A construction site personnel safety behavior analysis and control system

Abstract

The application discloses a construction site personnel safety behavior analysis and management and control system and relates to the technical field of building safety management. Through multi-modal heterogeneous data fusion collection, multi-source information such as video data, personnel positioning data, environment data and equipment state data of a construction site is acquired. Data preprocessing technology is used to clean, space-time align, feature extract and standardize different types, formats and frequencies of data, and unified format space-time behavior data is generated. On this basis, a standardized shadow behavior classification system is established, a large-scale shadow behavior sample library is constructed, a space-time causal graph neural network algorithm is adopted, chain causal correlation rules between shadow behaviors and accidents are mined, and an accident evolution chain model is constructed, which can accurately predict the probability and time of accidents. Through real-time analysis of the current personnel behavior sequence, a hierarchical early warning can be given 10-30 minutes before the occurrence of a violation behavior.

Description

Technical Field

[0001] This invention relates to the field of building safety management technology, specifically to a system for analyzing and controlling the safety behavior of personnel at construction sites. Background Technology

[0002] As a vital pillar of the national economy, the construction industry is characterized by complex environments, dense workforces, and diverse work types. On construction sites, due to the interplay of various factors, safety accidents occur frequently, causing significant losses to the lives of construction workers and company property. Therefore, the effective analysis and control of personnel safety behaviors on construction sites has always been a key task and core challenge in the field of construction safety management. With the continuous development of technology, emerging technologies such as computer vision, sensor technology, and artificial intelligence are gradually being applied to construction safety management, providing new ideas and methods for improving the level of safety control on construction sites.

[0003] Existing construction site personnel safety behavior analysis systems primarily rely on computer vision technology to identify explicit violations, such as not wearing a safety helmet, not wearing a safety belt, or entering a hazardous area. However, these traditional technologies have several serious drawbacks. First, warnings are severely delayed, only issuing alerts after the violation has actually occurred, by which time the risk of an accident is already extremely high, rendering them ineffective in preventing accidents. Second, the scope of identification is extremely limited, only able to identify predefined explicit violations. They lack the ability to identify implicit risk behaviors that do not violate safety regulations but have a very high probability of causing subsequent violations or accidents. Furthermore, they lack causal analysis capabilities; traditional technologies can only determine whether a single behavior is a violation, failing to deeply analyze the causal relationships and evolutionary trends between behaviors. This makes it difficult to predict accidents in advance, and the false alarm rate is high, easily misclassifying normal behavior as a violation. This not only significantly reduces the system's usability but also generates resistance from on-site personnel, further hindering the effective implementation of safety control measures. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a construction site personnel safety behavior analysis and control system. This system acquires multi-source information from construction sites, including video data, personnel positioning data, environmental data, and equipment status data, through multi-modal heterogeneous data fusion. Using data preprocessing techniques, it cleans, spatiotemporally aligns, extracts features, and standardizes data of different types, formats, and frequencies to generate unified-format spatiotemporal behavior data. Based on this, a standardized shadow behavior classification system is established, a large-scale shadow behavior sample library is constructed, and a spatiotemporal causal graph neural network algorithm is used to mine the chain-like causal relationship between shadow behavior and accidents, constructing an accident evolution chain model. This model can accurately predict the probability and timing of accidents. By analyzing the current personnel behavior sequence in real time, it issues tiered warnings 10-30 minutes before violations occur, shifting the safety control focus from post-event handling and in-event intervention to pre-event prediction.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a construction site personnel safety behavior analysis and control system, the system comprising: Multimodal data acquisition module: Collects video data, personnel location data, environmental data, and equipment status data from the construction site; Multi-source heterogeneous data preprocessing module: Cleans, aligns, extracts features and standardizes the collected multi-source heterogeneous data to generate spatiotemporal behavioral data in a unified format; Shadow Behavior Definition and Labeling Module: Establish a standardized shadow behavior classification system, label shadow behaviors in historical data, and build a large-scale shadow behavior sample library; Spatiotemporal behavior sequence construction module: converts continuous behavioral data of personnel into a four-dimensional behavior sequence containing time, space and semantic dimensions; Spatiotemporal causal chain association mining module: Employs an improved spatiotemporal causal graph neural network algorithm to mine the causal association patterns between shadow behaviors and subsequent violations and accidents, and constructs an accident evolution chain model; Dynamic risk assessment and graded early warning module: Analyzes the current behavior sequence of personnel in real time, predicts the probability and time of accident occurrence based on the accident evolution chain model, and issues graded early warnings; Intelligent closed-loop management module: Based on early warning information, it links on-site audible and visual alarm devices, broadcasting systems, access control systems, and smart wearable devices to achieve timely intervention and control of at-risk personnel; Incremental model update and optimization module: Continuously collects new shadow behavior samples and accident data to perform incremental learning and optimization of the model.

