Anomaly behavior detection method and system

Abstract

The application relates to the technical field of construction safety detection, in particular to an abnormal behavior detection system and method. The system comprises a multi-modal perception layer, an edge processing layer, a behavior cognition reasoning engine, a dynamic risk assessment and early warning module and a visual management platform connected in sequence. The multi-modal perception layer collects visual, Internet of Things and environmental data; the edge processing layer performs real-time target detection, behavior preliminary classification and space-time registration; the behavior cognition reasoning engine realizes behavior understanding and intention prediction by constructing a space-time graph and fusing rule reasoning and a deep learning model, and integrates a cause-and-effect analysis submodule to perform root cause tracing; the risk assessment module dynamically calculates a risk level according to a reasoning result and triggers early warning, and deduces risk evolution by using a risk transmission simulator; and the visual management platform realizes three-dimensional display and emergency linkage, and continuously optimizes a model by feedback data driving. The application realizes real-time and accurate identification and closed-loop risk control of abnormal behaviors on a construction site.

Description

Technical Field

[0001] This invention relates to the field of building construction safety detection technology, and more specifically, to a method and system for detecting abnormal behavior. Background Technology

[0002] Construction sites, as typical high-risk work environments, present significant challenges to safety management due to the complexity of their personnel, equipment, and environment. At the personnel level, there is a cross-flow of various roles, including construction workers, managers, equipment maintenance personnel, and temporary workers, with significant differences in safety awareness. This can lead to dangerous behaviors such as not wearing safety helmets / belts or illegally climbing scaffolding. At the equipment level, the site includes large machinery such as tower cranes, construction elevators, and concrete pump trucks, as well as smaller tools like electric drills and cutting machines. Inconsistent adherence to equipment operating procedures can easily result in risks such as machinery overloading, improper start-up and shutdown, and abnormal equipment displacement. At the environmental level, temporary structures (such as guardrails and temporary facilities) are frequently changed, and the site is further affected by factors such as changes in light intensity (e.g., nighttime construction, rainy weather), dust obstruction, and noise interference, exacerbating the uncertainty of the environment.

[0003] Currently, safety monitoring methods at construction sites remain relatively traditional, typically employing periodic manual inspections combined with a small number of fixed sensors (such as infrared sensors and smoke alarms) for localized monitoring. While some sites have introduced video surveillance systems, they primarily rely on manual real-time viewing of video footage or triggering simple alarms through preset rules (such as boundary crossing detection). Under this model, due to the significant dynamic changes in the construction site environment, such as equipment movement, material stacking, and weather conditions; and the diverse behaviors of personnel, such as operating equipment, handling materials, and engaging in overlapping operations, coupled with the sudden and covert nature of abnormal behaviors, traditional methods struggle to effectively predict potential hazards. This results in delayed accident warnings or a high false alarm rate, failing to establish a comprehensive and intelligent safety protection system. Summary of the Invention

[0004] The technical problem to be solved by this invention is that the accident early warning is delayed or the false alarm rate is high, making it impossible to form a comprehensive and intelligent safety protection system. In view of the above-mentioned defects of the existing technology, this invention provides an abnormal behavior detection method and system.

[0005] The technical solution adopted by this invention to solve its technical problem is: on the one hand An abnormal behavior detection system includes a multimodal perception layer, an edge processing layer, a behavior cognition reasoning engine, a dynamic risk assessment and early warning module, and a visualization management platform connected in sequence. The multimodal perception layer is deployed on the construction site to collect multi-source heterogeneous data, including visual data, IoT sensor data, and environmental data. The edge processing layer is communicatively connected to the multimodal perception layer and is used to perform real-time preprocessing on the multi-source heterogeneous data. The preprocessing includes target detection, initial behavior classification, and spatiotemporal registration of data. The behavioral cognitive reasoning engine is communicatively connected to the edge processing layer and is used to receive preprocessed data. By fusing rule-based reasoning and deep learning prediction, it understands the behavior and predicts the intentions of construction site personnel. The dynamic risk assessment and early warning module is connected to the behavioral cognition reasoning engine and is used to perform dynamic risk assessment and generate graded early warning signals based on the behavioral understanding and intention prediction results. The visualization management platform is connected to the dynamic risk assessment and early warning module to display risk information and execute early warning responses. It also records data throughout the process and feeds it back to the behavioral cognitive reasoning engine to drive the self-evolution of the system model.

