Automatic driving test fault case automatic generation method and vehicle comprehensive evaluation method

Abstract

The invention relates to the field of automatic driving test, in particular to an automatic generation method for an automatic driving test fault case and a vehicle comprehensive evaluation method, and the automatic generation method comprises the following steps: carrying out the fine adjustment of a pre-trained BERT model through employing professional vehicle safety text data; performing fine adjustment on the pre-trained Word2Vec model by adopting professional automobile safety text data; and inputting a test requirement into the BERT model and the Word2Vec model after fine tuning to output a fault case description text needing to be generated to represent a test fault case. According to the invention, through a combined framework of BERT and Word2Vec, a test requirement can be accurately identified and a proper fault case can be generated; according to the vehicle performance comprehensive evaluation method provided by the invention, two aspects of scene complexity and vehicle performance are comprehensively considered, and meanwhile, a subjective and objective comprehensive weighting method based on an optimal sequence graph algorithm and a commentator algorithm is adopted to ensure the flexibility and rationality of evaluation.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving testing, and in particular to a method for automatically generating autonomous driving test fault cases and a method for comprehensive vehicle evaluation. Background Art

[0002] With the rapid development of automotive electronics and advanced driver assistance technologies, autonomous driving, as an advanced stage of assisted driving technology, has become a key solution for future transportation. However, frequent accidents involving functional defects have sparked widespread public concern about autonomous driving safety. As the key to overcoming the bottlenecks in autonomous driving technology implementation, autonomous driving testing technology, through a systematic professional testing and scientific evaluation system, provides a core solution for ensuring autonomous driving safety.

[0003] To ensure the safety of autonomous vehicles, fault injection testing of autonomous vehicle systems in accordance with automotive safety standards such as ISO 26262 is required during autonomous vehicle testing to ensure system reliability. However, existing autonomous vehicle fault case testing suffers from difficulties in generating comprehensive fault test cases based on requirements, a single source of test metrics, difficulty determining metric weights, and a lack of a comprehensive process framework from fault case generation to test evaluation. To meet the rapidly evolving algorithmic requirements of autonomous vehicle testing, there is an urgent need to explore methods for generating fault cases and evaluating them to improve autonomous vehicle testing efficiency. Summary of the Invention

[0004] The purpose of the present invention is to overcome the shortcomings of the existing technology and propose an automatic generation method of autonomous driving test fault cases and a comprehensive vehicle evaluation method.

[0005] In order to achieve the above object, the technical solutions specifically adopted by the present invention are as follows:

[0006] A method for automatically generating fault cases for autonomous driving tests includes the following steps:

[0007] S1. Fine-tune the pre-trained BERT model using professional automotive safety text data.

[0008] First, initialize the pre-trained BERT-base-uncased model. Then, set the number of classification labels based on the number of columns in a professional automotive safety text dataset. After configuring key parameters, deploy the model to the optimal computing device to ensure training efficiency. Use the AdamW optimizer to fine-tune the optimization process and complete model training.

[0009] S2. Fine-tune the pre-trained Word2Vec model using professional automotive safety text data;

[0010] First, the professional automotive safety text is standardized, and the numeric conditions in the professional automotive safety text are identified by validating the regular expression of numbers and converted into a list of floating-point numbers; then the pre-trained word2vec-google-news-300 model is fine-tuned using the standardized professional automotive safety text.

[0011] S3. Input the test requirements into the fine-tuned BERT model and Word2Vec model to output the required fault case description text to represent the test fault case. Specifically, in step S3, the automatic generation of fault test cases is achieved through an input layer-processing layer-output layer architecture, wherein the input layer includes functional requirements text and fault injection condition text, which respectively describe the details of the safety task to be implemented by the system and the mathematical conditions for fault injection; in the processing layer, the BERT model performs multi-classification tasks based on functional safety requirements, identifies the fault impact location, and determines the condition trigger time based on the fault injection condition text; the Word2Vec model uses cosine similarity technology to intelligently match the mathematical conditions in the fault injection condition text to identify the accurate fault mathematical conditions. The condition trigger time and the fault mathematical conditions are combined to form the fault injection timing. Experts generate a fault type library, which contains possible faults for each sensor or actuator (such as random noise, pixel error, CRC error, etc.). The processing layer randomly selects the corresponding fault type based on the fault impact location and passes it to the output layer; the output layer combines the fault injection location, fault injection timing, and fault type input from the processing layer to generate a description text to represent the fault case.

