Power equipment vibration signal processing method based on interpretable wavelet threshold network

Abstract

The invention discloses a power equipment vibration signal processing method based on an interpretable wavelet threshold network. The method comprises the following steps: acquiring vibration data of key parts of power equipment; inputting the vibration data into a pre-trained fault diagnosis model, and detecting a fault feature matrix of each frequency band channel of the vibration data by using each wavelet threshold network; integrating the fault feature matrixes of different channels by using a channel-frequency dual-domain collaborative attention network, and performing weighted fusion on the fault feature matrixes of different frequency bands to obtain fusion features; using a global energy pooling layer to extract an average energy feature of each frequency band channel from the fusion feature as a discriminative feature; processing the discriminative features by using a fault classification layer, and predicting a fault type classification result; according to the method, noise robustness enhancement and fault sensitive feature depth extraction can be realized, and the fault classification precision and the reliability of a diagnosis result are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of signal processing, and particularly relates to a method for processing vibration signals of power equipment based on an interpretable wavelet threshold network. Background Art

[0002] The operating state of power equipment is crucial for ensuring the safe and stable operation of the power system. During its operation, affected by factors such as load, environment, and aging, faults or abnormal conditions may occur in the equipment, directly affecting the operation stability and safety of the power system. Therefore, real-time monitoring of the state of power equipment and fault diagnosis are important contents for power enterprises to manage power equipment. Traditional monitoring methods and fault diagnosis usually rely on manual inspections and regular maintenance. This method not only has low efficiency but also is difficult to detect potential faults or abnormal conditions.

[0003] Vibration signals are widely used in the research of power equipment fault diagnosis due to their high fault sensitivity and easy measurement. Generally, traditional data-driven intelligent machine learning fault methods mainly include three independent steps: vibration signal acquisition, fault feature extraction, and fault classification. These shallow machine learning methods usually need to use complex professional knowledge and prior knowledge to extract fault-related features, and their diagnostic performance highly depends on the quality of the extracted features, with poor generalization and robustness. In recent years, intelligent diagnosis technologies based on deep learning have received much attention in the academic community. With their flexible and diverse modular structures, powerful feature extraction and classification integrated learning capabilities, compared with traditional machine learning methods, they have better diagnostic performance, stronger scalability, and wider adaptability. In particular, the Convolutional Neural Network (CNN) has stood out among many deep learning methods and achieved remarkable results in various complex tasks. However, it still has the following limitations: (1) The "black box" characteristics of deep learning methods such as CNN seriously hinder their wide application in the industrial field. The low interpretability and opaque mechanism make it difficult for users to understand and trust the diagnostic results; (2) Due to complex mechanical structures, transmission paths, and working conditions, the vibration data collected by sensors are often contaminated by noise, affecting the feature extraction performance of CNN.

[0004] On the one hand, the traditional wavelet transform algorithm has rich mathematical theory and strong interpretability. The extracted fault features have characteristics such as clear physical meaning and robust and reliable prior knowledge. On the other hand, CNN has powerful learning ability. If the advantages of both can be combined and a CNN structure is constructed using wavelet transform, it can broaden the potential optimization direction for enhancing the interpretability of the CNN diagnosis model and improving the performance in the face of complex fault backgrounds. For example, in related technologies, the patent application document with publication number CN116256174A designed a wavelet transform integration layer that can be embedded at any position in the deep learning model to jointly participate in model training and gradient update. By alternately using the wavelet transform integration layer and the convolutional layer for signal decomposition and feature learning, the feature learning ability of the model is increased; however, in this scheme, the wavelet transform set layer and the attention mechanism are alternately used, the process is complex, the calculation amount is large, and the performance optimization is not obvious. Moreover, each layer of wavelet output needs to go through attention calculation, and weight allocation is repeatedly performed on the low-frequency band, resulting in calculation redundancy; the alternating stacking causes the backpropagation paths to intersect, there are gradient conflicts, and the network convergence speed is affected. The patent application document with publication number CN117725465A proposed to build a deep residual shrinkage network, and the core module of this network is the deep residual shrinkage module, which uses deep learning technology to implement soft threshold function noise reduction; while the patent application document with publication number CN116975527A proposed to implement soft threshold denoising based on the ResNet network; however, according to the characteristics of mechanical failures, there are situations where the intra-class differences of the vibration signal frequency band energy distributions of different fault types are small and the inter-class differences are large. The existing soft threshold denoising processes the feature map through the pooling layer, and it is difficult to obtain the optimal threshold. Summary of the Invention

[0005] The technical problem to be solved by the present invention is how to achieve noise reduction and enhancement of the vibration signals of the key components of power equipment under noise interference, and significantly improve the fault classification accuracy and the credibility of the diagnosis results.

[0006] The present invention solves the above technical problems through the following technical means:

[0007] A method for processing vibration signals of power equipment based on an interpretable wavelet threshold network is proposed, and the method includes:

[0008] Obtain the vibration data of the key components of the power equipment;

[0009] Input the vibration data into a pre-trained fault diagnosis model, wherein the fault diagnosis model includes a number of wavelet threshold networks connected in sequence. The number of wavelet threshold networks is the same as the number of discrete wavelet decomposition layers. The output of the last wavelet threshold network is connected to a channel-frequency dual-domain collaborative attention network, a global energy pooling layer, and a fault classification layer in sequence;

[0010] Use each wavelet threshold network to detect the fault feature matrix of each frequency band channel of the vibration data;

[0011] Use the channel-frequency dual-domain collaborative attention network to integrate the fault feature matrices of different channels, and perform weighted fusion on the fault feature matrices of different frequency bands to obtain the fusion features;

[0012] Use the global energy pooling layer to extract the average energy features of each frequency band channel from the fusion features as discriminative features;

[0013] Use the fault classification layer to process the discriminative features and predict the fault type classification result.

