Artificial intelligence-based construction site safety risk intelligent assessment method and system

Abstract

The present application belongs to the cross technical field of artificial intelligence and building safety engineering, especially a building site safety risk intelligent evaluation method and system based on artificial intelligence, aiming at the problems of incomplete information caused by single perception dimension, static risk model unable to reflect the dynamic evolution of the site, lack of effective constraint of physical law on data reasoning, no active preplan generation and closed-loop control mechanism, and no cross-site collaborative evolution capability of the system, the present application proposes the following scheme, which includes the following steps: S1: pre-processing multi-source heterogeneous sensing data on the edge side; the present application solves the problems of single perception, static model, lack of physical constraint, no active preplan and data island of the traditional scheme by constructing a full-link intelligent evaluation closed loop of perception-reasoning-decision-evolution, realizes that the coupling false alarm rate is reduced by more than 60%, the preplan generation is not more than 30 seconds, and the bandwidth saving is more than 70%.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of artificial intelligence and construction safety engineering, and in particular to an intelligent assessment method and system for construction site safety risks based on artificial intelligence. Background Technology

[0002] Construction is a high-risk industry. Statistics show that falls from heights, falling objects, machinery injuries, electric shocks, and collapses account for over 85% of all construction accidents nationwide. Traditional safety management relies on manual inspections and static safety procedures, which have inherent drawbacks such as low inspection frequency, large blind spots, and delayed risk prediction.

[0003] Existing technologies mainly focus on the following areas:

[0004] First, there's the issue of helmet and seatbelt wearing detection based on a single video stream. While object detection networks are used for violation identification, this method can only identify static violations that have already occurred, failing to detect non-visual risk signals such as vibration, gas, and displacement, and even less capable of proactive risk prediction. In complex environments with low visibility and severe occlusion, the reliability of a single-vision solution drops significantly.

[0005] Second, the safety scoring system based on rule engines relies on manually created safety checklists for item-by-item scoring, lacking dynamic environmental adaptability. Once the rule base is fixed, it cannot automatically adjust the reasoning logic as construction progresses. Rules applicable during the foundation construction stage often become invalid during the main structure or decoration and finishing stages, leading to a large number of false alarms or missed alarms.

[0006] Third, single-index threshold alarms based on IoT sensors. Each sensor has an independently set alarm threshold, and an alarm is triggered when the monitored value exceeds the limit. This solution cannot capture complex risk scenarios caused by the coupling of multiple factors—such as the nonlinear risk amplification effect caused by the superposition of "tower crane overload operation + sudden gust of wind + dense personnel operation below"—the single-index threshold solution is almost completely ineffective for such scenarios.

[0007] Existing technologies suffer from several problems, including incomplete information due to a single perception dimension, inability of static risk models to reflect the dynamic evolution of construction sites, lack of effective constraints on data reasoning based on physical laws, absence of proactive contingency plan generation and closed-loop management mechanisms, and lack of cross-construction site collaborative evolution capabilities. To address these issues, we propose an AI-based intelligent assessment method and system for construction site safety risks. Summary of the Invention

[0008] The purpose of this invention is to address the shortcomings of existing technologies, such as incomplete information due to a single perception dimension, inability of static risk models to reflect the dynamic evolution of construction sites, lack of effective constraints on data reasoning based on physical laws, absence of proactive contingency plan generation and closed-loop management mechanisms, and lack of cross-construction site collaborative evolution capabilities. Therefore, this invention proposes an intelligent assessment method and system for construction site safety risks based on artificial intelligence.

[0009] The intelligent assessment method and system for construction site safety risks based on artificial intelligence provided in this application adopts the following technical solution:

[0010] The AI-based intelligent assessment method for construction site safety risks includes the following steps:

[0011] S1: Preprocess multi-source heterogeneous sensor data at the edge;

[0012] S2: Construct a spatiotemporal graph from the skeletal sequence of construction workers and perform behavioral semantic recognition;

[0013] S3: Map and evolve BIM spatial entities with real-time operation status using a four-dimensional spatiotemporal dynamic knowledge graph;

[0014] S4: Couple finite element physical fields with Bayesian networks in two directions to quantify multi-factor risk indices;

[0015] S5: Import the generated scene into the digital sandbox and perform reinforcement learning, and then optimize and generate a contingency plan;

[0016] S6: Dynamically map the risk index matrix to quantile levels and perform closed-loop feedback self-evolution.

[0017] Furthermore, in S1, five types of sensors—visual, acoustic, gas, vibration, and displacement—are deployed at the construction site according to a grid density. The edge computing gateway adopts a heterogeneous chip architecture and performs two-level preprocessing on the raw data: hardware-level filtering and normalization, and multi-sensor timestamp alignment based on a precise time protocol. The data from each sensor is mapped to a unified three-dimensional spatial coordinate grid with the origin of the building information model as the reference. The gateway has a built-in lightweight neural network that extracts only the visual motion feature vectors and uploads them to the cloud. The raw image data is destroyed on the edge side in real time and is not uploaded to the cloud.