[0012] The present invention also provides a vehicle comprehensive performance evaluation method, which is implemented using a test fault case generated by the above-mentioned method for automatically generating an autonomous driving test fault case, and includes the following steps:

[0013] Step 1: Complete the autonomous vehicle fault use case test;

[0014] Step 2: Establish multi-source and multi-dimensional evaluation indicators, including scenario indicators and vehicle performance indicators. Scenario indicators include interaction intensity and traffic order deviation, while vehicle performance indicators include safety and comfort.

[0015] Step 3: Based on the multi-source and multi-dimensional evaluation indicators, the optimal order is adopted. Figure 1 The critic algorithm comprehensively weights the scene and vehicle indicators based on subjective and objective factors and calculates a comprehensive evaluation score.

[0016] Furthermore, in step 2, the calculation formula for the interaction strength index is as follows:

[0017]

[0018] Among them, n1 is the actual interaction pair number of all vehicles in the scene, and n1′ is the theoretical maximum interaction pair number of all vehicles in the scene;

[0019] The calculation formula for traffic order deviation index is as follows:

[0020]

[0021] Among them, n2 is the number of driving direction changes of all vehicles in the scene under the actual chaotic state, and n2′ is the number of driving direction changes of all vehicles in the scene under the ideal traffic state.

[0022] The safety index calculation formula is as follows:

[0023] Π S =s1(1-C)+s2e -E

[0024] Among them, s1 and s2 are balance parameters, C is the number of collisions, and E is the number of exposures to potentially dangerous driving environments.

[0025] Comfort C It is reflected by two dimensions: vertical jump and horizontal jump, which are defined as follows:

[0026]

[0027] Among them, J0 is the absolute value of the rate of change of the vehicle's longitudinal acceleration, reflecting the smoothness of the accelerator / brake and the longitudinal jump. a It is the absolute value of the rate of change of the vehicle's lateral acceleration, reflecting the comfort of steering and lateral jump. s3 and s4 are balance parameters.

[0028] Furthermore, the step 3 includes the following steps:

[0029] Step 3.1: The subjective weighting method based on the priority graph algorithm obtains the subjective weight vector ws = [ws1; ...; ws n ];

[0030] Step 3.2: The objective weighting method based on the critic algorithm obtains the objective weight wo = [wo1; ...; wo n ];

[0031] Step 3.3: Use the Lagrange multiplier method to find the optimal combination of subjective and objective weights and obtain the comprehensive weight:

[0032]

[0033] Among them, w i is the comprehensive weight of the i-th indicator, ws i is the subjective weight of the i-th indicator, woi is the objective weight of the i-th indicator; ws j is the subjective weight of the j-th indicator, w scoop is the objective weight of the j-th indicator;

[0034] Step 3.4: Calculate the comprehensive score S of the scene and vehicle effect evaluation:

[0035]

[0036]

[0037] in, is the value of the i-th indicator at time t, S (t) is the comprehensive score at time t, and T is the test time after the fault case is injected.

[0038] The present invention has the following characteristics and beneficial effects:

[0039] 1) The automatic generation method of autonomous driving test fault cases proposed in this invention can accurately identify test requirements and generate appropriate fault cases through the combined framework of BERT and Word2Vec, solving the problem of difficulty in generating complete fault test cases according to requirements in autonomous driving tests.

[0040] 2) The comprehensive vehicle performance evaluation method proposed in this invention comprehensively considers both scene complexity and vehicle performance. At the same time, it adopts a subjective and objective comprehensive weighting method based on the priority graph algorithm and the critic algorithm to ensure the flexibility and rationality of the evaluation, solving the problems of single evaluation indicators, rigid evaluation weights, and lack of rationality in autonomous driving tests.

