A multi-navigation sensor system online fault detection and repair method and system

Abstract

The application discloses a kind of multi-navigation sensor system online fault detection and repair method and system, based on multiple navigation sensors, constructs multi-navigation sensor system online fault detection and repair model, studies the joint detection of multiple types and multiple scales faults, adaptive robust filtering under non-smooth fault, the redundancy information acquisition and redundancy measurement voting method under slowly varying fault, the key algorithm such as online intelligent prediction method based on time series, designs the fault detection and information fusion scheme under the output information frequency inconsistency of each navigation sensor, simultaneously gives the numerical simulation verification and car test verification result of the online fault detection and repair method of full source navigation system.The application can complete the detection and identification of dynamic multiple faults, improve the system robustness and noise immunity, optimize the redundancy voting configuration and potential fault repair, realize the online fault detection and repair of multi-navigation sensor system.

Description

Technical Field

[0001] This invention relates to the field of multi-navigation sensors, and in particular to an online fault detection and repair method and system for multi-navigation sensor systems. Background Technology

[0002] Multi-navigation sensor systems aim to achieve navigation and positioning in the absence of satellite signals, requiring centralized processing of all information sources to maximize the advantages of multi-navigation sensor information fusion. Research on online fault detection, robustness improvement, redundancy voting, and intelligent active repair methods for multi-navigation sensor systems is crucial for enhancing their information quality and immunity to interference.

[0003] In the field of online fault detection algorithms for multi-navigation sensor systems, the complexity, hierarchy, and correlation of faults have driven the development of different types of fault detection methods. The basic idea of ​​fault detection is to perform online detection based on the residual characteristics represented by the measured signal when a fault occurs, thereby identifying the fault. Since the predicted residual sequence directly reflects the difference between system disturbances and faults and measurement noise and faults, it can represent the combined impact of both on the system. Therefore, the chi-square test of residuals is one of the most commonly used methods for fault detection in Kalman filters. However, due to the existence of false alarms in the chi-square test algorithm, its application in the self-filter of a federated Kalman filter can easily lead to frequent switching of the self-filter. Furthermore, the chi-square test of residuals has low sensitivity to slowly changing small-value soft fault information, failing to meet the high precision and high reliability requirements of multi-navigation sensor systems. In multi-navigation sensor systems, there are often multiple output sources for a single signal; therefore, redundant signal sources can serve as reference benchmarks for mutual detection of whether the output signals of their own signal sources have failed, i.e., redundant signal source mutual detection. In multi-navigation sensor systems, redundant information is obtained by measuring the same navigation parameters using multiple redundant sensors to acquire residuals. These residuals are then compared to set thresholds for fault detection, achieving high sensitivity and accuracy for some abrupt faults. However, its application is limited by space, weight, and cost. Furthermore, in multi-navigation sensor systems, there is a possibility that the output signal has only a single source, making fault detection algorithms based on signal redundancy unsuitable. Analytical model-based fault detection techniques can fully utilize deep-seated information within the system and avoid the low efficiency and high cost of hardware redundancy. However, accurate system modeling is difficult to achieve, and uncertainties such as modeling errors, parameter uncertainties, and unknown disturbances restrict the development of analytical model-based fault detection methods. The introduction of the Sequential Probability Ratio (SPRT) fault detection algorithm can fully utilize historical sensor data for self-checking and can also detect slowly changing faults to a certain extent. However, the current delay in slowly changing fault detection severely restricts the development and application of this technology. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for online fault detection and repair of multi-navigation sensor systems, thereby achieving online fault detection and repair of multi-navigation sensor systems and improving the accuracy of repair.

[0005] The technical solution to achieve the purpose of this invention is as follows:

[0006] To address the issue of degraded navigation signal quality caused by complex interference and the flexible movement of the carrier in complex environments, this invention proposes an online fault detection and repair method for multi-navigation sensor systems. It establishes an online fault detection and repair model for multi-navigation sensor systems, enabling the detection and identification of different types of faults in different sensors. This method improves the robustness of the multi-navigation sensor system, facilitates multi-sensor information fusion under optimal redundancy voting, and predicts and repairs potential faults. Specifically, it includes:

[0007] Collect navigation signals from all sensors;

[0008] The residual chi-square fault detection algorithm is used to detect sudden faults in the navigation signals of each sensor, and the sequential probability ratio algorithm is used to detect slowly changing faults in the navigation signals of each sensor. If a sudden fault is detected, the sudden fault is processed by robust filtering to obtain a normal signal. If only slowly changing faults are detected, the fault signal is reconstructed by redundancy voting algorithm to obtain a normal signal.

[0009] Training samples are constructed by calculating residuals based on navigation historical signals from each sensor. The residual signals at future times are predicted using an RBF neural network model. The predicted residual signals are then classified and identified using a BP neural network model to predict faults. Based on the fault prediction results, the normal signals of each sensor are then predicted.

[0010] Joint decision-making is performed. If the fault is abrupt, the normal signal obtained by processing the abrupt fault through robust filtering is used as the final sensor navigation signal. If a slowly changing fault is detected, the normal signal obtained by reconstructing the fault signal through a redundancy voting algorithm is used as the final sensor navigation signal. If the fault is predicted based on a neural network model, the predicted normal signals of each sensor are used as the final sensor navigation signal. If no fault is detected in any of the above cases, the normal signals of each sensor are used as the final sensor navigation signal.

[0011] An online fault detection and repair system for a multi-navigation sensor system includes a sensor data acquisition module, a fault detection module, a robustness enhancement module, a redundancy voting module, an active repair module, and a sensor signal output module, wherein:

[0012] The sensor data acquisition module is used to collect all sensor navigation signals and transmit them to the fault detection module and the active repair module.

[0013] The fault detection module performs abrupt fault detection on the navigation signals of each sensor based on the residual chi-square fault detection algorithm, and performs gradual fault detection on the navigation signals of each sensor using the sequential probability ratio algorithm. If an abrupt fault is detected, the data is processed through the robust enhancement module. If only a gradual fault is detected, the data is processed through the redundancy voting module.

[0014] The robust enhancement module processes sudden faults through robust filtering to obtain normal signals, which are then transmitted to the sensor signal output module.

[0015] The redundancy voting module reconstructs the fault signal to obtain the normal signal through the redundancy voting algorithm and transmits it to the sensor signal output module.

