Intelligent Fault Injection and Performance Quantification System and Method for Helicopter Engineering Simulator

Abstract

This invention discloses an intelligent fault injection and performance quantification system and method for helicopter engineering simulators, belonging to the field of flight simulation and testing technology. It includes a hierarchical composable fault model library module for flexible construction of faults from the component level to the system level and then to composite faults; an intelligent fault injection management module to implement intelligent fault injection strategies; a multi-dimensional performance index quantification module to collect data before and after fault injection in real time and set a performance evaluation index system; and an automated testing and visualization analysis module to compare various quantitative indicators of different control laws or different parameters of the same control law under the same fault scenario, generating a comprehensive evaluation report. The intelligent fault injection and performance quantification system and method for helicopter engineering simulators provided by this invention, through hierarchical fault models, intelligent injection strategies, and automated quantitative evaluation, achieves efficient, objective, and in-depth testing of the robustness, stability, and effectiveness of advanced flight control laws under special conditions.

Description

Technical Field

[0001] This invention relates to the field of flight simulation and testing technology, and in particular to a system and method for intelligent fault injection and performance quantification of helicopter engineering simulators. Background Technology

[0002] Helicopter engineering flight simulators are key equipment for the research and validation of advanced flight control laws (such as fault-tolerant control, adaptive control, and intelligent control). The performance of control laws under special conditions (especially fault states) directly affects flight safety, and therefore must be fully validated in the simulator. Thus, developing a system capable of closed-loop, intelligent, and automated fault injection and performance quantification evaluation has become an important research direction in the field of helicopter engineering simulators.

[0003] Fault injection technology is an important means of testing and evaluating the fault tolerance and reliability of a system, and can be divided into two main categories: hardware fault injection and software fault injection. Software Implemented Fault Injection (SWIFI) technology is widely used in simulation training systems due to its low cost, high flexibility, and lack of the need for additional hardware. It simulates hardware or software faults by modifying the memory image or simulation model parameters of the target system. SWIFI technology can inject a wide range of faults, simulating both hardware and software faults, and has a short development cycle, facilitating the tracking of target system execution and data recovery.

[0004] Current common fault injection testing methods have the following limitations: (1) Single fault scenario: mostly using preset, single fault scripts, lacking the ability to simulate complex coupled faults and time-series faults, making it difficult to reproduce real catastrophic chain faults.

[0005] (2) Open-loop or semi-open-loop testing: Many tests are conducted without or with only partial participation of the control system, which cannot truly simulate the dynamic response and interaction process of the control system after a fault occurs, resulting in low test confidence.

[0006] (3) Simple injection strategy: Fault triggering is mostly based on fixed time or simple conditions, and cannot intelligently and adaptively apply the most stringent fault test according to the current flight status or control system behavior.

[0007] (4) Subjective and one-sided evaluation: The evaluation of results often relies on the subjective analysis of engineers or only examines a few indicators (such as whether it is stable). There is a lack of a comprehensive, objective and automated quantitative indicator system to measure the overall performance of the control system under special conditions. It is inefficient and not easy to compare different control laws. Summary of the Invention

[0008] The purpose of this invention is to provide an intelligent fault injection and performance quantification system and method for helicopter engineering simulators, so as to solve the problems existing in the above-mentioned background art.

[0009] To achieve the above objectives, this invention provides an intelligent fault injection and performance quantification system for helicopter engineering simulators, integrated into a helicopter engineering flight simulator, comprising: A hierarchical composable fault model library module is used for flexible construction from component level to system level and then to complex faults; The intelligent fault injection management module adaptively adjusts faults based on the response performance of the control system, thereby realizing an intelligent fault injection strategy. The multi-dimensional performance index quantification module collects flight status data, control input data, and internal control system data in real time before and after fault injection, and sets up a performance evaluation index system for handling special situations. The automated testing and visualization analysis module is used to manage and run multiple test cases and record all process data. Through a graphical interface, it compares the quantitative indicators of different control laws or different parameters of the same control law under the same fault scenario, and generates a comprehensive evaluation report.

[0010] Preferably, the hierarchical composable fault model library module includes component-level fault models, system-level fault models, and composite fault models.

