A Smart Active Fault Tolerance Method for Gradually Changing Faults in Navigation Sensors

Abstract

This invention discloses an intelligent active fault-tolerance method for gradually changing faults in navigation sensors, comprising: constructing and training a standard signal prediction module using long short-term memory neurons as basic units to predict navigation signals; constructing and training a gradually changing fault diagnosis module using a multi-layer stacked convolutional neural network model, which extracts gradually changing fault trend information based on the residual between the predicted signal and the measured signal, and diagnoses the gradually changing fault in its early stages when its impact on the system is within a set range; if no gradually changing fault is diagnosed, the measured signal is used as the navigation signal; when a gradually changing fault is diagnosed, the predicted signal from the standard signal prediction module is used as the navigation signal, thus achieving active fault tolerance for gradually changing faults in the navigation signal. This invention can effectively predict lidar altitude information and solves the problems of slow gradually changing fault diagnosis and the limited application scenarios of traditional threshold methods.

Description

Technical Field

[0001] This invention belongs to the field of navigation, specifically relating to an intelligent active fault-tolerance method for gradually changing faults in navigation sensors. Background Technology

[0002] With increasing demands for navigation accuracy, high-precision navigation sensors are widely used in the navigation field. However, due to the complexity of their internal components and the variable operating environment, the signals acquired by high-precision navigation sensors are prone to slowly deviating from the true value over time. When the deviation exceeds the system's acceptable range, the navigation sensor experiences a slow-change fault, resulting in reduced navigation accuracy and impacting navigation performance. Currently, there is considerable research on active fault tolerance for rapid-change faults in navigation sensors, but relatively little research on slow-change faults. One reason is that slow-change faults initially have smaller amplitudes, and their changes over time are much smaller than those of rapid-change faults, making it difficult to diagnose them quickly. Another reason is that slow-change faults occur over a longer period, deviating significantly from the true value, making effective active fault tolerance difficult.

[0003] Research on slowly varying fault diagnosis algorithms for navigation systems typically employs backpropagation (BP) neural networks to extract fault features. For example, in the area of ​​inertial navigation unit (IMU) fault detection, Zhang Shufeng researched and optimized a deep belief network model for IMU fault detection and diagnosis. He proposed a deep learning model centered on a deep belief network, created a fault dataset for wheeled robots' IMUs, and evaluated the fault diagnosis performance of the optimized model. Li Gang et al. proposed a method for fault diagnosis of aircraft inertial navigation gyroscopes, using an improved empirical mode decomposition-permutation entropy algorithm for fault detection. The decomposed navigation signals were reconstructed as a training set, and a probabilistic neural network model with fast learning capabilities and high accuracy was used to establish a gyroscope fault diagnosis network for fault diagnosis. Zhang Zheng et al. used a BP neural network to predict standard sensor signals and construct residuals, then used a threshold method for navigation sensor fault diagnosis and signal reconstruction. While these algorithms can diagnose slowly varying faults, they are time-consuming. Furthermore, the use of BP neural networks can easily lead to overfitting of the training dataset, resulting in decreased diagnostic accuracy in practical applications. At the same time, the types of sensor signals that can be reconstructed using the threshold method are very limited, making it impossible to achieve multi-scenario applications.

[0004] Research on proactive repair of slowly varying fault signals in navigation systems often employs thresholding methods and backpropagation (BP) neural networks (LPNs) in engineering. However, research on using long short-term memory (LSTM) neural networks and convolutional neural networks for this purpose is limited. Thresholding methods, under a given threshold, use fuzzy membership functions to calculate whether the residual exceeds the threshold and diagnose the fault, thus using the residual for fault repair. However, this method has limited applicability, the threshold setting is fixed, and its generalization ability is insufficient. BP neural networks for proactive fault repair of navigation sensors utilize fully connected neural networks as the core of the deep learning model. A training dataset is constructed, and the trained model is used for proactive fault tolerance of navigation sensors. However, due to the relatively poor learning ability of BP neural networks, their strong dependence on the dataset, and the difficulty in constructing fault datasets, the accuracy of fault detection is also affected. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent active fault-tolerance method for gradually changing faults in navigation sensors, which enables timely and accurate diagnosis of gradually changing faults and provides repair signals; it also enables rapid diagnosis of gradually changing faults in navigation sensors and reconstruction of navigation signals with generalization capabilities, solving the problems of slow diagnosis of gradually changing faults and the limited application scenarios of traditional threshold methods.

[0006] The technical solution to achieve the purpose of this invention is as follows:

[0007] A smart active fault-tolerance method for slowly varying faults in navigation sensors, applicable to the diagnosis and repair of slowly varying faults in navigation sensors, includes:

[0008] A standard signal prediction module was built and trained using long short-term memory neurons as the basic unit, and the navigation signal was predicted through the standard signal prediction module.

[0009] A slowly varying fault diagnosis module is built and trained using a multi-layer stacked convolutional neural network model. The slowly varying fault diagnosis module extracts slowly varying fault trend information based on the residual between the predicted signal and the measured signal, and realizes the diagnosis of slowly varying faults in the early stage of the occurrence of the fault and when the impact on the system is within the set range.

[0010] If no gradual fault is diagnosed, the measured signal is used as the navigation signal. When a gradual fault is diagnosed, the predicted signal from the standard signal prediction module is used as the navigation signal to achieve active fault tolerance for gradual faults in the navigation signal.

