Neural network fault detection method based on generalized maximum relevance entropy Kalman filter

Abstract

The present invention relates to a neural network fault detection method based on generalized maximum correlation entropy Kalman filtering, which belongs to the field of integrated navigation technology, and includes: establishing a discrete state equation and an observation equation of INS error of a GNSS / INS tightly integrated navigation system; using generalized maximum correlation entropy Kalman filtering to perform INS recursive error estimation and recursive error correction; collecting a pseudo-range time series in the filtered new information data of a single satellite for normalization; performing GASF coding and MTF coding processing on the normalized time series to obtain a GASF image and an MTF image; establishing a neural network model for satellite normal and fault state classification; using GASF images and MTF images to train the neural network model respectively, and then performing feature fusion on the classification probability of satellite normal and fault state classification according to the GASF image and the MTF image to obtain the final classification result of the neural network model; and using the trained neural network model to detect single satellite faults in the positioning process. The present invention improves the fault detection accuracy.

Description

Technical Field

[0001] The invention relates to the technical field of integrated navigation, and in particular to a neural network fault detection method based on generalized maximum relevant entropy Kalman filtering. Background Art

[0002] With the development of technologies such as drones and autonomous driving, the market is paying more and more attention to the accuracy and reliability of navigation systems. Among them, integrated navigation has achieved good results in accuracy and reliability, especially tight integrated navigation is frequently used. During use, in order to further ensure the reliability, integrity and positioning accuracy of the system, fault detection and troubleshooting must be taken seriously.

[0003] In the past, the detection of combined navigation faults was mainly carried out by model methods, such as residual chi-square test, autonomous integrity monitoring extrapolation method, etc. Due to the limitations of the model itself, the accuracy and real-time performance of fault identification will be restricted. In addition, there is also a signal analysis method, which uses wavelet transform to distinguish signal faults and noise. However, the detection method based on signal analysis has a simple system design, which will result in poor results. Adding auxiliary modules will make the system complex and the detection time longer. Therefore, we need to consider new methods to obtain a method with a simple system, short detection time and high detection accuracy.

[0004] With the development of data-driven technology, the method of using neural networks for combined navigation fault detection has gradually been explored and applied. The application of neural networks makes our vision possible. Summary of the invention

[0005] In view of the above analysis, the present invention aims to disclose a neural network fault detection method based on generalized maximum relevance entropy Kalman filtering to improve the fault detection accuracy of the integrated navigation system.

[0006] The present invention discloses a neural network fault detection method based on generalized maximum relevant entropy Kalman filtering, comprising:

[0007] Step S1, establishing the discrete state equation and observation equation of the INS error of the GNSS / INS tightly integrated navigation system; using the generalized maximum relevant entropy Kalman filter to perform INS recursive error estimation and recursive error correction;

[0008] Step S2, collecting the pseudorange time series in the filtered new information data of a single satellite, and normalizing it to obtain a normalized time series;

[0009] Step S3, performing GASF coding and MTF coding processing on the normalized time series to obtain a GASF image and an MTF image;

[0010] Step S4, establishing a neural network model for satellite normal and fault state classification; after using GASF images and MTF images to train the neural network model respectively, the classification probabilities of satellite normal and fault state classification based on GASF images and MTF images are subjected to feature fusion to obtain the final classification result of the neural network model;

[0011] Step S5: using the trained neural network model to detect single satellite failures during the positioning process of the integrated navigation system.

[0012] Further, the neural network model for classifying normal and faulty states of satellites includes a first classification neural network, a second classification neural network and a feature fusion network;

[0013] The first classification neural network is used to classify the input GASF image into normal and faulty satellite states and output the classification probability. and in, To judge the probability that the input GASF image is in normal satellite status; To determine the probability that the input GASF image is in an abnormal satellite state;

[0014] The second classification neural network is used to classify the input MTF image into normal and faulty satellite states and output the classification probability and in, To determine the probability that the input MTF image is in a normal satellite state; To determine the probability that the input MTF image is in an abnormal satellite state;

[0015] Feature fusion network, used to classify the probability and Classification probability and Perform feature fusion to obtain the final probability p of the satellite being normal or abnormal 0 、p 1 Output.

