A virtual testing method for intelligent vehicles based on fault injection

Abstract

The present invention relates to a smart car testing method based on fault injection, which aims to discover the weaknesses of the regulator by quickly predicting key faults, so as to enhance the fault tolerance of the system; the method first injects faults into the safety boundary scene, wherein the faults that cause the scene to collide are defined as key faults, and then the key faults are quickly found through an adaptive experimental design method based on an agent model, and a neural network is trained to quickly predict the key faults of new scenes, and finally, the fault tolerance of the system is enhanced by using a PID control algorithm. Compared with the prior art, the present invention accelerates the efficiency of smart car testing through an adaptive method for searching for key faults, avoids calculation explosion, and achieves rapid prediction of key faults and enhances the fault tolerance of the system.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent vehicle virtual testing, and in particular to an intelligent vehicle virtual testing method based on fault injection. Background Art

[0002] Highly Automated Vehicles (HAV) have attracted extensive attention and research due to their great potential in driving safety and mobility. In order to ensure that HAV can operate safely in complex traffic flow environments, its evaluation method is crucial. Since there are a large number of black-box deep learning algorithms in the motion planning module of HAV, it is easily affected by errors or failures in hardware and software. Small faults in perception, decision-making or control systems may cause serious conflicts or even accidents. Therefore, before the large-scale promotion of HAV, it is crucial to ensure that HAV can still drive safely in the event of a fault. In order to avoid increased driving risks due to faults, it is necessary to improve the fault tolerance of HAV by searching and eliminating potential critical faults.

[0003] Currently, scenario-based testing has become the mainstream method for HAV testing. Compared with mileage-based testing or function-based testing, the scenario-based testing process is simple and avoids a large amount of redundant and meaningless test content. The test scenarios can be defined as functional, logical, and concrete scenarios according to their abstraction levels. Functional scenarios are a generalization of a class of scenarios based on semantics, logical scenarios describe specific parameter ranges or distributions, and concrete scenarios further determine the specific values ​​of each parameter based on logical scenarios. Scenario-based testing first defines functional scenarios based on test objectives, then defines several representative parameters and their ranges or distributions to determine logical scenarios, and finally, creates a series of concrete scenarios by iteratively searching the parameter space to test HAVs. Scenario-based testing can meet the needs of different optimization goals, such as maximizing scenario coverage, evaluating collision rates, and detecting key scenarios.

[0004] However, these testing methods usually assume that the states of the participants are completely deterministic, which ignores potential errors or failures that may occur in the real world. In other words, existing scenario-based methods cannot simulate failures, including input failures as well as parameter failures. Summary of the invention

[0005] The purpose of the present invention is to provide a smart car virtual testing method based on fault injection in order to overcome the defect that the existing scenario-based method in the prior art cannot simulate faults, including input faults and parameter faults.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] A virtual testing method for an intelligent vehicle based on fault injection comprises the following steps:

[0008] Pre-defined steps: Pre-establish a scenario description format that includes faults and define key faults;

[0009] The key fault search step specifically includes the following sub-steps:

[0010] S101: Given boundary scene and prior information;

[0011] S102: Initializing the fault space, which includes setting the number of faults and a method for generating faults; defining a proxy model and an acquisition function, wherein the proxy model is used to predict the test results of unlabeled samples, and the acquisition function is used to weigh the possibility of collision after injecting faults;

[0012] S103: Adjust the hyperparameters of the proxy model;

[0013] S104: Input the boundary scenario and the selected fault to be injected into the proxy model, the proxy model predicts the estimated danger value of each fault in the fault space in the boundary scenario, and calculates the corresponding acquisition function value through the acquisition function, selects the fault and boundary scenario with the highest acquisition function value for simulation testing, obtains the test results, and obtains the critical fault;

[0014] S105: Retrain the proxy model according to the test results;

[0015] S106: Determine whether the number of selected critical faults reaches a preset number threshold, if so, execute step S107, otherwise select the next critical fault and return to step S103;

[0016] S107: Output all critical faults and their corresponding boundary scenario information.

[0017] Furthermore, the critical fault is a fault that causes a scene collision.

