CN115964676A - Unsupervised automatic driving automobile fault detection method and system - Google Patents

Abstract

The invention discloses a fault detection method and system for an unsupervised automatic driving vehicle. Through real-time collection of driving state data of the automatic driving The encoder detection results are fused to obtain the encoder fusion detection results; at the same time, a class of support vector machine model, a local outlier factor model and an isolated forest model are used to detect the driving state data of the autonomous vehicle to obtain their respective detection results. The detection results of the support vector machine model, the local outlier factor model and the isolated forest model are further fused with the encoder fusion detection results to obtain the final detection results. From a data-driven perspective, an integration of multiple methods for solving fault detection problems is designed. A framework that can efficiently detect sensor data anomalies and faults in the operating state of autonomous vehicles.

Description

A method and system for fault detection of unsupervised self-driving cars

technical field

The invention belongs to the field of fault diagnosis of automobile systems, and in particular relates to a fault detection method and system for an unsupervised automatic driving automobile.

Background technique

Self-driving car technology integrates high-tech technologies in many fields such as sensor technology, computer technology, communication technology, information processing technology, and control technology, and has huge development potential. The safety of autonomous vehicles has attracted widespread attention. Existing fault detection methods for autonomous vehicles mainly include: model-based methods, signal-based methods, and data-driven methods. A single fault detection method is based on specific assumptions, and it is difficult to It is difficult to effectively detect the faults that occur in the operation of autonomous vehicles by comprehensively learning the sensor data patterns of autonomous vehicles.

Contents of the invention

The purpose of the present invention is to provide a fault detection method and system for an unsupervised automatic driving vehicle, so as to overcome the problem of low fault detection accuracy of the existing unsupervised automatic driving vehicle.

A fault detection method for an unsupervised self-driving car, comprising the following steps:

S1, collect the driving status data of the self-driving car in real time, use autoencoders with different structures to detect the acquired driving status data, and fuse the detection results of the autoencoders with different structures to obtain the encoder fusion detection results;

S2. At the same time, a class of support vector machine model, local outlier factor model and isolated forest model are used to detect the driving state data of autonomous vehicles to obtain their respective detection results. A class of support vector machine model, local outlier factor model The detection results of the isolation forest model and the encoder fusion detection results are further fused to obtain the final detection results.

Preferably, the vehicle driving state data in the normal driving state of the self-driving car is used as a training set to train multiple autoencoders with different structures.

Preferably, useful fields are extracted from the vehicle driving state data in the normal driving state of the self-driving car and data cleaning and data transformation are performed;

The extracted useful fields include protocol header, sampling instant, heading angle, easting speed, and northing speed.

Preferably, the speed information is obtained according to the transformation of the eastward speed and the northward speed, record the eastward speed as v _e , and the northward speed as v _n , then the magnitude of the speed value is:

The angular velocity information is obtained by calculating the difference quotient of the yaw angle at two consecutive sampling moments:

Preferably, the automatic encoder includes an encoder structure and a decoder structure, the encoder structure is described by formula (3), W and b are respectively the weight matrix and the bias vector of the encoder;

The input vector X=[x ₁ ,x ₂ ,…x _n ] ^T is encoded by the encoder structure as X′=[x′ ₁ ,x′ ₂ ,…x′ _n ] ^T ; the decoder structure operates as formula (4 ), W′ and b′ represent the weight matrix and bias vector of the decoder respectively, and the decoder decodes the encoded vector X′=[x′ ₁ ,x′ ₂ ,…x′ _n ] ^T as

X'＝f(WX+b) (8)

The goal of the autoencoder is to reconstruct the input data, the goal is as in formula (5), and the weight matrix and bias vector parameters are optimized through the backpropagation algorithm.

Preferably, the autoencoder constructed above is a non-fully connected autoencoder.