[0006] Preferably, the multimodal perception layer includes a visual perception unit, an Internet of Things sensor network, and a communication network; The visual perception unit includes high-definition cameras deployed in key areas of the construction site. The high-definition cameras integrate infrared and / or thermal imaging modules and support low-light enhancement and image denoising algorithms. The Internet of Things (IoT) sensor network includes equipment status sensors installed on large construction equipment, structural monitoring sensors deployed on temporary facilities and slopes, wearable devices worn by personnel, and environmental monitoring sensors. The communication network adopts a hybrid 5G and LoRa network to transmit data collected by the visual sensing unit and the Internet of Things sensor network.

[0007] Preferably, the edge processing layer includes multiple distributed edge computing nodes, which are used to perform at least one of the following processes: Based on a lightweight neural network model, real-time target detection and re-identification are performed on video streams. The identified targets include human bodies, safety equipment, and construction equipment. Based on the pose estimation algorithm, preliminary classification of human movements is performed; Filter and perform threshold screening on IoT sensor data to identify abnormal state signals; Based on positioning data and visual SLAM technology, time synchronization and spatial coordinate registration are performed on multi-source heterogeneous data.

[0008] Preferably, the behavioral cognitive reasoning engine includes: The spatiotemporal graph modeling unit is used to construct spatiotemporal graphs from personnel, equipment and environmental elements, where nodes represent entities and edges represent spatiotemporal or interactive relationships between entities. The hybrid reasoning unit includes a rule knowledge base based on security specification components and a behavior prediction model based on temporal deep learning model components, which are used for parallel reasoning on input data. The fusion decision unit is used to fuse the reasoning results of the rule knowledge base with the prediction results of the behavior prediction model to output the behavior understanding and risk assessment results.

[0009] Preferably, the behavioral cognitive reasoning engine further includes a causal analysis submodule, which is used to perform root cause analysis on abnormal behavioral events based on causal discovery algorithms and to quantitatively evaluate the effectiveness of different preventive measures based on counterfactual reasoning.

[0010] Preferably, the dynamic risk assessment and early warning module performs the following operations: A four-level risk assessment model is constructed, which divides risks into four levels: low, medium, high, and urgent, and presets corresponding response mechanisms for each level. Based on the output of the behavioral cognitive reasoning engine, the risk level of the current scene is calculated in real time; When the risk level exceeds a preset threshold, a corresponding early warning signal and mechanism are triggered. The dynamic risk assessment and early warning module also integrates a risk transmission simulator based on a dynamic Bayesian network, which is used to deduce the propagation path and probability of risk in the future period and generate optimal risk blocking strategy suggestions.

[0011] Preferably, the visualization management platform includes: The 3D digital twin module, based on BIM and GIS data, constructs a virtual simulation of the construction site, which is used to map the location and status of personnel, equipment, and risks in real time, and supports panoramic spatiotemporal retrospective analysis of historical early warning events using multimodal data. The emergency response interface is used to receive feedback annotations from management personnel on the early warning results and push the annotated data to the simulation training process to drive the model in the behavioral cognitive reasoning engine to perform incremental learning and optimization. The model evolution interface is used to receive feedback annotations from managers on the early warning results and push the annotated data to the model training process, driving the model in the behavioral cognitive reasoning engine to perform incremental learning and optimization.

[0012] Preferably, the model evolution interface supports an active learning mechanism. When the confidence level of the behavioral cognitive reasoning engine in predicting a specific scenario is lower than a set threshold, the scenario data is automatically marked as a sample to be labeled and pushed to the management platform.

[0013] on the other hand An abnormal behavior detection method, applied to any of the systems described above, includes the following steps: S1. Collect multi-source heterogeneous data from construction sites through a multimodal sensing layer; S2. The multi-source heterogeneous data is preprocessed in real time through the edge processing layer to obtain a structured event and state data stream; S3. The structured data stream is fused and analyzed through a behavioral cognitive reasoning engine to achieve behavioral understanding and intention prediction; S4. Through the dynamic risk assessment and early warning module, assess the real-time risk level based on the results of step S3, and trigger a graded early warning. S5. Display risks and coordinate responses through a visual management platform, while storing the entire process of early warning and handling data in the database; S6. Based on the feedback data and new case data accumulated in step S5, the model and knowledge base in the behavioral cognitive reasoning engine are periodically optimized and updated to complete the system closed loop.

