An autonomous vehicle anomaly detection method and related system

Abstract

The application discloses an automatic driving vehicle abnormality detection method and system, adopts system running data information of an automatic driving vehicle in a normal driving state to train a deep neural network abnormality detection model to obtain a trained abnormality detection model, detects real-time data of the automatic driving vehicle by using the abnormality detection model to obtain a detection result of the automatic driving vehicle, and can effectively support engineers or experts in the related field to investigate the abnormality predicted by a black-box abnormality detection model. In addition to helping save time and maintenance costs, the application also realizes the transparency of the model, thereby promoting engineers or experts in the related field to trust the proposed abnormality detection model. The application can well show features that have an important influence on the model output, and show the influence of the features on the model output, thereby improving the understanding of the deep learning model decision of the experts in the field, and more importantly, speeding up the positioning of the experts in the field on the abnormality and the diagnosis of the fault.

Description

Technical Field

[0001] This invention belongs to the field of system fault diagnosis and relates to the analysis of abnormal sensor data in autonomous vehicles. Specifically, it relates to an abnormal detection method and related system for autonomous vehicles. Background Technology

[0002] Autonomous vehicle technology integrates advanced technologies from numerous fields, including sensor technology, computer technology, communication technology, information processing technology, control technology, and artificial intelligence. It holds immense potential in the transportation sector, and its prospects are widely considered promising. Safety is the primary concern for autonomous vehicles. Before making control decisions, autonomous vehicles often need to receive and process vast amounts of sensor data; timely detection of sensor anomalies is crucial for ensuring their safety. Deep learning methods have proven to have significant potential in processing large amounts of high-dimensional data. However, due to their "black box" nature, traditional deep learning methods for detecting sensor anomalies in autonomous vehicles often make decisions that are not understandable to humans. The poor interpretability of deep learning used in anomaly detection for intelligent vehicles hinders the technology's acceptance by society, resulting in low accuracy in vehicle anomaly detection and seriously impacting the safety of autonomous vehicles. Summary of the Invention

[0003] The purpose of this invention is to provide an anomaly detection method and related system for autonomous vehicles, so as to overcome the problems of low anomaly detection accuracy and poor interpretability of existing intelligent detection methods.

[0004] An anomaly detection method for autonomous vehicles includes the following steps:

[0005] A well-trained anomaly detection model is obtained by training a deep neural network anomaly detection model using system operation data information of autonomous vehicles under normal driving conditions.

[0006] Anomaly detection models are used to detect real-time data from autonomous vehicles to obtain detection results for autonomous vehicles.

[0007] Preferably, the system operation data information used in the normal driving state of the autonomous vehicle includes: acquiring system operation data information in the normal driving state of the autonomous vehicle from the autonomous driving sensors of the autonomous vehicle, extracting autonomous vehicle integrated navigation data information from the acquired system operation data information, and preprocessing the extracted autonomous vehicle integrated navigation data information.

[0008] Preferably, the preprocessing of the extracted autonomous vehicle integrated navigation data information includes: retaining one duplicate record; and for samples with incomplete or incorrect fields, using linear interpolation of the corresponding fields of the two adjacent samples before and after the original sample to complete the incorrect or incomplete fields.

[0009] Preferably, the training process of the deep neural network anomaly detection model includes: using the difference between the system operation data information of the original input after compression and decompression of the deep neural network anomaly detection model and the original training data as the data for OCSVM fitting, fitting the input of the OCSVM model that has passed the hyperplane boundary, and using this to detect anomalies. Samples within the boundary are normal samples, and samples outside the boundary are anomalies.

[0010] Preferably, the deep long short-term memory autoencoder network is reconstructed: the output structure of the output layer of the deep neural network anomaly detection model is set to a single-dimensional output, and the weights of each layer of the trained deep neural network anomaly detection model are assigned to the newly constructed long short-term memory autoencoder to replace the original interpretable model for interpretability enhancement.

[0011] Preferably, the training data of the long short-term memory autoencoder deep neural network is used as the background set of the model interpreter. The interpreter uses the reconstructed deep long short-term memory autoencoder network to construct the model interpreter, and the SHAP values ​​of all features of all samples in the test set are calculated based on the model interpreter.