[0016] Furthermore, the first classification neural network and the second classification neural network VIT network model structure are trained with different training samples to obtain their respective network parameters;

[0017] The feature fusion network is a fully connected layer, and after training, it obtains the most appropriate classification weights.

[0018] Furthermore, the generation process of the input sequence of the VIT network model includes:

[0019] 1) Process the two-dimensional image in blocks;

[0020] Both GASF image and MTF image are single-channel images of size l*l, and the image is divided into patches of size P*P*1; P can divide l;

[0021] 2) Flatten each patch block to obtain a token vector;

[0022] After flattening, the token vector of each block has a dimension of P 2 , the length is

[0023] 3) Add a classification flag x before the token vector class , in P 2 The dimension vector is converted into P after adding the classification flag 2 +1 dimension;

[0024] 4) After the classification flag is set, perform Position Embedding operation to obtain the position vector;

[0025] 5) The token vector and position vector with the classification flag are added to form the input sequence of VIT.

[0026] Furthermore, in the feature fusion network, the classification probabilities obtained from the two image data are fused through the fully connected layer to train the most appropriate classification weight μ;

[0027]

[0028]

[0029] If p 0 If it is larger, it is judged as no fault, otherwise it is judged as a fault.

[0030] Furthermore, in the detection of a single star failure in step S5,

[0031] 1) During the test, a fixed time series length is used, and the pseudo-range time series of a single satellite is updated with a step length of 1, so that the end time of the pseudo-range time series is always the current time t, and the length of the time series to be tested is l, Γ t =[γ t-l+1 ,...,γ t ]; after normalization, we get the normalized time series

[0032] 2) Normalize the time series at the current moment t Perform two-dimensional encoding processes as in step S3 to obtain a two-dimensional image and

[0033] 3) Convert the two-dimensional image and The data is input into the trained neural network model to classify the satellite into normal and faulty states, and the classification with the highest probability is finally taken as the final judgment result.

[0034] When the probability of the label being "1" is relatively high, the satellite is determined to be faulty and is excluded from the navigation system to avoid the impact of the faulty satellite on navigation accuracy.

[0035] Furthermore, during the INS recursive error estimation process, the state vector of the GNSS / INS tightly integrated navigation system is established as:

[0036]

[0037] in, φ, b g , b a 、s g and a They are the position vector error, velocity error, misalignment angle, gyroscope bias, accelerometer bias, gyroscope scale factor error and accelerometer scale factor error of the INS state recursion result in the ECEF coordinate system;

[0038] The observed quantity is:

[0039]

[0040] Among them, ρ GNSSi is the pseudo-range observation of the i-th observable satellite at time k, is the three-dimensional coordinate of the i-th satellite position, is the three-dimensional coordinate of the recursive position of INS at time k; i = 1,…,m, where m is the number of observable satellites.

[0041] Further, in step S2, the new information data γ obtained by the filtering process k for:

[0042]

[0043] For a single observable satellite, the pseudo-range data in the new information data for a period of continuous length l is taken to form the pseudo-range time series Γ = [γ1,...,γ i ,...,γ l ];

[0044] After normalizing the pseudorange time series Γ, we get the normalized time series

[0045]

[0046] in,

[0047] Furthermore, the GASF-encoded image obtained by GASF encoding is:

[0048]

[0049] in

[0050] Furthermore, the process of obtaining the two-dimensional image data by MTF encoding includes:

[0051] 1) For normalized time series First, the interval is divided into k parts in the range [0,1] according to a certain size to obtain k quantile boxes q1,...,q j ,...,q k ; Normalize the time series Each element in is divided into the corresponding quantile bin q j middle,

[0052] 2) Construct a normalized time series The new time series of equal length Ω=[ω1,...,ω i ,...,ω l ], where ω i The value is the normalized time series The quantile bin corresponding to the i-th element in ;

[0053] 3) Construct a k×k weighted adjacency matrix W; the element w in the weighted adjacency matrix n,j The value is equal to the subsequence [q j ,q n ] frequency of occurrence,

[0054] 4) Perform normalization operation, let ∑ j w nj =1; then, by considering the time position, the matrix W that is not related to time is expanded into the M matrix to obtain the MTF coded image in the Markov transformation field:

[0055] The present invention can achieve one of the following beneficial effects:

[0056] The method disclosed by the present invention uses a generalized maximum correlation entropy Kalman filter to improve input data and reduce non-Gaussian noise in the data, thereby more accurately distinguishing between faults and non-faults.

