A method and system for intelligent assessment of safety risks at construction sites

Abstract

This application discloses an intelligent assessment method and system for safety risks at construction sites, belonging to the field of engineering supervision technology. The method includes: S100, real-time acquisition of multi-source data including equipment status, images and videos, personnel status, equipment distance, environmental parameters, and engineering text; S200, spatiotemporal alignment, preprocessing, and cross-modal feature fusion of the multi-source data; S300, outputting risk scores and risk level labels based on a dynamic risk knowledge graph and multi-model fusion algorithm; S400, generating tiered handling strategies according to the risk score hierarchy, and achieving closed-loop risk management through rectification verification and multi-party collaboration; S500, outputting a three-dimensional visualized risk distribution map and a compliance report. Through multi-source data fusion, dynamic risk knowledge graph, and intelligent closed-loop management technologies, the method achieves comprehensive perception, accurate assessment, efficient control, and compliance implementation of safety risks in engineering supervision, significantly reducing the incidence of safety accidents and providing reliable technical support for smart construction site construction.

Description

Technical Field

[0001] This application relates to the field of engineering supervision technology, and in particular to an intelligent assessment method and system for safety risks at construction sites. Background Technology

[0002] In the construction engineering field, project supervision is a core component in ensuring construction quality, safety, and compliance. Project supervision safety risks refer to potential threats to project safety, the safety of personnel and property, and the interests of the supervision unit itself, arising from various uncertainties during the project supervision process.

[0003] Traditional project supervision relies primarily on manual inspections, paper records, and experience-based judgment. This approach fails to promptly identify and warn of safety risks, suffers from poor real-time performance, low efficiency, and is highly prone to accidents. In some improved engineering supervision systems, IoT-based communication technologies are widely used in construction and management. These systems utilize IoT devices, such as sensors and surveillance cameras, to collect on-site data, enabling real-time monitoring of project progress, equipment status, and the working environment. IoT technology makes supervision more precise and efficient. Through real-time data analysis, supervisors can instantly understand various conditions on-site, thereby better guiding construction and ensuring project quality. Furthermore, it can identify and prevent on-site safety accidents to a certain extent.

[0004] While the aforementioned improvements have enhanced efficiency and safety accident prevention, the following technical shortcomings remain: fragmented data collection lacks effective integration, fails to consider comprehensive risk factors, and insufficient data fusion processing leads to one-sided risk assessment; static risk assessment models cannot dynamically respond to the complex and ever-changing environment of construction sites, and rule updates are not timely, resulting in insufficient adaptability; inefficient rectification of safety hazards lacks closed-loop management; lack of deep integration with regulations, compliance checks rely on manual comparison of clauses, which is time-consuming and prone to oversights, making compliance implementation difficult and posing significant legal risks; and insufficient integration with new technologies such as BIM, failing to intuitively display risk evolution trends. Summary of the Invention

[0005] Therefore, this application is proposed to address the problems and needs existing in the prior art. The objective of this application is achieved through the following technical solution:

[0006] In a first aspect, embodiments of this application provide an intelligent assessment method for safety risks at construction sites, comprising the following steps:

[0007] S100: Real-time acquisition of multi-source data including equipment status, images and videos, personnel status, equipment distance, environmental parameters, construction progress, and engineering text;

[0008] S200: Perform spatiotemporal alignment, preprocessing, and cross-modal feature fusion on multi-source data;

[0009] S300: Outputs risk scores and risk level labels based on dynamic risk knowledge graph and multi-model fusion algorithm;

[0010] S400: Generate graded disposal strategies based on risk scoring levels, and achieve closed-loop risk management through rectification verification and multi-party collaboration;

[0011] S500 outputs a 3D visualized risk distribution map and a compliance report.

[0012] In the aforementioned intelligent assessment method for construction site safety risks, step S300 includes:

[0013] S310. Input multi-source feature data into the dynamic risk knowledge graph and associate the "risk type-cause-effect" nodes in the graph;

[0014] S320. The XGBoost model is used to assign weights to static risk factors; the LSTM model is used to analyze high-frequency sensor data to predict short-term risks; and the Transformer model is used to fuse multi-source time series data to predict long-term risk trends.

[0015] S330. Calculate the risk score of the work unit based on the risk characteristic parameters, calculate the overall risk score of the project by weighted summation based on the work unit weight and associated risk value, and map the score results to the risk level and automatically associate them with the standard clauses.

[0016] S340. Pre-render a risk heat map in the BIM model and mark the location of high-risk areas and their causal chains.

[0017] In the above-mentioned intelligent assessment method for construction site safety risks, step S300 further includes updating the dynamic risk knowledge graph, specifically: updating new construction data to knowledge graph nodes in real time through incremental learning technology; dynamically optimizing the association weights between nodes based on graph attention network (GAT); and automatically expanding graph nodes and associated edges when a new risk pattern is detected.

[0018] In the aforementioned intelligent assessment method for safety risks at construction sites, the graded handling strategy in step S400 includes:

[0019] High-risk level (≥70 points): Immediately suspend work, implement a special rectification plan, and activate emergency response;

[0020] Medium risk level (30 points < risk score < 70 points): Rectification within a specified time limit, and inclusion in the key inspection list;

[0021] Low risk level (≤30 points): Log and update the monitoring plan, and review it regularly.