[0014] Preferably, the behavior understanding and prediction in step S3 specifically includes: Construct a spatiotemporal map of personnel, equipment, and environment; The current state is logically complied with using a rule knowledge base; Probabilistic predictions are made using behavioral prediction models and short-term behavioral sequences. The results of rule checks and behavior predictions are integrated, and environmental context factors are introduced for weighted correction. Output behavioral risk labels with confidence levels.

[0015] The beneficial effects of this invention are as follows: 1. By collaborating with visual perception, IoT sensors (wearable devices, status sensors), and environmental data, the system overcomes the limitations of single visual perception in scenarios such as occlusion, inclement weather, and nighttime. For example, thermal imaging compensates for low light conditions, UWB positioning compensates for lost visual tracking, and environmental data (such as wind speed) provides crucial context for risk assessment, significantly reducing false alarms and false negatives. Unlike simply recognizing superficial behaviors such as "falling" or "not wearing a helmet," the system, through spatiotemporal graph modeling, hybrid reasoning (rules + deep learning), and causal analysis, can understand the intent of behavior (such as the purpose of approaching a dangerous area) and its correlations (such as the causal relationship between equipment malfunction and personnel behavior), enabling the discovery of more complex and potential risk chains.

[0016] 2. Data preprocessing and initial screening at the edge significantly reduce transmission bandwidth pressure and enable near real-time completion of tasks such as target detection and initial classification, saving crucial time for subsequent deep reasoning and early warning. Behavioral prediction models (such as ST-GCN) can probabilistically predict behavioral sequences in the short term, achieving "early warning." Dynamic Bayesian network risk transmission simulators can further deduce the evolution path of risk events, upgrading management from responding to past events to predicting and intervening in potential events. Counterfactual reasoning can quantitatively evaluate the effectiveness of different preventive measures (e.g., "If evacuation were carried out in advance, the probability of casualties would be reduced by 70%), elevating early warning from "what happened" to "what should be done most effectively," providing managers with scientific decision support.

[0017] 3. The system collects manual feedback and annotations on early warning results (correcting false alarms / missed alarms) through a visual management platform and automatically pushes this feedback to the training process, driving incremental learning of the deep learning model and updating the rule base. This enables the system to continuously learn from real-world case studies, adapting to unknown risks brought by new construction sites, new processes, and new equipment. Performance improves over time, reducing long-term operation and maintenance optimization costs. When the system has low confidence in a specific new scenario, it can automatically identify and request expert annotations to specifically strengthen weak links and accelerate the system's maturation process in complex and ever-changing environments.

[0018] 4. The system seamlessly integrates the discrete stages of perception, analysis, early warning, response, and optimization, forming a complete management closed loop and solving the problems of information silos and disconnected handling processes in traditional solutions. Based on a BIM / GIS-based digital twin 3D platform, abstract data and alarms are transformed into intuitive, spatialized scene mappings, supporting panoramic spatiotemporal retrospective analysis. Users not only know "where the alarm occurred," but also clearly understand "why the alarm occurred, the surrounding situation, and whether it has occurred before," greatly improving the efficiency and scientific nature of command and decision-making. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the present invention will be further described below in conjunction with the accompanying drawings and embodiments. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Figure 1 This is a schematic diagram of the abnormal behavior detection system in Embodiment 1 of this application.

[0020] Figure 2 This is a schematic diagram of the steps of the abnormal behavior detection method in Embodiment 2 of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, a clear and complete description will be provided below in conjunction with the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the protection scope of the present invention.

[0022] Example 1 This application implements, for example Figure 1 As shown, an abnormal behavior detection system includes a multimodal perception layer, an edge processing layer, a behavior cognition reasoning engine, a dynamic risk assessment and early warning module, and a visualization management platform that are connected in sequence.

[0023] A multimodal perception layer is deployed on the construction site to collect multi-source heterogeneous data, including visual data, IoT sensor data, and environmental data. The multimodal perception layer includes a visual perception unit, an IoT sensor network, and a communication network. The visual perception unit includes high-definition cameras deployed in key areas of the construction site. As an optional embodiment, the high-definition cameras are optimized based on the principles of field of view coverage and key area overlap analysis to ensure blind-spot-free monitoring of high-risk areas such as high-altitude work areas and deep foundation pit edges. The high-definition cameras integrate infrared and / or thermal imaging modules to detect personnel positions at night or in smoky environments and to identify equipment overheating faults by recognizing abnormal thermal radiation. It also supports image preprocessing based on Retinex theory's low-light-intensity algorithm and a non-local means denoising algorithm.

