Fault diagnosis method and system for electric drive system based on phase space reconstruction and pulse neural network

Abstract

The application discloses a kind of based on phase space reconstruction and pulse neural network electric drive system fault diagnosis method and system, suitable for the motor system under a variety of complex conditions.The method is reconstructed by phase space to the vibration signal of each channel to obtain the exponential kernel matrix representing system dynamics, then the matrix data after fusion is converted into sparse event stream by pulse coding, finally the pulse convolutional neural network model with the fusion channel attention and time normalization mechanism constructed is trained and diagnosed.The method deeply combines the nonlinear dynamic system features with the pulse neural network time sequence processing capability, fully excavates the space-time dynamic characteristics of vibration signal, effectively solves the problem of insufficient fault feature extraction of traditional method and insufficient classification accuracy of pulse neural network, significantly improves the accuracy and reliability of motor fault diagnosis, and provides a strong guarantee for the safe operation of electric power carrying equipment electric drive system.

Description

Technical Field

[0001] This invention relates to the field of electromechanical equipment condition monitoring and fault diagnosis technology, specifically to a fault diagnosis method and system for electric drive systems based on phase space reconstruction and pulse neural networks. Background Technology

[0002] As a key component of electric transportation equipment such as heavy-duty mining trucks and new energy vehicles, the reliability of the electric drive system motor directly determines the service quality of the entire equipment. Accurate early fault diagnosis of the electric drive system motor is of paramount importance. However, electric drive system motors operate under complex, variable, and even extremely harsh conditions, posing a significant challenge to the diagnosis of early motor faults.

[0003] Existing fault diagnosis methods can be mainly divided into physical model-based methods and data-driven methods. Physical model-based methods require accurate mathematical models of the system, but constructing accurate analytical models is extremely difficult for complex, nonlinear electric drive systems. In recent years, data-driven methods, represented by deep learning, have shown significant advantages, automatically learning fault characteristics and patterns from historical data. However, conventional deep learning models (such as convolutional neural networks) struggle to effectively capture the inherent temporal dynamics and nonlinear evolution of one-dimensional time series signals, such as vibration signals. Although models like recurrent neural networks consider temporal relationships, they still suffer from problems such as complex training and limited ability to capture long-term dependencies.

[0004] Spiking Neural Networks (SNNs), as a third-generation neural network, possess biologically plausible neuron models. They encode and process information through pulse sequences, naturally possessing the ability to handle temporal information and exhibiting event-driven, low-power potential. However, SNNs also have significant drawbacks: their information representation relies on discrete pulse events, leading to insufficient data representation accuracy and low information density. Furthermore, the simplified spiking neuron model has limited dynamic diversity, making it difficult for SNNs to achieve the same accuracy as traditional artificial neural networks in complex classification tasks.

[0005] On the other hand, from the perspective of signal analysis, the vibration signal of a motor is essentially the response of its internal nonlinear dynamic system to external excitation. Traditional time-domain and frequency-domain analysis methods are insufficient to reveal its deep dynamic characteristics. Phase space reconstruction theory provides a powerful tool for analyzing such nonlinear time series, as it can map one-dimensional signals to a high-dimensional space to recover the dynamic characteristics of the original system. However, how to effectively integrate this profound dynamic system feature extraction method with advanced neural network models and overcome the bottlenecks of SNN in terms of accuracy and expressive power remains an unresolved problem in current technology.

[0006] Therefore, developing a novel fault diagnosis scheme that can deeply integrate the inherent characteristics of the power system and fully leverage the advantages of pulse neural network timing processing is of urgent need and significant engineering value for breaking through existing technological bottlenecks and improving the accuracy and reliability of motor fault diagnosis in electric drive systems. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a fault diagnosis method and system for electric drive systems based on phase space reconstruction and spiking neural networks. This method aims to solve the technical problems of insufficient capture of temporal dynamic features by traditional data-driven methods and the limited classification accuracy of spiking neural networks due to information sparsity and model simplification.

[0008] Therefore, the present invention provides the following technical solution:

[0009] On one hand, the present invention provides a fault diagnosis method for an electric drive system based on phase space reconstruction and spiking neural networks, comprising the following steps:

[0010] Step 1: Acquire multi-channel motor vibration signals;

[0011] Step 2: Multi-channel fusion kernel matrix construction. First, the motor vibration signal is reconstructed in multi-channel phase space. Then, based on the point-to-point distance matrix of phase space, the exponential kernel matrix of each channel is constructed and multi-channel weighted fusion is performed to obtain the multi-channel fusion kernel matrix.

[0012] Step 3: Sparse coding, using probabilistic pulse coding to convert the multi-channel fusion kernel matrix into a pulse signal to obtain a sparse event stream;

[0013] Step 4: Input the sparse event stream into the motor fault diagnosis model based on a spiking neural network to obtain the motor diagnosis results;

[0014] The training of the motor fault diagnosis model involves constructing an original dataset and fault labels using multi-channel motor vibration signals under normal and various fault conditions, and then performing supervised learning according to the data processing steps 2-4.

[0015] Preferably, in step 2, the exponential kernel matrix for each channel is constructed based on the point-to-point distance matrix in phase space as follows:

[0016] If the number of reconstructed phase space points is defined as N, then a symmetric matrix of size N*N is set.