[0016] The active repair module calculates residuals based on the navigation history signals of each sensor, predicts the residual signals at future times using an RBF neural network model, classifies and identifies the predicted residual signals using a BP neural network model, performs fault prediction, and then predicts the normal signals of each sensor based on the fault prediction results, and transmits them to the sensor signal output module.

[0017] The sensor signal output module performs joint decision-making on the input signals. If a sudden fault is detected, the normal signal obtained by the robust enhancement module is used as the final sensor navigation signal. If only a slowly changing fault is detected, the normal signal obtained by the redundancy voting module is used as the final sensor navigation signal. If a fault is predicted based on a neural network model, the normal signal predicted by the active repair module is used as the final sensor navigation signal. If no fault is detected, the normal signals of each sensor are used as the final sensor navigation signal.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0019] (1) A fault detection information preprocessing method was proposed, and a joint fault detection algorithm model of sudden change and slow change was established, realizing fast and accurate fault detection under unknown fault conditions;

[0020] (2) A filtering model was established for unknown disturbances in complex environments, which simultaneously estimated the disturbances and states in the sense of minimum variance and performed stability analysis on the system. The optimal state estimate can still be obtained even in the presence of unknown disturbances.

[0021] (3) A weighted average voting method based on statistical information was adopted, which uses multiple measurements of the same navigation information provided by multiple navigation devices or redundancy analysis to vote out a unique navigation information, thereby improving the accuracy and high reliability of navigation information.

[0022] (4) A data-driven approach is proposed to propose an online intelligent fault prediction method based on time series. At the same time, the analytical model is combined with the time series model based on deep network to establish a time series prediction model of system residuals. Fault trend information is extracted through residual analysis, and accurate prediction of the state and navigation information of multiple navigation sensor faults is realized. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the simulation verification platform architecture design in this invention.

[0024] Figure 2 This is a schematic diagram of the probability ratio function trajectory in this invention.

[0025] Figure 3 This is a schematic diagram illustrating the delay in fault detection using the SPRT method in this invention.

[0026] Figure 4 This is a schematic diagram of the improved SPRT joint fault detection algorithm scheme in this invention.

[0027] Figure 5 This is a schematic diagram of the adaptive robust filtering algorithm in this invention.

[0028] Figure 6 This is a schematic diagram of the fault prediction time window in this invention.

[0029] Figure 7 This is the actual test route map of the unmanned vehicle in this invention.

[0030] Figure 8 This is a diagram showing the satellite navigation signal fault detection results in the sports car test case one of this invention.

[0031] Figure 9 This is a diagram showing the results of satellite navigation signal fault handling in the sports car test case 1 of this invention.

[0032] Figure 10 This is a diagram showing the satellite navigation signal fault detection results in the second sports car test case of this invention.

[0033] Figure 11 This is a diagram showing the results of satellite navigation signal fault handling in the second sports car test case of this invention.

[0034] Figure 12 This is a diagram showing the failure detection results of the visual odometer in the sports car test case three of this invention.

[0035] Figure 13 This is a diagram showing the failure handling results of the visual odometer in the sports car test case three of this invention.

[0036] Figure 14This is a diagram showing the failure detection results of the visual odometer in the sports car test case four of this invention.

[0037] Figure 15 This is a diagram showing the failure handling results of the visual odometer in the fourth test case of the sports car in this invention.

[0038] Figure 16 This is a diagram showing the fault detection results of the laser odometer in the fifth test case of the sports car in this invention.

[0039] Figure 17 This is a diagram showing the failure handling results of the laser odometer in the fifth test case of the sports car in this invention. Detailed Implementation

[0040] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0041] This invention proposes an online fault detection and repair workflow suitable for multi-navigation sensor systems. Using an inertial navigation system (INS) as the main navigation system, the system calculates the trajectory based on gyroscope and accelerometer information measured by the inertial components and combined with Earth parameters. Simultaneously, the difference between sensor measurements and the INS calculation results is used as input for fault detection and processing. Finally, the sensor measurements after fault processing are used to correct the INS calculation results. This locally loosely coupled integrated navigation framework improves the overall accuracy of the multi-navigation sensor system. Addressing the challenges of traditional fault detection methods struggling to quickly and effectively identify abrupt and gradually changing faults under various unknown sensor interferences, and the poor effectiveness of subsequent fault processing, a joint fault detection algorithm is designed to accurately identify the fault type. Furthermore, a robust enhancement, redundancy voting, and active repair joint fault processing algorithm are designed based on the fault type. The system architecture design is as follows: Figure 1 As shown, it includes a sensor data acquisition module, a fault detection module, a robustness enhancement module, a redundancy voting module, an active repair module, and a sensor signal output module. The implementation process is as follows:

[0042] (1) The vehicle platform is equipped with different types of navigation sensors, which are input to the software of this invention for centralized processing through the sensor data acquisition module, and execute the core unit of multi-navigation sensor system fault detection and processing.

[0043] (2) First, the input signal enters the fault detection module. Based on the residual chi-square fault detection algorithm and the sequential probability ratio algorithm, the navigation signals of each sensor are subjected to sudden and gradual fault detection respectively. After the detection signal passes through the fault identification module, the fault detection result is given, and then it enters the fault repair module.

[0044] (3) Then, if the fault detection module diagnoses a sudden fault, the signal enters the robust enhancement module, the sudden fault is processed by robust filtering, and finally the normal signal processed by each sensor is output.

[0045] (4) If the fault detection module diagnoses a slowly changing fault, the signal enters the redundancy voting module, which constructs a hardware redundancy library and an analytical redundancy library through the full-source navigation system, reconstructs the fault signal through the redundancy voting algorithm, and finally outputs the normal signal processed by each sensor.

[0046] (5) At the same time, the input signal enters the active repair module. This module calculates the residual based on the navigation history signals of each sensor, predicts the residual signal at future time through the RBF neural network model, classifies and identifies the predicted residual signal through the BP neural network model, and realizes fault prediction. At the same time, it gives the normal output signal of each sensor at future time based on the fault prediction result.

[0047] (6) Finally, the signal processing results of the robust enhancement module, the redundancy voting module and the active repair module are fused and output to obtain the high-precision sensing navigation signal after fault detection and processing.

