Chemical industrial park safety risk assessment method based on agent model

Abstract

The invention relates to the technical field of computer application, and particularly discloses a chemical industry park safety risk assessment method based on an agent model. The method comprises the following steps: constructing a multi-agent collaborative perception framework, and fusing multi-source heterogeneous data of equipment, environment, materials and personnel to form a park panoramic view under a unified space-time reference; establishing an agent behavior model library embedded with the chemical process mechanism and the safety interlocking logic; an intelligent agent decision strategy is optimized through deep reinforcement learning, and multi-target training is carried out by taking a safety index as a reward function; executing dynamic risk deduction and accident chain simulation in a digital twin environment, and quantifying risk probability and consequences in combination with Monte Carlo sampling; and finally, a visual thermodynamic diagram and an interpretable risk assessment report are generated, weak links are identified, and intervention suggestions are provided. According to the technical scheme, dynamic, accurate and prospective evaluation of the safety risk of the chemical industry park can be realized, and the early warning timeliness and the evaluation reliability are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of computer application technology, specifically relating to a method for safety risk assessment of chemical industrial parks based on an intelligent agent model. Background Technology

[0002] With the deep integration of artificial intelligence and industrial safety, safety risk assessment in chemical industrial parks is gradually evolving from static rule-driven to dynamic intelligent decision-making. Chemical production processes are characterized by high temperature and pressure, flammability and explosiveness, multiphase coupling, and chain reactions, making their safety status susceptible to the interplay of multiple factors, including equipment aging, operational deviations, environmental disturbances, and emergency response capabilities. Traditional risk assessment methods primarily rely on manual or semi-automated processes such as HAZOP (Hazard and Operability Analysis), LOPA (Layer of Protection Analysis), or Quantitative Risk Assessment (QRA). These methods typically rely on offline calculations based on historical accident databases and preset failure modes, making it difficult to collaboratively perceive and proactively extrapolate the real-time operational status, dynamic personnel behavior, and sudden disturbances within the park. This results in delayed risk warnings, coarse-grained assessments, and a lack of scenario-adaptive capabilities.

[0003] Among these, safety risk assessment methods based on intelligent agent models have become a research hotspot in recent years. This method constructs virtual intelligent agents with perception, decision-making, and interaction capabilities to simulate the evolutionary behavior of personnel, equipment, materials, and the environment under complex working conditions, thereby reproducing potential accident chains and quantifying risk levels in the digital space. Ideally, such models should be able to integrate heterogeneous sensing data from DCS system parameters, video surveillance streams, gas leak sensor readings, and personnel location information, driving autonomous reasoning of the intelligent agent within a unified spatiotemporal framework and dynamically updating risk heatmaps and key vulnerability identification results. However, the application of existing intelligent agent modeling techniques in the field of chemical safety still faces significant bottlenecks.

[0004] In existing technologies, most agent models employ fixed behavioral rule bases or shallow reinforcement learning strategies, lacking embedded representations of the deep mechanisms of chemical processes, resulting in a semantic gap between simulated behavior and physical reality. Furthermore, multi-agent systems at the park level need to handle high-dimensional heterogeneous state spaces, but existing architectures generally lack standardized state coding mechanisms for chemical unit operations, making it difficult to accurately model risk transmission paths between different devices. In addition, current methods largely rely on post-event backtracking training, failing to achieve effective generalization and extrapolation of rare but high-consequence events under limited sample conditions, and risk outputs are typically single probability values, lacking explicit decoupling and visual representation of uncertainty sources. Therefore, in the highly complex, strongly coupled, and low-fault-tolerant environment of chemical industrial parks, there is an urgent need for an agent-driven safety risk assessment method that can deeply integrate domain knowledge, support real-time dynamic extrapolation, and possess interpretable risk tracing capabilities. Summary of the Invention

