Electric drive system motor temporal and spatial characteristics neural network fault diagnosis method and system

Abstract

The present invention discloses a method and system for diagnosing motor faults in an electric drive system using a neural network of spatiotemporal characteristics, which is applicable to motor systems under various complex working conditions. The method of the present invention collects vibration signals from various faulty motors as original signals, preprocesses the original signals into a data sample set, and then performs multi-scale decomposition on the data in the sample set to obtain spatiotemporal feature outputs, and then performs population encoding on the spatiotemporal features into time pulses. Finally, the constructed pulse convolutional network model is used to train the inputs. The gradient descent method is introduced into the training model and combined with the convolution and batch normalization time layer strategies, which effectively improves the stability and classification accuracy of the network. Among them, the multi-scale decomposition and extraction obtains the local features of the signal, which is consistent with the local nature of the motor fault characteristics, significantly improving the accuracy of motor fault diagnosis and effectively ensuring the reliable operation of the electric drive system of electric transport equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of motor fault detection, and in particular to a method and system for diagnosing motor faults in an electric drive system using a spatiotemporal characteristic neural network. Background Art

[0002] As a key and core component of the electromechanical hybrid transmission system of electric transportation equipment, electric drive systems have been widely used in electric transportation equipment such as heavy-duty mining trucks, new energy special vehicles, and heavy-duty automatic elevators for transportation. Their operating performance, reliability, and environmental suitability are key constraints on their development. As the application areas of electric transportation equipment continue to expand, new requirements are being placed on electric drive systems, especially in some special applications. The long-term service of electric drive systems in extreme natural environments has become the new normal, thus placing new requirements on electric drive system motor fault diagnosis, especially in applications with high reliability requirements such as rail transportation, aerospace, and military equipment.

[0003] Existing motor fault diagnosis methods fall into two main categories: data-driven and model-based. Current model-based motor fault diagnosis methods primarily rely on prior knowledge. However, the structure and function of electric drive systems are extremely complex, making it difficult to construct an accurate mathematical model to simulate their motor systems and components. Artificial neural networks, a widely used tool, successfully map fault symptoms to fault categories through inter-layer learning mechanisms. Consequently, data-driven fault classification methods based on deep learning and other approaches have been widely studied and applied.

[0004] However, because the signals analyzed in motor fault diagnosis are typically in the form of one-dimensional time series, convolutional neural networks are significantly inadequate for extracting temporal features. Spiking neural networks, which use biologically plausible neuron modeling, incorporate the concept of time series into their models, enabling them to capture richer spatiotemporal dynamic features. However, spiking neural networks offer limited information about their neuromorphic structure and are difficult to accurately represent data. The simplified spiking neuron model limits its expressive power and dynamic diversity, making it difficult for spiking neural networks to achieve high classification accuracy in classification tasks.

[0005] Therefore, how to effectively improve the accuracy of motor fault diagnosis in electric drive systems remains a hot topic in this field and is a key issue in ensuring the reliable operation of electric drive systems for electric transportation equipment. Summary of the Invention

[0006] The purpose of the present invention is to address the technical problem of improving the accuracy of motor fault diagnosis in electric drive systems, and to provide a method and system for motor fault diagnosis in electric drive systems using a spatiotemporal neural network. The method of the present invention combines the respective advantages of convolutional neural networks and spiking neural networks, creatively proposing a new spiking convolutional neural network for motor fault diagnosis. Specifically, a batch normalization temporal layer (BNTT) and an integral-integration-release model (LIF) are sequentially inserted after the convolutional and fully connected layers in the convolutional and fully connected modules.

[0007] To this end, the present invention provides the following technical solutions:

[0008] In one aspect, the present invention provides a method for diagnosing motor faults in an electric drive system using a spatiotemporal feature neural network, comprising the following steps:

[0009] Step 1: Collect motor vibration signals under normal and multiple fault conditions to form an original data set. Then, perform at least sample slicing and sample labeling on the samples of the original data set to obtain sample sets under different fault conditions and normal conditions.

[0010] Step 2: Convert the motor vibration signal of each sample into a pulse signal of each sample in a pulse pattern, and divide the sample set into a training set and a test set;

[0011] Step 3: Construct a fault diagnosis model based on a pulse convolutional neural network, and input the pulse signal of the sample into the pulse convolutional neural network for model training;

[0012] The network architecture of the pulse convolutional neural network is as follows: a convolution module, a pooling layer, and a fully connected module are sequentially provided, and a batch normalization time layer BNTT and an integration-discharge model LIF are sequentially inserted after at least one convolution layer in the convolution module and after a fully connected layer in the fully connected module;

[0013] Step 4: Collect the motor vibration signal of the motor to be monitored and convert it into a pulse signal, which is then input into the trained fault diagnosis model for fault diagnosis.