[0022] Secondly, embodiments of this application provide an intelligent assessment system for construction site safety risks, used to implement the aforementioned intelligent assessment method for construction site safety risks, including:

[0023] The multi-source data acquisition module is used to acquire IoT sensor data, image and video data, engineering document data, environmental parameter data, personnel biometric data, equipment distance monitoring data and construction progress data in real time at the construction site;

[0024] The heterogeneous data fusion module is used to perform spatiotemporal alignment, feature extraction, and cross-modal correlation analysis on multi-source data to generate a structured risk feature set.

[0025] The dynamic risk assessment module is used to output the overall project risk score, the work unit risk score, and the risk score between work units based on a dynamically updated risk knowledge graph and a multi-model fusion algorithm, and to generate risk level labels.

[0026] The intelligent closed-loop management module is used to generate graded disposal strategies based on risk scoring levels, and to achieve closed-loop risk control through rectification verification and multi-party collaboration.

[0027] The visualization and compliance module is used to dynamically display risk distribution through a 3D visualization interface and generate legal compliance reports that are aligned with industry standards.

[0028] In the aforementioned intelligent assessment system for safety risks at construction sites, the multi-source data acquisition module includes:

[0029] Personnel status monitoring unit is used to monitor personnel's heart rate, blood oxygen saturation, and fall behavior in real time;

[0030] The visual behavior recognition unit is used to perform real-time analysis of image and video streams based on a deep learning model, identify violations by construction workers, and output the behavior category and confidence level.

[0031] The equipment distance monitoring unit is used to calculate the spatial distance between personnel and equipment in real time; and to dynamically identify personnel entering the equipment safety warning area based on a depth perception algorithm.

[0032] The environmental and construction data acquisition unit includes the following sub-units:

[0033] The structural safety monitoring subunit is used to monitor the horizontal displacement and vertical settlement rate of the scaffolding in real time, as well as to collect settlement data of the foundation pit.

[0034] The equipment status monitoring subunit is used to monitor the tower tilt angle in real time; and to monitor the three-phase current balance and leakage current value in the temporary power circuit.

[0035] The environmental monitoring subunit is used to collect real-time data on wind speed, rainfall intensity, and groundwater level changes.

[0036] The construction data parsing subunit extracts safety hazard keywords from engineering documents based on OCR and Natural Language Processing (NLP) technologies.

[0037] The construction progress tracking sub-unit is integrated with the BIM model to compare design parameters with on-site construction progress in real time; and it collects component installation position deviation data through IoT sensors to generate progress deviation reports.

[0038] In the aforementioned intelligent assessment system for safety risks at construction sites, the heterogeneous data fusion module includes:

[0039] Spatiotemporal alignment unit is used to align multi-source data according to a unified spatiotemporal reference;

[0040] The data preprocessing unit is used to clean, filter, and standardize multi-source data.

[0041] The feature extraction unit extracts risk features for different data types;

[0042] The cross-modal feature interaction unit is used to map feature vectors of different dimensions to a unified N-dimensional space through a fully connected layer. It adopts a multi-channel Transformer model to perform cross-modal attention calculation and fuse multi-source features.

[0043] In the aforementioned intelligent risk assessment system for construction site safety, the dynamic risk assessment module includes:

[0044] The dynamic risk knowledge graph construction unit is used to build a "risk type-cause-effect" association network based on historical accident cases and safety regulations;

[0045] The multi-model fusion evaluation unit is used to assign weights to static risk factors using the XGBoost model; analyze high-frequency sensor data using the LSTM model to predict short-term risks; and fuse multi-source time-series data using the Transformer model to predict long-term risk trends.

[0046] The graded scoring unit is used to output the risk score of the work unit, map the risk score of the work unit to three levels of risk (low, medium and high) according to preset thresholds, and dynamically calculate the overall risk level of the project based on the risk score of the work unit and the associated risk values ​​between work units.

[0047] In the aforementioned intelligent assessment system for safety risks at construction sites, the intelligent closed-loop management module includes:

[0048] The tiered response strategy unit generates different response strategies based on the risk level, including:

[0049] High-risk level (≥70 points): Immediately suspend work, implement a special rectification plan, and activate emergency response;

[0050] Medium risk level (30 points < risk score < 70 points): Rectification within a specified time limit, and inclusion in the key inspection list;

[0051] Low risk level (≤30 points): Log and update monitoring plan, and review regularly;

[0052] The rectification verification unit verifies the completion of rectification measures through OCR and image comparison technology. If the standard is not met, a secondary handling process is triggered and the risk level is upgraded.

[0053] The multi-party collaborative interface unit synchronously pushes disposal instructions to the terminals of construction, supervision, and development parties, records response time and execution results, and generates a responsibility traceability log, which is associated with timestamps, geographical locations, and operator information.

[0054] In the aforementioned intelligent assessment system for construction site safety risks, the visualization and compliance module includes:

[0055] Three-dimensional risk heat map annotation unit, used to hierarchically annotate risk areas on the BIM model;

[0056] AR on-site navigation unit is used to locate potential hazards using augmented reality devices and display handling instructions;

[0057] The compliance binding unit is used to automatically match standard clauses and generate standard reports containing electronic signatures, legal basis indexes, and rectification evidence chains.