[0057] The pseudo-range time series (including fault series and non-fault series) in the new information data obtained by filtering are processed by two-dimensional coding using Gramman Angular Sum Field (GASF) and Markov Transition Field (MTF); images reflecting the time correlation in different time intervals of the time series and images reflecting the dynamic change characteristics of the time series are obtained respectively;

[0058] The images converted by GASF and MTF are processed separately by VIT network and then fused and judged to obtain more accurate fault detection results.

[0059] In the present invention, each satellite independently performs a fault determination process, and satellites determined to be faulty can be excluded from the navigation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] The drawings are only for the purpose of illustrating specific embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like components throughout the drawings.

[0061] Figure 1 is a flow chart of a neural network fault detection method based on generalized maximum relevance entropy Kalman filtering in an embodiment of the present invention;

[0062] FIG2( a ) is a schematic diagram of a time series before GASF encoding in an embodiment of the present invention;

[0063] FIG2( b ) is a schematic diagram of a GASF image after GASF encoding in an embodiment of the present invention;

[0064] FIG3( a ) is a schematic diagram of a time series before MTF encoding in an embodiment of the present invention;

[0065] FIG3( b ) is a schematic diagram of an MTF image after MTF encoding in an embodiment of the present invention;

[0066] Figure 4 The present invention is a block diagram of the connection of a neural network model for classifying normal and fault states of satellites in an embodiment of the present invention. DETAILED DESCRIPTION

[0067] The preferred embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used to illustrate the principles of the present invention together with the embodiments of the present invention.

[0068] One embodiment of the present invention discloses a neural network fault detection method based on generalized maximum relevance entropy Kalman filtering, such as Figure 1 As shown, including:

[0069] Step S1, establishing the discrete state equation and observation equation of the INS error of the GNSS / INS tightly integrated navigation system; using the generalized maximum relevant entropy Kalman filter to perform INS recursive error estimation and recursive error correction;

[0070] Step S2, collecting the pseudorange time series in the filtered new information data of a single satellite, and normalizing it to obtain a normalized time series;

[0071] Step S3, performing Gramian Summation Angular Field (GASF) encoding and Markov Transition Field (MTF) encoding on the normalized time series to obtain a GASF image and an MTF image;

[0072] Step S4, establishing a neural network model for satellite normal and fault state classification; after using GASF images and MTF images to train the neural network model respectively, the classification probabilities of satellite normal and fault state classification based on GASF images and MTF images are subjected to feature fusion to obtain the final classification result of the neural network model;

[0073] Step S5: using the trained neural network model to detect single satellite failures during the positioning process of the integrated navigation system.

[0074] Specifically, in step S1, the state vector of the established GNSS / INS tightly integrated navigation system is:

[0075]

[0076] in, φ, b g , b a 、s g and a They are the position vector error, velocity error, misalignment angle, gyroscope bias, accelerometer bias, gyroscope scale factor error and accelerometer scale factor error of the INS state recursion result in the ECEF coordinate system;

[0077] According to the error disturbance theory, the discrete state equation and state equation of INS error are established as follows:

[0078] X k =Φ k|k-1 X k-1 +ω k-1

[0079] Z k =H k X k +v k

[0080] X k , X k-1 are the state vectors at time k and k-1 respectively; ω k-1 is the system noise at time k-1, Φ k|k-1 is the state transfer matrix; Z k is the observed value at time k, v k is the observation noise at time k; H k is the observation matrix at time k.