[0048] This invention patent considers multiple navigation sensors such as Stim300, RTK, vision, and radar. It uses inertial navigation calculation as the state space equation and other measurement sensors as the measurement equation to realize local integrated navigation and perform online fault detection and repair of multi-navigation sensor systems. The implementation process includes the following 5 specific steps.

[0049] Step 1: Sensor navigation signal fault detection

[0050] (1) Multi-threshold residual chi-square mutation fault detection

[0051] The conventional residual chi-square test sets a threshold; if the detected quantity exceeds this threshold, it is considered faulty; otherwise, it is considered fault-free. Here, we further refine the fault threshold to reduce the false alarm probability based on Boolean logic's binary choice of either zero or one. First, we set T... D1 T D2 T D3 Three thresholds:

[0052] When the detection quantity λ k <T D1 When T is reached, the system is determined to be fault-free; here T D1 Set as the off-field threshold value determined by the false alarm rate in conventional chi-square detection;

[0053] When T D1 ≤ k <T D2 At that time, the system detected an infield value fault and issued an infield value alarm message. Here, T D2 =0.618T D3 T D3 Set as the threshold value determined by the false alarm rate in conventional chi-square detection;

[0054] When T D2 ≤ k <T D3 When the system detects a small field value fault, it issues a large field value alarm and initiates a fault alert process. Here, T... D3 Set as the threshold value determined by the false alarm rate in the conventional chi-square test method;

[0055] When λ k ≥T D3 When the system detects an over-the-counter (OTC) value fault, it will display an OTC value alarm message.

[0056] Based on the above analysis, the constructor is:

[0057]

[0058] The purpose of the design is to make μ(λ) k Mapping to the interval [0,1], we can then use μ(λ) k The value range distribution of the fault can intuitively determine the degree of fault. Furthermore, the fault degree discrimination and classification rules can be set as follows:

[0059]

[0060] Assuming the fault determination criterion is the occurrence of consecutive outliers, the continuous determination function based on probability density (membership degree) within the time window is designed as follows:

[0061]

[0062] Where, ρ f This represents the probability of a suspected fault signal occurring within the time window. This is the probability threshold for determining the occurrence of a fault signal within a time window. γ f =1 indicates that the current signal is a suspected fault signal; otherwise, γ f =0 indicates that the current signal is a normal signal.

[0063] (2) F-SPRT slow-varying fault detection based on fading factor

[0064] When the system is working normally, the innovation sequence is a zero-mean white noise sequence. Once a system fault occurs, the innovation sequence will change, and fault detection is performed by tracking this change. The residual η is generated using Kalman filtering. i =Z i -Z i,i-1 =Z i -H i X i,i-1 As the detection sample, where i = 0, 1, ..., k, η i Satisfies a normal distribution:

[0065] E(η i ) = 0

[0066] When a system failure occurs, both the mean and variance will change. To detect system failures, we make two assumptions:

[0067] H0: The innovation sequence is zero-mean white noise, and the system is operating normally. E(η) i )=0,Var(η i ) = 1

[0068] H1: The mean of the new information sequence is not zero or the variance is not 1, indicating a system malfunction. E(η) i )=a1≠0,Var(η i )=δ 2 ≠1.

[0069] Two test strengths are given, where α represents the allowable false alarm probability and β represents the allowable false alarm probability.

[0070] make

[0071] A=ln(β / (1-α)),B=ln((1-β) / α)

[0072] Calculate the probability ratio:

[0073]

[0074] The decision rule for fault detection is

[0075] ①If λ k If B ≥ B, then the system is considered to have failed before time k, and a system failure is indicated at time k.

[0076] ②If λ k If A ≤ k, then the system is considered to be normal before time k, and the system is indicated to be normal at time k.

[0077] ③If A < λ k If B < B, a potential fault is considered to exist. At time k, the system is considered normal, and the test continues. If the variance of the information sequence changes after the fault, further testing is conducted.

[0078] H1:E(η i )=0,Var(η i )=δ 2 >1

[0079] From the probability ratio equation, we obtain the recursive formula:

[0080]

[0081] If both the mean and variance change

[0082] H1:E(η i )=a1,Var(η i )=δ 2 >1

[0083] From the probability ratio equation, we obtain the recursive formula:

[0084]

[0085] Whether it's the change in mean or the change in variance, the curve shape is as follows: Figure 2 As shown:

[0086] From the functional relationship and the curve, we can see that when the system is working normally, generally speaking, λ k It tends to drift negatively, and when the system fails, λ k Tends to drift in the positive direction, see Figure 2 Therefore, there is always a delay between the system's normal operation and the occurrence of a fault, and between negative drift and positive drift. (See...) Figure 3 This is the shortcoming of the traditional SPRT method.

[0087] To address this deficiency, an improved SPRT fault detection algorithm is proposed when the detection quantity x follows a normal distribution. This algorithm ensures that the detection value λ(k) cannot be negative, thus avoiding the detection delay.

[0088] Let {x} be the k-order sequential independent samples of an unknown normal random variable x. i If |i=1,2,…,k}, then by the principles of probability theory and mathematical statistics, we can obtain an approximate result: Sample mean and variance They are respectively represented as

[0089]

[0090]

[0091] Assume the actual measured value of state variable x is x * .

[0092] In a fault-free state, x * =x0, where x0 is the mean of the true values ​​under fault-free conditions;

[0093] When there is a fault,

[0094] If H0 is defined as x * =x0;H1: Then the measurement sequence x1,x2,…,x kIt must belong to one of two sample classes, H0 (normal class) and H1 (fault class), and its probability density function is:

[0095]

[0096]

[0097] The likelihood ratio can be obtained as follows:

[0098]

[0099] Taking the natural logarithm of the above equation, we obtain the log-likelihood ratio as:

[0100]

[0101] After simplifying the above equation, we get:

[0102]

[0103] When the system is working normally, as k increases, the sample mean... The sample mean gradually approaches x0, i.e., λ(k)→0. When a system failure occurs, the sample mean decreases as k increases. The algorithm gradually approaches the true value under fault conditions, but this true value differs from x0. Analysis shows that λ(k) cannot be negative, thus eliminating the detection delay defect. However, the SPRT algorithm still suffers from the following problem: as k increases (k→∞), 1 / k→0, resulting in a very small proportion of the residual at time k, and the sample mean cannot reflect new measurement information in a timely manner. When a slowly changing fault occurs, even if the SPRT algorithm superimposes the faults, the sensitivity of the slowly changing fault detection system is significantly reduced due to the small weighting coefficient of the new measurement values. Therefore, the F-SPRT algorithm is proposed, introducing a fading factor s to gradually reduce the weight of historical information, increase the weight of new measurement values, and improve the sensitivity for detecting slowly changing faults.