[0006] Furthermore, the shadow behavior definition and annotation module, combined with the actual needs of construction site safety management, establishes a standardized shadow behavior classification system including four categories: tool operation, body posture, position movement, and interactive behavior, through case studies of construction safety accidents over the years and on-site work behavior surveys. It clarifies the judgment criteria and identification boundaries of various shadow behaviors. Then, based on the behavior data annotation tool, it annotates the historical monitoring videos, accident retrospective video data, and real-time behavior data collected on-site for each time period, accurately recording the occurrence time, spatial location, subject of the behavior, and potential risk association characteristics of each shadow behavior. It eliminates annotation deviations, missing data, and duplicate invalid samples, and integrates them into a large-scale shadow behavior sample library after verification and calibration.

[0007] Furthermore, the spatiotemporal behavior sequence construction module, combining the continuous characteristics of on-site personnel operations, transforms the continuous behavior data of each construction worker into a four-dimensional behavior sequence containing time, space, behavioral semantics, and environmental and equipment state dimensions, based on the standardized spatiotemporal behavior data output by the multi-source heterogeneous data preprocessing module. Specifically, this is represented as follows: ,in The three-dimensional spatial coordinates of the representative at time t correspond to the real-time location data collected by the UWB positioning system. Let be the feature vector of a person's behavioral state at time t. It is the environmental feature vector at time t. Let T be the device state feature vector at time t, where T represents the sequence length. Set the time granularity to 1 second / frame, select 300 frames (i.e. 5 minutes) as the length of a single sequence, and use a sliding window step size of 30 frames (i.e. 30 seconds) and an overlap rate of 90% for sequence truncation. At the same time, through a pre-trained behavior recognition model, the original visual features and trajectory features are encoded into a 128-dimensional behavior state vector to achieve a unified semantic representation of different types of behavior.

[0008] Furthermore, the spatiotemporal causal chain association mining module, as a core functional module of the system, uses an improved spatiotemporal causal graph neural network algorithm based on the preprocessed four-dimensional spatiotemporal behavior sequence to mine the chain causal association patterns between shadow behaviors and subsequent violations and accidents. This leads to the construction of an accident evolution chain model and a ternary behavior graph G=(V, E) containing personnel nodes, spatial nodes, and time nodes. Here, V represents the node set, encompassing personnel nodes Vp representing construction workers, spatial nodes Vs representing different work areas at the construction site, and time nodes Vt representing different time points. E represents the personnel-space, personnel-time, and space-time relationships between nodes. This is achieved by introducing spatiotemporal causality... An attention mechanism is used to calculate the causal contribution of each behavior node to the subsequent accident. Then, a graph convolutional network (GCN) is used to extract the spatial correlation features of the behavior graph, and a bidirectional LSTM network is used to capture the temporal dependency features of the behavior sequence. The attention mechanism achieves deep fusion of spatiotemporal features. The model is trained with shadow behavior sequences and corresponding accident labels as inputs and accident probability and expected occurrence time as outputs. A weighted fusion loss function is used to train the model. Adam optimizer and Dropout regularization are used to prevent the model from overfitting. At the same time, the causal contribution of each shadow behavior to the accident is calculated through backpropagation, and the key shadow behavior with the greatest impact on the accident is identified. Finally, a complete accident evolution chain model is constructed.

[0009] Furthermore, the spatiotemporal causal chain association mining module introduces a spatiotemporal causal attention mechanism to calculate the causal contribution of each behavioral node to the occurrence of subsequent accidents. The formula for calculating its attention weight is as follows: ,in It represents the spatiotemporal causal attention weight of behavior node j to node i, characterizing the degree of correlation and influence of the behavior corresponding to node j on subsequent accidents. , These are the feature vectors of nodes i and j in the behavior graph, respectively. It is the feature linear transformation matrix, used for dimension mapping and feature extraction of node features. It is the attention weight coefficient vector, which represents the learnable parameters of the model; is the set of neighboring nodes of node i in the behavioral graph; k is the traversal index in the set of neighboring nodes.

[0010] Furthermore, in the spatiotemporal causal chain association mining module, the expression for the weighted fusion loss function is: ,in, It is the total loss value of the accident evolution chain model. , These are loss weighting coefficients, used to balance the loss proportions of classification and regression tasks. It is the cross-entropy loss for the accident occurrence probability classification task. It is the mean square error loss of the regression task for the expected occurrence time of the accident.