[0017] Calculate the point-to-point distance matrix for each channel in the phase space; where the point-to-point distance matrix is ​​composed of the distance values ​​of all point pairs in the channel phase space;

[0018] An exponential transformation is performed on the point-to-point distance matrix to obtain the Gaussian kernel matrix representing the similarity of point trajectories in phase space, i.e., the exponential kernel matrix of the corresponding channel. The corresponding transformation function is:

[0019]

[0020] In the formula, σ is a hyperparameter, and e is the natural base. For phase space points in the channel phase space With phase space points distance, For phase space points With phase space points Gaussian kernel element values.

[0021] Preferably, the process of converting the multi-channel fusion kernel matrix into a pulse signal using probabilistic pulse coding in step 3 to obtain a sparse event stream is as follows:

[0022] The multi-channel fusion kernel matrix is ​​normalized, and the multi-channel fusion kernel matrix M is expressed as: K x K y K z The exponential kernel matrix for channels X, Y, and Z; These are the weighting coefficients corresponding to the X, Y, and Z channels;

[0023] Generate a value in Uniformly distributed random number matrix N is the number of phase space points;

[0024] For each time step t, the encoded sparse event stream is as follows:

[0025] ;

[0026] In the formula, The pulse value corresponding to time step t in a sparse event stream. The element values ​​of the multi-channel fusion kernel matrix M are normalized. A random number matrix The element values, i and j are the row and column labels of the matrix.

[0027] Preferably, a single channel of motor vibration signal is defined. n is the total number of time sampling points. For the data at the 1st, 2nd, and nth time sampling points; the points in the reconstructed phase space. Represented as: s is the point marker in the phase space, τ is the time delay, m is the embedding dimension, and N is the number of reconstructed points, representing the number of points in the phase space;

[0028] The process of multi-channel phase space reconstruction also includes: pre-optimizing the optimal time delay and / or embedding dimension m.

[0029] Preferably, the average mutual information method is introduced to determine the optimal time delay τ for phase space reconstruction, i.e., the average mutual information function is calculated. The average mutual information function When the first local minimum is reached, the corresponding s value is the optimal time delay τ;

[0030] Average mutual information function Defined as:

[0031]

[0032] In the formula, and The marginal probability distribution of signal values ​​throughout the entire sampling sequence. It is the joint probability distribution formed by the two. This represents the data from the i-th and i+s-th time sampling points.

[0033] Preferably, the Cao improved algorithm is introduced to determine the optimal embedding dimension m for phase space reconstruction;

[0034] For different embedding dimensions m, calculate the minimum point-to-neighbor distance between each phase space point and its nearest neighbor; then calculate the ratio of the average point-to-neighbor distance before and after the dimension increase for all points. The average point-to-distance ratio The embedding dimension m corresponding to the point of stabilization is taken as the optimal embedding dimension.

[0035] Preferably, the network architecture of the spiking convolutional neural network is as follows: a convolutional module with an attention mechanism, a pooling layer, and a fully connected classification module, and after the convolutional layer and the fully connected layer in the convolutional module with an attention mechanism and the fully connected classification module, a batch normalized time layer (BNTT) and a spiking neuron layer based on the leak integral-fire model (LIF) are inserted sequentially.

[0036] The processing procedure of the convolutional module with attention mechanism is as follows:

[0037] First, for the input feature map X, global average pooling is used to process each channel in its spatial dimension to obtain the channel description vector;

[0038] Secondly, for each channel description vector, local cross-channel interactions are captured by one-dimensional convolution to determine the attention weights S;

[0039] Finally, the attention weight S is multiplied channel by channel with the feature map X;

[0040] The kernel size k of the one-dimensional convolution is determined by an adaptive function. Confirmed, specifically:

[0041]

[0042] Where γ and b are hyperparameters, This means taking the nearest odd number. C The number of channels in feature map X corresponds to the time step;

[0043] Attention weights corresponding to channel c : , It is the channel description vector corresponding to channel c. The feature obtained by transposing [C,1,1] to [1,C,1] It is the Sigmoid activation function. This represents a one-dimensional convolution operation.

[0044] The technical solution of this invention can further and effectively solve the problem of balancing feature selection and information sparsity in spiking convolutional neural networks by introducing the ECA channel attention mechanism. That is, due to the discrete pulse characteristics and temporal dependence of spiking neural networks, they face the problems of low information density and easy submersion of key features when expressing features.

[0045] Preferably, the fault type includes any combination of eccentric fault, single-phase fault, broken bar fault, bearing fault, and short-circuit fault.

[0046] Secondly, the present invention also provides a system based on the above method, comprising:

[0047] The data acquisition module is used to acquire multi-channel motor vibration signals;

[0048] The multi-channel fusion kernel matrix construction module is used to first reconstruct the multi-channel phase space of the motor vibration signal, and then construct the exponential kernel matrix of each channel based on the point-to-point distance matrix of the phase space and perform multi-channel weighted fusion to obtain the multi-channel fusion kernel matrix.

[0049] The sparse coding module is used to convert the multi-channel fusion kernel matrix into a pulse signal using probabilistic pulse coding to obtain a sparse event stream.

[0050] The diagnostic module is used to input the sparse event stream into a motor fault diagnosis model based on a spiking neural network to obtain motor diagnosis results.

[0051] In three aspects, the present invention provides a computer terminal, which includes at least: one or more processors and a memory storing one or more computer programs;

[0052] The processor invokes a computer program to achieve the following:

[0053] The steps of a fault diagnosis method for electric drive systems based on phase space reconstruction and spiking neural networks.