[0018] Furthermore, the scenario description format containing the fault includes parameters describing the scenario, the type of injected fault, the time of fault injection, and the fault value.

[0019] Furthermore, the calculation expression of the acquisition function is:

[0020]

[0021] Where, Acquisition function i is the acquisition function value of the i-th scene, is the estimated value of the test result of the i-th scenario, is the minimum estimated value of the scene test result, a is the adjustment parameter, is the minimum Euclidean distance between the i-th scene and all test scenes.

[0022] Furthermore, the proxy model is a machine learning model.

[0023] Furthermore, in step S13, through grid search and K-fold cross validation, the hyperparameters of the proxy model are adjusted according to the boundary scenarios and fault spaces after simulation tests.

[0024] Furthermore, the method further comprises a neural network-based classification step: gradually loading the acquired boundary scenarios and their key faults into a pre-established neural network classifier for network training until a preset training stop condition is met;

[0025] The new boundary scenarios and fault spaces are loaded into the trained neural network classifier to predict the key faults of each boundary scenario. A simulation test is performed based on the prediction results to determine the prediction accuracy. If the prediction accuracy is greater than the preset accuracy threshold, the neural network classifier is used to predict the key faults of the boundary scenario, otherwise the network training continues.

[0026] Furthermore, the neural network classifier is a multi-layer feedforward neural network trained using an error back propagation algorithm.

[0027] Furthermore, the method further comprises a PID control step:

[0028] When a critical fault is determined to occur through simulation testing in step S104, the drive signal is continuously iteratively adjusted through the PID controller to modify the input parameters of the vehicle control until a preset number of iterations is reached, and the simulation test is performed using the modified input parameters.

[0029] Furthermore, the modification process of the input parameters of the vehicle control comprises the following steps:

[0030]

[0031]

[0032] In the formula, K P is the proportional gain, K I is the integral gain, K D is the differential gain, k is the number of iterations, e k is the systematic error, indicating that R and U k The difference between them, R is the controlled variable, U k is the driving signal, and is the input parameter before updating, is the updated input parameter.

[0033] Compared with the prior art, the present invention has the following advantages:

[0034] (1) The intelligent vehicle virtual testing method based on fault injection proposed in the present invention takes into account the possibility of faults occurring during the operation of the intelligent vehicle regulator during the traditional intelligent vehicle virtual testing process. The present invention adds a key fault search process based on adaptive experimental design, that is, by adding a machine learning proxy model, the estimated values ​​of each fault test result in the boundary scenario are obtained, and the acquisition function is used to weigh the high probability of collision after the injected fault and the exploration of the fault space, and then the fault scheme with the highest acquisition function is selected as the most likely key fault for simulation testing, thereby accelerating the search process. This method can test the fault tolerance of the regulator and fill the gap in existing testing methods.

[0035] (2) The neural network-based classifier proposed in the present invention can quickly predict key faults in various scenarios, providing support for discovering the weaknesses of regulators in various scenarios.

[0036] (3) Based on testing the fault tolerance of the regulator, the present invention provides a method for enhancing the fault tolerance of the regulator. By predicting the impact of a fault on the car, the deviation caused by the fault is corrected, which helps to enhance the fault tolerance of the smart car. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 A schematic diagram of a flow chart of a virtual testing method for an intelligent vehicle based on fault injection provided in an embodiment of the present invention;

[0038] Figure 2 A PID control technology roadmap provided in an embodiment of the present invention;

[0039] Figure 3 A neural network structure diagram of a classifier provided in an embodiment of the present invention;

[0040] In the figure, PID CONTROL is the PID controller, Input is the input, Hidden is the hidden layer, Dropout is the pooling layer, and Output is the output. DETAILED DESCRIPTION

[0041] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0042] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0044] Example 1

[0045] like Figure 1 As shown, this embodiment provides a virtual testing method for intelligent vehicles based on fault injection. First, based on the idea of ​​adaptive design of experiments (ADOE), the search process for key faults is accelerated by adaptively selecting scenarios with more test value. In view of the high-dimensionality of faults, a surrogate model (SM) is used instead of a simulation tool to provide an estimate of the test results and guide the selection of the next scenario. A classifier based on an artificial neural network (ANN) is further trained to quickly evaluate the safety of other faults in undeveloped concrete scenarios. The ultimate goal of fault injection is to evaluate the fault tolerance of the regulator module and enhance this fault tolerance. The proportional-integral-derivative (PID) controller is a widely used controller with the advantages of good robustness and simple algorithm. Therefore, the present invention further studies whether PID control can potentially enhance the fault tolerance of motion regulators.

