A method and system for detecting faults in unmanned vehicles

Abstract

This invention discloses an unmanned vehicle fault detection method and system. It utilizes positive example data from unmanned vehicles to construct and train a vehicle state monitor and a vehicle trajectory yaw monitor. GNSS data of the unmanned vehicle under test is acquired in real time. Field extraction, data cleaning, and data transformation are performed on the acquired GNSS information. This invention employs various data processing methods, splitting the unmanned vehicle fault diagnosis system into a state monitor and a trajectory yaw monitor. Corresponding fault detection methods are designed and implemented separately for each module, resulting in better system performance. A complete unmanned vehicle fault detection and error type inference system based on sensor GNSS information is constructed, which can effectively monitor the vehicle state and trajectory of the unmanned vehicle, provide early warnings when the unmanned vehicle state is abnormal, and infer possible error types.

Description

Technical Field

[0001] The present invention belongs to the field of system fault diagnosis, and relates to the processing of GNSS information in the integrated navigation system of driverless vehicles. Specifically, it relates to a method and system for fault detection of driverless vehicles. Background Art

[0002] Driverless technology integrates high-tech technologies in many fields such as sensor technology, computer technology, and information processing technology. Due to its huge development potential in the transportation field, its development prospect is generally看好. However, the safety issue of autonomous vehicles has一直阻碍着the large-scale commercial implementation of driverless technology. Therefore, designing a fault detection and diagnosis system配套with driverless vehicles is of crucial significance for the development of the autonomous driving industry.

[0003] Existing fault diagnosis methods for driverless vehicles are mostly migrated from fault diagnosis methods in traditional industrial processes, mainly including model-based methods, signal-based methods, and history data-based methods. However, for the complex system of driverless vehicles, it is difficult for a single method to effectively diagnose faults of driverless vehicles, and autonomous driving has high requirements for the real-time性 of data processing, and traditional methods are also difficult to effectively solve this problem. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and system for fault detection of driverless vehicles to overcome the problems of low fault diagnosis accuracy and low real-time data processing of existing driverless vehicles. The present invention can timely and accurately实现 the detection of vehicle status.

[0005] A method for fault detection of driverless vehicles includes the following steps:

[0006] S1, respectively construct and train a vehicle status monitor and a vehicle trajectory yaw monitor by using positive example data of the driverless vehicle;

[0007] S2,实时获取the GNSS data of the待测 driverless vehicle, and perform field extraction, data cleaning, and data transformation on the obtained GNSS information;

[0008] S3, use the vehicle status monitor and the vehicle trajectory yaw monitor to judge the transformed data, so as to obtain whether the driverless vehicle is faulty.

[0009] Preferably, in S1, based on a data acquisition device, obtain the GNSS data of the待测 driverless vehicle, perform field extraction, data cleaning, and data transformation on the obtained GNSS information, and the extracted useful fields include: protocol header, sampling time, latitude, longitude, heading angle, eastward speed, and northward speed.

[0010] It should be noted that there are some inaccuracies in the original text such as "一直阻碍着" which should be "has been hindering", "配套" which should be "matched with", "实时获取" which should be "Obtain in real time", and "待测" which should be "to-be-tested". These have been corrected in the translation.Preferably, incomplete records in the extracted useful fields are directly removed; duplicate records are removed, leaving only one; and complete records with non-standard formats are standardized.

[0011] The cleaned data is transformed to obtain the input data needed for model training or validation: the (X, Y) coordinates, velocity, and angular velocity of the autonomous vehicle in the Cartesian coordinate system at each sampling time in the autonomous driving state.

[0012] Preferably, the latitude and longitude data extracted from the field are used to obtain the (X, Y) coordinates in the rectangular coordinate system through Mercator projection;

[0013] Speed ​​information is obtained by transforming eastward and northward speeds:

[0014] Let the eastward velocity be v. e The northbound velocity is v n The value of the speed information is:

[0015]

[0016] Angular velocity information is obtained by calculating the difference between the heading angles at two consecutive sampling times.

