An autonomous fault diagnosis method for deep space probes based on spiking neural networks

Abstract

The present invention discloses a deep space detector autonomous fault diagnosis method based on a pulse neural network, comprising the following steps: preprocessing the fault data collected by the sensor in the deep space detector using the local mean decomposition method LMD based on the time-frequency domain; analyzing the processed raw data to extract the corresponding fault features; using the random forest algorithm to select the extracted fault features; using the random forest algorithm to select the extracted fault features; normalizing the selected fault features into a pulse sequence with time information; constructing a multi-layer pulse neural network, inputting the pulse sequence to train the pulse neural network, and completing the autonomous fault diagnosis with the trained model. The present invention realizes autonomous fault diagnosis of the detector by applying the pulse neural network to the field of deep space detection fault diagnosis for the first time and identifying fault information.

Description

Technical Field

[0001] The present invention belongs to the field of fault handling of spacecraft in deep space exploration missions, and in particular relates to an autonomous fault diagnosis method for deep space detectors based on a pulse neural network. Background Art

[0002] With the development of computer science and space technology in recent years, it has provided good support for the intelligent deep space exploration system. Due to the long distance between the moon and the planets and the harsh environmental conditions, the survival and health of machinery and equipment can hardly cope with many critical situations by relying solely on ground diagnosis, intervention and guidance modes. Therefore, it is necessary to study and develop intelligent survival and health management technologies so that deep space probes can autonomously detect and analyze the environment and health conditions and establish certain autonomous repair capabilities. With the rapid development of cutting-edge technologies such as artificial intelligence, the research on brain-like intelligence has become the focus of various countries. Domestic and foreign researchers have explored how biological neural systems can achieve intelligent computing through multidisciplinary research such as neuroscience, cognitive science and information science, thereby simulating biological neural systems and realizing brain-like computing intelligent systems.

[0003] In the existing technology, the more mainstream fault diagnosis technologies include knowledge-based fault diagnosis technology, model-based fault diagnosis technology and data-driven fault diagnosis technology. The most commonly used deep space probe fault diagnosis methods mainly include artificial ant colony algorithm and fuzzy algorithm. However, most traditional fault detection methods are aimed at identifying and searching for periodic faults, and their autonomy is relatively poor. In the 1990s, the Livingstone system developed by NASA in the United States used fault diagnosis technology based on qualitative models. The system can realize the fault diagnosis of Mars probes with minimal human intervention. The Beijing Institute of Control Engineering developed a fault diagnosis system SCRDES, which completes the online real-time monitoring and diagnosis of satellite control systems based on the fault diagnosis method of expert systems. In recent years, due to the vigorous development of deep learning, new ideas have been provided for autonomous fault diagnosis of deep space probes. Through heterogeneous deep neural networks, intelligent adaptive fault-tolerant control of the full neural network is established, thereby realizing fault diagnosis and fault-tolerant control. However, since spacecraft faults are usually discretely distributed, and traditional neural network training often consumes too much energy. Summary of the invention

[0004] The present invention provides a deep space detector autonomous fault diagnosis method based on pulse neural network to solve the problems of high limitations, low accuracy and high energy consumption in the existing deep space detector fault diagnosis methods. By using pulse neural network training and learning, the faults of deep space detectors are identified, thereby realizing autonomous fault diagnosis.

[0005] To achieve the above object, the technical solution adopted by the present invention is:

[0006] A deep space detector autonomous fault diagnosis method based on a spiking neural network comprises the following steps:

[0007] S1. Preprocess the raw data collected by sensors in deep space probes;

[0008] S2. Analyze the processed raw data and extract the corresponding fault features;

[0009] S3. Use random forest algorithm to select the extracted fault features;

[0010] S4. normalizing the selected fault features into a pulse sequence with time information;

[0011] S5. Construct a multi-layer pulse neural network, input a pulse sequence to train the pulse neural network, and use the trained model to complete autonomous fault diagnosis.