[0018] Furthermore, in S2, a pose estimation network is used to extract key points of the human skeleton and reconstruct three-dimensional coordinates. Each frame of skeleton is constructed as a graph structure with joints as nodes and skeleton connections as edges. Cross-frame connections are added to form a spatiotemporal graph topology. Human kinematic physical constraints are introduced into the graph edges. The proportion of skeleton length and the physiological range of joint angles are used as attention bias masks. The attention weight of those whose inference angles exceed the physiological range is forced to zero. A spatiotemporal graph convolutional network is used to classify dangerous behaviors and output the semantic labels and confidence scores of the behaviors. Those with low confidence scores are pushed to manual review.

[0019] Furthermore, in S3, the building information model is imported in an international standard format, the spatial topology layer is initialized, and the knowledge graph contains three semantic layers: spatial topology layer, operation status layer, and risk association layer. In addition to the current value, the node attributes of the operation status layer also store the rate of change and acceleration in the past time period, forming a time slicing mechanism. When the acceleration exceeds the threshold, a forward-looking early warning signal is triggered. The real-time danger coupling strength between entities is dynamically calculated using a time-series graph attention network as the edge weight. The knowledge graph is stored in a three-level hybrid architecture of graph database, time-series database, and cache.

[0020] Furthermore, in S4, a bidirectional coupling architecture is adopted: in the forward channel, the real-time load is input into the finite element solver to calculate the real-time safety margin of the structure and inject it as an observation variable into the hazard source state node of the Bayesian dynamic network; in the reverse channel, the confidence of the high-risk area output by Bayesian inference is fed back to the finite element solver to dynamically refine the mesh division granularity of the area. The Bayesian network topology adopts a scoring search algorithm to automatically prune or add directed edges for risk transmission according to the construction stage, realizing adaptive inference of the foundation, main structure, and decoration stages, and outputting the real-time risk index of each area;

[0021] Rating search algorithm formula:

[0022] ;

[0023] in, The log-likelihood of the current network structure G with respect to the training data reflects the degree to which the network fits the data; the larger the value, the better the fit. The penalty term for network structure complexity is positively correlated with the total number of independent parameters Dim(G) in the network; the more independent parameters there are, the greater the penalty. : Penalty coefficient weight, a positive real number, controls the severity of the penalty; the larger the value of λ, the more the algorithm tends to choose a simple network structure.

[0024] Scoring logic: The higher the BIC score, the better the network topology is in terms of the comprehensive trade-off between "fitting accuracy" and "structural simplicity". The algorithm selects the network topology that maximizes the BIC score from the preset network structure search space as the optimal structure for the current construction stage.

[0025] Search strategy: The algorithm starts from the initial topology (which can be predefined based on security expert experience or historical incident cases) and performs the following three local operations on the current network structure:

[0026] Adding an edge: Based on the current set of directed edges for risk transmission, add a new directed edge, which represents a new type of risk transmission relationship (for example, adding a directed transmission edge for "groundwater level change → foundation pit sidewall displacement" during the foundation construction stage).

[0027] Remove edge: Delete a directed edge from the current set of directed edges, which means canceling a type of risk transmission relationship (e.g., deleting the directed transmission edge "foundation pit displacement → collapse risk" during the decoration and renovation stage).

[0028] Edge Reversal: Reverses the direction of the current directed edge, indicating an adjustment to the causal transmission direction between risk factors;

[0029] Each time a local structural adjustment is performed, the BIC score of the new topology is calculated and compared with that before the adjustment. If the score increases, the structural adjustment is retained; if the score decreases, the process is rolled back iteratively until the structural adjustment fails to improve the BIC score after a preset number of consecutive adjustments (e.g., 100 times). At this point, the search is terminated and the current optimal network topology is output.

[0030] Furthermore, in S5, a conditional generative adversarial network is constructed. Based on the current state of the knowledge graph, an extreme composite perturbation scenario located at the edge of the historical data distribution is generated to expand the adversarial sample library. The adversarial samples are imported into a virtual sandbox based on a physics engine for rigid body and fluid dynamics simulation. The reinforcement learning agent executes a Markov decision process in the sandbox. The reward function is composed of three weighted factors: risk reduction, execution time, and action intensity. The optimal candidate set of solutions is selected through multi-objective Pareto optimization.

[0031] Furthermore, in S6, a five-color classification mapping is performed based on the risk index. The threshold for each level is dynamically determined based on the quantile of the risk statistical distribution over the past period, while also being compatible with fixed standard thresholds as a fallback. Through intelligent broadcasting, mobile push, and multi-channel linkage early warning with sound and light alarms, the residual between the expected risk reduction and the measured value is calculated. Through online learning feedback, positive rewards are given to update the reinforcement learning network. When the measured value is worse than expected, the Bayesian conditional probability table is triggered for emergency fine-tuning. A federated learning framework is used to coordinate multiple construction site nodes. Each node calculates the gradient locally and then uploads it with differential privacy encryption. After being aggregated by a robust aggregator, the global model is distributed and supports one-click rollback to historical versions.