[0017] Preferably, the obtained velocity and angular velocity information is used as training data, and the OCSVM method is adopted to obtain the decision boundary for determining whether the vehicle state is abnormal during autonomous driving, and a vehicle state monitor is constructed.

[0018] Preferably, the acquired velocity and angular velocity information are used as inputs to train the model to obtain the decision boundary for determining whether the vehicle state is abnormal during autonomous driving.

[0019] The decision boundary is generalized, and an alarm triggering mechanism is set based on the generalized boundary conditions.

[0020] Preferably, a vehicle trajectory yaw monitor is constructed using a Kalman filter based on the vehicle kinematics model and the extracted sampling time information, (X, Y) coordinate information and heading angle information.

[0021] Preferably, a one-step state transition matrix for the state prediction process in Kalman filtering is constructed based on the vehicle kinematics model;

[0022] Based on the obtained one-step state transition matrix, a Kalman filter for predicting vehicle trajectories is constructed.

[0023] Preferably, based on the positive example data and error sample data extracted from the useful fields, artificially injected error information data is also added; the above positive example data, error sample data and artificially injected error information data are used as training data samples for probabilistic inference to train an artificial neural network optimized fuzzy system, and the artificial neural network optimized fuzzy system is used to perform probabilistic inference on the causes of vehicle state errors to obtain the probability of the cause of the unmanned vehicle failure.

[0024] An unmanned vehicle fault detection system, comprising a data acquisition module and a detection module;

[0025] The detection module is used to store the vehicle state monitor and vehicle trajectory yaw monitor, which are constructed and trained using positive example data from the unmanned vehicle, respectively.

[0026] The data acquisition module acquires GNSS data of the unmanned vehicle under test in real time, performs field extraction, data cleaning and data transformation on the acquired GNSS information, and transmits the transformed data to the detection module. The detection module judges the transformed data and outputs the fault information of the unmanned vehicle.

[0027] Compared with the prior art, the present invention has the following beneficial technical effects:

[0028] This invention discloses a fault detection method for unmanned vehicles. It constructs and trains a vehicle state monitor and a vehicle trajectory yaw monitor using positive example data from unmanned vehicles. GNSS data of the unmanned vehicle under test is acquired in real time. The acquired GNSS information undergoes field extraction, data cleaning, and data transformation. This invention employs various data processing methods, splitting the unmanned vehicle fault diagnosis system into a state monitor and a trajectory yaw monitor. Corresponding fault detection methods are designed and implemented separately for each module, resulting in better system performance. A complete unmanned vehicle fault detection and error type inference system based on sensor GNSS information is constructed, effectively monitoring the vehicle state and trajectory, providing early warnings and inferring possible error types when the unmanned vehicle state is abnormal.

[0029] The collected samples are processed using the OCSVM algorithm to train the decision boundary for judging whether the vehicle driving state is abnormal, and a state diagnosis system is constructed. A Kalman filter model is used as the observer, and the residual distribution is inferred by Jaque-Bera test with the measured data. At the same time, interpolation techniques are used to construct a real-time vehicle trajectory monitoring system, which can accurately judge the acquired data.

[0030] This invention establishes an artificial neural network model based on collected vehicle status data, which can make probabilistic inferences on the types of detected vehicle status errors and achieve targeted and rapid detection of faults in autonomous vehicles.

[0031] This invention can provide probabilistic inferences about the types of vehicle status errors, which can provide some guidance for further investigation of errors. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the installation of the GNSS antenna and related equipment in an embodiment of the present invention.

[0033] Figure 2 This is a flowchart of a specific method in an embodiment of the present invention.

[0034] Figure 3 This is a schematic diagram illustrating the generalization of the decision boundary after obtaining it based on OCSVM in an embodiment of the present invention.

[0035] Figure 4 This is a schematic diagram of a linear vehicle bicycle model in an embodiment of the present invention.

[0036] Figure 5 This is a schematic diagram of the modified fully connected neural network in an embodiment of the present invention. Detailed Implementation

[0037] 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.