[0005] This invention relates to a method for safety risk assessment of chemical industrial parks based on an intelligent agent model, which includes the following specific steps: Step 1: Construct a multi-agent collaborative perception framework for the chemical industrial park. Through distributed control systems, video surveillance networks, gas concentration sensor arrays, and personnel positioning beacons deployed within the park, real-time data collection is conducted on equipment operating parameters, environmental status information, material flow data, and personnel behavior trajectories. This forms a multi-source heterogeneous perception data stream. Furthermore, a data fusion mechanism based on timestamp alignment and spatial coordinate registration is adopted to establish a panoramic view of the park's operational status under a unified spatiotemporal reference. Step 2: Design an intelligent agent behavior model library for chemical process mechanisms. Establish corresponding physicochemical process mathematical models for distillation columns, reactors, storage tanks, and pipeline systems. The models embed mass conservation, energy conservation, and momentum conservation equations and couple reaction kinetic parameters, heat and mass transfer coefficients, and safety interlock logic, so that the intelligent agent can simulate output behavior that conforms to actual physical laws based on the current input state. Step 3: Implement agent decision-making strategy optimization based on deep reinforcement learning, construct a multi-objective optimization problem with chemical safety indicators as reward functions, including the degree of deviation from the safe operation window, the potential leakage risk level and the emergency response time, and iteratively train the agent strategy in the simulation environment through the near-end policy optimization algorithm; Step 4: Perform dynamic risk simulation and accident chain simulation, drive multi-agent concurrent interaction in a digital twin environment, simulate the complete evolution process from the initial disturbance event to the spread of fire, explosion or toxic substances, and quantify the probability of occurrence and severity of consequences of different paths based on the Monte Carlo sampling method. Step 5: Generate an interpretable risk assessment report and a visual heat map. Calculate the risk level of each area in the park based on the comprehensive simulation results, identify key weaknesses and risk transmission paths, and dynamically display the risk distribution in the form of a heat map. At the same time, provide risk contribution analysis based on Shapley and interpretation.

[0006] Furthermore, the parameters collected by the distributed control system in step 1 include temperature, pressure, flow rate, and liquid level, with a sampling frequency of not less than 1 Hz; the gas concentration sensor array covers hydrogen sulfide, carbon monoxide, methane, and volatile organic compounds.

[0007] Furthermore, in step 2, the mathematical model of the distillation column includes the theoretical number of trays, reflux ratio, and feed composition variables; the reactor model includes the reaction rate constant, activation energy, and heat load parameters; the storage tank model includes the volatilization rate and pressure relief valve setpoint; and the pipeline system model includes fluid velocity, pressure drop, and corrosion rate.

[0008] Furthermore, in step 3, the deep reinforcement learning adopts an actor-critic architecture. Both the actor network and the critic network are fully connected neural networks with three hidden layers, and the number of neurons in each layer is 256, 128 and 64, respectively. The activation function is a modified linear unit, the learning rate is set to 0.001 and the discount factor is 0.99.

[0009] Furthermore, in step 4, the time step of the dynamic risk simulation is set to 1 second, the number of Monte Carlo samplings is no less than 10,000, and the state trajectory of the agent, the triggered event sequence and the final accident consequences are recorded for each sampling, and the frequency and severity indicators of each event chain are statistically analyzed.

[0010] Furthermore, in step 5, the risk level is divided into four levels: low risk, medium risk, high risk, and extremely high risk, based on the product of probability and consequence; the heatmap uses a color gradient from green to red to represent risk from low to high; Shapley and the interpretation are based on cooperative game theory, quantifying the marginal contribution of equipment status, environmental conditions, and personnel operation to risk output.

[0011] Furthermore, it also includes a real-time early warning and intervention suggestion module. When the simulation identifies a high-risk path or an impending safety accident, the system automatically generates early warning information and provides intervention suggestions based on the decision-making strategy of the intelligent agent, such as adjusting process parameters, activating emergency facilities, or evacuating personnel.

[0012] Furthermore, the method integrates model updating and self-learning mechanisms, and continuously collects actual operating data and accident cases to update the agent model and decision-making strategy online or offline.

[0013] Furthermore, the data fusion mechanism in step 1 includes a data quality assessment and repair submodule. This submodule detects missing values, outliers, and consistency conflicts, and uses a Kalman filter-based prediction method to fill in missing data and an isolated forest algorithm to identify outliers.

[0014] Furthermore, the safety interlocking logic in step 2 is implemented in the form of a rule base, containing more than 200 predefined rules, covering overpressure protection, overtemperature shutdown, leakage alarm and emergency release safety scenarios. The rule triggering conditions are directly related to the state variables perceived by the intelligent agent.