[0104] Introducing a fading memory factor, the sample mean of F-SPRT is obtained:

[0105]

[0106] The sample variance of F-SPRT:

[0107]

[0108] In summary, the recursive formula for calculating the likelihood ratio λ(k) is obtained as follows:

[0109]

[0110] If the prior variance of variable x is known to be σ 2 Then it can be used as a substitute. Therefore, we can conclude that:

[0111]

[0112] If the prior variance of variable x is unknown, the sample variance calculated in real time can be used instead.

[0113] Wald's sequential probabilistic algorithm provides two detection thresholds, T(H0) and T(H1). The improved algorithm described above only requires setting one detection threshold, T. FSPRT T FSPRT The calculation is consistent with the Wald algorithm, that is, it is based on the false alarm rate P. f and false alarm rate P m Sure:

[0114]

[0115] Therefore, the fault judgment criterion is:

[0116]

[0117] Clearly, the traditional SPRT algorithm is a special case when s=1. The choice of s value has a certain impact on the fault detection sensitivity. If it is too large, the gradual decay of memory cannot be achieved; if it is too small, noise will cause instability in the sample mean when there is no fault, leading to false alarms. Generally, s is taken as 0.95 to 1.

[0118] (3) Joint Fault Detection Algorithm

[0119] The above describes fault detection for abrupt and gradual faults. Since the fault type is not known in advance during actual operation, a combination of these two fault detection algorithms is needed to cover all fault types. A parallel approach is used to detect faults in the sensor's output signal. (See...) Figure 4 Because gradual and sudden faults can interfere with each other during the detection process, the algorithm will detect both gradual and sudden faults when a sudden fault occurs. However, it generally will not detect sudden faults when a gradual fault occurs. Therefore, sudden faults have a higher priority than gradual faults in the final comprehensive diagnostic result.

[0120] Both test results are input into the system fault comprehensive decision-making process for final fault determination, thus obtaining the final test result. "Yes" and "No" indicate whether a fault has occurred. The system fault comprehensive decision-making rules are shown in Table 1.

[0121] Table 1 System Fault Comprehensive Decision Rules

[0122] Test result 1 (mutation) Result 2 (Gradual Change) Comprehensive test results (joint) yes yes Yes (robust enhancement) yes no Yes (robust enhancement) no yes Yes (remaining vote) no no No (Active Repair)

[0123] (4) To address the issue of asynchronous data acquisition and recording frequencies among various sensors, time synchronization preprocessing research is required before algorithm verification. Four time synchronization schemes are considered and analyzed as follows:

[0124] Option 1: Using the inertial navigation system (INS) timestamp as a benchmark (INS has the highest update frequency), perform single-sample INS calculations and record the current timestamp. Compare this timestamp with the sampling timestamps of other sensors. If the INS sampling timestamp is greater than the timestamps of other sensors at the next moment, then perform loosely combined navigation algorithm processing on the current sensor's output information. If the INS timestamp is still less than the timestamps of other sensors, then only perform pure INS calculations.

[0125] Option 2: Using the inertial navigation system (INS) timestamp as the reference (highest INS update frequency), perform INS calculations at a sampling frequency of double sampling or higher. If there are signal outputs from other sensors between multiple INS calculation samplings, then execute the loosely combined navigation algorithm. If there are no signal outputs from other sensors between samplings, then only pure INS calculations are performed.

[0126] Option 3: In Option 1, there may be data duplication issues due to timestamp synchronization issues with the current timestamps of each sensor. Therefore, a difference threshold is introduced to constrain time synchronization. The currently sampled inertial navigation (INS) timestamp and the timestamps of other sensors are recorded and their differences are calculated. A threshold lower than the INS sampling step size is designed and compared with this threshold for time synchronization constraints. If the difference between the current INS timestamp and the timestamps of other sensors is lower than the threshold, the two sensors can be approximately considered to be in time synchronization, and a loosely combined navigation algorithm is executed. If the difference between the current INS timestamp and the timestamps of other sensors is higher than the threshold, the two timestamps are considered to be out of time, and only pure INS calculation is performed.

[0127] Option 4: In the above integrated navigation schemes, pure inertial navigation calculations are performed when low-frequency sensors have no output. Different frequency sensors approximate timestamps for alignment, making it difficult to guarantee high-precision navigation information. To ensure high-precision navigation information output, this scheme proposes to perform high-frequency reasonable extrapolation processing on low-frequency sensors. Taking sensor navigation information over a certain time series, when the inertial navigation system (highest frequency) has output but other auxiliary sensors have not yet output, historical data from the auxiliary sensors is used as a benchmark. Reasonable extrapolation is performed according to the current timestamp of the inertial navigation system to supplement the theoretical navigation information of the auxiliary sensors at the inertial navigation timestamp. Integrated navigation is then performed using a loosely combined approach. If the output frequency of the inertial navigation system and the output frequency of the auxiliary sensors have a least common multiple, then no extrapolation is needed, and integrated navigation can be performed directly.

[0128] By comparing and analyzing the above four schemes, Scheme 4 was selected as the time synchronization scheme in this invention.

[0129] Step 2: Adaptive robust filtering (robust enhancement) of navigation signals under non-gradually changing faults

[0130] (1) Anti-displacement factor

[0131] To construct a robustness factor, first, abrupt fault detection is performed on the measurement innovation. If r k ≤T D If the observation is normal, the sensor signal is not directly output and no repair is performed; if r k >T D It is determined that the observation contains outliers, requiring the construction of a variance inflation factor to handle these outliers and fully utilize the effectiveness of the sensor's output signal. The diagnostic results are given in step 1.

[0132] Now, construct the variance inflation factor:

[0133]

[0134] make

[0135] β = diag[β1 β2 … β] n ],0<β i ≤1, i=1,2,…,n

[0136] Therefore, the corresponding equivalent robust covariance matrix is

[0137]

[0138] but:

[0139]

[0140] When there is random interference in the measurement, robust Kalman filtering improves the output accuracy of the system by increasing the variance of the corresponding measurement noise and reducing the Kalman gain.

