Physical system risk prediction and simulation system based on hierarchical bounding and adversarial reasoning

Abstract

The application discloses a physical system risk prediction and simulation system based on hierarchical delimitation and adversarial deduction, and belongs to the technical field of simulation verification and trusted intelligence; comprising the following steps: S1, the boundary conditions of a complex physical system are determined in layers to form a four-layer boundary space; S2, adversarial deduction is carried out under the determined boundary conditions based on a world model; S3, risk prediction is carried out based on the deduction result; S4, a control baseline is constructed, the deduction result is compared with the control baseline, a verification evaluation index is calculated, and a verification evaluation conclusion is output; S5, whether the verification evaluation conclusion meets the requirements is judged; if not, abnormal attribution analysis is performed, otherwise, S6 is directly performed; S6, a risk prediction and verification evaluation report is generated, and the result is fed back to a model updating process; by adopting the above method, the boundary conditions can be determined systematically, an adversarial scene can be automatically generated, risks can be comprehensively evaluated, and a traceable verification conclusion can be formed.

Description

Technical Field

[0001] This invention belongs to the field of simulation verification and trusted intelligent technology, and in particular relates to a physical system risk prediction and simulation system based on hierarchical delimitation and adversarial inference. Background Technology

[0002] In safety-critical scenarios, the operation of intelligent systems must undergo rigorous verification and evaluation to demonstrate their safety and reliability under various boundary conditions and abnormal scenarios. However, the state space of complex physical systems has extremely high dimensionality, and the combinations of boundary conditions increase exponentially. Taking a nuclear power plant control system as an example, considering only three types of state variables—temperature, pressure, and flow rate—with 10 discrete values ​​for each type of variable, the number of combinations reaches [number missing]. If expanded to ten state variables, the number of combinations will reach... Traditional exhaustive testing methods are completely infeasible.

[0003] Existing verification and evaluation methods have the following specific shortcomings: First, boundary coverage is unsystematic. There is a lack of systematic methods for determining the state space boundaries of complex physical systems. The selection of boundary conditions relies heavily on expert experience, resulting in insufficient coverage and representativeness, and a large number of unexplored boundary blind spots. Existing methods typically only test design operating points and a few extreme operating points, often achieving less than 5% coverage of the boundary space.

[0004] Second, there is a lack of adversarial testing. Existing simulations are mostly conducted under normal or preset conditions, lacking adversarial simulations that address worst-case scenarios and fail to proactively identify system weaknesses. Attackers or extreme natural conditions may exploit these weaknesses to cause system failure, and conventional testing cannot expose these risks.

[0005] Third, risk assessment lacks foresight. Existing methods are mostly based on ex-post statistics or offline analysis, lacking the ability to conduct online risk assessments based on world models. This makes it impossible to predict potential risks before the system is operational, thus missing the opportunity to eliminate risks during the design phase.

[0006] Fourth, there is a disconnect between deduction and verification. The deduction results lack reproducible input records and verifiable output standards, making it difficult to trace verification conclusions and thus unable to support project acceptance and regulatory audits. When verification conclusions are questioned, there is a lack of a complete chain of evidence to support them.

[0007] Fifth, the evaluation metrics are not comprehensive. Existing evaluations focus more on functional correctness (whether the system works as expected), and lack systematic evaluation of inference stability (whether the results of multiple inferences are consistent), uncertainty calibration (whether the confidence interval of the model output is reliable), and model evolution quality (whether the performance improves after the model is updated).

[0008] Sixth, there is a lack of a closed-loop improvement mechanism. The verification and evaluation results are not effectively fed back into the model update process, and the problems found in the evaluation cannot systematically drive model improvement and form a closed loop of "evaluation-attribution-improvement". Summary of the Invention

[0009] The purpose of this invention is to provide a physical system risk prediction and simulation system based on hierarchical delimitation and adversarial inference, which can systematically determine boundary conditions, actively generate adversarial scenarios, comprehensively assess risks, and form traceable and verifiable conclusions.

[0010] To achieve the above objectives, this invention provides a physical system risk prediction and simulation system based on hierarchical delimitation and adversarial inference. The implementation of the system includes the following steps: S1. The boundary conditions of complex physical systems are determined in layers to form a four-layer boundary space; S2. Based on the world model, conduct adversarial simulations under defined boundary conditions; S3. Risk prediction based on simulation results; S4. Construct a control baseline, compare the extrapolation results with the control baseline, calculate the verification evaluation index, and output the verification evaluation conclusion; S5. Determine whether the verification and evaluation results meet the requirements. If they do not meet the requirements, perform anomaly attribution analysis; otherwise, proceed directly to S6. S6. Generate a risk prediction and verification evaluation report and feed the results back to the model update process.