[0012] An anomaly detection system for autonomous vehicles includes an anomaly detection module;

[0013] The anomaly detection module is used to store system operation data information under normal driving conditions of autonomous vehicles to train a deep neural network anomaly detection model to obtain a trained anomaly detection model; and to use the anomaly detection model to detect real-time data of autonomous vehicles to obtain the detection results of autonomous vehicles.

[0014] Preferably, the system operation data information used in the normal driving state of the autonomous vehicle includes: acquiring system operation data information in the normal driving state of the autonomous vehicle from the autonomous driving sensors of the autonomous vehicle, extracting autonomous vehicle integrated navigation data information from the acquired system operation data information, and preprocessing the extracted autonomous vehicle integrated navigation data information.

[0015] Preferably, the preprocessing of the extracted autonomous vehicle integrated navigation data information includes: retaining one duplicate record; and for samples with incomplete or incorrect fields, using linear interpolation of the corresponding fields of the two adjacent samples before and after the original sample to complete the incorrect or incomplete fields.

[0016] Preferably, the training process of the deep neural network anomaly detection model includes: using the difference between the system operation data information of the original input after compression and decompression of the deep neural network anomaly detection model and the original training data as the data for OCSVM fitting, fitting the input of the OCSVM model that has passed the hyperplane boundary, and using this to detect anomalies. Samples within the boundary are normal samples, and samples outside the boundary are anomalies.

[0017] Compared with the prior art, the present invention has the following beneficial technical effects:

[0018] This invention discloses an anomaly detection method for autonomous vehicles. It trains a deep neural network anomaly detection model using system operation data from the autonomous vehicle under normal driving conditions. The trained anomaly detection model is then used to detect anomalies in real-time data from the autonomous vehicle, yielding detection results. This invention effectively supports engineers or experts in related fields in investigating anomalies predicted by the black-box anomaly detection model. Besides saving time and maintenance costs, it also achieves model transparency, thereby promoting trust in the proposed anomaly detection model among engineers and experts.

[0019] This invention, through an interpretable and empowered deep learning model, can effectively demonstrate the features that have a significant impact on the model output, as well as the individual impact of these features on the model output. This enhances domain experts' understanding of the deep learning model's decision-making process and, more importantly, accelerates the domain experts' ability to locate anomalies and diagnose faults. Attached Figure Description

[0020] Figure 1 This is a system module diagram in an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the installation of the GNSS antenna and related equipment for an autonomous vehicle in an embodiment of the present invention.

[0022] Figure 3 This is a schematic diagram of the fully connected long-short-term autoencoder network structure in an embodiment of the present invention.

[0023] Figure 4 This is a schematic diagram of the model interpretation SHAP in this embodiment of the invention.

[0024] Figure 5 This is a model interpretation of the SHAP feature importance graph in this embodiment of the invention.

[0025] Figure 6 This is a flowchart of the anomaly detection method for autonomous vehicles in an embodiment of the present invention. Detailed Implementation

[0026] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0027] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0028] In this embodiment, as Figure 1-6 The implementation of this interpretable anomaly detection method for autonomous vehicles is described in detail. Taking integrated navigation (GNSS + inertial navigation) information as an example, the information selected for anomaly detection includes a high-precision MEMS (Micro-Electro-Mechanical Systems) integrated navigation system installed in the vehicle, a GNSS antenna for receiving satellite signals, an industrial control computer for recording the GNSS logs of the integrated navigation system, and power supply equipment for the aforementioned devices. The equipment installation is as follows... Figure 2 As shown, the calibrated integrated navigation system and industrial control computer are placed and fixed at the rear of the vehicle and connected to the GNSS antenna via a feeder cable. The GNSS antenna is screwed onto two strong magnetic chucks and fixed in the forward and reverse directions of the autonomous vehicle, respectively. It is placed as high as possible in the autonomous vehicle to ensure good GNSS signal reception. At the same time, it is necessary to ensure that the line connecting the phase centers of the two GNSS antennas is aligned with or parallel to the central axis of the test vehicle.

[0029] S1: Obtain system operation data information of autonomous vehicle under normal driving conditions from autonomous vehicle autonomous vehicle sensors, extract autonomous vehicle integrated navigation (GNSS + inertial navigation) data information from the acquired system operation data information, and preprocess the extracted autonomous vehicle integrated navigation data information.