[0032] This invention also proposes an intelligent assessment system for construction site safety risks based on artificial intelligence, used to implement the aforementioned intelligent assessment method for construction site safety risks based on artificial intelligence, including:

[0033] The multimodal perception and acquisition module is used to deploy five types of sensor arrays on construction sites. It filters, normalizes and preprocesses the raw data through an edge computing gateway, and uploads only the visual motion feature vectors to the cloud. The raw image data is destroyed on the edge side in real time.

[0034] The behavioral semantic parsing module is used to extract key points of the skeleton of construction workers and construct a spatiotemporal graph topology with physical constraints. It identifies dangerous behaviors through a spatiotemporal graph convolutional network, outputs semantic labels with confidence scores, and pushes low-confidence samples for manual review.

[0035] The digital twin graph management module is used to construct a three-layer dynamic knowledge graph based on building information model. The node attribute storage change rate and acceleration form a time slicing mechanism, and the dangerous coupling strength between entities is dynamically calculated through a time sequence graph attention network.

[0036] The risk quantification reasoning module is used to inject the structural safety margin calculated by the finite element solver into the Bayesian network, and at the same time feed back the high-risk areas of the Bayesian reasoning to the finite element to refine the mesh, so as to realize bidirectional coupled reasoning and output a real-time risk index matrix.

[0037] The contingency plan generation and optimization module is used to generate an expanded adversarial sample library of extreme complex scenarios based on the current map state. The module searches for the optimal contingency plan in a virtual sandbox through a reinforcement learning agent and outputs a candidate set after multi-objective Pareto optimization.

[0038] The graded early warning linkage module is used to dynamically classify risk indices using five-color quantiles and is compatible with fixed threshold fallback. It enables graded response and timeout upgrades through multi-channel linkage early warning and has built-in storm suppression and state machine to achieve graded response and timeout upgrades.

[0039] The human-machine collaborative decision-making module provides a digital twin 3D panoramic view, a list of behavioral alarms, and a contingency plan execution dashboard, supporting safety management personnel to perform manual intervention, contingency plan confirmation, and parameter configuration.

[0040] The evolutionary learning and federated aggregation module is used to calculate the residual between the expected and measured risk reduction and update the model through online learning. It also uses federated learning and differential privacy to achieve multi-site collaborative evolution and supports model version management and one-click rollback.

[0041] The present invention also proposes a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described intelligent assessment method for construction site safety risks based on artificial intelligence.

[0042] The present invention also proposes a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described intelligent assessment method for construction site safety risks based on artificial intelligence.

[0043] In summary, this application includes at least one of the following beneficial technical effects:

[0044] 1. This solution constructs a three-layer dynamic knowledge graph based on building information modeling. In addition to storing the current value, the node attributes also store the rate of change and acceleration, forming a time slicing mechanism. This enables the system to not only perceive the current risk status, but also predict the trend of risk changes, thus solving the core defect that static rule base cannot adapt to the dynamic evolution of construction sites.

[0045] 2. This scheme injects the finite element physical field as a hard constraint into the Bayesian network as an observation variable, and at the same time, the high-risk areas of Bayesian inference drive the finite element mesh refinement in reverse, realizing the deep integration of engineering physical mechanism and probabilistic inference. Compared with the pure single index threshold alarm scheme, the false alarm rate in multi-factor coupled scenarios is reduced by more than 60%.

[0046] 3. This solution uses conditional generative adversarial networks to generate rare long-tail scenarios at the edge of historical data distribution, effectively addressing the industrial bottleneck of sparse accident data; reinforcement learning agents combined with multi-objective Pareto optimization search for optimal contingency plans in a virtual sandbox, advancing safety management from post-event alarms to pre-event rehearsals and proactive intervention, with contingency plan generation time not exceeding thirty seconds.

[0047] This invention solves the problems of traditional solutions, such as single perception, static models, lack of physical constraints, lack of proactive plans, and data silos, by constructing a closed-loop intelligent evaluation system that integrates perception, reasoning, decision-making, and evolution. It achieves a reduction of over 60% in the coupled false alarm rate, a plan generation time of no more than 30 seconds, and bandwidth savings of over 70%, while also supporting cross-site privacy collaboration and continuous self-evolution. Attached Figure Description

[0048] Figure 1 The flowchart shows the intelligent assessment method for construction site safety risks based on artificial intelligence proposed in this invention.

[0049] Figure 2 This is a structural block diagram of the intelligent assessment system for construction site safety risks based on artificial intelligence proposed in this invention. Detailed Implementation

[0050] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0051] Example 1

[0052] Reference Figure 1 and Figure 2 An AI-based intelligent assessment method for construction site safety risks includes the following steps:

[0053] S1: Multi-source heterogeneous sensor data is preprocessed at the edge. Five types of sensors—visual, acoustic, gas, vibration, and displacement—are deployed at the construction site according to grid density. The edge computing gateway adopts a heterogeneous chip architecture and performs two-level preprocessing on the raw data: hardware-level filtering and normalization, and multi-sensor timestamp alignment based on a precise time protocol. The data from each sensor is mapped to a unified three-dimensional spatial coordinate grid based on the origin of the building information model. The gateway has a built-in lightweight neural network that extracts only the visual motion feature vectors and uploads them to the cloud. The raw image data is destroyed on the edge and is not uploaded to the cloud.