[0012] Furthermore, the step S1 includes the following steps:

[0013] S11. According to the collected original signal data, the time-frequency domain signal decomposition algorithm LMD is used to decompose the original fault data. Let k be a constant for recording the sequence number of the local extreme point, and find all the local extreme points n of the fault signal x(t). k (k=1,2,…), calculate every two consecutive extreme points n k and n k+1 The mean value m k and the envelope estimate a k , where: m k =(n k +n k+1 ) / 2,a k =(|n k -n k+1 |) / 2; perform curve fitting on each discrete point to obtain the local mean function m 11 (t) and the local envelope function a 11 (t);

[0014] S12. Subtract the local mean function m from the fault signal x(t) 11 (t) Get the remaining signal h 11 (t), the remaining signal h 11 (t) divided by the local envelope function a 11 (t) get the frequency modulated signal s 11 (t); the frequency modulated signal s 11 (t) As a new fault signal, execute S21 and S22 repeatedly until a pure FM signal s is obtained. 1n (t), let n be the number of decompositions, and the frequency modulation signal obtained at the nth time is a 1n(t), where: a 1n (t) = 1;

[0015] S13. Let j be a constant that records the serial number of the local envelope function, j = 1, 2, 3..., all local envelope functions a 1j (t) is multiplied to obtain the envelope signal a1(t), represents the cumulative multiplication symbol, then Substituting a1(t) and s 1n (t) is multiplied to obtain the first instantaneous signal component PF1(t), PF1(t) = a1(t) × s 1n (t); Separate PF1(t) and the first residual signal r1(t) from the original fault signal as a new signal, and repeat steps S11 to S13 until r1(t) is monotonic. Let I be the number of cycles, i represents a constant from 1 to I, i=1,2…I, and obtain the decomposition result of the fault signal x(t):

[0016] Furthermore, the step S2 includes the following steps:

[0017] S21. Since the PF component of the instantaneous signal obtained after LMD decomposition still has a huge amount of data, directly using it as a feature vector will increase the complexity of the model. The original fault signal is divided into m segments, and each segment is decomposed into 4 instantaneous signal components by LMD, that is, PF1 is the first instantaneous signal component, PF2 is the second instantaneous signal component, PF3 is the third instantaneous signal component, PF4 is the fourth instantaneous signal component and the residual signal r(t). The PF component is feature extracted, and the residual signal is not feature extracted. Among them: the original fault signal is S, T is the signal length s1, s2…s T is the fault signal component obtained after decomposition, that is, S = (s1, s2, ... s T ), is the mean, σ is the standard deviation;

[0018] S22. Assume t S=(s1,s2,…s T ), t represents a constant from 1 to T t = 1, 2…T, and the skewness is obtained by feature extraction Kurtosis Kurtosis index Waveform indicators Pulse Indicator Margin Index Get the extracted features.

[0019] Furthermore, the step S3 includes the following steps:

[0020] S31. In a high-dimensional deep space environment, since it includes a large amount of irrelevant and redundant information, it is necessary to select fault features. The random forest algorithm is used to select fault features. The fault features extracted in S22 are obtained. Each PF component is used as a node. Each node has f features. A part of the features n on the node is randomly selected. sub , using the randomly selected sample features, select an optimal left and right subtree division as the decision tree;

[0021] S32. Input the fault feature set D, the number of iterations P, perform e random sampling on the fault feature set, and obtain a sampling set D containing e samples. t , use the sampling set to train the decision tree model G t (x) Each time, a part of all sample features on the training decision tree model node is selected, and the best one is selected from these randomly selected sample features as the left and right subtree divisions. After the final iteration is completed, the fault feature set is obtained and the feature selection is completed.

[0022] Further, the step S4 includes the following steps:

[0023] S41. Since the input and activation values ​​of each layer are transmitted in the form of pulse signals in the spiking neural network, the input sample data needs to be encoded. In order to encode the data, the fault feature vector is normalized using the min-max normalization method. The formula is: in: is the normalized value of the pth feature, x p is the pth feature, x max is the maximum value of the pth feature, x min is the minimum value of the pth feature;

[0024] S42. Each fault feature vector has 6 features (x1, x2, ... x6). The Gaussian group coding method is used to encode the input features into pulse time information that can be processed by the pulse neural network. Three Gaussian acceptance thresholds are used to encode each input feature into the firing time of the pulse neuron. That is, the firing time of three neurons is used to represent one feature. The original fault data is mapped to a higher-dimensional data space to expand the gap between the data.