[0030] The extracted integrated navigation data for autonomous vehicles includes protocol headers, sampling times, latitude, longitude, heading angle, eastward speed, and northward speed. However, the extracted data files still suffer from incomplete, missing, and duplicate records, making them unsuitable as direct input for model training. This invention retains only one duplicate record. For samples with incomplete or incorrect fields, linear interpolation of the corresponding fields from the two adjacent samples is used to complete the incorrect or incomplete fields. Subsequent model training or validation requires input data including the vehicle's speed and angular velocity in autonomous driving mode, which cannot be directly obtained from the extracted fields and must be transformed.

[0031] Specific transformation methods include:

[0032] i. Velocity information can be obtained by transforming eastward velocity and northward velocity; let the eastward velocity be v. e The northbound velocity is v n Then the magnitude of the velocity value is:

[0033]

[0034] ii. Angular velocity information is obtained by calculating the difference quotient of yaw angles at two consecutive sampling times:

[0035]

[0036] After the extracted data is completed or modified, its structure needs to be changed to make it suitable for the next step of the deep neural network anomaly detection model.

[0037] S2, Construction and training of a deep neural network anomaly detection model.

[0038] S2.1, the deep neural network anomaly detection model includes a Long Short-Term Memory (LSTM) autoencoder and an OCSVM anomaly detection module. First, the LSTM autoencoder is constructed: a four-layer fully connected deep learning autoencoder (AE) is built using a Long Short-Term Memory (LSTM) neural network, as follows... Figure 3As shown, an autoencoder is a symmetric network architecture, a feedforward neural network consisting of an encoder and a decoder. It compresses data and extracts features from the input layer up to the encoding layer at the center of the architecture; this part is called the encoder. The information extracted from the original input data is called the code or implicit representation, where the dimension change from the original input data to the code is between 100 and 20. The part that expands the data from the code to the output layer is called the decoder. The output of the decoder is generally called the reconstruction data, where the dimension change from the encoding to the reconstruction data is between 20 and 100.

[0039] Starting from the input layer, data is compressed to extract features up to the encoding layer at the center of the architecture. The encoded data is then expanded and reconstructed down to the output layer. This structure ensures that the output remains consistent with the original data structure. Different numerical values ​​are represented by different data, and the reconstruction error constrains model training, which helps detect whether the output data carries a large amount or all of the information from the original data. Under this architecture, the trained and optimized anomaly detection model is applied to actual autonomous vehicle anomaly detection tasks. If anomalies exist in the autonomous vehicle system's operational data, its reconstruction error will exceed the threshold under normal operating conditions, enabling rapid detection of anomalies in the data.

[0040] S2.2, Training of the Long Short-Term Memory Autoencoder: The encoder operation of the autoencoder is described by equation (3), where W and b are the weight matrix and bias vector of the encoder, respectively. Input vector X = [x1, x2, ... x n ] T The encoder encodes it as X′=[x′1,x′2,…x′] n ] T The decoder operation of the autoencoder is as shown in equation (4), where W′ and b′ represent the weight matrix and bias vector of the decoder, respectively. The decoder encodes the vector X′=[x′1,x′2,…x′…]. n ] T Decoded as

[0041] X′=f(WX+b) (3)

[0042]

[0043] The goal of the autoencoder is to reconstruct the input data. The learning objective can be defined by equation (5). At the start of training, the weight parameters and bias parameters are randomly initialized. The back propagation (BP) algorithm is used during training, and the adaptive momentum (Adam) algorithm is used for optimization. The training process is conducted offline. The training data is data extracted and preprocessed under the normal operating conditions of the autonomous vehicle, and the iteration period is 100. After training, the autoencoder can reconstruct data with similar patterns to the training data very well. Conversely, the autoencoder cannot reconstruct data with large errors compared to the training data (abnormal data) very well. For abnormal data, the reconstruction error of the autoencoder will be significantly greater than that of normal data. Based on the magnitude of the reconstruction error of the input data, a judgment can be made as to whether the sample is abnormal.

