An industrial safety risk assessment method based on behavior simulation and artificial intelligence

Abstract

The application provides an industrial safety risk assessment method based on behavior simulation and artificial intelligence. The method obtains industrial field basic data through multiple protocols, obtains real-time data through clock unification, multi-scale window alignment and quality detection; uses multi-agent behavior simulation to enumerate disturbances, constructs contrast scene data, maps various data to a unified feature space for domain adaptive training; constructs a space-time graph containing multiple types of nodes and multiple relationship edges, inputs into a space-time graph neural network for online inference calculation of abnormal score and output of Top-K explanation; combines soft prior learning to obtain a root cause chain through intervention verification; fuses the abnormal score and the model uncertainty to determine an adaptive threshold, and divides the risk level into four levels of no risk, early stage, serious and emergency; outputs the risk level and the root cause chain, and executes the corresponding safety strategy according to the level. The method realizes accurate assessment and hierarchical control of industrial safety risk.

Description

Technical Field

[0001] The embodiments of this application relate to the field of industrial safety risk assessment technology, and in particular to an industrial safety risk assessment method based on behavioral simulation and artificial intelligence. Background Technology

[0002] Industrial safety risk assessment is a core component of industrial production and operation. Its purpose is to identify, predict, and control potential hazards to prevent accidents, protect personnel safety, and reduce property damage. With the acceleration of industrialization and the increasing complexity of processes, industrial safety has evolved from traditional passive protection to proactive prevention and intelligent management. This places higher demands on the real-time performance, accuracy, and interpretability of industrial safety risk assessments.

[0003] Existing industrial safety risk assessment technologies mainly rely on the following methods. The first is a real-time monitoring method based on sensor networks and video surveillance. This involves deploying sensors for temperature, pressure, gas concentration, etc., along with video surveillance equipment at the production site. Assessments are triggered manually or by a rule engine based on preset thresholds to detect abnormal events. The second method is a safety training and scenario simulation method based on virtual simulation and multi-agent behavior modeling. This method utilizes a multi-agent system to construct a virtual environment that simulates personnel behavior, work processes, and accident scenarios for safety drills and decision support.

[0004] However, these technologies all have significant shortcomings. Real-time monitoring methods rely on preset rules and fixed thresholds, lacking the ability to dynamically interact with real-time data, making it impossible to predict accidents in their early stages, and struggling to explain the root causes of anomalies. Scene data generated based on virtual simulation and multi-agent behavior modeling methods lacks effective integration with real production data, leading to a disconnect between simulation results and actual risk conditions. Furthermore, existing technologies fail to uniformly model the complex interrelationships of "human-machine-material-method-environment," resulting in insufficient ability to identify cross-domain coupled risks. In addition, the risk assessment outputs of existing technologies lack a closed-loop feedback mechanism with the physical world, failing to provide actionable response strategies and severely limiting their practicality. Summary of the Invention

[0005] To address the aforementioned technical issues, embodiments of this application propose an industrial safety risk assessment method based on behavioral simulation and artificial intelligence. This method supplements scarce tail samples through multi-agent behavioral simulation, models the spatiotemporal dependencies of "human-machine-material-method-environment" using a unified spatiotemporal graph neural network, outputs an interpretable root cause chain by combining causal structure learning, and learns intervention strategies in the simulation environment using reinforcement learning. This forms a closed-loop disposal system that links recommendations with approval, thereby achieving intelligent industrial safety management that is assessable, interpretable, and executable.

[0006] To achieve the above objectives, embodiments of this application propose an industrial safety risk assessment method based on behavioral simulation and artificial intelligence. The method includes: acquiring basic data from the industrial site, including time-series signals, video streams, and personnel trajectory data, through multi-protocol access; unifying the clock reference based on the NTP protocol; performing multi-scale window unification on basic data with different sampling rates; and performing missing data detection and drift identification to obtain actual sampled data; enumerating disturbances including valve position misalignment, equipment degradation, and human-caused violations through multi-agent behavioral simulation to construct control scenario data and supplement scarce tail samples; wherein the constructed control scenario data includes at least normal scenario data, abnormal scenario data, and extreme scenario data; mapping the control scenario data, actual sampled data, and historical actual sampled data to a unified feature space; and performing domain adaptive training according to quantile thresholds for equipment type, work group, and seasonal maintenance; and using measurement points, equipment, areas, personnel, and work tickets as nodes. Using process relationships, spatial relationships, business relationships, and behavioral relationships as multiple relation edges, a spatiotemporal graph is constructed and input into a spatiotemporal graph neural network for online inference to calculate anomaly scores. Simultaneously, node-level Top-K explanations and subgraph-level Top-K explanations are output. A pre-generated prior graph is used as a soft constraint to learn causal structures under this constraint. The node-level and subgraph-level Top-K explanations are then validated through intervention to obtain the root cause chain. Anomaly scores, quantile thresholds, and the uncertainty of the spatiotemporal graph neural network are fused to determine an adaptive threshold. Based on the anomaly scores and the adaptive threshold, the industrial safety risk level is determined. The industrial safety risk level is divided into four levels: no risk, early stage, severe, and urgent. The industrial safety risk level and root cause chain are output, and corresponding safety policies are executed. The no-risk and early stage levels directly execute the corresponding safety policies, while the severe and urgent levels execute their policies only after operational approval.

[0007] To achieve the above objectives, embodiments of this application also propose an industrial safety risk assessment system based on behavioral simulation and artificial intelligence. The system includes: a data acquisition module for acquiring basic data from the industrial site, including time-series signals, video streams, and personnel trajectory data, through multi-protocol access; a data governance module for unifying the clock reference based on the NTP protocol, performing multi-scale window unification on basic data with different sampling rates, and performing missing data detection and drift identification to obtain actual sampling data; a behavioral simulation module for enumerating disturbances including valve position misalignment, equipment degradation, and human-caused violations through multi-agent behavioral simulation to construct control scenario data and supplement scarce tail samples; the constructed control scenario data includes at least normal scenario data, abnormal scenario data, and extreme scenario data; a domain adaptive training module for mapping the control scenario data, actual sampling data, and historical actual sampling data to a unified feature space, and performing domain adaptive training according to quantile thresholds for equipment type, work group, and seasonal maintenance; and a space-time anomaly detection module for detecting anomalies based on measurement points, equipment, areas, and personnel. Using work tickets as nodes and process relationships, spatial relationships, business relationships, and behavioral relationships as multi-relationship edges, a spatiotemporal graph is constructed and input into a spatiotemporal graph neural network for online inference to calculate anomaly scores. Simultaneously, it outputs node-level Top-K explanations and subgraph-level Top-K explanations. A causal inference module uses a pre-generated prior graph as a soft constraint to learn the causal structure under this constraint and intervenes to verify the node-level and subgraph-level Top-K explanations, obtaining the root cause chain. A risk grading module fuses anomaly scores, quantile thresholds, and the uncertainty of the spatiotemporal graph neural network to determine an adaptive threshold, and then determines the industrial safety risk level based on the anomaly score and the adaptive threshold. The industrial safety risk level is divided into four levels: no risk, early stage, severe, and urgent. A policy execution and approval linkage module outputs the industrial safety risk level and root cause chain, and executes the corresponding safety policy. The no-risk and early stage levels directly execute the corresponding safety policy, while the severe and urgent levels execute the corresponding safety policy only after operation and maintenance approval.

[0008] To achieve the above objectives, embodiments of this application also propose an electronic device, including: a processor and a memory, wherein the memory stores instructions executable by the processor, and the processor is configured to execute the instructions such that the electronic device can implement the industrial safety risk assessment method based on behavioral simulation and artificial intelligence as described above.

[0009] To achieve the above objectives, embodiments of this application also propose a computer-readable storage medium storing a computer program that, when executed by a processor, enables an industrial safety risk assessment method based on behavioral simulation and artificial intelligence as described above.