[0011] Preferably, the boundary delimitation process for the four-layer boundary space in S1 is as follows: State-space layer boundary determination: Based on physical laws, equipment specifications, and operational constraints, determine the system's boundary. State variables The range of values ​​for each state variable; Determine its physical feasible region and dynamic constraints: ; ; in and These represent the minimum and maximum threshold values ​​of the preset physical feasible region of the state variables, respectively. Represents state variables The derivative with respect to time, Represents state variables The maximum permissible rate of change; Determine the set of combined constraints between state variables. ; ; in Indicates the first A coupling constraint function, To constrain the total number; Input space layer delimitation: Determines the range of inputs the system may receive, and delimits the input space. Divided into four subdomains: ; in This is a normal input field. For extreme input domains, This is an abnormal input field. For adversarial input domain; Environmental stratification: Determining the range of variation in environmental conditions. ; ; in Within the range of meteorological conditions, Electromagnetic environment range, Within the range of terrain conditions, This is within the scope of human interference; Time-level delimitation: Determining the time range for evaluation. And it is divided into three scales; ; in For short-term transient response window, This is the medium-term steady-state operating window. This serves as a window into a long-term degradation trend.

[0012] Preferably, after the four-layer boundary space delimitation in S1 is completed, a hybrid strategy combining Latin hypercube sampling and risk-based sampling is used to generate a set of boundary condition combinations. .

[0013] Preferably, the process of S2 is as follows: S21. Adversarial Boundary Generation: This involves using an optimization algorithm to search for the combination of boundary conditions in the world model that maximizes the risk index; defining the optimal adversarial boundary. as follows: ; ; in Indicate boundary conditions, This represents the state risk measurement function. To estimate the time range, Indicates the first The state variables at time 1 The value, As a safety threshold, Risk weights for each state variable, Indicates the first The maximum value of each state variable; S22. After obtaining the adversarial boundary, perform five types of simulations, including normal scenario simulation, boundary scenario simulation, abnormal scenario simulation, cascade scenario simulation and long-term degradation simulation. S23. Based on the simulation results, output the candidate state trajectory set, risk distribution, key node identifier, abnormal event sequence and recovery path.

[0014] Preferably, the five types of deductions in S22 are as follows: Normal scenario deduction: under normal boundary conditions The system's operating trajectory is then simulated. Verify the correctness of the function; Boundary scenario simulation: under boundary conditions The system response was simulated to verify its boundary handling capabilities. Represents the boundary surface of the boundary space; Abnormal scenario simulation: under abnormal input Simulate the system response under any of the conditions of sudden environmental changes to verify robustness; Cascading scenario simulation: Simulating a single fault Cascade of effects To verify the system's isolation and recovery capabilities; Long-term degradation projection: Projecting the performance degradation trend of the system after long-term operation. This verifies the ability to issue early warnings of degradation.

[0015] Preferably, the S3 risk prediction process is as follows: S31. Conduct risk identification, and identify a set of potential risk events from the simulation results. Risk event types include functional failure, performance degradation, security breaches, and cascading failures; S32. After risk identification is completed, risk quantification is performed for each type of risk event. Calculate its probability of occurrence and severity of consequences Define risk value The process is as follows: ; ; ; in To calculate the total number of simulations, For the first The set of risk events that occurred in this simulation Risk event For the first The impact function of each consequence dimension Weights for the consequences dimension; S33. Perform risk sorting: Sort risk events in descending order according to risk value and generate a risk priority list. S34. Conduct risk acceptability assessment: compare the risk value with the acceptable threshold. Compare; ; when When a risk is deemed unacceptable, mitigation recommendations are generated, including design modification recommendations, operational restriction recommendations, and enhanced monitoring recommendations. when At that time, the risk is acceptable, so proceed with S4.

[0016] Preferably, the process of S4 is as follows: S41. Construct four types of control baselines, as follows: S411. Construction of statistical distribution based on historical operating data, including the construction of historical baseline using mean, variance, and quantiles. ; S412. Construct a model baseline based on theoretical predictions from simplified models or analytical solutions. ; S413. Construct a baseline for the control experiment based on the measured data from the controlled experiment. ; S414. Constructing expert baselines based on expected results judged by domain experts. ; S42. After the baseline is constructed, the validation and evaluation metrics are calculated. The process is as follows: S421. Accuracy of Calculation and Deduction ; ; in Indicates the reference state in the baseline. Indicates the actual state; S422, Calculation and Deduction Stability ; ; in This represents the variance of the results from multiple simulations. This represents the mean of the second moments from multiple derivations; S423, Calculate the uncertainty calibration degree ; ; in For prediction Confidence interval For the validation set, Represents the actual state value; S424. Calculate the quality of risk prediction ; ; in To improve risk identification recall rate, To improve the accuracy of risk identification, To ensure consistency in risk ranking, , and These represent the corresponding risk weight parameters; S425, Calculate boundary coverage ; ; in This is the set of covered boundary regions. For the total boundary space; S426, Quality of Computational Model Evolution ; ; in and These represent the overall performance indicators before and after the simulation; S43. Evaluate each evaluation indicator, give a pass / fail judgment, and attach the basis and confidence level of the judgment.