[0141] (2) Adaptive robust algorithm

[0142] The innovation of the adaptive robust algorithm lies in introducing an adaptive factor to allocate the weights of observations and measurements during information fusion, thereby ensuring accuracy. The difference between adaptive filtering and standard Kalman filtering lies in α. k The carrier is inevitably subject to external interference, resulting in deviations in the predicted state. Therefore, an adaptive factor α is introduced on the basis of standard Kalman filtering. k and α k The state covariance matrix applied to the measurement update at time k plays a significant role in reducing the error of the state estimate during filtering. Adaptive factor α k By changing the prior covariance in the Kalman filter, the Kalman gain is affected, and finally the weight of the innovation vector in the observation correction process is determined to ensure accuracy.

[0143] The prior covariance P of the standard Kalman filter k,k-1 Can be updated to

[0144]

[0145] To obtain the value of the adaptive factor α, the following calculation is performed:

[0146]

[0147] In the formula, tr[·] denotes the trace of a matrix.

[0148]

[0149]

[0150]

[0151] In the formula, V0 is the initial parameter for calculating the robustness factor; c is the free scaling parameter, and the value of c makes the filter calculation more flexible, ensuring the overall tracking capability and accuracy of the filter; H k Let Q be the filter measurement matrix. k Let R be the variance matrix of the system noise sequence. k The variance matrix of the measurement noise sequence; the adaptive factor satisfies 0 < α k ≤1. When α k When α = 1, the filtering is standard Kalman filtering; when the carrier is in a non-stationary state or when there are dynamic model errors, α k Satisfying 0 < α k <1, at this point the adaptive factor will simultaneously increase both the prior state covariance matrix and the state perturbation covariance matrix by 1 / α. k The factor 'c' adaptively adjusts the fusion weights of the dynamic model information and the observation information to improve positioning accuracy. The value of 'c' makes the filtering calculation more flexible, ensuring the overall tracking capability and accuracy of the filter.

[0152] Based on the above analysis, using the adaptive factor α k While it can improve the convergence and stability of the filter to some extent, it cannot effectively resist the influence of accidental measurement interference on the filter output. Therefore, a robustness factor needs to be introduced into the adaptive Kalman filter. The robust adaptive filtering algorithm flow within a single cycle is as follows: Figure 5 Specifically, based on traditional Kalman filtering technology, an adaptive factor α is introduced during the time update process to adaptively adjust the prior state covariance, while a robust factor β is introduced during the measurement update process to reflect the actual noise scale characteristics of the current test, thereby improving the accuracy of the traditional Kalman filter.

[0153] Step 3: Navigation signal redundancy voting under gradual fault conditions

[0154] Since the sensitivity of each redundancy signal source to error is different, the quality of their measurement signals (mean error, standard deviation and confidence level) varies. Therefore, the final voting signal is calculated by weighting the measurement signals of each redundancy.

[0155]

[0156] In the formula, n is the information redundancy number, and V i The effective factor (V) for the i-th redundancy measurement information i =0 indicates a fault in the measurement value, V i =1 indicates that the measurement value is normal), h i (t) represents the measured value of the i-th redundant signal; K i Let be the weighting factor for the i-th redundancy. This is the voting value for information.

[0157] Introducing a weighting factor can correlate the voting value with the actual performance of each redundancy signal, increasing the proportion of high-quality sensor output signals (smaller mean and standard deviation of error) in the repaired navigation information, and reducing the proportion of low-quality sensor output signals in the repaired navigation information, so that the final voting value is closer to the voting value of the real signal.

[0158] The method for calculating the weighting factor:

[0159]

[0160] In the formula, V i The effective factor (V) for the i-th redundancy measurement information i =0 indicates a fault in the measurement value, V i =1 indicates that the measurement value is normal, μ i Let σ be the mean error of the i-th redundant sensor. i Let be the error variance of the i-th redundant sensor.

[0161] Considering the prior confidence information, which can also reflect the quality of sensor output information, we introduce it into the weighting factor calculation formula:

[0162]

[0163] In the formula, V i The effective factor (V) for the i-th redundancy measurement information i =0 indicates a fault in the measurement value, V i =1 indicates that the measurement value is normal), λ i Let μ be the confidence level of the i-th redundant sensor. iLet σ be the mean error of the i-th redundant sensor. i Let be the error variance of the i-th redundant sensor.

[0164] In practical systems, the mean error is generally smaller than the variance error, so the mean error term in the above equation can be ignored. Therefore, the equation can be rewritten as:

[0165]

[0166] It can be seen that, compared to the median voting method and the arithmetic mean voting method, the improved redundancy signal voting method considers the actual quality of each redundancy signal and can improve the contribution value of high-precision sensors. However, if there is a sudden change in the confidence level of the navigation sensor, the weighted mean voting method cannot handle such situations well. Therefore, a sliding window is added to reduce the impact of problems such as sudden changes in confidence level.

[0167]

[0168] Among them, K j (k) represents the weighting factor at the current time, n is the sliding window length, and K j (ki) represents the weighting factor at the previous time step.

[0169] As can be seen from the above formula, compared with using the influence factor coefficients (standard deviation and confidence level) of the sensor at a certain moment, the improved voting method by adding a sliding window comprehensively considers the redundancy voting value in historical information, thereby improving the stability of the repair signal voting value.

[0170] Step 4: Navigation signal prediction based on neural networks

[0171] (1) Constructing neural network training samples

[0172] Since obtaining comprehensive fault data samples of a full-source integrated navigation system under multi-task profiles online is impractical, a combination of theoretical analysis and mathematical simulation will be used to establish a navigation database based on typical scenarios and fault types. A combination of fault propagation mechanism analysis and feature extraction will be employed to uncover the complex correlation mapping between potential system faults and system parameters, constructing a fault mode library and a system for evaluating system operating status. This database will be used to generate a training sample set to support fault prediction model training, fault feature extraction and identification, and the acquisition of fault evolution patterns. To this end, a simulation analysis platform for a full-source integrated navigation system based on task profiles will be built. By simulating the injection of slowly varying faults at different scales into different navigation sensors at different times, the simulation will obtain fault signal propagation results from the device level to the system level, generating a fault signal propagation database under typical faults, and constructing a fault database accordingly. Currently, only one type of slowly varying fault is considered as a typical sample.