[0015] The present invention has the following beneficial effects: By constructing a multi-agent model that deeply integrates the mechanism of chemical processes, and combining real-time sensing data and deep reinforcement learning optimization, it realizes dynamic, accurate and forward-looking assessment of safety risks in chemical industrial parks. It effectively overcomes the shortcomings of traditional methods in terms of real-time performance, adaptability and interpretability, significantly improves the timeliness of risk warning and the reliability of assessment results, and provides strong technical support for safety management and emergency decision-making in chemical industrial parks. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall technical solution architecture of the chemical industrial park safety risk assessment method based on the intelligent agent model proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the collaborative optimization of intelligent agent behavior model and deep reinforcement learning decision strategy for chemical process mechanism in this invention. Detailed Implementation

[0017] The features and exemplary embodiments of various aspects of the present invention will now be described in detail. In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] This invention proposes a multi-agent model that deeply integrates chemical process mechanisms and combines real-time sensing data with deep reinforcement learning optimization to achieve dynamic, accurate, and forward-looking assessment of safety risks in chemical industrial parks. It effectively overcomes the shortcomings of traditional methods in terms of real-time performance, adaptability, and interpretability, significantly improves the timeliness of risk warnings and the reliability of assessment results, and is applied to a chemical industrial park safety risk assessment method based on the agent model.

[0019] Reference Appendix Figure 1 The overall technical architecture of the chemical industrial park safety risk assessment method based on the intelligent agent model proposed in this invention includes a multi-source perception layer, an intelligent agent modeling layer, a decision optimization layer, a risk inference layer, and a visualization output layer. Information interaction and closed-loop feedback are achieved between these layers through a data bus under a unified spatiotemporal reference. In the above-mentioned chemical industrial park safety risk assessment method based on the intelligent agent model, step (1) constructs a multi-agent collaborative perception framework for the chemical industrial park. Through distributed control systems, video surveillance networks, gas concentration sensor arrays, and personnel positioning beacons deployed within the park, real-time data is collected on equipment operating parameters, environmental status information, material flow data, and personnel behavior trajectories, forming a multi-source heterogeneous perception data stream. A data fusion mechanism based on timestamp alignment and spatial coordinate registration is then used to establish a panoramic view of the park's operating status under a unified spatiotemporal reference.

[0020] Specifically, in step (1), the parameters collected by the distributed control system include temperature, pressure, flow rate, and liquid level, with a sampling frequency of not less than 1 Hz; the video monitoring network uses high-definition infrared cameras with a resolution of 1920 x 1080 pixels and a frame rate of 30 frames per second; the gas concentration sensor array covers hydrogen sulfide, carbon monoxide, methane, and volatile organic compounds, with a detection accuracy of parts per million; the personnel positioning beacon uses ultra-wideband technology, with a positioning accuracy better than 0.3 meters. All sensing devices are connected to the edge computing node through industrial Ethernet or 5G private network. The raw data stream is first compressed and standardized at the edge, and then transmitted to the central data fusion engine.

[0021] The data fusion mechanism includes a data quality assessment and repair submodule. This submodule detects missing values, outliers, and consistency conflicts, and uses a Kalman filter-based prediction method to fill in missing data, an isolated forest algorithm to identify outliers, and data consistency rules to resolve conflicts, ensuring the reliability of the data input to the intelligent agent model. Timestamp alignment employs a nanosecond-level hardware synchronization protocol, and spatial coordinate alignment is based on a high-precision 3D Geographic Information System (GIS) model of the park, mapping all sensor observations to a unified world coordinate system with coordinate errors controlled within ±0.1 meters. The final panoramic view of the park's operational status is stored in a structured time-series database, with each time step corresponding to a complete status snapshot, including equipment status vectors, environmental field distribution matrices, material flow topology diagrams, and personnel location lists.

[0022] In the above method, step (2) designs an intelligent agent behavior model library for chemical process mechanisms. For different types of chemical unit operations, including distillation towers, reactors, storage tanks and pipeline systems, corresponding physical and chemical process mathematical models are established. The models embed mass conservation, energy conservation and momentum conservation equations, and couple reaction kinetic parameters, heat and mass transfer coefficients and safety interlock logic, so that the intelligent agent can simulate output behavior that conforms to actual physical laws based on the current input state.

[0023] Specifically, in step (2), the mathematical model of the distillation column includes the theoretical plate number, reflux ratio, and feed composition variables; the reactor model includes the reaction rate constant, activation energy, and heat load parameters; the storage tank model includes the volatilization rate and pressure relief valve setpoint; and the piping system model includes fluid velocity, pressure drop, and corrosion rate. All models are calibrated using historical operating data, and the goodness-of-fit index is greater than or equal to 0.95. For example, for a typical continuous stirred tank reactor (CSTR), its material balance equation is:

[0024] in, The concentration of reactant A For feed flow rate, For reactor volume, The feed concentration is... Pre-exponential factor, For activation energy, The gas constant is The reaction temperature is represented by this equation, which, coupled with the energy balance equation, forms the core dynamic model of the reactor's intelligent system.