[0054] In four aspects, the present invention provides a computer-readable storage medium storing a computer program, which is invoked by a processor to implement:

[0055] The steps of a fault diagnosis method for electric drive systems based on phase space reconstruction and spiking neural networks.

[0056] Compared with existing methods, the advantages of the present invention are:

[0057] This invention creatively proposes a fault diagnosis method for electric drive systems based on phase space reconstruction and spiking neural networks. Its core lies in reconstructing one-dimensional vibration signals into phase space, accurately capturing the system's inherent nonlinear dynamic characteristics using an exponential kernel matrix, fusing multi-channel features to obtain a multi-channel fusion kernel matrix, and then using pulse coding to transform it into a sparse event stream. This method deeply integrates dynamic system theory and brain-inspired computing, fully exploring the deep evolutionary laws of vibration signals in the spatiotemporal dimension. It effectively solves the problem that traditional methods cannot deeply characterize and adequately represent the dynamic characteristics of faults under complex operating conditions, thus achieving accurate identification and diagnosis of early weak faults and high-dimensional mixed faults.

[0058] The vibration signal of a motor is essentially the response of its internal nonlinear dynamic system to external excitation, and traditional time-domain or frequency-domain analysis often fails to uncover its deeper underlying mechanisms. The method proposed in this invention maps the one-dimensional motor vibration signal to a higher-dimensional space through phase space reconstruction. Utilizing the average mutual information method and the delay coordinate method, it effectively recovers the topological structure of the motor's internal nonlinear dynamic system, enabling a deeper understanding of the nonlinear evolution within the motor compared to traditional time-frequency analysis. This approach transforms scattered time-domain points into trajectories characterizing the overall dynamic behavior of the system, thereby accurately capturing minute dynamic shifts caused by early, minor faults.

[0059] Furthermore, the technical solution of this invention incorporates an efficient channel attention (ECA) mechanism, aiming to address the challenges of low information density and easy obscuring of key features in pulsed convolutional neural networks (SCNNs) due to the discrete pulse characteristics. By inserting ECA into the convolutional module, the system can capture the deep dependencies between multi-channel kernel matrices after phase space reconstruction, adaptively recalibrate the importance of each channel feature, thereby accurately locking the key signals characterizing motor dynamic anomalies in a sparse event stream and effectively suppressing the interference of background noise and operating condition fluctuations. This mechanism forms a deep technical synergy with the aforementioned phase space reconstruction, Euclidean distance matrix calculation, and weighted fusion. Attached Figure Description

[0060] Figure 1 This is a schematic diagram of the basic process of the method in an embodiment of the present invention.

[0061] Figure 2 This is a schematic diagram of a fault diagnosis method for an electric drive system based on phase space reconstruction and a spiking neural network, according to an embodiment of the present invention.

[0062] Figure 3 This is a flowchart illustrating the training process of the pulsed convolutional neural network according to an embodiment of the present invention. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the application. It should be noted that the reference to "embodiment" herein means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The presentation of this phrase in various locations throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. Those skilled in the art will understand that the embodiments described herein can be combined with other embodiments. The term "and / or" as used herein refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations.

[0064] This invention provides a fault diagnosis method for electric drive systems based on phase space reconstruction and spiking neural networks. It is a novel fault diagnosis paradigm applied to the fault diagnosis of motors in electric drive systems. Its core lies in mapping multi-channel vibration signals into a phase space distance matrix that characterizes the system dynamics, converting it into a sparse event stream through pulse coding, and finally using a spiking convolutional neural network that integrates channel attention and time normalization mechanisms for efficient classification.

[0065] Example 1:

[0066] like Figure 1 As shown, a fault diagnosis method for an electric drive system based on phase space reconstruction and spiking neural networks includes the following steps:

[0067] Step 1: Collect multi-channel motor vibration signals under normal and various fault conditions to form the original dataset. Then, perform sample slicing and labeling on the samples in the original dataset to obtain sample sets under different fault conditions and normal conditions.

[0068] This embodiment collects vibration signals of normal motors and five different motor faults under multi-channel conditions on a motor fault experimental platform. The selected operating conditions include variable speed and variable load conditions. The five types of motor faults defined in this embodiment are eccentricity fault, single-phase fault, broken bar fault, bearing fault, and short circuit fault. Each fault is represented by a triaxial acceleration vibration signal. Preferably, the sequence of each channel is downsampled to significantly reduce the data size while preserving the fault characteristics. Other feasible embodiments are not limited to the above-mentioned fault types and sources of motor fault vibration data.

[0069] The collected vibration signals form a three-dimensional data matrix (number of samples, sequence length, number of channels). In this embodiment, the horizontal x, axial y, and vertical z acceleration signals in the triaxial acceleration vibration signals each correspond to one channel. The vibration signals are sliced ​​into samples, and samples with the same fault are labeled with the same label and randomly shuffled into a sample set.

[0070] Step 2: Perform multi-channel phase space reconstruction and exponential kernel matrix and fusion calculation.

[0071] In this embodiment, the following method is preferably used for multi-channel phase space reconstruction and exponential kernel matrix calculation:

[0072] S2-1: Perform phase space reconstruction independently on each channel signal after downsampling, and upgrade each channel signal from a one-dimensional time series to an m-dimensional dynamic system phase space to capture its intrinsic dynamic characteristics.