[0054] S2: Construct a spatiotemporal graph from the skeletal sequence of construction workers and perform behavioral semantic recognition. Use a pose estimation network to extract key points of the human skeleton and reconstruct three-dimensional coordinates. Construct a graph structure for each frame of skeleton with joints as nodes and skeleton connections as edges, and add cross-frame connections to form a spatiotemporal graph topology. Introduce human kinematic physical constraints into the graph edges. Use the proportion of skeleton length and the physiological range of joint angles as attention bias masks. For those whose inference angles exceed the physiological range, their attention weights are forcibly reset to zero. Use a spatiotemporal graph convolutional network to classify dangerous behaviors and output behavioral semantic labels and confidence scores. Those with low confidence scores are pushed to manual review.

[0055] S3: Map and evolve BIM spatial entities and real-time operation status using a four-dimensional spatiotemporal dynamic knowledge graph. Import the BIM into the building information model in an international standard format and initialize the spatial topology layer. The knowledge graph contains three semantic layers: spatial topology layer, operation status layer, and risk association layer. In addition to the current value, the node attributes of the operation status layer also store the rate of change and acceleration in past time periods, forming a time slicing mechanism. When the acceleration exceeds the threshold, a forward-looking early warning signal is triggered. The real-time danger coupling strength between entities is dynamically calculated using a time-series graph attention network as the edge weight. The knowledge graph is stored in a three-level hybrid architecture of graph database, time-series database, and cache.

[0056] S4: The finite element physical field and Bayesian network are bidirectionally coupled for inference to quantify the multi-factor risk index. The bidirectional coupling architecture is adopted: In the forward channel, the real-time load is input into the finite element solver to calculate the real-time safety margin of the structure and inject it as an observation variable into the hazard source state node of the Bayesian dynamic network; In the reverse channel, the confidence of the high-risk area output by Bayesian inference is fed back to the finite element solver to dynamically refine the mesh granularity of the area. The Bayesian network topology adopts a scoring search algorithm to automatically prune or add directed edges for risk transmission according to the construction stage, realizing adaptive inference of the foundation, main structure and decoration stages, and outputting the real-time risk index of each area.

[0057] Rating search algorithm formula:

[0058] ;

[0059] in, The log-likelihood of the current network structure G with respect to the training data reflects the degree to which the network fits the data; the larger the value, the better the fit. The penalty term for network structure complexity is positively correlated with the total number of independent parameters Dim(G) in the network; the more independent parameters there are, the greater the penalty. : Penalty coefficient weight, a positive real number, controls the severity of the penalty; the larger the value of λ, the more the algorithm tends to choose a simple network structure.

[0060] Scoring logic: The higher the BIC score, the better the network topology is in terms of the comprehensive trade-off between "fitting accuracy" and "structural simplicity". The algorithm selects the network topology that maximizes the BIC score from the preset network structure search space as the optimal structure for the current construction stage.

[0061] Search strategy: The algorithm starts from the initial topology (which can be predefined based on security expert experience or historical incident cases) and performs the following three local operations on the current network structure:

[0062] Adding an edge: Based on the current set of directed edges for risk transmission, add a new directed edge, which represents a new type of risk transmission relationship (for example, adding a directed transmission edge for "groundwater level change → foundation pit sidewall displacement" during the foundation construction stage).

[0063] Remove edge: Delete a directed edge from the current set of directed edges, which means canceling a type of risk transmission relationship (e.g., deleting the directed transmission edge "foundation pit displacement → collapse risk" during the decoration and renovation stage).

[0064] Edge Reversal: Reverses the direction of the current directed edge, indicating an adjustment to the causal transmission direction between risk factors;

[0065] Each time a local structure adjustment is performed, the BIC score of the new topology is calculated and compared with that before the adjustment. If the score increases, the structure adjustment is retained; if the score decreases, the iteration is rolled back. The above process is executed iteratively until the structure adjustments for a consecutive preset number of times (e.g., 100 times) all fail to improve the BIC score, then the search is terminated, and the current optimal network topology is output;

[0066] S5: Import the generated scenarios into the digital sandbox for reinforcement learning, and perform plan optimization generation. Construct a conditional generative adversarial network, which takes the current knowledge graph state as a condition to generate extreme composite disturbance scenarios at the edge of the historical data distribution to expand the adversarial sample library. Import the adversarial samples into the physics engine-based virtual sandbox for rigid body and fluid dynamics simulation. The reinforcement learning agent executes a Markov decision process in the sandbox, and the reward function is composed of three weighted items: risk reduction amount, execution time, and action intensity. The optimal plan candidate set is screened out through multi-objective Pareto optimization;