[0011] Furthermore, the spatiotemporal causal chain correlation mining module uses backpropagation to trace back the accident occurrence probability and expected occurrence time indicators output by the model to each shadow behavior feature node constituting the spatiotemporal behavior sequence. It calculates the gradient contribution value of the feature vector corresponding to each shadow behavior to the final prediction result of the model, thereby quantifying the causal contribution of each shadow behavior to the occurrence of the accident. The larger the gradient contribution value, the stronger the driving effect of the shadow behavior on the occurrence of the accident. Combined with the actual needs of safety management and control at the construction site, a causal contribution threshold is preset. All shadow behaviors involved in the calculation are sorted from high to low contribution. Shadow behaviors with contribution values ​​higher than the threshold are selected. Then, false correlation behaviors caused by data deviation are eliminated through cross-validation. Finally, the key shadow behaviors with the greatest impact on the occurrence of the accident are determined.

[0012] Furthermore, the dynamic risk assessment and graded early warning module captures and preprocesses the behavioral data of construction personnel in real time based on multi-source data collected at the construction site, quickly generates corresponding four-dimensional spatiotemporal behavioral sequences, and inputs them into the pre-trained accident evolution chain model in real time. Through the accident evolution chain model, the probability P of the accident caused by the current personnel behavior and the expected occurrence time are accurately predicted, thereby realizing the dynamic assessment of construction safety risks. In addition, a three-level graded early warning mechanism is established in combination with the actual needs of construction site safety management.

[0013] Furthermore, the three-tiered early warning mechanism is as follows: A Level 1 warning is triggered when the probability of an accident is 30% ≤ P < 50% and the expected occurrence time is 20-30 minutes. A Level II warning is triggered when the probability of an accident is 50% ≤ P < 80% and the expected occurrence time is 10-20 minutes. A Level 3 warning is triggered when the probability of an accident (P) is greater than or equal to 80% and the expected time of occurrence is less than 10 minutes.

[0014] Compared with existing technologies, this construction site personnel safety behavior analysis and control system has the following beneficial effects: This invention proposes a shadow behavior model and constructs a construction site safety management system based on multimodal data fusion and spatiotemporal causal reasoning, realizing a paradigm shift from passive post-event handling to proactive pre-event prediction. Through multi-source heterogeneous data fusion and standardized processing, it can accurately capture shadow behaviors that do not violate regulations but have accident precursor characteristics. Based on an improved spatiotemporal causal graph neural network algorithm, a chain evolution model of behavior-accident is constructed, which can provide graded early warnings 10-30 minutes before violations occur, significantly advancing the accident prevention window. Through intelligent device linkage and closed-loop management mechanisms, a full-process management chain is formed, which not only improves the initiative and accuracy of construction site safety management, but also ensures the continuous optimization capability of the technical solution through an incremental model update mechanism.

[0015] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0017] Figure 1 This is a structural block diagram of a construction site personnel safety behavior analysis and control system. Figure 2 A flowchart of a spatiotemporal causal chain correlation mining module for a construction site personnel safety behavior analysis and control system; Figure 3 This is a flowchart of a construction site personnel safety behavior analysis and control system. Detailed Implementation

[0018] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0019] This invention provides a system for analyzing and controlling the safety behavior of personnel at construction sites, such as... Figure 1As shown, the system includes a multimodal data acquisition module, a multi-source heterogeneous data preprocessing module, a shadow behavior definition and annotation module, a spatiotemporal behavior sequence construction module, a spatiotemporal causal chain correlation mining module, a dynamic risk assessment and graded early warning module, an intelligent closed-loop management and control module, and an incremental model update and optimization module. Through multimodal heterogeneous data fusion acquisition, it obtains multi-source information such as video data, personnel positioning data, environmental data, and equipment status data from the construction site. Using data preprocessing technology, it cleans, spatiotemporally aligns, extracts features, and standardizes data of different types, formats, and frequencies to generate spatiotemporal behavior data in a unified format. Based on this, a standardized shadow behavior classification system is established, a large-scale shadow behavior sample library is constructed, and a spatiotemporal causal graph neural network algorithm is used to mine the chain causal correlation between shadow behavior and accidents, constructing an accident evolution chain model. This model can accurately predict the probability and timing of accidents. By analyzing the current personnel behavior sequence in real time, it issues graded early warnings 10-30 minutes before violations occur, shifting the safety control focus from post-event handling and in-event intervention to pre-event prediction.

[0020] In key areas of the construction site, such as high-altitude operation areas, hot work areas, hoisting operation areas, temporary power supply areas, entrances and exits, and main passages, video acquisition equipment, ultra-wideband personnel positioning base stations, environmental sensors, and equipment IoT terminals are deployed to achieve data collection in all scenarios and at all times. Video data acquisition: Real-time video streams from the construction site are acquired through video acquisition equipment. The frame rate is set to 25 frames per second, and the resolution is adapted to the on-site monitoring needs. It covers all high-risk work areas and areas with dense personnel activity, focusing on capturing personnel's movements, tool usage, and interactive behaviors.