[0044]

[0045] S2.3, Construction of the OCSVM Anomaly Detection Module. The reconstructed data output from the input data after encoding and decoding by the Long Short-Term Memory (LSTM) autoencoder in the previous step needs to be compared with the input data, and the difference is checked. If the reconstruction error exceeds a set threshold, the sample point can be considered an anomaly. In this process, a suitable threshold needs to be found: the difference between the LSM autoencoder input data and the reconstructed data is used as the input data for the One-Class Support Vector Machine (OCSVM) anomaly detection module, used to fit a hyperplane with the maximum margin separation between normal data points and the origin. For new input reconstruction errors, if the error is too large, it will separate the reconstructed data obtained by comparing the reconstruction error obtained from the LSM autoencoder encoding and decoding of the autonomous vehicle's operating data under normal conditions with the hyperplane fitted by the reconstruction error obtained from the comparison with the original input data, thereby detecting anomalies in the input data.

[0046] The objective function of a class of support vector machine algorithms is defined as:

[0047]

[0048] st (w·Φ(x i ))≥ρ-ξ i ,ξ i ≥0, i=1,...n (7)

[0049] Where w is the feature vector of the high-dimensional feature space, and ξ i It is a slack variable that allows some data points to lie within the margin, ρ is the offset, v∈(0,1) is the trade-off parameter controlling the boundary, n is the number of samples, and x iIt is the i-th input training data, φ(x) i () is a nonlinear mapping function that maps low-dimensional raw data points to high-dimensional data. v sets the upper bound for the scores of outliers and the lower bound for the scores of support vectors. For a sample x, the function s determines whether it is a fault. OCSVM Defined by equation (8).

[0050]

[0051] Where x is the sample to be tested, x i For the i-th input training data, σ∈R is the parameter that determines the radial range of the function, α i Equation (7) is a series of support vectors obtained using the Lagrange technique. If the result of the sample to be detected is negative, it may be an abnormal sample.

[0052] Reconstructing a deep long short-term memory autoencoder network:

[0053] Since SHAP only accepts a single output vector for interpretation, the output structure of the Long Short-Term Memory (LSTM) autoencoder deep learning model in step S2.1 needs to be modified. In this invention, a new LSM autoencoder with the same structure is constructed, except that its output layer's output structure is set to a single dimension. The weights of each layer of the LSM autoencoder deep learning model trained in step S2.1 are assigned to the newly constructed LSM autoencoder, thereby replacing the original interpretable model and enabling interpretability.

[0054] Using SHAP to interpret deep anomaly detection models eliminates the "black box" (or incomprehensibility) of deep learning models. SHAP provides techniques such as feature importance assessment, data point interpretation, and visualization, which can help domain experts better understand the prediction results of deep learning models, thus improving the interpretability of the models and making deep learning-based anomaly detection methods more acceptable to the public.

[0055] SHAP stands for Shapley Additive exPlanations, a general ex-post interpretability method for model-agnostic situations. SHAP is based on the concept of Shapley Value, a game theory concept used to solve payoff distribution problems in multi-player cooperative games. In machine learning, the concept of Shapley Value has been introduced into SHAP interpretation techniques to explain the predictions of machine learning models. This method determines the importance of an individual by calculating its contribution in the cooperation. Here, Shapley represents the Shapley value calculated for each feature variable of each sample; Additive indicates that the Shapley value corresponding to the feature variable is additive for each sample; and explanation represents the interpretation of a single sample, i.e., how each feature variable affects the model's prediction. The goal of SHAP is to explain the prediction of instances by calculating the contribution of each feature to the prediction. The Shapley value interpretation is represented as an additive feature attribution method, expressed by the formula:

[0056]

[0057] Where g is the ex post facto explanation model, and z′∈{0,1} M This indicates whether the corresponding feature "exists" or "does not exist" (1 or 0), where M is the number of input features. φ is the Shapley value for each feature, and φ0 is the predicted mean of all samples in the explanatory model. Since the input data of the model in this invention is preprocessed structured data, all features exist, that is, z′ should be a vector with all values ​​of 1, and according to the local accuracy of the explanatory model g, for a single sample x, g(x) = f(x). Therefore, formula (9) can be simplified to:

[0058]

[0059] The attribution value φ of each feature in formula (10) j It can be calculated according to equation (11):

[0060]

[0061] Where {x1,…,x p Let {x1, ..., x} be the set of all input features, and p be the number of all input features. p}\{x j} is excluding {x j The set of all possible input features, fx (S) is the prediction of the feature subset S. The weights of this expression... This can be understood in the following ways:

[0062] Denominator: p features have p! combinations when ordered arbitrarily.