[0046] The specific steps include:

[0047] Predefined step A1: Pre-establish the scenario description format containing faults, including: the parameters describing the scenario, the type of injected fault, the time of fault injection and the fault value. Define the critical fault as the fault that causes the scenario to crash;

[0048] Key fault search step A2, the key fault search step specifically includes the following sub-steps:

[0049] S101: Given boundary scene and prior information;

[0050] S102: Initialize the fault space, which includes setting the number of faults and the method of generating faults; define a proxy model and an acquisition function, wherein the proxy model is used to predict the test results of unlabeled samples, and the acquisition function is used to weigh the possibility of collision after injecting faults; the proxy model is a machine learning model that can be trained through a data set.

[0051] S103: adjusting the hyperparameters of the proxy model according to the boundary scenarios and fault spaces after simulation tests through grid search and K-fold cross validation;

[0052] S104: Input the boundary scenario and the selected fault to be injected into the proxy model, the proxy model predicts the estimated danger value of each fault in the fault space in the boundary scenario, and calculates the corresponding acquisition function value through the acquisition function, selects the fault and boundary scenario with the highest acquisition function value for simulation testing, obtains the test results, and obtains the critical fault;

[0053] S105: Retrain the proxy model according to the test results;

[0054] S106: Determine whether the number of selected critical faults reaches a preset number threshold, if so, execute step S107, otherwise select the next boundary scenario and return to step S103;

[0055] S107: Output all critical faults and their corresponding boundary scenario information.

[0056] The key fault search step of this embodiment is first based on the idea of ​​Adaptive Design of Experiment (ADOE), which accelerates the key fault search process by adaptively selecting scenarios with more test value. In view of the high-dimensionality of faults, a surrogate model (SM) is used instead of the simulation tool to provide an estimated value of the test results and guide the selection of the next scenario.

[0057] Neural network-based classification step A3: gradually loading the acquired boundary scenes and their key faults into a pre-established neural network classifier for network training until a preset training stop condition is met; the neural network classifier is a multi-layer feedforward neural network trained using an error back propagation algorithm.

[0058] The new boundary scenarios and fault spaces are loaded into the trained neural network classifier to predict the key faults of each boundary scenario. A simulation test is performed based on the prediction results to determine the prediction accuracy. If the prediction accuracy is greater than the preset accuracy threshold, the neural network classifier is used to predict the key faults of the boundary scenario, otherwise the network training continues.

[0059] The neural network-based classification step can train a neural network-based classifier according to the key faults of different concrete scenarios that have been searched, so as to quickly indicate the key faults of other unexplored concrete scenarios.

[0060] PID control steps:

[0061] When a critical fault is determined to occur through simulation testing in step S104, the drive signal is continuously iteratively adjusted through the PID controller to modify the input parameters of the vehicle control until a preset number of iterations is reached, and the simulation test is performed using the modified input parameters.

[0062] The ultimate goal of fault injection is to evaluate the fault tolerance of the motion controller module and enhance this fault tolerance. The proportional-integral-derivative (PID) controller is a widely used controller with advantages such as good robustness and simple algorithm. Therefore, this embodiment further studies whether PID control can potentially enhance the fault tolerance of the motion controller.

[0063] The implementation process of this embodiment is specifically as follows:

[0064] 1) First, based on the ADOE intelligent vehicle fault injection method, multiple key faults in concrete scenarios are quickly discovered;

[0065] 1.1) Establish a basic scenario description format that includes faults. The description parameters include the following:

[0066] ① Parameters describing the scene;

[0067] ② Type of injected fault;

[0068] ③Fault injection time and fault value;

[0069] Based on the above definition of the parameters of the scenario containing faults, a general expression of the scenario containing faults can be given, which is the basis for searching for key faults.