[0011] Furthermore, the training data set of the standard signal prediction module uses a sliding window method to extract and obtain the constructed time-series information.

[0012] Furthermore, the standard signal prediction module employs an LSTM recurrent neural network, stacking multiple recurrent networks to increase the depth of the recurrent neural network. The input of the l-th layer is the output of the (l-1)-th layer, and a Dropout mechanism is added for propagation between layers.

[0013] Furthermore, each layer is normalized using Max-Min, and the output layer is denormalized.

[0014] Furthermore, the slowly varying fault diagnosis module employs a model that stacks convolutional neural networks and fully connected neural networks to extract fault features and classify faults from the residuals, respectively. This model includes a first CNN layer, a batchnorm layer, a second CNN layer, a ReLU rectifier layer, a dropout layer, a fully connected layer, and a cross-entropy loss function layer connected in sequence. The dropout layer serves as the output layer, and the cross-entropy loss function layer propagates back to the dropout layer to update the weight information of each unit in the model.

[0015] Furthermore, the training dataset of the gradually changing fault diagnosis module is the residual between the predicted signal and the measurement signal. It is obtained by processing the predicted signal dataset and the measurement navigation information dataset containing fault information through the Delta calculation module and residual processing. When processing the predicted signal data, the Delta calculation module uses the predicted Δ value instead of the predicted true navigation information. The Δ value is obtained by subtracting the navigation information of the previous sampling point from the navigation information of the next sampling point. The residual information is sequentially segmented, and the data is segmented vertically according to the window size. The segmented matrix is ​​labeled.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0017] (1) This invention utilizes a long short-term memory neural network (LSTM) that is more sensitive to time-related data to build a standard signal prediction module. The propagation between layers is enhanced by the addition of the Dropout mechanism, which weakens the joint adaptability between neural nodes and enhances the generalization ability, thus obtaining measurement estimates that are not affected by faults. The training data set is divided into batches for training, which improves training efficiency and avoids the overfitting phenomenon of the neural network.

[0018] (2) This invention constructs a residual matrix based on the characteristics of slowly varying faults to extract slowly varying fault features. The navigation sensor measurement information is correlated before and after the time sequence. A standard signal prediction module based on a long short-term memory neural network (LSTM) is used to extract the time sequence features of the navigation sensor signals. The LSTM neural network improves the long-range dependency problem of ordinary recurrent neural networks to the greatest extent. A gating mechanism is introduced to control the accumulation speed of information, including selectively adding new information and selectively forgetting previously accumulated information. At the same time, it solves the gradient vanishing and gradient explosion problems of simple recurrent neural networks. Then, the slowly varying fault features are extracted from the residual information obtained by model prediction and actual acquisition to realize the diagnosis of slowly varying faults. The slowly varying fault diagnosis module based on a convolutional neural network (CNN) is used to classify the residual information. The CNN convolution kernel is shared, which can process two-dimensional residual data information without pressure. At the same time, the convolutional neural network can automatically extract features from the generated navigation residual matrix information. The required features (slowly varying fault features) are extracted by convolution of the convolution kernel on the residual matrix. The deep learning model can better complete the classification work based on the extracted feature information. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the sliding window capture of training data for the standard signal prediction module in this invention.

[0020] Figure 2 This is a schematic diagram of the long short-term memory network neurons in this invention.

[0021] Figure 3 This is a network structure diagram of the standard signal prediction module in this invention.

[0022] Figure 4 This is a flowchart of the training data generation process in this invention.

[0023] Figure 5 This is a flowchart of the fault diagnosis module model generation process in this invention.

[0024] Figure 6 This is a structural diagram of the fault diagnosis module model in this invention.

[0025] Figure 7 This is a schematic diagram of the output layer structure of the fault diagnosis module in this invention.

[0026] Figure 8 This is a flowchart of the active fault-tolerant data flow in this invention.

[0027] Figure 9 This is a simulation diagram of the motion trajectory in this invention.

[0028] Figure 10 This is a diagram showing the attitude angle changes in this invention.

[0029] Figure 11 This is a speed variation diagram in the present invention.

[0030] Figure 12 This is a diagram showing the positional changes in this invention.

[0031] Figure 13 This is a diagram showing the results of the active latitude-tolerance function of the infrared camera in this invention.

[0032] Figure 14 This is a diagram showing the latitude active tolerance error analysis of the infrared camera in this invention.

[0033] Figure 15 This is a diagram showing the results of the active longitude tolerance of the infrared camera in this invention.

[0034] Figure 16 This is a diagram showing the active longitude tolerance error analysis of the infrared camera in this invention.

[0035] Figure 17 This is a diagram showing the results of the active fault tolerance of the infrared camera in this invention.

[0036] Figure 18 This is a diagram showing the active error tolerance analysis of the infrared camera in this invention. Detailed Implementation

[0037] When auxiliary navigation sensors experience a series of minor disturbances, the navigation information can gradually deviate from the true information, with the deviation showing a slow increasing trend. When the difference between the navigation information and the true information becomes too large, it can be considered that the navigation sensor has experienced a slow-varying fault. Neural network systems can extract slow-varying fault characteristics from the deviation between the navigation information and the true navigation information.