[0011] Preferably, the component-level fault models include at least high-fidelity mathematical models of sensor faults (bias, jamming, noise, drift), actuator faults (jamming, saturation, efficiency loss, phase lag), engine / transmission system faults (partial / complete power loss, surge, torque fluctuation), and rotor system faults (blade mass imbalance, pitch control linkage failure); the system-level fault models include at least models of flight control computer channel failure, data bus error, hydraulic system pressure leakage, and electrical system power supply interruption; the composite fault model is used to provide a graphical or scripted editing interface, allowing users to combine multiple basic fault models according to specific timing and logical relationships (such as AND, OR, NOT) to generate complex cascading fault scenarios (e.g., after the engine power drops, an attempt to significantly increase the pitch leads to transmission system torque exceeding limits and failure).

[0012] Preferably, the intelligent fault injection management module implements three injection modes through the set injection strategy controller, including script mode, state trigger mode and adaptive injection mode; and provides a real-time data interface synchronized with the main simulation loop of the simulator through the set closed-loop injection interface, so as to inject the fault model output into the specified sensor signal, actuator command or system state without interrupting the simulation.

[0013] Preferably, the three injection modes are as follows: Script mode: Executes the fault injection sequence according to a preset timeline; Status-triggered mode: Automatically triggers associated faults based on real-time flight status (such as altitude, airspeed, rate of descent, attitude angle); Adaptive injection mode: Real-time monitoring of the response performance of the control system (such as tracking error, control output, stability index), and dynamic adjustment of the intensity and type of faults or introduction of new faults based on the feedback response performance. It is used to actively find the performance boundaries and robustness limits of the control system. The logic design of adaptive fault injection is based on the collaboration of reinforcement learning and meta-model, which transforms the fault injection process into an "exploration-exploitation" optimization problem, allowing the algorithm to actively and dynamically find the weak links of the control system like an experienced test engineer.

[0014] Preferably, the performance evaluation index system and its detailed indicators in the multi-dimensional performance index quantification module include at least the following: Stability metrics: overshoot, settling time, number of oscillations, and stability margin; Controllability indicators: command tracking error, flight quality specification compliance, and inter-axis coupling; Robustness metrics: degree of performance degradation, recovery time and recovery rate, maximum deviation; Energy consumption indicators: control energy consumption, engine power demand, and servo motor activity frequency; Task performance metrics: task completion rate, task accuracy, and task time; Human-machine ergonomic indicators: driver's operating load index and frequency of attention capture.

[0015] This invention also provides a method for intelligent fault injection and performance quantification of a helicopter engineering simulator, comprising the following steps: Step 1: The user selects or combines fault scenarios from the fault model library, sets the injection strategy (script, state triggering or adaptive), and selects the control law version to be evaluated; Step 2: Start the adaptive fault injection algorithm in the adaptive injection mode to obtain the optimal fault injection strategy; Step 3: Start the flight simulation. The intelligent fault injection management module injects the fault effects into the closed-loop system during the simulation process according to the optimal fault injection strategy obtained in Step 2. Step 4: The multi-dimensional performance index quantification module synchronously collects all key data throughout the simulation process; after the simulation ends, the multi-dimensional performance index quantification module calculates the preset performance indices. Step 5: The automated testing and visualization analysis module presents the results in a visual format, supports horizontal comparison of multiple test results, and generates evaluation conclusions.

[0016] Preferably, the adaptive fault injection algorithm in step 2 employs a reinforcement learning algorithm, and the specific steps are as follows: Step 21: Initialize the adaptive fault injection algorithm, including initializing the policy network and value network of the reinforcement learning agent, the meta-model (such as a neural network or Gaussian process), and the created empty experience replay buffer. The policy network is used to select fault injection actions based on the current state. Its network parameters are usually randomly initialized or pre-trained and loaded. The meta-model is used to predict the evolution of the system's next state and the immediate reward under a given state and action. During the initialization phase, the meta-model's predictive ability is weak and needs to be updated through subsequent data collection. The experience replay buffer is used to store the experience data generated by the agent's interaction with the simulation environment. Then, the environment state is initialized by resetting the helicopter engineering simulator to the initial test state (such as stable hovering) and obtaining the initial state. Step 22: Enter the intelligent fault injection main loop. The main loop is the core of the algorithm's online learning and decision-making. At each decision moment, the reinforcement learning agent observes the current state from the simulation environment. Step 23: The reinforcement learning agent and the meta-model learn and update collaboratively, asynchronously and in parallel with the main loop of Step 22, and periodically sample data from the experience replay buffer to update the model.