[0017] Preferably, the process of S5 is to receive the judgment result of S4, and perform anomaly attribution when the evaluation result does not meet the requirements; otherwise, proceed to S6. The process of attributing anomalies is as follows: S51. Check the quality of the input data and calculate the data quality score. Perform data attribution; ; in This is the original data. As the expected value, Confidence level for data points; when Data quality below a preset data quality threshold is considered data attribution. S52. Examine the model structure and parameters, identify insufficient capacity, overfitting, and underfitting of the model through cross-validation, and attribute the model to the cause. ; in and These represent the model's loss on the training set and the test set, respectively. This represents the cross-validation results of the model. when When the cross-validation identification threshold is exceeded, it is determined to be a model attribution; S53. Check the boundary condition settings, identify cases of boundary omissions and overly strict boundaries through sensitivity analysis, and perform boundary attribution. ; in Represents state variables Sensitivity; S54. Examine the deduction process, identify numerical errors, truncation errors and convergence problems, and attribute the causes of the deduction. S55. Generate an attribution analysis report based on the attribution results, and identify the sources of deviation and directions for improvement, then proceed to S6.

[0018] Preferably, the risk prediction and verification evaluation report for S6 includes the following: Evaluation scope description: boundary space definition, sampling strategy, and inference scenario set; Deduction process record: input conditions, deduction parameters and output results for each deduction; Risk prediction conclusions: list of risk events, risk quantification results, and risk priority ranking; Verification and evaluation indicators: calculated values, pass / fail criteria, and criteria for each indicator; Attribution analysis results: sources of deviation and directions for improvement; Improvement suggestions: Specific improvement measures for unacceptable risks and items that fail to meet evaluation standards.

[0019] Preferably, the system is implemented through a physical system risk prediction and simulation system, including a layered boundary module for determining the boundary conditions of complex physical systems in four layers. A boundary combination generation module for generating a set of boundary condition combinations using a hybrid sampling strategy; An adversarial simulation module that performs five types of simulations and generates adversarial boundaries based on a world model; A risk prediction module that identifies, quantifies, ranks, and determines acceptability of risks based on the simulation results; The validation and evaluation module is used to construct the control baseline, calculate the evaluation indicators, and generate validation conclusions. An anomaly attribution module for performing four-dimensional attribution analysis on items that fail to meet evaluation criteria; The evaluation report generation module is used to generate complete risk prediction and verification evaluation reports.

[0020] Therefore, the present invention employs the above-mentioned physical system risk prediction and simulation system based on hierarchical delimitation and adversarial inference, which has the following beneficial effects: (1) By using the four-layer hierarchical boundary method, the systematic determination and comprehensive coverage of the boundary space were achieved, eliminating the limitations of relying on expert experience; (2) By using the adversarial boundary generation algorithm, the weakest link of the system is actively searched, and the risk discovery is upgraded from "passive waiting" to "active search"; (3) Through five types of simulation systems, comprehensively assess the risks of the system under normal, boundary, abnormal, cascading and degradation scenarios; (4) Through four types of baseline comparisons and a six-dimensional evaluation index system, reproducible, verifiable and traceable verification conclusions are formed; (5) The model is continuously improved through four-dimensional attribution analysis and closed-loop feedback mechanism.

[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0022] Figure 1 This is an overall flowchart of the physical system risk prediction and simulation system based on hierarchical delimitation and adversarial inference of the present invention. Figure 2 This is a system diagram of the physical system risk prediction and simulation system based on hierarchical delimitation and adversarial inference of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Specific model specifications need to be selected and determined according to the actual specifications of the device, etc. The specific selection calculation method adopts existing technology in the art, and therefore will not be described in detail.

[0024] Example 1; like Figure 1 and Figure 2 As shown, this invention provides a physical system risk prediction and simulation system based on hierarchical delimitation and adversarial inference. The implementation of the system includes the following steps: S1. The boundary conditions of complex physical systems are determined in layers to form a four-layer boundary space; State-space layer boundary determination: Based on physical laws, equipment specifications, and operational constraints, determine the system's boundary. State variables The range of values ​​for each state variable; Determine its physical feasible region and dynamic constraints: ; ; in and These represent the minimum and maximum threshold values ​​of the preset physical feasible region of the state variables, respectively. Represents state variables The derivative with respect to time, Represents state variables The maximum permissible rate of change; Determine the set of combined constraints between state variables. ; ; in Indicates the first A coupling constraint function, To constrain the total number; Input space layer delimitation: Determines the range of inputs the system may receive, and delimits the input space. Divided into four subdomains: ; in This is a normal input field. For extreme input domains, This is an abnormal input field. For adversarial input domain; Environmental stratification: Determining the range of variation in environmental conditions. ; ; in Within the range of meteorological conditions, Electromagnetic environment range, Within the range of terrain conditions, This is within the scope of human interference; Time-level delimitation: Determining the time range for evaluation. And it is divided into three scales; ; in For short-term transient response window, This is the medium-term steady-state operating window. This serves as a window into a long-term degradation trend.