[0025] The safety interlocking logic is implemented in the form of a rule base, containing over 200 predefined rules covering various safety scenarios such as overpressure protection, overtemperature shutdown, leakage alarm, and emergency release. The rule triggering conditions are directly related to the state variables perceived by the agent, achieving seamless integration of the mechanism model and safety logic. For example, when the pressure inside the reactor exceeds a set threshold of 1.2 MPa and the temperature is above 150 degrees Celsius, the interlocking rule automatically triggers the cooling water valve to fully open and the heating power to be cut off. Each agent instance independently runs its mechanism model in the simulation environment and updates its internal state variables in real time based on external input states, outputting key process parameters including equipment outlet flow rate, component concentration, temperature, and pressure.

[0026] In the above method, step (3) implements agent decision-making strategy optimization based on deep reinforcement learning, constructs a multi-objective optimization problem with chemical safety indicators as reward functions, the safety indicators include the degree of deviation from the safe operation window, the potential leakage risk level and the emergency response time, and iteratively trains the agent strategy in the simulation environment through the near-end policy optimization algorithm so that it learns to take the optimal safe operation under complex disturbances.

[0027] Specifically, in step (3), deep reinforcement learning adopts an actor-critic architecture. The actor network is responsible for outputting the operation actions, and the critic network is responsible for evaluating the state value. The network structure is a fully connected neural network with three hidden layers, and the number of neurons in each layer is 256, 128, and 64, respectively. The activation function is a modified linear unit, the learning rate is set to 0.001, and the discount factor is 0.99. The state space is formed by feature extraction from the panoramic view generated in step (1), and has a dimension of 512. The action space includes discrete-continuous mixed operations such as adjusting valve opening, starting and stopping pumps, and switching backup lines, with a total of 64 selectable actions.

[0028] The reward function is defined as:

[0029] in, The normalized distance of the current state from the safe operating window. This is a potential leakage risk index calculated based on a gas diffusion model. The normalized value of the response time from the current state to the nearest emergency facility, with weighting coefficients. , , During training, the agent experiences millions of perturbation events (such as pump failure, cooling failure, and misoperation) in a digital twin environment. The policy network parameters are updated through a proximal policy optimization algorithm until the policy converges. The convergence criterion is that the average cumulative reward fluctuation is less than 0.01 over 1000 consecutive training rounds.

[0030] In the above method, step (4) performs dynamic risk simulation and accident chain simulation, drives multiple agents to interact concurrently in a digital twin environment, simulates the complete evolution process from initial disturbance events, such as equipment failure or operational errors, to possible accident consequences, such as fire, explosion or toxic substance spread, and quantifies the probability of occurrence and severity of consequences of different paths based on the Monte Carlo sampling method.

[0031] Specifically, in step (4), the time step of the dynamic risk simulation is set to 1 second, and the number of Monte Carlo samplings is no less than 10,000. Each sampling begins with a randomly selected initial disturbance event, such as the reactor agitator stopping, the distillation column reboiler leaking, or personnel accidentally entering a high-pressure area. During the simulation, all agents operate autonomously according to their mechanism models and optimized decision-making strategies, exchanging status information (such as material supply interruption, utility pressure drop) with each other through a message passing mechanism. The system monitors key safety indicators in real time, and once the accident judgment conditions are triggered (such as the concentration of combustible gas exceeding 20% ​​of the lower explosive limit, or the temperature rise rate exceeding 10 degrees Celsius per second), the complete trajectory of the accident chain is recorded.

[0032] For each sampling, the system records the agent's state trajectory, the triggered event sequence, and the final accident consequences, and calculates the frequency and severity index of each event chain. The severity index comprehensively considers the expected value of casualties, the estimated value of property damage, and the environmental impact index, and is calculated using a weighted summation method. Finally, the system generates a probability-consequence joint distribution matrix containing all possible accident paths for subsequent risk quantification analysis.

[0033] In the above method, step (5) generates an interpretable risk assessment report and a visual heat map. Based on the comprehensive deduction results, the risk level of each area in the park is calculated, key weak links and risk transmission paths are identified, and the risk distribution is dynamically displayed in the form of a heat map. At the same time, risk contribution analysis based on Shapley and interpretation is provided to indicate the influence weight of each factor on the overall risk.