[0058] Compared with the prior art, the beneficial effects of the present invention are at least as follows:

[0059] This application presents an intelligent assessment method and system for construction site safety risks. By integrating multi-source data such as IoT sensors, visual recognition, and BIM models, and employing spatiotemporal alignment and cross-modal feature fusion technologies, it effectively solves the data fragmentation problem in traditional supervision. The system comprehensively covers risk factors such as equipment status, personnel behavior, and environmental parameters, avoiding the one-sidedness of assessments caused by single data dimensions, significantly improving the accuracy of safety hazard identification, and enhancing the comprehensiveness of risk assessment. Based on a dynamic risk knowledge graph and multi-model fusion algorithm, the system can respond to complex changes at the construction site in real time, achieving a leap from static rules to dynamic perception, demonstrating strong adaptability and significantly improving the timeliness of risk warnings. The system automatically generates graded handling strategies based on risk levels and verifies the rectification effect through technologies such as OCR image comparison. Multi-party collaborative interfaces push instructions to the construction, supervision, and construction parties in real time. The system records response time and execution results, forming a complete accountability log, significantly shortening rectification response time and greatly improving closed-loop management efficiency. Through compliance-bound units, the system automatically matches legal clauses and generates standard reports containing electronic signatures, legal basis indexes, and rectification evidence chains, avoiding omissions in manual comparison and ensuring 100% compliance inspection coverage, reducing legal risks by over 90%. It dynamically renders risk heat maps on the BIM model and supports AR on-site navigation, combined with timeline animations to display risk trends, helping managers intuitively grasp the overall risk distribution and evolution path, significantly improving decision-making efficiency. Through dynamic calculation of risk scores and weighting coefficients between work units, the system can quantify the cross-unit risk transmission effect, avoiding the limitations of isolated assessments, effectively improving the coverage of high-risk area identification, and significantly reducing the overall risk assessment error rate.

[0060] In summary, this application, through technologies such as multi-source data fusion, dynamic risk knowledge graph, and intelligent closed-loop management, has achieved comprehensive perception, accurate assessment, efficient control, and compliant implementation of safety risks in engineering supervision, significantly reducing the incidence of safety accidents and providing reliable technical support for the construction of smart construction sites. Attached Figure Description

[0061] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The accompanying drawings are used to provide a further understanding of the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof.

[0062] Figure 1 This is a schematic diagram of the structure of an intelligent assessment system for safety risks at construction sites provided in an embodiment of this application;

[0063] Figure 2 This is a schematic diagram illustrating the steps of an intelligent assessment method for safety risks at construction sites provided in an embodiment of this application;

[0064] Figure 3 This is a schematic diagram illustrating the steps of outputting a risk score based on a dynamic risk knowledge graph and a multi-model fusion algorithm provided in an embodiment of this application. Detailed Implementation

[0065] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only a part of the embodiments of this application, and this application is not limited to the exemplary embodiments described herein.

[0066] Example 1

[0067] In existing technologies, data collection is fragmented and lacks effective integration. It does not consider all risk factors and the level of data fusion processing is insufficient, leading to one-sided risk assessment. The risk assessment model is static and cannot dynamically respond to the complex and ever-changing environment of the construction site. The rules are not updated in a timely manner and lack adaptability. The rectification of safety hazards is inefficient and lacks closed-loop management. It is not deeply integrated with laws and regulations. Compliance checks rely on manual comparison of clauses, which is time-consuming and prone to omissions, making it difficult to implement compliance and posing significant legal risks. It is not well integrated with new technologies such as BIM and cannot intuitively display the trend of risk evolution.

[0068] Therefore, to address the above problems, this application provides an intelligent assessment system for safety risks at construction sites, such as... Figure 1 As shown, it includes:

[0069] The multi-source data acquisition module 100 is used to acquire IoT sensor data, image and video data, engineering document data, environmental parameter data, personnel biometric data, equipment distance monitoring data and construction progress data at the construction site in real time.

[0070] The heterogeneous data fusion module 200 is used to perform spatiotemporal alignment, feature extraction, and cross-modal correlation analysis on multi-source data to generate a structured risk feature set;

[0071] The dynamic risk assessment module 300 is used to output the overall project risk score, the work unit risk score, and the risk score between work units based on a dynamically updated risk knowledge graph and a multi-model fusion algorithm, and to generate risk level labels.

[0072] The intelligent closed-loop management module 400 is used to generate graded disposal strategies based on risk scoring levels, and to achieve closed-loop risk control through rectification verification and multi-party collaboration.

[0073] The Visualization and Compliance Module 500 is used to dynamically display risk distribution through a 3D visualization interface and generate legal compliance reports that are aligned with industry standards.