[0038] This invention discloses a fault detection method for unmanned vehicles, based on GNSS data from an integrated navigation system for fault detection and error type probability inference. The data acquisition device includes a high-precision MEMS integrated navigation system installed inside the vehicle, a GNSS antenna for receiving satellite signals, an industrial control computer for recording the GNSS logs of the integrated navigation system, and a power supply device. The power supply can be connected to the vehicle's onboard power supply or use an independent generator. The data acquisition device is fixedly installed on the unmanned vehicle under test.

[0039] Data acquisition device installation can be as follows Figure 1 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 fixedly positioned in the forward and backward directions of the test vehicle, respectively. The GNSS antenna is placed at the highest point of the test vehicle to ensure good GNSS signal reception, while ensuring 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.

[0040] like Figure 2 As shown, this invention discloses a method for fault detection and error type probability inference of unmanned vehicles based on GNSS data from an integrated navigation system. This application includes the following steps from model training to fault detection and judgment output:

[0041] S1, GNSS data acquisition: Real-time acquisition of GNSS data of the unmanned vehicle under test based on the data acquisition device;

[0042] In autonomous driving mode, the industrial control computer reads the real-time GNSS information of the unmanned vehicle from the high-precision integrated navigation system via an RS232 serial data cable and writes the acquired GNSS information to a log file as collected data. The sampling interval is 10ms, and the serial port baud rate is set to 230400 baud per second.

[0043] S2 involves preprocessing the collected GNSS information to prepare it for subsequent model training. The preprocessing steps include useful field extraction, data cleaning, and data transformation. The transformed data is then used for model training. Specifically, this includes the following steps:

[0044] S21. Extract useful fields from GNSS information: The industrial control computer reads GNSS information from the data acquisition device via an RS232 serial port. The log file is a CSV (comma-separated values) file, where each record includes vehicle-related information at the sampling time. Useful fields needed for training the model are extracted from the CSV file, specifically: protocol header, sampling time, latitude, longitude, heading angle, eastward velocity, and northward velocity. The extracted useful fields are saved in CSV file format.

[0045] S22, Data Cleaning: The extracted useful field files still have many problems and cannot be directly used as input for model training. These problems include incomplete records, missing records, duplicates, and some records having non-standard formats. Therefore, it is necessary to directly remove incomplete records; remove duplicate parts from duplicate records, keeping only one record; and standardize the format of complete records with non-standard formats.

[0046] S23, Data Transformation: The input data required for model training or validation includes the (X, Y) coordinates, velocity, and angular velocity of the autonomous vehicle in the Cartesian coordinate system at each sampling time in autonomous driving mode. These attributes cannot be directly obtained from the extracted and cleaned fields; therefore, the extracted and cleaned field data is transformed to obtain the input data required for model training or validation.

[0047] The specific transformation method includes the following steps:

[0048] (1) Using the latitude and longitude data extracted from the fields, the (X, Y) coordinates in the rectangular coordinate system are obtained through Mercator projection, with the origin of the (X, Y) coordinates being the vehicle's starting point. This application utilizes the geographic feature coordinate projection function mfwdtran() integrated into the MATLAB R2016a platform, selecting Mercator projection as the projection method and the vehicle's starting point latitude and longitude data as the origin. By taking the latitude and longitude data as input, the corresponding (X, Y) coordinates can be obtained.

[0049] (2) The velocity information is obtained by transforming the eastward velocity and the northward velocity:

[0050] Let the eastward velocity be v. e The northbound velocity is v n The value of the speed information is:

[0051]

[0052] (3) Angular velocity information is obtained by calculating the difference between the heading angles at two consecutive sampling times:

[0053] Considering that heading angles of 360° and 0° actually represent the same heading angle, but numerically they appear as abrupt changes, phase unwrapping is used in MATLAB R2016a to avoid such abrupt changes between two consecutive time frames before calculating the difference quotient. The calculation formula is as follows:

[0054]

[0055] ω k t represents the angular velocity at time k. k , t k-1 Represents two consecutive sampling times. These represent the heading angles at time k and time k-1, respectively.