[0038] This method addresses the problem of navigation sensors experiencing gradual faults due to complex operating environments, unknown interference, and diverse motion patterns, leading to a decline in information source quality or even unavailability. It focuses on active fault tolerance for gradual faults in navigation systems. To solve the problem of decreased navigation accuracy caused by gradual faults in navigation sensors, this patent proposes an intelligent active fault tolerance algorithm for navigation sensor gradual faults. This algorithm achieves timely and accurate diagnosis of gradual faults and provides repair signals. It enables rapid diagnosis of gradual faults in navigation sensors and reconstruction of navigation signals with generalization capabilities.

[0039] To address the challenge of detecting and preventing slowly changing faults in navigation sensors in their early stages, a slowly changing fault identification algorithm based on a deep learning model is proposed. This algorithm extracts and identifies fault features in the early stages of a slowly changing fault in the navigation system. Simultaneously, by extracting features from historical navigation information, it predicts the standard navigation signal for a future period. This allows for correction of navigation information in the early stages of a slowly changing fault, achieving proactive fault tolerance for sensor slowly changing faults.

[0040] A standard signal prediction module is built using a Long Short-Term Memory (LSTM) neural network, which is more sensitive to time-related data. This module predicts the standard navigation information for the current time window based on historical navigation information and constructs the residuals. A slowly varying fault diagnosis module is built using a Convolutional Neural Network (CNN), which can quickly extract matrix feature information, to classify the residual information. When the fault diagnosis module detects a slowly varying fault, it uses the predicted signal from the standard information prediction module to repair the fault signal.

[0041] This algorithm can diagnose slowly occurring faults when the resulting navigation deviations are within the system's tolerance range, thus solving the problem of slow fault diagnosis. Furthermore, it replaces the faulty navigation signal with a standard navigation signal predicted by a neural network model with strong generalization capabilities, overcoming the limitation of traditional threshold methods in terms of application scenarios.

[0042] The aforementioned intelligent active fault-tolerance method for gradually changing faults in navigation sensors specifically includes:

[0043] 1. Standard navigation information prediction

[0044] To extract slowly changing fault features from the residual between the theoretical true value and the measured signal, a standard signal prediction module is constructed to predict standard navigation signals that are close to the theoretical true value. At the same time, when a slowly changing fault occurs, the output of this module is used to replace the fault signal.

[0045] The navigation data acquired by the sensors is fed into the standard navigation information prediction module. This module, based on the LSTM neural network algorithm, constructs a deep learning model to predict standard (fault-free) navigation information for the current window time based on historical navigation information.

[0046] The standard signal prediction module can predict the non-standard navigation information (fault-free navigation information) for the current time window based on historical navigation information. This yields measurement estimates unaffected by faults. Measurement estimate The residual between the data and the actual navigation data Z(t) obtained from the sensor sampling information contains slowly varying fault information.

[0047] Z(t)=M(t)+E(t)+f t (t)

[0048]

[0049] The navigation data Z(t) obtained from the sampling and calculation by the sensors on the device contains the device's motion state information M(t), the sensor error information E(t) on the device, and the trend information f of the gradually changing fault. t (t).

[0050] Navigation data information obtained from the standard signal prediction module Includes predicted equipment motion state information And the error information of the sensors on the predicted device. The error between the fitting results of the deep learning model and the actual measurement data is negligible. Z(t) and The difference between the two values ​​is the extracted measurement residual ΔZ(t). The information contained in ΔZ(t) is the trend information of the slowly changing fault.

[0051] During the dataset construction phase, the training dataset for the standard signal prediction module uses a sliding window approach, such as... Figure 1 The constructed temporal information is extracted. In deep neural networks, the input to an intermediate layer is the output of the layer above. Therefore, changes in the parameters of previous layers can lead to significant differences in the distribution of the input to the current layer. The solution is to perform a normalization operation on each layer to stabilize its distribution. This model uses the Max-Min normalization method. Inverse normalization is performed at the output stage to complete the prediction of fault-free measurement signals.

[0052] During the model building phase, the standard signal prediction module, in order to process time-series data and utilize its historical information, needs to enable the deep learning network to have short-term memory capabilities. Therefore, an LSTM recurrent neural network is adopted. LSTM neural networks significantly improve upon the long-range dependency problem of ordinary recurrent neural networks by introducing a gating mechanism to control the rate of information accumulation, including selectively adding new information and selectively forgetting previously accumulated information. Its neuron structure is as follows: Figure 2 It includes forget gates, output gates, new memory units, and input gates, with the following processing schemes for empty units in each gate:

[0053] Forget gate f: The forget gate controls the internal state c of the previous time step. t-1 How much information needs to be forgotten?

[0054]

[0055] Where σ is the sigmoid activation function, and the variables computed by the integrated forgetting gate are mapped between (0, 1), x t The input sequence for the forget gate is given at time t. h represents the weight of the neuron corresponding to the forgetting gate. t-1 This is a transfer from the previous moment's hidden state via the Forgotten Gate. b represents the hidden state weight corresponding to the forget gate unit. f The bias coefficient corresponding to the forget gate;

[0056] Output gate o: controls the candidate state information c at the current moment. t How much information needs to be output to the external state?

[0057]

[0058]

[0059] Where σ is the sigmoid activation function, the variables computed by the integrated output gate are mapped between (0, 1), and x t The input sequence to the output gate at time t is the input of the output gate input sequence. h represents the weights of the neurons corresponding to the output gate. t-1 This is passed in from the previous moment when the output gate was hidden. b represents the hidden state weight corresponding to the output gate unit. o The bias coefficient corresponding to the output gate is c. The hidden state of the output gate needs to be weighted with the memory state. t The current state of the output gate is represented by tanh, where tanh is the hyperbolic tangent activation function.