[0041] 3) The method for generating autonomous driving test fault cases and comprehensively evaluating vehicle performance proposed in the present invention constructs a specific autonomous driving fault test process framework, starting from the on-demand generation of fault cases, the confirmation of multiple and multi-dimensional indicators after the test is completed, and then to the subjective and objective weighting to determine the evaluation score, thus filling the gap in the field of autonomous driving testing in this regard. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments with reference to the following drawings:

[0043] Figure 1 This is a flowchart of a method for automatically generating fault cases in autonomous driving tests and a method for comprehensive vehicle evaluation according to an embodiment of the present invention;

[0044] Figure 2 A flowchart for automatically generating a test failure case using BERT and Word2Vec in collaboration in an embodiment of the present invention;

[0045] Figure 3 It is a multi-source and multi-dimensional evaluation index in the embodiment of the present invention;

[0046] Figure 4 The figure is a flow chart of a vehicle comprehensive performance evaluation method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0047] The present invention is described in detail below in conjunction with specific embodiments. The following examples will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other.

[0048] First, in order to solve the problem that it is difficult to generate complete fault test cases according to requirements in autonomous driving tests, an embodiment of the present invention provides a method for automatically generating fault test cases for autonomous driving tests, such as Figure 1 As shown, the following steps are included:

[0049] S1. Fine-tuning the BERT model

[0050] During the model building phase, a pre-trained BERT-base-uncased model is initialized. The number of classification labels in the model is then dynamically adjusted based on the number of columns in specialized automotive safety text datasets such as ISO 26262. Following strict adherence to BERT's 512-token input length specification, key parameters such as batch size and learning rate are configured before the model is deployed to an optimized computing device for training. This allows the pre-trained BERT model to be precisely tuned for specific classification tasks.

[0051] During the model optimization phase, the AdamW optimizer was used to fine-tune the BERT model. Its learning rate setting effectively controlled the magnitude of weight updates to minimize the loss function. Through targeted model configuration and customized optimization strategies, the entire process achieved efficient transfer learning of the BERT model for the functional requirements classification task in autonomous driving testing.

[0052] The fine-tuned BERT model can perform multiple classification and recognition on the input functional requirement text, and obtain various fault locations and fault injection times in the text.

[0053] S2. Fine-tuning the Word2Vec model

[0054] First, we standardized professional automotive safety text, such as ISO 26262, and used regular expressions to identify numeric conditions within the text, converting it into a list of floating-point numbers. We then used the standardized text to fine-tune the pre-trained word2vec-google-news-300 model.

[0055] The fine-tuned Word2Vec model converts the fault injection condition text and the condition text in the database into word vectors. It then calculates the cosine similarity between the fault injection condition and the average vector of each candidate condition, accurately capturing the semantic associations in the mathematical context. Ultimately, it returns the key value corresponding to the condition with the highest similarity, thus obtaining the mathematical condition for the fault.

[0056] S3. BERT and Word2Vec collaborate to automatically generate test failure cases

[0057] like Figure 2 As shown, the present invention achieves automatic generation of fault test cases through a three-layer architecture. The input layer includes functional requirements text and fault injection condition text, which respectively describe the details of the system's safety task and the mathematical conditions for fault injection. At the processing layer, the BERT model performs multi-classification tasks based on the functional safety requirements, identifies the fault impact location (sensor or actuator signal), and determines the condition trigger time based on the fault injection condition text. The Word2Vec model uses cosine similarity technology to intelligently match the mathematical conditions in the fault injection condition text to identify the exact fault mathematical condition. The condition trigger time and the fault mathematical condition are combined to form the fault injection timing. Based on the fault impact location, the processing layer randomly selects the corresponding fault type from an expert-generated fault type library and transmits it to the output layer. The fault type library contains all possible faults (such as random noise, pixel error, CRC error, etc.) for each sensor or actuator, as well as the fault impact location. The output layer combines the fault injection location, fault injection timing, and fault type input from the processing layer to generate a textual description representing the fault case. Based on this fault case, the autonomous vehicle can be simulated or subjected to real-world fault injection testing to verify the safety and other performance of the autonomous vehicle system.