[0014] Preferably, in the convolution module and the fully connected module, the data processing models of the network layer including the convolution layer / fully connected layer, the batch normalization time layer BNTT and the integration-discharge model LIF are as follows:

[0015]

[0016] Where, are the output membrane potentials of neuron i at time step t and time step t-1, respectively. Corresponding to the output of the network layer, neuron i is regarded as the output neuron / output channel and serves as the input of a network layer; is the pulse input of neuron j at time step t, that is, the input of the corresponding network layer, and neuron j is regarded as the input neuron / input channel; ω ij is the connection weight connecting neurons j and i; γ t is the learnable parameter at time step t, v is the voltage threshold, and λ is the leakage factor, which is used to simulate the leakage behavior of the LIF model. Represents the output of the batch normalization time layer BNTT;

[0017] Unnormalized membrane potential is the input of the batch normalization time layer BNTT, and the data processing of the batch normalization time layer BNTT is expressed as:

[0018]

[0019] Where, represents the normalized membrane potential, and are the mean and variance of the sample input of the batch normalization time layer BNTT corresponding to neuron i at time step t, and ε is a pre-given minimum value.

[0020] Preferably, the learnable parameter γ at the next time step is t+1 The learnable parameter γ from the previous time step t Update, the update formula is: γ t+1 =γ t -ηΔγ t , η is the preset coefficient;

[0021] Among them, the update difference Δγ t The calculation of is as follows:

[0022]

[0023] In the formula, m represents the total number of samples in a batch of model training, b is the sample index in the corresponding batch, and u t is the network output of the convolutional / fully connected module corresponding to time step t, The network output membrane potential corresponding to sample b at time step t is obtained by the output of all neurons i corresponding to sample b. constitute; is the output of the batch normalization time layer BNTT corresponding to sample b at time step t, which is composed of all neurons i corresponding to sample b. Composition; L is the loss function, is the symbol of partial derivative.

[0024] The mean of all neurons i and variance Construct the mean vector μ t and variance vector σ t , the calculation formula is:

[0025]

[0026] Where, It is the input of the batch normalization time layer BNTT corresponding to sample b at time step t, that is, the input of all neurons i corresponding to sample b Composition, m represents the total number of samples in a batch of model training, and b is the sample index in the corresponding batch.

[0027] Preferably, the model training process uses the time back propagation method to perform network training;

[0028] Among them, the weight gradient formula of the connected neurons in the network layer without the batch normalization time layer BNTT is:

[0029]

[0030] The weight gradient formula of the connected neurons in the network layer containing the batch normalization time layer BNTT is:

[0031]

[0032] Where Δω ij is the connection weight ω connecting neurons j and i ij The weight gradient is used to update the connection weight ω ij ; T is the total number of time steps set, L is the loss function, is the corresponding network layer output of neuron i at time step t, is the output membrane potential of neuron i at the network layer at time step t; It is the input of the batch normalization time layer BNTT corresponding to sample b at time step t; represents the output of the batch normalization time layer BNTT corresponding to sample b and sample k; m represents the total number of samples in a batch in model training, b and k are the sample indices in the corresponding batch, ε is a pre-given minimum value, and σ t It is the variance of the sample input of the batch normalization time layer BNTT under time step t, T and t represent the total number of time steps and time step index respectively.

[0033] Preferably, there are:

[0034]

[0035] Among them, the proxy gradient descent method of piecewise linear approximation is used to solve as follows:

[0036]

[0037] Where the hyperparameter ξ is the attenuation factor of the back-propagation gradient, and v is the set threshold.

[0038] Preferably, in step 2, the process of converting the motor vibration signal of each sample into a pulse signal of each sample encoded in a pulse mode is as follows:

[0039] The motor vibration signal of each sample is extracted using low-pass and high-pass filters. The approximation coefficients of the sample data wavelet decomposition and the bi-norm of all detail coefficients are combined into a one-dimensional array as the fault feature output of the sample.

[0040] The Gaussian function is used to perform population coding on the fault feature output of each sample to obtain the pulse signal of each sample.

[0041] Among them, the technical solution of the present invention uses Gaussian population coding to encode input features into discrete pulses for communication, laying the foundation for the subsequent discretized pulse neuron model, so that the neuron will only be activated when the pulse signal received by the neuron reaches the threshold. This is an accurate imitation of the biological brain and is also more energy-efficient.

[0042] Preferably, the fault type includes any combination of an eccentricity fault, a single-phase fault, a broken bar fault, a bearing fault and a short circuit fault.

[0043] In another aspect, the present invention further provides a diagnostic system based on the above-mentioned fault diagnosis method, comprising at least:

[0044] The data acquisition and processing module is used to collect motor vibration signals under normal and multiple fault conditions to form an original data set, and then perform at least sample slicing and sample labeling on the samples of the original data set to obtain sample sets under different fault conditions and normal conditions;

[0045] An encoding module, configured to convert the motor vibration signal of each sample into a pulse signal of each sample in a pulse mode, and divide the sample set into a training set and a test set;

[0046] A model building and training module is used to build a fault diagnosis model based on a pulse convolutional neural network and input the pulse signal of the sample into the pulse convolutional neural network for model training;

[0047] The network architecture of the pulse convolutional neural network is as follows: a convolution module, a pooling layer, and a fully connected module are sequentially provided, and a batch normalization time layer BNTT and an integration-discharge model LIF are sequentially inserted after the convolution layer and the fully connected layer in the convolution module and the fully connected module;

[0048] The diagnostic module is used to collect the motor vibration signal of the motor to be monitored, convert it into a pulse signal, and then input it into the trained fault diagnosis model for fault diagnosis.