[0021] Personnel positioning data acquisition: Through the ultra-wideband personnel positioning system, the three-dimensional spatial coordinates of the positioning tags worn by construction personnel are obtained in real time. The positioning update frequency is 1Hz, which can accurately distinguish the positions of personnel on different floors and different work surfaces.

[0022] Environmental data acquisition: Environmental parameters at the construction site are collected in real time through environmental sensors such as temperature, humidity, wind speed, dust, and toxic and harmful gases. The acquisition frequency is set to 1-5 minutes / time according to the characteristics of environmental changes. When environmental parameters exceed the safety threshold, the acquisition frequency is automatically increased.

[0023] Equipment status data acquisition: IoT terminals installed on the construction equipment collect status parameters such as operating speed, load, start / stop status, fault codes, and vibration amplitude at a frequency of 1Hz, allowing for real-time monitoring of equipment operation. All collected data is transmitted to the system's backend server for further processing via on-site industrial Ethernet and wireless communication networks.

[0024] The multi-source heterogeneous data preprocessing module sequentially cleans, aligns, extracts features, and standardizes the collected multi-source heterogeneous data to generate spatiotemporal behavioral data in a unified format. The specific steps are as follows: Data cleaning: For video data, image filtering and target detection algorithms are used to remove invalid frames that are blurry, occluded, or too dark / too bright, while background frames without human activity are also filtered out; for location data, Kalman filtering is used to remove jump points, drift points, and duplicate data, and linear interpolation is used to complete location data that is missing for no more than 5 seconds, while segments missing for more than 5 seconds are marked as invalid data; for environmental and equipment data, the 3σ criterion is used to remove outliers, and missing values ​​are completed using the average of adjacent sensors within the same time period.

[0025] Spatiotemporal alignment: Based on the unified timestamp of the system server, video, positioning, environmental and equipment data with different acquisition frequencies are synchronized in time to ensure that various types of data at the same time can correspond one by one; at the same time, a unified three-dimensional spatial coordinate system is established based on the BIM model of the construction site, and all spatial data are mapped to this coordinate system to achieve spatial data alignment of different acquisition devices.

[0026] Feature extraction: For video data, the coordinates of 17 key skeletal points of the person are extracted using a human keypoint detection algorithm. At the same time, visual features such as the person's movement trajectory, tool holding status, and helmet wearing status are also extracted. For positioning data, trajectory features such as the person's movement speed, acceleration, dwell time, and activity range are extracted. For environmental data, a comprehensive environmental risk index is calculated by weighted summation. For equipment data, status features such as equipment operating load rate, abnormal vibration amplitude, and continuous operating time are extracted.

[0027] Standardization process: The Z-score standardization method is used to convert all feature values ​​with different dimensions and different value ranges into standard data with a mean of 0 and a variance of 1, thereby eliminating the difference in dimensions between features and generating spatiotemporal behavior data in a unified format.

[0028] The shadow behavior definition and annotation module, combined with the actual needs of construction site safety management, establishes a standardized shadow behavior classification system and constructs a large-scale sample library. The specific implementation process is as follows: We analyzed various safety accident cases in the construction industry over the past 10 years and extracted all warning behaviors within one hour before the accident occurred. At the same time, we conducted an in-depth field survey on construction sites for three months to observe the routine and abnormal behaviors of different trades such as carpenters, steelworkers, scaffolders, and electricians in different work scenarios. Based on this, we established a standardized shadow behavior classification system that includes four categories: tool operation, body posture, position movement, and interactive behavior.

[0029] Using behavioral data annotation tools, historical monitoring videos from the construction site over the past three years, accident retrospective video data, and real-time behavioral data collected on-site were imported and annotated at a time granularity of 1 second / frame. The annotation content included the occurrence time of the shadow behavior, its spatial location, the job type and unique identifier of the entity performing the behavior, the behavior type, the duration of the behavior, and potential risk association characteristics. After annotation, a two-person cross-verification mechanism was adopted, with two annotation engineers independently annotating the same batch of data. After removing annotation biases, missing data, and duplicate or invalid samples, all valid samples were categorized and stored according to behavior type, work scenario, and risk level, constructing a large-scale shadow behavior sample library.