[0056] S3 uses the velocity and angular velocity information obtained in S2 as training data and adopts the OCSVM (One-Class Support Vector Machine) method to obtain the decision boundary for determining whether the vehicle state is abnormal during autonomous driving and to construct a vehicle state monitor.

[0057] Specifically, the following steps are included:

[0058] S31. The acquired velocity and angular velocity information are used as inputs to train the model. The aforementioned velocity and angular velocity information are samples collected when the vehicle is in an unmanned driving state and has not passed through tunnels or other areas that affect GNSS information reception; these are positive example data.

[0059] The objective function of the model is expressed as:

[0060]

[0061] w is a feature vector in a high-dimensional feature space, which is a representation of the input vector in a high-dimensional space after mapping; ξ i It is a slack variable that affects how many data points will exist outside the decision boundary; ρ is the bias; μ∈(0,1) is the profit / loss coefficient controlling the decision boundary; n is the number of samples; x i φ(x) represents the i-th input vector. i ) represents a nonlinear mapping function that can map an input vector in a low-dimensional space to a high-dimensional space.

[0062] Define the state vector of the vehicle in its autonomous driving state as the input vector, expressed as:

[0063]

[0064] Here, v is the instantaneous velocity of the vehicle, reflecting its longitudinal state, and ω is the instantaneous angular velocity of the vehicle, reflecting its lateral state. The nonlinear mapping function is represented by a kernel function (K(x,x)). i )=φ(x) T φ(x i K (the kernel function) simplifies the mapping process. Here, the Gaussian radial basis function (RBF) is chosen as the kernel function:

[0065]

[0066] ||x-x'|| is the Euclidean distance between vector x and center point x', and σ determines the size of the radial range of the kernel function.

[0067] Substituting (4) and (5) into (3) and using the Lagrange multiplier method to solve the minimization problem of (3), we can obtain the decision function for determining the state of the autonomous vehicle:

[0068]

[0069] Here α i That is, the Lagrange multipliers of each input vector can also be seen as the weights of each input sample. For simplicity, we can assume that each sample is equally important, and the α of all samples... i All values ​​are set to 1; χ² is a sample to be tested. i For the i-th training data (χ) i (Do not take χ).

[0070] Set parameters μ = 0.01, ρ = 0, and all ξ i=0.1, σ=0.1, and the sample data is fed into the model for training, so that the decision boundary for determining whether the vehicle state is abnormal during autonomous driving can be obtained.

[0071] S32, Generalizing the Decision Boundary: After obtaining the decision boundary in S31, since the decision boundary closely follows the edge of the data point distribution, directly using this boundary will lead to a high false alarm rate. Therefore, it is necessary to extend the boundary outwards to a certain extent, conformally, based on the actual situation, on the basis of the original hard boundary. According to the Newman-Pearson criterion, the extension can be stopped when the false alarm rate β0 is lower than the set threshold (here, β0 is set to 0.1). The boundary extension is illustrated below. Figure 3 As shown.

[0072] S33, Alarm Trigger Mechanism Design: Due to various interferences and noise, even under normal circumstances, there may be sample points falling outside the extended boundary. In order to reduce the false alarm rate, the alarm trigger mechanism is designed as follows: within 2 seconds (i.e., 200 sampling cycles), if more than 100 sample points fall outside the extended boundary, an alarm for abnormal vehicle status will be triggered.

[0073] S4, based on the vehicle kinematics model and the sampling time information, (X, Y) coordinate information and heading angle information extracted from S2, uses a Kalman filter to construct a vehicle trajectory yaw monitor.

[0074] Specifically, the following steps are included:

[0075] S41, based on the vehicle kinematics model, construct the one-step state transition matrix for the state prediction process in Kalman filtering.