[0010] Optionally, basic data from the industrial site, including time-series signals, video streams, and personnel trajectory data, is acquired through multi-protocol access. A unified clock reference is established based on the NTP (Network Time Protocol). Multi-scale windowing is performed on basic data with different sampling rates, and missing data detection and drift identification are conducted to obtain the actual collected data. This includes data from the industrial site's DCS (Distributed Control System), PLC (Programmable Logic Controller), and SCADA (Supervisory Control and Data Acquisition) systems via protocols including OPC UA (OPC Unified Architecture), Modbus, MQTT (Message Queuing Telemetry Transport), and GB28181. The system acquires basic data from data acquisition and monitoring control systems (data acquisition and monitoring control), sensors, cameras, access control devices, and personnel positioning devices, and configures the basic data with data source identifiers, sampling timestamps, and quality identifiers. The sampling rates of basic data from different sources vary. Based on the NTP protocol, the sampling timestamps of the basic data are uniformly calibrated to obtain calibration timestamps under a unified clock reference, and the basic data is aligned to a unified time axis according to a preset time granularity. Multi-scale windowing processing is performed on the basic data on the unified time axis using sliding windows with at least two different time scales. The system aggregates, interpolates, and resamples the basic data within the same sliding window to obtain multi-scale window data. Based on the quality identifier, zero-order hold duration and threshold, run length, and lag ratio in the multi-scale window data, it performs missing data detection and lag detection, and performs neighborhood interpolation or model interpolation on the detected missing segments to obtain interpolated multi-scale window data. Based on the statistical distribution changes within the sliding window at different time scales, it performs drift identification on the interpolated multi-scale window data and writes the drift identifier into the corresponding multi-scale window data to obtain the sampled data carrying the quality identifier and drift identifier.

[0011] Optionally, through multi-agent behavior simulation, disturbances including valve position misalignment, equipment degradation, and human error are enumerated to construct comparative scenario data and supplement scarce tail samples. This includes: constructing a multi-agent simulation model including personnel, equipment, environment, and work tasks, and configuring state parameters, behavioral rules, and interaction constraints for different agents; establishing a disturbance parameter set based on the industrial site's process flow, equipment operating boundaries, and operating procedures, wherein the disturbance parameter set includes at least valve position misalignment parameters, equipment degradation parameters, and human error parameters; and performing single-factor enumeration and / or multi-factor combination enumeration based on the disturbance parameter set according to preset disturbance intensity, disturbance duration, disturbance occurrence sequence, and disturbance coupling relationship. The simulation process generates simulation disturbance instances. These instances are then input into a multi-agent simulation model, driving different agents to update their states and evolve interactively under a unified simulation clock, resulting in corresponding simulation process data. Based on the disturbance type, disturbance intensity, and risk evolution results corresponding to the simulation process data, the simulation process data is divided into normal, abnormal, and extreme scenarios. Scenario labels, disturbance labels, and risk labels are then assigned to each control scenario. Simulation process data carrying scenario labels, disturbance labels, and risk labels are used as control scenario data. Samples belonging to low-frequency, high-risk disturbance combinations in the control scenario data are identified as tail samples to supplement the scarce industrial safety risk samples in historical data, i.e., to supplement the scarce tail samples.

[0012] Optionally, the comparison scenario data, actual collected data, and historical actual collected data are mapped to a unified feature space, and domain-adaptive training is performed according to the quantile thresholds of equipment type, work group, and seasonal maintenance, including: Window-level feature extraction was performed on the comparison scenario data, actual data, and historical actual data to obtain multiple initial feature vectors; among them, the initial feature vectors include at least time-frequency domain features, personnel trajectory features, operation compliance features, and topological association features; According to the unified feature encoding rules, the initial feature vector is normalized, missing mask encoded, quality identifier encoded and scene label encoded to obtain the sample feature vector in the unified feature space. All sample feature vectors are divided into domains according to equipment type, work group and season to obtain multiple training domains. For each training domain, a quantile threshold is maintained based on the abnormal distribution statistics corresponding to the sample feature vectors in that domain. The quantile thresholds include the first-level quantile threshold, the second-level quantile threshold and the third-level quantile threshold. Domain adaptive training based on sample feature vectors reduces the feature distance between sample feature vectors with the same scene label and the same risk label in different training domains within a unified feature space, while increasing the feature distance between sample feature vectors with different scene labels and different risk labels within the unified feature space, thus obtaining a domain adaptive feature representation. Among these, the domain adaptive feature representation, quantile threshold, and sample feature vectors are the basis for subsequent construction of spatiotemporal maps and determination of adaptive thresholds. Among them, the Each sample feature vector Represented as: ; ; in, Indicates the first 1 initial feature vector For feature mapping function, This indicates the comparison scene data. This indicates the actual collected data. This represents historical data collected. Represents the characteristics of time-series signals. This indicates the event characteristics corresponding to the video stream. Indicates the characteristics of personnel trajectories; Indicates the characteristics of operational compliance. Indicates topological association features, For missing mask, For quality labeling, The tags include scene tags, disturbance tags, and risk tags; Record No. The device type, the first The class group, the The training domain for each season is , express: ; in, Indicates the device type. Indicates work group, Indicates the season; First-order quantile threshold Second-level quantile threshold and third-order quantile threshold These can be expressed as formulas: ; ; ; in, It is a quantile function. Represents the training domain The set of training scores corresponding to the feature vectors of each sample within the set. , , For the preset quantiles, .

[0013] Optionally, a spatiotemporal graph is constructed using measurement points, equipment, areas, personnel, and work permits as nodes, and process relationships, spatial relationships, business relationships, and behavioral relationships as multi-relationship edges, including: Based on domain-adaptive feature representation and the correspondence between sample feature vectors and industrial sites, a spatiotemporal map is constructed for the current time window. ;in, , For the current time window, Represents a set of nodes. Denotes the set of edges. Represents a set of edge relation types. Represents a multi-relation adjacency matrix. Represents the node feature matrix; Map measurement points, equipment, areas, personnel, and work permits as sets of nodes. The nodes in the diagram are mapped to a set of edge relationship types, representing technological, spatial, business, and behavioral relationships. Different types of edges, node feature matrices Automatic generation of multi-relation adjacency matrices based on domain-adaptive feature representation Dynamic updates are based on process connection relationships, spatial distance, operational relationships, and personnel behavior relationships; The spatiotemporal graph is input into a spatiotemporal graph neural network for online inference, anomaly scores are calculated, and node-level Top-K explanations and subgraph-level Top-K explanations are output, including: According to the preset online inference window, candidate subgraph sequences are extracted from the spatiotemporal spectra under multiple consecutive time windows. Represented as: ; in, Indicates the preset length of the online inference window; Will The input is fed into the spatiotemporal graph neural network. First, the features of neighbor nodes under different relation types are weighted and aggregated based on the multi-relation graph aggregation layer to obtain the node spatial representation. Then, the node spatial representation under continuous time windows is temporally encoded based on the temporal coding layer to obtain the node spatiotemporal representation and the subgraph spatiotemporal representation. Based on the spatiotemporal representations of nodes and subgraphs, reconstruction and prediction tasks are performed separately to obtain node reconstruction residuals, node prediction residuals, subgraph reconstruction residuals, and subgraph prediction residuals. Node anomaly scores and subgraph anomaly scores are then calculated using the following formulas: ; , , ; ; , , ; in, Indicates the node anomaly score. , , The preset weighting coefficients, Represents the node reconstruction residual. For node empty time representation, This represents the spacetime representation of node reconstruction. This indicates taking the L2 norm. Represents the node prediction residual. This represents the true spacetime representation of a node in the next time window. This represents the predicted spatiotemporal representation of a node in the next time window. This is an anomaly in the node structure. This is the adjacency vector of the nodes. This is the reconstructed adjacency vector of the nodes. This represents the subgraph anomaly score. Represents the subgraph reconstruction residual. For subgraphs, the empty time representation is used. Reconstruct the spacetime representation of the subgraph. This represents the subplot prediction residual. This represents the true spacetime representation of the subgraph in the next time window. This represents the spatial-temporal representation of the subgraph in the next time window. This is an anomaly in the subgraph structure. Let be the adjacency relation matrix of the subgraph. This is the reconstructed adjacency matrix of the subgraph; Based on node anomaly scores and subgraph anomaly scores, combined with quantile thresholds, candidate nodes and candidate subgraphs for anomalies within the current time window are determined. The explanatory contributions of candidate nodes and candidate subgraphs are calculated separately to obtain the node feature contribution, edge relationship contribution, and time window contribution. The top K explanatory items are selected according to the magnitude of the explanatory contribution, and the node-level Top-K explanation and the subgraph-level Top-K explanation are output. The node-level Top-K explanation includes at least the abnormal node identifier, abnormal feature item, explanatory contribution, and corresponding time window. The subgraph-level Top-K explanation includes at least the abnormal subgraph identifier, high-contribution edge, high-contribution path, associated node sequence, explanatory contribution, and corresponding time window.