[0173] In the context of all-source navigation, assuming the reliability of the main signal source's output signal is guaranteed, the output error (system error) of the sub-sensors and the main sensor is calculated as the detection quantity for slowly varying faults. This way, the error does not update over time, achieving a more ideal fault detection effect. Due to the influence of the complex working environment, navigation system faults are characterized by unpredictability, uncertainty, and propagation. Therefore, a Boolean logic-based method will be used to determine the mapping rules between fault characteristics and fault modes, establishing a fault mode database for fault detection.

[0174] Training and test samples are established based on a fault mode library. In this invention, 90% of the data is randomly selected as training samples and 10% as test samples. Based on the training samples, the output signal is first predicted using an RBF neural network. Then, based on the system state prediction signal output by the RBF network, a BP network is used for fault detection, identification, and repair.

[0175] (2) Signal prediction based on RBF network

[0176] Radial Basis Function (RBF) neural networks utilize RBFs as the "basis" for hidden units to construct the hidden layer space. The hidden layers transform the input vector, converting the low-dimensional pattern input data to a high-dimensional space. This makes linearly inseparable problems in the low-dimensional space linearly separable in the high-dimensional space, allowing the input vector to be directly mapped to the hidden space (i.e., without weighted connections). Once the center point of the RBF is determined, this mapping relationship is also determined. The mapping from the hidden layer space to the output space is linear; that is, the network output is a linear weighted sum of the hidden unit outputs. Here, the weights are the network's adjustable parameters. Therefore, overall, the network's input-to-output mapping is non-linear, while the network output is linear with respect to the adjustable parameters. This allows the network weights to be directly solved from linear equations, significantly accelerating the learning speed and avoiding local minima.

[0177] The RBF neural network learning algorithm requires solving for three parameters: the center and variance of the basis functions, and the weights from the hidden layers to the output layer. Depending on the method used to select the RBF function centers, RBF networks employ various learning methods, such as random center selection, self-organized center selection, supervised center selection, and orthogonal least squares. This scheme adopts the self-organized center selection RBF neural network learning method. This method consists of two stages: the first is the self-organized learning stage, which is an unsupervised learning process, solving for the center and variance of the hidden layer basis functions (using K-means clustering); the second is the supervised learning stage, which solves for the weights from the hidden layers to the output layer. The specific algorithm is as follows:

[0178] The RBF function commonly used in RBF neural networks is the Gaussian function; therefore, the activation function of an RBF neural network can be expressed as:

[0179]

[0180] In the formula, ||x p -c i || is the Euclidean norm, c is the center of the Gaussian function, and σ is the variance of the Gaussian function.

[0181] The output of the RBF neural network is

[0182]

[0183] In the formula, This is the p-th input sample, where p = 1, 2, ..., P represents the total number of samples, and c i ω is the center of the hidden layer node in the network. ij The connection weights from the hidden layer to the output layer are y, where i = 1, 2, ..., h is the number of nodes in the hidden layer, and y is the connection weight from the hidden layer to the output layer. i This is the actual output of the j-th output node of the network corresponding to the input sample.

[0184] Let the slope d be the expected output value of the sample, then the variance of the basis function can be expressed as:

[0185]

[0186] The basis function centers c are solved using the K-means clustering method. The specific steps of the algorithm are as follows:

[0187] ①Network initialization

[0188] h training samples are randomly selected as cluster centers c i (i = 1, 2, ..., h)

[0189] ② Group the input training samples geometrically according to the nearest neighbor rule.

[0190] According to x p With center c i The Euclidean distance between them will be x p The various cluster sets θ assigned to the input samples p In (p=1,2,…,P).

[0191] ③ Readjust the cluster centers.

[0192] Calculate the θ of each cluster set p The average value of the training samples, i.e., the new cluster center c. i If the new cluster centers do not change, then the obtained c iThis is the final basis function center of the RBF neural network; otherwise, return to (1) and proceed to the next round of center solving.

[0193] ④ Solve for the variance σ i

[0194] The basis functions of this RBF neural network are Gaussian functions, therefore the variance σ i It can be solved by the following formula.

[0195]

[0196] In the formula, c max This represents the maximum distance between the selected centers.

[0197] ⑤ Calculate the weights between the hidden layer and the output layer.

[0198] The connection weights of neurons between the hidden layer and the output layer can be directly calculated using the least squares method, as shown in the following formula.

[0199]

[0200] The methods described above improve network performance by optimizing the parameters of the RBF, especially the center of the RBF, under the premise that the network structure is as simple as possible (i.e., the number of hidden layer units is as small as possible). These methods can achieve an infinite approximation effect, but they are relatively complex and cumbersome to calculate.

[0201] In actual fault prediction, it is necessary to perform system analysis based on several sample signals from the current and previous time windows, and then use an RBF neural network to solve for and predict the signal state at future times. A flowchart of the fault prediction process is shown below. Figure 6 Here, t represents the current time, Δt1 represents the observation time window, Δt represents the time interval, and Δt2 represents the prediction time window. The fault prediction task process is as follows: starting from time t, observe the operating status of the network system within the time window Δt1 before t, and then predict the operating status of the network system within the time window Δt2 after the time interval Δt based on the operating status of the network system at this time. It should be noted that increasing the prediction time window Δt2 can obtain more data, thus improving the accuracy of fault prediction. However, if the time window Δt2 is too large, that is, the prediction time range is too wide, and it is impossible to know when the fault will occur, which will defeat the purpose of prediction. At the same time, the time interval Δt cannot be too small. If Δt is too small, it will be impossible to take effective preventive measures before the fault occurs, which will also defeat the purpose of prediction.

[0202] (3) Fault identification based on BP neural network

[0203] The basic BP neural network algorithm consists of two aspects: forward propagation of the signal and backward propagation of the error. That is, the actual output is calculated from input to output, while the weights and thresholds are adjusted from output to input. The calculation process of the BP neural network is shown below.

[0204] ①The forward propagation process of the signal

[0205] The input net of the i-th node in the hidden layer i for

[0206]

[0207] The output y of the i-th node in the hidden layer i for

[0208]

[0209] The input net of the kth node in the output layer k for

[0210]

[0211] The output of the kth node in the output layer k for

[0212]

[0213] ② Backpropagation process of error

[0214] Backpropagation of error involves first calculating the output error of each neuron in the output layer, and then adjusting the weights and thresholds of each layer using the error gradient descent method, so that the final output of the modified network can approach the expected value.