[0070] In this embodiment, a generalized test matrix can be established by establishing a fault space based on the selected concrete scenario. The three-parameter following scenario can be used to calculate the vehicle speed V0, the front vehicle speed V f , relative position to describe Dis. For the fault type, according to existing research, the possible situations can be divided into three categories for the fault value, namely:

[0071] Vanishing faults: The selected variable is set to zero or a null value. Vanishing faults can reflect conditions such as packet loss.

[0072] Stuck-at faults: Selected variables are set to fixed values ​​corresponding to their variable types (e.g., the speed of the leading vehicle is set to the maximum or minimum supported value). By injecting stuck-at faults, the maximum fault tolerance capability is verified.

[0073] Random failures: Selected variables are randomly modified to values ​​within a limited range (e.g., the speed of a vehicle ranges from 0 to 65 mph). Random failures can reflect a variety of failure types in reality.

[0074] Table 1 Different types of fault values

[0075]

[0076] 1.2) Define critical failures. A critical failure is defined as a failure that causes an originally safe boundary scenario to become dangerous; equivalent to a failure that causes a scenario to crash.

[0077] The critical failure set can be defined as:

[0078] F critical ={f:Y<0&Y f ≥0}

[0079] Among them, Y and Y f They represent the simulation test results when there is no fault and when there is a fault respectively. If Y is positive, it means that a collision occurs in the scene, otherwise no collision occurs.

[0080] 1.3) ADOE-based key fault search method, ADOE is a new structure based on the traditional idea of ​​experimental design (DOE). ADOE adaptively updates the next sample to be tested in the fault space based on all previous experience. Before searching, the agent model and acquisition function need to be defined. The agent model is used to predict the test results of unlabeled samples, and the acquisition function is used to solve the trade-off between selecting faults with a high probability of making the scene dangerous and exploring the fault space. The specific search steps are:

[0081] 1.3.1) Initialization. Since there is no prior information about the critical faults of the boundary scenarios, it is necessary to generate random faults for a given safety boundary scenario to establish an initial set. Initialization requires determining two things: the number of initial faults and the method used to generate these faults. The number of initial faults needs to be selected empirically. Considering that ADOE can gradually explore the fault space, a small amount of initial data is sufficient to generate reliable initial results. For the latter part, methods such as random sampling can be used. Since there is no prior information about the critical faults of the boundary scenarios. A random sampling method is used to randomly select 10 groups of faults in the fault space as the initial set.

[0082] 1.3.2) Adjust the hyperparameters of the proxy model. After initialization or adding new tested faults, the entire fault space is divided into two groups: one group consists of tested faults and the other group consists of untested faults. Among the possible combinations of hyperparameter values, grid search and K-fold cross validation are used to select the best hyperparameter combination for the proxy model based on all prior information. In this example, K is 10. Considering that F1-score can take into account both accuracy and recall, it is used as a performance evaluation indicator here.

[0083] First, candidate values ​​for the hyperparameters were defined. Then, the F1–score for each hyperparameter combination was calculated using cross-validation, which consists of two steps: first, all tested scenarios containing faults were split into 10 parts; second, one of the parts was used as the test set and all other parts were used as the training set. The surrogate model was trained and the F1–score was recorded. This process was repeated 10 times, and the average F1–score was used as the final score for the combination. The combination with the highest final score was used.

[0084] 1.3.3) Train the proxy model. After obtaining new and better hyperparameters, retrain the proxy model using all the acquired prior information. Based on the retrained proxy model and the tested fault injection scenario, the test results of the scenario after the fault injection can be predicted.

[0085] 1.3.4) Set termination conditions. Simulation testing is time-consuming and test resources are limited. When the number of injected fault scenarios reaches N limit When N limit Take 1000.

[0086] 1.3.5) Select the next fault to be injected. The acquisition function is used to select the next fault from the candidate set, and the fault solution with the highest acquisition function value is selected. The acquisition function is used to balance the high probability of collision after injecting the fault and exploring the fault space. After selecting and testing the new fault, return to step (1.3.2) and repeat this process until the termination criteria are met.