[0049] In a third aspect, the present invention provides a computer terminal comprising at least:

[0050] one or more processors and memory storing one or more computer programs;

[0051] The processor calls the computer program to implement:

[0052] Load a pre-trained fault diagnosis model or build and train a fault diagnosis model;

[0053] Obtain the motor vibration signal of the motor to be monitored and convert it into a pulse signal, which is then input into the trained fault diagnosis model for fault diagnosis;

[0054] The fault diagnosis model is constructed based on a pulse convolutional neural network. The network architecture of the pulse convolutional neural network is as follows: a convolution module, a pooling layer, and a fully connected module are sequentially provided, and a batch normalization time layer BNTT and an integration-discharge model LIF are inserted after the convolution layer and the fully connected layer in the convolution module and the fully connected module;

[0055] The input of the fault diagnosis model is the pulse signal, and the output is the fault diagnosis result.

[0056] In a fourth aspect, the present invention provides a computer-readable storage medium storing a computer program, wherein the computer program is invoked by a processor to implement:

[0057] Load a pre-trained fault diagnosis model or build and train a fault diagnosis model;

[0058] Obtain the motor vibration signal of the motor to be monitored and convert it into a pulse signal, which is then input into the trained fault diagnosis model for fault diagnosis;

[0059] The fault diagnosis model is constructed based on a pulse convolutional neural network. The network architecture of the pulse convolutional neural network is as follows: a convolution module, a pooling layer, and a fully connected module are sequentially provided, and a batch normalization time layer BNTT and an integration-discharge model LIF are inserted after the convolution layer and the fully connected layer in the convolution module and the fully connected module;

[0060] The input of the fault diagnosis model is the pulse signal, and the output is the fault diagnosis result.

[0061] Beneficial effects

[0062] Compared with the existing method, the advantages of the present invention are:

[0063] The technical solution of the present invention creatively proposes a new pulse convolutional neural network to realize motor fault diagnosis, that is, the batch normalization time layer BNTT and the integration-emission model LIF are inserted in sequence after the convolution layer and the fully connected layer in the convolution module and the fully connected module, integrating the respective advantages of the convolutional neural network and the pulse neural network, converting the motor vibration signal of the motor into a pulse signal and integrating it into the pulse neural network, making full use of the spatiotemporal dynamic characteristics of the original motor vibration signal, effectively solving the problems of insufficient local feature extraction of the motor vibration signal and slow processing of large-scale data, and effectively improving the stability and classification accuracy of the neural network. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 Schematic diagram of the basic process of the method of the embodiment of the present invention.

[0065] Figure 2 Schematic diagram of a method for diagnosing motor faults in an electric drive system based on a neural network using spatiotemporal characteristics according to an embodiment of the present invention.

[0066] Figure 3 Schematic diagram of the “drain-integrate-discharge” behavior of the LIF provided by an embodiment of the present invention.

[0067] Figure 4 This is a training flow chart of a pulse convolutional neural network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0068] The electric drive system motor spatiotemporal feature neural network fault diagnosis method and system provided by the present invention are applied to the fault diagnosis of the electric drive system motor. The core of the method is to propose a pulse convolutional neural network that integrates the convolutional neural network and the pulse neural network, and introduce the batch normalization time layer BNTT.

[0069] The present invention will be further described below with reference to the embodiments.

[0070] Example 1:

[0071] like Figure 1 As shown, this embodiment provides a method for diagnosing motor faults in an electric drive system using a spatiotemporal characteristic neural network, including the following steps:

[0072] Step 1: Collect motor vibration signals under normal and multiple fault conditions to form an original data set. Then, perform sample slicing, sample labeling, and sample shuffling on the samples of the original data set to obtain sample sets under different fault states and normal states.

[0073] In this embodiment, normal and five different motor fault vibration signals collected on the motor fault experimental platform are used, and the selected working conditions include variable speed working conditions and variable load working conditions. Figure 3 Schematic diagram of a faulty motor. In this embodiment, the faults set are eccentricity, single-phase, broken bars, bearing, and short circuit. For each fault, sensors are used to collect four sets of vibration signals at different locations. Other feasible embodiments are not limited to the above fault types.

[0074] The collected vibration signal is a one-dimensional time series, which is sliced into samples. Then, the samples with the same fault are labeled with the same label and randomly shuffled into a sample set.

[0075] Step 2: Convert the motor vibration signal of each sample into a pulse signal of each sample encoded in a pulse pattern, and divide the sample set into a training set and a test set.