[0081] Among them, the state transfer matrix Φ k|k-1 for:

[0082]

[0083] Where Δt is the discrete time interval of the system, is the Earth's rotation angular velocity vector, is the direction cosine matrix from the carrier coordinate system b to the ECEF coordinate system, f b is the specific force output by the carrier accelerometer, T bg 、T ba , T sg 、T sa The gyro bias error, accelerometer bias error, gyro scale factor error, and accelerometer scale factor error are modeled as the relevant time parameters of the first-order Markov process respectively;

[0084] Among them, ρ GNSSi is the pseudo-range observation of the i-th observable satellite at time k, ρ INSi is the geometric distance calculated based on the position presented by INS at time k and the position of the i-th observable satellite; i = 1,…,m; m is the number of observable satellites;

[0085] v k is the observation noise at time k; H k is the observation matrix at time k, and the optimal estimate is obtained from the known state at time k-1 and are calculated from the observable satellite positions at time k solved from the ephemeris solution and are known quantities.

[0086] The relationship between the observable satellite pseudo-range observation and the INS recursive state error (i.e., state quantity X) is established through the observation equation; that is, the observation at time k:

[0087]

[0088] Among them, ρ GNSSi is the pseudorange observed for the i-th observable satellite at time k, is the three-dimensional coordinate of the position of the i-th observable satellite, is the three-dimensional coordinate of the recursive position of INS at time k.

[0089] Specifically, in step S1, the process of using generalized correlation entropy Kalman filtering to perform INS recursive error estimation includes:

[0090] 1) Perform one-step state prediction to obtain one-step estimation at time k The covariance matrix P(k|k-1) is:

[0091]

[0092]

[0093] is the optimal estimate of the error state at time k-1, For use and the state transfer matrix Φ k|k-1 The one-step estimate of the error state at time k is calculated, and P(k|k-1) is the one-step estimate The covariance matrix of , P(k-1|k-1) is the estimation error covariance matrix at time k-1, and Q(k-1) is the system noise covariance matrix at time k-1;.

[0094] Here and P(k|k-1) will be used to calculate the optimal estimate of the error state at time k. When the system starts, the first moment k=1 is considered, and the moment k-1 corresponds to the moment 0. The state at the moment 0 is the preset initial state. The determination of the initial state x(0) and the corresponding Q(0) and P(0|0) depends on the characteristics of the adopted physical system and prior knowledge.

[0095] 2) One-step estimation based on k moment The covariance matrix P(k|k-1) of and the covariance matrix R(k) of the observation noise are decomposed by Cholesky;

[0096]

[0097] R(k) is the covariance matrix of the observation noise in the observation equation.

[0098] 3) Based on the Cholesky decomposition results, one-step estimation based on k moments Perform iterative calculation of the filter residual e(k,t);

[0099] e(k,t)=D(k)-W(k)x(k,t);

[0100] in, z(k) is the observed value at time k;

[0101] Hk is the observation matrix at time k;

[0102] When step 3) is executed for the first time at the same time, x(k, t) takes the value in step 1) Then the calculation is performed according to the iteration value; t represents the number of iterations, and t=1 at the initial time.

[0103] 4) According to the filtering residual of t iterations, the generalized correlation entropy weighted value of each observation channel of t iterations is calculated to form a generalized correlation entropy weighted matrix

[0104]

[0105] h i (e i (k,t))=|e i (k,t)| α-2 exp(-β -α |e i (k,t) α |)

[0106] β is the generalized Gaussian kernel function h of the i-th observation channel i (e i (k, t)) is the kernel bandwidth, α is the shape parameter, and when the filter robustness is concerned, h i (e i The α parameter in (k, t) is set to 2, which is the form of Gaussian distribution, and the β parameter is set to 0.1 to make it robust.

[0107] When using the optimal estimation filter under the maximum generalized correlation entropy criterion, the Kalman gain needs to consider the correlation entropy weighting based on the observation noise covariance. The weight h i (e i (k, t)) has a first-order decay in the interval with small residuals, and a high-order decay in the interval with large residuals and decays to 0 quickly. Compared with the traditional Kalman filter that only uses covariance as a fixed weight, GMCKF has excellent anti-error capability for observations with gross error interference.

[0108] 5) Using the generalized correlation entropy weighted matrix Calculate the Kalman gain K(k,t);

[0109]

[0110]

[0111] in, is the observation error covariance matrix of each observation channel weighted by relevant entropy;

[0112] 6) According to the Kalman gain K(k,t), the state is estimated to obtain the t+1th iteration value When the difference between the t+1th iteration value and the tth iteration value is greater than the threshold, return to step 3) and execute again; otherwise, the best estimated value of the state quantity k at the current moment is obtained.