Preferably, multiple non-fully-connected autoencoders evaluate the input samples respectively, and respectively obtain the reconstruction errors of the samples. Note that the reconstruction error of the i-th non-fully-connected autoencoder for the current sample is e _i , then multiple The reconstruction error of the non-fully connected autoencoder for the current sample is E=[e ₁ ,e ₂ ,…e _n ] ^T .

Preferably, a class of support vector machine (One-Class Support Vector Machine, OCSVM) algorithm is used to construct a hyperplane with maximum interval separation between the normal data point and the origin for fault detection, and the data point is located within the boundary, then Think of making normal samples.

Preferably, the fault score after fusion is mapped to the interval (0,1) through the sigmoid function, and whether a fault occurs is judged according to the value after mapping, and the fault is on the interval (0,0.5), and (0.5,1) The range is normal.

An unsupervised self-driving car fault detection system, a data acquisition module and a detection module;

The data acquisition module is used to collect the driving status data of the self-driving car in real time, and transmit the collected data to the detection module;

The detection module stores autoencoders of different structures, a class of support vector machine models, local outlier factor models and isolated forest models, uses autoencoders of different structures to detect the acquired driving status data, and automatically encodes different structures The encoder detection results are fused to obtain the encoder fusion detection results; at the same time, a class of support vector machine model, local outlier factor model and isolated forest model are used to detect the driving state data of the autonomous vehicle to obtain their respective detection results. The detection results of support vector machine model, local outlier factor model and isolation forest model are further fused with the encoder fusion detection results to obtain the final detection results.

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

The invention discloses an unsupervised self-driving car fault detection method, which collects the driving state data of the self-driving car in real time, uses automatic encoders of different structures to detect the acquired driving state data, and performs detection results of the automatic encoders of different structures. Fusion to obtain the encoder fusion detection results; at the same time, a class of support vector machine model, local outlier factor model and isolated forest model are used to detect the driving state data of the autonomous vehicle to obtain their own detection results, and a class of support vector machine model , the local outlier factor model and the isolated forest model detection results are further fused with the encoder fusion detection results to obtain the final detection results. From a data-driven perspective, an integrated framework that integrates multiple methods for solving fault detection problems is designed, which can effectively It can accurately detect sensor data anomalies and faults in the operating state of autonomous vehicles.

This application uses integrated autoencoder, OCSVM, LOF and IF to simultaneously monitor the motion state of the vehicle, which can effectively avoid the singleness of data processing by a single model. This application will vote on the detection results of the vehicle motion state to build an integrated fault detection framework , can verify and consider the final inspection results from multiple angles, and improve the accuracy of vehicle fault detection.

Description of drawings

FIG. 1 is an overall framework diagram of an unsupervised fault detection method and system in an embodiment of the present invention.

Fig. 2 is a schematic diagram of the installation of GNSS antennas and related equipment in the embodiment of the present invention.

Fig. 3 is the network structure of the automatic encoder in the embodiment of the present invention.

Fig. 4 is a schematic diagram of a network structure of a non-fully connected autoencoder in an embodiment of the present invention.

Fig. 5 is a network structure diagram of a voter based on an autoencoder in an embodiment of the present invention.

Detailed ways

In order to enable those skilled in the art to better understand the solutions of the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only It is an embodiment of a part of the present invention, but not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts shall fall within the protection scope of the present invention.

It should be noted that the terms "first" and "second" in the description and claims of the present invention and the above drawings are used to distinguish similar objects, but not necessarily used to describe a specific sequence or sequence. It is to be understood that the data so used are interchangeable under appropriate circumstances such that the embodiments of the invention described herein can be practiced in sequences other than those illustrated or described herein. Furthermore, the terms "comprising" and "having", as well as any variations thereof, are intended to cover a non-exclusive inclusion, for example, a process, method, system, product or device comprising a sequence of steps or elements is not necessarily limited to the expressly listed instead, may include other steps or elements not explicitly listed or inherent to the process, method, product or apparatus.