[0060] New memory unit g:

[0061]

[0062] Where tanh is the hyperbolic tangent activation function, x t The input sequence for the new memory cell at time t. h represents the weight of the neuron corresponding to the new memory unit. t-1 The hidden state of the new memory unit from the previous moment is passed in. b represents the weight of the hidden state corresponding to the new memory unit. g The bias coefficient corresponding to the new memory unit;

[0063] Input gate I: controls the candidate state at the current time step. How much information needs to be saved?

[0064]

[0065]

[0066] Where σ is the sigmoid activation function, x t The input sequence of the input gate unit at time t is the input of the input gate unit. h represents the weights of the neurons corresponding to the input gate units. t-1 This is the input gate cell's hidden state passed in at the previous moment. b represents the hidden state weights corresponding to the input gate unit. i The bias coefficients corresponding to the input gate units;

[0067] The approach used in this module to increase the depth of the recurrent neural network is to stack multiple recurrent networks together, such as... Figure 3 The stacked recurrent neural network structure of this module is given, where the input of the l-th layer is the output of the (l-1)-th layer.

[0068] Feature extraction of temporal information is achieved through layer stacking, with the input of layer l being the output of layer (l-1). To enhance the model's generalization ability, a Dropout mechanism is added for inter-layer propagation. During training, this mechanism randomly ignores some neurons. It forces a neuron to work with randomly selected other neurons, achieving better results, reducing the joint adaptability between neural nodes, and enhancing generalization ability.

[0069] The training process uses batch training, with a defined batch size: Batch_size (this concept only appears during the learning process). The learning process calculates the loss function value for the training data and aims to find parameters that minimize the loss function value. Instead of calculating the loss function for individual data points during training, the training data set is divided into groups, which are called "batches".

[0070] Suppose there are N independent data points in a batch. The batch size is N. If we use cross-entropy as the loss function, the result would be:

[0071]

[0072] Where t nk y represents the value of the k-th element at the n-th sampling point. nk It is the output of the neural network, t nk This is supervised data. Finally, it is averaged by dividing by N to obtain the average loss function for each data point.

[0073] Neural networks learn by selecting a batch of data (called a mini-batch) from the training data and then learning from each mini-batch. While not strictly equivalent to the entire dataset, it is an approximation of the whole. This improves training efficiency while avoiding overfitting in neural networks.

[0074] 2. Gradual Fault Identification

[0075] After obtaining the standard navigation signal, it is necessary to compare the residual between the standard navigation signal and the measurement data to diagnose whether the current sensor has experienced a gradual change fault. Therefore, a gradual change fault diagnosis module is constructed to extract features from the residual information and diagnose whether a gradual change fault has occurred.

[0076] The gradual fault diagnosis module extracts gradual fault trend information by measuring residuals, performs data classification tasks, and completes the prediction of gradual faults.

[0077] Slowly changing faults initially cause small deviations in navigation information, but gradually deviate from the true information, with the deviation amplitude showing a slow increasing trend. A neural network is used to extract this feature; it is used to identify slowly changing faults in their early stages; and the fault is identified while the deviation amplitude is still within the system's allowable range.

[0078] The sampling length of the residual matrix input to the model is less than the system's tolerable time for a slowly changing fault. This allows for the detection (prediction) of slowly changing faults when the deviation in navigation information caused by the fault is less than the system's tolerable deviation. In practical applications, the model needs to process and analyze the sensor sampling data and predict whether a slowly changing fault will occur to an intolerable level in the future. Therefore, during model training, the sampling window of the residual matrix needs to be compressed to ensure that the sampling window time and the model prediction time are always less than the system's tolerable time for a slowly changing fault.

[0079] A model combining convolutional neural networks and fully connected neural networks is used to extract fault features and classify faults from the residual matrix.

[0080] Design a deep learning model to identify gradually changing faults before they occur, which cause navigation information to deviate too much from the real information and become unusable.

[0081] The training and input data for this module are both navigation information residual matrices. These residual matrices represent the difference between the navigation information predicted by the standard signal prediction module and the real-time navigation measurements collected by the sensors. However, they require some processing before they can be used as training data for the gradually changing fault diagnosis module and as input information in actual use. The data processing flow is as follows: Figure 4 , 5 The data consists of two parts: training data and test data. Both parts are derived from a standard navigation information dataset generated from the same path information and a navigation information dataset containing fault information, after Delta processing and residual processing.

[0082] The navigation data (fault-free data obtained after processing by the standard signal prediction module) is a navigation information dataset that does not contain faults. The vertical axis represents the time series; the horizontal axis represents the navigation information. The navigation data (including faults) is a dataset containing fault information generated by adding abrupt and gradual faults to the navigation information generated along the same path. Sampling points are labeled to indicate whether a fault has been added and what type of fault it is. This dataset has the same vertical information as the standard navigation information dataset. The horizontal information adds a one-dimensional fault label to indicate whether a fault has occurred and its type. In the construction and training of the standard signal prediction module, the Delta calculation module uses predicted Δ values ​​instead of predicted true navigation information to improve model accuracy and reduce the difficulty of normalization. This method is still used here to reduce the complexity of model data normalization. The navigation information of the previous sampling point is subtracted from the navigation information of the next sampling point to obtain the Δ value. The residual navigation information is split (sequential splitting, not random splitting) to obtain two parts: test information and training information. Both parts are labeled. Configure the window for the fault identification module and segment the data vertically according to the window size. Label the segmented matrix.