[0025] Furthermore, the step S5 includes the following steps:

[0026] S51. Using the leaky integrate fire (LIF) neuron model, the input is divided into two parts, I R is the current flowing to the resistor R, I Cis the current flowing to capacitor C, let: t is time, u is the neuron membrane voltage, I(t) is the current function that changes with time, u(t) is the function that changes with time of neuron membrane voltage, is the time derivative of the neuron membrane voltage, Multiplying both sides of the equation by R gives Where: time constant τ m =RC;

[0027] S52. Update weights using gradient descent Among them: w is the weight, w new is the updated weight value, w post is the weight value before update, E is the loss function, α is the learning rate, and the cross entropy function E(y,y′) is selected as the loss function, E(y,y′)=-[ylogy′+(1-y)log(1-y′)], y is the actual output sequence obtained by the sample input data in the network, y′ is the target output sequence, and the network parameters of each layer are updated. Represents the weight of neuron a connected to neuron b in layer l. When the membrane voltage exceeds the threshold θ, the neuron sends a pulse, after which the neuron is at a resting potential u rest , usually u rest =0, Indicates whether the neurons in layer l emit pulses at time t, and takes values ​​of 0 or 1. When the membrane voltage exceeds the threshold θ, The value is 1, and then the voltage is at a resting potential of 0. If the voltage value does not reach the threshold, then Set the value to 0, continue updating the weight, and accumulate the voltage until the pulse can be triggered;

[0028] S53. Construct a three-layer pulse neural network model, including an input layer, a hidden layer and an output layer, input the encoded pulse data into the input layer, and set different threshold voltages θ for the hidden layer and the output layer respectively. h and θ o ,The connection mode between each two layers of neurons is fully connected, and the number of neurons in the output layer is equal to the number of fault types. The spiking neural network model is trained to complete autonomous fault diagnosis.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] The present invention applies the pulse neural network to the autonomous fault diagnosis field of deep space exploration of spacecraft for the first time, and performs autonomous fault diagnosis by identifying fault data. First, the original fault data collected by the detector is preprocessed, and the segmented signal data is obtained by time-frequency domain decomposition method; then the signal data obtained after decomposition is subjected to feature extraction; and the extracted fault features are selected by the random forest algorithm to remove redundant information; then the fault feature data is converted into a pulse sequence by Gaussian group coding; finally, a pulse neural network model is constructed, and the normalized data is input into the pulse neural network model for training, thereby completing autonomous fault diagnosis. The advantage of the present invention is that since the deep space detector fault is discrete, the discreteness of the fault can be effectively simulated by the pulse neuron; at the same time, the pulse neural network is an event-driven mechanism, and only when the pulse is emitted will there be energy loss, so the energy consumption can be greatly reduced; and experiments show that identifying the fault features and effectively selecting the features, and using the pulse neural network for fault diagnosis training can effectively complete autonomous fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flow chart of the present invention.

[0032] Figure 2 This is the fault data signal processing flow chart.

[0033] Figure 3 This is a fault diagnosis model diagram based on pulse neural network. DETAILED DESCRIPTION

[0034] The present invention will be further described below in conjunction with the embodiments.

[0035] like Figure 1 As shown, a deep space detector autonomous fault diagnosis method based on a pulse neural network includes the following steps:

[0036] S1. Preprocess the raw data collected by sensors in deep space probes;

[0037] S11. According to the collected original signal data, the time-frequency domain signal decomposition algorithm LMD is used to decompose the original fault data. Let k be a constant for recording the sequence number of the local extreme point, and find all the local extreme points n of the fault signal x(t). k (k=1,2,…), calculate every two consecutive extreme points n k and n k+1 The mean value m k and the envelope estimate a k , where: m k =(n k +n k+1 ) / 2,a k =(|nk -n k+1 |) / 2; perform curve fitting on each discrete point to obtain the local mean function m 11 (t) and the local envelope function a 11 (t);

[0038] S12. Subtract the local mean function m from the fault signal x(t) 11 (t) Get the remaining signal h 11 (t), the remaining signal h 11 (t) divided by the local envelope function a 11 (t) get the frequency modulated signal s 11 (t); the frequency modulated signal s 11 (t) As a new fault signal, execute S21 and S22 repeatedly until a pure FM signal s is obtained. 1n (t), let n be the number of decompositions, and the frequency modulation signal obtained at the nth time is a 1n (t), where: a 1n (t) = 1;