[0030] The spatiotemporal behavior sequence construction module, taking into account the continuous nature of on-site personnel operations, transforms the continuous behavior data of each construction worker into a four-dimensional spatiotemporal behavior sequence. Specifically, it involves: acquiring standardized spatiotemporal behavior data from a multi-source heterogeneous data preprocessing module; grouping the data according to the unique identifier of each construction worker; and extracting the continuous behavior records for each worker. The mathematical expression for constructing the four-dimensional behavior sequence is as follows: The three-dimensional spatial coordinates of the representative at time t correspond to the real-time location data collected by the UWB positioning system. The feature vector of a person’s behavioral state at time t includes features such as the coordinates of key points on the human body, action type, and tool usage status. It is the environmental feature vector at time t, which includes features such as temperature, humidity, wind speed, and dust concentration; Let T be the equipment state feature vector at time t, which includes features such as equipment operating speed, load, and fault codes; T represents the length of a single behavior sequence.

[0031] A time granularity of 1 second / frame was set, and 300 frames were selected as the length of a single behavior sequence. A sliding window step of 30 frames and an overlap rate of 90% were used to extract the sequence of continuous behavior data to ensure the continuity and integrity of the behavior. At the same time, a pre-trained human behavior recognition model was used to encode the original visual features and trajectory features into a 128-dimensional behavior state vector, realizing a unified semantic representation of different types and dimensions of behavior features, and providing standardized input for subsequent causal association mining.

[0032] The spatiotemporal causal chain correlation mining module employs an improved spatiotemporal causal graph neural network algorithm to uncover causal correlation patterns between shadow behaviors and subsequent violations and accidents, constructing an accident evolution chain model. The specific steps are as follows: Ternary Behavior Graph Construction: Based on the preprocessed four-dimensional spatiotemporal behavior sequence, a ternary behavior graph G=(V, E) is constructed, containing personnel nodes, spatial nodes, and temporal nodes. The node set V includes: Personnel nodes are characterized by attributes such as the type of work, length of service, safety training records, and historical violation records of construction workers. Spatial nodes are characterized by attributes such as the risk level of the work area, the type of work, the distribution of surrounding equipment, and the historical accident rate. Time nodes are characterized by attributes such as work period, weather conditions, construction progress, and personnel density.

[0033] The edge set E represents the relationships between nodes, including personnel-space relationships (the time a person stays in a certain area and the frequency of their activities), personnel-time relationships (the work behavior of a person during a certain period of time), and space-time relationships (the work intensity and accident rate of a certain area during a certain period of time).

[0034] Spatiotemporal causal attention mechanism calculation: A spatiotemporal causal attention mechanism is introduced to calculate the causal contribution of each behavioral node to the occurrence of subsequent events. The attention weight calculation formula is as follows: ,in, It is the spatiotemporal causal attention weight of behavior node j to node i, which represents the degree of correlation between the behavior corresponding to node j and subsequent accidents; , , respectively, are the feature vectors of node i and node j in the behavior graph; W is a learnable feature linear transformation matrix, used for dimensionality mapping and feature purification of node features; a is a learnable attention weight coefficient vector; It is the set of neighboring nodes of node i in the behavior graph. Through this mechanism, the model can adaptively focus on the behavior nodes that have a greater impact on the occurrence of the accident and weaken the interference of irrelevant nodes.

[0035] Spatiotemporal Feature Fusion and Model Training: A Graph Convolutional Network (GCN) is used to perform convolution operations on the ternary behavior graph to extract spatial correlation features between nodes in the behavior graph, capturing the spatial dependencies between different people, regions, and times. Simultaneously, a Bi-Short Memory Network (Bi-LSTM) is used to process the four-dimensional spatiotemporal behavior sequence, extracting forward and reverse temporal dependency features respectively. The extracted spatial correlation and temporal dependency features are then deeply fused using an attention mechanism to obtain the fused spatiotemporal feature vector.

[0036] Using shadow behavior sequences and corresponding accident labels (whether an accident occurred, and the time of occurrence) as input, and the accident occurrence probability and expected occurrence time as output, an accident evolution chain model is constructed. A weighted fusion loss function is used to train the model, and the loss function expression is: ,in, It is the total loss value of the model; , These are the loss weight coefficients, set to 0.6 and 0.4 respectively, to balance the loss proportions of classification and regression tasks; It is the cross-entropy loss for the accident probability classification task; The mean squared error loss is used for the accident prediction time regression task. The Adam optimizer is used during training, with an initial learning rate of 0.001 and a batch size of 32. Dropout regularization is also introduced with a dropout rate of 0.3 to prevent model overfitting.