[0215] For each sample p, the quadratic error criterion function is E p

[0216]

[0217] The system's total error criterion function for p training samples is:

[0218]

[0219] The correction amount Δw of the output layer weights is adjusted sequentially according to the error gradient descent method. ki The correction amount Δα for the output layer threshold and the correction amount Δw for the hidden layer weights. ij The correction amount Δθ for the hidden layer threshold i .

[0220] ③ Performance indicators

[0221] F(x)=E[e T e]=E[(ta) T (ta)]

[0222] ④ Approximate performance indicators

[0223]

[0224] ⑤ Sensitivity feedback

[0225]

[0226] ⑥ Forward propagation

[0227] a 0 =p

[0228] a m+1 =f m+1 (W m+1 a m +b m+1 ), m=0,1,…,M-1

[0229] a = a M

[0230] ⑦ Backpropagation

[0231]

[0232]

[0233] in,

[0234]

[0235] ⑧ Weight update (approximate steepest descent)

[0236] W m (k+1)=W m (k)-αs m (a m-1 ) T

[0237] b m (k+1)=b m (k)-αs m

[0238] In summary, an online fault detection and repair method for a multi-navigation sensor system includes the following design points: (1) fusion of measurement information from multiple navigation sensors and fault detection; (2) design of a method to improve the robustness of navigation information under different fault conditions (abrupt and gradual faults); (3) design of a redundancy voting method for the multi-navigation sensor system to handle fault signals that are difficult to repair from a single information source; and (4) optimization of fault detection strategy to achieve intelligent prediction and repair of gradual faults.

[0239] Example

[0240] Based on sports car testing, the claims of this invention are verified. The sports car's trajectory is as follows: Figure 7 As shown. Through analysis of the motion characteristics and sensor fault characteristics of autonomous vehicles in different scenarios, the invention points for fault detection and handling were verified under five typical cases. Specific experimental results are as follows:

[0241] Satellite interference navigation signal divergence case: Select the 150-400s trajectory of the sports car test to verify the correctness of the fault detection and handling of the full-source navigation system when the output information of the GPS integrated navigation module deviates. Figures 8-9 These are the results of GPS fault detection and handling. Specifically, fault detection flag 1 indicates the use of a robust boosting module, and flag 2 indicates the use of a redundancy voting module for fault handling. When the satellite navigation signal diverged, the tested software ran the redundancy voting module, and the satellite navigation signal was effectively repaired.

[0242] Abnormal values ​​of satellite navigation signal mutation: The trajectory of 700-1000s in the sports car test was selected to verify the correctness of the fault detection and handling of the full-source navigation system under the instability of satellite navigation output signal. Figures 10-11 The figures show the results of satellite navigation signal fault detection and processing. As can be seen from the figures, a sudden outlier occurred in the satellite navigation signal. The tested software ran the robust boosting processing module, and the satellite navigation signal was effectively repaired.

[0243] Visual odometer signal offset: The trajectory of the sports car test from 350 to 500 seconds was selected to verify the correctness of the fault detection and handling of the full-source navigation system under visual odometer signal offset. Figures 12-13 The figures show the fault detection and processing results for the visual odometer navigation signal. As can be seen from the figures, the visual odometer signal was offset. The tested software used the redundancy voting processing module, and the visual odometer navigation signal was effectively repaired.

[0244] Visual odometer signal bias: The trajectory of the sports car test from 700 to 1000 seconds was selected to verify the correctness of the fault detection and handling of the full-source navigation system under the bias of the visual odometer navigation output signal. Figures 14-15The figures show the fault detection and processing results for the visual odometry navigation signal. As can be seen from the figures, when the visual odometry signal was biased, the test software ran the redundancy voting processing module, and the visual odometry navigation signal was effectively repaired.

[0245] Laser odometry feature extraction failure: The trajectory of the sports car test from 350 to 500 seconds was selected to verify the correctness of the fault detection and handling of the full-source navigation system under the condition of laser odometry feature extraction failure. Figures 16-17 The figures show the fault detection and processing results of the laser odometry output signal. As can be seen from the figures, the laser odometry feature extraction failed. The tested software ran the redundancy voting processing module, and the laser odometry navigation signal was effectively repaired.

Claims

1. A method for online fault detection and repair of a multi-navigation sensor system, characterized in that, Including the following steps: Collect navigation signals from all sensors; The residual chi-square fault detection algorithm is used to detect sudden faults in the navigation signals of each sensor, and the sequential probability ratio algorithm is used to detect slowly changing faults in the navigation signals of each sensor. If a sudden fault is detected, the sudden fault is processed by robust filtering to obtain a normal signal. If only a slowly varying fault is detected, the fault signal is reconstructed using a redundancy voting algorithm to obtain a normal signal; Training samples are constructed by calculating residuals based on navigation historical signals from each sensor. The residual signals at future times are predicted using an RBF neural network model. The predicted residual signals are then classified and identified using a BP neural network model to predict faults. Based on the fault prediction results, the normal signals of each sensor are then predicted. Joint decision-making is performed. If the fault is abrupt, the normal signal obtained by processing the abrupt fault through robust filtering is used as the final sensor navigation signal. If a gradually changing fault is detected, the normal signal obtained by reconstructing the fault signal through a redundancy voting algorithm is used as the final sensor navigation signal. If a fault is predicted based on a neural network model, the predicted normal signals from each sensor will be used as the final sensor navigation signals. If no fault is detected above, the normal signals from each sensor will be used as the final sensor navigation signals. The residual chi-square fault detection algorithm is used to detect sudden faults in the navigation signals of each sensor, specifically including: set up , , Three thresholds Let this be the Ono threshold value determined by the false alarm rate in chi-square detection. , Set as the threshold value determined by the false alarm rate in chi-square detection; Constructor for: in, For the collected sensor navigation signal values, Mapped to the interval [0,1], according to The value range distribution is used to determine the degree of abrupt failure. The rules for determining and classifying the degree of abrupt failure are as follows: Assuming the fault determination criterion is the occurrence of consecutive outliers, the continuous determination function based on probability density within the time window is designed as follows: in, For based on The probability of a suspected sudden fault signal occurring within a time window determined by the value range distribution. This is the probability threshold for determining the occurrence of a sudden fault signal within a time window. hour This indicates that the current signal is a suspected sudden fault signal; otherwise... This indicates that the current signal is a normal signal.

