Fault-tolerant method for SINS and GNSS integrated navigation system based on LSTM neural network

Abstract

The invention provides a fault-tolerant method for an SINS (strapdown inertial navigation system) and GNSS (global navigation satellite system) integrated navigation system based on an LSTM (long short-term memory) neural network, and provides a fault detection and system fault-tolerant method based on assistance of the LSTM neural network for sudden change faults and slow change faults in an SINS / GNSS loose integrated navigation system. An innovation mahalanobis distance assisted chi-square test algorithm is utilized to detect abrupt change faults in the system in real time, and then a sequential probability ratio algorithm is utilized to detect slow change faults in the system. And the robust Kalman filtering based on M estimation is used for completing robustness enhancement work of the system, so that the influence of the sudden change fault on the system is effectively inhibited. And an LSTM neural network substitution mode is used to effectively inhibit the influence of a slowly changing fault on the system, and signal restoration is realized. Through detection and fault tolerance of two faults existing in the GNSS signal, the precision of integrated navigation positioning is improved.

Description

Technical Field

[0001] The present application relates to the technical field of detection and fault tolerance of sudden faults and slowly changing faults, and in particular to a fault tolerance method for a SINS and GNSS integrated navigation system based on an LSTM neural network. Background Art

[0002] In the integrated navigation system involved in actual application scenarios, due to the complex types and large number of sensors, the sensor information often carries sudden (wild value) faults and slowly changing faults. Slowly changing faults often take the form of small amplitude and weak trend, and exist in the sensors of the integrated navigation system. Multi-source unknown interference and device aging will also affect the accuracy of the final result of the integrated navigation. Therefore, designing a fault-tolerant solution that can detect and effectively handle sudden faults and slowly changing faults in the integrated navigation system in real time plays a vital role in the accuracy of the final positioning result of the integrated navigation.

[0003] In recent years, among the commonly used positioning technologies, the Strapdown Inertial Navigation System (SINS) has been favored by scholars from various countries due to its strong autonomy, good concealment, high accuracy in a short period of time, fast update frequency and comprehensive navigation information. GNSS (Global Navigation Satellite System) has also been widely used in many technical fields due to its real-time and high-precision advantages. Considering that a single navigation method can no longer meet the needs of today's actual tasks, and that inertial navigation has the characteristics of divergence over time, combined navigation technology has become a core technology that must be broken through and developed sustainably in the field of navigation. SINS / GNSS is a commonly used combined navigation method. GNSS can provide SINS with real-time and high-precision navigation information, assist SINS information in timely correction, and prevent errors from accumulating over time.

[0004] However, the accuracy of the SINS / GNSS integrated navigation system is still affected by abnormal GNSS measurement signals. In actual application scenarios, factors such as complex environments and time-varying noise often affect GNSS measurement signals (such as position, speed, etc.). According to the degree of impact and the characteristics of the fault, it is divided into sudden faults (wild values) and slow-varying faults. Slow-varying faults are difficult to detect due to their small changes and unclear characteristics, which have a great impact on the positioning results. Although sudden faults are relatively easy to detect, the detection time and the degree of fault tolerance also affect the final positioning results. Summary of the invention

[0005] The present application provides a SINS and GNSS integrated navigation system fault tolerance method based on LSTM neural network, which can be used to solve the technical problem that two types of faults in the SINS / GNSS integrated navigation system cannot be accurately detected and processed.

[0006] The present application provides a SINS and GNSS integrated navigation system fault tolerance method based on LSTM neural network, the method comprising:

[0007] Step 1: Establish a SINS and GNSS integrated navigation model; then, measure the wild value of the sudden fault carried in the GNSS position data, and use the residual chi-square test method assisted by the Mahalanobis distance of the new information sequence, that is, the MD chi-square test, to determine whether there is a sudden fault;

[0008] Step 2: For the slowly varying faults with probability in the GNSS position data, the sequential probability ratio algorithm SPRT is used. When a slowly varying fault occurs, the sequential probability ratio increment is constructed, and a slowly varying fault discrimination model based on the sequential probability ratio increment is established. A threshold is set. When the sequential probability ratio increment value is negative, it is determined that the SINS and GNSS systems contain slowly varying faults.

[0009] Step 3: According to the results of the MD chi-square algorithm and the SPRT algorithm in step 1 and step 2, determine whether there is a fault in the SINS and GNSS systems, and determine whether there is a sudden fault or a slow-changing fault; as long as a sudden fault or a slow-changing fault occurs in the system, it is determined that the SINS and GNSS systems have a fault;

[0010] Step 4, perform fault tolerance processing on the SINS and GNSS systems; when the GNSS position information is detected to contain a slowly varying fault, introduce the LSTM neural network model; the formation of the LSTM neural network is divided into a training phase and a prediction phase; when the GNSS position information is fault-free, the LSTM training phase is completed; when the GNSS position information contains a slowly varying fault, the original GNSS position information is replaced with the LSTM trained position information to complete the online reconstruction of the system;

[0011] Step 5: When the GNSS position information is detected to contain only mutation faults, in order to suppress the impact of measured wild values ​​on the integrated navigation system, a weighted matrix restriction and a modified innovation matrix are introduced;

[0012] Step 6: Perform SINS and GNSS combined navigation and output the navigation results.