[0067] S6: Perform dynamic quantile grading mapping on the risk index matrix and carry out closed-loop feedback self-evolution. Perform five-color grading mapping according to the risk index, and the thresholds of each grade dynamically define quantiles based on the risk statistical distribution in the past period, while being compatible with fixed specification thresholds as a fallback. Perform multi-channel linked early warning through intelligent broadcasting, mobile push, and acousto-optic alarm. Calculate the residual between the expected risk reduction amount and the measured value, and transmit it back through online learning to give a positive reward to update the reinforcement learning network. When the measured value is worse than expected, trigger emergency fine-tuning of the Bayesian conditional probability table. Adopt a federated learning framework to coordinate multiple construction site nodes. After each node calculates the gradient locally, it uploads the gradient after differential privacy encryption, and the global model is distributed after aggregation by the robust aggregator, supporting one-click rollback to historical versions.

[0068] This embodiment further proposes an artificial intelligence-based intelligent safety risk assessment system for construction sites, which is used to implement the above-mentioned artificial intelligence-based intelligent safety risk assessment method for construction sites, comprising:

[0069] A multi-modal perception collection module, configured to deploy five types of sensor arrays at the construction site, perform preprocessing including filtering, normalization and time synchronization on raw data through the edge computing gateway, upload only visual motion feature vectors to the cloud, and destroy original image data immediately on the edge side;

[0070] A behavior semantic parsing module, configured to extract skeletal key points of construction personnel and construct a spatio-temporal graph topology with physical constraints, identify dangerous behaviors through a spatio-temporal graph convolutional network, output semantic labels with confidence, and push low-confidence samples for manual review;

[0071] The digital twin graph management module is used to construct a three-layer dynamic knowledge graph based on building information model. The node attribute storage change rate and acceleration form a time slicing mechanism, and the dangerous coupling strength between entities is dynamically calculated through a time sequence graph attention network.

[0072] The risk quantification reasoning module is used to inject the structural safety margin calculated by the finite element solver into the Bayesian network, and at the same time feed back the high-risk areas of the Bayesian reasoning to the finite element to refine the mesh, so as to realize bidirectional coupled reasoning and output a real-time risk index matrix.

[0073] The contingency plan generation and optimization module is used to generate an expanded adversarial sample library of extreme complex scenarios based on the current map state. The module searches for the optimal contingency plan in a virtual sandbox through a reinforcement learning agent and outputs a candidate set after multi-objective Pareto optimization.

[0074] The graded early warning linkage module is used to dynamically classify risk indices using five-color quantiles and is compatible with fixed threshold fallback. It enables graded response and timeout upgrades through multi-channel linkage early warning and has built-in storm suppression and state machine to achieve graded response and timeout upgrades.

[0075] The human-machine collaborative decision-making module provides a digital twin 3D panoramic view, a list of behavioral alarms, and a contingency plan execution dashboard, supporting safety management personnel to perform manual intervention, contingency plan confirmation, and parameter configuration.

[0076] The evolutionary learning and federated aggregation module is used to calculate the residual between the expected and measured risk reduction and update the model through online learning. It also uses federated learning and differential privacy to achieve multi-site collaborative evolution and supports model version management and one-click rollback.

[0077] This embodiment also proposes a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described intelligent assessment method for construction site safety risks based on artificial intelligence.

[0078] This embodiment also proposes a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described intelligent assessment method for construction site safety risks based on artificial intelligence.

[0079] Example 2

[0080] The difference between this embodiment and Embodiment 1 is that the multimodal perception acquisition module is connected to a data quality governance module, which is used to perform data quality assessment and cleaning on the multi-source feature vectors uploaded by the edge computing gateway, including missing value detection, outlier removal, data integrity verification, and sensor drift compensation; data sources with quality scores below a preset threshold are automatically marked and downweighted for use, and a data quality report is generated and pushed to the human-machine collaborative decision-making module to ensure that the input data of the risk quantification reasoning module is reliable and effective.

[0081] Example 3

[0082] The difference between this embodiment and Embodiment 1 is that the digital twin graph management module is connected to the digital twin 3D visualization engine module, which is used to receive entity relationship data and real-time risk index matrix from the knowledge graph, construct a panoramic digital twin scene of the construction site based on the 3D graphics rendering engine, and overlay the risk heat map as a semi-transparent color layer on the corresponding spatial coordinate position of the 3D view of the building information model. It supports view rotation, zoom, sectioning and floor switching operations, and marks the warning status in the corresponding area in a dynamic highlighting and flashing manner, driving the presentation of the 3D panoramic view of the human-machine collaborative decision-making module.

[0083] Example 4

[0084] The difference between this embodiment and Embodiment 1 is that the digital twin graph management module is connected to a knowledge graph real-time query acceleration module, which is used to preload the frequently accessed spatial topological relationships and real-time operation status node attributes in the knowledge graph into the memory cache, and accelerate the multi-hop relationship query response based on graph structure indexing and adjacency list compression technology; when the hierarchical early warning linkage module triggers an alarm, the module completes the first-degree and second-degree relationship retrieval of the related entities around the dangerous area within a preset millisecond time, and pushes the retrieval results to the early warning linkage module as context information for contingency plan decision-making.