[0076] This embodiment preferably uses the following method to perform conversion encoding:

[0077] S2-1: Use low-pass filters and high-pass filters to extract features from the motor vibration signal of each sample, and combine the approximation coefficients of the wavelet decomposition of the sample data and the bi-norm of all detail coefficients into a one-dimensional array as the fault feature output of the sample.

[0078] Among them, feature extraction is performed on the sample set data under various motor states, and the features of each fault are output at different frequencies and time scales. In order to extract the frequency features and time scale features, low-pass filters and high-pass filters are used to obtain multi-level approximation coefficients and detail coefficients, as follows:

[0079]

[0080] Then, the approximation coefficient and the bi-norm of all detail coefficients are combined into a one-dimensional array as the fault feature output Y of the sample: Y = [a j [k j ],b j [k j ],b j-1 [k j-1 ]...,b1[k1]]

[0081] Where A j [n] is the approximation coefficient of the jth layer, A0[n] is the input data, b j [n] is the detail coefficient of the jth layer, h[n] is the low-pass filter, g[n] is the high-pass filter, n represents the sampling point index, k j+1 Indicates the number of decomposition coefficients of this layer. In this embodiment, it is preferred that the number k of each decomposition layer is not consistent. The present invention obtains the characteristics of the data at different frequencies and time scales through the above multi-level decomposition.

[0082] S2-2: Use the Gaussian function to perform population encoding on the fault feature output of each sample to obtain the pulse signal of each sample.

[0083] In this embodiment, the extracted fault features are first simply normalized, and then the input normalized features are rationally encoded using multiple overlapping Gaussian functions, mapping each input variable to the pulse time of multiple input neurons. Preferably, each input feature is encoded by a set of Gaussian receptive fields containing M Gaussian functions, and the mean C of the i-th Gaussian function is i and standard deviation σ i The formula is defined as follows:

[0084]

[0085] In the above formula, I max and I min are the maximum and minimum values of the input features corresponding to each sample, and N are adjustable constants for the standard deviation σ, and M is the number of Gaussian functions. The encoding process of this embodiment can be understood as calculating the activation values of the M neurons encoding the variable x. Highly activated neurons will fire prematurely, while less activated neurons will fire later or not at all. The encoding variable x is a normalized feature.

[0086] It should be understood that this embodiment utilizes the Gaussian receptive fields of M Gaussian functions for encoding, and then encodes the normalized fault signature output into a discrete pulse signal, with each characteristic variable corresponding to a set of M encoded values. In other feasible embodiments, other feasible methods for encoding the fault signature output into a pulse signal also fall within the scope of protection of the present invention.

[0087] Among them, the processed sample set is divided into a training set and a test set, and an appropriate batch size is selected for training. In this embodiment, the pulsed fault samples are concentrated, and then randomly divided into a training set and a test set according to a certain ratio, and then the parameters are compared to select an appropriate batch size for training.

[0088] Step 3: Construct a fault diagnosis model based on a pulse convolutional neural network, and input the pulse signal of the sample into the pulse convolutional neural network for model training.

[0089] like Figure 4 As shown in the figure, the network architecture of the pulse convolutional neural network is as follows: a convolution module, a pooling layer, and a fully connected module are provided in sequence, and a batch normalization time layer (BNTT) and an integral-discharge model LIF are inserted in sequence after the convolution layer and the fully connected layer in the convolution module and the fully connected module.

[0090] The above-mentioned spiking convolutional neural network is constructed based on the fusion of spiking neural network and convolutional neural network, that is, the LIF model is combined with the convolutional layer and the fully connected layer. In each time step cycle, membrane potential leakage, pulse emission and membrane potential reset operations are added after each convolutional layer and fully connected layer. Figure 3 Shown is the “drain-integrate-discharge” behavior of the LIF model.

[0091] The batch normalization time layer BNTT introduced in this paper is used to optimize the time-dependent learnable parameter γ to capture the input distribution that changes over time, thereby stabilizing the reverse gradient during training and ensuring better deep learning of SNN.

[0092] Based on the above structure, the data processing models of the network layers including the convolutional layer / fully connected layer, the batch normalization time layer BNTT, and the integration-discharge model LIF in the convolutional module and the fully connected module are as follows:

[0093]

[0094] Where, are the output membrane potentials of neuron i at time step t and time step t-1, respectively. Corresponding to the output of the network layer, neuron i is regarded as the output neuron / output channel; is the pulse input of neuron j at time step t, that is, the input of the corresponding network layer, and neuron j is regarded as the input neuron / input channel; ω ij is the connection weight connecting neurons j and i; γ t is the learnable parameter at time step t, v is the voltage threshold, and λ is the leakage factor, which is used to simulate the leakage behavior of the LIF model. Represents the output of the batch normalization time layer BNTT;

[0095] Unnormalized membrane potential is the input of the batch normalization time layer BNTT, and the data processing of the batch normalization time layer BNTT is expressed as:

[0096]

[0097] Where, represents the normalized membrane potential, and are the mean and variance of the sample input of the batch normalization time layer BNTT corresponding to neuron i at time step t, and ε is a pre-given minimum value.