[0076] The one-step state transition matrix is ​​the core of the Kalman filter prediction process, and this invention obtains this transition matrix based on the vehicle kinematics model. For example... Figure 4 As shown, in the rectangular coordinate system XOY, using x, y, The motion of the vehicle is characterized by three degrees of freedom. The state vector of the linear system is defined. The discrete form of the vehicle-bicycle model can be described as follows:

[0077]

[0078] Δt represents the sampling interval, δ represents the front wheel steering angle, v represents the longitudinal velocity of the vehicle, l represents half of the front and rear wheelbase, and Ω(k) represents the vehicle's motion state at time k, which has the relationship with the vehicle's state at time k-1 as described in equation (7). From the extracted GNSS information, it is not possible to directly obtain... This information per degree of freedom, but the yaw angle and There is a simple linear correspondence between them. If the vehicle's starting point is taken as the origin of the coordinate system, the two differ numerically by the initial heading angle phase at the starting point. Therefore, as long as the initial heading angle of the vehicle at the starting point is measured, This information about the degree of freedom can then be obtained.

[0079] Treating the vehicle as a point mass to predict its trajectory, it can be represented on a plane using the X and Y coordinates. After obtaining... After obtaining this degree of freedom information, equation (7) can be simplified to:

[0080]

[0081] here,

[0082]

[0083] (8) The coefficient matrix in equation (8) is the one-step state transition matrix of the Kalman filter in this invention, denoted as F.

[0084] S42, based on the one-step state transition matrix obtained in S41, constructs a Kalman filter to predict the vehicle trajectory. The calculation process is as follows:

[0085] Based on the previous state, a priori state prediction is given:

[0086] Ω(k|k-1)=FΩ(k-1|k-1)+BU(k) (9)

[0087] In equation (9), Ω(k|k-1) is the result predicted based on the previous state, which in this invention represents the prior prediction of the current position; F is the one-step state transition matrix obtained in S41; Ω(k-1|k-1) is the optimal result of the previous state; U(k) is the state control quantity, and B is the system parameter. Since there is no additional input control quantity, U(k) = 0.

[0088] Prior prediction of the input data covariance matrix:

[0089] P(k|k-1)=FP(k-1|k-1)F′+Q (10)

[0090] In equation (10), P(k|k-1) is the covariance matrix corresponding to Ω(k|k-1); P(k-1|k-1) is the covariance matrix corresponding to Ω(k-1|k-1); F′ is the transpose of F; Q is the covariance matrix corresponding to the process noise, which can be taken as...

[0091]

[0092] In this invention, the initial value of P is taken as the identity matrix.

[0093] Calculation of Kalman gain:

[0094] K g (k)=P(k|k-1)H′ / (HP(k|k-1)H′+R) (11)

[0095] Where K g (k) is the Kalman gain at time k, H is the observation matrix, which is set as the identity matrix in this invention, and R is the observation noise covariance matrix, which can take the same value as Q.

[0096] Updated state estimate:

[0097] Ω(k|k)=Ω(k|k-1)+K g (k)(Z(k)-HΩ(k|k-1)) (12)

[0098] Where Z(k) is the sensor's measurement of the vehicle's position at time k.

[0099] Update the estimate of the input data covariance matrix:

[0100] P(k|k=(IK) g (k)H)P(k|k-1) (13)

[0101] In the formula, I is the identity matrix.

[0102] Once an initial value is given, the above five steps can be used to iterate continuously and obtain the corrected trajectory prediction information.

[0103] S43 involves inferring the residual distribution between the trajectory prediction information after Kalman filtering correction and the actual sensor measurements to monitor whether the vehicle trajectory has deviated. If there are enough noise sources that are independent and identically distributed, the noise process is a zero-mean Gaussian process according to the Central Limit Theorem. In reality, there are many noise sources that are independent of each other, and we can assume that the assumption of "enough noise sources that are independent and identically distributed" is satisfied. Therefore, the residual between the predicted value and the actual observation value of the Kalman filter should approximately conform to a zero-mean Gaussian distribution. If the residual distribution differs significantly from the zero-mean Gaussian distribution, the trajectory may have deviated.