[0085] Example 5

[0086] The difference between this embodiment and Embodiment 1 is that the evolutionary learning and federated aggregation module is connected to a mobile on-site verification and closed-loop execution module, which is used to push verification tasks to the handheld terminals of on-site security personnel, including confirmation of warning location, confirmation of risk handling actions, and collection and uploading of on-site photos and short videos; the collected verification results and execution feedback data are sent back to the evolutionary learning and federated aggregation module as input for residual calculation, and at the same time, unidentified risk points found during the verification process are fed back to the digital twin graph management module as new entities, driving the dynamic incremental update of the knowledge graph.

[0087] Experimental Example

[0088] I. Experimental Objective

[0089] Through comparative experiments with traditional single-index threshold alarm schemes and vision + rule engine schemes, the actual technical effects of the present invention in multi-source sensor fusion, dynamic knowledge graph evolution, physical-data bidirectional coupling reasoning, reinforcement learning pre-plan optimization, and edge privacy protection are verified. It is demonstrated that the present invention can effectively solve the core defects of existing technologies such as single perception dimension, static risk model, lack of physical constraints, lack of proactive pre-plans, and data silos.

[0090] II. Experimental Design

[0091] Experimental Groups:

[0092] Control Group 1 (Traditional Single-Indicator Threshold Scheme): Five identical types of sensors (visual, acoustic, gas, vibration, and displacement) are deployed. Each sensor has an independently set alarm threshold, and an alarm is triggered when any monitored value exceeds the limit. Manual inspections and fixed safety checklists are used for auxiliary evaluation.

[0093] Control group 2 (vision + rule engine solution): Deploy binocular cameras, use YOLOv8 object detection network to detect the wearing of safety helmets and safety belts, and combine manually defined static rule engine for safety scoring. The rule base is fixed and does not change with the construction stage.

[0094] The present invention solution group adopts the complete method and system described in Embodiment 1 of the present invention, which includes six major steps and eight core modules, enabling edge-side preprocessing, physical constraint behavior recognition, four-dimensional spatiotemporal dynamic knowledge graph, finite element-Bayes bidirectional coupled reasoning, GAN-reinforcement learning scheme optimization, dynamic quantile hierarchical early warning and federated learning self-evolution functions.

[0095] Experimental period: 90 consecutive days.

[0096] Experimental site: Three independent construction sections (Section A, Section B, and Section C) of the same project. The construction stages, work content, and personnel configuration of each section are basically the same. Three sets of plans are deployed separately to avoid cross-interference.

[0097] Evaluation indicators:

[0098] Missed Report Rate (%): The percentage of actual incidents or violations that occur but for which the system does not generate any warnings. The true value is based on the daily inspection records and video recordings manually reviewed by on-site safety management personnel.

[0099] False alarm rate (%): The percentage of warnings issued by the system but confirmed by manual review to have no actual risk out of the total number of warnings;

[0100] Average warning lead time (minutes): The time interval between the first effective warning issued by the system and the actual occurrence of the danger or escalation of the risk, reflecting the system's risk prediction capability;

[0101] Multi-factor coupling risk identification rate (%): The identification rate of complex scenarios involving three or more risk factors at the same time, based on the assessment of human experts.

[0102] Contingency plan effectiveness (%): After the system-recommended contingency plan is implemented, the actual risk reduction reaches more than 80% of the expected risk reduction.

[0103] III. Experimental Results

[0104] Evaluation indicators Control group 1 (single-index threshold scheme) Control Group 2 (Visual + Rule Engine Solution) Invention scheme group Missed report rate (%) 68.4 52.7 26.2 False alarm rate (%) 42.1 35.6 18.3 Average warning lead time (minutes) — (Alarms are only triggered after the threshold is exceeded; they cannot be triggered in advance) 2.1 17.6 Multi-factor coupled risk identification rate (%) 18.6 31.2 79.4 Contingency plan effectiveness (%) — (No contingency plan generation function) — (No contingency plan generation function) 86.5 Average daily bandwidth usage (GB / day) 7.8 (Full upload of original video) 8.2 (Upload the full original video) 2.2 (Feature vector upload only) System response time (seconds) 3.2 4.8 1.6

[0105] IV. Results Analysis

[0106] (1) Comparison of false negative rates: The false negative rate of control group 1 was as high as 68.4%, meaning that more than two-thirds of the dangerous situations were not detected by any sensor. The main reason is that the single-index threshold scheme cannot identify scenarios with multiple overlapping factors. The false negative rate of control group 2 was 52.7%. The visual scheme can detect some behavioral violations, but it cannot perceive non-visual risks such as vibration and displacement. The false negative rate of the scheme of this invention was 26.2%, which is about 61.7% lower than that of control group 1 and about 50.3% lower than that of control group 2, verifying the effectiveness of multimodal fusion and bidirectional coupled reasoning;

[0107] (2) Comparison of false alarm rates: The false alarm rate of control group 1 was 42.1%, with a large number of transient spikes from single sensors causing invalid alarms. The false alarm rate of control group 2 was 35.6%, with the static rule base generating a large number of false alarms at different construction stages. The false alarm rate of the proposed solution group was 18.3%, a significant reduction, thanks to the stage-adaptive reasoning of the dynamic knowledge graph and the conditional probability constraints of the Bayesian network.