[0017] Preferably, the current state in step 22 includes flight state (such as helicopter altitude, airspeed, attitude angle, angular rate, etc.) and control system performance indicators (such as attitude tracking error, control surface deflection, real-time identified stability indicators, etc.). Based on its current policy and the prediction results of the meta-model, the reinforcement learning agent selects a fault injection action, including discrete actions (such as sensor jamming, actuator efficiency loss of 50%, etc.) and continuous actions (such as setting parameters like fault intensity and duration). Then, it integrates a realistic fault simulation environment (containing a high-fidelity helicopter flight dynamics model and the advanced flight control law to be tested) and runs one or more simulation steps until the system response stabilizes or reaches the next decision point. The simulation environment generates a new state and calculates the immediate reward obtained from this action according to a pre-designed reward function. Finally, the experience tuples generated from this interaction are stored in the experience replay buffer. These data form the basis for subsequent updates to the agent and meta-model.

[0018] Preferably, in step 23, the reinforcement learning agent is first updated. A small batch of empirical data is periodically sampled randomly from the experience replay buffer. Using this data, the agent updates its policy network and value network through a reinforcement learning algorithm. The update objective is to maximize the cumulative expected reward, enabling the agent to learn to select fault injection strategies that more effectively test the performance boundaries of the control system. Then, data is periodically sampled from the experience replay buffer to train and update the meta-model. The meta-model's learning objective is to minimize the error between its predictions and the actual data. The updated meta-model can more accurately predict the control system's response to faults, thus providing the reinforcement learning agent with more reliable and rapid predictions in action decision-making, guiding it to conduct more efficient exploration, and significantly reducing reliance on time-consuming high-fidelity simulations.

[0019] Therefore, the present invention employs the above-mentioned intelligent fault injection and performance quantification system and method for helicopter engineering simulators, which has the following beneficial effects: (1) By establishing a hierarchical and composable fault model library, it supports the flexible construction of fault models from the component level to the system level and then to the composite fault level; (2) The intelligent fault injection strategy not only supports script and state triggering, but also can adaptively adjust the fault according to the response performance of the control system and actively find the performance boundary (adaptive fault injection), which is a significant leap compared with the simple preset fault. (3) Construct an automated quantitative evaluation index system, transform the indicators in flight quality specifications (such as ADS-33E) into quantitative parameters that can be automatically calculated and compared, and realize visualization analysis, so as to efficiently and objectively evaluate the performance of different control laws under fault conditions. (4) Form a complete closed loop from fault injection to control law response and then to performance evaluation.

[0020] 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

[0021] Figure 1 This is the overall architecture diagram of the intelligent fault injection and performance quantification system for helicopter engineering simulators of the present invention; Figure 2 This is a schematic diagram of the hierarchical composable fault model library module in an embodiment of the present invention; Figure 3 This is a flowchart illustrating the workflow of the intelligent fault injection management module in an embodiment of the present invention. Figure 4 This is a flowchart illustrating the workflow of the multi-dimensional performance index quantification module in an embodiment of the present invention. Figure 5 This is a schematic diagram of a typical example of an embodiment of the present invention. Detailed Implementation

[0022] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0023] Please see Figures 1-5 A smart fault injection and performance quantification system for helicopter engineering simulators, integrated into a helicopter engineering flight simulator, includes: A hierarchical composable fault model library module is used for flexible construction from component level to system level and then to complex faults; The intelligent fault injection management module adaptively adjusts faults based on the response performance of the control system, thereby realizing an intelligent fault injection strategy. The multi-dimensional performance index quantification module collects flight status data, control input data, and internal control system data in real time before and after fault injection, and sets up a performance evaluation index system for handling special situations. The automated testing and visualization analysis module is used to manage and run multiple test cases and record all process data. Through a graphical interface, it compares the quantitative indicators of different control laws or different parameters of the same control law under the same fault scenario, and generates a comprehensive evaluation report.

[0024] This invention can be integrated into existing helicopter engineering simulator management software as a software module. The fault model library is written in C++ / Python, and model parameters can be configured via XML files. The intelligent injection module exchanges data with the main simulation loop via real-time Ethernet (such as HLA or DDS) or shared memory to ensure real-time injection. The performance quantization module utilizes MATLAB / Simulink or Python data processing libraries for online or offline calculations. The visualization interface can be developed using Qt or Web technologies.