[0113] Specifically,

[0114] It is the iteration value obtained at the t+1th iteration, which is calculated from the result of the tth iteration.

[0115] At this time, if:

[0116]

[0117] Jump to step 3) and re-execute, otherwise get the best estimate of the state quantity at the current moment ε is the set threshold.

[0118] The estimated error value is used to feedback correct the recursive carrier motion state.

[0119] Use the best estimate of the state at time k In φ correction pure inertial navigation state recursion state: position speed Direction cosine matrix Get the position at the current time (assuming it is time k) after the estimated error correction speed Direction Cosine Matrix

[0120]

[0121]

[0122] Among them, the pure inertial navigation state recursive state: position speed Direction Cosine Matrix The conventional pure inertial navigation state recursion method can be used to obtain the state, and the use of the conventional pure inertial navigation state recursion method to obtain the recursive state does not affect the protection scope of the present invention.

[0123] In the preferred solution, the generalized Gaussian kernel function h of each observation channel is adapted to the observation noise that varies dynamically with the carrier motion. i (e i (k,t)) with shape parameter α iDynamically adjust the generalized Gaussian kernel function h of each observation channel i (e i The shape characteristics of (k,t)) match the observation noise and autonomously adapt to the changes of observation noise.

[0124] Specifically, in step S2, the new information data γ obtained by the filtering process k for:

[0125]

[0126] For a single observable satellite, the pseudo-range data in the new information data for a period of continuous length l is taken to form the pseudo-range time series Γ = [γ1,...,γ i ,...,γ l ];

[0127] After normalizing the pseudorange time series Γ, we get the normalized time series

[0128]

[0129] in,

[0130] Specifically, the process of performing GASF encoding on the normalized time series to obtain the GASF image in step S3 includes:

[0131] 1) Encode the normalized time series as the angle cosine φ i ;

[0132]

[0133] There is an important consequence of encoding the normalized time series as the cosine of an angle: It is monotonic in the interval [0,1], so each time series must have and only generates one encoding method.

[0134] 2) Considering the cosine of the angle sum between each image point to reflect the time correlation in different time intervals to establish a two-dimensional image;

[0135]

[0136] Specifically, the time series before encoding is shown in FIG2( a ), and the GASF image after GASF encoding is shown in FIG2( b ).

[0137] The process of obtaining the MTF image by MTF encoding the normalized time series includes:

[0138] 1) For normalized time series First, the interval is divided into k parts in the range [0,1] according to a certain size to obtain k quantile boxes q1,...,q j ,...,q k ; Normalize the time series Each element in is divided into the corresponding quantile bin q j middle;

[0139] 2) Construct a normalized time series The new time series of equal length Ω=[ω1,...,ω i ,...,ω l ], where ω i The value is the normalized time series The quantile bin corresponding to the i-th element in ;

[0140] 3) Construct a k×k weighted adjacency matrix W; the element w in the weighted adjacency matrix n,j The value is equal to the subsequence [q j ,q n ] frequency of occurrence,

[0141] 4) Perform normalization operation, let ∑ j w nj =1; then, by considering the time position, the matrix W that is not related to time is expanded into the M matrix to obtain the MTF coded image in the Markov transformation field:

[0142]

[0143] For example, Γ′=[1,5,7,3,9], the normalized In the range of [0,1], the bins q1, q2, q3 are constructed with 0, 0.3, 0.7 and 1 as the boundaries; the constructed and normalized time series New time series of equal length Ω = [q1, q2, q3, q1, q3]; construct a 3×3 weighted adjacency matrix After normalization Then perform MTF coding to obtain

[0144] Specifically, the time series before encoding is shown in FIG3( a ), and the MTF image after MTF encoding is shown in FIG3( b ).