[0013] Further, in step 1, a SINS and GNSS integrated navigation model is established; then, for the sudden fault measurement wild value carried in the GNSS position data, a residual chi-square test method assisted by the Mahalanobis distance of the new information sequence, i.e., MD chi-square, is used to test to determine whether there is a sudden fault; including:

[0014] Step 11, establishing a SINS / GNSS integrated navigation system model;

[0015] According to the analysis of the SINS error model, the Northeastern sky geographic coordinate system is used as the navigation coordinate system to establish the SINS / GNSS integrated navigation system. The discrete state space model is:

[0016] X k =Φ k,k-1 X k-1 +Γ k-1 W k-1

[0017] In the formula, Φ k,k-1 represents the state one-step transfer matrix of the SINS / GNSS integrated navigation system; Γ k-1 represents the noise distribution or driving matrix of the system; W k-1 is the system noise; and W k-1 ~N(0,Q k ) satisfies; X k With X k-1 are the state variables of the integrated navigation system at time k and time k-1 respectively, where:

[0018]

[0019] In the formula, vector φ=[φ E φ N φ U ] T Represents the angular error information of the inertial navigation platform; vector δv = [δv E δv N δv U ] T Represents the error information of inertial navigation velocity; vector δp=[δλδLδh] T is the error information of the inertial navigation position; ε b , They are the first-order Malvian processes of the three-axis errors of the gyroscope and accelerometer;

[0020] The measurement equation of the integrated navigation system is expressed as:

[0021] Z k =H k X k +V k

[0022] In the formula, the measurement vector Z k Taken as the difference in three-dimensional position and velocity between SINS and GNSS; H k is the measurement matrix; V k is the measurement noise, which is approximately white noise, and V k Meet V k ~N(0,R k );

[0023] Step 12, using the Mahalanobis distance of the new information sequence to assist the chi-square test;

[0024] The Mahalanobis distance for measuring innovation is established as:

[0025]

[0026] Among them, δZ k =Z k -H k X kk-1 Represents the measurement innovation, and the variance is:

[0027]

[0028] In the formula, μ z Represents the measurement vector Z k The mean value, P k,k-1 represents the error covariance;

[0029] Construction Statistics That is M k It follows a chi-square distribution with parameter 1;

[0030] When the measurement information output by GNSS contains outliers, that is, when there is a sudden fault in the system, the square of the Mahalanobis distance of the measurement information will no longer satisfy the chi-square distribution;

[0031] The hypothesis test idea in probability statistics is used to evaluate the system; first, the significance level α of the chi-square test is set, and the Mahalanobis distance under the condition of significance level α is generally satisfied Further establish the conditions for judging whether there is a sudden fault in the GNSS output information:

[0032] like It is determined that the SINS / GNSS integrated navigation system has a sudden fault;

[0033] like It is determined that the SINS / GNSS integrated navigation system has no sudden fault;

[0034] Furthermore, for the slowly varying faults with probability in GNSS position data, the sequential probability ratio algorithm SPRT is used. When a slowly varying fault occurs, the sequential probability ratio increment is constructed, and a slowly varying fault discrimination model based on the sequential probability ratio increment is established. The threshold is set. When the sequential probability ratio increment value is negative, it is determined that the SINS and GNSS systems contain slowly varying faults, including:

[0035] First, assume that the residual value at time k is v(k); the sequence of independent random samples belonging to it at consecutive k times is {v(1), v(2), …, v(k)}. According to the central limit theorem of probability theory, we know that: in and (σ(k)) 2 They represent the sample mean and sample variance respectively; when the GNSS output information has no slowly varying fault, that is, the null hypothesis H0, and has a slowly varying fault, that is, the alternative hypothesis H1, the maximum likelihood functions are:

[0036]

[0037] Based on the above binary hypothesis testing theory, we now take the residual sequence {v(1), v(2), …, v(k)} at k independent moments, and the likelihood ratio is:

[0038]

[0039] Take the logarithm of both sides of the likelihood ratio function and then use the sample mean Replacing μ approximately, we get the likelihood ratio as follows:

[0040]

[0041] The sequential probability ratio increment is:

[0042] Considering the concepts of false alarm rate and missed detection rate of fault detection, the following inspection threshold is set:

[0043]

[0044] Among them, P M represents the missed detection rate, P F represents the false alarm rate;

[0045] Based on the above analysis, the following conditions are established to determine whether the GNSS output information has a slowly varying fault:

[0046] If λ(k)≥T(H1), the null hypothesis H0 is judged to be false, and the alternative hypothesis H1 is true;

[0047] If λ(k)≤T(H0), then the null hypothesis H0 is judged to be true, and the alternative hypothesis H1 is false;

[0048] If T(H0)≤λ(k)≤T(H1), more GNSS data information needs to be added to continue the inspection.