[0083] The model building in this module adopts a design pattern of stacking convolutional layers and fully connected layers, such as... Figure 6 The residual matrix information obtained from the standard signal prediction module and sensor measurements serves as the input to the slowly varying fault diagnosis module. A more defined feature matrix is ​​initially extracted from the slowly varying fault features through CNN layer 1. This information is then integrated using a batch normalization layer and fed into CNN layer 2 for further slowly varying fault feature extraction. Convolutional operations in convolutional layers are used to optimize redundant parameters through weight sharing, reducing the possibility of model overfitting. A dropout layer is added to improve the model's generalization ability. A fully connected layer is used for integration operations to construct a suitable output and complete the classification task. The output layer design employs... Figure 7 The structure shown, after filtering through the softmax layer model output, in actual use, integrates and outputs the label of the data with the highest probability based on the probabilities of the softmax layer. During model training, one-hot labels are compared, and backpropagation is performed after calculating cross-entropy to update the weight information of each unit of the model.

[0084] 3. Active fault tolerance for gradually changing faults

[0085] After predicting and fitting the navigation information using a standard signal prediction module, residuals are calculated between the predicted and fitted navigation information. This allows for the detection of slowly changing faults before they occur and affect the overall system stability. The fault signal is then repaired using the fitting results from the labeled standard signal prediction module. The data flow for the active fault-tolerant navigation information scheme is as follows: Figure 8 .

[0086] Meanwhile, the learning of the fault diagnosis module needs to be performed under interference generated by the standard signal prediction module, which better reflects the model's performance in real-world applications. However, this increases the difficulty of feature extraction and classification in the fault identification module. The residual matrix generated during training uses navigation information (including prediction error) Z(t). LSTM And navigation information (including sudden and gradual faults) Z(t) Err The difference:

[0087] Z(t) LSTM =M(t)+E(t)+LSTM Err (t)

[0088] Z(t) Err =M(t)+E(t)+f t (t)

[0089] Where E(t) is the noise information of the sensor, f t (t) represents the added fault information, LSTM Err (t) represents the error information generated by the neural network after passing through the standard signal prediction module, f t (t) represents the added fault information.

[0090] This invention addresses the difficulty of actively repairing slowly changing faults by employing a residual method to extract slowly changing fault features from navigation sensors and using a deep learning model to achieve active fault tolerance. First, a deep learning model based on a Long Short-Term Memory (LSTM) neural network predicts the standard navigation information for the current time window based on historical navigation information, thus constructing residuals. Next, a deep learning model based on a Convolutional Neural Network (CNN) is used to classify the residual information. When the fault diagnosis module detects a slowly changing fault, the predicted signal from the standard signal prediction module is used to repair the fault signal, achieving active fault tolerance for slowly changing faults.

[0091] In the gradual fault diagnosis stage, a residual matrix is ​​first constructed based on the characteristics of the gradual fault to extract its features. Navigation sensor measurement information is correlated before and after a time series. A standard signal prediction module based on a Long Short-Term Memory (LSTM) neural network is used to extract the temporal features of the navigation sensor signals. The LSTM neural network significantly improves upon the long-range dependency problem of ordinary recurrent neural networks by introducing a gating mechanism to control the information accumulation rate, including selectively adding new information and selectively forgetting previously accumulated information. It also solves the gradient vanishing and gradient exploding problems of simple recurrent neural networks. Then, the gradual fault features are extracted from the residual information obtained by the model prediction and the actual acquisition results to achieve gradual fault diagnosis. A gradual fault diagnosis module based on a Convolutional Neural Network (CNN) is used to classify the residual information. CNN convolutional kernels are shared, allowing for effortless processing of two-dimensional residual data. Simultaneously, the CNN can automatically extract features from the generated navigation residual matrix information. By convolving the residual matrix with convolutional kernels, the required features (gradual fault features) are extracted. This deep learning model can better complete the classification work based on the extracted feature information.

[0092] During the gradual fault repair phase, the diagnostic results provided by the gradual fault diagnosis module are used to determine whether repair should proceed. If no gradual fault is diagnosed, the signal acquired by the navigation sensor is used as the signal output. If a gradual fault is diagnosed, the result of the standard signal prediction module is used as the output of the navigation sensor. This achieves proactive fault tolerance for gradual faults in the navigation sensor.

[0093] Example

[0094] In this embodiment, the navigation system is designed with synchronized sensor sampling frequencies, and the sensors sample external information. Gradually changing faults occur randomly, and the possible motion states of the moving body carrying the sensor are shown in the table below.

[0095] Table 1: Motion Status

[0096]

[0097]

[0098] Step 1: Building and training the standard signal prediction module

[0099] For model training, a path training dataset containing various motion states was constructed. Paths were constructed based on the six motion states listed in the table above. During path construction, pseudo-randomization was used to maximize the diversity of states; to maximize the diversity of state transitions, a full permutation of motion states was performed while ensuring the feasibility of the device's motion states (deceleration cannot be performed at the initial moment, uniform deceleration cannot continue, causing the device's speed to become negative, etc.).

[0100] The simulated navigation data consists of 9 dimensions: pitch angle, roll angle, yaw angle, eastward speed, northward speed, celestial speed, latitude, longitude, and altitude.