[0108] (3) Average warning lead time. Control group 1 only triggered the alarm after the monitored value exceeded the limit, and could not predict in advance. Control group 2 could detect the risk on average about 2.1 minutes in advance by visually identifying the violation. The solution group of this invention issued the warning on average 17.6 minutes in advance by using the time slicing mechanism (node ​​attribute change rate and acceleration) and the forward-looking warning signal, leaving sufficient response and handling window for on-site personnel;

[0109] (4) Multi-factor coupling risk identification rate: Control group 1 had a rate of only 18.6%, and control group 2 had a rate of only 31.2%. Both schemes failed to capture the nonlinear amplification effect of multiple risk factors superimposed. The scheme of this invention reached 79.4%, which is a significant improvement and verifies the core advantages of finite element-Bayes bidirectional coupled reasoning in multi-factor coupling scenarios;

[0110] (5) Effectiveness of the contingency plan: None of the control groups had the function of generating contingency plans. The contingency plan effectiveness rate of the proposed solution group reached 86.5%, indicating that the contingency plans searched and generated by the reinforcement learning agent in the virtual sandbox can achieve the expected risk reduction goal in most scenarios, verifying the effectiveness of the GAN-reinforcement learning contingency plan optimization mechanism;

[0111] (6) Bandwidth usage and system response: The solution group of this invention destroys the original image in real time at the edge and only uploads the feature vector, with an average daily bandwidth usage of 2.2GB, which is about 72% lower than the control group. The system response time is 1.6 seconds, which is better than the control group, thanks to the acceleration of edge preprocessing and the three-level caching architecture of knowledge graph.

[0112] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An intelligent assessment method for safety risks at construction sites based on artificial intelligence, characterized by: Includes the following steps: S1: Preprocess multi-source heterogeneous sensor data at the edge; S2: Construct a spatiotemporal graph from the skeletal sequence of construction workers and perform behavioral semantic recognition; S3: Map and evolve BIM spatial entities with real-time operation status using a four-dimensional spatiotemporal dynamic knowledge graph; S4: Couple finite element physical fields with Bayesian networks in two directions to quantify multi-factor risk indices; S5: Import the generated scene into the digital sandbox and perform reinforcement learning, and then optimize and generate a contingency plan. S6: Dynamically map the risk index matrix to quantile levels and perform closed-loop feedback self-evolution.

2. The intelligent assessment method for construction site safety risks based on artificial intelligence according to claim 1, characterized in that: In S1, five types of sensors—visual, acoustic, gas, vibration, and displacement—are deployed at the construction site according to a grid density. The edge computing gateway adopts a heterogeneous chip architecture and performs two-level preprocessing on the raw data: hardware-level filtering and normalization, and multi-sensor timestamp alignment based on a precise time protocol. The data from each sensor is mapped to a unified three-dimensional spatial coordinate grid with the origin of the building information model as the reference. The gateway has a built-in lightweight neural network that extracts only the visual motion feature vectors and uploads them to the cloud. The raw image data is destroyed on the edge side in real time and is not uploaded to the cloud.

3. The intelligent assessment method for construction site safety risks based on artificial intelligence according to claim 2, characterized in that: In S2, a pose estimation network is used to extract key points of the human skeleton and reconstruct three-dimensional coordinates. Each frame of skeleton is constructed as a graph structure with joints as nodes and skeleton connections as edges. Cross-frame connections are added to form a spatiotemporal graph topology. Human kinematic physical constraints are introduced into the graph edges. The proportion of skeleton length and the physiological range of joint angles are used as attention bias masks. The attention weight of those whose inference angles exceed the physiological range is forcibly reset to zero. A spatiotemporal graph convolutional network is used to classify dangerous behaviors and output the semantic labels and confidence scores of the behaviors. Those with low confidence scores are pushed to manual review.

4. The intelligent assessment method for construction site safety risks based on artificial intelligence according to claim 3, characterized in that: In S3, the building information model is imported in the international standard format, the spatial topology layer is initialized, and the knowledge graph contains three semantic layers: spatial topology layer, operation status layer and risk association layer. In addition to the current value, the node attributes of the operation status layer also store the rate of change and acceleration in the past time period, forming a time slicing mechanism. When the acceleration exceeds the threshold, a forward-looking early warning signal is triggered. The real-time danger coupling strength between entities is dynamically calculated as the edge weight using a time-series graph attention network. The knowledge graph is stored in a three-level hybrid architecture of graph database, time-series database and cache.