[0073] For a dynamic system that meets specific conditions, if the embedding dimension m is greater than or equal to 2d+1 (d is the minimum embedding dimension of the input system, and d is 3 in this embodiment), then the phase space trajectory reconstructed by the delayed coordinate method will maintain the topology of the original system. Define a channel signal time series. n is the total number of time sampling points, and the reconstructed phase space points. Represented as:

[0074] (1)

[0075] Where τ is the time delay, m is the embedding dimension, and N is the number of reconstruction points, expressed as N=n-(m-1)τ. Represents the reconstructed phase space, and is determined by constitute.

[0076] In some embodiments, to determine the optimal time delay τ, this invention employs the average mutual information method. This average mutual information (MI) method finds the delay time τ with the minimum information content by quantifying the nonlinear correlation between the original signal and its delayed version, at which point the statistical independence between the two time points is optimal. (False average mutual information function) Defined as:

[0077] (2)

[0078] In the above formula (2), and The marginal probability distribution of signal values ​​throughout the entire sampling sequence. It is the joint probability distribution formed by the two. In actual calculations, by considering all possible values ​​of s ( Iterate through the data and calculate the corresponding values. According to information theory logic, when When the first local minimum is reached, the corresponding s value is the optimal time delay τ. At this point, the redundant information between sampling points is minimized, and sufficient dynamic correlation is preserved, providing the most representative trajectory features for subsequent phase space reconstruction.

[0079] In some embodiments, when determining the optimal embedding dimension m, the present invention selects the Cao improved algorithm. This algorithm determines the optimal embedding dimension by observing the elimination of "pseudo-nearest neighbor" points during the mapping process from low-dimensional to high-dimensional. First, for each point in the m-dimensional reconstruction space... It is necessary to find its corresponding nearest neighbor. (in Define the minimum point-to-point distance of this point in m-dimensional space. for:

[0080] (3)

[0081] When the embedding dimension increases from m to At that time, using the nearest neighbor index k already determined in m-dimensional space, calculate its index k in m-dimensional space. Distance in 3D space Subsequently, the average point-to-point distance ratio was calculated for all points before and after the dimension increase. Determine the minimum embedding dimension:

[0082] (4)

[0083] In equation (4) above, N represents the total number of reconstructed points. As the embedding dimension m increases, the observed average ratio... The changing trend. When When the value of m approaches saturation, meaning it no longer changes significantly with increasing m, it indicates that the "pseudo-nearest neighbor" points caused by excessively low dimensions have been eliminated in the phase space, and the dynamic characteristics of the original system can be fully recovered. The corresponding m at this point is the optimal minimum embedding dimension. Notably, the nearest neighbor point used from m dimensions to m+1 dimensions is the same.

[0084] In some embodiments, the following further action is taken: determining the chaotic time series of the vibration signal using the maximum Lyapunov exponent. It should be understood that this embodiment only limits the calculation of the maximum Lyapunov exponent λ; specific applications can refer to existing technologies, and this invention does not limit this.

[0085] One of the fundamental characteristics of chaotic systems is their extreme sensitivity to initial values. Trajectories generated by two nearly identical initial values ​​diverge exponentially over time. The maximum Lyapunov exponent is an important quantitative indicator for measuring the dynamic characteristics of a system. It represents the average exponential rate of convergence or divergence between adjacent orbits in phase space. Whether a system exhibits dynamic chaos can be intuitively determined by whether λ (the maximum Lyapunov exponent) is greater than zero. The Rosenstein method is used to obtain the maximum Lyapunov exponent. Formula (3) is used to calculate the nearest neighbor of each reconstructed point, and then the divergence distance over time is calculated for each pair of neighboring points. The mean logarithmic divergence of effective orbital pairs The calculation formula is as follows:

[0086] (5)

[0087] (6)

[0088] In the formula, t satisfies K is the average number of initial neighbor pairs calculated.

[0089] Finally, the maximum Lyapunov exponent λ is obtained through linear regression, and its calculation formula is expressed as covariance. With variance The ratio:

[0090] (7)

[0091] In the formula, cov represents the calculation of covariance, and var represents the calculation of variance.

[0092] S2-2: Exponential Kernel Matrix and Multi-channel Fusion Calculation. The purpose of this step is to construct an N*N distance matrix based on N reconstruction points, where each matrix element represents the distance between two reconstruction points, thus obtaining the distance matrix between all point pairs in the phase space trajectory of each channel. An exponential transformation is then performed to obtain the Gaussian kernel matrix and multi-channel fusion.

[0093] In this embodiment, a block-based computation strategy is preferably adopted to efficiently calculate the Euclidean distance matrix between all point pairs in the phase space trajectory of each channel, and then perform an exponential transformation on it to obtain a Gaussian kernel matrix representing the similarity of trajectory points. This matrix simultaneously contains information about the temporal evolution and spatial structure of the system.