[0039] S13. Let j be a constant that records the serial number of the local envelope function, j = 1, 2, 3..., all local envelope functions a 1j (t) is multiplied to obtain the envelope signal a1(t), represents the cumulative multiplication symbol, then Substituting a1(t) and s 1n (t) is multiplied to obtain the first instantaneous signal component PF1(t), PF1(t) = a1(t) × s 1n (t); Separate PF1(t) and the first residual signal r1(t) from the original fault signal as a new signal, and repeat steps S11 to S13 until r1(t) is monotonic. Let I be the number of cycles, i represents a constant from 1 to I, i=1,2…I, and obtain the decomposition result of the fault signal x(t):

[0040] S2. Analyze the processed raw data and extract the corresponding fault features;

[0041] S21. Since the PF component of the instantaneous signal obtained after LMD decomposition still has a huge amount of data, directly using it as a feature vector will increase the complexity of the model. The original fault signal is divided into m segments, and each segment is decomposed into 4 instantaneous signal components by LMD, that is, PF1 is the first instantaneous signal component, PF2 is the second instantaneous signal component, PF3 is the third instantaneous signal component, PF4 is the fourth instantaneous signal component and the residual signal r(t). The PF component is feature extracted, and the residual signal is not feature extracted. Among them: the original fault signal is S, T is the signal length s1, s2…s Tis the fault signal component obtained after decomposition, that is, S = (s1, s2, ... s T ), is the mean, σ is the standard deviation;

[0042] S22. Assume t S=(s1,s2,…s T ), t represents a constant from 1 to T t = 1, 2…T, and the skewness is obtained by feature extraction Kurtosis Kurtosis index Waveform indicators Pulse Indicator Margin index Get the extracted features.

[0043] Specifically, if Figure 2 As shown, the original fault data is first decomposed by LMD to obtain four PF components and a residual signal, and then feature extraction is performed on each PF component. The residual data is not processed. Six features are extracted from each PF component, including skewness S, kurtosis K, kurtosis index CF, waveform index SF, pulse index IF, and margin index CLF, based on which the extracted features are obtained.

[0044] S3. Use random forest algorithm to select the extracted fault features;

[0045] S31. In a high-dimensional deep space environment, since it includes a large amount of irrelevant and redundant information, it is necessary to select fault features. The random forest algorithm is used to select fault features. The fault features extracted in S22 are obtained. Each PF component is used as a node. Each node has f features. A part of the features n on the node is randomly selected. sub , using the randomly selected sample features, select an optimal left and right subtree division as the decision tree;

[0046] S32. Input the fault feature set D, the number of iterations P, perform e random sampling on the fault feature set, and obtain a sampling set D containing e samples. t , use the sampling set to train the decision tree model G t (x) Each time, a part of all sample features on the training decision tree model node is selected, and the best one is selected from these randomly selected sample features as the left and right subtree divisions. After the final iteration is completed, the fault feature set is obtained and the feature selection is completed.

[0047] Specifically, the extracted fault features are selected using the random forest algorithm to determine the importance of each fault feature. When establishing each decision tree, the out-of-bag data error is calculated based on the corresponding out-of-bag data (OOB), that is, the data that is not used when establishing the decision tree. Then, some interference is added to the out-of-bag data sample features, and the out-of-bag error is calculated again. The difference between the two out-of-bag errors is divided by the number of decision trees to obtain the importance of each fault feature. The fault features are sorted according to their importance, and a certain proportion of features are selected for elimination.

[0048] S4. normalizing the selected fault features into a pulse sequence with time information;

[0049] S41. Since the input and activation values ​​of each layer are transmitted in the form of pulse signals in the spiking neural network, the input sample data needs to be encoded. In order to encode the data, the fault feature vector is normalized using the min-max normalization method. The formula is: in: is the normalized value of the pth feature, x p is the pth feature, x max is the maximum value of the pth feature, x min is the minimum value of the pth feature;

[0050] S42. Each fault feature vector has 6 features (x1, x2, ... x6). The Gaussian group coding method is used to encode the input features into pulse time information that can be processed by the pulse neural network. Three Gaussian acceptance thresholds are used to encode each input feature into the firing time of the pulse neuron. That is, the firing time of three neurons is used to represent one feature. The original fault data is mapped to a higher-dimensional data space to expand the gap between the data.