[0098] Among them, the mean of all neurons i and variance Construct the mean vector μ t and variance vector σt , the calculation formula is:

[0099]

[0100] Where, It is the input of the batch normalization time layer BNTT corresponding to sample b at time step t, that is, the input of all neurons i corresponding to sample b Composition, m represents the total number of samples in a batch of model training, and b is the sample index in the corresponding batch.

[0101] Use different γ at all time steps t value, γ t Find the optimal value at each time step to achieve efficient reasoning. That is, the learnable parameter γ in the next time step t+1 The learnable parameter γ from the previous time step t Update, Update | t γ t+1 =γ t -ηΔγ t , where Δγ t The calculation of is as follows:

[0102]

[0103] Where u t is the network output membrane potential of the convolutional / fully connected module corresponding to time step t, The network output membrane potential corresponding to sample b at time step t is obtained by the output of all neurons i corresponding to sample b. constitute; is the output of the batch normalization time layer BNTT corresponding to sample b at time step t, which is composed of all neurons i corresponding to sample b. constitute.

[0104] It should be understood that formulas (3)-(6) are all applicable to the operations of the convolution module and the fully connected module that integrate BNTT and the integral-disbursement model LIF. The data of the fully connected module is the classification result of the fault diagnosis module after the Softmax layer.

[0105] The model training of this embodiment adopts the time back propagation method to perform network training. Therefore, the operation of the pulse neurons in the pulse convolution model can be expanded along the time dimension, and the state of the neurons will be updated at each time step. Therefore, the weight gradient connecting the neurons is calculated as the accumulation at each time step.

[0106] Among them, the weight gradient formula of the connected neurons in the local network part that does not include the batch normalization time layer BNTT is:

[0107]

[0108] Adding a BNTT layer to a LIF neuron changes the gradient calculation. For the sample input signal of the BNTT layer, the gradient value of the BNTT layer through the next layer can be calculated. Therefore, the weight gradient formula of the connected neurons in the local network part containing the batch normalization time layer BNTT is:

[0109]

[0110] And there exists:

[0111] The above formula omits a neuron index i, where Δω ij is the connection weight ω connecting neurons j and i ij The weight gradient is used to update the connection weight ω ij ; T is the total number of time steps set, L is the loss function, and this embodiment selects the classification cross entropy loss function; is the corresponding network output of neuron i at time step t, is the network output membrane potential of neuron i at time step t; It is the input of the batch normalization time layer BNTT corresponding to sample b at time step t; represents the output of the batch normalization time layer BNTT corresponding to sample b and sample k; L is the loss function, m represents the total number of samples in a batch in model training, b and k are the sample indices in the corresponding batch, ε is a pre-given minimum value, σ t It is the variance of the sample input of the batch normalization time layer BNTT under time step t, T and t represent the total number of time steps and time step index respectively.

[0112] Since the threshold function o i The derivative of is a Dirkat function that is undefined at the peak of the pulse and is 0 elsewhere, so the gradient Δω ij It is difficult to calculate. The proxy gradient descent method using piecewise linear approximation is as follows:

[0113]

[0114] In the above formula, the hyperparameter ξ is the attenuation factor of the back-propagation gradient, and v is the threshold. The hyperparameter ξ should be set according to the size of the time step T. Because gradients accumulate at each time step, a smaller ξ is used for larger T to avoid gradient explosion.

[0115] It should be noted that the technical solution of the present invention effectively solves the application problem of the standard back-propagation algorithm in the pulse neural network by introducing the proxy gradient method. That is, due to the particularity of the output mode of the pulse neural network and its dependence on time, it is necessary to process pulse signals in actual application. Therefore, it cannot directly apply the gradient descent back-propagation algorithm based on differential calculation; in other feasible embodiments, other technical means can be adopted to solve this problem.

[0116] Based on the above theoretical analysis, the network model of the present invention is as follows:

[0117] Set a time step parameter T, for each time step t = 0, 1, ... T-1 and each layer of neurons l = 1, 2, ... L-1, have their own connection weight tensor W l , BNTT parameter γ l and the voltage threshold v of the lth network layer l . Update the membrane potential set by forward propagation When the membrane potential exceeds the threshold v l When the output spike The specific implementation method is as follows:

[0118]

[0119] In the above formula, capital letters represent vectors and lowercase letters represent scalar values. Formula (10) corresponds to the network layer with BNTT, and formula (11) corresponds to the network layer without BNTT. and are the membrane potential and output pulse of all neurons in the first layer with time step t, W l is the weight w of all neurons connecting the lth layer and the l-1th layer ij where λ is a constant leakage factor for the membrane potential. The membrane potential in the final layer is accumulated without the leakage factor and without using a threshold function. This allows the output value to remain continuous, and finally a Softmax layer is used for diagnosis to identify the fault type.