[0058] Secondly, to address the issues of single evaluation indicators, rigid evaluation weights, and lack of rationality in autonomous driving tests, the present invention provides a vehicle comprehensive performance evaluation method based on the aforementioned autonomous driving test fault case, comprising the following steps:

[0059] S1. Complete autonomous vehicle fault use case testing

[0060] S2. Establish multi-source and multi-dimensional evaluation indicators

[0061] like Figure 3As shown, the scenario indicators and vehicle performance indicators that need to be evaluated are first defined. The scenario indicators include two dimensions: interaction intensity and traffic order deviation, and the vehicle performance indicators include two dimensions: safety and comfort.

[0062] The interaction intensity and traffic order deviation of a scene are important indicators for determining the dangerousness of a scene. This paper defines the interaction intensity ∏I as the ratio of the actual interaction pairs of all vehicles in the scene to the theoretical maximum interaction pair. The theoretical maximum interaction pair assumes that all unpassed conflict points in front of each vehicle have interaction objects. Conflict points are locations where the planned driving trajectories of multiple vehicles overlap. The interaction intensity index is calculated as follows:

[0063]

[0064] Among them, n1 is the actual interaction pair number of all vehicles in the scene, and n1′ is the theoretical maximum interaction pair number of all vehicles in the scene.

[0065] In an open road environment without traffic signal control, traffic participants often determine the road right allocation mechanism through autonomous game, which will cause significant traffic sequence uncertainty. Under ideal traffic conditions, a single conflict node will only have one directional priority transfer (for example, car A goes first and car B follows); while in the non-steady-state interaction process, multi-vehicle competition will cause multiple priority alternations at the same node (for example, vehicles A and B repeatedly compete for the right of way). The more times the change occurs, the higher the scene chaos. The present invention uses the traffic order deviation degree Π D It is defined as the ratio of the number of driving direction changes under actual chaotic conditions to the number of driving direction changes under ideal traffic conditions. The traffic order deviation index is calculated as follows:

[0066]

[0067] Among them, n2 is the number of driving direction changes of all vehicles in the scene under the actual chaotic state, and n2′ is the number of driving direction changes of all vehicles in the scene under the ideal traffic state.

[0068] Safety and comfort are the two most important evaluation indicators for autonomous vehicles. S The number of collisions (C) and exposure to potentially dangerous driving situations (E) is measured. Typically, dangerous situations are caused by short predicted collision times or large decelerations. This paper assumes that if the predicted collision time is less than 2s, or the deceleration of the tested vehicle is less than -6m / s 2 , then the driving environment is considered potentially dangerous. Therefore, the safety index calculation formula is:

[0069] Π S =s1(1-C)+s2e -E (3)

[0070] Among them, s1 and s2 are balance parameters, C is the number of collisions, and E is the number of exposures to potentially dangerous driving environments.

[0071] Comfort C It is reflected by two dimensions: vertical jump and horizontal jump, which are defined as follows:

[0072]

[0073] Among them, J0 is the absolute value of the rate of change of the vehicle's longitudinal acceleration, reflecting the smoothness of the accelerator / brake and the longitudinal jump. a It is the absolute value of the rate of change of the vehicle's lateral acceleration, reflecting the comfort of steering and lateral jump. s3 and s4 are balance parameters.

[0074] S3. Conduct subjective and objective comprehensive evaluation

[0075] like Figure 4 As shown, the present invention adopts the priority order Figure 1 The critic algorithm assigns subjective and objective weights to scene and vehicle indicators and calculates a comprehensive evaluation score. Both algorithms require collecting evaluation indicator data through multiple tests to establish an evaluation indicator database.

[0076] First, the relative importance matrix A between each indicator is quantified by the expert evaluation index database and the 1-9 scale (as shown in Table 1), and the weight coefficient ws is obtained by summing each row and normalizing it. i :

[0077]

[0078] A ki *A ik =1 (6)

[0079] Among them, n is the number of indicators, k is the number of rows of matrix A, i, j are the number of columns of matrix A, ki is the relative importance of the kth indicator and the ith indicator, A kj is the relative importance of the kth indicator and the jth indicator.

[0080] Calculate the largest eigenvalue λ of matrix A max , the reliability of the results is verified by the consistency index and random consistency index:

[0081]

[0082]

[0083] Where CI is the consistency index, λ maxis the maximum eigenvalue of the judgment matrix, and RI is the average random consistency index, which is given by experts. If CR < 0.1, the judgment matrix is ​​considered consistent, otherwise it is inconsistent.