[0094] For a given delay time τ and embedding dimension m, the phase space reconstruction point in Formula 1 is obtained. Because the amount of motor vibration data is large, directly calculating the full distance matrix would lead to memory overflow. Therefore, this embodiment preferably adopts a block-based calculation strategy combined with a three-axis data fusion mechanism to improve diagnostic efficiency. Specifically, the N reconstruction points are divided into K1 sub-blocks of size B (where... Let I and J represent the sets of index points of the p-th and q-th sub-blocks, respectively. Calculate the Euclidean distance between points in the sub-blocks using nested loops to generate local submatrices. any element The calculation formula is:

[0095] (8)

[0096] In formula (8) After obtaining the submatrix Then, it is filled into the corresponding positions in the global distance matrix D. Based on the physical symmetry of the distance matrix, only the calculation is needed. The method divides the data into upper triangular blocks and then symmetrically fills the lower triangular blocks with the results. This approach halves the computational overhead while maintaining the same computational accuracy and effectively controls peak memory usage, significantly improving the real-time performance of the diagnostic system.

[0097] In addition, an exponential transformation is used to construct a "similarity" metric, with the transformation function as follows:

[0098] (9)

[0099] In the above formula, σ is a hyperparameter that controls the range of "similarity". By calculating this Gaussian kernel matrix K, a "similarity map" is generated.

[0100] Motor faults often exhibit spatial coupling. This invention calculates the similarity kernel matrix K for the X, Y, and Z channels of the motor, respectively. x K y K z And perform weighted fusion:

[0101] (10)

[0102] Finally, we get a dimension of The multi-channel fusion kernel matrix M integrates the vibration evolution information of the motor in three-dimensional space, providing extremely rich spatiotemporal feature inputs for subsequent SCNN.

[0103] By pairwise associating the reconstructed dynamic trajectory points, the evolution distance of the system state in phase space is intuitively quantified. Employing a block-based computation strategy and leveraging the symmetry of the distance matrix, only half of the point-to-point distances need to be calculated to complete the full matrix filling, significantly reducing the computational overhead when processing large-scale vibration data and meeting the real-time monitoring requirements of electric drive systems. Furthermore, the distance matrix is ​​transformed into an exponential kernel matrix using an exponential transformation (Gaussian kernel function), which not only enhances the nonlinear mapping but also generates a "similarity map" that simultaneously contains information on temporal evolution and spatial structure, providing high-quality, high-density feature representations for subsequent models.

[0104] Multi-channel weighted fusion can effectively improve the comprehensiveness and reliability of diagnosis. Motor vibration under complex operating conditions often exhibits spatial directionality. By weighted fusion of the exponential kernel matrices of the X, Y, and Z channels, dynamic characteristics from different dimensions can be aggregated, compensating for potential feature gaps from single sensors. The fused multi-channel kernel matrix exhibits stronger robustness compared to single-channel data, effectively shielding against random noise interference and ensuring the system can accurately identify early, subtle faults even under complex and harsh conditions such as varying loads and speeds.

[0105] Step 3: Convert the fused matrix data into pulse signals and divide the sample set into training and test sets;

[0106] To ensure that the extracted dynamic features conform to the processing mechanism of SNN, this invention employs a probabilistic pulse coding method to convert continuous values ​​into pulse sequences. First, the fused similarity matrix M is linearly normalized, mapping its element values ​​to... Interval:

[0107] (11)

[0108] In the above equation (11), The normalized similarity value. and ε and ε' are the minimum and maximum values ​​in the similarity matrix M, respectively, and ε is a minimal constant to prevent the denominator from being zero. Then, within the time window T, a random matrix with the same shape as the normalized matrix is ​​generated. Its elements are in The pulses are uniformly distributed. For each time step t, the pulse firing pattern is determined by the following formula:

[0109] (12)

[0110] This encoding method converts the similarity map into a sparse pulse stream S with temporal attributes. This conversion preserves the similarity features of the phase space trajectories while fully utilizing the low power consumption and high robustness advantages of SNN in processing time-series pulses.

[0111] In this embodiment, the processed sample set is divided into a training set and a test set, and an appropriate batch size is selected for training. The pulsed fault samples are then concentrated, 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.

[0112] Step 4: Construct a fault diagnosis model based on a pulse convolutional neural network, and input the encoded pulse signal into the pulse convolutional neural network for model training;

[0113] This invention introduces an efficient channel attention (ECA) mechanism, aiming to address the challenges of low information density and easily obscured key features in pulsed convolutional neural networks (SCNNs) due to their discrete pulse characteristics. It provides an adaptive feature enhancement method without dimensionality reduction. By inserting ECA into the convolutional module, the system can capture the deep dependencies between multi-channel kernel matrices after phase space reconstruction, adaptively recalibrating the importance of each channel feature. This allows for precise identification of key signals characterizing motor dynamic anomalies in a sparse event stream, effectively suppressing background noise and operational fluctuations. This mechanism forms a deep technical synergy with the aforementioned phase space reconstruction, Euclidean distance matrix calculation, and weighted fusion. Phase space reconstruction transforms one-dimensional vibration signals into high-dimensional dynamic trajectories, which are then mapped to a "similarity map" rich in spatiotemporal information via exponential kernel transformation. The ECA mechanism acts as a "smart filter" in the pulse domain, efficiently identifying the most fault-sensitive dynamic components in the fusion matrix through adaptive one-dimensional convolutional kernels, significantly enhancing the SCNN's ability to express early, weak fault features. This combination not only preserves the dynamic nature of nonlinear systems, but also utilizes the attention weight allocation mechanism of brain-like computing to achieve high-precision and robust diagnosis of motor faults under complex and variable operating conditions.