[0074] The multi-source data acquisition module 100 collects construction site data in real time through various sensors and terminal devices, and includes the following units:

[0075] The personnel status monitoring unit 110 collects real-time data on the heart rate (sampling rate 100Hz), blood oxygen saturation, and fall behavior of construction workers through PPG sensors (such as model MAX30102) and six-axis gyroscopes (such as model MPU6050) deployed on safety helmets. When an abnormal heart rate is detected (such as three consecutive tests exceeding 120 bpm) or the fall acceleration threshold (the acceleration threshold can be 2.5 to 3g), an alarm signal is triggered. When an abnormal heart rate or fatigue index is detected, a rest instruction is automatically pushed to the personal terminal and the high-risk operation privileges are suspended.

[0076] The visual behavior recognition unit 120, by deploying a high-definition camera (such as the Hikvision DS-2CD3T86 series), and based on the YOLOv8 deep learning model (input resolution 640×640, confidence threshold 0.7), performs real-time analysis on the image and video streams captured by the camera, identifies behaviors such as not wearing a safety helmet and / or not fastening a safety rope, and illegally crossing the edge protection, and outputs the behavior category and coordinate frame information;

[0077] The equipment distance monitoring unit 130, through a UWB positioning tag (such as Decawave DW1000, accuracy ±10cm) and a receiver, is installed at the end of the tower crane boom and the excavator cab. It uses a TOF algorithm to calculate the three-dimensional spatial distance between personnel and the tower crane and excavator in real time; and a depth perception algorithm based on a binocular camera (such as Intel RealSense D435, depth resolution 1280×720) to dynamically identify personnel entering the tower crane rotation radius warning zone (such as a preset safe distance of 5 meters). When personnel enter the equipment's dangerous area, it sends an emergency stop signal to the equipment operator and records the violation to the accountability system.

[0078] Environmental and construction data acquisition unit 140 includes the following sub-units:

[0079] The structural safety monitoring subunit 141 monitors horizontal displacement and vertical settlement rate in real time through displacement sensors (range ±50mm, accuracy 0.1mm) deployed at key nodes of the scaffolding; and collects foundation pit settlement data based on a preset frequency (e.g., once every 10 minutes) using a static level (accuracy ±0.01mm) installed in the deep foundation pit.

[0080] The equipment status monitoring subunit 142 is used to monitor the tower tilt angle in real time through the tower crane tilt sensor (accuracy ±0.1°). If it exceeds the preset threshold (longitudinal tilt threshold 2° or lateral tilt threshold 1.5°), an early warning is triggered. It also detects the three-phase current imbalance (threshold ±10%) and leakage current value (threshold 30mA) in the temporary power circuit using the current transformer (such as Honeywell CT-1000 model).

[0081] The environmental monitoring subunit 143 is used to collect real-time data on wind speed, rainfall intensity, and groundwater level changes through a wind speed sensor (range 0-60m / s), a rainfall meter (resolution 0.2mm), and a groundwater level monitor (accuracy ±0.1m), and dynamically compare them with environmental safety thresholds (such as wind speed threshold level 6 and hourly rainfall threshold level 30mm).

[0082] Construction data parsing subunit 144 uses OCR and natural language processing (NLP) technologies to extract safety hazard keywords (such as "collapse", "leakage", "support cracks", "unrectified") from engineering documents and combines them with the BERT model for semantic association analysis.

[0083] Construction progress tracking sub-unit 145 is integrated with the BIM model to compare design parameters with on-site construction progress in real time; and collects component installation position deviation data through IoT sensors (such as RFID tags (reading and writing distance 10m) and laser scanners) to generate progress deviation reports (color-mapped deviation range: green <5mm, red ≥10mm).

[0084] The heterogeneous data fusion module 200 processes multi-source data to generate a structured risk feature set, and includes the following unit modules:

[0085] The spatiotemporal alignment unit 210 is used to align multi-source data according to a unified spatiotemporal reference. It uses a GPS synchronized clock (time accuracy ±1ms) to add a unified timestamp to all data or uses a sliding window mechanism to unify data with different sampling frequencies to the same timestamp. Spatial alignment is achieved through coordinate system transformation (WGS84 to construction local coordinate system). For example, the ICP algorithm is used to register the BIM model with UAV point cloud data, and the sensor position is bound to the BIM component ID.

[0086] The data preprocessing unit 220 is used to clean, filter and standardize multi-source data. For example, for IoT sensor time series data, noise filtering (such as Kalman filtering) and missing value interpolation are performed; video data is decoded into H.264 format by FFmpeg; text data is cleaned by regular expressions; and image data is processed by image enhancement (histogram equalization).

[0087] The feature extraction unit 230 extracts risk features for different data types. For example, for sensor data, it extracts time-domain features through mean, variance, and peak value, and extracts frequency-domain features through FFT spectral energy. For image or video data, it outputs "safety helmet wearing confidence" and "number of missing edge protection" based on YOLOv8 target detection, and calculates abnormal personnel movement trajectories (such as entering restricted areas) through optical flow analysis. For text data, it extracts semantic features through the BERT model (such as mapping "excessive displacement of support structure" to a risk vector) and performs rule matching (such as associating with the clause number of the "Code for High-Altitude Operations in Building Construction").