[0101] The method for constructing training data (X, Tag) for the standard signal prediction module is to slide the data from the starting position to the ending position with a single step size. For each window, the sequence information of the data within the window serves as the model input X during the training process, and the navigation data at the next time step of the window serves as the supervision label.

[0102] Simultaneously, a normalization operation is performed on each layer of the neural network to stabilize its distribution. The model's normalization operation uses the Max-Min normalization method, distributing the values ​​of the training and test data within the [0,1] interval. An inverse normalization operation is performed at the output stage to complete the prediction of fault-free measurement signals. The formula is as follows:

[0103]

[0104] Where x represents the original data, x BN This is the normalized data.

[0105] The predicted ΔMonentState is used instead of MomentState. Taking the northbound velocity as an example, the prediction of ΔV is used instead of the prediction of V. The navigation information resulting from the prediction is as follows:

[0106] Construct a dataset for ΔV, reduce the dataset, and use ΔV of the sampling points within the window as the input of the network model, and use the sampling points of the next time step of the window as the data tags.

[0107] Define the hidden state of each layer as h, and the shape of the input vector as X. T = [x1, x2... x t ] T , where x tThe information is for a single sampling point, with a shape of [1×1]. Because the output layer is a fully connected layer, the features of the 9-dimensional input data will be blurred during the operations of the fully connected layer. Therefore, the model is trained separately for each dimension. This also solves the problem of inconsistent arrival times of the 9-dimensional data during actual testing. Input data X T :[Batch_size × t_size × 1]. Initial hidden state h0:[num_layer × Batch_size × h], initially set to all zeros. Initial memory unit c0:[num_layer × Batch_size × h], initially set to all zeros. Output information output LSTM :[Batch_size×h].

[0108] Temporal information feature extraction is achieved by stacking num_layers, where the input of layer l is the output of layer l-1. To enhance the model's generalization ability, a dropout mechanism is added for propagation between layers, reducing the joint adaptability between neurons and improving generalization. The input of the fully connected layer is the output of the LSTM layer, i.e., the output... LSTM Fully connected layer parameter dimension reduction: linear_dim = [Batch_size × h × 1]. Fully connected layer output dimension reduction: output linear = [Batch_size × 1]. The training process adopts a batch training method, using mean squared error as the loss function.

[0109] The initialization method uses the Xavier initialization method to initialize the weights and biases of the fully connected layer, ensuring that the weights follow a normal distribution.

[0110] W~U[-a,a]

[0111]

[0112]

[0113] Where, n i and n i+1 These represent the number of input neurons and the number of output neurons, respectively.

[0114] Step 2: Building and Training the Gradually Changing Fault Diagnosis Module

[0115] The training data label processing method is as follows:

[0116] Label 0: No fault has been added to this sampling point;

[0117] Label 1: A sudden fault has been added to this sampling point;

[0118] Label 2: A slowly varying fault has been added to this sampling point;

[0119] The Delta calculation module subtracts the navigation information of the previous sampling point from the navigation information of the subsequent sampling point to obtain the Δ value. The normalization process uses a maximum-minimum normalization method, applying different normalization parameters to each of the nine columns of data to normalize the values ​​to the range (0, 1). Normalization is not performed within this module.

[0120] The navigation data at the current time (the t-th sampling point) is Z(t). The navigation data at the next time (the (t+1)-th sampling point) is Z(t+1). The processing result of the Delta module on the navigation data at the current time is: ΔZ(t) = Z(t+1) - Z(t).

[0121] The Delta processing module's result for processing navigation data at the initial time is:

[0122] ΔZ(t0)=0

[0123] Z(t)=M(t)+E(t)+f t (t)

[0124]

[0125] The navigation data Z(t) obtained from the sampling and calculation by the sensors on the device contains information including the device's motion state information M(t), the sensor error information E(t) on the device, and the trend information f of the slowly changing fault. t (t).

[0126] Navigation data information obtained from the standard signal prediction module Includes predicted equipment motion state information And the error information of the sensors on the predicted device. The error between the fitting results of the deep learning model and the actual measurement data is negligible. Z(t) and The difference between the two datasets is the extracted measurement residual ΔZ(t). ΔZ(t) contains the trend information of the slowly varying fault. The residual calculation unit calculates the residual containing the slowly varying fault information using the two datasets mentioned above. This residual is then provided to the slowly varying fault diagnosis module for fault feature extraction and classification.

[0127] The matrix segmented from the residual navigation information is labeled.

[0128] The annotation method is as follows:

[0129] Label according to the tenth column (sampling point fault label) in the window:

[0130] All labels are zero: [000...000] No errors in the information within the window, label type 1;

[0131] Tag 2 appears: The information in the window shows a slow change fault, labeled as category 2 (this category is also the key category to be extracted);

[0132] All labels are 2: [222...222] A slow-change fault is occurring in the window, labeled as type 3;

[0133] Tag 1 appears: A sudden fault is occurring in the window, labeled as type 4;

[0134] All labels are 1: [111...111] A sudden fault is occurring in the window, labeled as type 5;

[0135] Other: Label type 6;

[0136] A stacked design pattern of convolutional and fully connected layers is adopted. The convolutional operations of the convolutional layers are used to share weights and optimize redundant parameters, reducing the possibility of model overfitting. The integration operation of the fully connected layers is used to construct a suitable output to complete the classification task.

[0137] A slowly changing fault diagnosis module was built, using a CNN convolutional neural network as the core part of the model.