[0114] like Figure 3 As shown, the network architecture of the pulse convolutional neural network is as follows: it consists of a convolutional module with an attention mechanism, a pooling layer, and a fully connected classification module. Figure 3The system employs a fully connected layer module, and after the convolutional layers in the attention-based convolutional module and the fully connected classification module, a batch normalization through time (BNTT) layer and a spiking neuron layer based on the leaky integral-and-fire (LIF) model are sequentially inserted. The attention-based convolutional module combines efficient channel attention (ECA) with the convolutional layers, enhancing feature representation capabilities through adaptive channel weighting. The batch normalization through time layer maintains independent batch normalization parameters for each time step to adapt to the temporal dynamics of the spiking neural network; this is a prior art and will not be elaborated further. The LIF neurons achieve bio-inspired spatiotemporal information processing through membrane potential integration, threshold comparison, and spiking mechanisms; this is also a prior art and will not be elaborated further.

[0115] The aforementioned spiking convolutional neural network was constructed based on a fusion of spiking neural networks and convolutional neural networks, specifically by combining the LIF model with convolutional layers and fully connected layers, respectively. In each time step loop, membrane potential leakage, pulse firing, and membrane potential reset operations were added after each convolutional layer and fully connected layer.

[0116] It should be understood that the improvement of the pulse convolutional neural network in this invention lies in the inclusion of an attention mechanism convolutional module. The processing of other network layers refers to the prior art, and this invention will not elaborate on this, as can be found in the published patent document: CN119598261A-Method and System for Fault Diagnosis of Motor Spatiotemporal Features in Electric Drive System.

[0117] This invention introduces an ECA (Efficient Channel Attention) mechanism within a convolutional module with an attention mechanism, aiming to efficiently enhance the feature representation capabilities of convolutional spiking neural networks. Its core idea is to adaptively recalibrate the importance of channel features by capturing the dependencies between channels, without requiring complex fully connected layers or dimensionality reduction operations. The specific implementation principle of ECA is as follows:

[0118] For an input feature map X (number of channels C, height H, width W), global average pooling (GAP) is applied to each channel's spatial dimension to obtain the channel description vector:

[0119] (12)

[0120] In the formula, This indicates the feature map X corresponding to channel c. Location characteristics. Channel C refers to the number of channels corresponding to the convolution operation, which is the time step T (time window) used in the last encoding step above.

[0121] Then, for each channel description vector, local cross-channel interactions are captured through one-dimensional convolution to determine the attention weights s;

[0122] Finally, the attention weight S is multiplied with the feature map X channel by channel.

[0123] The kernel size k of the one-dimensional convolution is determined by the adaptive function. Confirmed, specifically:

[0124] (13)

[0125] Where γ and b are hyperparameters, This indicates taking the nearest odd number.

[0126] Attention weights corresponding to channel c The calculation is as follows:

[0127] (14)

[0128] in, It is the channel description vector The feature obtained by transposing [C,1,1] to [1,C,1] It is the Sigmoid activation function.

[0129] Attention weights corresponding to channel c With feature map Features are obtained by multiplying each channel. :

[0130] (15)

[0131] It should be noted that the technical solution of this invention effectively solves the problem of balancing feature selection and information sparsity in spiking convolutional neural networks by introducing the ECA channel attention mechanism. That is, due to the discrete pulse characteristics and temporal dependence of spiking neural networks, they face the problem of low information density and easy submersion of key features when expressing features. Therefore, the feature enhancement methods in traditional convolutional neural networks are difficult to be directly applied to spiking neural networks. In other feasible embodiments, other lightweight attention mechanisms or feature enhancement methods can be adopted to solve this problem.

[0132] Based on the above theoretical statements, this invention uses sample data to train an ECA-enhanced spiking convolutional neural network. The weight parameters of the convolutional layer, ECA attention module, and BNTT layer are updated through temporal backpropagation. Through multiple rounds of iterative optimization, a spiking neural network model for spatiotemporal data classification is obtained. The test set is input into the trained classification model, and the classification accuracy performance index of the test set is obtained based on the accumulation of the output layer membrane potential.

[0133] Step 5: Collect the motor vibration signal of the motor to be monitored and convert it into a pulse signal, then input it into the trained fault diagnosis model for fault diagnosis.

[0134] In summary, the technical solution of this invention fully extracts the spatiotemporal features of the original signal, effectively solving the problems of insufficient local feature extraction and slow processing speed of large-scale data in motor vibration signals. Gaussian population coding is used to encode input features into discrete pulses for communication; neurons are only activated when the received pulse signal reaches a threshold, which is a precise imitation of the biological brain and is also more energy-efficient. Combining convolutional pooling layers with BNTT with SNN and employing an alternative gradient method in backpropagation effectively improves the stability and classification accuracy of the spiking neural network. Furthermore, to verify its performance, comparative experiments were conducted on feature extraction and model training, ultimately finding that it not only achieves high accuracy on two different datasets but also has low computational cost.

[0135] Example 2:

[0136] This embodiment provides a diagnostic system based on the above-described fault diagnosis method, which includes at least: a data acquisition module with mutual communication connection / sequential communication connection, a multi-channel fusion kernel matrix construction module, a sparse coding module, and a diagnostic module.