[0088] The cross-modal feature interaction unit 240 is used to map feature vectors of different dimensions to a unified 128-dimensional space through a fully connected layer. It adopts a multi-channel Transformer model (8 heads, 512 hidden layer dimensions) to perform cross-modal attention calculation on sensor time series data, image features, text embedding, personnel status, equipment distance, etc. (such as associating "wind speed > level 6" sensor data, personnel heart rate of 130 beats / minute with "high-altitude operation not stopped" image behavior), fuse multi-source features, generate a structured risk feature set, and adjust the fusion weight according to the real-time risk level (such as increasing the weight of sensor data to 70% in high-risk states).

[0089] The dynamic risk assessment module 300 includes:

[0090] The dynamic risk knowledge graph construction unit 310 is used to build a "risk type-cause-effect" association network based on historical accident cases and safety regulations. For example, scaffolding displacement rate > 5 mm / h is associated with the "collapse risk" node; the keyword "unrectified" in text data is associated with the "management dereliction of duty" node; hourly rainfall of 50 mm (threshold 30 mm) is associated with the "rainstorm water accumulation" node; foundation pit settlement rate of 0.8 mm / h (threshold 0.5 mm / h) is associated with the "support failure" node; personnel heart rate > 120 beats / minute → increased probability of operational error and personnel distance from tower crane boom < 5 meters → increased collision risk, is associated with the "personnel status-equipment distance-injury type" node, and is also associated with the clauses of the "Construction Safety Inspection Standard" and corresponding disposal measures.

[0091] The multi-model fusion evaluation unit 320 is used to assign weights to static risk factors (such as construction plan compliance, material strength, foundation pit depth, scaffolding materials, etc.) using an XGBoost model (learning rate 0.1, tree depth 6); to analyze high-frequency sensor data using an LSTM model (128 hidden layers) to predict short-term risks (such as predicting the probability of sudden changes in tower crane tilt angle in the next 5-10 minutes, and the settlement rate of deep foundation pits under sudden rainstorms in the next 10 minutes); and to fuse multi-source time-series data using a Transformer model (4 encoder layers) to predict long-term risk trends (such as predicting the change in the rate of water accumulation in foundation pits caused by continuous rainstorms in the next 2-6 hours).

[0092] The graded scoring unit 330 is used to output the risk score of the work unit, map the risk score of the work unit to three levels of risk (low, medium and high) according to a preset threshold, and dynamically calculate the overall risk level of the project based on the risk score of the work unit and the associated risk values ​​between work units.

[0093] Specifically, the risk score for a work unit is calculated using the following formula:

[0094]

[0095] Where S is the risk score of the work unit, w i The weights assigned to XGBoost, f i S is the eigenvalue, α is the associated risk weight coefficient, the initial value is set according to engineering safety specifications and historical accident cases, and is dynamically adjusted according to the construction stage, environmental complexity or historical data. 关联 To score the risk of inter-work unit association (e.g., if work unit A is adjacent to high-risk unit B (such as a tower crane operation area), the association risk of A will increase due to the influence of B), the S association risk can be calculated using the following formula:

[0096]

[0097] Among them, R j For the risk score of the j-th work unit associated with the current work unit, β j The correlation weight is calculated by the GAT model based on real-time data (e.g., the higher the dependence and the closer the physical distance, the larger βj will be).

[0098] The overall project risk is calculated using the following formula:

[0099]

[0100] Where, γ k S represents the weight of work unit k. k Let β be the risk score for work unit k, and ∑S be the global correlation risk coefficient. 关联,kThis represents the total associated risk value.

[0101] The risk scoring thresholds are: high risk level (≥70 points), medium risk level (30 points < risk score < 70 points), and low risk level (≤30 points). Based on the thresholds, the risk scores of the work units and the overall project risk scores are mapped to three levels of risk: low, medium, and high.

[0102] The intelligent closed-loop management module 400 includes:

[0103] The tiered response strategy unit 410 generates different response strategies based on the risk level (matching pre-defined standardized response procedures from the response measures library), including:

[0104] High-risk level (≥70 points): Immediately suspend work, implement a special rectification plan, and activate emergency response;

[0105] Medium risk level (30 points < risk score < 70 points): Rectification within a specified time limit, and inclusion in the key inspection list;

[0106] Low risk level (≤30 points): Log and update the monitoring plan (e.g., adjust the sensor sampling frequency from 1Hz to 0.2Hz), and review regularly;

[0107] The rectification verification unit 420 verifies the completion of rectification measures (such as the number of scaffolding reinforcement points and the status of leakage protection device replacement) through OCR and image comparison technology. If the standard is not met (similarity <90%), a secondary handling process is triggered and the risk level is upgraded (such as from medium risk to high risk).

[0108] The multi-party collaboration interface unit 430 synchronously pushes the handling instructions to the terminals of the construction, supervision and construction parties, records the response time and execution results, and generates a responsibility traceability log, which is associated with timestamps, geographical locations and operator information.

[0109] The visualization and compliance module 500 includes:

[0110] The 3D risk heat map annotation unit 510 is used to hierarchically annotate risk areas on the BIM model (e.g., red corresponds to high risk, yellow corresponds to medium risk, and green corresponds to low risk).