[0049] Further, step 3, according to the results of the MD chi-square algorithm and the SPRT algorithm in step 1 and step 2, it is determined whether there is a fault in the SINS and GNSS systems, and whether there is a sudden fault and a slow-changing fault; as long as a sudden fault or a slow-changing fault occurs in the system, it is determined that the SINS and GNSS systems have a fault; including:

[0050] For the MD chi-square-sequential probability ratio algorithm, it is possible to simultaneously detect sudden faults and slowly varying faults in the GNSS output information; when the sequential probability ratio increment is a positive value, it is considered that the system has a slowly varying fault, and step 4 is executed to perform online reconstruction using the LSTM neural network method; when the sequential probability ratio increment is a negative value, step 5 is executed to perform robust enhancement based on the M-estimation robust Kalman filtering algorithm to eliminate the impact of sudden faults on the positioning results.

[0051] Furthermore, in step 4, the formation of the LSTM neural network includes the following steps:

[0052] Step 41, training phase: when no fault occurs in the system, the output of the accelerometer and the gyroscope is used as the input of the LSTM, and the position information output by the GNSS is used as the output of the LSTM to complete the LSTM training phase;

[0053] Step 42, prediction phase: when the system contains a slowly varying fault, the output of the trained LSTM model is used to replace the GNSS information containing the slowly varying fault to complete the reconstruction of the system; after the reconstruction work is completed, the integrated navigation is performed again.

[0054] Furthermore, in step 5, when the GNSS position information is detected to contain only a sudden fault, in order to suppress the influence of the measured wild value on the integrated navigation system, a weighted matrix restriction and a modified innovation matrix are introduced, including:

[0055] When the sequential probability ratio increment is negative, the MD chi-square test result is referred to; if the MD chi-square test result indicates that the system has a sudden fault, the robust Kalman filter algorithm based on M estimation is used to enhance the system robustness;

[0056] Define the criterion function as shown below:

[0057]

[0058] where the function ρ(·) is a properly selected continuous convex function, is the i-th element of the new information at time k, and m is the dimension of the quantity measurement; a robust scale adjustment parameter is introduced To standardize the new information, the criterion function is rewritten as:

[0059]

[0060] in It also becomes the mean square error estimate of the i-th element of the new information at time k, It is called standardized news; Calculated by the following formula:

[0061]

[0062] Where MAD stands for median absolute deviation and is defined as:

[0063]

[0064] In the formula, med represents the median, That is, the new information sequence of the i-th measurement component as of the current moment;

[0065] Taking the derivative of with respect to the estimated quantity X and setting it to 0 yields:

[0066]

[0067] in

[0068] definition and matrix ω(·) is called the weight function, so the matrix form is written as:

[0069] Η k ω k e k =0

[0070] The above is the principle of M estimation; therefore, the application of M estimation in Kalman filtering is to introduce the matrix ω k As the weighted matrix of the new information, it is shown as follows:

[0071]

[0072] The robust Kalman filtering algorithm based on M estimation is obtained.

[0073] When the SINS / GNSS system only contains sudden faults (measurement wild values), considering that the existence of wild values ​​will pollute the GNSS data innovation, which will lead to uncontrollable errors in the combined results, the robust KF filtering algorithm based on M estimation is adopted. By defining a new criterion function, a robust scale adjustment parameter is introduced to standardize the innovation. The constructed weight function is used as the weighting matrix of the innovation to suppress the wild values ​​in the position information output by the GNSS, thereby improving the accuracy of the SINS / GNSS combined navigation.

[0074] The present invention uses the fault detection method of step 1 and step 2 to judge the sudden fault and slow-changing fault in the SINS / GNSS integrated navigation system, and then uses the fault-tolerant scheme proposed in step 4 and step 5 to perform online reconstruction and robustness enhancement on the system, thereby improving the accuracy of the integrated navigation.

[0075] The advantages of the present invention compared with the prior art are: according to the new SINS / GNSS combined navigation system fault-tolerant solution provided by the present application, it is possible to better detect sudden faults in combination with the Mahalanobis distance of the new information. Although the traditional chi-square detection algorithm can identify sudden faults, it is still not sensitive enough. If the system is detected to have only slow-varying faults, the robustness enhancement processing is performed using the M-estimation-based robust Kalman filtering algorithm. The MD chi-square-sequential probability ratio fault diagnosis machine proposed in the present invention can be sensitive to slow-varying faults, and the proposed LSTM network is used to reconstruct the system state at this time online to ensure that the combined navigation work can proceed smoothly while ensuring high accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 The overall framework diagram of the method provided in the embodiment of the present application;

[0077] Figure 2 A structural diagram of the LSTM provided in the embodiment of the present application;

[0078] Figure 3 The combined navigation error diagram of the algorithm before and after the improvement provided in the embodiment of the present application;

[0079] Figure 4 A schematic diagram of the method flow provided in an embodiment of the present application. DETAILED DESCRIPTION

[0080] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below with reference to the accompanying drawings.