[0145] Specifically, the neural network model used for satellite normal and fault status classification in step S4 is as follows: Figure 4 As shown, it includes a first classification neural network, a second classification neural network and a feature fusion network;

[0146] The first classification neural network is used to classify the input GASF image into normal and faulty satellite states and output the classification probability. and in, To judge the probability that the input GASF image is in normal satellite status; To determine the probability that the input GASF image is in an abnormal satellite state;

[0147] The second classification neural network is used to classify the input MTF image into normal and faulty satellite states and output the classification probability and in, To determine the probability that the input MTF image is in a normal satellite state; To determine the probability that the input MTF image is in an abnormal satellite state;

[0148] Feature fusion network, used to classify the probability and Classification probability and Perform feature fusion to obtain the final probability p of the satellite being normal or abnormal 0 、p 1 Output.

[0149] More specifically, the first classification neural network and the second classification neural network are VIT (Vision Transformer) network model structures, and their respective network parameters are obtained after being trained with different training samples;

[0150] The feature fusion network is a fully connected layer, and after training, it obtains the most appropriate classification weights.

[0151] Among them, when the first classification neural network and the second classification neural network are trained, the input image is first divided into blocks, flattened and classified to obtain a one-dimensional input vector. After the block processing, network training is performed to determine the final output category probability.

[0152] Specifically, the generation process of the input sequence of the VIT network model includes:

[0153] 1) Process the two-dimensional image in blocks;

[0154] Both GASF image and MTF image are single-channel images of size l*l, and the image is divided into patches of size P*P*1; P can divide l;

[0155] 2) Flatten each patch block to obtain a token vector;

[0156] After flattening, the token vector of each block has a dimension of P 2 , the length is

[0157] 3) Add a classification flag (x) before the token vector class ), P obtained by Patch Embedding 2 The dimension vector is converted into P after adding the classification flag 2 +1 dimension;

[0158] The classification flag is a learnable parameter that can determine the final output category after continuous network training;

[0159] 4) After the classification flag is set, the Position Embedding operation needs to be performed to obtain the position vector;

[0160] After the image is segmented and rearranged, the position information is lost. Embedding the position code into the image block can express the position information of the image block in the original image.

[0161] 5) The token vector and position vector with the classification flag added constitute the input sequence of ViT;

[0162] The input sequence z of ViT is as follows:

[0163]

[0164]

[0165] where x class It is the classification flag. Represents each dimension vector of the block, E pos Represents positional encoding.

[0166] The input sequence z is passed into the Transformer Encoder of the VIT network model, where the Transformer Encoder is mainly composed of MHA (multi-head attention) and MLP (two-layer fully connected network using GELU activation function). The Norm layer and residual link are added before MSA and MLP; after TE, the classification probability is obtained through MLP Head.

[0167] In the feature fusion network, the classification probabilities obtained from the two image data are fused through the fully connected layer to train the most appropriate classification weight μ.

[0168]

[0169]

[0170] If p 0 If it is larger, it is judged as no fault, otherwise it is judged as a fault.

[0171] Specifically, during the training process,

[0172] The normal data label is set to 0, the abnormal data label is set to 1, the loss function is the binary classification cross entropy function, and the model parameters are optimized by the gradient descent method. The input data is the training data prepared in advance, in which the number of normal data and abnormal data is consistent, so as to avoid the problem of unbalanced model training. The number of loops is set to 15 during training, and the model is trained in a loop, iterating the data set, updating the model weights, and calculating the loss and accuracy.

[0173] Finally, the model weights with the least loss and the highest accuracy are saved.

[0174] Specifically, in the detection of a single-star fault in step S5,.

[0175] 1) During the test, a fixed time series length is used, and the pseudo-range time series of a single satellite is updated with a step length of 1, so that the end time of the pseudo-range time series is always the current time t, and the length of the time series to be tested is l, Γ t =[γ t-l+1 ,...,γ t ]; after normalization, we get the normalized time series

[0176] 2) Normalize the time series at the current moment t Perform two-dimensional encoding processes as in step S3 to obtain a two-dimensional image and

[0177] 3) Convert the two-dimensional image and The data is input into the trained neural network model to classify the satellite into normal and faulty states, and finally the classification with the higher probability is taken as the final judgment result.

[0178] When the probability of the label being "1" is high, the satellite is judged to be faulty and is excluded from the navigation system, thus avoiding the impact of faulty satellites on navigation accuracy.

[0179] In summary, the method disclosed in the embodiment of the present invention uses a generalized maximum relevant entropy Kalman filter to improve input data and reduce non-Gaussian noise in the data, thereby more accurately distinguishing between faults and non-faults.