[0111] AR on-site navigation unit 520 is used to locate potential hazards using augmented reality devices and display handling instructions;

[0112] The compliance binding unit 530 is used to automatically match standard clauses and generate a standard report containing electronic signatures, legal basis indexes, and rectification evidence chains.

[0113] The intelligent construction site safety risk assessment system in this embodiment integrates multi-source data such as IoT sensors, visual recognition, and BIM models, and employs spatiotemporal alignment and cross-modal feature fusion technologies. This effectively solves the data fragmentation problem in traditional supervision, enabling the system to comprehensively cover risk factors such as equipment status, personnel behavior, and environmental parameters. This avoids the bias caused by single data dimensions, significantly improving the accuracy of safety hazard identification and enhancing the comprehensiveness of risk assessment. Based on a dynamic risk knowledge graph and multi-model fusion algorithms, the system can respond to complex changes at the construction site in real time, achieving a leap from static rules to dynamic perception. This results in strong adaptability and significantly improved timeliness of risk warnings. The system automatically generates tiered response strategies based on risk levels and verifies rectification effects through technologies such as OCR image comparison. Multi-party collaborative interfaces push instructions to the terminals of the construction, supervision, and construction parties in real time. The system records response time and execution results, forming a complete accountability log, significantly shortening rectification response time and greatly improving closed-loop management efficiency. Through compliance binding units, the system automatically matches legal clauses and generates standard reports containing electronic signatures, legal basis indexes, and rectification evidence chains, avoiding omissions in manual comparison, ensuring 100% compliance inspection coverage, and reducing legal risks by more than 90%. It dynamically renders risk heat maps on the BIM model and supports AR on-site navigation, combined with timeline animations to display risk trends, helping managers intuitively grasp the overall risk distribution and evolution path, significantly improving decision-making efficiency. Through the dynamic calculation of risk scores and weight coefficients between work units, the system can quantify the cross-unit risk transmission effect, avoiding the limitations of isolated assessments, effectively improving the coverage of high-risk area identification, and significantly reducing the overall risk assessment error rate.

[0114] Example 2

[0115] This application also provides an intelligent assessment method for safety risks at construction sites, such as... Figure 2 As shown, it includes S100-S500.

[0116] S100: Real-time acquisition of multi-source data including equipment status, images and videos, personnel status, equipment distance, environmental parameters, construction progress, and engineering text;

[0117] S200: Perform spatiotemporal alignment, preprocessing, and cross-modal feature fusion on multi-source data;

[0118] S300: Outputs risk scores and risk level labels based on dynamic risk knowledge graph and multi-model fusion algorithm;

[0119] S400: Generate graded disposal strategies based on risk scoring levels, and achieve closed-loop risk management through rectification verification and multi-party collaboration;

[0120] S500 outputs a 3D visualized risk distribution map and a compliance report.

[0121] Among them, step S300, such as Figure 3 As shown, it further includes the following sub-steps:

[0122] S310. Input multi-source feature data into the dynamic risk knowledge graph and associate the "risk type-cause-effect" nodes in the graph;

[0123] S320. The XGBoost model is used to assign weights to static risk factors; the LSTM model is used to analyze high-frequency sensor data to predict short-term risks; and the Transformer model is used to fuse multi-source time series data to predict long-term risk trends.

[0124] S330. Calculate the risk score of the work unit based on the risk characteristic parameters, calculate the overall risk score of the project by weighted summation based on the work unit weight and associated risk value, and map the score results to the risk level and automatically associate them with the standard clauses.

[0125] S340. Pre-render a risk heat map in the BIM model and mark the location of high-risk areas and their causal chains.

[0126] Taking a scenario of sudden heavy rain in a deep foundation pit as an example, environmental data such as hourly rainfall of 50mm (exceeding the threshold of 30mm) is associated with the "heavy rain and water accumulation" node, and sensor data of foundation pit settlement rate of 0.8mm / h (exceeding the threshold of 0.5mm / h) is associated with the "support failure" node. LSTM predicts that the settlement rate will reach 1.2mm / h in the next 10 minutes, and Transformer predicts that the groundwater level will rise to the warning line after 6 hours of continuous rainfall. A new "heavy rain → support failure → collapse" association edge is added with an initial weight of 0.7. The foundation pit operation unit is scored as 90 (high risk) after calculation and is bound to Article 4.3.2 of the "Technical Specification for Foundation Pit Support". The overall project score is 75 (high risk) after weighted calculation. The foundation pit is marked as a red area in the BIM model, showing the associated risk chain "heavy rain → support failure → collapse".

[0127] The specific method for updating the dynamic risk knowledge graph is as follows:

[0128] Incremental learning techniques (such as online random forest algorithms) are used to update new construction data (such as hazard records and rectification feedback) to knowledge graph nodes in real time, triggered by conditions such as more than 100 new data entries or a risk event frequency of more than 3 times per day. Based on graph attention networks (GAT), the association weights between nodes are dynamically optimized (e.g., the weight of the causal chain "rainstorm → support displacement" is increased by 20%), and implicit associations (e.g., "personnel fatigue → operational errors → equipment collision") are discovered. When a new risk pattern is detected, the graph nodes and associated edges are automatically expanded. For example, in the scenario of a new rainstorm causing water accumulation in the foundation pit, the knowledge graph automatically adds an associated edge "rainstorm → rising groundwater level → support displacement", with the weight adjusted from 0.6 to 0.72. When a new prefabricated component installation risk is detected, the graph node is expanded and associated with "defects in hoisting scheme → loose bolts → overturning risk".