[0024] The IoT sensor network includes equipment status sensors installed on large construction equipment, structural monitoring sensors deployed in temporary facilities and slopes, smart wearable devices worn by personnel, and environmental monitoring sensors. The communication network employs a hybrid 5G and LoRa network for transmitting data collected by the visual sensing units and the IoT sensor network. Specifically, the 5G network transmits high-bandwidth real-time video streams, while the LoRa network transmits low-power, wide-coverage IoT sensor data. A data optimization priority scheduling protocol is deployed at the network layer to ensure low-latency transmission of early warning data. As an optional embodiment, equipment status sensors are used to detect vibration, tilt angle, pressure, hydraulic oil temperature, and engine speed; structural monitoring sensors can be configured as inclinometers, crack gauges, stress-strain sensors, etc.; smart wearable devices may include integrated UWB / Bluetooth beacons for precise outdoor positioning and monitoring the wearer's heart rate, body posture, and electronic fence status; environmental monitoring sensors can be used to monitor PM2.5, wind speed, noise, temperature, and humidity.

[0025] The edge processing layer is used for real-time preprocessing of multi-source heterogeneous data. Preprocessing includes object detection, initial behavior classification, and spatiotemporal registration of data. The edge processing layer comprises multiple distributed edge computing nodes, which are industrial-grade AI computing boxes deployed on the construction site, with built-in GPU or NPU acceleration units. These edge computing nodes perform at least one of the following processes: real-time target detection and re-identification of video streams based on lightweight YOLOv5s or MobileNet-SSD neural network models, targeting objects including humans, safety equipment, and construction equipment, and using the DeepSORT algorithm for continuous target tracking across cameras; extraction of human skeletal key points based on OpenPose or AlphaPose pose estimation algorithms, and preliminary classification of basic actions such as "running," "climbing," and "falling" based on rule-based thresholds or simple temporal convolutional networks; Kalman filtering and threshold screening of IoT sensor data to identify abnormal state signals; and millisecond-level timestamping of all data using a BeiDou / GPS time server as a unified time source, constructing a sparse point cloud map of the construction site based on visual SLAM technology, and registering UWB and other positioning coordinate systems to this map coordinate system to achieve spatiotemporal synchronization and spatial coordinate registration of multi-source data, generating structured event and state data streams.

[0026] The behavioral cognitive reasoning engine receives preprocessed data and, through the fusion of rule-based reasoning and deep learning prediction, understands and predicts the intentions of construction site personnel. As an optional embodiment, the behavioral cognitive reasoning engine includes a spatiotemporal graph modeling unit, a hybrid reasoning unit, a fusion decision-making unit, and a causal analysis submodule. The spatiotemporal graph modeling unit constructs a spatiotemporal graph of personnel, equipment, and environmental elements, where nodes represent entities (personnel node attributes include ID, location, speed, and posture; equipment node attributes include equipment type, working status, and location; environmental node attributes include area type and risk level), and edges represent the spatiotemporal proximity or specific interaction relationships between entities (such as "operation" or "entry"). The hybrid reasoning unit includes a rule knowledge base built on safety regulations and a behavior prediction model built on a temporal deep learning model, used for parallel reasoning on input data. The rule knowledge base contains explicit logical rules such as "IF (personnel activity area = 'aerial work platform') AND (detection status = 'not wearing a seat belt') THEN (behavior = 'serious violation')". The behavior prediction model is a spatiotemporal graph convolutional network (ST-GCN) or a Transformer encoder, used to take historical spatiotemporal graph sequences as input and output the probability of abnormal behavior occurring in the near future.

[0027] The fusion decision unit is used to fuse the reasoning results from the rule knowledge base with the prediction results from the behavior prediction model using DS evidence theory or a weighted voting fusion algorithm, and introduces environmental context factors (such as rainy days and nighttime) for weighted correction, ultimately outputting the behavior understanding and risk assessment results. The causal analysis submodule is used to perform root cause analysis between abnormal behaviors based on the PC algorithm or causal discovery algorithms such as NOTEARS, and to quantitatively evaluate different preventive measures based on structural causal models (SCM) and counterfactual reasoning (e.g., simulating "if the instruction is broadcast in advance, the probability of injury or death will be reduced by X%").