[0104] This invention utilizes the Jarque-Bera test to examine the normality of the residual distribution. Assuming a sample comes from normally distributed data, its expected skewness is 0 and kurtosis is 3. The normality of the sample is tested by assessing the fit between the sample's skewness and kurtosis and the normal distribution. The test statistic is denoted as JB and defined as follows:

[0105]

[0106] In equation (14), n is the degrees of freedom. If we assume that each sample is independent, then n is equal to the number of samples; S represents skewness, and K represents kurtosis. The formulas are as follows:

[0107]

[0108] here and These are the third and fourth center moments of the samples, respectively. It is the sample mean. It is the sample standard deviation. If the test statistic JB deviates significantly from 0, it indicates that the tested sample does not conform to a normal distribution.

[0109] Cubic spline interpolation expands the sample size for residual distribution inference: Distribution inference generally requires a large sample size to ensure a certain confidence level, but autonomous driving has high real-time requirements, making it impossible to collect a large number of samples in a short time. To balance sample size and real-time performance, this invention employs cubic spline interpolation to appropriately expand the number of samples collected in a short period. The residual distribution inference is also based on the interpolated short-term collected samples, specifically including the following steps:

[0110] The data collected every 2 seconds (200 sampling periods) and the predicted values ​​of the Kalman filter are first subjected to residual operation. The residuals are then subjected to cubic spline interpolation in segmented intervals. The interpolated data is then tested for normality using Jarque-Bera detection. If three consecutive Jarque-Bera detection samples fail to pass, a trajectory yaw alarm is triggered.

[0111] Artificial neural networks are used to optimize the membership function of a fuzzy system, enabling it to probabilistically infer the causes of vehicle state errors. Specifically, the steps include:

[0112] Collection and creation of training data for probabilistic inference: There are three sources of training data in this step, such as... Figure 2As shown, in addition to utilizing the same positive example data as in S31, error sample data is also acquired in real time, and error information is manually injected as training data samples for probabilistic inference. The actual acquired error sample data refers to samples collected when the unmanned vehicle passes through a short tunnel (at this time, although GNSS messages are abnormal or even unreceived, other sensors, such as inertial navigation, will ensure the unmanned vehicle's safe passage through the short tunnel). After data preprocessing similar to S2, velocity and angular velocity information are extracted. The manually injected error information refers to artificially adding high-power Gaussian white noise to the positive example samples, causing a certain degree of distortion. All positive example samples are labeled 1, and the collected error samples and manually injected error information samples are labeled 0. Then, all training samples are randomly ordered to avoid the adverse effects of consecutive labels on training. The method is as follows: a sequence containing only elements 0 and 1, with the same number of samples, is randomly generated. The row containing element 1 stores the information of positive examples, and the row containing element 0 stores the information of error samples. 70% of all samples are used as the training set, and the remaining 30% are used as the test set.

[0113] Initialization of the membership function in a fuzzy system: This membership function consists of two functions. One reflects the change in the probability of longitudinal state error as the vehicle's speed *v* changes, since the longitudinal state is reflected by speed *v*. The other reflects the change in the probability of lateral state error as the vehicle's angular velocity *ω* changes, since the lateral state is reflected by angular velocity *ω*. To give the membership function probabilistic properties, its value should be within the range of 0 to 1. Define the "speed runaway" membership function M(v):

[0114]

[0115] "Heading loss" membership function M(ω):

[0116]

[0117] Here, μ v ,μ ω It refers to the velocity and angular velocity thresholds. In a certain state, if the velocity is greater than μ... v Or angular velocity greater than μ ω This means that the vehicle is considered to be in an abnormal state and may be out of control; σ v and σ ω These parameters determine the steepness of the transition region (the region around an error probability of 0.5) of the membership function; the larger the value, the steeper the transition. All four parameters are dynamically changing and will vary depending on the vehicle's state, as will be further demonstrated in later steps. During initialization, μ can be set... v =16.67m / s, μ ω =25° / s, σv =σ ω =0.5.