[0025] After the four-layer boundary space delimitation in S1 is completed, a hybrid strategy combining Latin hypercube sampling and risk-based sampling is used to generate a set of boundary condition combinations. .

[0026] Latin hypercube sampling specifically involves: sampling each boundary dimension... Divide the space into N non-overlapping intervals, and randomly select a sample point in each interval to ensure that the N sample points in each dimension cover all intervals, thereby ensuring uniform coverage in the boundary space.

[0027] Risk-based sampling specifically refers to: in historically high-risk areas and causal sensitive areas Encrypted sampling.

[0028] Encryption sampling density Defined as: ; in, Based on sampling density, The historical risk heat at boundary point b, The causal sensitivity at boundary point b, and The encrypted weights are based on historical risk and causal sensitivity, respectively.

[0029] S2. Based on the world model, conduct adversarial simulations under defined boundary conditions; S21. Adversarial Boundary Generation: This involves using an optimization algorithm to search the world model for the combination of boundary conditions that maximizes the risk index; defining the optimal adversarial boundary. as follows: ; ; in Indicate boundary conditions, This represents the state risk measurement function. To estimate the time range, Indicates the first The state variables at time 1 The value, As a safety threshold, Risk weights for each state variable, Indicates the first The maximum value of each state variable; Search algorithms include Bayesian optimization methods and gradient-guided search methods; Bayesian optimization method: Constructing a Gaussian process surrogate model for risk indicators The next evaluation point is selected using the expected improvement function. : ; in The maximum known risk value. To achieve the desired improvement, A Gaussian process surrogate model representing risk indicators.

[0030] Gradient-guided search method: When the world model is differentiable, search along the direction of increasing risk. ; in For the search step size, Let be the gradient with respect to boundary condition b; S22. After obtaining the adversarial boundary, perform five types of simulations, including normal scenario simulation, boundary scenario simulation, abnormal scenario simulation, cascade scenario simulation and long-term degradation simulation. The five categories of deductions are as follows: Normal scenario deduction: under normal boundary conditions The system's operating trajectory is then simulated. Verify the correctness of the function; Boundary scenario simulation: under boundary conditions The system response was simulated to verify its boundary handling capabilities. Represents the boundary surface of the boundary space; Abnormal scenario simulation: under abnormal input Simulate the system response under any of the conditions of sudden environmental changes to verify robustness; Cascading scenario simulation: Simulating a single fault Cascade of effects To verify the system's isolation and recovery capabilities; Long-term degradation projection: Projecting the performance degradation trend of the system after long-term operation. To verify the degradation early warning capability, This represents the degradation rate coefficient in long-term degradation projections.

[0031] S23. Output a set of candidate state trajectories based on the deduction results. Risk distribution Key node identifiers, abnormal event sequences, and recovery paths.

[0032] S3. Conduct risk prediction based on the simulation results; S31. Conduct risk identification, and identify a set of potential risk events from the simulation results. Risk event types include functional failure, performance degradation, security breaches, and cascading failures; S32. After risk identification is completed, risk quantification is performed for each type of risk event. Calculate its probability of occurrence and severity of consequences Define risk value The process is as follows: ; ; ; in To calculate the total number of simulations, For the first The set of risk events that occurred in this simulation Risk event For the The impact function of each consequence dimension Weights for the consequences dimension; S33. Perform risk sorting: Sort risk events in descending order according to risk value and generate a risk priority list. S34. Conduct risk acceptability assessment: compare the risk value with the acceptable threshold. Compare; ; when When a risk is deemed unacceptable, mitigation recommendations are generated, including design modification recommendations, operational restriction recommendations, and enhanced monitoring recommendations. when At that time, the risk is acceptable, so proceed with S4.

[0033] S4. Construct a control baseline, compare the extrapolation results with the control baseline, calculate the verification evaluation index, and output the verification evaluation conclusion; S41. Construct four types of control baselines, as follows: S411. Construction of statistical distribution based on historical operating data, including the construction of historical baseline using mean, variance, and quantiles. ; S412. Construct a model baseline based on theoretical predictions from simplified models or analytical solutions. ; S413. Construct a baseline for the control experiment based on the measured data from the controlled experiment. ; S414. Constructing expert baselines based on expected results from domain expert judgments. ; S42. After the baseline is constructed, the validation and evaluation metrics are calculated. The process is as follows: S421. Accuracy of Calculation and Deduction ; ; in Indicates the reference state in the baseline. Indicates the actual state; S422, Calculation and Deduction Stability ; ; in This represents the variance of the results from multiple simulations. This represents the mean of the second moments from multiple derivations; S423, Calculate the uncertainty calibration degree ; ; in For prediction Confidence interval For the validation set, Represents the actual state value; S424, Calculate the quality of risk prediction ; ; in To improve risk identification recall rate, To improve the accuracy of risk identification, To ensure consistency in risk ranking, , and These represent the corresponding risk weight parameters; S425, Calculate boundary coverage ; ; in This is the set of covered boundary regions. For the total boundary space; S426, Quality of Computational Model Evolution ; ; in and These represent the overall performance indicators before and after the simulation; S43. Evaluate each evaluation indicator, give a pass / fail judgment, and attach the basis and confidence level of the judgment.