[0028] The dynamic risk assessment and early warning module is used to dynamically assess risks and generate tiered early warning signals based on behavioral understanding and intent prediction results. The module performs the following operations: constructs a four-level risk assessment model, classifying risks into four levels: low (e.g., a single person not wearing a safety helmet), medium (e.g., personnel entering a restricted area), high (e.g., people below a hoisting area), and emergency (e.g., structural displacement exceeding limits or open flame), and presets corresponding response mechanisms for each level; calculates the risk level of the current scenario in real time based on the output of the behavioral cognitive reasoning engine; triggers corresponding early warning signals and mechanisms when the risk level exceeds a preset threshold; the module also integrates a risk transmission simulator based on a dynamic Bayesian network (DBN), used to predict the propagation path and probability of secondary risk events in the future period through forward reasoning, using current risk events as evidence, and generating optimal risk blocking strategy suggestions through backward reasoning.

[0029] The visualization management platform displays risk information and executes early warning responses. It also records data throughout the entire process and feeds it back to the behavioral cognitive inference engine, driving the self-evolution of the system model. The platform includes a 3D digital twin module, an emergency response interface, and a model evolution interface. The 3D digital twin module constructs a virtual simulation of the construction site based on BIM and GIS data. It uses WebGL or a game engine to map the location and status of personnel, equipment, and risks in real time, and supports panoramic spatiotemporal retrospective analysis of historical early warning events using multimodal data. The emergency response interface automatically triggers on-site audible and visual alarms based on the warning level, pushes information to the responsible person's mobile terminal, or links with the access control system. It also receives feedback annotations from management personnel on the warning results and pushes the annotated data to the simulation training process, driving incremental learning and optimization of the model in the behavioral cognitive inference engine. The model evolution interface is used to store the entire process data of early warning processing (including raw data, inference results, handling actions and final results) into the spatiotemporal event database, and periodically push high-quality data packages labeled by management personnel to the model training process, driving the prediction model in the behavioral cognition inference engine to perform incremental learning and optimize and update the rule knowledge base; the model evolution interface supports an active learning mechanism. When the prediction confidence of the behavioral cognition inference engine for a specific scenario is lower than a set threshold, it automatically marks the data of that scenario as a sample to be labeled and pushes it to the management platform.

[0030] Example 2 An abnormal behavior detection method, applied to any of the systems in Example 1, includes the following steps: S1. Collect multi-source heterogeneous data from the construction site through a multimodal perception layer; specifically, this includes: collecting visual data through optimized deployment of high-definition cameras, collecting IoT data through equipment status, structural monitoring, smart wearables and environmental sensors, and using a 5G and LoRa hybrid network for hierarchical transmission.

[0031] S2. Real-time predictive processing of multi-source heterogeneous data is performed through the edge processing layer to obtain structured event and state data streams. Specifically, this includes: using lightweight AI models for video target monitoring and tracking on edge computing nodes, using pose estimation algorithms for initial classification of personnel movements, filtering and thresholding sensor data, and registering and encapsulating all data based on a unified spatiotemporal reference.

[0032] S3. The structured data stream is fused and analyzed using a behavioral cognitive reasoning engine to achieve behavioral understanding and intent prediction. As an optional embodiment, behavioral understanding and prediction specifically include: S31. Dynamically construct a spatiotemporal diagram representing the relationship between personnel, equipment, and environment using event windows; S32. Perform logical compliance checks on the current state using a rule knowledge base; simultaneously, use a behavior prediction model to make probabilistic predictions based on the behavior sequence in the near future. S33. A confidence fusion algorithm is used to fuse the rule check results and behavior prediction results, and environmental context factors are introduced for weighted correction to output the final behavior understanding and risk assessment results. For complex anomalies, causal analysis is initiated to trace the root causes and simulate intervention measures.

[0033] S4. Through the dynamic risk assessment and early warning module, based on the real-time risk level assessed in step S3, a graded early warning is triggered. Specifically, the judgment result is matched with a four-level risk assessment model. If the risk level exceeds the threshold, an early warning signal containing the location, type, and recommended measures is immediately generated, and the risk transmission simulator is simultaneously driven to simulate the risk evolution.