[0081] In order to effectively handle the two types of faults in the SINS / GNSS integrated navigation system, this scheme proposes a SINS / GNSS integrated navigation system fault tolerance method based on neural network processing, builds a set of SINS / GNSS integrated navigation system fault detection, robust enhancement and online reconstruction models, performs fault detection on the position information output by GNSS, and effectively determines whether it contains sudden faults and slow-changing faults. For sudden faults, the robust Kalman filter algorithm based on M estimation is used for robustness enhancement; for slow-changing faults, considering that LSTM can effectively process time series data in SINS / GNSS integrated navigation, especially in dealing with nonlinearity, long-term dependence and signal loss. LSTM neural network can be familiar with the time series pattern of sensor data through its own learning, and provide more accurate and reliable navigation information when the GNSS signal is interrupted or the data is unreliable. Therefore, for slow-changing faults, when its existence causes the position information output by GNSS to be affected, the robust filtering algorithm cannot effectively suppress the wild value, and LSTM needs to be used for system online reconstruction processing. After the above fault-tolerant scheme is applied, the influence of two types of GNSS faults on the results of the SINS / GNSS integrated navigation system is effectively suppressed, and the accuracy of the SINS / GNSS integrated navigation is improved.

[0082] Step 1: Establish a SINS and GNSS integrated navigation model; then, measure the wild value of the sudden fault carried in the GNSS position data, and use the residual chi-square test method assisted by the Mahalanobis distance of the new information sequence, that is, the MD chi-square test, to determine whether there is a sudden fault.

[0083] Step 2: For the slowly varying faults with probability in the GNSS position data, the sequential probability ratio algorithm SPRT is used. When a slowly varying fault occurs, the sequential probability ratio increment is constructed, and a slowly varying fault discrimination model based on the sequential probability ratio increment is established. The threshold is set. When the sequential probability ratio increment value is negative, it is determined that the SINS and GNSS systems contain slowly varying faults.

[0084] Step 3: According to the results of the MD chi-square algorithm and the SPRT algorithm in step 1 and step 2, determine whether there are faults in the SINS and GNSS systems, and determine whether there are sudden faults and slow-changing faults; as long as the system has a sudden fault or a slow-changing fault, it can be determined that the SINS and GNSS systems have faults.

[0085] Step 4: Perform fault-tolerance processing on the SINS and GNSS systems. When the GNSS position information is detected to contain a slowly varying fault, the LSTM neural network model is introduced. The formation of the LSTM neural network is divided into a training phase and a prediction phase. When the GNSS position information is fault-free, the LSTM training phase is completed. When the GNSS position information contains a slowly varying fault, the original GNSS position information is replaced with the LSTM trained position information to complete the online reconstruction of the system.

[0086] Step 5: When the GNSS position information is detected to contain only mutation faults, in order to suppress the impact of measurement wild values ​​on the integrated navigation system, weighted matrix restrictions and modified information matrices are introduced to effectively reduce the impact of measurement wild values ​​on the integrated navigation results and complete the robust enhancement of the system.

[0087] Step 6: Perform SINS and GNSS combined navigation and output the navigation results.

[0088] The present invention provides a SINS and GNSS integrated navigation system fault tolerance method based on neural network processing, such as Figure 1 As shown, based on the MD chi-square detection and sequential probability ratio algorithm, an MD chi-square-sequential probability ratio joint fault diagnosis machine is formed to diagnose the GNSS output information, mainly for the sudden change faults and slow change faults contained in the GNSS output information.

[0089] When a slowly varying fault is detected in the SINS / GNSS system, LSTM is used for online reconstruction. When the GNSS data is normal, the LSTM training phase is completed first; when the GNSS data contains a slowly varying fault, the output of the trained LSTM is used to replace the GNSS fault data to complete the reconstruction of the system. When the SINS / GNSS system is detected to contain only a sudden fault, the robust Kalman filter based on M estimation is used to enhance the system robustness.