[0051] S5. Construct a multi-layer pulse neural network, input a pulse sequence to train the pulse neural network, and use the trained model to complete autonomous fault diagnosis.

[0052] S51. Using the leaky integrate fire (LIF) neuron model, the input is divided into two parts, I R is the current flowing to the resistor R, I C is the current flowing to capacitor C, let: t is time, u is the neuron membrane voltage, I(t) is the current function that changes with time, u(t) is the function that changes with time of neuron membrane voltage, is the time derivative of the neuron membrane voltage, Multiplying both sides of the equation by R gives Where: time constant τ m =RC;

[0053] S52. Update weights using gradient descent Among them: w is the weight, w new is the updated weight value, w post is the weight value before update, E is the loss function, α is the learning rate, and the cross entropy function E(y,y′) is selected as the loss function, E(y,y′)=-[ylogy′+(1-y)log(1-y′)], y is the actual output sequence obtained by the sample input data in the network, y′ is the target output sequence, and the network parameters of each layer are updated. Represents the weight of neuron a connected to neuron b in layer l. When the membrane voltage exceeds the threshold θ, the neuron sends a pulse, after which the neuron is at a resting potential u rest , usually u rest =0, Indicates whether the neurons in layer l emit pulses at time t, and takes values ​​of 0 or 1. When the membrane voltage exceeds the threshold θ, The value is 1, and then the voltage is at a resting potential of 0. If the voltage value does not reach the threshold, then Set the value to 0, continue updating the weight, and accumulate the voltage until the pulse can be triggered;

[0054] S53. Construct a three-layer pulse neural network model, including an input layer, a hidden layer and an output layer, input the encoded pulse data into the input layer, and set different threshold voltages θ for the hidden layer and the output layer respectively. h and θ o ,The connection mode between each two layers of neurons is fully connected, and the number of neurons in the output layer is equal to the number of fault types. The spiking neural network model is trained to complete autonomous fault diagnosis.