[0014] Optionally, a pre-generated prior graph is used as a soft constraint to learn the causal structure under the soft constraint. Intervention verification is performed on the node-level Top-K interpretation and the subgraph-level Top-K interpretation to obtain the root cause chain, including: constructing a prior graph based on equipment connection relationships, material flow relationships, and control loop relationships in the P&ID (Piping and Instrumentation Diagram) and deviation relationships, causal relationships, and consequence relationships in the HAZOP (Hazard and Operability Study); mapping the node-level Top-K interpretation and the subgraph-level Top-K interpretation to candidate causal variables, and constructing an observation matrix based on the candidate causal variables; using the weight matrix corresponding to the prior graph as a soft constraint, based on NOTEARS (No TEARS from...)... The DAGs algorithm learns the causal structure between candidate causal variables in the observation matrix to obtain the target causal weight matrix. Based on the target causal weight matrix, candidate causal edges and their directions are determined, and the candidate causal edges are sorted according to causal weight, temporal sequence, and explanatory contribution to obtain candidate causal paths. Intervention verification is performed on the candidate causal variables in the candidate causal paths, and the risk change before and after the intervention is calculated. Candidate causal variables whose risk change meets the preset conditions and their corresponding candidate causal paths are determined as valid root cause paths. Root cause chains are generated based on the triggering events, affected node sequences, edge relationship types, time delays, risk changes, and confidence levels in all valid root cause paths.

[0015] Optionally, the uncertainty of the anomaly score, quantile threshold, and spatiotemporal graph neural network is fused to determine an adaptive threshold. Then, the industrial safety risk level is determined based on the anomaly score and the adaptive threshold, including: determining a comprehensive anomaly score based on node anomaly scores and subgraph anomaly scores; calculating model uncertainty based on the anomaly scores output by the spatiotemporal graph neural network under multiple inferences; calling the corresponding quantile threshold according to the equipment type, shift, and season corresponding to the current time window, and correcting the quantile threshold based on the model uncertainty to obtain an adaptive threshold; and determining the industrial safety risk level based on the comprehensive anomaly score and the adaptive threshold. Specifically, when the comprehensive anomaly score is less than the first-level adaptive threshold, the industrial safety risk level is determined to be risk-free; when the comprehensive anomaly score is not less than the first-level adaptive threshold and less than the second-level adaptive threshold, the industrial safety risk level is determined to be early; when the comprehensive anomaly score is not less than the second-level adaptive threshold and less than the third-level adaptive threshold, the industrial safety risk level is determined to be severe; and when the comprehensive anomaly score is not less than the third-level adaptive threshold, the industrial safety risk level is determined to be urgent.

[0016] This application proposes an industrial safety risk assessment method based on behavioral simulation and artificial intelligence. Addressing five types of deficiencies in industrial settings—heterogeneous multi-source strong coupling, scarcity of tail risks, difficulty in characterizing human dynamics, insufficient explanation and closed-loop mechanisms, and a balance between real-time and compliance—this method constructs an integrated system that is "advance risk assessment, explainable, executable, and auditable" through a combined technical solution of "behavioral simulation + spatiotemporal diagram + causal inference + strategy learning + domain-specific thresholds + edge-cloud collaboration." This application achieves the following improvements.

[0017] First, it covers tail risks and enhances cold start capability. This application uses multi-agent behavior simulation to enumerate disturbances such as valve position misalignment, equipment degradation, and human error violations, generating normal, abnormal, and extreme control scenarios. This addresses the problem of insufficient low-frequency, high-risk samples in real industrial settings, enabling the system to still have risk identification capabilities during the cold start phase when there are few historical accident samples.

[0018] Second, it enables the identification and location of coupled risks across equipment, personnel, and regions. This application constructs a spatiotemporal graph using measurement points, equipment, regions, personnel, and work tickets as nodes, and process, space, business, and behavioral relationships as multi-relationship edges. It then performs online inference through a spatiotemporal graph neural network to uniformly characterize the associated risks between people, machines, regions, and operations, and can output node-level and subgraph-level anomaly location results.

[0019] Third, the causal explanation is clear, reducing handling costs and enhancing audit evidence collection capabilities. This application uses the prior graph formed by P&ID and HAZOP as a soft constraint, combined with the NOTEARS algorithm and intervention verification, to generate a root cause chain, enabling risk assessment results to point to triggering events, propagation paths, and key causes, reducing ineffective investigations, and providing evidence for incident review and compliance audits.

[0020] Fourth, this application sets domain-specific adaptive thresholds in conjunction with uncertainty to suppress risk storms. It uses thresholds based on equipment type, work group, and seasonal maintenance, and integrates the uncertainty of a space-time graph neural network to dynamically adjust the thresholds, avoiding concentrated risk outbreaks caused by fixed thresholds and improving the stability of risk classification.

[0021] Fifth, achieve a closed-loop linkage between timely handling and approval. This application matches security strategies based on risk level and root cause chain. No risk and early risks are directly executed with the corresponding strategies, while serious and urgent risks are executed with control strategies after approval by operations and maintenance, forming a closed loop of identification, interpretation, approval, execution and confirmation.

[0022] Sixth, comprehensive data governance and unified standards reduce maintenance costs. This application achieves consistent data and feature standards for different data through multi-protocol access, NTP time synchronization, unified multi-scale windows, missing data detection, drift identification, and unified feature mapping, facilitating model training, threshold updates, and long-term operation and maintenance. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies of this application will be briefly introduced below. Obviously, the following drawings are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. The drawings described herein are only used to explain this application and are not intended to limit this application.

[0024] Figure 1 This is a flowchart of an industrial safety risk assessment method based on behavioral simulation and artificial intelligence provided in one embodiment of this application; Figure 2 This is a schematic diagram of data acquisition and governance provided in one embodiment of this application; Figure 3 This is a schematic diagram of multi-agent behavior simulation provided in one embodiment of this application; Figure 4 This is a schematic diagram of online inference of a space-time graph neural network provided in one embodiment of this application; Figure 5 This is a schematic diagram of causal structure learning and intervention verification provided in one embodiment of this application; Figure 6 This is a schematic diagram of the structure of an industrial safety risk assessment system based on behavioral simulation and artificial intelligence, provided in another embodiment of this application; Figure 7 This is a schematic diagram of the structure of an electronic device provided in another embodiment of this application. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the various embodiments of this application will be described in detail below with reference to the accompanying drawings. Those skilled in the art will understand that many technical details have been provided in the embodiments of this application to facilitate better understanding. However, the technical solutions claimed in this application can be implemented even without these technical details and various variations and modifications based on the following embodiments. The division of the following embodiments is for ease of description and should not constitute any limitation on the specific implementation of this application. The following embodiments can be combined with and referenced by each other without contradiction.

[0026] One embodiment of this application proposes an industrial safety risk assessment method based on behavioral simulation and artificial intelligence. The implementation details of the industrial safety risk assessment method based on behavioral simulation and artificial intelligence proposed in this embodiment are described in detail below. The following implementation details are provided for ease of understanding and are not necessary for implementing this solution.

[0027] The specific process of the industrial safety risk assessment method based on behavioral simulation and artificial intelligence proposed in this embodiment can be described as follows: Figure 1 As shown, it includes: Step 11: Acquire basic data from the industrial site, including time-series signals, video streams, and personnel trajectory data, through multi-protocol access. Based on the NTP protocol, unify the clock reference, perform multi-scale window unification on basic data with different sampling rates, and perform missing detection and drift identification to obtain the actual sampled data.

[0028] Specifically, the basis of industrial safety risk assessment is data from the industrial site. In this embodiment, basic data from the industrial site, including time-series signals, video streams and personnel trajectory data, is obtained through multi-protocol (communication protocol) access. Then, based on the NTP protocol, a unified clock reference is established, and the basic data with different sampling rates are unified by multi-scale windows. Missing data detection and drift identification are also performed to obtain the actual sampled data.

[0029] Data collection and data governance processes, such as Figure 2This embodiment acquires basic data from DCS, PLC, SCADA, sensors, cameras, access control devices, and personnel positioning devices configured in the industrial field using protocols including OPC UA, Modbus, MQTT, and GB28181 (communication protocols). It also configures data source identifiers, sampling timestamps, and quality identifiers for the acquired basic data. The sampling rates for basic data from different sources vary.

[0030] For the basic data collected at different sampling rates, this embodiment performs unified calibration of the sampling timestamps of the basic data based on the NTP protocol to obtain calibration timestamps under a unified clock reference, and aligns the basic data to a unified time axis according to a preset time granularity.

[0031] Next, the basic data on the same time axis is processed by multi-scale windowing according to at least two different time scales (such as 1s, 10s and 60s), and the basic data located in the same sliding window are processed by including aggregation, interpolation and resampling to obtain multi-scale window data.