[0118] S53, Construction of an Artificial Neural Network Optimized Fuzzy System: System structure as follows Figure 5 As shown, it includes an input layer, a membership function layer, a hidden layer, and an output layer;

[0119] The output layer has two nodes, corresponding to the two components of the input feature: velocity v and angular velocity ω. Each of the two hidden layers has six nodes, and their activation function is the sigmoid function. The output layer also has two nodes, and the target output is in one-hot format, meaning that if the input is a normal sample, the expected output is [1,0]. T If the input is an abnormal sample, the expected output is [0,1]. T The membership function layer is a variant of the hidden layer, and the activation function uses the membership function from S52, which is equivalent to performing a feature transformation on the input features. Here, the connection between the input layer and the hidden layer only serves to transfer features; the weights w and biases b of the connections between nodes in all subsequent layers are randomly initialized at the beginning and continuously corrected as training progresses. Backpropagation (BP) is used during training, and gradient descent is employed during the backpropagation process. Training stops when the mean squared error of backpropagation is less than 1e-3.

[0120] Black-box testing updates the parameters of the membership function. The information of a well-trained artificial neural network optimizing a fuzzy system is reflected in the weights and biases of each node's connections. The purpose of this step is to obtain a membership function form equivalent to the trained network through approximate fitting; in other words, to update the membership function parameters in S52. The trained network is treated as a black box, and the parameter update process is as follows:

[0121] The velocity v is limited to a range of 0–20, and the angular velocity ω is limited to a range of 0–50. Both are equally spaced with a step size of 0.1 to obtain a series of nodes {v}. i |i=1,2,…,201},{ω j |j=1,2,…,501}, the nodes of the two dimensions are paired to form a series of state vectors χ. i =[v i ,ω j ] T This serves as the input to the black box; the output of the black box is a series of one-hot values ​​indicating whether an anomaly has occurred, i.e., [1, 0]. T This represents normal, [0,1] T This represents an anomaly. Following the dimension of v, with a fixed v value each time, the search output consists of [1, 0]. TIt becomes [0,1] T The corresponding angular velocity ω is stored, so that each v will correspond to an ω value that causes the state to change. This ω value is the threshold μ in equation (16). ω As can be seen, v and μ ω There is a one-to-one correspondence, and this relationship can be fitted using a function, that is:

[0122]

[0123] Here, g is the function to be fitted, max(μ) ω ) represents the maximum value among the saved angular velocity thresholds, v mid To make the function g take the maximum value of μ ω The velocity value at () / 2. Similarly, along the dimension of ω, fixing one ω value each time, the search output is [1,0]. T It becomes [0,1] T The corresponding angular velocity v is stored, so that each ω will correspond to a v value that causes the state to change. This v value is the threshold μ in equation (15). v ω and μ v Similarly, it can be fitted using a functional relationship as follows:

[0124]

[0125] Where f is the function to be fitted, max(μ) v ) represents the maximum value among the saved speed thresholds, v mid To make the function f take the maximum value of μ v The velocity value at ) / 2.

[0126] After obtaining g(v) and f(ω), the other two parameters σ can be obtained through equations (15) and (16). v and σ ω Update. When the velocity reaches a certain upper bound v up Or the angular velocity reaches a certain upper bound ω up At that time, it is assumed that the vehicle must have malfunctioned, and the probability of error is infinitely close to 1, that is:

[0127]

[0128] Let M(v) = 1 - ε1, M(ω) = 1 - ε2, where ε1 and ε2 are both very small positive numbers, then we can obtain:

[0129]

[0130]

[0131] At this point, all four parameters of the membership function have been updated, given any vehicle state χ = [v, ω]. T The probability of v causing an error (or the probability that the error is caused by a vertical problem) and the probability of ω causing an error (or the probability that the error is caused by a horizontal problem) can both be calculated from the updated membership function.

[0132] This invention employs a Kalman filter model as the observer and uses the Jaque-Bera test to infer the residual distribution of the measured data. Simultaneously, interpolation techniques are used to construct a real-time vehicle trajectory monitoring system. Based on the collected vehicle state data, an improved artificial neural network model, and black-box testing technology, a fuzzy system model is designed to probabilistically infer the types of detected vehicle state errors. This invention utilizes various data processing techniques to construct a relatively complete unmanned vehicle fault detection and error type inference system based on sensor GNSS information. It can effectively monitor the unmanned vehicle's state and trajectory, providing early warnings and inferring possible error types when the unmanned vehicle's state is abnormal. By combining the advantages of model-based fault diagnosis methods and data-driven methods, the designed fault diagnosis system is more targeted towards autonomous vehicles. Probabilistic inferences of vehicle state error types can provide guidance for further error investigation.