[0034] S5. Determine whether the verification and evaluation results meet the requirements. If they do not meet the requirements, perform anomaly attribution analysis; otherwise, proceed directly to S6. The process of S5 involves receiving the judgment result of S4. If the evaluation result does not meet the requirements, anomaly attribution is performed; otherwise, S6 is performed. The process of attributing anomalies is as follows: S51. Check the quality of the input data and calculate the data quality score. Perform data attribution; ; in The original data, As the expected value, Confidence level for data points; when Data quality below a preset data quality threshold is considered data attribution. S52. Examine the model structure and parameters, identify insufficient capacity, overfitting, and underfitting of the model through cross-validation, and attribute the model to the cause. ; in and These represent the model's loss on the training set and the test set, respectively. This represents the cross-validation results of the model. when When the cross-validation identification threshold is exceeded, it is determined to be a model attribution; S53. Check the boundary condition settings, identify cases of boundary omissions and overly strict boundaries through sensitivity analysis, and perform boundary attribution. ; in Represents state variables Sensitivity; S54. Examine the deduction process, identify numerical errors, truncation errors and convergence problems, and attribute the causes of the deduction. S55. Generate an attribution analysis report based on the attribution results, and identify the sources of deviation and directions for improvement, then proceed to S6.

[0035] S6. Generate a risk prediction and verification evaluation report and feed the results back to the model update process.

[0036] The risk prediction and verification assessment report includes the following: Evaluation scope description: boundary space definition, sampling strategy, and inference scenario set; Deduction process record: input conditions, deduction parameters and output results for each deduction; Risk prediction conclusions: list of risk events, risk quantification results, and risk priority ranking; Verification and evaluation indicators: calculated values, pass / fail criteria, and criteria for each indicator; Attribution analysis results: sources of deviation and directions for improvement; Improvement suggestions: Specific improvement measures for unacceptable risks and items that fail to meet evaluation standards.

[0037] The system is realized through a physical system risk prediction and simulation system, including a layered boundary module for determining the boundary conditions of complex physical systems in four layers. A boundary combination generation module for generating a set of boundary condition combinations using a hybrid sampling strategy; An adversarial simulation module that performs five types of simulations and generates adversarial boundaries based on a world model; A risk prediction module that identifies, quantifies, ranks, and determines acceptability of risks based on the simulation results; The validation and evaluation module is used to construct the control baseline, calculate the evaluation indicators, and generate validation conclusions. An anomaly attribution module for performing four-dimensional attribution analysis on items that fail to meet evaluation criteria; The evaluation report generation module is used to generate complete risk prediction and verification evaluation reports.

[0038] Taking the water level control system of the secondary loop steam generator in a pressurized water reactor nuclear power plant as an example in the safety verification of nuclear power control systems, the specific implementation process of the method of the present invention is explained; Implementation of step S1: State-space layer boundary definition: Determining the boundaries of the six core state variables—steam generator water level. Water supply flow rate Main steam flow rate Steam pressure Water supply temperature Steam generator metal temperature Determine the coupling constraints: thermodynamic coupling between steam pressure and water level, and mass conservation constraints between feedwater flow rate and main steam flow rate.

[0039] Input space layer delimitation: Normal input domain is the control command within the design operating condition range; extreme input domain is the control command that exceeds the design range but is physically possible (such as the water supply valve being fully open / fully closed); abnormal input domain is the erroneous water level signal (±20% deviation) caused by sensor failure; and adversarial input domain is the input sequence that causes the water level control system to produce the maximum deviation.

[0040] Environmental layer delineation: taking into account cooling water temperature changes (15-35℃), ambient temperature changes (-20 to 50℃), and power grid frequency fluctuations (49.5-50.5 Hz).

[0041] Time-layer delineation: Transient response window is 0-30 seconds (rapid water level changes); steady-state operation window is 30 seconds-2 hours (normal power operation); degradation trend window is 1-30 days (sensor drift, valve wear).

[0042] Boundary combination strategy: 2000 basic boundary combinations are generated using Latin hypercube sampling, and 500 points are sampled in historical high-risk areas (areas where the water level is close to the high / low protection setpoint), for a total of 2500 boundary condition combinations.