[0084] After the test, the final subjective weight vector ws = [ws1; ...; ws n ].

[0085] Table 1 1 to 9 scale

[0086]

[0087] The critic algorithm is an objective weight allocation method based on the entropy weight optimization method. This method analyzes the discreteness and correlation of indicator data, quantifies the independent differences of indicators and the interaction effects between indicators, and then integrates multi-dimensional data features to evaluate the value of indicator information, ultimately achieving scientific weight allocation. First, the initial data matrix B is constructed based on the evaluation indicator database and standardized:

[0088]

[0089] Where n is the number of indicators, m is the number of selected tests, and B ij Indicates the jth indicator value in the i-th test, B′ ij is the jth standardized index value in the i-th test.

[0090] Next, calculate the correlation coefficient between indicators. The larger , the higher the information overlap between the j1th and j2th indicators, and the lower the weight:

[0091]

[0092] in, is the average value of the standardized j1th and j2th indicators in m tests. is the j1th normalized index value in the i-th test, is the j2th standardized index value in the i-th test.

[0093] Through the comprehensive deviation coefficient σ j Correlation coefficient r ij Determine each indicator x j The objective weight of:

[0094]

[0095]

[0096]

[0097] in, is the average value of the jth index in m tests after standardization, c j is the conflict between the jth indicator and other indicators, wo i is the objective weight of the i-th indicator. Through calculation, we can finally get the objective weight of the critic algorithm wo = [wo1;…;wo n ].

[0098] After obtaining the subjective and objective weights, the present invention analyzes the correlation between the two through the Pearson correlation test. After determining that there is no significant correlation between the two, the Lagrange multiplier method is used to solve the optimal combination of subjective and objective weights to obtain the comprehensive weight:

[0099]

[0100] Among them, w i is the comprehensive weight of the i-th indicator, ws i is the subjective weight of the i-th indicator, wo i is the objective weight of the i-th indicator; ws j is the subjective weight of the jth indicator, wo j is the objective weight of the jth indicator;

[0101] Finally, the comprehensive score S for the scene and vehicle effect evaluation is obtained:

[0102]

[0103]

[0104] in, is the value of the i-th indicator at time t, S (t) is the comprehensive score at time t, and T is the test time after the fault case is injected.

[0105] Thirdly, to address the lack of a framework integrating fault case generation and test evaluation in autonomous driving testing, the present invention provides a comprehensive vehicle performance evaluation system. First, a program generates text describing the fault case, which can then be used for fault injection testing on the dSPACE simulation platform. This process is divided into three phases: the initialization phase activates the simulation control software MotionDesk, then configures the Model Desk to build the experimental environment and start the data capture system. The fault injection phase initiates the test scenario, records system behavior in real time, triggers fault injection when conditions are met, captures test process and result data, and resets the environment. The cleanup phase sequentially stops data capture, closes the Model Desk and Motion Desk, terminates the animation demonstration, and archives the test data, completing the closed-loop process. Once the test data is acquired, it is input into the evaluation program for automatic evaluation and generates a final evaluation score.

[0106] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for automatically generating fault cases for autonomous driving tests, characterized by: The steps include: S1. Fine-tune the pre-trained BERT model using professional automotive safety text data. S2. Fine-tune the pre-trained Word2Vec model using professional automotive safety text data; S3. Input the test requirements into the fine-tuned BERT model and Word2Vec model to output the fault case description text that needs to be generated to represent the test fault case.

2. The method for automatically generating fault cases for autonomous driving tests according to claim 1, wherein: In step S1, first, the pre-trained BERT-base-uncased model is initialized; then the number of classification labels is set according to the number of columns of the professional automotive safety text dataset; after completing the configuration of key parameters, the model is deployed to the optimal computing device to ensure training efficiency, and the AdamW optimizer is used to complete the fine-tuning of the optimization link to complete the model training.

