CN-116242393-B - Intelligent active fault tolerance method for slow-changing fault of navigation sensor

CN116242393BCN 116242393 BCN116242393 BCN 116242393BCN-116242393-B

Abstract

The invention discloses an intelligent active fault-tolerant method for a slow-changing fault of a navigation sensor, which comprises the steps of constructing a standard signal prediction module by taking a long-period memory neuron as a basic unit, training the standard signal prediction module, predicting a navigation signal, constructing a slow-changing fault diagnosis module by adopting a multi-layer stacked convolutional neural network model, training the slow-changing fault diagnosis module, extracting slow-changing fault trend information based on residual errors between the prediction signal and a measurement signal, realizing the diagnosis of the slow-changing fault in a set range at the early stage of the occurrence of the slow-changing fault and the influence on a system, and adopting the measurement signal as the navigation signal if the slow-changing fault does not occur, and adopting the prediction signal of the standard signal prediction module as the navigation signal when the occurrence of the slow-changing fault is diagnosed, so as to realize the active fault-tolerant of the slow-changing fault of the navigation signal. The method can effectively predict the laser radar height information, and solves the problems of slow-change fault diagnosis and single application scene of the traditional threshold method.

Inventors

CHEN WEI
You Runyi
SUN RUISHENG
Qian Menghao
CHEN ZIJIE

Assignees

南京理工大学

Dates

Publication Date: 20260505
Application Date: 20221230

Claims (7)

1. The intelligent active fault-tolerant method for the slow-change fault of the navigation sensor is characterized by comprising the following steps of: building a standard signal prediction module by taking a long-period memory neuron as a basic unit, training, and predicting a navigation signal through the standard signal prediction module; Constructing and training a slow-change fault diagnosis module by adopting a multi-layer stacked convolutional neural network model, extracting slow-change fault trend information by the slow-change fault diagnosis module based on residual errors between a predicted signal and a measured signal, and realizing the diagnosis of the slow-change fault in a set range in the early stage of occurrence of the slow-change fault and on the influence of the system; if the slow-change fault is diagnosed, the measurement signal is used as a navigation signal, and when the slow-change fault is diagnosed, the prediction signal of the standard signal prediction module is used as the navigation signal, so that the active fault tolerance of the slow-change fault of the navigation signal is realized; The slow-change fault diagnosis module adopts a model stacked by a convolutional neural network and a fully-connected neural network to extract and classify fault characteristics of residual errors, and comprises a first cnn layer, a batchnorm layer, a second cnn layer, a Relu rectifying layer, a dropout layer, a fully-connected layer and a cross entropy loss function layer which are sequentially connected, wherein the dropout layer is used as an output layer, the cross entropy loss function layer is reversely propagated to the dropout layer, and weight information of each unit of the model is updated; the dropout layer is for a given test input Estimating probability values for each category using a hypothesis function, which is to output one The vector of dimensions represents this estimated probability value P, k as the number of output nodes, i.e. the number of classes, in the form of a function: Wherein the method comprises the steps of Representing model parameters by The probability distribution is normalized so that the sum of all probabilities is 1 and T is the parameter identification of the current time period.
2. The intelligent active fault-tolerant method for the navigation sensor slow-changing fault as claimed in claim 1, wherein the training data set of the standard signal prediction module adopts a sliding window method to intercept and acquire the constructed time sequence information.
3. The intelligent active fault-tolerant method for navigation sensor ramp fault as claimed in claim 2, wherein said standard signal prediction module employs LSTM cyclic neural network, stacks multiple cyclic networks to increase cyclic neural network depth, the first The input of the layer is the first The output of the layer, the propagation between the layer and the layer add Dropout mechanism.
4. A navigation sensor ramp fault intelligent active fault tolerance method according to claim 3, wherein said neurons comprise forgetting gates, output gates, new memory units and input gates, wherein: Forgetting door Controlling the internal state of the last moment How much information needs to be forgotten, expressed as: Wherein the method comprises the steps of For sigmoid activation functions, the variable map integrating the forgetting gate computation is between (0, 1), Input sequence for forgetting door The input of the moment of time is made, To forget the weights of the corresponding neurons of the gate, The hidden state of the door is forgotten for the last moment, For forgetting the hidden state weight corresponding to the gate unit, The bias coefficient corresponding to the forgetting door; output door Controlling current time candidate state information How much information needs to be output to the external state, expressed as: Wherein, the Inputting sequences for output gates The input of the moment of time is made, To output the weights of the corresponding neurons, To output the hidden state weight corresponding to the gate unit, To output the corresponding bias factor of the gate, For the input of the hidden state of the output gate at the previous time, the hidden state of the output gate needs to be weighted with the memory state, Outputting a memory state of the door at the current moment, wherein tan is a hyperbolic tangent activation function; The new memory cell g is: where tan h is the hyperbolic tangent activation function, Input sequence for new memory cell The input of the moment of time is made, The weights of the corresponding neurons for the new memory cells, The hidden state of the new memory cell is transmitted to the previous time, Is the hidden state weight corresponding to the new memory cell, The bias coefficient corresponding to the new memory cell; Input gate I controls the candidate state at the current time How much information needs to be saved is: Wherein the method comprises the steps of The input sequence for the input gate unit is as follows The input of the moment of time is made, To input the weights of the corresponding neurons of the gate units, The hidden state of the input door unit is transmitted for the last moment, For inputting the hidden state weight corresponding to the gate unit, Is the bias coefficient corresponding to the input gate unit.
5. The intelligent active fault tolerance method for the navigation sensor slow-changing fault of claim 4, wherein the standard signal prediction module training process adopts batch training, adopts cross entropy as a loss function, and the loss function is as follows: Wherein the method comprises the steps of In order to obtain the number of batches, the number of batches is, Represent the first The first sampling point The value of the individual element(s), Is the output of the neural network and, Is the supervision data.
6. The navigation sensor slow-changing fault-tolerant intelligent active fault-tolerant method according to claim 1, wherein the cross entropy loss function is: Wherein, the For input, P estimates a probability value for each category output by the Softmax module, and Q is the value of the corresponding training data tag.
7. The intelligent active fault-tolerant method for the slow-changing fault of the navigation sensor according to claim 6, wherein the data set trained by the slow-changing fault diagnosis module is a residual error between a predicted signal and a measured signal, the predicted signal data set and the measured navigation information data set containing fault information are obtained through Delta calculation module processing and residual error processing, the Delta calculation module adopts a predicted Delta value to replace predicted real navigation information when the predicted signal data is processed, the Delta value is obtained by subtracting the navigation information of the previous sampling point from the navigation information of the next sampling point, the residual error information is sequentially segmented, the data is longitudinally segmented according to the window size, and the segmented matrix is marked.