[0090] The specific steps to implement the method are as follows:

[0091] Step 1, SINS / GNSS model establishment and mutation fault detection includes:

[0092] Step 11: Mathematical model of SINS / GNSS integrated navigation system

[0093] According to the analysis of the SINS error model, the Northeastern sky geographic coordinate system is used as the navigation coordinate system to establish the SINS / GNSS integrated navigation system. The discrete state space model is:

[0094] X k =Φ k,k-1 X k-1 +Γ k-1 W k-1

[0095] In the formula, Φ k,k-1 represents the state one-step transfer matrix of the SINS / GNSS integrated navigation system; Γ k-1 represents the noise distribution or driving matrix of the system; W k-1 is the system noise; and W k-1 ~N(0,Q k ) satisfies; X k With X k-1 are the state variables of the integrated navigation system at time k and time k-1 respectively, where:

[0096]

[0097] In the formula, vector φ=[φ E φ N φ U ] T Represents the angular error information of the inertial navigation platform; vector δv = [δv E δv N δv U ] T Represents the error information of inertial navigation velocity; vector δp=[δλδLδh] T is the error information of the inertial navigation position; ε b , They are the first-order Malvian processes of the three-axis errors of the gyroscope and accelerometer;

[0098] The measurement equation of the integrated navigation system is expressed as:

[0099] Z k =H k X k +V k

[0100] In the formula, the measurement vector Z k Taken as the difference in three-dimensional position and velocity between SINS and GNSS; H k is the measurement matrix; V k is the measurement noise, which is approximately white noise, and V k Meet V k ~N(0,R k );

[0101] Step 12, using the Mahalanobis distance of the new information sequence to assist the chi-square test;

[0102] In order to track the characteristics of sudden faults more accurately and quickly, the Mahalanobis distance of the measurement information is established as:

[0103]

[0104] Among them, δZ k=Z k -H k X kk-1 Represents the measurement innovation, and the variance is:

[0105]

[0106] In the formula, μ z Represents the measurement vector Z k The mean value, P k,k-1 represents the error covariance;

[0107] Construction Statistics That is M k It follows a chi-square distribution with parameter 1.

[0108] When the measurement information output by GNSS contains outliers, that is, when there is a sudden fault in the system, the square of the Mahalanobis distance of the measurement information will no longer satisfy the chi-square distribution;

[0109] Based on the analysis of the above principles, the hypothesis test idea in probability statistics is used to evaluate the system; first, the significance level α of the chi-square test is set, and the Mahalanobis distance under the condition of significance level α is generally satisfied Further establish the conditions for judging whether there is a sudden fault in the GNSS output information:

[0110] like It is determined that the SINS / GNSS integrated navigation system has a sudden fault;

[0111] like It is determined that the SINS / GNSS integrated navigation system has no sudden fault;

[0112] Step 2, detection of slow-changing faults, specifically includes:

[0113] In order to detect whether the GNSS measurement information contains slowly varying faults, this scheme uses the sequential probability ratio algorithm. The main steps of the sequential probability ratio algorithm are as follows.

[0114] First, assume that the residual value at time k is v(k); the sequence of independent random samples belonging to it at consecutive k times is {v(1), v(2), …, v(k)}. According to the central limit theorem of probability theory, we know that: in and (σ(k)) 2 They represent the sample mean and sample variance respectively; when the GNSS output information has no slowly varying fault, that is, the null hypothesis H0, and has a slowly varying fault, that is, the alternative hypothesis H1, the maximum likelihood functions are:

[0115]

[0116] Based on the above binary hypothesis testing theory, we now take the residual sequence {v(1), v(2), …, v(k)} at k independent moments, and the likelihood ratio is:

[0117]

[0118] Take the logarithm of both sides of the likelihood ratio function and then use the sample mean Replacing μ approximately, we get the likelihood ratio as follows:

[0119]

[0120] The sequential probability ratio increment is:

[0121] Considering the concepts of false alarm rate and missed detection rate of fault detection, the following inspection threshold is set:

[0122]

[0123] Among them, P M represents the missed detection rate, P F represents the false alarm rate;

[0124] Based on the above analysis, the following conditions are established to determine whether the GNSS output information has a slowly varying fault:

[0125] If λ(k)≥T(H1), the null hypothesis H0 is judged to be false, and the alternative hypothesis H1 is true;

[0126] If λ(k)≤T(H0), then the null hypothesis H0 is judged to be true, and the alternative hypothesis H1 is false;

[0127] If T(H0)≤λ(k)≤T(H1), more GNSS data information needs to be added to continue the inspection.

[0128] Step 3, determination of the test results of the MD chi-square-sequential probability ratio algorithm, including:

[0129] For the MD chi-square-sequential probability ratio algorithm, it is possible to simultaneously detect sudden faults and slowly varying faults in the GNSS output information; the chi-square detection algorithm based on Mahalanobis distance is more sensitive to sudden faults, and at the same time, the sequential probability ratio algorithm can track the slowly changing characteristics of slowly varying faults in real time; when the sequential probability ratio increment is a positive value, it is considered that a slowly varying fault has occurred in the system, and step 4 is executed to perform online reconstruction using the LSTM neural network method; when the sequential probability ratio increment is a negative value, step 5 is executed to perform robust enhancement based on the M-estimation robust Kalman filtering algorithm to eliminate the impact of sudden faults on the positioning results.