[0034] S5. Display risks and coordinate responses through a visual management platform, and store the entire process of early warning and handling data in the database. Specifically, this includes: highlighting risks in a 3D digital twin scene, executing automatic or manual responses through an emergency linkage interface, and storing the entire chain of data from data collection to handling feedback in a structured and spatial database.

[0035] S6. Based on the feedback data and new case data accumulated in step S5, periodically optimize and update the model and knowledge base in the cognitive behavioral reasoning engine to complete the system closed loop. Specifically, this includes: periodically extracting labeled data from the database to retrain the behavior prediction model, reviewing and updating the rule knowledge base, and requesting labels for low-confidence scenarios through an active learning mechanism to achieve adaptive evolution of the system.

[0036] The implementation principle of an abnormal behavior detection system according to an embodiment of this application is as follows: First, multi-modal sensor networks (high-definition vision, IoT, environmental sensors) and hybrid communication networks (5G / LoRa) deployed at the construction site are used to collect multi-source heterogeneous raw data from the physical world in real time and collaboratively. Then, distributed edge computing nodes are used to perform lightweight AI model processing (target detection, tracking, initial classification) and precise spatiotemporal registration on the raw data, transforming it into a standardized structured event stream. The core layer's behavior cognition reasoning engine receives this data stream, dynamically constructs a spatiotemporal graph to represent the complex relationships between elements at the construction site, and runs knowledge base reasoning based on explicit rules and sequence prediction based on deep learning in parallel. Through confidence fusion and environmental context weighting, the system outputs an understanding of the current behavioral intent and risk, and can trace the root cause through causal analysis. Based on this understanding, the dynamic risk assessment module uses quantitative models and dynamic Bayesian networks to classify risks and simulate future evolution, triggering tiered early warnings. The visualization management platform simultaneously presents a three-dimensional view of the risk and drives on-site emergency linkages such as sound and light alarms and information push. Finally, the system systematically stores the data and management feedback of the entire handling process into the database. Through periodic model retraining and knowledge base updates, the cognitive model is continuously evolved, thus forming a continuously self-improving intelligent cycle from physical world perception to digital world optimization decision-making, and then back to physical world risk intervention.

[0037] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. An abnormal behavior detection system, characterized in that, It includes a multimodal perception layer, an edge processing layer, a behavioral cognition reasoning engine, a dynamic risk assessment and early warning module, and a visualization management platform, which are connected in sequence. The multimodal perception layer is deployed on the construction site to collect multi-source heterogeneous data, including visual data, IoT sensor data, and environmental data. The edge processing layer is communicatively connected to the multimodal perception layer and is used to perform real-time preprocessing on the multi-source heterogeneous data. The preprocessing includes target detection, initial behavior classification, and spatiotemporal registration of data. The behavioral cognitive reasoning engine is communicatively connected to the edge processing layer and is used to receive preprocessed data. By fusing rule-based reasoning and deep learning prediction, it understands the behavior and predicts the intentions of construction site personnel. The dynamic risk assessment and early warning module is connected to the behavioral cognition reasoning engine and is used to perform dynamic risk assessment and generate graded early warning signals based on the behavioral understanding and intention prediction results. The visualization management platform is connected to the dynamic risk assessment and early warning module to display risk information and execute early warning responses. It also records data throughout the process and feeds it back to the behavioral cognitive reasoning engine to drive the self-evolution of the system model.

2. The abnormal behavior detection system according to claim 1, characterized in that, The multimodal perception layer includes a visual perception unit, an Internet of Things (IoT) sensor network, and a communication network. The visual perception unit includes high-definition cameras deployed in key areas of the construction site. The high-definition cameras integrate infrared and / or thermal imaging modules and support low-light enhancement and image denoising algorithms. The Internet of Things (IoT) sensor network includes equipment status sensors installed on large construction equipment, structural monitoring sensors deployed on temporary facilities and slopes, wearable devices worn by personnel, and environmental monitoring sensors. The communication network adopts a hybrid 5G and LoRa network to transmit data collected by the visual sensing unit and the Internet of Things sensor network.