[0063] Molecular weight: After determining a subset S, there are |S|! (p-|S|-1)! combinations of p features under a specific ordering. Once subset S is determined, the set of features should be {x1,…,x}. |S| ,x j ,x |S|+2 ,…,x p}, subset S{x1,…,x |S| The} itself has |S|! possible combinations of sequences, and must be followed by feature j, then the remaining features {x} |S|+2 ,…,x p Then there are (p-|S|-1)! combinations, and after determining the subset S, there are |S|! (p-|S|-1)! combinations. Therefore... Let S be the proportion of feature combinations of subset S, and the sum of the proportions of feature combinations of all possible subset S equals 1.

[0064] The GradientSHAP interpretation method used in this invention interprets a model using the expected gradient (an extension of the ensemble gradient). The expected gradient is an extension of the ensemble gradient method, a feature attribution method designed for differentiable models. It combines the ideas of the ensemble gradient, SHAP, and smoothing gradient into a single expected value equation. This allows the use of the entire dataset instead of a single reference value as the background distribution and allows for local smoothing. If we approximate the model with a linear function between each background data sample and the current input to be interpreted, and we assume that the input features are independent, then the expected gradient will compute an approximate SHAP value.

[0065] Build the model interpreter and calculate the SHAP value:

[0066] The training data of a deep neural network using a Long Short-Term Memory (LSTM) autoencoder is used as the background set for the model interpreter. The interpreter is constructed using a reconstructed LSM autoencoder network. Since the interpreter focuses on the model's predictions for the test set, it calculates the SHAP values ​​of all features for all samples in the test set. The model interpretation is then visualized, and the possible causes of detected anomalies are analyzed.

[0067] Based on the SHAP values ​​of all features of all samples in the test set calculated by the model interpreter, these SHAP values ​​can be visualized using specific methods within the SHAP framework. This allows for intuitive analysis of how data features influence the model output, thus providing an explanation to the outside world. Specifically, such as... Figure 5 The diagram shown is a schematic representation of the model interpretation SHAP in an embodiment of the present invention. Figure 5 This paper demonstrates the impact of all features (including anomalous patterns) of all samples in the test set on the predicted output of the second neuron (in this embodiment, the second predictive output neuron of the model is selected; other neurons can be selected for analysis) of the reconstructed Long Short-Term Memory Autoencoder deep neural network model in step 3, and analyzes these impacts from a global perspective. Figure 5 The left side displays all feature names in the data, and the gradient bar on the right represents the change in feature values ​​from low to high. Each point extending left and right from the axis represents a sample in the dataset. Features are arranged in descending order of importance, and the horizontal line at the bottom indicates whether the feature's impact (i.e., the SHAP value) is related to the lower or higher predicted value of the sample. Therefore, a negative SHAP value indicates that the higher the feature value, the lower the predicted value, and vice versa. It can be seen that for all features in the test set, higher feature values ​​have a greater impact on the model's predicted output, while relatively lower feature values ​​have a smaller impact.

[0068] More specifically, feature importance maps can be generated using SHAP, such as... Figure 6 The average of the absolute values ​​of the SHAP values ​​of each feature is taken as the importance of that feature, resulting in a stacked bar chart that more intuitively shows the ranking of the importance of each feature in the test set to the predicted output of the second neuron of the model. Figure 6 As can be seen, "longitude," "eastward speed," and "altitude" are the three main influencing features of the model's second neuron's predicted output, meaning that these three features exhibit more anomalies. In practical applications, anomalies are caused by various threats, but their occurrence requires engineers to inspect the entire vehicle system to verify whether the anomaly is due to a fault in the physical component represented by the anomaly feature. This process consumes significant manpower, resources, and funds. However, it is possible to... Figure 5 and Figure 6The illustrated model interpretation visualization overcomes this problem, allowing for the inference that errors should be prioritized in examining physical components represented by several features that significantly impact model prediction anomalies. Therefore, the interpretable deep learning-based anomaly detection method for autonomous vehicles in this invention can effectively support engineers or domain experts in investigating anomalies predicted by black-box anomaly detection models. Besides saving time and maintenance costs, it also achieves model transparency, thereby fostering trust in the proposed anomaly detection model among engineers or domain experts.