[0043] Implementation of step S2: Adversarial boundary generation: Bayesian optimization was used to search for combinations that maximized the water level deviation among 2500 boundary conditions. After 50 iterations, the worst boundary conditions were found: a -15% deviation of the feedwater flow sensor, a 20% step increase in the main steam flow rate, and a 3-second response delay of the feedwater valve. Under these conditions, the water level deviation reached 0.8m (the safety threshold was 0.5m).

[0044] Five types of simulation execution: - Normal scenario simulation: Under design conditions, the water level control accuracy is ±0.05m, and the function is correct; - Boundary scenario simulation: When the water level is close to the high protection setpoint (5.2m), the control system response time increases by 40%; - Abnormal scenario simulation: Simulate the situation when the water level sensor fails (deviation +10%). The system detects the sensor abnormality within 8 seconds and switches to the redundant sensor. - Cascade scenario simulation: Water pump failure → Water flow rate decrease → Water level decrease → Steam generator heat transfer deterioration → Steam pressure decrease. The cascade effect propagates to three subsystems within 45 seconds. - Long-term degradation simulation: The simulation shows that the water level sensor drifts by 0.1% per day over 30 days, and the drift reaches the alarm threshold on the 22nd day.

[0045] Implementation of step S3: Five risk events were identified: water level exceeding limit ( Sensor malfunction not detected ), cascading faults ( ), control delay ( Sensor degradation ).

[0046] Risk ranking: Sensor degradation > Control delay > Water level exceeding limit > Undetected sensor malfunction > Cascade failure.

[0047] Acceptability determination: Let If the control delay (0.32) and sensor degradation (0.45) are considered unacceptable risks, then the control delay (0.32) and sensor degradation (0.45) are considered unacceptable risks.

[0048] Implementation of step S4: Constructing baseline comparisons: historical baseline (statistical data from the past 5 years of operation), model baseline (analytical solution of simplified thermodynamic model), and expert baseline (judgment by experts in the nuclear power field).

[0049] Verification and evaluation results: accuracy of the deduction (Pass); Deductive stability (Pass); Uncertainty calibration (Qualified); Risk assessment quality (Qualified); Boundary coverage (Qualified, but 28% of the area remains uncovered); Model evolution quality (qualified).

[0050] Implementation of step S5: The boundary coverage was 72%, with uncovered areas mainly concentrated in the extreme combination region of steam pressure and feedwater temperature coupling. Sensitivity analysis showed that steam pressure Sensitivity It is significantly higher than other variables (average 0.12), indicating that the boundary settings for this variable need to be expanded.

[0051] Implementation of step S6: The verification and evaluation report is generated, and the following recommendations are made: (1) To address the risk of control delay, it is recommended to optimize the response time of the water supply valve (from 3 seconds to 1.5 seconds); (2) To address the risk of sensor degradation, it is recommended to add an online calibration mechanism; (3) To address the insufficient boundary coverage, it is recommended to add 200 sampling points in the steam pressure-water supply temperature coupling area.

[0052] Example 2: Boundary testing of an autonomous driving system; Take the emergency braking function of a Level 3 autonomous driving system as an example.

[0053] Layered Boundaries: The state space includes vehicle speed (0-200 km / h), distance to the vehicle ahead (0-100 m), road surface friction coefficient (0.1-0.9), and driver reaction time (0-3 s). The input space includes the deceleration of the vehicle ahead (0-10 m / s²). 2 The environmental layer includes lighting conditions (0-100,000 lux) and weather conditions (sunny / rainy / snowy / foggy). The time layer is the 0-10 second emergency braking window.

[0054] Adversarial simulation: The worst-case scenario is found through gradient-guided search: vehicle speed 120 km / h, the vehicle in front suddenly accelerates at 8 m / s. 2 Deceleration, road surface friction coefficient 0.3 (wet and slippery road surface), light intensity 500 lux (dusk). Under these conditions, the braking distance is 68m, exceeding the safe distance of 55m.

[0055] Risk prediction: Identify the risk of exceeding braking distance limits ( (Unacceptable). Recommendation: In wet and slippery road conditions, trigger the warning earlier and increase the warning distance from 55m to 75m.

[0056] Example 3: Cybersecurity Verification of Industrial Control Systems; Taking a chemical industry DCS system as an example, the robustness of the system under network security attacks is verified.

[0057] Layered delimitation: Adversarial input domains include: injection of false sensor data, control command tampering, communication delay attacks, and denial-of-service attacks.

[0058] Adversarial simulation: Bayesian optimization was used to identify the worst-case attack combination: +3°C deviation in temperature sensor data injection, 2-second delay in control commands, and a 30% packet loss rate in communication. Under this attack, the reactor temperature rises by 8°C within 5 minutes (the safety threshold is 5°C).

[0059] Risk prediction: Identifying undetected attack risks ( (Unacceptable). Recommendation: Add a multi-sensor cross-validation mechanism and communication integrity verification.