[0120] Iterate the dataset by formula (10) to calculate the model gradient by back-propagating the loss L through the unfolded neural network. Here, the loss L is the loss of the last layer. The categorical cross entropy function between the cumulative membrane potential and the ground truth label Y (the score for each class of faults) is defined as:

[0121]

[0122] Based on the above theoretical statements, the present invention uses sample data to train the model, updates the weight parameters, obtains the data classification model of the network through continuous iteration, inputs the test set into the data classification model, and obtains the classification accuracy of the test set.

[0123] Step 4: Collect the motor vibration signal of the motor to be monitored and convert it into a pulse signal, which is then input into the trained fault diagnosis model for fault diagnosis.

[0124] In summary, the technical solution of the present invention fully extracts the spatiotemporal features of the original signal, effectively solving the problems of insufficient extraction of local features of motor vibration signals and slow processing of large-scale data. Gaussian population coding is used to encode input features into discrete pulses for communication. Neurons are activated only when the pulse signal received by the neuron reaches the threshold. This is an accurate imitation of the biological brain and is more energy-efficient. Combining the convolutional pooling layer with BNTT with SNN and using the alternative gradient method in back propagation effectively improves the stability and classification accuracy of the pulse neural network. In addition, in order to test its performance, relevant comparative experiments were carried out in feature extraction and model training, and it was found that not only higher accuracy can be obtained in two different data sets, but also the computational cost is low.

[0125] Example 2:

[0126] This embodiment provides a diagnostic system based on the above-mentioned fault diagnosis method, which at least includes: a data acquisition and processing module, an encoding module, a model construction and training module, and a diagnostic module that are communicatively connected to each other / communicatively connected in sequence.

[0127] Among them, the data acquisition and processing module is used to collect motor vibration signals under normal and multiple types of fault conditions to form an original data set, and then perform sample slicing and sample labeling on the samples of the original data set to obtain sample sets under different fault conditions and normal conditions. The encoding module is used to convert the motor vibration signal of each sample into a pulse signal of each sample in a pulse mode, and divide the sample set into a training set and a test set. The model construction and training module is used to construct a fault diagnosis model based on a pulse convolutional neural network, and input the pulse signal of the sample into the pulse convolutional neural network for model training; the diagnosis module is used to collect the motor vibration signal of the motor to be monitored, convert it into a pulse signal, and then input it into the trained fault diagnosis model for fault diagnosis.

[0128] Among them, the network architecture of the pulse convolutional neural network is: a convolution module, a pooling layer and a fully connected module are provided in sequence, and a batch normalization time layer BNTT and an integration-discharge model LIF are inserted in sequence at least after one convolution layer in the convolution module and after one fully connected layer in the fully connected module.

[0129] Please refer to the above method content for the specific implementation process of each module. The present invention will not go into details here, and the division of the above functional modules is only for illustration. In some embodiments, some functional modules can be merged, and some functional modules can be split. Each functional module can be implemented in software or hardware or a combination of software and hardware. Among them, the software and hardware equipment includes but is not limited to general-purpose computer equipment, programmable gate arrays, digital signal processors, microprocessors and their corresponding programming or burning software.

[0130] Example 3:

[0131] This embodiment provides a computer terminal comprising at least one or more processors and a memory storing one or more computer programs. The processors invoke the computer programs to: load a pre-trained fault diagnosis model or construct and train a fault diagnosis model; obtain a motor vibration signal of a motor to be monitored and convert it into a pulse signal, which is then input into the trained fault diagnosis model for fault diagnosis;

[0132] Among them, the fault diagnosis model is constructed based on a pulse convolutional neural network. The network architecture of the pulse convolutional neural network is: a convolution module, a pooling layer and a fully connected module are provided in sequence, and a batch normalization time layer BNTT and an integration-discharge model LIF are inserted in sequence after at least one convolution layer in the convolution module and after a fully connected layer in the fully connected module; the input of the fault diagnosis model is a pulse signal, and the output is a fault diagnosis result.

[0133] For the specific implementation process of each step, please refer to the description of the above method.

[0134] It should be understood that in the embodiments of the present invention, the processor referred to may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The memory may include a read-only memory and a random access memory, and provides instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.

[0135] Example 4:

[0136] This embodiment provides a computer-readable storage medium storing a computer program, which is called by a processor to implement: loading a pre-trained fault diagnosis model or building and training a fault diagnosis model; obtaining a motor vibration signal of a motor to be monitored and converting it into a pulse signal, which is then input into the trained fault diagnosis model for fault diagnosis;

[0137] Among them, the fault diagnosis model is constructed based on a pulse convolutional neural network. The network architecture of the pulse convolutional neural network is: a convolution module, a pooling layer and a fully connected module are provided in sequence, and a batch normalization time layer BNTT and an integration-discharge model LIF are inserted in sequence after at least one convolution layer in the convolution module and after a fully connected layer in the fully connected module; the input of the fault diagnosis model is a pulse signal, and the output is a fault diagnosis result.