[0025] The reinforcement learning algorithm for the adaptive fault injection algorithm was built using the PyTorch environment, and the state space, action space, reward function, and meta-model were designed: The state space includes flight status and control system performance indicators; The actions in the action space are fault injection instructions, designed as a hybrid space, including discrete actions and continuous actions; The reward function is an evaluator that guides the behavior of a reinforcement learning agent. Its design principle is to reward injection strategies that effectively expose system weaknesses without directly causing system collapse, expressed as: ; in, , , All are weighting coefficients; This is a performance bonus item, which is positively correlated with the degree of deterioration of the control system's performance indicators (such as the integral of tracking error). As a trend reward, it rewards the trend of continuous performance deterioration. Such as the first or second derivative of the error can be used to reflect this trend, encouraging the agent to find unstable injection points. The penalties include major and minor penalties. Major penalties are designed to result in a large negative reward and termination of the test round if the injection causes the system to diverge (e.g., crash). Minor penalties encourage agents to achieve the test objective with the smallest possible failure intensity, improving test efficiency.

[0026] The metamodel takes the current state and action as input and outputs the predicted next state and predicted reward. The metamodel is modeled using either a neural network (such as an MLP) or a Gaussian process. Neural networks are fast and suitable for high-dimensional spaces; Gaussian processes provide predictive uncertainty, facilitating guided exploration.

[0027] In the initial training phase of the meta-model, a preliminary meta-model is trained by accumulating a small amount of data through random injection. Subsequently, during reinforcement learning, the meta-model is periodically updated online using the latest collected real simulation data to ensure its predictive accuracy. Through multiple rounds of reinforcement learning training, the optimal fault injection strategy is obtained and used as the injection strategy for the adaptive fault injection management model.

[0028] The working steps of the multi-dimensional performance indicator quantification module include: (1) Data acquisition and synchronization: The module subscribes to and synchronizes the required data stream in real time through a high-speed data bus (such as HLA or DDS).

[0029] (2) Preprocessing: Filter the data (such as low-pass filtering to remove noise) and format it.

[0030] (3) Parallel computing: Each sub-unit in the index calculation engine calculates the index it is responsible for in parallel. For time-domain indexes (such as adjustment time), a new data window needs to be opened for calculation after a fault injection or instruction change trigger signal is detected.

[0031] (4) Dynamic baseline management: The module stores or receives the baseline values ​​of each indicator under fault-free operating conditions in real time, which serve as the benchmark for calculating performance degradation, etc.

[0032] The automated testing and visualization analysis module integrates data and displays the results, including: (1) Real-time output: Some key indicators (such as overshoot and current tracking error) can be streamed to the visualization dashboard in real time for monitoring personnel to observe.

[0033] (2) Post-test analysis: After a test case is completed, the module automatically generates a structured data file (such as JSON or XML) containing the final calculation results of all metrics.

[0034] (3) Visualization: The automated evaluation report generator reads the result file and automatically generates a comparison report of multiple rounds of testing, in the form of radar charts, trend curves, bar charts or detailed tables.

[0035] The computational modules for performance quantification and visualization utilize C++ / Python (high-performance numerical computing libraries such as NumPy and SciPy) to ensure computational efficiency. Algorithms involving online identification (such as RLS) and machine learning (for human-computer interaction evaluation) need to be optimized to ensure real-time performance.

[0036] This embodiment uses the adaptive fault injection and control law evaluation of tail rotor efficiency loss in helicopter hovering state as an example to describe in detail the complete workflow of the above system.

[0037] (a) Test objectives To assess whether a novel adaptive fault-tolerant control law can, when faced with a sudden loss of tail rotor performance,: (1) Stabilize the helicopter's attitude and prevent it from entering an uncontrollable spin.

[0038] (2) By coordinating other control surfaces (such as main propeller pitch and cyclic pitch) to automatically compensate for yaw moment, hovering is maintained.

[0039] (3) Provide pilots with clear control prompts and reduce their workload.

[0040] (II) Test Initialization (1) Set baseline: The system is initialized and the helicopter hovers stably at 50 feet above the ground in a fault-free state.

[0041] (2) Select fault: The engineer selects the "tail rotor efficiency loss" fault model in the intelligent fault injection management module.

[0042] (3) Configuration strategy: Select adaptive injection mode and set the target, such as finding the combination of fault parameters that can cause the yaw rate to exceed 30deg / s.

[0043] (4) Define indicators: The performance quantification module automatically loads indicators related to yaw control, such as yaw rate overshoot, recovery time, control energy consumption, ADS-33E bandwidth requirements, etc.