A fault detection method for an unsupervised self-driving car, comprising the following steps:

S1, collect the driving status data of the self-driving car in real time, use autoencoders with different structures to detect the acquired driving status data, and fuse the detection results of the autoencoders with different structures to obtain the encoder fusion detection results;

Taking the vehicle driving state data in the normal driving state of the self-driving car as the training set, multiple autoencoders with different structures are trained.

S2. At the same time, a class of support vector machine model, local outlier factor model and isolated forest model are used to detect the driving state data of autonomous vehicles to obtain their respective detection results. A class of support vector machine model, local outlier factor model The detection results of the isolation forest model and the encoder fusion detection results are further fused to obtain the final detection results.

The vehicle driving state data acquisition in this application uses sensor information, and specifically selects integrated navigation (GNSS+inertial navigation) information as an example. For integrated navigation information, the sensors used for data acquisition include high-precision MEMS (micro-electromechanical system) integrated navigation system and a set of GNSS antennas for receiving satellite signals, an industrial computer for recording GNSS logs of the integrated navigation system and the supporting power supply equipment for the above devices. The specific installation is shown in Figure 1. The calibrated integrated navigation system and industrial control computer are placed at the rear of the compartment and fixed, and connected to the GNSS antenna through a feeder. The GNSS antennas are respectively screwed onto two strong magnetic chucks and fixedly placed in the forward direction and backward direction of the self-driving car, and placed at the highest point of the self-driving car to ensure good GNSS signals can be received, and at the same time Ensure that the connection line formed by the phase centers of the two GNSS antennas is consistent with or parallel to the central axis of the test carrier.

Data extraction and preprocessing: From the data collected by the above sensors, extract useful fields and perform data cleaning and data transformation to prepare for subsequent model training.

For the data information collected by the sensor, the extracted useful fields include protocol header, sampling time, heading angle, eastward speed and northward speed.

The files extracted from the useful fields still have many problems such as incomplete, missing, and duplicate records, and cannot be directly used as input to the model for training. In the present invention, only one of the repeated records in the extracted useful fields is reserved, and for samples of incomplete fields or wrong fields extracted from the useful fields, the linear interpolation of the corresponding fields of the two adjacent samples before and after the sample is used to complete the incomplete field or error field.

The input data required for model training or verification includes: the velocity and angular velocity of the vehicle in the unmanned state. For data that cannot be obtained directly from the extracted useful fields, it needs to be obtained after transformation using the extracted effective fields. The specific conversion method includes the following steps:

(1) Velocity information is obtained by transforming eastward velocity and northward velocity. Record the eastward speed as v _e and the northward speed as v _n , then the magnitude of the speed value is:

(2) The angular velocity information is obtained by calculating the difference quotient of the yaw angle at two consecutive sampling moments.

Using the speed data and angular velocity data obtained by the above method during the normal operation of the self-driving car, train multiple autoencoders with different structures, and compare the new samples with multiple autoencoders after training, and detect differences with the normal operating state. Data with large deviations. Specifically include the following steps:

S21, training of autoencoder: autoencoder is essentially a feed-forward neural network including encoder structure and decoder structure, as shown in Figure 2, the encoder structure operation of autoencoder is described by formula (3), where W and b are the weight matrix and bias vector of the encoder, respectively.

The input vector X=[x ₁ ,x ₂ ,…x _n ] ^T is encoded as X′=[x′ ₁ ,x′ ₂ ,…x′ _n ] ^T through the encoder structure. The decoder structure operation of the autoencoder is shown in formula (4), where: W' and b' represent the weight matrix and bias vector of the decoder respectively, and the decoder encodes the vector X'=[x' ₁ ,x' ₂ ,…x′ _n ] ^T is decoded as

X'＝f(WX+b) (13)

The goal of the autoencoder is to reconstruct the input data. The goal can be defined by formula (5), and the weight matrix and bias vector parameters are optimized through the backpropagation algorithm.