[0034] Specifically, in step (5), the risk level is divided into four levels: low risk, medium risk, high risk, and extremely high risk, based on the product of probability and consequence. The corresponding risk value ranges are [0, 0.1), [0.1, 0.3), [0.3, 0.7), and [0.7, 1.0], respectively. The heat map uses a color gradient from green to red to represent risks from low to high, and is overlaid on the three-dimensional digital map of the park, with a refresh rate of once every 5 seconds. Shapley and the interpretation are based on cooperative game theory, quantifying the marginal contribution of each feature, including equipment status, environmental conditions, and personnel operation, to the risk output. For example, for a certain high-risk area, the system can indicate that "abnormal reactor temperature" contributes 45% of the risk value, "insufficient cooling water pressure" contributes 30%, and "operator failure to respond to alarm in a timely manner" contributes 25%.

[0035] In addition, it includes a real-time early warning and intervention suggestion module. When the simulation identifies a high-risk path or an impending safety accident, the system automatically generates early warning information and provides specific intervention suggestions based on the decision-making strategy of the intelligent agent. These suggestions include adjusting process parameters, activating emergency facilities, or evacuating personnel. The early warning information is pushed in real time through the large screen and mobile terminals of the park control center, with a response delay of less than 5 seconds. For example, when the system predicts that a vapor cloud explosion may occur in a certain tank area within the next 30 seconds, it immediately pushes a Level 1 early warning to the control room and suggests "closing the feed valve, activating the nitrogen inerting system, and evacuating personnel within a 50-meter radius."

[0036] The method also integrates a model update and self-learning mechanism. By continuously collecting actual operational data and accident cases, the agent model and decision-making strategy are updated online or offline. Incremental learning algorithms are used to adapt to equipment aging and process changes, ensuring the timeliness and accuracy of the risk assessment model. The model update cycle can be set daily, weekly, or monthly as needed. Online updates use a sliding window mechanism, retaining only the most recent 30 days of high-quality operational data for fine-tuning; offline updates perform full retraining during non-production periods to ensure that model performance does not degrade.

[0037] To further illustrate the technical effects of this invention, a specific application example is constructed: A large petrochemical park includes two ethylene cracking units, three styrene storage tanks, and supporting utilities. After implementing the method of this invention, the system detected abnormal temperature fluctuations (standard deviation exceeding 3 times) at the outlet of cracking furnace No. 1 during a routine inspection. Traditional QRA methods did not issue an early warning because they did not incorporate real-time data, but the multi-agent framework of this invention immediately initiated dynamic simulation. After 10,000 Monte Carlo samplings, the system identified a high-probability accident chain: temperature fluctuation → increased coking → local overheating of furnace tubes → rupture and leakage → explosion of ethylene cloud upon contact with open flame, with an occurrence probability of 0.18, a severity of consequences of 0.85, and a comprehensive risk value of 0.153, belonging to the high-risk level. The system then generated intervention suggestions: "reduce feed load by 10%, increase dilution steam flow, and arrange infrared thermal imaging detection," and highlighted the area of ​​cracking furnace No. 1 on the control center's large screen with an orange heat map. After the operator adopted the suggestions, the temperature fluctuations returned to normal, successfully avoiding a potentially major accident. Post-hoc analysis showed that Shapleyga and his explanation accurately identified "insufficient dilution ratio" as the main risk contributor, accounting for 62%.

[0038] The above examples fully demonstrate the significant advantages of this invention in terms of real-time risk perception, forward-looking projection, and operability of decision support. By meticulously engineering each technical aspect, this embodiment comprehensively covers the entire chain of technical details from data acquisition, model building, strategy optimization to risk output, ensuring the feasibility of the solution and the verifiability of its technical effects.