Claims

1. A method for detecting faults in unmanned vehicles, characterized in that, Includes the following steps: S1. Using positive example data from unmanned vehicles, a vehicle state monitor and a vehicle trajectory yaw monitor are constructed and trained respectively. The GNSS data of the unmanned vehicle under test is acquired by the data acquisition device. The acquired GNSS information is then processed by field extraction, data cleaning and data transformation. The extracted useful fields include: protocol header, sampling time, latitude, longitude, heading angle, eastward speed and northward speed. S2 acquires GNSS data of the unmanned vehicle under test in real time, and performs field extraction, data cleaning and data transformation on the acquired GNSS information; Incomplete records in the extracted useful fields are directly removed; duplicate records are removed, keeping only one; and complete records with non-standard formats are standardized. The cleaned data is transformed to obtain the input data needed for model training or validation: the (X, Y) coordinates, velocity, and angular velocity of the autonomous vehicle in the Cartesian coordinate system at each sampling time in the autonomous driving state; Using the latitude and longitude data extracted from the fields, the (X, Y) coordinates in the rectangular coordinate system are obtained through Mercator projection; Speed ​​information is obtained by transforming eastward and northward speeds: Let the eastward velocity be v. e The northbound velocity is v n The value of the speed information is: Angular velocity information is obtained by calculating the difference quotient of heading angles at two consecutive sampling times; Using the obtained velocity and angular velocity information as training data, the OCSVM method is adopted to obtain the decision boundary for determining whether the vehicle state is abnormal during autonomous driving, and a vehicle state monitor is constructed. S3 uses the vehicle status monitor and vehicle trajectory yaw monitor to judge the transformed data, thereby determining whether the unmanned vehicle is malfunctioning.

2. The method for detecting faults in an unmanned vehicle according to claim 1, characterized in that, The acquired velocity and angular velocity information are used as inputs to train the model and obtain the decision boundary for determining whether the vehicle state is abnormal during autonomous driving. The decision boundary is generalized, and an alarm triggering mechanism is set based on the generalized boundary conditions.

3. The method for detecting faults in an unmanned vehicle according to claim 1, characterized in that, Based on the vehicle kinematics model and the extracted sampling time information, (X, Y) coordinate information and heading angle information, a vehicle trajectory yaw monitor is constructed using a Kalman filter.

4. The method for detecting faults in an unmanned vehicle according to claim 3, characterized in that, Based on the vehicle kinematics model, a one-step state transition matrix for the state prediction process in Kalman filtering is constructed; based on the obtained one-step state transition matrix, a Kalman filter for predicting vehicle trajectory is constructed.

5. The method for detecting faults in an unmanned vehicle according to claim 1, characterized in that, Based on the positive example data and error sample data extracted from the useful fields, and with the addition of manually injected error information data, the positive example data, error sample data, and manually injected error information data are used as training data samples for probabilistic inference. An artificial neural network optimized fuzzy system is trained, and the artificial neural network optimized fuzzy system is used to perform probabilistic inference on the causes of vehicle state errors, thereby obtaining the probability of the cause of the unmanned vehicle malfunction.

6. An unmanned vehicle fault detection system for the unmanned vehicle fault detection method of claim 1, characterized in that, Includes a data acquisition module and a detection module; The detection module is used to store the vehicle state monitor and vehicle trajectory yaw monitor, which are constructed and trained using positive example data from the unmanned vehicle, respectively. The data acquisition module acquires GNSS data of the unmanned vehicle under test in real time, performs field extraction, data cleaning and data transformation on the acquired GNSS information, and transmits the transformed data to the detection module. The detection module judges the transformed data and outputs the fault information of the unmanned vehicle.

Images