[0037] After model training, backpropagation is used to trace back the accident probability and predicted time of occurrence output by the model to each shadow behavior feature node constituting the spatiotemporal behavioral sequence. The gradient contribution value of the feature vector corresponding to each shadow behavior to the final prediction result of the model is calculated, thereby quantifying the causal contribution of each shadow behavior to the accident. The preset causal contribution threshold is 0.15. All shadow behaviors are sorted from high to low contribution, and shadow behaviors with a contribution value higher than the threshold are selected. Then, 5-fold cross-validation is used to eliminate spurious correlation behaviors caused by data bias. Finally, the key shadow behaviors with the greatest impact on the accident are determined. Based on these key shadow behaviors and their causal relationships, a complete accident evolution chain model is constructed, clearly showing the complete evolution path from the appearance of shadow behaviors to the occurrence of violations and then to the outbreak of accidents.

[0038] The dynamic risk assessment and tiered early warning module performs real-time dynamic risk assessments on the behavior of construction workers and issues tiered early warnings based on the assessment results. Specifically, it is implemented as follows: Real-time acquisition of multi-source data from the construction site, processed by a multi-source heterogeneous data preprocessing module, rapidly generates a four-dimensional spatiotemporal behavior sequence of the current construction personnel. This sequence is then input into a pre-trained accident evolution chain model in real time. Through inference calculations, the model outputs the probability of an accident caused by the current personnel's behavior and the estimated time of occurrence within one second, achieving millisecond-level dynamic assessment of construction safety risks.

[0039] Based on the actual needs of safety management at construction sites, a three-tiered early warning mechanism is established: Level 1 Warning: When the probability of an accident is 30% ≤ P < 50% and the expected occurrence time is 20-30 minutes, a Level 1 warning is triggered. The warning information is pushed to the mobile terminal of the on-site safety officer in the form of a text message, prompting the safety officer to go to the relevant area to verbally remind and conduct safety checks on the personnel involved.

[0040] Level 2 warning: When the probability of an accident is 50%≤P<80% and the expected occurrence time is 10-20 minutes, a Level 2 warning is triggered. The warning information is simultaneously pushed to the mobile terminals of the on-site safety officer and the construction team leader. The on-site audible and visual alarm equipment emits a yellow warning light and a slow prompt sound, and the broadcast system plays targeted safety prompts.

[0041] Level 3 warning: When the probability of an accident is P≥80% and the expected occurrence time is less than 10 minutes, a Level 3 warning is triggered. The warning information is pushed to the mobile terminals of the on-site safety officer, construction team leader and project safety manager. The on-site sound and light alarm equipment emits red warning lights and urgent alarm sounds, and the broadcast system plays emergency evacuation prompts in a loop.

[0042] Based on different levels of early warning information, the intelligent closed-loop control module automatically links various on-site control devices to achieve tiered intervention and closed-loop control. Level 1 Early Warning and Control: The system automatically records early warning information, information on involved personnel, and behavioral data, generates standardized safety inspection work orders, and pushes them to on-site safety officers. After completing on-site inspections and rectifications, safety officers upload rectification photos and explanations to the system. The system automatically performs closed-loop confirmation and stores the rectification records in the safety control ledger.

[0043] Level 2 Early Warning and Control: In addition to implementing the control measures for Level 1 early warning, the system automatically retrieves real-time monitoring videos of the relevant areas and displays them in a pop-up window on the large screen in the central control room, facilitating remote monitoring of the on-site situation by management personnel. Simultaneously, the system records the entire rectification process via video data, enabling subsequent traceability and review.

[0044] Level 3 Early Warning and Control: The system immediately initiates emergency intervention measures, linking with the access control system to close entrances and exits to the hazardous area to prevent unauthorized personnel from entering; smart wearable devices worn by construction workers emit continuous vibrations and voice alarms, directly reminding at-risk personnel to immediately stop work and evacuate to a safe area. Simultaneously, the system automatically generates an emergency response plan, pushes it to the project's emergency command team, and initiates the emergency response process.

[0045] All early warning information, intervention measures, rectification results, and emergency response processes will be fully recorded by the system, forming a traceable safety management and control ledger.

[0046] The incremental model update and optimization module enables the system to continuously self-optimize during operation, ensuring that the model performance adapts to the dynamic changes at the construction site: The system collects new shadow behavior samples and accident data in real time, including new types of shadow behaviors labeled on site, abnormal behaviors not recognized by the model, and actual accident data. When the number of newly collected valid samples reaches 1,000, the system automatically starts the incremental learning process, fine-tunes the accident evolution chain model based on the new sample data, and only updates the parameters of the last few layers of the model without retraining the entire model, which greatly improves the efficiency of model update.

[0047] Meanwhile, the system conducts a comprehensive monthly evaluation of the model's prediction accuracy, false positive rate, and false negative rate, and dynamically adjusts the model's hyperparameters, including learning rate, batch size, and dropout rate, based on the evaluation results. Furthermore, the system updates its shadow behavior classification system and sample library quarterly, incorporating new work methods and tool usage-related new shadow behaviors into the classification system, continuously enriching the diversity and representativeness of the sample library to ensure the model can consistently and accurately identify various safety risks.