[0044] (III) Adaptive Fault Injection and Closed-Loop Evaluation Process Reinforcement learning is applied to optimize the adaptive fault injection strategy, and then flight simulation is initiated, dynamically injecting fault effects into the closed-loop system during the simulation. The required data streams are subscribed to and synchronized in real time via a high-speed data bus, and then the various sub-units in the index calculation engine are used to calculate and store the indices they are responsible for in parallel.

[0045] (iv) Performance Quantification and Report Generation After the test, the performance quantification module automatically generates a report, including radar charts and weighted indicators for six key metrics: stability, operability, robustness, energy consumption, task efficiency, and ergonomics. The report compares the performance of the new fault-tolerant control law with the old version under the same fault across these six dimensions, concluding that the new control law has significant advantages in stability and ergonomics, but consumes more energy and may trigger engine over-temperature warnings. It is recommended that designers optimize the efficiency of the fault-tolerant algorithm.

[0046] Therefore, this invention employs the aforementioned intelligent fault injection and performance quantification system and method for helicopter engineering simulators. Through composite faults and intelligent injection strategies, it can simulate more extreme, complex, and realistic fault scenarios, fully exploring the potential defects of the control system. The automated testing process and quantitative index system replace traditional manual interpretation, significantly improving testing efficiency and making the evaluation results more comprehensive, objective, and comparable, greatly accelerating the verification and confirmation process of control laws. The adaptive injection mode can proactively explore the performance of the control system under unknown scenarios, discovering vulnerable links that were not considered during the design, and improving the overall robustness of the control system. As an integral part of the helicopter engineering simulator, it can seamlessly collaborate with environmental models, aerodynamic models, etc., providing a complete testing ecosystem.

[0047] 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 helicopter engineering simulator intelligent fault injection and performance quantification system, integrated into a helicopter engineering flight simulator, characterized in that: include: A hierarchical composable fault model library module is used for flexible construction from component level to system level and then to complex faults; The intelligent fault injection management module adaptively adjusts faults based on the response performance of the control system. It adopts an adaptive fault injection algorithm to realize an intelligent fault injection strategy. The reinforcement learning algorithm of the adaptive fault injection algorithm includes state space, action space, reward function and meta-model. The reward function is an evaluator that guides the behavior of a reinforcement learning agent. Its design principle is to reward injection strategies that effectively expose system weaknesses without directly causing system collapse, expressed as: ； in, , , All are weighting coefficients; This is a performance incentive item, which is positively correlated with the degree of deterioration in the performance indicators of the control system. This is a trend-based reward item; The penalties include major and minor penalties. Major penalties are designed to give a huge negative reward and terminate the current test if the injection causes the system to diverge. Minor penalties encourage agents to achieve the test objective with small failure intensity, thereby improving test efficiency. The multi-dimensional performance index quantification module collects flight status data, control input data, and internal control system data in real time before and after fault injection, and sets up a performance evaluation index system for handling special situations. The automated testing and visualization analysis module is used to manage and run multiple test cases and record all process data. Through a graphical interface, it compares the quantitative indicators of different control laws or different parameters of the same control law under the same fault scenario, and generates a comprehensive evaluation report.

2. The intelligent fault injection and performance quantification system for helicopter engineering simulators according to claim 1, characterized in that, The hierarchical composable fault model library module includes component-level fault models, system-level fault models, and composite fault models.

3. The intelligent fault injection and performance quantification system for helicopter engineering simulators according to claim 2, characterized in that: Component-level fault models should include mathematical models for at least sensor faults, actuator faults, engine / transmission system faults, and rotor system faults. System-level fault models include at least flight control computer channel failure, data bus error, hydraulic system pressure leakage, and electrical system power supply interruption models; composite fault models are used to provide a graphical or scripted editing interface, allowing users to combine multiple basic fault models according to timing and logical relationships to generate complex cascading fault scenarios.

4. The intelligent fault injection and performance quantification system for helicopter engineering simulators according to claim 1, characterized in that: The intelligent fault injection management module implements three injection modes through the set injection strategy controller, including script mode, state trigger mode and adaptive injection mode; it provides a real-time data interface synchronized with the main simulation loop of the simulator through the set closed-loop injection interface, and injects the fault model output into the specified sensor signals, actuator commands or system states without interrupting the simulation.