[0137] The data acquisition module is used to acquire multi-channel motor vibration signals;

[0138] The multi-channel fusion kernel matrix construction module is used to first reconstruct the multi-channel phase space of the motor vibration signal, then construct an exponential kernel matrix based on the point-to-point distance matrix of the phase space and perform multi-channel weighted fusion to obtain the multi-channel fusion kernel matrix;

[0139] The sparse coding module is used to convert the multi-channel fusion kernel matrix into a pulse signal using probabilistic pulse coding to obtain a sparse event stream.

[0140] The diagnostic module is used to input the sparse event stream into a motor fault diagnosis model based on a spiking neural network to obtain motor diagnosis results.

[0141] In some embodiments, the system further includes a model building and training module for building and training a motor fault diagnosis model based on a spiking neural network.

[0142] For the specific implementation process of each module, please refer to the above method content. This invention will not repeat it here. The above division of functional modules is only for illustrative purposes. In some embodiments, some functional modules can be merged and some functional modules can be split. Each functional module can be implemented in software, hardware, or a combination of software and hardware. Among them, software and hardware devices include, but are not limited to, general-purpose computer equipment, programmable gate arrays, digital signal processors, microprocessors and their corresponding programming or burning software.

[0143] Example 3:

[0144] This invention provides a computer terminal, comprising at least one or more processors and a memory storing one or more computer programs; wherein the processor invokes the computer programs to implement the steps of a fault diagnosis method for an electric drive system based on phase space reconstruction and spiking neural networks. Specifically:

[0145] Step 1: Acquire multi-channel motor vibration signals;

[0146] Step 2: Multi-channel fusion kernel matrix construction. First, the motor vibration signal is reconstructed in multi-channel phase space. Then, an exponential kernel matrix is ​​constructed based on the point-to-point distance matrix in phase space and multi-channel weighted fusion is performed to obtain the multi-channel fusion kernel matrix.

[0147] Step 3: Sparse coding. Probabilistic pulse coding is used to convert the multi-channel fusion kernel matrix into a pulse signal to obtain a sparse event stream.

[0148] Step 4: Input the sparse event stream into the motor fault diagnosis model based on the spiking neural network to obtain the motor diagnosis results.

[0149] For details on the implementation of each step, please refer to the explanation of the method described above.

[0150] It should be understood that, in the embodiments of the present invention, the processor may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor. The memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of the memory may also include non-volatile random access memory. For example, the memory may also store device type information.

[0151] Example 4:

[0152] This embodiment provides a computer-readable storage medium storing a computer program. The computer program is called by a processor to: load a pre-trained fault diagnosis model or construct and train a fault diagnosis model; acquire the motor vibration signal of the motor to be monitored and convert it into a pulse signal, then input it into the trained fault diagnosis model for fault diagnosis. Specifically:

[0153] Step 1: Acquire multi-channel motor vibration signals;

[0154] Step 2: Multi-channel fusion kernel matrix construction. First, the motor vibration signal is reconstructed in multi-channel phase space. Then, an exponential kernel matrix is ​​constructed based on the point-to-point distance matrix in phase space and multi-channel weighted fusion is performed to obtain the multi-channel fusion kernel matrix.

[0155] Step 3: Sparse coding. Probabilistic pulse coding is used to convert the multi-channel fusion kernel matrix into a pulse signal to obtain a sparse event stream.

[0156] Step 4: Input the sparse event stream into the motor fault diagnosis model based on the spiking neural network to obtain the motor diagnosis results.

[0157] Please refer to the explanation of the method above for the specific implementation process of each step.

[0158] The computer-readable storage medium can be an internal storage unit of the hardware or software device in any of the foregoing embodiments, such as the hard disk or memory of the controller. The computer-readable storage medium can also be an external storage device of the controller, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the controller. Furthermore, the computer-readable storage medium can include both internal storage units and external storage devices of the controller. The computer-readable storage medium is used to store computer programs and other programs and data required by the controller. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0159] Based on this understanding, the technical solution of the present invention, in essence, or the part 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 several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned readable storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0160] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied 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. This application refers to flowchart illustrations and / or instructions executed by a processor of a method, apparatus (system), and computer program product according to embodiments of this application to create means for implementing the functions specified in one or more flowchart illustrations and / or one or more block diagrams. These computer program instructions may also be stored in a computer-readable storage medium capable of directing a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means that implement the functions specified in one or more flowchart illustrations and / or one or more block diagrams. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, such that the instructions, which execute on the computer or other programmable apparatus, provide steps for implementing the functions specified in one or more flowcharts and / or one or more blocks of a block diagram.

[0161] It should be emphasized that the examples described in this invention are illustrative rather than limiting. Therefore, this 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 solutions of this invention, without departing from the spirit and scope of this invention, whether modifications or substitutions, are also within the protection scope of this invention.

Claims

1. A fault diagnosis method for an electric drive system based on phase space reconstruction and spiking neural networks, characterized in that: Includes the following steps: Step 1: Acquire multi-channel motor vibration signals; Step 2: Multi-channel fusion kernel matrix construction. First, the motor vibration signal is reconstructed in multi-channel phase space. Then, based on the point-to-point distance matrix of phase space, the exponential kernel matrix of each channel is constructed and multi-channel weighted fusion is performed to obtain the multi-channel fusion kernel matrix. Step 3: Sparse coding, using probabilistic pulse coding to convert the multi-channel fusion kernel matrix into a pulse signal to obtain a sparse event stream; Step 4: Input the sparse event stream into the motor fault diagnosis model based on a spiking neural network to obtain the motor diagnosis results; The training of the motor fault diagnosis model involves constructing an original dataset and fault labels using multi-channel motor vibration signals under normal and various fault conditions, and then performing supervised learning according to the data processing steps 2-4.