[0180] The pseudorange time series (including fault series and non-fault series) in the filtered new information data are two-dimensionally encoded using the Gramman Angular Sum Field (GASF) and Markov Transition Field (MTF); images reflecting the time correlation in different time intervals of the time series and images reflecting the dynamic change characteristics of the time series are obtained respectively; the images converted by GASF and MTF are processed separately through the VIT network and then fused and judged, so as to obtain more accurate fault detection results.

[0181] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A neural network fault detection method based on generalized maximum relevance entropy Kalman filtering, characterized in that: include: Step S1, establishing the discrete state equation and observation equation of INS error of GNSS / INS tightly integrated navigation system; Generalized maximum correlation entropy Kalman filter is used for INS recursive error estimation and recursive error correction; Step S2, collecting the pseudorange time series in the filtered new information data of a single satellite, and normalizing it to obtain a normalized time series; Step S3, performing GASF coding and MTF coding processing on the normalized time series to obtain a GASF image and an MTF image; Step S4, establishing a neural network model for satellite normal and fault state classification; after using GASF images and MTF images to train the neural network model respectively, the classification probabilities of satellite normal and fault state classification based on GASF images and MTF images are subjected to feature fusion to obtain the final classification result of the neural network model; Step S5: using the trained neural network model to detect single satellite failures during the positioning process of the integrated navigation system; In step S1, the process of using generalized correlation entropy Kalman filtering to perform INS recursive error estimation includes: 1) Perform one-step state prediction to obtain one-step estimation at time k The covariance matrix P(k|k-1) of is: 2) One-step estimation based on k moment The covariance matrix P(k|k-1) of and the covariance matrix R(k) of the observation noise are decomposed by Cholesky; 3) Based on the Cholesky decomposition results, one-step estimation based on k moments Perform iterative calculation of the filter residual e(k,t); t represents the number of iterations, and t=1 initially; 4) According to the filtering residual of t iterations, the generalized correlation entropy weighted value of each observation channel of t iterations is calculated to form a generalized correlation entropy weighted matrix h i (e i (k,t))=|e i (k,t)| α-2 exp(-β -α |e i (k,t) α |) β is the kernel bandwidth, α is the shape parameter; the generalized Gaussian kernel function h of each observation channel is i (e i (k, t)) is dynamically adjusted so that the generalized Gaussian kernel function h of each observation channel i (e i The shape characteristics of (k,t)) match the observation noise and autonomously adapt to the changes of observation noise; i = 1,…,m; m is the number of observable satellites; 5) Using the generalized correlation entropy weighted matrix Calculate the Kalman gain K(k,t); 6) According to the Kalman gain K(k,t), the state is estimated to obtain the t+1th iteration value When the difference between the t+1th iteration value and the tth iteration value is greater than the threshold, return to step 3) and execute again; otherwise, the best estimated value of the state quantity k at the current moment is obtained.

2. The neural network fault detection method based on generalized maximum relevance entropy Kalman filtering according to claim 1 is characterized in that: A neural network model for classifying normal and faulty states of satellites, including a first classification neural network, a second classification neural network and a feature fusion network; The first classification neural network is used to classify the input GASF image into normal and faulty satellite states and output the classification probability. and in, To judge the probability that the input GASF image is in normal satellite status; To determine the probability that the input GASF image is in an abnormal satellite state; The second classification neural network is used to classify the input MTF image into normal and faulty satellite states and output the classification probability and in, To determine the probability that the input MTF image is in a normal satellite state; To determine the probability that the input MTF image is in an abnormal satellite state; Feature fusion network, used to classify the probability and Classification probability and Perform feature fusion to obtain the final probability p of the satellite being normal or abnormal 0 、p 1 Output.

3. The neural network fault detection method based on generalized maximum relevance entropy Kalman filtering according to claim 2 is characterized in that: The first classification neural network and the second classification neural network are VIT network model structures, and their respective network parameters are obtained after being trained with different training samples; The feature fusion network is a fully connected layer, and after training, it obtains the most appropriate classification weights.