[0087] The acquisition function is calculated based on the trained proxy model and is as follows:

[0088]

[0089] Where, Acquisition function i is the acquisition function value of the i-th scene, is the estimated value of the test result of the i-th scenario, is the minimum estimated value of the scene test result, a is the adjustment parameter, is the minimum Euclidean distance between the i-th scene and all test scenes.

[0090] Finally, for adaptive search using different proxy models, the search results and algorithm efficiency are shown in Table 2.

[0091] Table 2 Comparison of key fault search efficiency under different proxy models

[0092]

[0093]

[0094] 2) Building a fast classifier based on neural network

[0095] 2.2) The second stage is mainly the neural network training process. In the present invention, the ANN classifier used is a back propagation neural network, which is a multi-layer feedforward neural network trained using the error back propagation algorithm. The structure of the proposed ANN-based classifier is as follows: Figure 3 As shown in the figure, a total of 4 hidden layers are configured, and the ReLU activation function is used: f(x) = max(x, 0).

[0096] 2.2) Train the classifier. Since the time cost of training the classifier is higher than the time cost of searching for critical faults, a step-by-step approach is used to construct the training set. First, select some safety boundary scenarios and their critical faults, and then train the ANN-based classifier. Gradually add more safety boundary scenarios and their critical faults, and repeat the training until the F1–score reaches more than 90%. The training results are as follows;

[0097] Table 3 Classifier training results

[0098]

[0099] 3) Using PID control to enhance system fault tolerance

[0100] The third stage is mainly the PID control process, and the effectiveness of the PID controller is verified by using certain boundary scenarios and their corresponding faults. PID control is used to enhance the robustness of smart cars, where the controlled device is the HAV simulation test. Due to the high cost of simulation tests, a proxy model is used as a simple substitute for HAV simulation tests. The control variable (i.e., the output to be obtained) is the predicted result of the proxy model for the non-fault scenario at the moment of fault injection, where the controlled variable is the predicted result after fault injection, and the predicted result changes as the scenario develops. The drive signal is a linear combination of proportional, integral, and differential terms. The PID controller attempts to minimize the system error by iteratively adjusting the input.

[0101] The PID controller can correct the test results but cannot detect faults, so fault detection is performed before the PID controller is implemented. The feasible input value range is estimated by inputting parameters with prior information to determine whether the input parameters at a certain moment may fail. The workflow of the PID controller with fault detection function established in the present invention is as follows: Figure 2 shown.

[0102] 3.1) Fault Detection

[0103] At time T, if the estimated input parameters are consistent with the actual input parameters within the error range, no fault is marked. Conversely, the parameters whose actual values ​​do not meet the constraints are considered to have failed. If a fault occurs, proceed to step 2, and then perform the HAV simulation test at the next time (T+1) to determine whether a fault occurs.

[0104] 3.2) Modify input parameters through PID control

[0105] The PID controller can continuously adjust the drive signal U k To modify the failed input parameters, and then adjust the parameters until the number of iterations is reached. The process is shown in the following equation.

[0106]

[0107]

[0108] In the formula, K P is the proportional gain, K I is the integral gain, K D is the differential gain, k is the number of iterations, e k is the systematic error, indicating that R and U k The difference between them, R is the controlled variable, U k is the driving signal, and is the input parameter before updating, is the updated input parameter.

[0109] 3.3) Perform HAV simulation test using modified parameters

[0110] Once the PID controller is terminated, the modified input parameters are obtained and the modified parameters are transmitted to the HAV simulation test.

[0111] The safety of the fault-free boundary scenario is compared with the safety of the faulty scenario, where the PID parameters are selected as shown in Table 4. The analysis results are shown in Table 4. Columns 4-5 of the table show a typical case, which compares the test results of the FI scenario with and without PID control. Different fault types have different effects on the performance of the controller. This example uses 100 groups of faults to verify the average performance of the controller. The collision rate of the FI scenario with fixed faults decreased by 5%, the disappearing faults decreased by 7%, and the random faults decreased by 3%. This shows that the PID controller can improve the fault tolerance of the motion controller;

[0112] Table 4 Effect of PID control algorithm

[0113]

[0114]

[0115] In summary, this intelligent vehicle testing method based on fault injection can simulate actual fault conditions, provide assistance for enhancing the safety of intelligent vehicles, and provide a reference for the research of intelligent vehicle testing.