The data used to train the autoencoder model is the health data of the normal operation state of the self-driving car. After training, the autoencoder is able to reconstruct data that is similar in pattern to the training data. For faulty data, the reconstruction error of autoencoder for this data will be significantly larger than that of normal data. According to the magnitude of the reconstruction error of the input data, it can be judged whether the sample is faulty or not.

S22, Construction of integrated autoencoder: The number of network layers and the number of neurons in each layer have a great influence on the fault detection effect of the autoencoder, and it is difficult to determine the appropriate network scale to achieve the optimal fault detection effect. To solve this problem, multiple autoencoders with different network structures are used to build an integrated autoencoder, and the integrated autoencoder is used to fuse the detection results of the autoencoders with different structures. Considering that a typical autoencoder is composed of fully connected layers, even if the number of network layers is changed, the difference between multiple autoencoders is small due to the same network structure. Therefore, this application uses a non-fully connected autoencoder as a learner to build an ensemble autoencoder. On the basis of the fully connected autoencoder, some connections are randomly disconnected to form a non-fully connected autoencoder, as shown in Figure 3.

Multiple non-fully-connected autoencoders evaluate the input samples separately, and obtain the reconstruction errors of the samples respectively. Note that the reconstruction error of the i-th non-fully-connected autoencoder for the current sample is e _i , then multiple (a group of ) The reconstruction error of the non-fully connected autoencoder for the current sample is E=[e ₁ ,e ₂ ,…e _n ] ^T .

Integrated autoencoder detection result fusion: an integrated autoencoder composed of a group of non-fully connected autoencoders, which gives a set of reconstruction errors for input samples. A group of non-fully connected autoencoders vote on the reconstruction error of the input sample, and finally fuse into a reconstruction error. The voter is composed of the encoder part of the autoencoder. For the reconstruction error vector E=[e ₁ ,e ₂ ,…e _n ] ^T obtained above, it is input to the bottleneck as an automatic neuron as an input vector. Encoder, as shown in Figure 4. After voting, a group of reconstruction errors is fused into one reconstruction error, and the specific operation is described by formula (6), where f _encoder is the encoder operation, and E' is the reconstruction error after fusion. According to the magnitude of the reconstruction error after fusion, it is judged whether the input sample is faulty or not.

Build heterogeneous integration fault detection:

Fault detection algorithms are usually based on specific assumptions, and the detection results of different fault detection algorithms are quite different. In order to further improve the accuracy of fault detection, a heterogeneous integrated fault detection is constructed on the basis of the integrated autoencoder constructed above. Select a class of support vector machine, local outlier factor, isolation forest, and integrated autoencoder as the learner, use the learner to detect in parallel, and then fuse the test results of the learner.

Specific details include:

Using a class of support vector machine (One-Class Support Vector Machine, OCSVM) algorithm to construct a hyperplane with the maximum interval separation between the normal data point and the origin for fault detection, if the data point is within the boundary, it is considered to be normal sample. The objective function of the algorithm is defined as:

where w is the eigenvector of the high-dimensional feature space, _ξi is the slack variable that allows some data points to lie within the margin, ρ is the offset, μ∈(0,1) is the trade-off parameter controlling the boundary, n is the number of samples, _xi is the i-th input training data, and φ( _xi ) is a non-linear mapping function that maps low-dimensional raw data points to high-dimensional. μ sets an upper bound on the score of outliers and a lower bound on the score of support vectors. For the sample χ, the function s _OCSVM to determine whether the sample is a fault is defined by formula (9).

Among them, χ is the sample to be tested, χ _i is the i-th input training data, σ∈R is the parameter that determines the radial range of the function, α _i is a series of supports obtained by using Lagrangian technology in formula (7) vector. If the result of the sample to be tested is negative, it is an abnormal sample.

The process of fault detection using the Isolation Forest algorithm is: recursively and randomly split the data set until all samples are isolated or reach the limited tree height. Faulty data points have shorter paths than normal points due to the feature that faulty data points are sparsely distributed and farther away from denser populations. Given a dataset containing n samples, the average path length of the tree is:

Among them, H(i) is the harmonic number, which can be calculated by ln(i)+γ (Euler's constant), and c(n) is the average value of the path length when the number of samples n is given. The failure score for the sample is:

where h(x) is the number of edges passing through x from the root node of the tree to the leaf node, and E(h(x)) is the expected path length of sample x in a batch of trees.