4. The neural network fault detection method based on generalized maximum relevance entropy Kalman filtering according to claim 3 is characterized in that: The generation process of the input sequence of the VIT network model includes: 1) Process the two-dimensional image in blocks; Both GASF image and MTF image are single-channel images of size l*l, and the image is divided into patches of size P*P*1; P can divide l; 2) Flatten each patch block to obtain a token vector; After flattening, the token vector of each block has a dimension of P 2 , the length is 3) Add a classification flag x before the token vector class , in P 2 The dimension vector is converted into P after adding the classification flag 2 +1 dimension; 4) After the classification flag is set, perform Position Embedding operation to obtain the position vector; 5) The token vector and position vector with the classification flag are added to form the input sequence of VIT.

5. The neural network fault detection method based on generalized maximum relevance entropy Kalman filtering according to claim 3 is characterized in that: In the feature fusion network, the classification probabilities obtained from the two image data are fused through the fully connected layer to train the most appropriate classification weight μ; If p 0 If it is larger, it is judged as no fault, otherwise it is judged as a fault.

6. The neural network fault detection method based on generalized maximum relevance entropy Kalman filtering according to claim 3 is characterized in that: In the detection of single star failure in step S5, 1) During the test, a fixed time series length is used, and the pseudo-range time series of a single satellite is updated with a step length of 1, so that the end time of the pseudo-range time series is always the current time t, and the length of the time series to be tested is l, Γ t =[γ t-l+1 ,...,γ t ]; After normalization, we get the normalized time series 2) Normalize the time series at the current moment t Perform two-dimensional encoding processes as in step S3 to obtain a two-dimensional image and 3) Convert the two-dimensional image and The data is input into the trained neural network model to classify the satellite into normal and faulty states, and the classification with the highest probability is taken as the final judgment result. When the probability of the label being "1" is high, the satellite is determined to be faulty and is excluded from the navigation system to avoid the impact of the faulty satellite on navigation accuracy.

7. The neural network fault detection method based on generalized maximum relevance entropy Kalman filtering according to any one of claims 1 to 6, characterized in that: During the INS recursive error estimation process, the state vector of the established GNSS / INS tightly integrated navigation system is: in, φ, b g 、b a 、s g and a They are the position vector error, velocity error, misalignment angle, gyroscope bias, accelerometer bias, gyroscope scale factor error and accelerometer scale factor error of the INS state recursion result in the ECEF coordinate system; The observed quantity is: Among them, ρ GNSSi is the pseudo-range observation of the i-th observable satellite at time k, is the three-dimensional coordinate of the position of the i-th satellite, is the three-dimensional coordinate of the recursive position of INS at time k; i = 1,…,m, where m is the number of observable satellites.

8. The neural network fault detection method based on generalized maximum relevance entropy Kalman filtering according to claim 7 is characterized in that: In step S2, the new information data γ obtained by the filtering process k for: For a single observable satellite, the pseudo-range data in the new information data for a period of continuous length l is taken to form the pseudo-range time series of the satellite Γ = [γ1,...,γ i ,...,γ l ]; After normalizing the pseudorange time series Γ, we get the normalized time series in, 9. The neural network fault detection method based on generalized maximum relevance entropy Kalman filtering according to claim 8 is characterized in that: GASF encoded image obtained by GASF encoding: in 10. The neural network fault detection method based on generalized maximum relevance entropy Kalman filtering according to claim 8, characterized in that: The process of obtaining two-dimensional image data by MTF encoding includes: 1) For normalized time series First, the interval is divided into k parts in the range [0,1] according to a certain size to obtain k quantile boxes q1,...,q j ,...,q k ; Normalize the time series Each element in is divided into the corresponding quantile bin q j middle, 2) Construct a normalized time series The new time series of equal length Ω=[ω1,...,ω i ,...,ω l ], where ω i The value is the normalized time series The quantile bin corresponding to the i-th element in ; 3) Construct a k×k weighted adjacency matrix W; the element w in the weighted adjacency matrix n,j The value of is equal to the subsequence [q j ,q n ] frequency of occurrence, 4) Perform normalization operation, let ∑ j w nj =1; then, by considering the time position, the matrix W that is not related to time is expanded into the M matrix to obtain the MTF coded image in the Markov transformation field:

Images