[0060] Therefore, the present invention employs a physical system risk prediction and simulation system based on hierarchical delimitation and adversarial deduction, which has the following beneficial effects: (1) By using the four-layer hierarchical boundary method, the systematic determination and comprehensive coverage of the boundary space were achieved, eliminating the limitations of relying on expert experience; (2) By using the adversarial boundary generation algorithm, the weakest link of the system is actively searched, and the risk discovery is upgraded from "passive waiting" to "active search"; (3) Through five types of simulation systems, comprehensively assess the risks of the system under normal, boundary, abnormal, cascading and degradation scenarios; (4) Through four types of baseline comparisons and a six-dimensional evaluation index system, reproducible, verifiable and traceable verification conclusions are formed; (5) The model is continuously improved through four-dimensional attribution analysis and closed-loop feedback mechanism.

[0061] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A physical system risk prediction and simulation system based on hierarchical delimitation and adversarial inference, characterized in that, The implementation of the system includes the following steps: S1. The boundary conditions of complex physical systems are determined in layers to form a four-layer boundary space; S2. Based on the world model, conduct adversarial simulations under defined boundary conditions; S3. Conduct risk prediction based on the simulation results; S4. Construct a control baseline, compare the extrapolation results with the control baseline, calculate the verification evaluation index, and output the verification evaluation conclusion; S5. Determine whether the verification and evaluation results meet the requirements. If they do not meet the requirements, perform anomaly attribution analysis; otherwise, proceed directly to S6. S6. Generate a risk prediction and verification evaluation report and feed the results back to the model update process.

2. The physical system risk prediction and simulation system based on hierarchical delimitation and adversarial deduction as described in claim 1, characterized in that, The process of defining the four boundary spaces in S1 is as follows: State-space layer boundary determination: Based on physical laws, equipment specifications, and operational constraints, determine the system's boundary. State variables The range of values ​​for each state variable; Determine its physical feasible region and dynamic constraints: ； ； in and These represent the minimum and maximum threshold values ​​of the preset physical feasible region of the state variables, respectively. Represents state variables The derivative with respect to time, Represents state variables The maximum permissible rate of change; Determine the set of combined constraints between state variables. ; ； in Indicates the first A coupling constraint function, To constrain the total number; Input space layer delimitation: Determines the range of inputs the system may receive, and delimits the input space. Divided into four subdomains: ； in This is a normal input field. For extreme input domains, This is an abnormal input field. For adversarial input domain; Environmental stratification: Determining the range of variation in environmental conditions. ; ； in Within the range of meteorological conditions, Electromagnetic environment range, Within the range of terrain conditions, This is within the scope of human interference; Time-level delimitation: Determining the time range for evaluation. And it is divided into three scales; ； in For short-term transient response window, This is the medium-term steady-state operating window. This serves as a window into a long-term degradation trend.

3. The physical system risk prediction and simulation system based on hierarchical delimitation and adversarial inference as described in claim 2, characterized in that: After the four-layer boundary space delimitation in S1 is completed, a hybrid strategy combining Latin hypercube sampling and risk-based sampling is used to generate a set of boundary condition combinations. .

4. The physical system risk prediction and simulation system based on hierarchical delimitation and adversarial deduction as described in claim 3, characterized in that, The process of S2 is as follows: S21. Adversarial Boundary Generation: This involves using an optimization algorithm to search the world model for the combination of boundary conditions that maximizes the risk index; defining the optimal adversarial boundary. as follows: ； ； in Indicate boundary conditions, This represents the state risk measurement function. To estimate the time range, Indicates the first The state variables at time 1 The value, As a safety threshold, Risk weights for each state variable, Indicates the first The maximum value of each state variable; S22. After obtaining the adversarial boundary, perform five types of simulations, including normal scenario simulation, boundary scenario simulation, abnormal scenario simulation, cascade scenario simulation and long-term degradation simulation. S23. Based on the simulation results, output the candidate state trajectory set, risk distribution, key node identifier, abnormal event sequence and recovery path.

5. The physical system risk prediction and simulation system based on hierarchical delimitation and adversarial inference as described in claim 4, characterized in that, The five types of deductions in S22 are as follows: Normal scenario deduction: under normal boundary conditions The system's operating trajectory is then simulated. Verify the correctness of the function; Boundary scenario simulation: under boundary conditions The system response was simulated to verify its boundary handling capabilities. Represents the boundary surface of the boundary space; Abnormal scenario simulation: under abnormal input Simulate the system response under any of the conditions of sudden environmental changes to verify robustness; Cascading scenario simulation: Simulating a single fault Cascade of effects To verify the system's isolation and recovery capabilities; Long-term degradation projection: Projecting the performance degradation trend of the system after long-term operation. To verify the degradation early warning capability.