5. The intelligent assessment method for construction site safety risks based on artificial intelligence according to claim 4, characterized in that: In S4, a bidirectional coupling architecture is adopted: in the forward channel, the real-time load is input into the finite element solver to calculate the real-time safety margin of the structure and inject it as an observation variable into the hazard source state node of the Bayesian dynamic network; in the reverse channel, the confidence of the high-risk area output by Bayesian inference is fed back to the finite element solver to dynamically refine the mesh division granularity of the area. The Bayesian network topology adopts a scoring search algorithm to automatically prune or add directed edges for risk transmission according to the construction stage, realizing adaptive inference of the foundation, main structure and decoration stages, and outputting the real-time risk index of each area; Rating search algorithm formula: ； in, The log-likelihood of the current network structure G with respect to the training data reflects the degree to which the network fits the data; the larger the value, the better the fit. The penalty term for network structure complexity is positively correlated with the total number of independent parameters Dim(G) in the network; the more independent parameters there are, the greater the penalty. : Penalty coefficient weight, a positive real number, controls the severity of the penalty; the larger the value of λ, the more the algorithm tends to choose a simple network structure. Scoring logic: The higher the BIC score, the better the network topology is in terms of the comprehensive balance between "fitting accuracy" and "structural simplicity". The algorithm selects the network topology that maximizes the BIC score from the preset network structure search space as the optimal structure for the current construction stage.

6. The intelligent assessment method for construction site safety risks based on artificial intelligence according to claim 5, characterized in that: In S5, a conditional generative adversarial network is constructed. Based on the current state of the knowledge graph, an extreme composite perturbation scenario located at the edge of the historical data distribution is generated to expand the adversarial sample library. The adversarial samples are imported into a virtual sandbox based on a physics engine for rigid body and fluid dynamics simulation. The reinforcement learning agent executes a Markov decision process in the sandbox. The reward function is composed of three weighted factors: risk reduction, execution time, and action intensity. The optimal candidate set of solutions is selected through multi-objective Pareto optimization.

7. The intelligent assessment method for construction site safety risks based on artificial intelligence according to claim 6, characterized in that: In S6, a five-color classification mapping is performed based on the risk index. The threshold of each level is dynamically determined based on the quantile of the risk statistical distribution in the past period, while also being compatible with fixed standard thresholds as a fallback. Through intelligent broadcasting, mobile push, and multi-channel linkage early warning with sound and light alarms, the residual between the expected risk reduction and the measured value is calculated. Through online learning feedback, positive rewards are given to update the reinforcement learning network. When the measured value is worse than expected, the Bayesian conditional probability table is triggered for emergency fine-tuning. A federated learning framework is used to coordinate multiple construction site nodes. Each node calculates the gradient locally and then uploads it with differential privacy encryption. After being aggregated by a robust aggregator, the global model is distributed and supports one-click rollback to historical versions.

8. An AI-based intelligent assessment system for construction site safety risks, used to implement the AI-based intelligent assessment method for construction site safety risks as described in any one of claims 1 to 7, characterized in that, include: The multimodal perception and acquisition module is used to deploy five types of sensor arrays on construction sites. It filters, normalizes and preprocesses the raw data through an edge computing gateway, and uploads only the visual motion feature vectors to the cloud. The raw image data is destroyed on the edge side in real time. The behavioral semantic parsing module is used to extract key points of the skeleton of construction workers and construct a spatiotemporal graph topology with physical constraints. It identifies dangerous behaviors through a spatiotemporal graph convolutional network, outputs semantic labels with confidence scores, and pushes low-confidence samples for manual review. The digital twin graph management module is used to construct a three-layer dynamic knowledge graph based on building information model. The node attribute storage change rate and acceleration form a time slicing mechanism, and the dangerous coupling strength between entities is dynamically calculated through a time sequence graph attention network. The risk quantification reasoning module is used to inject the structural safety margin calculated by the finite element solver into the Bayesian network, and at the same time feed back the high-risk areas of the Bayesian reasoning to the finite element to refine the mesh, so as to realize bidirectional coupled reasoning and output a real-time risk index matrix. The contingency plan generation and optimization module is used to generate an expanded adversarial sample library of extreme complex scenarios based on the current map state. The module searches for the optimal contingency plan in a virtual sandbox through a reinforcement learning agent and outputs a candidate set after multi-objective Pareto optimization. The graded early warning linkage module is used to dynamically classify risk indices using five-color quantiles and is compatible with fixed threshold fallback. It enables graded response and timeout upgrades through multi-channel linkage early warning and has built-in storm suppression and state machine to achieve graded response and timeout upgrades. The human-machine collaborative decision-making module provides a digital twin 3D panoramic view, a list of behavioral alarms, and a contingency plan execution dashboard, supporting safety management personnel to perform manual intervention, contingency plan confirmation, and parameter configuration. The evolutionary learning and federated aggregation module is used to calculate the residual between the expected and measured risk reduction and update the model through online learning. It also uses federated learning and differential privacy to achieve multi-site collaborative evolution and supports model version management and one-click rollback.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent assessment method for construction site safety risks based on artificial intelligence, as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the intelligent assessment method for construction site safety risks based on artificial intelligence, as described in any one of claims 1 to 7.

Images