5. The intelligent fault injection and performance quantification system for helicopter engineering simulators according to claim 4, characterized in that, The three injection modes are as follows: Script mode: Executes the fault injection sequence according to a preset timeline; Status-triggered mode: Automatically triggers associated faults based on real-time flight status; Adaptive injection mode: Real-time monitoring of the response performance of the control system, and dynamic adjustment of the intensity and type of faults or introduction of new faults based on the feedback response performance, in order to actively find the performance boundaries and robustness limits of the control system.

6. The intelligent fault injection and performance quantification system for helicopter engineering simulators according to claim 1, characterized in that, The performance evaluation index system in the multi-dimensional performance index quantification module, and its detailed indicators, include at least the following: Stability metrics: overshoot, settling time, number of oscillations, and stability margin; Controllability indicators: command tracking error, flight quality specification compliance, and inter-axis coupling; Robustness metrics: degree of performance degradation, recovery time and recovery rate, maximum deviation; Energy consumption indicators: control energy consumption, engine power demand, and servo motor activity frequency; Task performance metrics: task completion rate, task accuracy, and task time; Human-machine ergonomic indicators: driver's operating load index and frequency of attention capture.

7. A method for intelligent fault injection and performance quantification of a helicopter engineering simulator, applied to the intelligent fault injection and performance quantification system for a helicopter engineering simulator as described in any one of claims 1-6, characterized in that, Includes the following steps: Step 1: The user selects or combines fault scenarios from the fault model library, sets the injection strategy, and selects the control law version to be evaluated; Step 2: Start the adaptive fault injection algorithm in the adaptive injection mode to obtain the optimal fault injection strategy; The reinforcement learning algorithm for adaptive fault injection includes a state space, action space, reward function, and meta-model. The reward function is an evaluator that guides the behavior of a reinforcement learning agent. Its design principle is to reward injection strategies that effectively expose system weaknesses without directly causing system collapse, expressed as: ； in, , , All are weighting coefficients; This is a performance incentive item, which is positively correlated with the degree of deterioration in the performance indicators of the control system. This is a trend-based reward item; The penalties include major and minor penalties. Major penalties are designed to give a huge negative reward and terminate the current test if the injection causes the system to diverge. Minor penalties encourage agents to achieve the test objective with small failure intensity, thereby improving test efficiency. Step 3: Start the flight simulation. The intelligent fault injection management module injects the fault effects into the closed-loop system during the simulation process according to the optimal fault injection strategy obtained in Step 2. Step 4: The multi-dimensional performance index quantification module synchronously collects all key data throughout the simulation process; after the simulation ends, the multi-dimensional performance index quantification module calculates the preset performance indices. Step 5: The automated testing and visualization analysis module presents the results in a visual format, supports horizontal comparison of multiple test results, and generates evaluation conclusions.

8. The intelligent fault injection and performance quantification method for helicopter engineering simulators according to claim 7, characterized in that, The adaptive fault injection algorithm in step 2 uses a reinforcement learning algorithm, and the specific steps are as follows: Step 21: Initialize the adaptive fault injection algorithm, including initializing the policy network and value network of the reinforcement learning agent, the meta-model and the created empty experience replay buffer, and then initialize the environment state by resetting the helicopter engineering simulator to the initial test state and obtaining the initial state. Step 22: Enter the intelligent fault injection main loop. At each decision moment, the reinforcement learning agent observes the current state from the simulation environment. Step 23: The reinforcement learning agent and the meta-model learn and update collaboratively, asynchronously and in parallel with the main loop of Step 22, and periodically sample data from the experience replay buffer to update the model.

9. The intelligent fault injection and performance quantification method for helicopter engineering simulators according to claim 8, characterized in that: The current status in step 22 includes flight status and control system performance indicators; The reinforcement learning agent selects a fault injection action, including discrete and continuous actions, based on its current policy and the prediction results of the meta-model. Then, it runs one or more simulation steps in a real simulation environment that integrates the fault until the system response tends to stabilize or the next decision point is reached. The simulation environment generates a new state and calculates the immediate reward obtained from this action according to a pre-designed reward function. Finally, the experience tuples generated from this interaction are stored in the experience replay buffer.

10. The intelligent fault injection and performance quantification method for helicopter engineering simulators according to claim 8, characterized in that: In step 23, the reinforcement learning agent is updated first. A small batch of experience data is randomly sampled from the experience replay buffer periodically. The agent uses the data to update its policy network and value network through the reinforcement learning algorithm. Then, data is sampled from the experience replay buffer periodically to train and update the meta-model.

Images