standardized risk event objects and matching logic rules such as current time, allowed operation period, and regional permissions. Lightweight digital twin module: Based on time-series on-site images, it generates a visual timeline of project progress and associates it with evidence of key milestones.

[0058] The intelligent analysis engine module communicates with the rule engine and the handling and scheduling module through a standardized data format for risk event objects.

[0059] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0060] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A construction safety monitoring method for small-scale, scattered projects, characterized in that, Includes the following steps: Obtain the reporting information of small-scale projects, determine the terminal type based on the reporting information, and calculate the number of terminals based on the construction area, risk level, and construction stage in the reporting information; generate a visual installation point map based on the terminal type and number of terminals, and bind terminal devices according to the visual installation point map; synchronously collect terminal data through the terminal devices; input the terminal data into a pre-built multimodal deep learning model, and output standardized risk event objects; input the standardized risk event objects into a rule engine, match them with preset logical rules, and generate early warning instructions.

2. The construction safety monitoring method for small-scale projects according to claim 1, characterized in that, The number of terminals is calculated based on the construction area, risk level, and construction stage in the reported information, including: Determine a basic number N based on the engineering type in the report information base Calculate an area coefficient a according to the construction area area Determine a risk coefficient b according to the risk level risk Determine a stage coefficient g according to the construction stage stage ; The method for calculating the number of terminals is expressed as follows: N=N base ×α area ×β risk ×γ stage ; The number of terminals is N rounded up.

3. The construction safety monitoring method for small-scale projects according to claim 2, characterized in that, Generate a visual installation point map based on the terminal type and number of terminals, including: The construction site plan is divided into several functional areas and assigned attribute labels, and the plan is discretized into a grid. Based on the attribute labels of the functional areas and the risk factors mapped according to the terminal type, calculate the risk weight W for each grid. grid =Σ(α i ×β i ×γ i ), generate a risk heatmap, where α i For the regional basic weights, β i γ is the risk factor coefficient. i This is the distance attenuation factor; Based on the number of terminals and the risk heat map, a set coverage algorithm is used to greedily select the set of points that maximizes the coverage risk weight from the candidate points, and the set of points is modified in combination with preset installation constraint rules to output a visual installation point map.

4. The construction safety monitoring method for small-scale projects according to claim 1, characterized in that, The terminal data includes video streams, audio streams, sensor timing data, and engineering metadata; The terminal data is input into a pre-built multimodal deep learning model, which outputs standardized risk event objects, including: The video stream, audio stream, sensor time-series data, and engineering metadata are respectively input into the feature extraction branch corresponding to the multimodal deep learning model to obtain multimodal feature vectors; The multimodal feature vectors are concatenated and input into a fusion network to output a standardized risk event object containing risk category, risk level, confidence level, and spatiotemporal location.

5. The construction safety monitoring method for small-scale projects according to claim 4, characterized in that, The terminal data is input into a pre-built multimodal deep learning model, which outputs standardized risk event objects and also includes cross-validation of the multimodal data. When features extracted from video streams identify violations of personnel safety protection, and features extracted from sensor time-series data detect abnormal physical environment parameters in the corresponding area, the risk level of the standardized risk event object is increased. When hot work features are identified based on features extracted from video streams, and abnormal temperature and humidity are detected based on features extracted from sensor time-series data but no fire-fighting equipment is identified, a standardized risk event object indicating non-compliance of hot work operations is output. When mechanical operation sounds are identified based on features extracted from the audio stream, and the current time is determined to be outside the permitted construction period based on project metadata, a standardized risk event object for construction during the violation period is output.

6. The construction safety monitoring method for small-scale projects according to claim 4, characterized in that, It also includes privacy-conscious dynamic video processing steps: Real-time blurring of video streams that have been identified as non-public areas; When a preset type of emergency event is detected in the non-public area based on the standardized risk event object, temporary permission is requested to remove the obfuscation. Real-time obfuscation processing is restored after the event ends.

7. The construction safety monitoring method for small-scale projects according to claim 4, characterized in that, The standardized risk event object is input into the rule engine, matched with preset logical rules, and an early warning instruction is generated, including: The standardized risk event object, combined with the current time and the permitted work period and area permissions in the reporting information, is input into the rule engine; When a scenario that meets the preset exemption conditions is detected, the risk level of the standardized risk event object is reduced; When the risk level of a standardized risk event object exceeds a preset threshold and alarms are triggered continuously in the same area, the weight coefficient of the corresponding logical rule is dynamically increased and an early warning instruction is generated.

8. The construction safety monitoring method for small-scale projects according to claim 7, characterized in that, It also includes closed-loop optimization steps: Collect misjudgment cases, new risk events, and handling results arising from the flow of the aforementioned early warning instructions; The misjudged cases are periodically extracted as incremental data to fine-tune the multimodal deep learning model, and the logical rule thresholds and weights in the rule engine are optimized simultaneously.

9. A construction safety monitoring system for small-scale, scattered projects, characterized in that, The system is used to implement the construction safety monitoring method for small-scale engineering projects as described in any one of claims 1 to 8, the system comprising: The terminal deployment configuration module is used to obtain the reporting information of small-scale projects, determine the terminal type based on the reporting information, calculate the number of terminals based on the construction area, risk level and construction stage in the reporting information, generate a visual installation point map based on the terminal type and the number of terminals, and bind the terminal devices according to the visual installation point map. A multimodal data acquisition module is used to synchronously acquire terminal data through the terminal device; The intelligent analysis engine module is used to input the terminal data into a pre-built multimodal deep learning model and output standardized risk event objects; The rules engine and the handling and scheduling module are used to input the standardized risk event objects into the rules engine, match them with preset logical rules, and generate early warning instructions.

10. The construction safety monitoring system for small-scale engineering projects according to claim 9, characterized in that, The multimodal data acquisition module includes a smart video terminal that supports a unified interface and plug-and-play design, as well as various IoT sensors; the smart analysis engine module communicates with the rule engine and the handling and scheduling module through a standardized data format for risk event objects.

Images