3. The method for automatically generating fault cases for autonomous driving tests according to claim 1, wherein: In step S2, the professional automobile safety text is first standardized, and the numerical conditions in the professional automobile safety text are identified by verifying the regular expression of the numbers and converted into a floating-point number list; then the pre-trained word2vec-google-news-300 model is fine-tuned using the standardized professional automobile safety text.

4. The method for automatically generating fault cases for autonomous driving tests according to claim 1, wherein: In step S3, automatic generation of fault test cases is achieved through an input layer-processing layer-output layer architecture, wherein the input layer includes a functional requirement text and a fault injection condition text, which respectively describe the details of the safety task to be implemented by the system and the mathematical conditions of the fault injection: in the processing layer, the BERT model performs multi-classification tasks based on functional safety requirements, identifies the fault impact location, and determines the condition trigger time according to the fault injection condition text; the Word2Vec model uses cosine similarity technology to intelligently match the mathematical conditions in the fault injection condition text to identify the accurate fault mathematical conditions; the processing layer generates a fault injection timing based on the condition trigger time and the fault mathematical conditions, and randomly selects the corresponding fault type from the fault type library according to the fault impact location and passes it to the output layer; the output layer generates a description text to represent the fault use case based on the fault injection location, fault injection timing and fault type passed in by the processing layer.

5. A vehicle comprehensive evaluation method, characterized by: The method for automatically generating fault cases for autonomous driving tests is implemented based on any one of claims 1 to 4.

6. The vehicle comprehensive evaluation method according to claim 5, wherein: The steps include: Step 1: Complete the autonomous vehicle fault use case test; Step 2: Establish multi-source and multi-dimensional evaluation indicators, including scenario indicators and vehicle performance indicators. Scenario indicators include interaction intensity and traffic order deviation, while vehicle performance indicators include safety and comfort. Step 3: Based on the multi-source and multi-dimensional evaluation indicators, a priority graph-critic algorithm is used to perform subjective and objective comprehensive weighting on the scene and vehicle indicators and calculate the comprehensive evaluation score.

7. The vehicle comprehensive evaluation method according to claim 6, wherein: In step 2, the calculation formula of the interaction strength index is as follows: Among them, n1 is the actual interaction pair number of all vehicles in the scene, and n1′ is the theoretical maximum interaction pair number of all vehicles in the scene; The calculation formula for traffic order deviation index is as follows: Among them, n2 is the number of driving direction changes of all vehicles in the scene under the actual chaotic state, and n2′ is the number of driving direction changes of all vehicles in the scene under the ideal traffic state.

8. The vehicle comprehensive evaluation method according to claim 6, wherein: In step 2, the safety index calculation formula is: P S =s1(1-C)+s2e -E Among them, s1 and s2 are balance parameters, C is the number of collisions, and E is the number of exposures to potentially dangerous driving environments.

9. The vehicle comprehensive evaluation method according to claim 6, wherein: In the step 2, comfort Π C It is reflected by two dimensions: vertical jump and horizontal jump, which are defined as follows: Among them, J0 is the absolute value of the rate of change of the vehicle's longitudinal acceleration, reflecting the smoothness of the accelerator / brake and the longitudinal jump. a It is the absolute value of the rate of change of the vehicle's lateral acceleration, reflecting the comfort of steering and lateral jump. s3 and s4 are balance parameters.

10. The vehicle comprehensive evaluation method according to claim 6, wherein: Described step 3 comprises the following steps: Step 3.1: The subjective weighting method based on the priority graph algorithm obtains the subjective weight vector ws = [ws1; ...; ws n ]; Step 3.2: The objective weighting method based on the critic algorithm obtains the objective weight wo = [w01; ...; wo n ]; Step 3.3: Use the Lagrange multiplier method to find the optimal combination of subjective and objective weights and obtain the comprehensive weight: Among them, w i is the comprehensive weight of the i-th indicator, ws i is the subjective weight of the i-th indicator, wo i is the objective weight of the i-th indicator; ws j is the subjective weight of the jth indicator, wo j is the objective weight of the jth indicator; Step 3.4: Calculate the comprehensive score S of the scene and vehicle effect evaluation: in, is the value of the i-th indicator at time t, S (t) is the comprehensive score at time t, and T is the test time after the fault case is injected.

Images