The Local Outlier Factor (LOF) algorithm calculates the local reachability density of each point, and then calculates the local outlier factor of each point, and selects the n sample points with the highest outlier degree.

The k-th distance d _k ( _xi ) is the distance between the kth point x _j closest to point x _i and point x _i , d _k ( _xi )=d( _xi , x _j ), x _j satisfies: in There are at least k points x′ _j not including _xi in the set, so that d( _xi , x′ _j )≤d( _xi , x _j ); there are at most points not including _xi in the set k-1 points x′ _j , such that d( _xi , o')<d( _xi , x _j ).

The k-th distance neighborhood: the k-th distance neighborhood N _k ( _xi ) of point x _i , the k-th distance of point x _i and the set of all points within the k-th distance.

The k-th reachable distance: rd _k (x _i , x _j )=max{d _k (x _i ),d(x _i , x _j )}, that is, the k-th reachable distance of sample points x _i and x _j is The greater of the k-th distance of point x _i and the distances of points x _i and x _j .

The kth local reachable density:

The fault score s ^LOF of the LOF algorithm is described by formula (13). The larger the fault score is, the more likely the sample point is a fault point.

Combined with the integrated autoencoder constructed by the above method, the failure score of the input data is used as the input of the autoencoder voter, and the input vector s=[s ₁ ,s ₂ ,…,s _n ] ^T where s _i is the i-th The failure score given by the learner. The failure scores of multiple learners form a vector, which is used as the input data to an autoencoder whose bottleneck is a neuron. The autoencoder needs to be pre-trained, and the input of the pre-training stage is the abnormal scores of multiple autoencoders for fault detection, input s=[s ₁ ,s ₂ ,…,s _n ] ^T

是编码器部分的权重矩阵。自动编码器网络中使用的非线性激活函数为ReLu(Rectified Linear Unit)。s′ is the one-dimensional data of the input s encoded by the autoencoder. In order to combine the reconstruction errors of multiple input autoencoders into an abnormal score, the bottleneck of the autoencoder used to combine the reconstruction errors is Set to one neuron, ie s′∈s ¹ .

is the weight matrix of the encoder part. The nonlinear activation function used in the autoencoder network is ReLu (Rectified Linear Unit).

The fused fault score is mapped to the interval (0,1) through the sigmoid function, and the sigmoid function is described as formula (15). Judging whether there is a fault according to the value after mapping, the fault is in the interval (0,0.5), and the fault is in the interval (0.5,1). When a failure of the self-driving car is detected, the system will record the failure data to the log and issue an alarm.

The invention discloses an unsupervised self-driving car fault detection system based on a class of support vector machine (OCSVM), local outlier factor (LOF), isolation forest (IF) and integrated autoencoder. Among them, the integrated autoencoder fault detection module includes multiple autoencoders. The data in the normal driving state of the vehicle is collected as training data, and multiple autoencoders are constructed to detect the vehicle motion state, and the detection results of multiple autoencoders are voted to build an integrated autoencoder. At the same time, the OCSVM, LOF, and IF fault detection modules are constructed using the data of normal driving conditions. The integrated autoencoder, OCSVM, LOF and IF simultaneously monitor the motion state of the vehicle, and vote on the detection results of the vehicle motion state to build an integrated fault detection framework. This framework is mainly aimed at the motion state of autonomous vehicles. From a data-driven perspective, an integrated framework that combines multiple methods for solving fault detection problems is designed, which can effectively detect abnormal sensor data and faults in the operating state of autonomous vehicles.