[0032] Subsequently, based on the quality identifier, zero-order hold duration and threshold, run length and lag ratio in the multi-scale window data, missing data detection and lag detection are performed, and the detected missing segments are interpolated by neighborhood interpolation or model interpolation to obtain the interpolated multi-scale window data.

[0033] Finally, based on the statistical distribution changes within the sliding window at different time scales, drift identification is performed on the interpolated multi-scale window data, and the drift identifier is written into the corresponding multi-scale window data to obtain the actual collected data carrying quality and drift identifiers.

[0034] exist Figure 2In this system, the data acquisition domain supports protocols such as OPC UA, Modbus, MQTT, and GB28181, aggregating time series, events, video vectors, and personnel trajectories at the edge gateway. The edge computing domain is responsible for lightweight feature extraction (FFT / wavelet, pose keypoints), caching, and de-identification using industrial PCs / embedded GPUs. The data bus and storage domain includes a message bus (Kafka / Flink), a TSDB (time series database) for storing process / equipment signals, Lakehouse for storing events and metadata, a graph database for storing spatiotemporal graphs, and a vector library for storing features. The AI ​​simulation computing domain, equipped with a simulation engine (ABM + mechanistic model / proxy model), is responsible for STGNN (spatiotemporal neural network) training / inference, causal structure learning, and reinforcement strategy learning. The application and linkage domain is responsible for risk scoring and tiered risk assessment, SOP (standard operating procedure) recommendations, work order / electronic fence / interlocking interfaces (priority recommendation + approval writing), and audit trails. The security and compliance domain is responsible for OT / IT partitioning, one-way gateways, access tiers, end-to-end encryption, and traceability.

[0035] Step 12: Through multi-agent behavior simulation, enumerate disturbances including valve position misalignment, equipment degradation, and human error to construct comparative scenario data and supplement the scarce tail samples.

[0036] Specifically, while collecting and managing data, this embodiment also needs to enumerate disturbances, including valve position misalignment, equipment degradation, and human error, through multi-agent behavior simulation to construct control scenario data and supplement the scarce tail samples. The constructed control scenario data includes at least normal scenario data, abnormal scenario data, and extreme scenario data.

[0037] Please refer to the multi-agent behavior simulation process. Figure 3 First, a multi-agent simulation model is constructed, including personnel, equipment, environment, and work tasks. State parameters, behavioral rules, and interaction constraints are configured for each agent. Simultaneously, a set of disturbance parameters is established based on the industrial process flow, equipment operating boundaries, and operating procedures. The disturbance parameters in this set include at least valve position misalignment parameters, equipment degradation parameters, and human error parameters.

[0038] Subsequently, based on the preset disturbance intensity, preset disturbance duration, preset disturbance occurrence order, and preset disturbance coupling relationship, single-factor enumeration and / or multi-factor combination enumeration are performed on the disturbance parameter set to generate simulation disturbance instances.

[0039] Next, the simulation disturbance instance is input into the multi-agent simulation model to drive different agents to perform state updates and interactive evolution under a unified simulation clock, thereby obtaining the corresponding simulation process data.

[0040] Finally, based on the disturbance type, disturbance intensity, and risk evolution results corresponding to the simulation process data of the simulated disturbance instances, the simulation process data is divided into normal scenarios, abnormal scenarios, and extreme scenarios. Scenario labels, disturbance labels, and risk labels are configured for each control scenario. The simulation process data carrying scenario labels, disturbance labels, and risk labels are used as control scenario data. Samples belonging to low-frequency, high-risk disturbance combinations in the control scenario data are identified as tail samples to supplement the scarce industrial safety risk samples in the historical data. This process successfully supplemented the scarce tail samples.

[0041] Step 13: Map the comparison scenario data, actual data, and historical data to a unified feature space, and perform domain adaptive training according to the quantile thresholds of equipment type, work group, and seasonal maintenance.

[0042] Specifically, after constructing the reference scenario data, the reference scenario data, actual data, and historical data can be mapped to a unified feature space, and domain adaptive training can be performed according to the quantile thresholds of equipment type, work group, and seasonal maintenance.

[0043] The premise of unifying the feature space is to extract the required features from the comparison scenario data, actual data, and historical data. In this embodiment, window-level feature extraction is performed on the comparison scenario data, actual data, and historical data to obtain multiple initial feature vectors. Among them, the initial feature vectors include at least time-frequency domain features, personnel trajectory features, operation compliance features, and topological association features.

[0044] Next, following the unified feature encoding rules, the initial feature vector is normalized, missing mask encoded, quality identifier encoded, and scene label encoded to obtain the sample feature vector in the unified feature space.

[0045] Subsequently, all sample feature vectors were divided into domains according to equipment type, work group, and season to obtain multiple training domains. For each training domain, a quantile threshold was maintained based on the abnormal distribution statistics corresponding to the sample feature vectors within that domain. The quantile threshold is specifically composed of a first-level quantile threshold, a second-level quantile threshold, and a third-level quantile threshold, which corresponds to the subsequent four levels of industrial safety risk.

[0046] Finally, domain-adaptive training is performed based on the sample feature vectors. This reduces the feature distance between sample feature vectors with the same scene label and risk label in different training domains within a unified feature space, while increasing the feature distance between sample feature vectors with different scene labels and risk labels within the unified feature space, resulting in a domain-adaptive feature representation. The domain-adaptive feature representation, quantile threshold, and sample feature vectors serve as the basis for subsequent construction of the spatiotemporal map and determination of the adaptive threshold.

[0047] In one example, the first Each sample feature vector Represented as: ; ; in, Indicates the first 1 initial feature vector For feature mapping function, This indicates the comparison scene data. This indicates the actual collected data. This represents historical data collected. Represents the characteristics of time-series signals. This indicates the event characteristics corresponding to the video stream. Indicates the characteristics of personnel trajectories; Indicates the characteristics of operational compliance. Indicates topological association features, For missing mask, For quality labeling, The tags include scene tags, disturbance tags, and risk tags.

[0048] Record No. The device type, the first The class group, the The training domain for each season is , express: ; in, Indicates the device type. Indicates work group, Indicates the season.

[0049] First-order quantile threshold Second-level quantile threshold and third-order quantile threshold These can be expressed as formulas: ; ; ; in, It is a quantile function. Represents the training domain The set of training scores corresponding to the feature vectors of each sample within the set. , , For the preset quantiles, .

[0050] Step 14: Using measurement points, equipment, areas, personnel, and work tickets as nodes, and process relationships, spatial relationships, business relationships, and behavioral relationships as multi-relationship edges, construct a spatiotemporal graph, and input it into the spatiotemporal graph neural network for online inference, calculate the anomaly score, and output node-level Top-K explanations and subgraph-level Top-K explanations.

[0051] Specifically, after completing the domain adaptive training, a spatiotemporal graph can be constructed using measurement points, equipment, areas, personnel, and work tickets as nodes, and process relationships, spatial relationships, business relationships, and behavioral relationships as multi-relationship edges. The constructed spatiotemporal graph is then input into the spatiotemporal graph neural network for online inference, calculating anomaly scores, and simultaneously outputting node-level Top-K explanations and subgraph-level Top-K explanations.

[0052] refer to Figure 4 This embodiment constructs a spatiotemporal map for the current time window based on domain-adaptive feature representation and the correspondence between sample feature vectors and industrial sites. .in, , For the current time window, Represents a set of nodes. Denotes the set of edges. Represents a set of edge relation types. Represents a multi-relation adjacency matrix. This represents the node feature matrix.

[0053] Map measurement points, equipment, areas, personnel, and work permits as sets of nodes. The nodes in the diagram are mapped to a set of edge relationship types, representing technological, spatial, business, and behavioral relationships. Different types of edges, node feature matrices Automatic generation of multi-relation adjacency matrices based on domain-adaptive feature representation Dynamic updates are performed based on process connection relationships, spatial distance, operational relationships, and personnel behavior relationships.

[0054] Subsequently, following the preset online inference window, candidate subgraph sequences are extracted from the spatiotemporal spectra across multiple consecutive time windows. Represented as: ; in, This indicates the length of the preset online inference window.

[0055] Will The input is fed into a space-time graph neural network. First, the features of neighbor nodes under different relation types are weighted and aggregated based on a multi-relation graph aggregation layer to obtain the node spatial representation. Then, the node spatial representation under continuous time windows is temporally encoded based on a time coding layer to obtain the node space-time representation and the subgraph space-time representation.