[0138] For the specific implementation process of each step, please refer to the description of the above method.

[0139] The readable storage medium is a computer-readable storage medium, which can be an internal storage unit of the software and hardware device described in any of the aforementioned embodiments, such as a hard disk or memory of a controller. The readable storage medium can also be an external storage device of the controller, such as a plug-in hard disk equipped on the controller, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. Furthermore, the readable storage medium can also include both an internal storage unit of the controller and an external storage device. The readable storage medium is used to store the computer program and other programs and data required by the controller. The readable storage medium can also be used to temporarily store data that has been output or is to be output.

[0140] Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes instructions for causing a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned readable storage medium includes various media that can store program code, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0141] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-readable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The present application is a flow chart according to the method, device (system), and computer program product of the embodiment of the present application and / or the instructions executed by the processor to generate a device for realizing the function specified in one flow chart or multiple flows and / or one box or multiple boxes of the block diagram. These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a product comprising an instruction device, which realizes the function specified in one flow chart or multiple flows and / or one box or multiple boxes of the block diagram. These computer program instructions may also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0142] It should be emphasized that the examples described in the present invention are illustrative rather than restrictive. Therefore, the present invention is not limited to the examples described in the specific embodiments. Any other embodiments derived by those skilled in the art based on the technical solution of the present invention that do not depart from the purpose and scope of the present invention, whether modified or replaced, also fall within the scope of protection of the present invention.

Claims

1. A method for fault diagnosis of electric drive system motors based on spatiotemporal characteristics of neural networks, characterized by: The following steps are involved: Step 1: Collect motor vibration signals under normal and multiple fault conditions to form an original data set. Then, perform at least sample slicing and sample labeling on the samples of the original data set to obtain sample sets under different fault conditions and normal conditions. Step 2: Convert the motor vibration signal of each sample into a pulse signal of each sample encoded in a pulse pattern, and divide the sample set into a training set and a test set; Step 3: Construct a fault diagnosis model based on a pulse convolutional neural network, and input the pulse signal of the sample into the pulse convolutional neural network for model training; The network architecture of the pulse convolutional neural network is as follows: a convolution module, a pooling layer, and a fully connected module are sequentially provided, and a batch normalization time layer BNTT and an integration-discharge model LIF are sequentially inserted after at least one convolution layer in the convolution module and after a fully connected layer in the fully connected module; Step 4: Collect the motor vibration signal of the motor to be monitored and convert it into a pulse signal, which is then input into the trained fault diagnosis model for fault diagnosis; The process of converting the motor vibration signal of each sample into a pulse signal of each sample in the pulse mode in step 2 is as follows: The motor vibration signal of each sample is extracted using a low-pass filter and a high-pass filter. The approximation coefficients of the sample data wavelet decomposition and the two-norm of all detail coefficients are combined into a one-dimensional array as the fault feature output Y of the sample, as follows: Y=[a j [k j ],b j [k j ],b j-1 [k j-1 ]...,b1[k1]] The fault feature output of each sample is first normalized, and then multiple overlapping Gaussian functions are used for population coding to obtain the pulse signal of each sample, that is, each sample is encoded by a set of Gaussian receptive fields containing M Gaussian functions, and the mean value C of the i-th Gaussian function is i and standard deviation σ i The formula is defined as follows: In the above formula, I max and I min are the maximum and minimum values of the input features corresponding to each sample, and N are adjustable constants of the standard deviation σ, M is the number of Gaussian functions, and i represents the i-th Gaussian function; The model training process uses the time back propagation method to train the network; Among them, the weight gradient formula of the connected neurons in the network layer without the batch normalization time layer BNTT is: The weight gradient formula of the connected neurons in the network layer containing the batch normalization time layer BNTT is: Where Δω ij is the connection weight ω connecting neurons j and i ij The weight gradient is used to update the connection weight ω ij ; T is the total number of time steps set, L is the loss function, is the corresponding network layer output of neuron i at time step t, is the output membrane potential of neuron i at the network layer at time step t; It is the input of the batch normalization time layer BNTT corresponding to sample b at time step t; represents the output of the batch normalization time layer BNTT corresponding to sample b and sample k; m represents the total number of samples in a batch in model training, b and k are the sample indices in the corresponding batch, ε is a pre-given minimum value, and σ t It is the variance of the sample input of the batch normalization time layer BNTT under time step t, where T and t represent the total number of time steps and time step index respectively.