[0138] The first layer of the model is a CNN layer with 1 input channel: a channel refers to an image. The data being processed is a residual matrix, and the model processes one matrix at a time, so the number of channels is set to 1.

[0139] Output dimension h: Similar to the hidden layer, it is also the size of the output dimension of this CNN layer.

[0140] Kernel settings: For nine-dimensional navigation data, because the three dimensions of the navigation data have mutually influential features, and to minimize the kernel size as much as possible to reduce the number of parameters, improve training efficiency, and prevent overfitting, a small (3*3) convolutional kernel is used.

[0141] Kernel stride: (3*1) The convolution kernel slides with a stride of 3 horizontally and a stride of 1 vertically. Horizontally, it represents navigation information; the kernel extracts features related to 3D navigation information using a stride of 3. Vertically, it represents sampling points; since the features between sampling points are correlated, a stride of 1 is used.

[0142] Padding: The padding strategy uses 1 padding. 1 padding is chosen to avoid ignoring edge data.

[0143] The data undergoes the first layer of shape transformation: [batch_size,1,9,200]→[batch_size,h,3,200]

[0144] The second layer of the model is the BatchNorm layer, which is used to solve the covariance shift problem in the intermediate layers of multi-layer neural networks, and this layer only plays a role during training.

[0145] The operational form of this layer:

[0146] For the input: Input = {x1, x2, x3... x Batch_Size}

[0147] Output: Output = {y i =BN γ ,β(x i )}

[0148]

[0149]

[0150]

[0151]

[0152] A batch of data consists of Batch_Size samples, and the parameters to be learned are γ and β. In deep networks containing convolutional operations (and multi-channel networks), for each channel (each feature map), the mean and variance of the data in the specified channel for this Batch_Size samples are calculated. The third equation is the normalization operation, where ε is a very small number to prevent numerical problems in the calculation (the case where the denominator is 0). The last equation is an operation to scale and shift the output. The last step is also called affine, which ensures that the output after Batch Normalization can at least return to the state of the input (i.e., when γ and β are 1 and 0 respectively), so that the Batch Normalization operation does not degrade the model's performance.

[0153] The third layer of the model is a CNN layer, with the number of input channels h: the output state set in the previous layer is h, and the number of input channels in this layer is set to h.

[0154] Output channel set to 2h: This is to increase the number of neurons within a reasonable range to ensure learning accuracy.

[0155] The fourth layer of the model is a ReLU activation function layer, with the function expression f(x) = max(0,x). This layer is added to increase the non-linearity of the neural network model; without activation functions, each layer would be equivalent to matrix multiplication. The output of each layer is a linear function of the input of the previous layer. Regardless of the number of layers in the neural network, the output is always a linear combination of the inputs.

[0156] The fifth layer of the model is a fully connected layer. Its function is to integrate the output by linearly integrating the neurons that processed the previous inputs. For the target category (6 categories) of the problem to be solved (classification problem), the input from the previous layer undergoes an affine transformation, making the output dimension 6. This corresponds to the probabilities of the 6 categories obtained from network analysis.

[0157] The loss function layer of the model is the fifth layer. This layer's function is to calculate probabilities from the six outputs of the fully connected layer, convert the labels to one-hot encoding, calculate cross-entropy with the neural network outputs, and backpropagate based on the cross-entropy value. The formula for cross-entropy is:

[0158] Loss(P,Q)=-∑(P(x)log Q(x))

[0159] Where x is the input, P is the estimated probability value of each category output by the Softmax module, and Q is the value of the corresponding training data label.

[0160] This layer integrates one-hot encoding of labeled data; softmax calculation of model-generated data; and cross-entropy calculation of model-generated data and labeled one-hot data (to verify the distribution similarity between model-generated data and labeled data). The result of the loss function is backpropagated to the model, modifying the model parameters until the loss function converges to a state where the model can solve the classification problem. The Softmax module, for a given test input x, estimates the probability value P(y = j|x) for each class using a hypothesis function. That is, it estimates the probability of each classification result. Therefore, the hypothesis function outputs a k-dimensional vector (with the sum of its elements being 1) to represent this estimated probability value. The function form is as follows:

[0161]

[0162] Where θ represents the model parameters, using The probability distribution is normalized so that the sum of all probabilities is 1, and T is the parameter identifier for the current time period.

[0163] The model achieved an accuracy of 85% on the test set (out of 20 test datasets, 17 of which demonstrated active fault tolerance).

[0164] By analyzing the data collected from the sports car's movement, the trajectory and parameter changes are as follows: Figure 9 ,in Figure 10 The changes in the pitch, roll, and yaw angles of the sports car. Figure 11 This refers to the changes in eastward velocity, northward velocity, celestial velocity, and absolute velocity. Figure 12 This is a trend chart showing the latitude, longitude, and altitude data of a sports car. The above data is the navigation data information of the sports car, and the feasibility of active fault tolerance of the infrared camera sensor is simulated using this data. Figure 13 , 15 Figure 17 shows the active fault-tolerant results of the latitude, longitude, and altitude information from the infrared camera. It can be seen that after 25 seconds, the gradual fault began to appear and was detected by the active fault-tolerant system. The sensor measurement error increased, and the active fault-tolerant system began to process the faults in the data (longitude, latitude, and altitude) collected by the navigation sensor. Figure 14 , 16 The height output error of the infrared camera and the height prediction error of the active fault-tolerant algorithm are compared. It can be seen that the measurement error is significantly lower than the prediction error and the trend is basically consistent. The active fault-tolerant algorithm can effectively predict the height information of the lidar.