[0048] The specific operation steps of the construction site personnel safety behavior analysis and control system provided by this invention are as follows: Multimodal data acquisition: Simultaneously acquire four types of data—video, personnel positioning, environment, and equipment status—through various on-site sensing devices; Multi-source heterogeneous data preprocessing: The collected raw data is cleaned, spatiotemporally aligned, feature extracted and standardized to generate spatiotemporal behavioral data in a unified format; Construction of the shadow behavior sample library: Based on historical data and field research, establish a standardized shadow behavior classification system, complete data annotation and sample verification, and form a large-scale sample library; Four-dimensional spatiotemporal behavior sequence construction: transforming the continuous behavior data of a single construction worker into a four-dimensional behavior sequence that includes time, space, semantics, environment, and equipment status; Accident evolution chain model construction: An improved spatiotemporal causal graph neural network algorithm is used to mine the causal relationship between shadow behaviors and accidents, identify key shadow behaviors, and construct an accident evolution chain model. Dynamic risk assessment and graded early warning: Input the current personnel behavior sequence in real time, and predict the probability and time of the accident through model reasoning, triggering the corresponding level of early warning; Intelligent closed-loop management: Based on the warning level, the system will link on-site equipment, implement graded intervention measures, and complete rectification acceptance and full-process data archiving. Incremental model updates: Continuously collect new samples and accident data, and regularly fine-tune the model, evaluate performance, and update the classification system to achieve system self-optimization.

[0049] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A system for analyzing and controlling the safe behavior of personnel at construction sites, characterized in that, The system includes: Multimodal data acquisition module: Collects video data, personnel location data, environmental data, and equipment status data from the construction site; Multi-source heterogeneous data preprocessing module: Cleans, aligns, extracts features and standardizes the collected multi-source heterogeneous data to generate spatiotemporal behavioral data in a unified format; Shadow Behavior Definition and Labeling Module: Establish a standardized shadow behavior classification system, label shadow behaviors in historical data, and build a large-scale shadow behavior sample library; Spatiotemporal behavior sequence construction module: converts continuous behavioral data of personnel into a four-dimensional behavior sequence containing time, space and semantic dimensions; Spatiotemporal causal chain association mining module: Employs an improved spatiotemporal causal graph neural network algorithm to mine the causal association patterns between shadow behaviors and subsequent violations and accidents, and constructs an accident evolution chain model; Dynamic risk assessment and graded early warning module: Analyzes the current behavior sequence of personnel in real time, predicts the probability and time of accident occurrence based on the accident evolution chain model, and issues graded early warnings; Intelligent closed-loop management module: Based on early warning information, it links on-site audible and visual alarm devices, broadcasting systems, access control systems, and smart wearable devices to achieve timely intervention and control of at-risk personnel; Incremental model update and optimization module: Continuously collects new shadow behavior samples and accident data to perform incremental learning and optimization of the model.

2. The construction site personnel safety behavior analysis and control system according to claim 1, characterized in that, The shadow behavior definition and annotation module, combined with the actual needs of construction site safety management, establishes a standardized shadow behavior classification system based on years of construction safety accident cases and on-site work behavior surveys. This system includes four categories: tool operation, body posture, position movement, and interactive behavior. It clarifies the judgment criteria and identification boundaries of various shadow behaviors. Then, using behavior data annotation tools, it annotates historical monitoring videos, accident retrospective video data, and real-time behavior data collected on-site for each time period. This accurately records the occurrence time, spatial location, subject of the behavior, and potential risk association characteristics of each shadow behavior. It eliminates annotation biases, missing data, and duplicate invalid samples. After verification and calibration, the data is integrated to form a large-scale shadow behavior sample library.

3. The construction site personnel safety behavior analysis and control system according to claim 1, characterized in that, The spatiotemporal behavior sequence construction module, combining the continuous characteristics of on-site personnel operations, transforms the continuous behavior data of each construction worker into a four-dimensional behavior sequence containing time, space, behavioral semantics, and environmental and equipment state dimensions, based on the standardized spatiotemporal behavior data output by the multi-source heterogeneous data preprocessing module. Specifically, it is represented as follows: ,in The three-dimensional spatial coordinates of the representative at time t correspond to the real-time location data collected by the UWB positioning system. Let be the feature vector of a person's behavioral state at time t. It is the environmental feature vector at time t. Let be the device state feature vector at time t, and T represent the sequence length.