[0056] Next, based on the node spatiotemporal representation and the subgraph spatiotemporal representation, reconstruction and prediction tasks are performed respectively to obtain node reconstruction residuals, node prediction residuals, subgraph reconstruction residuals, and subgraph prediction residuals. The node anomaly score and subgraph anomaly score are then calculated according to the following formulas: ; , , ; ; , , ; in, Indicates the node anomaly score. , , The preset weighting coefficients, Represents the node reconstruction residual. For node empty time representation, This represents the spacetime representation of node reconstruction. This indicates taking the L2 norm. Represents the node prediction residual. This represents the true spacetime representation of a node in the next time window. This represents the predicted spatiotemporal representation of a node in the next time window. This is an anomaly in the node structure. This is the adjacency vector of the nodes. This is the reconstructed adjacency vector of the nodes. This represents the subgraph anomaly score. Represents the subgraph reconstruction residual. For subgraphs, the empty time representation is used. Reconstruct the spacetime representation of the subgraph. This represents the subplot prediction residual. This represents the true spacetime representation of the subgraph in the next time window. This represents the spatial-temporal representation of the subgraph in the next time window. This is an anomaly in the subgraph structure. Let be the adjacency relation matrix of the subgraph. This is the reconstructed adjacency matrix of the subgraph.

[0057] Finally, based on the node anomaly score and subgraph anomaly score, combined with the quantile threshold, candidate anomaly nodes and candidate anomaly subgraphs within the current time window are determined. Then, the explanatory contribution of each candidate anomaly node and candidate anomaly subgraph is calculated to obtain the node feature contribution, edge relation contribution, and time window contribution. The top K explanatory items are selected according to their explanatory contributions, outputting node-level Top-K explanations and subgraph-level Top-K explanations. The node-level Top-K explanation includes at least the anomaly node identifier, anomaly feature items, explanatory contribution, and the corresponding time window. The subgraph-level Top-K explanation includes at least the anomaly subgraph identifier, high-contribution edges, high-contribution paths, associated node sequences, explanatory contribution, and the corresponding time window.

[0058] In one example, for abnormal candidate nodes The node features, associated edges, and time windows are perturbed or occluded respectively, and the change in anomaly scores before and after occlusion is calculated. If the anomaly score decreases more significantly after occluding a certain feature, edge, or time window, it indicates that the feature, edge, or time window contributes more to the anomaly score, and vice versa.

[0059] The contribution of node features can be expressed as: ; in, Indicates abnormal candidate nodes The The explanatory contribution of each feature term. Indicates abnormal candidate nodes Node anomaly scores before occlusion Indicates abnormal candidate nodes Concealing the first The node anomaly score is obtained by recalculating each feature item.

[0060] The contribution of a relation edge can be expressed as: ; in, Representing relation edges Explanation of contribution, Represents anomaly candidate subgraph Subgraph anomaly score before occlusion Represents anomaly candidate subgraph occlusion relationship edge The subgraph anomaly score was then recalculated.

[0061] The contribution of the time window can be expressed as: ; in, Indicates the first The explanatory contribution of each time window Represents anomaly candidate subgraph Concealing the first The subgraph anomaly score is obtained by recalculating the graph features corresponding to each time window.

[0062] Step 15: Using the pre-generated prior graph as a soft constraint, learn the causal structure under the soft constraint, and intervene to verify the node-level Top-K explanation and the subgraph-level Top-K explanation to obtain the root cause chain.

[0063] Specifically, after calculating the anomaly score and obtaining the node-level Top-K explanation and the subgraph-level Top-K explanation, the pre-generated prior graph needs to be used as a soft constraint. Under the soft constraint, the causal structure is learned, and the node-level Top-K explanation and the subgraph-level Top-K explanation are intervened and verified to obtain the root cause chain.

[0064] refer to Figure 5 The prior diagram is generated based on P&ID and HAZOP. First, the prior diagram is constructed based on the equipment connection relationships, material flow relationships, and control loop relationships in P&ID, as well as the deviation relationships, cause relationships, and consequence relationships in HAZOP.

[0065] Next, the node-level Top-K interpretation and the subgraph-level Top-K interpretation are mapped to candidate causal variables, and an observation matrix is ​​constructed based on the candidate causal variables.

[0066] Subsequently, using the weight matrix corresponding to the prior graph as a soft constraint, the causal structure between candidate causal variables in the observation matrix is ​​learned based on the NOTEARS algorithm to obtain the target causal weight matrix.

[0067] Next, based on the target causal weight matrix, candidate causal edges and their directions are determined, and the candidate causal edges are sorted according to causal weight, temporal order, and explanatory contribution to obtain candidate causal paths.

[0068] Finally, intervention verification is performed on the candidate causal variables in the candidate causal path, the change in risk before and after the intervention is calculated, and the candidate causal variables whose change in risk meets the preset conditions and their corresponding candidate causal paths are determined as valid root cause paths. Root cause chains are generated based on the triggering events, affected node sequences, edge relationship types, time delays, change in risk, and confidence levels in all valid root cause paths.

[0069] In one example, the target causal weight matrix The following optimization objectives were determined: ; in, Represents the observation matrix. Indicates based on The structural fitting loss, and The preset weighting coefficients, As a soft constraint, This indicates taking the L1 norm. This indicates taking the F-norm.

[0070] In one example, the change in risk is calculated using the following formula: ; in, Representing candidate causal variables The corresponding change in risk, Indicates the candidate causal variable The probability of risk after intervention. This indicates the probability of risk without intervention.

[0071] Step 16: The uncertainty of the anomaly score, quantile threshold and spatiotemporal graph neural network is fused to determine the adaptive threshold. Then, the industrial safety risk level is determined based on the anomaly score and the adaptive threshold, which is divided into four levels: no risk, early stage, severe and emergency.

[0072] Specifically, while generating the root cause chain, the uncertainty of the anomaly score, quantile threshold and spatiotemporal graph neural network are fused to determine the adaptive threshold. Then, the industrial safety risk level is determined based on the anomaly score and the adaptive threshold, which is divided into four levels: no risk, early stage, severe and emergency.

[0073] The classification of industrial safety risk levels is based on anomaly scores, quantile thresholds, and the uncertainty of a spatiotemporal graph neural network. First, a comprehensive anomaly score is determined based on node anomaly scores and subgraph anomaly scores. Then, model uncertainty is calculated based on the anomaly scores output by the spatiotemporal graph neural network after multiple inferences. Next, according to the equipment type, work group, and season corresponding to the current time window, the corresponding quantile threshold is applied and adjusted based on the model uncertainty to obtain an adaptive threshold. Finally, the industrial safety risk level is determined based on the comprehensive anomaly score and the adaptive threshold. Specifically, when the comprehensive anomaly score is less than the first-level adaptive threshold, the industrial safety risk level is determined to be no risk; when the comprehensive anomaly score is not less than the first-level adaptive threshold and less than the second-level adaptive threshold, the industrial safety risk level is determined to be early; when the comprehensive anomaly score is not less than the second-level adaptive threshold and less than the third-level adaptive threshold, the industrial safety risk level is determined to be severe; and when the comprehensive anomaly score is not less than the third-level adaptive threshold, the industrial safety risk level is determined to be urgent.

[0074] In one example, industrial safety risk levels are classified using the following formula: ; ; , , ; ; in, , , , These are categorized as no-risk, early-stage, severe, and emergency levels. To summarize abnormal scores, The preset comprehensive weight, , , These represent the first-level adaptive threshold, the second-level adaptive threshold, and the third-level adaptive threshold, respectively. , , These are the preset correction parameters. Indicating model uncertainty, The spatiotemporal graph neural network represents the first... The comprehensive abnormal score under the next inference Representing a space-time graph neural network The average score of the comprehensive abnormality in the reasoning process. Indicates the total number of inferences.

[0075] Step 17: Output the industrial safety risk level and root cause chain, and execute the corresponding safety policy. For no-risk and early-stage levels, execute the corresponding safety policy directly. For severe and emergency levels, execute the corresponding safety policy after the operation and maintenance approval.

[0076] Specifically, after obtaining the industrial security risk level, the industrial security risk level and root cause chain can be output, and the corresponding security policy can be executed. The corresponding security policy is executed directly for the no-risk level and the early level, while the corresponding security policy is executed after the operation and maintenance approval is obtained for the severe level and the emergency level.