Claims (10)

1. A fault detection method for an unsupervised automatic driving automobile is characterized by comprising the following steps:

s1, acquiring running state data of an automatic driving automobile in real time, detecting the acquired running state data by using automatic encoders with different structures, and fusing detection results of the automatic encoders with different structures to obtain a coder fusion detection result;

and S2, simultaneously adopting a first-class support vector machine model, a local outlier factor model and an isolated forest model to respectively detect the driving state data of the automatic driving automobile to obtain respective detection results, and further fusing the detection results of the first-class support vector machine model, the local outlier factor model and the isolated forest model with the encoder fusion detection results to obtain a final detection result.

2. The method as claimed in claim 1, wherein the vehicle driving status data of the autonomous vehicle in normal driving status is used as a training set to train a plurality of different automatic encoders.

3. The unsupervised automatic driving vehicle fault detection method according to claim 2, wherein useful fields are extracted from vehicle driving state data in a normal driving state of the automatic driving vehicle and data cleaning and data transformation are performed;

the extracted useful fields include a protocol header, a sampling time, a heading angle, an east speed, and a north speed.

4. A method as claimed in claim 3, wherein the velocity information is obtained from the conversion of east and north velocities, and the east velocity is recorded as v _e North velocity is v _n Then the magnitude of the velocity value is:

angular velocity information is obtained by calculating the difference quotient of the yaw angles at two consecutive sampling moments:

5. the unsupervised autopilot vehicle fault detection method of claim 4 wherein the automatic encoder comprises an encoder structure and a decoder structure, the encoder structure is described by equation (3), W and b are the weight matrix and offset vector of the encoder, respectively;

input vector X = [ X ] ₁ ,x ₂ ,…x _n ] ^T Encoded by the encoder structure as X' = [ X = ₁ ′,x ₂ ′,…x _n ′] ^T (ii) a The decoder architecture operates as equation (4), W 'and b' representing the weight matrix and offset vector, respectively, of the decoder that will encode the vector X '= [ X' ₁ ,x′ ₂ ,…x′ _n ] ^T Decode into

X'＝f(WX+b) (3)

The goal of the auto-encoder is to reconstruct the input data, with the goal of optimizing the weight matrix and offset vector parameters by the back-propagation algorithm, as in equation (5).

6. The method of claim 5, wherein the automatic encoder is a non-fully-connected automatic encoder.

7. The method of claim 6, wherein the plurality of non-fully connected automatic encoders respectively evaluate the input samples to respectively obtain reconstruction errors of the samples, and the reconstruction error of the ith non-fully connected automatic encoder on the current sample is recorded as e _i Then the reconstruction error of the current sample by the multiple non-fully connected automatic encoders is E = [ E ] ₁ ,e ₂ ,…e _n ] ^T 。

8. The method of claim 1, wherein fault detection is performed by constructing a hyperplane with maximum separation between a normal data point and an origin using a Class-Support Vector Machine (OCSVM) algorithm, and the data point is considered to be a normal sample if the data point is within a boundary.

9. The method as claimed in claim 1, wherein the fused fault score is mapped to the interval (0, 1) by a sigmoid function, and whether a fault occurs is determined according to the mapped value, wherein the fault occurs in the interval (0, 0.5) and the fault occurs in the interval (0.5, 1).

10. A fault detection system of an unsupervised automatic driving automobile is characterized by comprising a data acquisition module and a detection module;

the data acquisition module is used for acquiring the driving state data of the automatic driving automobile in real time and transmitting the acquired data to the detection module;

the detection module stores automatic encoders with different structures, a first-class support vector machine model, a local outlier factor model and an isolated forest model, the automatic encoders with different structures are used for detecting the acquired driving state data, and detection results of the automatic encoders with different structures are fused to obtain an encoder fusion detection result; meanwhile, a first-class support vector machine model, a local outlier factor model and an isolated forest model are adopted to respectively detect the driving state data of the automatic driving automobile to obtain respective detection results, and the detection results of the first-class support vector machine model, the local outlier factor model and the isolated forest model are further fused with the encoder fusion detection results to obtain a final detection result.

Images