[0116] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A virtual testing method for intelligent vehicles based on fault injection, characterized in that: The following steps are involved: Pre-defined steps: Pre-establish the scenario description format including faults and define key faults; The key fault search step specifically includes the following sub-steps: S101: Given boundary scene and prior information; S102: Initializing the fault space, which includes setting the number of faults and a method for generating faults; defining a proxy model and an acquisition function, wherein the proxy model is used to predict the test results of unlabeled samples, and the acquisition function is used to weigh the possibility of collision after injecting faults; S103: Adjust the hyperparameters of the proxy model; S104: Input the boundary scenario and the selected fault to be injected into the proxy model, the proxy model predicts the estimated danger value of each fault in the fault space in the boundary scenario, and calculates the corresponding acquisition function value through the acquisition function, selects the fault and boundary scenario with the highest acquisition function value for simulation testing, obtains the test results, and obtains the critical fault; S105: Retrain the proxy model according to the test results; S106: Determine whether the number of selected critical faults reaches a preset number threshold, if so, execute step S107, otherwise select the next critical fault and return to step S103; S107: Output all critical faults and their corresponding boundary scenario information.

2. The intelligent vehicle virtual testing method based on fault injection according to claim 1 is characterized in that: The critical fault is a fault that causes a scene collision.

3. The intelligent vehicle virtual testing method based on fault injection according to claim 1, characterized in that: The scenario description format including the fault includes parameters describing the scenario, the type of injected fault, the time of fault injection, parameters of the fault, and fault value.

4. The intelligent vehicle virtual testing method based on fault injection according to claim 1, characterized in that: The calculation expression of the acquisition function is: Where, Acquisition function i is the acquisition function value of the i-th scene, is the estimated value of the test result of the i-th scenario, is the minimum estimated value of the scene test result, a is the adjustment parameter, is the minimum Euclidean distance between the i-th scene and all test scenes.

5. The intelligent vehicle virtual testing method based on fault injection according to claim 1, characterized in that: The proxy model is a machine learning model.

6. The intelligent vehicle virtual testing method based on fault injection according to claim 1, characterized in that: In step S13, through grid search and K-fold cross validation, the hyperparameters of the proxy model are adjusted according to the boundary scenarios and fault spaces after simulation tests.

7. The intelligent vehicle virtual testing method based on fault injection according to claim 1, characterized in that: The method further comprises a neural network-based classification step: gradually loading the acquired boundary scenarios and their key faults into a pre-established neural network classifier for network training until a preset training stop condition is met; The new boundary scenarios and fault spaces are loaded into the trained neural network classifier to predict the key faults of each boundary scenario. A simulation test is performed based on the prediction results to determine the prediction accuracy. If the prediction accuracy is greater than the preset accuracy threshold, the neural network classifier is used to predict the key faults of the boundary scenario, otherwise the network training continues.

8. The intelligent vehicle virtual testing method based on fault injection according to claim 7, characterized in that: The neural network classifier is a multi-layer feedforward neural network trained using an error back propagation algorithm.

9. The intelligent vehicle virtual testing method based on fault injection according to claim 1, characterized in that: The method further comprises a PID control step: When a critical fault is determined to occur through simulation testing in step S104, the drive signal is continuously iteratively adjusted through the PID controller to modify the input parameters of the vehicle control until a preset number of iterations is reached, and the simulation test is performed using the modified input parameters.

10. The intelligent vehicle virtual testing method based on fault injection according to claim 1, characterized in that: The modification process of the input parameters of the vehicle control comprises the following steps: In the formula, K P is the proportional gain, K I is the integral gain, K D is the differential gain, k is the number of iterations, e k is the systematic error, indicating that R and U k The difference between them, R is the controlled variable, U k is the driving signal, and is the input parameter before updating, is the updated input parameter.

Images