2. The method according to claim 1, characterized in that: In step 2, the exponential kernel matrix for each channel is constructed based on the point-to-point distance matrix in phase space as follows: If the number of reconstructed phase space points is defined as N, then a symmetric matrix of size N*N is set. Calculate the point-to-point distance matrix for each channel in the phase space; where the point-to-point distance matrix is ​​composed of the distance values ​​of all point pairs in the channel phase space; An exponential transformation is performed on the point-to-point distance matrix to obtain the Gaussian kernel matrix representing the similarity of point trajectories in phase space, i.e., the exponential kernel matrix of the corresponding channel. The corresponding transformation function is: ； In the formula, σ is a hyperparameter, and e is the natural base. For phase space points in the channel phase space With phase space points distance, For phase space points With phase space points Gaussian kernel element values.

3. The method according to claim 1, characterized in that: Step 3 involves using probabilistic pulse coding to convert the multi-channel fusion kernel matrix into a pulse signal to obtain a sparse event stream. The multi-channel fusion kernel matrix is ​​normalized, and the multi-channel fusion kernel matrix M is expressed as: K x K y K z The exponential kernel matrix for channels X, Y, and Z; These are the weighting coefficients corresponding to the X, Y, and Z channels; Generate a value in Uniformly distributed random number matrix N is the number of phase space points; For each time step t, the encoded sparse event stream is as follows: ； In the formula, The pulse value corresponding to time step t in a sparse event stream. The element values ​​of the multi-channel fusion kernel matrix M are normalized. A random number matrix The element values, i and j are the row and column labels of the matrix.

4. The method according to claim 1, characterized in that: Define a channel of motor vibration signal n is the total number of time sampling points. For the 1st, 2nd, and nth time sampling points; reconstruct the phase space points. Represented as: s is the point marker in the phase space, τ is the time delay, m is the embedding dimension, and N is the number of reconstructed points, representing the number of points in the phase space; The process of multi-channel phase space reconstruction also includes: pre-optimizing the optimal time delay and / or embedding dimension m.

5. The method according to claim 4, characterized in that: The average mutual information method is introduced to determine the optimal time delay τ for phase space reconstruction, i.e., to calculate the average mutual information function. The average mutual information function When the first local minimum is reached, the corresponding s value is the optimal time delay τ; Average mutual information function Defined as: ； In the formula, and The marginal probability distribution of signal values ​​throughout the entire sampling sequence. It is the joint probability distribution formed by the two. This represents the data from the i-th and i+s-th time sampling points.

6. The method according to claim 4, characterized in that: Introducing the Cao improved algorithm to determine the optimal embedding dimension m for phase space reconstruction; For different embedding dimensions m, calculate the minimum point-to-neighbor distance between each phase space point and its nearest neighbor; then calculate the ratio of the average point-to-neighbor distance before and after the dimension increase for all points. The average point-to-distance ratio The embedding dimension m corresponding to the point of stabilization is taken as the optimal embedding dimension.

7. The method according to claim 1, characterized in that: The network architecture of the spiking convolutional neural network is as follows: a convolutional module with an attention mechanism, a pooling layer, and a fully connected classification module. After the convolutional layer and the fully connected layer in the convolutional module with an attention mechanism and the fully connected classification module, a batch normalized time layer (BNTT) and a spiking neuron layer based on the leak integral-fire model (LIF) are inserted in sequence. The processing procedure of the convolutional module with attention mechanism is as follows: First, for the input feature map X, global average pooling is used to process each channel in its spatial dimension to obtain the channel description vector; Secondly, for each channel description vector, local cross-channel interactions are captured by one-dimensional convolution to determine the attention weights S; Finally, the attention weight S is multiplied channel by channel with the feature map X; The kernel size k of the one-dimensional convolution is determined by an adaptive function. Confirmed, specifically: ； Where γ and b are hyperparameters, This means taking the nearest odd number. C The number of channels in feature map X corresponds to the time step; Attention weights corresponding to channel c : , It is the channel description vector corresponding to channel c. The feature obtained by transposing [C,1,1] to [1,C,1] It is the Sigmoid activation function. This represents a one-dimensional convolution operation.

8. A system based on the method of any one of claims 1-7, characterized in that: include: The data acquisition module is used to acquire multi-channel motor vibration signals; The multi-channel fusion kernel matrix construction module is used to first reconstruct the multi-channel phase space of the motor vibration signal, and then construct the exponential kernel matrix of each channel based on the point-to-point distance matrix of the phase space and perform multi-channel weighted fusion to obtain the multi-channel fusion kernel matrix. The sparse coding module is used to convert the multi-channel fusion kernel matrix into a pulse signal using probabilistic pulse coding to obtain a sparse event stream. The diagnostic module is used to input the sparse event stream into a motor fault diagnosis model based on a spiking neural network to obtain motor diagnosis results.

9. A computer device, characterized in that: include: One or more processors; A memory that stores one or more computer programs; The processor invokes a computer program to achieve the following: The steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that: The computer program is stored and is invoked by the processor to implement: The steps of the method according to any one of claims 1-7.

Images