2. The method according to claim 1, wherein: In the convolution module and the fully connected module, the data processing models of the network layer including the convolution layer / fully connected layer, the batch normalization time layer BNTT and the integration-discharge model LIF are as follows: Where, are the output membrane potentials of neuron i at time step t and time step t-1, respectively. Corresponding to the output of the network layer, neuron i is regarded as the output neuron / output channel; is the pulse input of neuron j at time step t, that is, the input of the corresponding network layer, and neuron j is regarded as the input neuron / input channel; ω ij is the connection weight connecting neurons j and i; γ t is the learnable parameter at time step t, v is the voltage threshold, and λ is the leakage factor, which is used to simulate the leakage behavior of the integration-release model LIF. Represents the output of the batch normalization time layer BNTT; Unnormalized membrane potential is the input of the batch normalization time layer BNTT, and the data processing of the batch normalization time layer BNTT is expressed as: Where, represents the normalized membrane potential, and are the mean and variance of the sample input of the batch normalization time layer BNTT corresponding to neuron i at time step t, and ε is a pre-given minimum value.

3. The method according to claim 2, wherein: The learnable parameter γ at the next time step t+1 The learnable parameter γ from the previous time step t Update, the update formula is: γ t+1 =γ t -ηΔγ t , η is the preset coefficient; Among them, the update difference Δγ t The calculation of is as follows: In the formula, m represents the total number of samples in a batch of model training, b is the sample index in the corresponding batch, and u t is the network output membrane potential of the convolutional / fully connected module corresponding to time step t, is the output membrane potential of the network layer corresponding to sample b at time step t, which is the output membrane potential of all neurons i corresponding to sample b. constitute; is the output of the batch normalization time layer BNTT corresponding to sample b at time step t, which is composed of all neurons i corresponding to sample b. Composition; L is the loss function, is the symbol of partial derivative.

4. The method according to claim 2, wherein: The mean of all neurons i and variance Construct the mean vector μ t and variance vector σ t , the calculation formula is: Where, It is the input of the batch normalization time layer BNTT corresponding to sample b at time step t, that is, the input of all neurons i corresponding to sample b Composition, m represents the total number of samples in a batch of model training, and b is the sample index in the corresponding batch.

5. The method according to claim 1, wherein: exist: Among them, the proxy gradient descent method of piecewise linear approximation is used to solve as follows: Where the hyperparameter ξ is the attenuation factor of the back-propagation gradient, and v is the set threshold.

6. A diagnostic system based on the method according to any one of claims 1 to 5, characterized in that: At least: The data acquisition and processing module is used to collect motor vibration signals under normal and multiple fault conditions to form an original data set, and then perform at least sample slicing and sample labeling on the samples of the original data set to obtain sample sets under different fault conditions and normal conditions; An encoding module, configured to convert the motor vibration signal of each sample into a pulse signal of each sample in a pulse mode, and divide the sample set into a training set and a test set; A model building and training module is used to build a fault diagnosis model based on a pulse convolutional neural network and input the pulse signal of the sample into the pulse convolutional neural network for model training; The network architecture of the pulse convolutional neural network is as follows: a convolution module, a pooling layer, and a fully connected module are sequentially provided, and a batch normalization time layer BNTT and an integration-discharge model LIF are sequentially inserted after the convolution layer and the fully connected layer in the convolution module and the fully connected module; The diagnostic module is used to collect the motor vibration signal of the motor to be monitored, convert it into a pulse signal, and then input it into the trained fault diagnosis model for fault diagnosis.

7. A computer terminal, characterized in that: At least: one or more processors and memory storing one or more computer programs; The processor calls the computer program to implement: Loading a pre-trained fault diagnosis model or constructing and training a fault diagnosis model, wherein the fault diagnosis model is constructed and trained according to steps 1 to 3 of the electric drive system motor spatiotemporal characteristic neural network fault diagnosis method according to claim 1; Obtain the motor vibration signal of the motor to be monitored and convert it into a pulse signal, which is then input into the trained fault diagnosis model for fault diagnosis; The fault diagnosis model is constructed based on a pulse convolutional neural network. The network architecture of the pulse convolutional neural network is as follows: a convolution module, a pooling layer, and a fully connected module are sequentially provided, and a batch normalization time layer BNTT and an integration-discharge model LIF are inserted after the convolution layer and the fully connected layer in the convolution module and the fully connected module; The input of the fault diagnosis model is the pulse signal, and the output is the fault diagnosis result.

8. A computer-readable storage medium, characterized in that: A computer program is stored, which is called by a processor to implement: Loading a pre-trained fault diagnosis model or constructing and training a fault diagnosis model, wherein the fault diagnosis model is constructed and trained according to steps 1 to 3 of the electric drive system motor spatiotemporal characteristic neural network fault diagnosis method according to claim 1; Obtain the motor vibration signal of the motor to be monitored and convert it into a pulse signal, which is then input into the trained fault diagnosis model for fault diagnosis; The fault diagnosis model is constructed based on a pulse convolutional neural network. The network architecture of the pulse convolutional neural network is as follows: a convolution module, a pooling layer, and a fully connected module are sequentially provided, and a batch normalization time layer BNTT and an integration-discharge model LIF are inserted after the convolution layer and the fully connected layer in the convolution module and the fully connected module; The input of the fault diagnosis model is the pulse signal, and the output is the fault diagnosis result.

Images