[0130] Step 4, an online reconstruction method of a SINS / GNSS integrated navigation system based on LSTM, including:

[0131] When the detection result of the chi-square-sequential probability ratio algorithm is positive, it is determined that the system has a slow-varying fault. At this time, the signal containing the slow-varying fault is reconstructed. This solution introduces an LSTM neural network to complete the online reconstruction of the system. Figure 2 As shown in the figure, the LSTM neural network has the characteristics of strong learning ability and high fitting degree, and can provide relatively accurate original information. It processes the GNSS position information containing slow-varying faults through online reconstruction to reduce its impact on the combined positioning results. Online reconstruction using LSTM neural network is divided into the following two stages:

[0132] Step 41, training phase: when no fault occurs in the system, the output of the accelerometer and the gyroscope is used as the input of the LSTM, and the position information output by the GNSS is used as the output of the LSTM to complete the LSTM training phase;

[0133] Step 42, prediction phase: when the system contains a slowly varying fault, the output of the trained LSTM model is used to replace the GNSS information containing the slowly varying fault to complete the reconstruction of the system; after the reconstruction work is completed, the integrated navigation is performed again.

[0134] Step 5: Enhance the robustness of the system, including:

[0135] When the sequential probability ratio increment is negative, the MD chi-square test result is referred to at this time; if the MD chi-square test result indicates that the system has a sudden fault, the robust Kalman filter algorithm based on M estimation is used to enhance the system robustness. Its main principle is to define a new criterion function and introduce a robust scale adjustment parameter to standardize the new information. The constructed weight function is used as the weighting matrix of the new information to suppress the wild values ​​in the position information output by GNSS. Compared with the traditional robust filtering algorithm, this algorithm can more accurately adjust the weight to suppress sudden faults and improve the accuracy of SINS / GNSS combined navigation positioning results. The main process of the robust Kalman filter algorithm based on M estimation is as follows.

[0136] Taking the robust filtering algorithm based on Huber-M estimation as an example, the robustness enhancement algorithm of the integrated navigation system is performed.

[0137] M estimation, or generalized maximum likelihood estimation, is a commonly used robust estimation method. It is different from the least squares estimation, which uses the minimum residual sum of squares as the criterion function. Instead, it defines the criterion function as shown in the following formula:

[0138]

[0139] where the function ρ(·) is a properly selected continuous convex function, is the i-th element of the new information at time k, and m is the dimension of the quantity measurement; in order to further optimize the robustness of the above criterion function, a robust scale adjustment parameter is often introduced To standardize the new information, the criterion function is rewritten as:

[0140]

[0141] in It also becomes the mean square error estimate of the i-th element of the new information at time k, It is called standardized news; Calculated by the following formula:

[0142]

[0143] Where MAD stands for median absolute deviation and is defined as:

[0144]

[0145] In the formula, med represents the median, That is, the new information sequence of the i-th measurement component as of the current moment;

[0146] Taking the derivative of with respect to the estimated quantity X and setting it to 0 yields:

[0147]

[0148] in

[0149] definition and matrix ω(·) is called the weight function, so the matrix form is written as:

[0150] Η k ω k e k =0

[0151] The above is the principle of M estimation; therefore, the application of M estimation in Kalman filtering is to introduce the matrix ω k As the weighted matrix of the new information, it is shown as follows:

[0152]

[0153] The robust Kalman filter algorithm (MKF for short) based on M estimation is obtained. By introducing a weighted matrix to restrict and correct the new information, the influence of gross measurement errors on the filtering accuracy can be effectively reduced, and the robustness of the filtering can be enhanced.

[0154] Step 6, combined navigation after fault tolerance processing, includes:

[0155] According to steps (1) to (5), sudden change faults and slow change faults in the position information output by the GNSS sensor can be detected and fault-tolerantly processed.

[0156] According to the above description, for the sudden wild values ​​in the system, the MD chi-square test can be used to detect it. For slowly varying faults with small amplitude and small change, the sequential probability ratio algorithm is used for fault detection. After the detection of the two types of faults is completed, the most systematic fault tolerance processing is required. For the sudden wild values ​​in the system, the above-mentioned robust Kalman filter algorithm based on M estimation can effectively suppress the influence of wild values ​​on the combined navigation positioning results. For the slowly varying faults in the system, the robust Kalman filter algorithm based on M estimation can be used to effectively suppress the influence of wild values ​​on the combined navigation positioning results. For the slowly varying faults in the system, the robust filter algorithm alone can no longer achieve the effect of repairing the system. Therefore, this scheme uses the LSTM neural network to reconstruct the signal affected by the slowly varying fault. When the system does not contain sudden wild values ​​and slowly varying faults, the LSTM neural network is trained in advance. The training phase of the LSTM neural network is completed using normal information. When the GNSS signal is detected to contain a slowly varying fault, the output of the trained LSTM can be used to replace the GNSS signal originally affected by the slowly varying fault.