[0129] The specific processing procedures for each step and sub-step in the above-described intelligent assessment method for safety risks at construction sites have been detailed in the intelligent assessment system for safety risks at construction sites described above. Therefore, their repeated description will be omitted here.

[0130] Example 3

[0131] Specifically, the graded treatment strategy in step S400 of Embodiment 2 includes:

[0132] High-risk level (≥70 points): Immediately suspend work, implement a special rectification plan, and activate emergency response;

[0133] Medium risk level (30 points < risk score < 70 points): Rectification within a specified time limit, and inclusion in the key inspection list;

[0134] Low risk level (≤30 points): Log and update the monitoring plan, and review it regularly.

[0135] In summary, this application, through technologies such as multi-source data fusion, dynamic risk knowledge graph, and intelligent closed-loop management, has achieved comprehensive perception, accurate assessment, efficient control, and compliant implementation of safety risks in engineering supervision, significantly reducing the incidence of safety accidents and providing reliable technical support for the construction of smart construction sites.

[0136] The basic principles of this application have been described above with reference to specific embodiments. It should be understood that the specific details disclosed above are for illustrative and illustrative purposes only, and are not intended to limit the scope of protection of this invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for intelligent assessment of safety risks at construction sites, characterized in that, Includes the following steps: S100: Real-time acquisition of multi-source data including equipment status, images and videos, personnel status, equipment distance, environmental parameters, construction progress, and engineering text; Multi-source data includes: real-time monitoring of personnel heart rate, blood oxygen saturation, and fall behavior; real-time analysis of image and video streams based on deep learning models to identify construction worker violations and output behavior categories and confidence levels; real-time calculation of spatial distance between personnel and equipment; dynamic identification of personnel entering equipment safety warning zones based on deep perception algorithms; real-time monitoring of scaffold horizontal displacement and vertical settlement rate; collection of foundation pit settlement data; real-time monitoring of tower tilt angle; monitoring of three-phase current balance and leakage current values ​​in temporary power circuits; real-time collection of wind speed, rainfall intensity, and groundwater level change data; extraction of safety hazard keywords from engineering documents based on OCR and natural language processing technologies; real-time comparison of design parameters with on-site construction progress; and collection of component installation position deviation data through IoT sensors. S200. Spatiotemporal alignment, preprocessing, and cross-modal feature fusion of multi-source data: Multi-source data is aligned according to a unified spatiotemporal benchmark. A GPS synchronized clock is used to add a unified timestamp to all data, or a sliding window mechanism is used to unify data from different sampling frequencies to the same timestamp. Spatial alignment is achieved through coordinate system transformation. The ICP algorithm is used to register the BIM model with UAV point cloud data, and sensor locations are bound to BIM component IDs. Risk features are extracted for different data types. For sensor data, temporal features are extracted using mean, variance, and peak value; frequency domain features are extracted using FFT spectral energy. For images or visual data… For frequency data, YOLOv8-based target detection is used to output "helmet wearing confidence" and "number of missing edge protection". Optical flow analysis is used to calculate abnormal personnel movement trajectories. For text data, semantic features are extracted using the BERT model and rule matching is performed. Feature vectors of different dimensions are mapped to a unified 128-dimensional space through a fully connected layer. A multi-channel Transformer model is used to perform cross-modal attention calculations on sensor time-series data, image features, text embeddings, personnel status, and equipment distance. Multi-source features are fused to generate a structured risk feature set, and the fusion weights are adjusted according to the real-time risk level. S300. Output risk scores and risk level labels based on dynamic risk knowledge graphs and multi-model fusion algorithms; including: S310. Inputting multi-source feature data into the dynamic risk knowledge graph and associating the "risk type-cause-effect" nodes in the graph; S320. Using the XGBoost model to assign weights to static risk factors; using the LSTM model to analyze high-frequency sensor data and predict short-term risks; using the Transformer model to fuse multi-source time-series data and predict long-term risk trends; S330. Calculating the risk score of the work unit based on risk feature parameters, calculating the overall project risk score by weighted summation based on the work unit weights and associated risk values, and mapping the score results to risk levels and automatically associating them with standard clauses; S340. Pre-rendering a risk heat map in the BIM model and marking the location of high-risk areas and causal chains; S400: Generate graded disposal strategies based on risk scoring levels, and achieve closed-loop risk management through rectification verification and multi-party collaboration; S500 outputs a 3D visualized risk distribution map and a compliance report.

2. The intelligent assessment method for construction site safety risks as described in claim 1, characterized in that, Step S300 further includes: The dynamic risk knowledge graph is updated in the following ways: new construction data is updated to the knowledge graph nodes in real time through incremental learning technology; and the association weights between nodes are dynamically optimized based on graph attention networks. When a new risk pattern is detected, the graph nodes and associated edges are automatically expanded.