[0055] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A deep space detector autonomous fault diagnosis method based on pulse neural network, characterized in that: The following steps are involved: S1. Preprocess the raw data collected by sensors in deep space probes; S2. Analyze the processed raw data and extract the corresponding fault features; S3. Use random forest algorithm to select the extracted fault features; S4. normalizing the selected fault features into a pulse sequence with time information; S5. Construct a multi-layer spiking neural network, input a pulse sequence to train the spiking neural network, and use the trained model to complete autonomous fault diagnosis; The step S1 comprises the following steps: S11. According to the collected original signal data, the time-frequency domain signal decomposition algorithm LMD is used to decompose the original fault data. Let k be a constant for recording the sequence number of the local extreme point, and find all the local extreme points n of the fault signal x(t). k (k=1,2,…), calculate every two consecutive extreme points n k and n k+1 The mean value m k and the envelope estimate a k , where: m k =(n k +n k+1 ) / 2,a k =(|n k -n k+1 |) / 2; perform curve fitting on each discrete point to obtain the local mean function m 11 (t) and the local envelope function a 11 (t); S12. Subtract the local mean function m from the fault signal x(t) 11 (t) Get the remaining signal h 11 (t), the remaining signal h 11 (t) divided by the local envelope function a 11 (t) get the frequency modulated signal s 11 (t); the frequency modulated signal s 11 (t) As a new fault signal, execute S21 and S22 repeatedly until a pure FM signal s is obtained. 1n (t), let n be the number of decompositions, and the frequency modulation signal obtained at the nth time is a 1n (t), where: a 1n (t) = 1; S13. Let j be a constant that records the serial number of the local envelope function, j = 1, 2, 3..., all local envelope functions a 1j (t) is multiplied to obtain the envelope signal a1(t), represents the cumulative multiplication symbol, then Substituting a1(t) and s 1n (t) is multiplied to obtain the first instantaneous signal component PF1(t), PF1(t) = a1(t) × s 1n (t); Separate PF1(t) and the first residual signal r1(t) from the original fault signal as a new signal, and repeat steps S11 to S13 until r1(t) is monotonic. Let I be the number of cycles, i represents a constant from 1 to I, i=1,2…I, and obtain the decomposition result of the fault signal x(t): The step S2 comprises the following steps: S21. Since the PF component of the instantaneous signal obtained after LMD decomposition still has a huge amount of data, directly using it as a feature vector will increase the complexity of the model. The original fault signal is divided into m segments, and each segment is decomposed into 4 instantaneous signal components by LMD, that is, PF1 is the first instantaneous signal component, PF2 is the second instantaneous signal component, PF3 is the third instantaneous signal component, PF4 is the fourth instantaneous signal component and the residual signal r(t). The PF component is feature extracted, and the residual signal is not feature extracted. Among them: the original fault signal is S, T is the signal length s1, s2…s T is the fault signal component obtained after decomposition, that is, S = (s1, s2, ... s T ), is the mean, σ is the standard deviation; S22. Assume t S=(s1,s2,…s T ), t represents a constant from 1 to T t = 1, 2…T, and the skewness is obtained by feature extraction Kurtosis Kurtosis index Waveform indicators Pulse Indicator Margin index Get the extracted features; The step S3 comprises the following steps: S31. In a high-dimensional deep space environment, since it includes a large amount of irrelevant and redundant information, it is necessary to select fault features. The random forest algorithm is used to select fault features. The fault features extracted in S22 are obtained. Each PF component is used as a node. Each node has f features. A part of the features n on the node is randomly selected. sub , using the randomly selected sample features, select an optimal left and right subtree division as the decision tree; S32. Input the fault feature set D, the number of iterations P, perform e random sampling on the fault feature set, and obtain a sampling set D containing e samples. t , use the sampling set to train the decision tree model G t (x) Each time, a part of all sample features on the training decision tree model node is selected, and the best one among these randomly selected sample features is selected as the left and right subtree divisions. After the final iteration is completed, the fault feature set is obtained, and the feature selection is completed; The step S4 comprises the following steps: S41. Since the input and activation values ​​of each layer in the spiking neural network are transmitted in the form of pulse signals, the input sample data needs to be encoded. In order to encode the data, the fault feature vector is normalized using the min-max normalization method, and the formula is: in: is the normalized value of the pth feature, x p is the pth feature, x max is the maximum value of the pth feature, x min is the minimum value of the pth feature; S42. Each fault feature vector has 6 features (x1, x2, ... x6). The Gaussian group coding method is used to encode the input features into pulse time information that can be processed by the pulse neural network. Three Gaussian acceptance thresholds are used to encode each input feature into the firing time of the pulse neuron, that is, the firing time of three neurons is used to represent one feature, and the original fault data is mapped to a higher-dimensional data space to expand the gap between the data; The step S5 comprises the following steps: S51. Using the leaky release integral neuron model, the input is divided into two parts, I R is the current flowing to the resistor R, I C is the current flowing to capacitor C, let: t is time, u is the neuron membrane voltage, I(t) is the current function that changes with time, u(t) is the function that changes with time of neuron membrane voltage, is the time derivative of the neuron membrane voltage, Multiplying both sides of the equation by R gives Where: time constant t m =RC; S52. Update weights using gradient descent Among them: w is the weight, w new is the updated weight value, w post is the weight value before update, E is the loss function, α is the learning rate, and the cross entropy function E(y,y′) is selected as the loss function, E(y,y′)=-[ylogy′+(1-y)log(1-y′)], y is the actual output sequence obtained by the sample input data in the network, y′ is the target output sequence, and the network parameters of each layer are updated. Represents the weight of neuron a connected to neuron b in layer l. When the membrane voltage exceeds the threshold θ, the neuron sends a pulse, after which the neuron is at a resting potential u rest , usually u rest =0, Indicates whether the neurons in layer l emit pulses at time t, and takes values ​​of 0 or 1. When the membrane voltage exceeds the threshold θ, The value is 1, and then the voltage is at a resting potential of 0. If the voltage value does not reach the threshold, then Set the value to 0, continue updating the weight, and accumulate the voltage until the pulse can be triggered; S53. Construct a three-layer pulse neural network model, including an input layer, a hidden layer and an output layer, input the encoded pulse data into the input layer, and set different threshold voltages θ for the hidden layer and the output layer respectively. h and θ o ,The connection mode between each two layers of neurons is fully connected, and the number of neurons in the output layer is equal to the number of fault types. The spiking neural network model is trained to complete autonomous fault diagnosis.

Images