Claims

1. An autonomous vehicle anomaly detection method, characterized by, Includes the following steps: A well-trained anomaly detection model is obtained by training a deep neural network anomaly detection model using system operation data information of autonomous vehicles under normal driving conditions. Anomaly detection models are used to detect real-time data from autonomous vehicles to obtain detection results for autonomous vehicles. The training process of the deep neural network anomaly detection model includes: using the difference between the original system operation data information compressed and decompressed by the deep neural network anomaly detection model and the original training data as the data for OCSVM fitting; fitting the OCSVM model through the hyperplane boundary as input to detect anomalies; samples within the boundary are normal samples, and samples outside the boundary are anomalies. The deep neural network anomaly detection model includes a long short-term memory autoencoder and an OCSVM anomaly detection module; Reconstructing a deep long short-term memory autoencoder network: The output structure of the output layer of the deep neural network anomaly detection model is set to a single-dimensional output, and the weights of each layer of the trained deep neural network anomaly detection model are assigned to the newly constructed long short-term memory autoencoder to replace the original interpretable model and enable interpretability. The training data of the Long Short-Term Memory Autoencoder Deep Neural Network is used as the background set of the model interpreter. The interpreter is constructed by using the reconstructed Long Short-Term Memory Autoencoder Network and the SHAP values ​​of all features of all samples in the test set are calculated based on the model interpreter.

2. The method for detecting anomalies in an autonomous vehicle according to claim 1, characterized in that, The system operation data information adopted under the normal driving state of autonomous vehicles includes: acquiring system operation data information under the normal driving state of autonomous vehicles from the autonomous driving sensors of autonomous vehicles, extracting autonomous vehicle integrated navigation data information from the acquired system operation data information, and preprocessing the extracted autonomous vehicle integrated navigation data information. 3.The method of claim 2, wherein, Preprocessing of the extracted autonomous vehicle integrated navigation data includes: retaining one duplicate record; and for samples with incomplete or incorrect fields, using linear interpolation of the corresponding fields of the two adjacent samples before and after the original sample to complete the incorrect or incomplete fields.

4. An autonomous vehicle anomaly detection system, comprising: Includes an anomaly detection module; The anomaly detection module is used to store system operation data information under normal driving conditions of autonomous vehicles to train a deep neural network anomaly detection model and obtain a trained anomaly detection model. Anomaly detection models are used to detect real-time data from autonomous vehicles to obtain detection results for autonomous vehicles. The training process of the deep neural network anomaly detection model includes: using the difference between the original system operation data information compressed and decompressed by the deep neural network anomaly detection model and the original training data as the data for OCSVM fitting; fitting the OCSVM model through the hyperplane boundary as input to detect anomalies; samples within the boundary are normal samples, and samples outside the boundary are anomalies. The deep neural network anomaly detection model includes a long short-term memory autoencoder and an OCSVM anomaly detection module; Reconstructing deep long short-term memory auto-encoder network: setting the output structure of the output layer of the deep neural network anomaly detection model to single dimension output, and assigning each layer weight of the trained deep neural network anomaly detection model to the newly built long short-term memory auto-encoder, so as to replace the original explained model to enable explainability; Using the training data of the long short-term memory auto-encoder deep neural network as the background set of the model interpreter, and using the reconstructed deep long short-term memory auto-encoder network to explain the model, so as to build the model interpreter, and calculate the SHAP value of all features of all samples in the test set based on the model interpreter.

5. The automatic driving vehicle anomaly detection system according to claim 4, wherein the system running data information of the automatic driving vehicle in the normal driving state comprises: obtaining the system running data information of the automatic driving vehicle in the normal driving state from the automatic driving sensor of the automatic driving vehicle, extracting the automatic driving vehicle combined navigation data information from the obtained system running data information, and preprocessing the extracted automatic driving vehicle combined navigation data information. The preprocessing of the extracted automatic driving vehicle combined navigation data information comprises: retaining one record for repeated records, and using linear interpolation of corresponding fields of two adjacent samples before and after the sample to complete the error or incomplete field for the sample with incomplete field or error field.

6. The system of claim 4, wherein the system is configured to: ​

Images