[0077] In one example, this embodiment matches corresponding security policies from a preset security policy library based on the industrial safety risk level and the root cause chain. These security policies include at least monitoring and maintenance policies, alarm notification policies, inspection work order policies, electronic fence policies, equipment load reduction policies, redundant equipment switching policies, bypass valve control policies, and shutdown policies. When the industrial safety risk level is no risk, the monitoring and maintenance policy or the recording policy is executed directly, and the current industrial site status, comprehensive anomaly score, and root cause chain are written to the audit log. When the industrial safety risk level is early level, the corresponding alarm notification policy, inspection work order policy, or electronic fence policy is executed directly, and the triggering event, affected node sequence, and time delay in the root cause chain are used as policy execution parameters. When the industrial safety risk level is severe or urgent, an operation and maintenance approval request is generated. The operation and maintenance approval request includes at least the industrial safety risk level, root cause chain, security policy to be executed, policy impact scope, and evidence snapshot. After receiving the operation and maintenance approval information, the security policy corresponding to the severe or urgent level is executed, and a policy execution receipt is generated after execution. Finally, the strategy execution receipts, approval information, execution time, execution results, and risk changes are written back and stored for subsequent auditing, incremental learning, and quantile threshold updates.

[0078] This embodiment proposes an industrial safety risk assessment method based on behavioral simulation and artificial intelligence. Addressing five types of deficiencies in industrial settings—heterogeneous multi-source strong coupling, scarcity of tail risks, difficulty in characterizing human dynamics, insufficient explanation and closed-loop mechanisms, and a balance between real-time and compliance—this method constructs an integrated system that is "advance risk assessment, explainable, executable, and auditable" through a combined technical solution of "behavioral simulation + spatiotemporal graph + causal inference + strategy learning + domain-specific thresholds + edge-cloud collaboration." This application achieves the following improvements.

[0079] First, it covers tail risks and enhances cold start capability. This embodiment uses multi-agent behavior simulation to enumerate disturbances such as valve position misalignment, equipment degradation, and human error violations, generating normal, abnormal, and extreme control scenarios. This addresses the problem of insufficient low-frequency, high-risk samples in real industrial settings, enabling the system to still have risk identification capabilities during the cold start phase when there are few historical accident samples.

[0080] Second, it enables the identification and location of coupled risks across equipment, personnel, and regions. This embodiment uses measurement points, equipment, regions, personnel, and work tickets as nodes, and constructs a spatiotemporal graph with process, space, business, and behavioral relationships as multiple relational edges. It then performs online inference through a spatiotemporal graph neural network to uniformly characterize the associated risks between people, machines, regions, and operations, and can output node-level and subgraph-level anomaly location results.

[0081] Third, the causal explanation is clear, reducing handling costs and enhancing audit evidence collection capabilities. This embodiment uses the prior graph formed by P&ID and HAZOP as a soft constraint, combined with the NOTEARS algorithm and intervention verification, to generate a root cause chain, enabling risk assessment results to point to triggering events, propagation paths, and key causes, reducing ineffective investigations, and providing evidence for incident review and compliance audits.

[0082] Fourth, a combination of domain-specific adaptive thresholds and uncertainty is used to suppress risk storms. This embodiment sets thresholds based on equipment type, work group, and seasonal maintenance, and integrates the uncertainty of a space-time graph neural network to dynamically adjust the thresholds, avoiding concentrated risk outbreaks caused by fixed thresholds and improving the stability of risk classification.

[0083] Fifth, it achieves a closed-loop linkage between timely handling and approval. This embodiment matches security policies according to risk level and root cause chain. No risk and early risks directly execute the corresponding policies, while serious and urgent risks are subject to operation and maintenance approval before control policies are executed, forming a closed loop of identification, interpretation, approval, execution and confirmation.

[0084] Sixth, comprehensive data governance and unified standards reduce maintenance costs. This embodiment achieves consistent data and feature standards for different data through multi-protocol access, NTP time synchronization, unified multi-scale windows, missing data detection, drift identification, and unified feature mapping, facilitating model training, threshold updates, and long-term operation and maintenance.

[0085] The steps described above are merely for clarity in describing the technical solution. In actual implementation, they can be combined into one step, or certain steps can be broken down into multiple steps, as long as they involve the same logical relationship, they are all within the scope of protection of this application. Any insignificant modifications or designs added to the algorithm or process, as long as they do not change the core of the algorithm or process, are also within the scope of protection of this application.

[0086] Another embodiment of this application proposes an industrial safety risk assessment system based on behavioral simulation and artificial intelligence. The details of the industrial safety risk assessment system based on behavioral simulation and artificial intelligence proposed in this embodiment are described in detail below. The following content is only for the convenience of understanding and is not necessary for implementing this solution. Figure 6 This is a schematic diagram of the structure of an industrial safety risk assessment system based on behavioral simulation and artificial intelligence proposed in this embodiment, including: a data acquisition module 21, a data governance module 22, a behavioral simulation module 23, a domain adaptive training module 24, a space-time graph anomaly detection module 25, a causal inference module 26, a risk classification module 27, and a strategy execution and approval linkage module 28.

[0087] The data acquisition module 21 is used to acquire basic data from the industrial site, including time-series signals, video streams, and personnel trajectory data, through multi-protocol access.

[0088] The data governance module 22 is used to unify the clock reference based on the NTP protocol, perform multi-scale window unification on the basic data with different sampling rates, and perform missing detection and drift identification to obtain the actual sampled data.

[0089] The behavior simulation module 23 is used to enumerate disturbances including valve position misalignment, equipment degradation, and human-caused violations through multi-subject behavior simulation in order to construct control scenario data and supplement the scarce tail samples; the constructed control scenario data includes at least normal scenario data, abnormal scenario data, and extreme scenario data.

[0090] The domain adaptive training module 24 is used to map the comparison scenario data, actual data and historical actual data to a unified feature space, and perform domain adaptive training according to the quantile thresholds of equipment type, work group and seasonal maintenance.

[0091] The space-time graph anomaly detection module 25 is used to construct a space-time graph with measurement points, equipment, areas, personnel and work tickets as nodes and process relationships, spatial relationships, business relationships and behavioral relationships as multi-relationship edges. The graph is then input into the space-time graph neural network for online inference to calculate anomaly scores. At the same time, it outputs node-level Top-K explanations and subgraph-level Top-K explanations.

[0092] The causal inference module 26 is used to learn the causal structure under the soft constraints by using the pre-generated prior graph as a soft constraint, and to intervene and verify the node-level Top-K explanation and the subgraph-level Top-K explanation to obtain the root cause chain.

[0093] The risk grading module 27 is used to fuse the uncertainty of anomaly scores, quantile thresholds and spatiotemporal graph neural networks to determine an adaptive threshold, and then determine the industrial safety risk level based on the anomaly scores and the adaptive threshold; among which, the industrial safety risk level is divided into four levels: no risk, early stage, severe and emergency.

[0094] The strategy execution and approval linkage module 28 is used to output the industrial security risk level and root cause chain, and execute the corresponding security policy. Among them, the corresponding security policy is executed directly for the no-risk level and early level, while the corresponding security policy is executed after the operation and maintenance approval is approved.

[0095] It is worth noting that all modules involved in this embodiment are logical modules. In practical applications, a logical module can be a physical module, a part of a physical module, or an organic combination of multiple physical modules. Furthermore, to highlight the innovative aspects of this application, this embodiment does not introduce modules that are not closely related to solving the technical problems proposed in this application. However, this does not mean that other modules are absent from this embodiment.

[0096] It is not difficult to see that this embodiment is a system embodiment corresponding to the above method embodiments, and this embodiment can be implemented in conjunction with the above method embodiments. The relevant technical details and technical effects mentioned in the above method embodiments are still valid in this embodiment, and will not be repeated here to reduce repetition. Accordingly, the relevant technical details mentioned in this embodiment can also be applied to the above method embodiments.

[0097] Another embodiment of this application provides an electronic device, such as Figure 7 As shown, it includes a processor 31 and a memory 32. The memory 32 stores instructions that the processor 31 can execute. When the processor 31 is configured to execute the instructions, the electronic device can realize an industrial safety risk assessment method based on behavioral simulation and artificial intelligence as described in the above method embodiment.

[0098] The memory and processor are connected via a bus, which includes any number of interconnecting buses and bridges. The bus can connect various circuits of one or more processors and memories, as well as other circuits such as peripherals, voltage regulators, and power management circuits—all well-known in the art and therefore not described further herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be a single component or multiple components, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by the processor is transmitted over the wireless medium via an antenna, which also receives and transmits data to the processor.