6. The physical system risk prediction and simulation system based on hierarchical delimitation and adversarial inference as described in claim 5, characterized in that, The process of S3 risk prediction is as follows: S31. Conduct risk identification, and identify a set of potential risk events from the simulation results. Risk event types include functional failure, performance degradation, security breaches, and cascading failures; S32. After risk identification is completed, risk quantification is performed for each type of risk event. Calculate its probability of occurrence and severity of consequences Define risk value The process is as follows: ； ； ； in To calculate the total number of simulations, For the first The set of risk events that occurred in this simulation Risk event For the The impact function of each consequence dimension Weights for the consequences dimension; S33. Perform risk sorting: Sort risk events in descending order according to risk value and generate a risk priority list. S34. Conduct risk acceptability assessment: compare the risk value with the acceptable threshold. Compare; ； when When a risk is deemed unacceptable, mitigation recommendations are generated, including design modification recommendations, operational restriction recommendations, and enhanced monitoring recommendations. when At that time, the risk is acceptable, so proceed with S4.

7. The physical system risk prediction and simulation system based on hierarchical delimitation and adversarial inference as described in claim 6, characterized in that, The process of S4 is as follows: S41. Construct four types of control baselines, as follows: S411. Construction of statistical distribution based on historical operating data, including the construction of historical baseline using mean, variance, and quantiles. ; S412. Construct a model baseline based on theoretical predictions from simplified models or analytical solutions. ; S413. Construct a baseline for the control experiment based on the measured data from the controlled experiment. ; S414. Constructing expert baselines based on expected results from domain expert judgments. ; S42. After the baseline is constructed, the validation and evaluation metrics are calculated. The process is as follows: S421. Accuracy of Calculation and Deduction ; ； in Indicates the reference state in the baseline. Indicates the actual state; S422, Calculation and Deduction Stability ; ； in This represents the variance of the results from multiple simulations. This represents the mean of the second moments from multiple derivations; S423, Calculate the uncertainty calibration degree ; ； in For prediction Confidence interval For the validation set, Represents the actual state value; S424, Calculate the quality of risk prediction ; ； in To improve risk identification recall rate, To improve the accuracy of risk identification, To ensure consistency in risk ranking, , and These represent the corresponding risk weight parameters; S425, Calculate boundary coverage ; ； in This is the set of covered boundary regions. For the total boundary space; S426, Quality of Computational Model Evolution ; ； in and These represent the overall performance indicators before and after the simulation; S43. Evaluate each evaluation indicator, give a pass / fail judgment, and attach the basis and confidence level of the judgment.

8. The physical system risk prediction and simulation system based on hierarchical delimitation and adversarial inference as described in claim 7, characterized in that: The process of S5 involves receiving the judgment result of S4. If the evaluation result does not meet the requirements, anomaly attribution is performed; otherwise, S6 is performed. The process of attributing anomalies is as follows: S51. Check the quality of the input data and calculate the data quality score. Perform data attribution; ； in The original data, For the expected value, Confidence level for data points; when Data quality below a preset data quality threshold is considered data attribution. S52. Examine the model structure and parameters, identify insufficient capacity, overfitting, and underfitting of the model through cross-validation, and attribute the model to the cause. ； in and These represent the model's loss on the training set and the test set, respectively. This represents the cross-validation results of the model. when When the cross-validation identification threshold is exceeded, it is determined to be a model attribution; S53. Check the boundary condition settings, identify cases of boundary omissions and overly strict boundaries through sensitivity analysis, and perform boundary attribution. ； in Represents state variables Sensitivity; S54. Examine the deduction process, identify numerical errors, truncation errors and convergence problems, and attribute the causes of the deduction. S55. Generate an attribution analysis report based on the attribution results, and identify the sources of deviation and directions for improvement, then proceed to S6.

9. The physical system risk prediction and simulation system based on hierarchical delimitation and adversarial deduction as described in claim 8, characterized in that, The S6 risk prediction and verification evaluation report includes the following: Evaluation scope description: boundary space definition, sampling strategy, and inference scenario set; Deduction process record: input conditions, deduction parameters and output results for each deduction; Risk prediction conclusions: list of risk events, risk quantification results, and risk priority ranking; Verification and evaluation indicators: calculated values, pass / fail criteria, and criteria for each indicator; Attribution analysis results: sources of deviation and directions for improvement; Improvement suggestions: Specific improvement measures for unacceptable risks and items that fail to meet evaluation standards.

10. The physical system risk prediction and simulation system based on hierarchical delimitation and adversarial inference as described in claim 9, characterized in that: The system is realized through a physical system risk prediction and simulation system, including a layered boundary module for determining the boundary conditions of complex physical systems in four layers. A boundary combination generation module for generating a set of boundary condition combinations using a hybrid sampling strategy; An adversarial simulation module that performs five types of simulations and generates adversarial boundaries based on a world model; A risk prediction module that identifies, quantifies, ranks, and determines acceptability of risks based on the simulation results; The validation and evaluation module is used to construct the control baseline, calculate the evaluation indicators, and generate validation conclusions. An anomaly attribution module for performing four-dimensional attribution analysis on items that fail to meet evaluation criteria; The evaluation report generation module is used to generate complete risk prediction and verification evaluation reports.

Images