[0039] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A chemical industrial park safety risk assessment method based on an agent model, characterized in that: The method comprises the following specific steps: Step 1: Construct a multi-agent collaborative perception framework for a chemical industrial park, collect equipment operation parameters, environmental state information, material flow data and personnel behavior trajectories in real time through distributed control systems, video monitoring networks, gas concentration sensor arrays and personnel positioning beacons deployed in the park, form multi-source heterogeneous perception data streams, and establish a panoramic view of the park operation state under a unified time-space reference by using a data fusion mechanism based on timestamp alignment and spatial coordinate registration; Step 2: Design an agent behavior model library oriented to chemical process mechanisms, establish corresponding physical and chemical process mathematical models for distillation columns, reaction kettles, storage tanks and pipeline systems respectively, the models embed mass conservation, energy conservation and momentum conservation equations, and couple reaction kinetics parameters, heat and mass transfer coefficients and safety interlocking logic, so that the agent can simulate output behaviors conforming to actual physical laws based on current input states; Step 3: Implement agent decision strategy optimization based on deep reinforcement learning, construct a multi-objective optimization problem with chemical safety indicators as reward functions, the safety indicators include the degree of deviation from the safety operation window, the level of potential leakage risk and the emergency response time, and iteratively train the agent strategy in the simulation environment through the proximal policy optimization algorithm; Step 4: Perform dynamic risk deduction and accident chain simulation, drive concurrent interaction of multiple agents in the digital twin environment, simulate the complete evolution process from the initial disturbance event to fire, explosion or toxic substance diffusion, and quantify the occurrence probability and consequence severity of different paths based on the Monte Carlo sampling method; Step 5: Generate an interpretable risk assessment report and a visualized heat map, calculate the risk level of each area in the park based on the deduction results, identify key weak links and risk transmission paths, and dynamically display the risk distribution in the form of a heat map while providing a Shapley-based explanation of the risk contribution analysis.

2. The agent model-based chemical industry park safety risk assessment method according to claim 1, characterized in that: The parameters collected by the distributed control system in step 1 include temperature, pressure, flow rate and liquid level, and the sampling frequency is not less than 1 Hz; the gas concentration sensor array covers hydrogen sulfide, carbon monoxide, methane and volatile organic compounds.

3. The agent model-based chemical industry park safety risk assessment method according to claim 1, characterized in that: The mathematical model of the distillation column in step 2 includes theoretical tray number, reflux ratio and feed composition variables; the reaction kettle model includes reaction rate constant, activation energy and heat load parameter; the storage tank model includes evaporation rate and pressure relief valve set value; the pipeline system model includes fluid flow rate, pressure drop and corrosion rate.

4. The agent model-based chemical industry park safety risk assessment method according to claim 1, characterized in that: The deep reinforcement learning in step 3 uses an actor-critic architecture, both the actor network and the critic network are fully connected neural networks containing 3 hidden layers, the number of neurons in each layer is 256, 128 and 64 respectively, the activation function uses a rectified linear unit, the learning rate is set to 0.001, and the discount factor is 0.

99.

5. The agent model based safety risk assessment method for chemical industry park according to claim 1, characterized in that: The time step of the dynamic risk deduction in step 4 is set to 1 second, the number of Monte Carlo sampling is not less than 10,000 times, the state trajectory of the agent, the triggered event sequence and the final accident consequence are recorded for each sampling, and the frequency and severity indicators of each event chain are counted.

6. The agent model-based chemical industry park safety risk assessment method according to claim 1, characterized in that: The risk level in step 5 is divided into four levels of low risk, medium risk, high risk and extremely high risk according to the product of probability and consequence; The heat map uses a color gradient from green to red to represent the risk from low to high; The Shapley value and explanation are based on cooperative game theory, quantifying the marginal contribution of equipment status, environmental conditions and personnel operation to the risk output.

7. The agent model based chemical industry park safety risk assessment method according to claim 1, characterized in that: A real-time warning and intervention suggestion module is also included, which generates warning information and intervention suggestions for adjusting process parameters, starting emergency facilities or evacuating personnel based on the decision-making strategy of the agent when the deduction identifies a high-risk path or an imminent safety accident. 8.The safety risk assessment method for chemical industry park based on agent model according to claim 1, characterized in that: The method integrates model updating and self-learning mechanism, which updates the agent model and decision-making strategy online or offline by continuously collecting actual operation data and accident cases. 9.The safety risk assessment method for chemical industry park based on agent model according to claim 1, characterized in that: The data fusion mechanism in step 1 includes a data quality assessment and repair sub-module that detects missing values, outliers and consistency conflicts, and uses a Kalman filter-based prediction method to fill in missing data and an isolation forest algorithm to identify outliers.

10. The agent model based safety risk assessment method for chemical industry park according to claim 1, characterized in that: The safety interlocking logic in step 2 is implemented in the form of a rule base, containing more than 200 pre-defined rules covering overpressure protection, over-temperature shutdown, leakage alarm and emergency relief safety scenarios, and the rule trigger conditions are directly associated with the state variables perceived by the agent.

Images