[0157] The novel fault diagnosis and fault-tolerant processing method provided by the present invention, on the one hand, can detect the sudden change faults and slowly changing faults carried in the GNSS output signal in a more real-time and accurate manner. On the other hand, after detecting the two types of faults, it can perform effective fault-tolerant processing to reduce the impact on the final positioning result. Compared with traditional algorithms, it has better real-time performance and accuracy. Therefore, for sudden change wild values ​​and slowly changing faults, the present invention can be used to perform real-time detection and effective fault-tolerant processing, and then re-perform combined navigation, and verify the reliability of this solution through the results of combined navigation. Figure 3 As shown in Figure 2, this is the combined navigation error diagram of the algorithm before and after improvement. Figure 3 It shows that compared with the KF method, the algorithm proposed in this scheme has a smaller error in combined navigation and has effectively improved the estimation accuracy.

[0158] The above-described embodiments of the present application do not constitute a limitation on the protection scope of the present application.

Claims

1. A fault-tolerant method for SINS and GNSS integrated navigation system based on LSTM neural network, characterized in that: The method comprises: Step 1: Establish a SINS and GNSS integrated navigation model; then, measure the wild value of the sudden fault carried in the GNSS position data, and use the residual chi-square test method assisted by the Mahalanobis distance of the new information sequence, that is, the MD chi-square test, to determine whether there is a sudden fault; Step 2: For the slowly varying faults with probability in the GNSS position data, the sequential probability ratio algorithm SPRT is used. When a slowly varying fault occurs, the sequential probability ratio increment is constructed, and a slowly varying fault discrimination model based on the sequential probability ratio increment is established. A threshold is set. When the sequential probability ratio increment value is negative, it is determined that the SINS and GNSS systems contain slowly varying faults. Step 3: According to the results of the MD chi-square algorithm and the SPRT algorithm in step 1 and step 2, determine whether there is a fault in the SINS and GNSS systems, and determine whether there is a sudden fault or a slow-changing fault; as long as a sudden fault or a slow-changing fault occurs in the system, it is determined that the SINS and GNSS systems have a fault; Step 4, perform fault tolerance processing on the SINS and GNSS systems; when the GNSS position information is detected to contain a slowly varying fault, introduce the LSTM neural network model; the formation of the LSTM neural network is divided into a training phase and a prediction phase; when the GNSS position information is fault-free, the LSTM training phase is completed; when the GNSS position information contains a slowly varying fault, the original GNSS position information is replaced with the LSTM trained position information to complete the online reconstruction of the system; Step 5: When the GNSS position information is detected to contain only mutation faults, in order to suppress the impact of measured wild values ​​on the integrated navigation system, a weighted matrix restriction and a modified innovation matrix are introduced; Step 6: Perform SINS and GNSS combined navigation and output the navigation results.

2. The method according to claim 1, characterized in that Step 1: Establish a SINS and GNSS integrated navigation model; then, for the sudden fault measurement wild value carried in the GNSS position data, use the residual chi-square test method assisted by the Mahalanobis distance of the new information sequence, i.e., MD chi-square, to test whether there is a sudden fault; including: Step 11, establishing a SINS / GNSS integrated navigation system model; According to the analysis of the SINS error model, the Northeastern sky geographic coordinate system is used as the navigation coordinate system to establish the SINS / GNSS integrated navigation system. The discrete state space model is: X k =Φ k,k-1 X k-1 +C k-1 W k-1 In the formula, Φ k,k-1 represents the state one-step transfer matrix of the SINS / GNSS integrated navigation system; Γ k-1 represents the noise distribution or driving matrix of the system; W k-1 is the system noise; and W k-1 ~N(0,Q k ) satisfies; X k With X k-1 are the state variables of the integrated navigation system at time k and time k-1 respectively, where: In the formula, vector φ=[φ E φ N φ U ] T Represents the angular error information of the inertial navigation platform; vector δv = [δv E δv N δv U ] T Represents the error information of inertial navigation velocity; vector δp=[δλ δL δh] T is the error information of the inertial navigation position; ε b , They are the first-order Malvian processes of the three-axis errors of the gyroscope and accelerometer; The measurement equation of the integrated navigation system is expressed as: Z k =H k X k +V k In the formula, the measurement vector Z k Taken as the difference in three-dimensional position and velocity between SINS and GNSS; H k is the measurement matrix; V k is the measurement noise, which is approximately white noise, and V k Meet V k ~N(0,R k ); Step 12, using the Mahalanobis distance of the new information sequence to assist the chi-square test; The Mahalanobis distance for measuring innovation is established as: Among them, δZ k =Z k -H k X kk-1 Represents the measurement innovation, and the variance is: In the formula, μ z Represents the measurement vector Z k The mean value, P k,k-1 represents the error covariance; Construction Statistics That is M k It follows a chi-square distribution with parameter 1; When the measurement information output by GNSS contains outliers, that is, when there is a sudden fault in the system, the square of the Mahalanobis distance of the measurement information will no longer satisfy the chi-square distribution; The hypothesis test idea in probability statistics is used to evaluate the system; first, the significance level α of the chi-square test is set, and the Mahalanobis distance under the condition of significance level α is generally satisfied Further establish the conditions for judging whether there is a sudden fault in the GNSS output information: like It is determined that the SINS / GNSS integrated navigation system has a sudden fault; like It is determined that the SINS / GNSS integrated navigation system has no sudden fault.