Description

Intelligent active fault tolerance method for slow-changing fault of navigation sensor Technical Field The invention belongs to the field of navigation, and particularly relates to an intelligent active fault-tolerant method for a slow-changing fault of a navigation sensor. Background Along with the improvement of the navigation precision requirement, the high-precision navigation sensor is widely applied to the navigation field. However, the high-precision navigation sensor is easy to slowly deviate from a true value along with time change of the collected signals due to the complex working environment of the internal components of the high-precision navigation sensor. When the deviation value exceeds the acceptable range of the system, the navigation sensor is slowly changed, the navigation precision is reduced, and the navigation performance is affected. There are many studies currently on active fault tolerance for navigation sensor snap-down faults, but relatively few related studies on snap-down faults. One reason is that the initial amplitude of occurrence of the gradual failure is small, and the amplitude of change with time is far smaller than that of the rapid failure, so that diagnosis of the gradual failure is difficult to complete in a short time. Secondly, the occurrence time of the slow-change fault is longer, the deviation from the true value is too large, and effective active fault tolerance is difficult to carry out. Aiming at researches related to a slow-change fault diagnosis algorithm of a navigation system, students at home and abroad usually adopt BP neural networks to extract fault characteristics. For example, in the aspect of inertial navigation unit fault detection Zhang Shufeng, research and optimization are performed on a IMUs fault detection and diagnosis model deep confidence network, a deep learning model with the deep confidence network as a core is adopted, a IMUs fault dataset of the wheeled robot is created for model training and testing, and optimized model fault diagnosis performance evaluation is completed. Li Gang et al propose a method for fault diagnosis of an aircraft inertial navigation gyroscope, which adopts an improved empirical mode decomposition-permutation entropy algorithm to perform fault detection, reconstructs decomposed navigation signals as a training set, and establishes a gyroscope fault diagnosis network to perform fault diagnosis by using a probabilistic neural network model with rapid learning capability and high accuracy. And (3) performing standard sensor signal prediction by using a BP neural network, constructing residual errors, and performing navigation sensor fault diagnosis and signal reconstruction by using a threshold method. The above algorithm can diagnose the slow-changing fault but takes a long time. And because the BP neural network is used, overfitting to the training data set is easy to cause, so that the diagnosis accuracy in practical application is reduced. Meanwhile, the sensor signal types reconstructed by the threshold method are limited, and multi-scene application cannot be realized. Aiming at the research of the active repair of the slow-changing fault signal of the navigation system, the threshold method and the BP neural network are adopted in engineering to perform the active repair of the slow-changing fault signal of the navigation sensor. Little research has been done in the active repair of navigation sensor creep failure using long and short term memory neural networks and convolutional neural networks. The threshold method is to calculate whether the residual error exceeds the threshold value by using a fuzzy membership function under the condition of stipulating the threshold value and diagnose faults so as to use the residual error for fault repair, but the method has less applicable scenes and the setting of the threshold value has stationarity. The navigation sensor fault active repair by adopting the BP neural network is the core of using the fully-connected neural network as a deep learning model. And constructing a training data set, and performing fault active fault tolerance on the navigation sensor by using the trained model. Because the learning ability of the BP neural network is relatively poor, the dependence on the data set is too strong and the fault data set is difficult to construct, the precision of fault detection is also affected. Disclosure of Invention The invention aims to provide an intelligent active fault-tolerant method for the slow-change fault of a navigation sensor, which realizes timely and accurate diagnosis of the slow-change fault and provides a repair signal, realizes rapid diagnosis of the slow-change fault of the navigation sensor and reconstruction of the navigation signal with generalization capability, and solves the problems of too slow diagnosis of the slow-change fault and single application scene of the traditional threshold method. The technical solution for realizing