4. The construction site personnel safety behavior analysis and control system according to claim 1, characterized in that, The spatiotemporal causal chain association mining module, as the core functional module of the system, is based on the preprocessed four-dimensional spatiotemporal behavior sequence. It adopts an improved spatiotemporal causal graph neural network algorithm to mine the chain causal association pattern between shadow behavior and subsequent violations and accidents. Then, it constructs an accident evolution chain model and builds a ternary behavior graph G=(V,E) containing personnel nodes, spatial nodes, and time nodes. Here, V represents the node set, which includes personnel nodes Vp representing construction personnel, spatial nodes Vs representing different work areas of the construction site, and time nodes Vt representing different time points. E represents the personnel-space, personnel-time, and space-time association relationships between each node. By introducing a spatiotemporal causal attention mechanism, the causal contribution of each behavioral node to the subsequent accident is calculated. Then, a graph convolutional network (GCN) is used to extract the spatial correlation features of the behavioral graph, and a bidirectional LSTM network is used to capture the temporal dependency features of the behavioral sequence. The attention mechanism is used to achieve deep fusion of spatiotemporal features. The model is trained with shadow behavioral sequences and corresponding accident labels as inputs and accident probability and expected occurrence time as outputs. A weighted fusion loss function is used to train the model. Adam optimizer and Dropout regularization measures are used to prevent the model from overfitting. At the same time, the causal contribution of each shadow behavior to the accident is calculated through backpropagation, and the key shadow behavior with the greatest impact on the accident is identified. Finally, a complete accident evolution chain model is constructed.

5. The construction site personnel safety behavior analysis and control system according to claim 4, characterized in that, The spatiotemporal causal chain association mining module introduces a spatiotemporal causal attention mechanism to calculate the causal contribution of each behavioral node to the occurrence of subsequent accidents. The formula for calculating its attention weight is as follows: ,in It represents the spatiotemporal causal attention weight of behavior node j to node i, characterizing the degree of correlation and influence of the behavior corresponding to node j on subsequent accidents. , These are the feature vectors of nodes i and j in the behavior graph, respectively. It is a linear transformation matrix for features, used for dimension mapping and feature extraction of node features. It is the attention weight coefficient vector, which represents the learnable parameters of the model; is the set of neighboring nodes of node i in the behavioral graph; k is the traversal index in the set of neighboring nodes.

6. The construction site personnel safety behavior analysis and control system according to claim 4, characterized in that, In the spatiotemporal causal chain association mining module, the expression for the weighted fusion loss function is: ,in, It is the total loss value of the accident evolution chain model. , These are the loss weighting coefficients, used to balance the loss proportions between classification and regression tasks. It is the cross-entropy loss for the accident occurrence probability classification task. It is the mean square error loss of the regression task for the expected occurrence time of the accident.

7. The construction site personnel safety behavior analysis and control system according to claim 4, characterized in that, The spatiotemporal causal chain correlation mining module uses backpropagation to trace back the accident occurrence probability and expected occurrence time indicators output by the model to each shadow behavior feature node that constitutes the spatiotemporal behavior sequence. It calculates the gradient contribution value of the feature vector corresponding to each shadow behavior to the final prediction result of the model, thereby quantifying the causal contribution of each shadow behavior to the occurrence of the accident. The larger the gradient contribution value, the stronger the driving effect of the shadow behavior on the occurrence of the accident. Based on the actual needs of safety management at the construction site, a causal contribution threshold is preset. All shadow behaviors involved in the calculation are sorted from high to low according to their contribution. Shadow behaviors with a contribution higher than the threshold are selected. Then, false correlation behaviors caused by data deviation are eliminated through cross-validation. Finally, the key shadow behaviors with the greatest impact on the occurrence of the accident are determined.

8. The construction site personnel safety behavior analysis and control system according to claim 1, characterized in that, The dynamic risk assessment and graded early warning module captures and preprocesses the behavioral data of construction personnel in real time based on multi-source data collected at the construction site, quickly generating corresponding four-dimensional spatiotemporal behavioral sequences. These sequences are then input into a pre-trained accident evolution chain model. The model accurately predicts the probability P of an accident caused by the current personnel behavior and the expected time of occurrence, thereby achieving dynamic assessment of construction safety risks. Furthermore, a three-level graded early warning mechanism is established in conjunction with the actual needs of safety management at the construction site.

9. A construction site personnel safety behavior analysis and control system according to claim 8, characterized in that, The three-tiered early warning mechanism is as follows: A Level 1 warning is triggered when the probability of an accident is 30% ≤ P < 50% and the expected occurrence time is 20-30 minutes. A Level II warning is triggered when the probability of an accident is 50% ≤ P < 80% and the expected occurrence time is 10-20 minutes. A Level 3 warning is triggered when the probability of an accident (P) is greater than or equal to 80% and the expected time of occurrence is less than 10 minutes.

Images