2. The online fault detection and repair method for a multi-navigation sensor system according to claim 1, characterized in that, The sequential probability ratio algorithm for detecting slowly varying faults in navigation signals from various sensors specifically includes: Suppose an unknown normal random variable of Sequential independent samples are Then there is Sample mean and variance They are respectively: Assuming state variables The actual measured value is In the absence of a gradually changing fault state, , This represents the average of the true values ​​under fault-free conditions; under slowly varying fault conditions, ; If a normal class is defined : Fault Class : Then the measurement sequence Must belong to two sample classes and One of them, and its probability density function is: The likelihood ratio can be obtained as follows: Taking the natural logarithm of the above equation, we obtain the log-likelihood ratio as: After simplifying the above equation, we get: The decision rule for slowly changing fault detection is: ①If Then it is believed A slow-change fault had already occurred some time ago. The fault is a slow-changing fault. ②If Then it is believed Things were normal before. The time is normal; ③If If so, it is considered that a potential slowly changing fault exists. The system keeps indicating normal status; please continue testing. Among them, the threshold, , This represents the permissible probability of a false alarm. This represents the allowed false alarm probability.

3. The online fault detection and repair method for a multi-navigation sensor system according to claim 2, characterized in that, Introducing fading memory factor A sequential probability ratio algorithm is used to detect slowly varying faults in the navigation signals of each sensor. Specifically, this includes designing the log-likelihood ratio as follows: Set a detection threshold , , For false alarm rate, The fault judgment criterion is as follows: (This is to determine the false alarm rate.) 。 4. The online fault detection and repair method for a multi-navigation sensor system according to claim 3, characterized in that, Robust filtering is used to process abrupt faults and obtain normal signals. Specifically, this involves introducing an adaptive factor during the time update process based on the Kalman filter algorithm. The prior state covariance is adaptively adjusted, and a robust factor is introduced during the measurement update process. Reflecting the current actual noise scale characteristics, among which, The adaptive factor for: In the formula, This represents finding the trace of a matrix, where matrix M is the trace of a matrix. k N k for: In the formula, the adaptive factor satisfies , Calculate the initial parameters for the robustness factor; For free scaling parameters; This is the filter measurement matrix. Let Variance be the variance matrix of the system noise sequence. The variance matrix of the measurement noise sequence; The robustness factor for: in, For the log-likelihood ratio, T D The threshold value is used.

5. The online fault detection and repair method for a multi-navigation sensor system according to claim 4, characterized in that, Reconstructing normal signals from fault signals using a redundancy voting algorithm specifically involves: introducing a weighting factor into the redundancy voting algorithm to correlate the voting value with the actual performance of each redundancy signal. The formula for calculating the weighting factor is as follows: in, For the first Effective factor of redundancy measurement information For the first Confidence level of the redundancy sensor For the first Error variance of the redundancy sensor This represents the information redundancy.

6. The online fault detection and repair method for a multi-navigation sensor system according to claim 5, characterized in that, The weighting factor is calculated by incorporating a sliding window, and is as follows: in, Represented as the weighting factor at the current moment, The length of the sliding window. This represents the weighting factor at the previous time step.

7. The online fault detection and repair method for a multi-navigation sensor system according to claim 6, characterized in that, The RBF neural network model employs a self-organizing center selection RBF neural network learning method, comprising two stages: a self-organizing learning stage, in which K-means clustering is used to solve for the center and variance of the hidden layer basis functions; and a supervised learning stage, in which least squares is used to solve for the connection weights from the hidden layer to the output layer. The activation function of the RBF neural network model is: In the formula, It is the Euclidean norm. The center of the Gaussian function, Let be the variance of the Gaussian function; The output of the RBF neural network model is In the formula, It is the first One input sample, As the center of the hidden layer nodes in the network, These are the connection weights from the hidden layer to the output layer. , This represents the number of nodes in the hidden layer. The network corresponding to the input sample The actual output of each output node.

8. The online fault detection and repair method for a multi-navigation sensor system according to claim 7, characterized in that, The BP neural network model calculates the actual output in the direction from input to output, while the weight and threshold are corrected in the direction from output to input. Specifically, the weight and threshold correction is performed by first calculating the output error of each neuron in each layer, starting from the output layer, and then adjusting the weights and thresholds of each layer according to the error gradient descent method so that the final output of the modified network meets the accuracy requirements.

9. A multi-navigation sensor system online fault detection and repair system for implementing the online fault detection and repair method for any one of claims 1-8, characterized in that, It includes a sensor data acquisition module, a fault detection module, a robustness enhancement module, a redundancy voting module, an active repair module, and a sensor signal output module, among which: The sensor data acquisition module is used to collect all sensor navigation signals and transmit them to the fault detection module and the active repair module. The fault detection module performs abrupt fault detection on the navigation signals of each sensor based on the residual chi-square fault detection algorithm, and performs gradual fault detection on the navigation signals of each sensor using the sequential probability ratio algorithm. If an abrupt fault is detected, the data is processed through the robust enhancement module. If only a gradual fault is detected, the data is processed through the redundancy voting module. The robust enhancement module processes sudden faults through robust filtering to obtain normal signals, which are then transmitted to the sensor signal output module. The redundancy voting module reconstructs the fault signal to obtain the normal signal through the redundancy voting algorithm and transmits it to the sensor signal output module. The active repair module calculates residuals based on the navigation history signals of each sensor, predicts the residual signals at future times using an RBF neural network model, classifies and identifies the predicted residual signals using a BP neural network model, performs fault prediction, and then predicts the normal signals of each sensor based on the fault prediction results, and transmits them to the sensor signal output module. The sensor signal output module performs joint decision-making on the input signals. If a sudden fault is detected, the normal signal obtained by the robust enhancement module is used as the final sensor navigation signal. If only a slowly changing fault is detected, the normal signal obtained by the redundancy voting module is used as the final sensor navigation signal. If a fault is predicted based on a neural network model, the normal signal predicted by the active repair module is used as the final sensor navigation signal. If no fault is detected, the normal signals of each sensor are used as the final sensor navigation signal.

Images