[0099] The processor manages the bus and handles general processing, providing various functions, including but not limited to timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory, on the other hand, is used to store data used by the processor during operation.

[0100] Another embodiment of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, enables an industrial safety risk assessment method based on behavioral simulation and artificial intelligence as described in the above method embodiments.

[0101] That is, those skilled in the art will understand that all or part of the steps in the above method embodiments can be implemented by a program instructing related hardware. The program is stored in a storage medium and includes several instructions to cause a device (such as a microcontroller, chip, etc.) or processor to execute all or part of the steps of the method described in the method embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0102] It will be understood by those skilled in the art that the above embodiments are specific implementations of this application, and various changes in form and detail can be made in practical applications without departing from the spirit and scope of this application. For those skilled in the art, several improvements and modifications can be made without departing from the principles of this application, and these improvements and modifications are also considered to be within the scope of protection of this application.

Claims

1. An industrial safety risk assessment method based on behavioral simulation and artificial intelligence, characterized in that, The method includes: The system acquires basic data from the industrial site, including time-series signals, video streams, and personnel trajectory data, through multi-protocol access. Based on the NTP protocol, it unifies the clock reference, performs multi-scale window unification on basic data with different sampling rates, and performs missing detection and drift identification to obtain the actual sampled data. By using multi-agent behavior simulation, disturbances including valve position misalignment, equipment degradation, and human error are enumerated to construct control scenario data and fill in the scarce tail samples; the constructed control scenario data includes at least normal scenario data, abnormal scenario data, and extreme scenario data. The comparison scenario data, actual data, and historical data are mapped to a unified feature space, and domain adaptive training is performed according to the quantile thresholds of equipment type, work group, and seasonal maintenance. Using measurement points, equipment, areas, personnel, and work tickets as nodes, and process relationships, spatial relationships, business relationships, and behavioral relationships as multi-relationship edges, a spatiotemporal graph is constructed and input into a spatiotemporal graph neural network for online inference to calculate anomaly scores. At the same time, node-level Top-K explanations and subgraph-level Top-K explanations are output. The pre-generated prior graph is used as a soft constraint. Under the soft constraint, the causal structure is learned, and the node-level Top-K explanation and the subgraph-level Top-K explanation are verified by intervention to obtain the root cause chain. The uncertainty of anomaly scores, quantile thresholds, and spatiotemporal graph neural networks is fused to determine an adaptive threshold. Then, the industrial safety risk level is determined based on the anomaly scores and the adaptive threshold. The industrial safety risk level is divided into four levels: no risk, early stage, severe, and emergency. Output the industrial safety risk level and root cause chain, and execute the corresponding security policy; among them, the corresponding security policy is executed directly for the no-risk level and the early level, while the corresponding security policy is executed after the operation and maintenance approval is obtained.

2. The industrial safety risk assessment method based on behavioral simulation and artificial intelligence as described in claim 1, characterized in that, Basic data from the industrial site, including time-series signals, video streams, and personnel trajectory data, is acquired through multi-protocol access. A unified clock reference is established based on the NTP protocol. Multi-scale windowing is applied to basic data with different sampling rates, and missing data detection and drift identification are performed to obtain the actual collected data, including: Basic data is acquired from DCS, PLC, SCADA, sensors, cameras, access control devices, and personnel positioning devices in the industrial field through protocols including OPC UA, Modbus, MQTT, and GB28181. Data source identifiers, sampling timestamps, and quality identifiers are configured for the basic data. The sampling rate of basic data from different sources varies. Based on the NTP protocol, the sampling timestamps of the basic data are uniformly calibrated to obtain the calibration timestamps under a unified clock reference, and the basic data is aligned to a unified time axis according to a preset time granularity. Multi-scale windowing is performed on the basic data on the same time axis using sliding windows with at least two different time scales, and the basic data within the same sliding window are aggregated, interpolated, and resampled to obtain multi-scale window data. Based on the quality identifier, zero-order hold duration and threshold, run length and lag ratio in the multi-scale window data, missing data detection and lag detection are performed, and the detected missing segments are interpolated by neighborhood interpolation or model interpolation to obtain the interpolated multi-scale window data. Based on the statistical distribution changes within sliding windows at different time scales, drift identification is performed on the interpolated multi-scale window data, and the drift identifier is written into the corresponding multi-scale window data to obtain the actual collected data carrying quality and drift identifiers.

3. The industrial safety risk assessment method based on behavioral simulation and artificial intelligence as described in claim 2, characterized in that, By employing multi-agent behavior simulation, disturbances including valve position misalignment, equipment degradation, and human error are enumerated to construct comparative scenario data and supplement scarce tail samples, including: Construct a multi-agent simulation model that includes personnel, equipment, environment, and work tasks, and configure state parameters, behavior rules, and interaction constraints for different agents; Based on the process flow, equipment operating boundaries and operating procedures of the industrial site, a set of disturbance parameters is established. The set of disturbance parameters includes at least valve position misalignment parameters, equipment degradation parameters and human error parameters. Based on the preset disturbance intensity, disturbance duration, disturbance occurrence sequence and disturbance coupling relationship, single-factor enumeration and / or multi-factor combination enumeration are performed on the disturbance parameter set to generate simulated disturbance instances. The simulation disturbance instance is input into the multi-agent simulation model, driving different agents to perform state updates and interactive evolution under a unified simulation clock, and obtaining the corresponding simulation process data; Based on the disturbance type, disturbance intensity, and risk evolution results corresponding to the simulation process data of the simulation disturbance instances, the simulation process data is divided into normal scenarios, abnormal scenarios, and extreme scenarios, and scenario labels, disturbance labels, and risk labels are configured for each control scenario; Simulation process data carrying scenario labels, disturbance labels, and risk labels are used as reference scenario data. Samples belonging to low-frequency, high-risk disturbance combinations in the reference scenario data are identified as tail samples to supplement the scarce industrial safety risk samples in historical data, i.e., to supplement the scarce tail samples.

4. The industrial safety risk assessment method based on behavioral simulation and artificial intelligence as described in claim 3, characterized in that, The comparison scenario data, actual collected data, and historical actual collected data are mapped to a unified feature space. Domain adaptive training is then performed according to the quantile thresholds for equipment type, work group, and seasonal maintenance, including: Window-level feature extraction was performed on the comparison scenario data, actual data, and historical actual data to obtain multiple initial feature vectors; among them, the initial feature vectors include at least time-frequency domain features, personnel trajectory features, operation compliance features, and topological association features; According to the unified feature encoding rules, the initial feature vector is normalized, missing mask encoded, quality identifier encoded and scene label encoded to obtain the sample feature vector in the unified feature space. All sample feature vectors are divided into domains according to equipment type, work group and season to obtain multiple training domains. For each training domain, a quantile threshold is maintained based on the abnormal distribution statistics corresponding to the sample feature vectors in that domain. The quantile thresholds include the first-level quantile threshold, the second-level quantile threshold and the third-level quantile threshold. Domain adaptive training based on sample feature vectors reduces the feature distance between sample feature vectors with the same scene label and the same risk label in different training domains within a unified feature space, while increasing the feature distance between sample feature vectors with different scene labels and different risk labels within the unified feature space, thus obtaining a domain adaptive feature representation. Among these, the domain adaptive feature representation, quantile threshold, and sample feature vectors are the basis for subsequent construction of spatiotemporal maps and determination of adaptive thresholds. Among them, the Each sample feature vector Represented as: ； ； in, Indicates the first 1 initial feature vector For feature mapping function, This indicates the comparison scene data. This indicates the actual collected data. This represents historical data collected. Represents the characteristics of time-series signals. This indicates the event characteristics corresponding to the video stream. Indicates the characteristics of personnel trajectories; Indicates the characteristics of operational compliance. Indicates topological association features, For missing mask, For quality labeling, The tags include scene tags, disturbance tags, and risk tags; Record No. The device type, the first The class group, the The training domain for each season is , express: ； in, Indicates the device type. Indicates work group, Indicates the season; First-order quantile threshold Second-level quantile threshold and third-order quantile threshold These can be expressed as formulas: ； ； ； in, It is a quantile function. Represents the training domain The set of training scores corresponding to the feature vectors of each sample within the set. , , For the preset quantiles, .