3. The abnormal behavior detection system according to claim 2, characterized in that, The edge processing layer includes multiple distributed edge computing nodes, which are used to perform at least one of the following processes: Based on a lightweight neural network model, real-time target detection and re-identification are performed on video streams. The identified targets include human bodies, safety equipment, and construction equipment. Based on the pose estimation algorithm, preliminary classification of human movements is performed; Filter and perform threshold screening on IoT sensor data to identify abnormal state signals; Based on positioning data and visual SLAM technology, time synchronization and spatial coordinate registration are performed on multi-source heterogeneous data.

4. The abnormal behavior detection system according to claim 1, characterized in that, The behavioral cognitive reasoning engine includes: The spatiotemporal graph modeling unit is used to construct spatiotemporal graphs from personnel, equipment and environmental elements, where nodes represent entities and edges represent spatiotemporal or interactive relationships between entities. The hybrid reasoning unit includes a rule knowledge base based on security specification components and a behavior prediction model based on temporal deep learning model components, which are used for parallel reasoning on input data. The fusion decision unit is used to fuse the reasoning results of the rule knowledge base with the prediction results of the behavior prediction model to output the behavior understanding and risk assessment results.

5. The abnormal behavior detection system according to claim 4, characterized in that, The behavioral cognitive reasoning engine also includes a causal analysis submodule, which is used to perform root cause analysis on abnormal behavioral events based on causal discovery algorithms and to quantitatively evaluate the effectiveness of different preventive measures based on counterfactual reasoning.

6. The abnormal behavior detection system according to claim 1, characterized in that, The dynamic risk assessment and early warning module performs the following operations: A four-level risk assessment model is constructed, which divides risks into four levels: low, medium, high, and urgent, and presets corresponding response mechanisms for each level. Based on the output of the behavioral cognitive reasoning engine, the risk level of the current scene is calculated in real time; When the risk level exceeds a preset threshold, a corresponding early warning signal and mechanism are triggered. The dynamic risk assessment and early warning module also integrates a risk transmission simulator based on a dynamic Bayesian network, which is used to deduce the propagation path and probability of risk in the future period and generate optimal risk blocking strategy suggestions.

7. As described in claim 1, characterized in that An abnormal behavior detection system, wherein the visualization management platform includes: The 3D digital twin module, based on BIM and GIS data, constructs a virtual simulation of the construction site, which is used to map the location and status of personnel, equipment, and risks in real time, and supports panoramic spatiotemporal retrospective analysis of historical early warning events using multimodal data. The emergency response interface is used to receive feedback annotations from management personnel on the early warning results and push the annotated data to the simulation training process to drive the model in the behavioral cognitive reasoning engine to perform incremental learning and optimization. The model evolution interface is used to receive feedback annotations from managers on the early warning results and push the annotated data to the model training process, driving the model in the behavioral cognitive reasoning engine to perform incremental learning and optimization.

8. The abnormal behavior detection system according to claim 7, characterized in that, The model evolution interface supports an active learning mechanism. When the confidence level of the behavioral cognitive reasoning engine in predicting a specific scenario is lower than a set threshold, the scenario data is automatically marked as a sample to be labeled and pushed to the management platform.

9. An abnormal behavior detection method, applied to the system as described in any one of claims 1-8, characterized in that, Includes the following steps: S1. Collect multi-source heterogeneous data from construction sites through a multimodal sensing layer; S2. The multi-source heterogeneous data is preprocessed in real time through the edge processing layer to obtain a structured event and state data stream; S3. The structured data stream is fused and analyzed through a behavioral cognitive reasoning engine to achieve behavioral understanding and intention prediction; S4. Through the dynamic risk assessment and early warning module, assess the real-time risk level based on the results of step S3, and trigger a graded early warning. S5. Display risks and coordinate responses through a visual management platform, while storing the entire process of early warning and handling data in the database; S6. Based on the feedback data and new case data accumulated in step S5, the model and knowledge base in the behavioral cognitive reasoning engine are periodically optimized and updated to complete the system closed loop.

10. The abnormal behavior detection method according to claim 9, characterized in that, The behavioral understanding and prediction in step S3 specifically includes: Construct a spatiotemporal map of personnel, equipment, and environment; The current state is logically complied with using a rule knowledge base; Probabilistic predictions are made using behavioral prediction models and short-term behavioral sequences. The results of rule checks and behavior predictions are integrated, and environmental context factors are introduced for weighted correction. Output behavioral risk labels with confidence levels.

Images