3. The intelligent assessment method for construction site safety risks as described in claim 2, characterized in that, The graded handling strategy in step S400 includes: High-risk level: Risk score ≥ 70 points, work must be stopped immediately, a special rectification plan must be issued and an emergency response must be initiated; Medium risk level: 30 points < risk score < 70 points, rectification within a time limit, included in the key inspection list; Low risk level: Risk score ≤ 30 points, log and update monitoring plan, and review regularly.

4. A construction site safety risk intelligent assessment system, used to implement the construction site safety risk intelligent assessment method according to any one of claims 1-3, characterized in that, include: The multi-source data acquisition module is used to acquire IoT sensor data, image and video data, engineering document data, environmental parameter data, personnel biometric data, equipment distance monitoring data and construction progress data in real time at the construction site; The heterogeneous data fusion module is used to perform spatiotemporal alignment, feature extraction, and cross-modal correlation analysis on multi-source data to generate a structured risk feature set. The dynamic risk assessment module is used to output the overall project risk score, the work unit risk score, and the risk score between work units based on a dynamically updated risk knowledge graph and a multi-model fusion algorithm, and to generate risk level labels. The intelligent closed-loop management module is used to generate graded disposal strategies based on risk scoring levels, and to achieve closed-loop risk control through rectification verification and multi-party collaboration. The visualization and compliance module is used to dynamically display risk distribution through a 3D visualization interface and generate legal compliance reports that are aligned with industry standards.

5. The intelligent assessment system for construction site safety risks as described in claim 4, characterized in that, The multi-source data acquisition module includes: Personnel status monitoring unit is used to monitor personnel's heart rate, blood oxygen saturation, and fall behavior in real time; The visual behavior recognition unit is used to perform real-time analysis of image and video streams based on a deep learning model, identify violations by construction workers, and output the behavior category and confidence level. The equipment distance monitoring unit is used to calculate the spatial distance between personnel and equipment in real time; and to dynamically identify personnel entering the equipment safety warning area based on a depth perception algorithm. The environmental and construction data acquisition unit includes the following sub-units: The structural safety monitoring subunit is used to monitor the horizontal displacement and vertical settlement rate of the scaffolding in real time, as well as to collect settlement data of the foundation pit. The equipment status monitoring subunit is used to monitor the tower tilt angle in real time; and to monitor the three-phase current balance and leakage current value in the temporary power circuit. The environmental monitoring subunit is used to collect real-time data on wind speed, rainfall intensity, and groundwater level changes. The construction data parsing sub-unit extracts safety hazard keywords from engineering documents based on OCR and natural language processing technologies. The construction progress tracking sub-unit is integrated with the BIM model to compare design parameters with on-site construction progress in real time; and it collects component installation position deviation data through IoT sensors to generate progress deviation reports.

6. The intelligent assessment system for construction site safety risks as described in claim 4, characterized in that, The heterogeneous data fusion module includes: Spatiotemporal alignment unit is used to align multi-source data according to a unified spatiotemporal reference; The data preprocessing unit is used to clean, filter, and standardize multi-source data. The feature extraction unit extracts risk features for different data types; The cross-modal feature interaction unit is used to map feature vectors of different dimensions to a unified N-dimensional space through a fully connected layer. It adopts a multi-channel Transformer model to perform cross-modal attention calculation and fuse multi-source features.

7. The intelligent assessment system for construction site safety risks as described in claim 4, characterized in that, The dynamic risk assessment module includes: The dynamic risk knowledge graph construction unit is used to build a "risk type-cause-effect" association network based on historical accident cases and safety regulations; The multi-model fusion evaluation unit is used to assign weights to static risk factors using the XGBoost model; analyze high-frequency sensor data using the LSTM model to predict short-term risks; and fuse multi-source time-series data using the Transformer model to predict long-term risk trends. The graded scoring unit is used to output the risk score of the work unit, map the risk score of the work unit to three levels of risk (low, medium and high) according to preset thresholds, and dynamically calculate the overall risk level of the project based on the risk score of the work unit and the associated risk values ​​between work units.

8. The intelligent assessment system for construction site safety risks as described in claim 4, characterized in that, The intelligent closed-loop management module includes: The tiered response strategy unit generates different response strategies based on the risk level, including: High-risk level: Immediately halt work, implement a specific rectification plan, and activate emergency response; Medium-risk level: Rectify within a time limit and include in the key inspection list; Low-risk level: Record logs and update monitoring plans, and review regularly; The rectification verification unit verifies the completion of rectification measures through OCR and image comparison technology. If the standard is not met, a secondary handling process is triggered and the risk level is upgraded. The multi-party collaborative interface unit synchronously pushes disposal instructions to the terminals of construction, supervision, and development parties, records response time and execution results, and generates a responsibility traceability log, which is associated with timestamps, geographical locations, and operator information.

9. The intelligent assessment system for construction site safety risks as described in claim 4, characterized in that, The visualization and compliance module includes: Three-dimensional risk heat map annotation unit, used to hierarchically annotate risk areas on the BIM model; AR on-site navigation unit is used to locate potential hazards using augmented reality devices and display handling instructions; The compliance binding unit is used to automatically match standard clauses and generate standard reports containing electronic signatures, legal basis indexes, and rectification evidence chains.

Images