5. The industrial safety risk assessment method based on behavioral simulation and artificial intelligence as described in claim 4, characterized in that, Using measurement points, equipment, areas, personnel, and work permits as nodes, and process relationships, spatial relationships, business relationships, and behavioral relationships as multi-relationship edges, a spatiotemporal graph is constructed, including: Based on domain-adaptive feature representation and the correspondence between sample feature vectors and industrial sites, a spatiotemporal map is constructed for the current time window. ;in, , For the current time window, Represents a set of nodes. Denotes the set of edges. Represents a set of edge relation types. Represents a multi-relation adjacency matrix. Represents the node feature matrix; Map measurement points, equipment, areas, personnel, and work permits as sets of nodes. The nodes in the diagram are mapped to a set of edge relationship types, representing technological, spatial, business, and behavioral relationships. Different types of edges, node feature matrices Automatic generation of multi-relation adjacency matrices based on domain-adaptive feature representation Dynamic updates are based on process connection relationships, spatial distance, operational relationships, and personnel behavior relationships; The spatiotemporal graph is input into a spatiotemporal graph neural network for online inference, anomaly scores are calculated, and node-level Top-K explanations and subgraph-level Top-K explanations are output, including: According to the preset online inference window, candidate subgraph sequences are extracted from the spatiotemporal spectra under multiple consecutive time windows. Represented as: ； in, Indicates the preset length of the online inference window; Will The input is fed into the spatiotemporal graph neural network. First, the features of neighbor nodes under different relation types are weighted and aggregated based on the multi-relation graph aggregation layer to obtain the node spatial representation. Then, the node spatial representation under continuous time windows is temporally encoded based on the temporal coding layer to obtain the node spatiotemporal representation and the subgraph spatiotemporal representation. Based on the spatiotemporal representations of nodes and subgraphs, reconstruction and prediction tasks are performed separately to obtain node reconstruction residuals, node prediction residuals, subgraph reconstruction residuals, and subgraph prediction residuals. Node anomaly scores and subgraph anomaly scores are then calculated using the following formulas: ； ， ， ； ； ， ， ； in, Indicates the node anomaly score. , , The preset weighting coefficients, Represents the node reconstruction residual. For node empty time representation, This represents the spacetime representation of node reconstruction. This indicates taking the L2 norm. Represents the node prediction residual. This represents the true spacetime representation of a node in the next time window. This represents the predicted spatiotemporal representation of a node in the next time window. This is an anomaly in the node structure. This is the adjacency vector of the nodes. This is the reconstructed adjacency vector of the nodes. This represents the subgraph anomaly score. Represents the subgraph reconstruction residual. For subgraphs, the empty time representation is used. Reconstruct the spacetime representation of the subgraph. This represents the subplot prediction residual. This represents the true spacetime representation of the subgraph in the next time window. This represents the spatial-temporal representation of the subgraph in the next time window. This is an anomaly in the subgraph structure. Let be the adjacency relation matrix of the subgraph. This is the reconstructed adjacency matrix of the subgraph; Based on node anomaly scores and subgraph anomaly scores, combined with quantile thresholds, candidate nodes and candidate subgraphs for anomalies within the current time window are determined. The explanatory contributions of candidate nodes and candidate subgraphs are calculated separately to obtain the node feature contribution, edge relationship contribution, and time window contribution. The top K explanatory items are selected according to the magnitude of the explanatory contribution, and the node-level Top-K explanation and the subgraph-level Top-K explanation are output. The node-level Top-K explanation includes at least the abnormal node identifier, abnormal feature item, explanatory contribution, and corresponding time window. The subgraph-level Top-K explanation includes at least the abnormal subgraph identifier, high-contribution edge, high-contribution path, associated node sequence, explanatory contribution, and corresponding time window.

6. The industrial safety risk assessment method based on behavioral simulation and artificial intelligence as described in claim 5, characterized in that, Using a pre-generated prior graph as a soft constraint, the causal structure is learned under this constraint. Intervention verification is performed on both node-level Top-K explanations and subgraph-level Top-K explanations to obtain the root cause chain, including: Based on the equipment connection relationships, material flow relationships, and control loop relationships in P&ID, and the deviation relationships, cause relationships, and consequence relationships in HAZOP, a priori diagram is constructed. Node-level Top-K interpretations and subgraph-level Top-K interpretations are mapped to candidate causal variables, and an observation matrix is ​​constructed based on the candidate causal variables; Using the weight matrix corresponding to the prior graph as a soft constraint, the causal structure between candidate causal variables in the observation matrix is ​​learned based on the NOTEARS algorithm to obtain the target causal weight matrix. Based on the target causal weight matrix, candidate causal edges and their directions are determined, and the candidate causal edges are sorted according to causal weight, temporal order, and explanatory contribution to obtain candidate causal paths; Intervention verification is performed on candidate causal variables in the candidate causal path, the change in risk before and after the intervention is calculated, and the candidate causal variables whose change in risk meets the preset conditions and their corresponding candidate causal paths are determined as valid root cause paths. Root cause chains are generated based on the triggering events, affected node sequences, edge relationship types, time delays, change in risk and confidence in all valid root cause paths.

7. The industrial safety risk assessment method based on behavioral simulation and artificial intelligence as described in claim 6, characterized in that, The uncertainty of outlier scores, quantile thresholds, and spatiotemporal graph neural networks is fused to determine an adaptive threshold. Then, based on the outlier scores and the adaptive threshold, the industrial safety risk level is determined, including: Determine the comprehensive anomaly score based on node anomaly scores and subgraph anomaly scores; The uncertainty of the model is calculated based on the anomaly score output by the spatiotemporal graph neural network under multiple inferences. Based on the equipment type, work group, and season corresponding to the current time window, the corresponding quantile threshold is called, and the quantile threshold is corrected based on the model uncertainty to obtain an adaptive threshold; The industrial safety risk level is determined based on the comprehensive anomaly score and adaptive threshold. Specifically, when the comprehensive anomaly score is less than the first-level adaptive threshold, the industrial safety risk level is determined to be no risk; when the comprehensive anomaly score is not less than the first-level adaptive threshold and is less than the second-level adaptive threshold, the industrial safety risk level is determined to be early; when the comprehensive anomaly score is not less than the second-level adaptive threshold and is less than the third-level adaptive threshold, the industrial safety risk level is determined to be severe; and when the comprehensive anomaly score is not less than the third-level adaptive threshold, the industrial safety risk level is determined to be urgent.

8. An industrial safety risk assessment system based on behavioral simulation and artificial intelligence, characterized in that, The system includes: The data acquisition module is used to acquire basic data from the industrial site, including time-series signals, video streams, and personnel trajectory data, through multi-protocol access. The data governance module is used to unify the clock reference based on the NTP protocol, perform multi-scale window unification on basic data with different sampling rates, and perform missing data detection and drift identification to obtain the actual sampled data. The behavior simulation module is used to enumerate disturbances, including valve position misalignment, equipment degradation, and human error, through multi-agent behavior simulation to construct control scenario data and fill in the scarce tail samples. The constructed control scenario data includes at least normal scenario data, abnormal scenario data, and extreme scenario data. The domain adaptive training module is used to map the comparison scenario data, actual data and historical actual data to a unified feature space, and perform domain adaptive training according to the quantile thresholds of equipment type, work group and seasonal maintenance. The space-time graph anomaly detection module is used to construct a space-time graph with measurement points, equipment, areas, personnel and work tickets as nodes and process relationships, spatial relationships, business relationships and behavioral relationships as multi-relationship edges. The graph is then input into the space-time graph neural network for online inference to calculate anomaly scores. At the same time, it outputs node-level Top-K explanations and subgraph-level Top-K explanations. The causal inference module is used to learn the causal structure under the soft constraints by using the pre-generated prior graph as a soft constraint, and to intervene and verify the node-level Top-K explanation and the subgraph-level Top-K explanation to obtain the root cause chain. The risk grading module is used to fuse the uncertainty of anomaly scores, quantile thresholds, and spatiotemporal graph neural networks to determine an adaptive threshold, and then determine the industrial safety risk level based on the anomaly scores and the adaptive threshold; among which, the industrial safety risk level is divided into four levels: no risk, early stage, severe, and emergency. The strategy execution and approval linkage module is used to output the industrial security risk level and root cause chain, and execute the corresponding security policy. Among them, the corresponding security policy is executed directly for the no-risk level and early level, while the corresponding security policy is executed after the operation and maintenance approval is approved.

9. An electronic device, characterized in that, include: A processor and a memory, wherein the memory stores instructions that the processor can execute, and the processor is configured to, when executing the instructions, enable the electronic device to implement an industrial safety risk assessment method based on behavioral simulation and artificial intelligence as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it can implement an industrial safety risk assessment method based on behavioral simulation and artificial intelligence as described in any one of claims 1 to 7.

Images