Claims

1. A method for intelligent active fault tolerance of slowly varying faults in navigation sensors, characterized in that, include: A standard signal prediction module was built and trained using long short-term memory neurons as the basic unit, and the navigation signal was predicted through the standard signal prediction module. A slowly varying fault diagnosis module is built and trained using a multi-layer stacked convolutional neural network model. The slowly varying fault diagnosis module extracts slowly varying fault trend information based on the residual between the predicted signal and the measured signal, and realizes the diagnosis of slowly varying faults in the early stage of the occurrence of the fault and when the impact on the system is within the set range. If no gradual fault is diagnosed, the measurement signal is used as the navigation signal. When a gradual fault is diagnosed, the prediction signal of the standard signal prediction module is used as the navigation signal to achieve active fault tolerance for gradual faults in the navigation signal. The slowly varying fault diagnosis module uses a model that stacks convolutional neural networks and fully connected neural networks to extract fault features and classify faults from the residuals, respectively. It includes a first CNN layer, a batchnorm layer, a second CNN layer, a ReLU rectifier layer, a dropout layer, a fully connected layer, and a cross-entropy loss function layer connected in sequence. The dropout layer is used as the output layer, and the cross-entropy loss function layer is backpropagated to the dropout layer to update the weight information of each unit of the model. The dropout layer for a given test input The hypothesis function estimates the probability value for each category; the hypothesis function will output a... The estimated probability value P is represented by a dimensional vector, where k is the number of output nodes, i.e., the number of categories. The function form is: in To represent model parameters, use The probability distribution is normalized so that the sum of all probabilities is 1, and T is the parameter identifier for the current time period.

2. The intelligent active fault-tolerance method for gradually changing faults in navigation sensors according to claim 1, characterized in that, The training data set of the standard signal prediction module is obtained by using a sliding window method to extract the constructed time series information.

3. The intelligent active fault-tolerance method for gradually changing faults in navigation sensors according to claim 2, characterized in that, The standard signal prediction module employs an LSTM recurrent neural network, stacking multiple recurrent networks to increase the depth of the recurrent neural network. The input of the layer is the first Add a Dropout mechanism to the output of layers and the propagation between layers.

4. The intelligent active fault-tolerance method for gradually changing faults in navigation sensors according to claim 3, characterized in that, The neuron includes a forgetting gate, an output gate, a new memory unit, and an input gate, wherein: Forgotten Gate Controlling the internal state of the previous time step The amount of information that needs to be forgotten is expressed as: in The variable mapping, calculated using the sigmoid activation function and the forgetting gate, is between (0, 1). Input sequence for forget gate Input at any time The weights of the neurons corresponding to the forget gate. This is a transfer from the previous moment's hidden state via the Forgotten Gate. The hidden state weights corresponding to the forget gate units. The bias coefficient corresponding to the forget gate; Output gate Control the candidate state information at the current moment The amount of information that needs to be output to the external state is represented as: in, For the input sequence of the output gate in Input at any time The weights of the neurons corresponding to the output gate. The hidden state weights corresponding to the output gate units. This represents the bias coefficient corresponding to the output gate. The hidden state of the output gate from the previous moment is passed in. The hidden state of the output gate needs to be weighted with the remembered state. The current state of the output gate is represented by tanh, where tanh is the hyperbolic tangent activation function. The new memory unit g is: Where tanh is the hyperbolic tangent activation function. The input sequence for the new memory unit in Input at any time The weights of the neurons corresponding to the new memory units. The hidden state of the new memory unit from the previous moment is passed in. The hidden state weights corresponding to the new memory units. The bias coefficient corresponding to the new memory unit; Input gate I controls the candidate state at the current time step. The amount of information that needs to be saved is: in The input sequence of the input gate unit is in Input at any time These are the weights of the neurons corresponding to the input gate units. This is the input gate cell's hidden state passed in at the previous moment. The hidden state weights are the values ​​corresponding to the input gate units. This represents the bias coefficient corresponding to the input gate unit.

5. The intelligent active fault-tolerance method for gradually changing faults in navigation sensors according to claim 4, characterized in that, The standard signal prediction module is trained using batch training, with cross-entropy as the loss function. The loss function is: in For batch number, Indicates the first The sampling point of the first sampling point Each element value It is the output of the neural network. It is supervisory data.

6. The intelligent active fault-tolerance method for gradually changing faults in navigation sensors according to claim 1, characterized in that, The cross-entropy loss function is: in, As input, P is the probability value estimated for each category from the output of the Softmax module, and Q is the value of the corresponding training data label.

7. The intelligent active fault-tolerance method for gradually changing faults in navigation sensors according to claim 6, characterized in that, The dataset trained by the gradually changing fault diagnosis module is the residual between the predicted signal and the measurement signal. It is obtained by processing the predicted signal dataset and the measurement navigation information dataset containing fault information through the Delta calculation module and residual processing. When processing the predicted signal data, the Delta calculation module uses the predicted Δ value instead of the predicted true navigation information. The Δ value is obtained by subtracting the navigation information of the previous sampling point from the navigation information of the next sampling point. The residual information is sequentially segmented, and the data is segmented vertically according to the window size. The segmented matrix is ​​labeled.

Images