3. The method according to claim 2, characterized in that In view of the slow-varying faults with probability in GNSS position data, the sequential probability ratio algorithm SPRT is used. When a slow-varying fault occurs, the sequential probability ratio increment is constructed, and a slow-varying fault discrimination model based on the sequential probability ratio increment is established, and a threshold is set. When the sequential probability ratio increment value is negative, it is determined that the SINS and GNSS systems contain slow-varying faults, including: First, assume that the residual value at time k is v(k); the sequence of independent random samples belonging to it at consecutive k times is {v(1), v(2), …, v(k)}. According to the central limit theorem of probability theory, we know that: in and (σ(k)) 2 They represent the sample mean and sample variance respectively; when the GNSS output information has no slowly varying fault, that is, the null hypothesis H0, and has a slowly varying fault, that is, the alternative hypothesis H1, the maximum likelihood functions are: Based on the above binary hypothesis testing theory, we now take the residual sequence {v(1), v(2), …, v(k)} at k independent moments, and the likelihood ratio is: Take the logarithm of both sides of the likelihood ratio function and then use the sample mean Replacing μ approximately, we get the likelihood ratio as follows: The sequential probability ratio increment is: Considering the concepts of false alarm rate and missed detection rate of fault detection, the following inspection threshold is set: Among them, P M represents the missed detection rate, P F represents the false alarm rate; Based on the above analysis, the following conditions are established to determine whether the GNSS output information has a slowly varying fault: If λ(k)≥T(H1), the null hypothesis H0 is judged to be false, and the alternative hypothesis H1 is true; If λ(k)≤T(H0), then the null hypothesis H0 is judged to be true, and the alternative hypothesis H1 is false; If T(H0)≤λ(k)≤T(H1), more GNSS data information needs to be added to continue the inspection.

4. The method according to claim 3, characterized in that Step 3: According to the detection results of the MD chi-square algorithm and the SPRT algorithm in step 1 and step 2, determine whether there are faults in the SINS and GNSS systems, and determine whether there are sudden change faults and slow change faults; As long as the system has a sudden fault or a slow fault, it is determined that the SINS and GNSS systems have a fault, including: For the MD chi-square-sequential probability ratio algorithm, it is possible to simultaneously detect the sudden change faults and slow change faults in the GNSS output information; When the sequential probability ratio increment is positive, it is considered that the system has a slowly varying fault, and step 4 is executed to perform online reconstruction using the LSTM neural network method; when the sequential probability ratio increment is negative, step 5 is executed to perform robust enhancement based on the M-estimation robust Kalman filtering algorithm to eliminate the impact of sudden faults on the positioning results.

5. The method according to claim 4, characterized in that Step 4, the formation of the LSTM neural network includes the following steps: Step 41, training phase: when no fault occurs in the system, the output of the accelerometer and the gyroscope is used as the input of the LSTM, and the position information output by the GNSS is used as the output of the LSTM to complete the LSTM training phase; Step 42, prediction phase: when the system contains a slowly varying fault, the output of the trained LSTM model is used to replace the GNSS information containing the slowly varying fault to complete the reconstruction of the system; after the reconstruction work is completed, the integrated navigation is performed again.

6. The method according to claim 5, characterized in that Step 5: When the GNSS position information is detected to contain only mutation faults, in order to suppress the impact of measured wild values ​​on the integrated navigation system, a weighted matrix restriction and a modified innovation matrix are introduced, including: When the sequential probability ratio increment is negative, the MD chi-square test result is referred to; if the MD chi-square test result indicates that the system has a sudden fault, the robust Kalman filter algorithm based on M estimation is used to enhance the system robustness; Define the criterion function as shown below: where the function ρ(·) is a properly selected continuous convex function, is the i-th element of the new information at time k, and m is the dimension of the quantity measurement; a robust scale adjustment parameter is introduced To standardize the new information, the criterion function is rewritten as: in It also becomes the mean square error estimate of the i-th element of the new information at time k, It is called standardized news; Calculated by the following formula: Where MAD stands for median absolute deviation and is defined as: In the formula, med represents the median, That is, the new information sequence of the i-th measurement component as of the current moment; Taking the derivative of with respect to the estimated quantity X and setting it to 0 yields: in definition and matrix ω(·) is called the weight function, so the matrix form is written as: OR k oh k e k =0 The above is the principle of M estimation; therefore, the application of M estimation in Kalman filtering is to introduce the matrix ω k As the weighted matrix of the new information, it is shown as follows: The robust Kalman filtering algorithm based on M estimation is obtained.

Images