[0014] Furthermore, the wavelet threshold network includes a discrete wavelet transform layer, an adaptive threshold residual shrinkage layer, and a convolutional layer stacked in sequence, where:

[0015] Use the discrete wavelet transform layer to perform wavelet decomposition on the vibration data to obtain multiple wavelet coefficient frequency band channels;

[0016] Use the adaptive threshold residual shrinkage layer to adaptively estimate the thresholds of each wavelet coefficient frequency band channel;

[0017] Use the convolutional layer to locate and detect the fault feature matrix of each frequency band channel.

[0018] Furthermore, the discrete wavelet transform layer uses a low-pass filter and a high-pass filter to decompose the vibration data;

[0019] A Dropout layer is connected after the discrete wavelet transform layer.

[0020] Furthermore, the adaptive threshold residual shrinkage layer includes a convolutional layer Conv_1, a batch normalization layer BN_1, and a soft threshold learning branch network. The convolutional layer Conv_1 is connected to the batch normalization layer BN_1. The output of the batch normalization layer BN_1 is respectively connected to the soft threshold function and the soft threshold learning branch network. The output of the soft threshold learning branch network is connected to the soft threshold function. The output features of the soft threshold function are added element-wise to the input features of the convolutional layer Conv_1 to obtain the channel-level adaptive threshold;

[0021] An activation function is also connected after the batch normalization layer BN_1.

[0022] Furthermore, the soft threshold learning branch network includes a global power pooling layer, a fully connected layer FC_1, a batch normalization layer BN_2, and a fully connected layer FC_2 connected in sequence. The output features of the fully connected layer FC_2 are added element-wise to the output features of the global power pooling layer and then used as the input of the soft threshold function;

[0023] After the batch normalization layer BN_2 and the fully connected layer FC_2, an activation function is connected.

[0024] Furthermore, the channel-frequency dual-domain collaborative attention network includes a channel attention module and a frequency attention module, where:

[0025] The channel attention module is used to transpose the received fault feature matrix and multiply it with the fault feature matrix, and then output it to the SoftMax function to obtain a feature matrix that only retains the channel dimension;

[0026] The frequency attention module is used to perform frequency domain transformation and local convolution on the feature matrix that only retains the channel dimension and the fault feature matrix to obtain a fused feature.

[0027] Furthermore, the frequency attention module includes a global average pooling layer GAP, a convolutional layer Conv_2, and a convolutional layer Conv_3 connected in sequence;

[0028] The feature matrix that only retains the channel dimension and the fault feature matrix are reshaped to obtain a feature matrix F as the input of the global average pooling layer GAP. The feature matrix F is element-wise multiplied with the output feature of the convolutional layer Conv_3 and then element-wise added to the feature matrix F to obtain the fused feature;

[0029] After the convolutional layer Conv_2 and the convolutional layer Conv_3, an activation function is connected.

[0030] Furthermore, the formula for the global energy pooling layer to extract the average energy feature of each frequency band channel from the fused feature is expressed as:

[0031]

[0032] In the formula: represents the i-th channel feature map output by the channel-frequency dual-domain serial attention module, N is the length of the feature map, y i is the average energy feature.

[0033] Furthermore, the fault classification layer uses a fully connected layer FC_3, and a Softmax activation function is connected after the fully connected layer FC_3.

[0034] In addition, the present invention also proposes a power equipment vibration signal processing system based on an interpretable wavelet threshold network, including:

[0035] A data acquisition module, which is used to acquire the vibration data of key components of power equipment and output the vibration data to the fault diagnosis module;

[0036] The fault diagnosis module is internally deployed with a pre-trained fault diagnosis model. The fault diagnosis model includes several wavelet threshold networks connected in sequence. The number of wavelet threshold networks is the same as the number of discrete wavelet decomposition layers. The output of the last wavelet threshold network is sequentially connected to a channel-frequency dual-domain collaborative attention network, a global energy pooling layer, and a fault classification layer;

[0037] Each wavelet threshold network is used to detect the fault feature matrix of each frequency band channel of the vibration data;

[0038] The channel-frequency dual-domain collaborative attention network is used to integrate the fault feature matrices of different channels, and perform weighted fusion on the fault feature matrices of different frequency bands to obtain fused features;

[0039] The global energy pooling layer is used to extract the average energy features of each frequency band channel from the fused features as discriminative features;

[0040] The fault classification layer is used to process the discriminative features and predict the fault type classification result.

[0041] The advantages of the present invention are as follows:

[0042] (1) The present invention realizes the decomposition of vibration data through wavelet threshold networks, further strengthens the differential weight fusion of multi-band features through the channel-frequency dual-domain collaborative attention network, and quantifies the frequency band energy mutation characteristics through the global energy pooling layer. Finally, an interpretable wavelet threshold network is constructed to enhance noise robustness and deeply extract fault-sensitive features, significantly improving the fault classification accuracy and the credibility of the diagnosis result.

[0043] (2) The wavelet threshold network realizes the expansion of the network in the depth direction by alternately stacking discrete wavelet layers, adaptive threshold residual shrinkage layers, and convolutional layers multiple times, decomposes the signal into different frequency band channels, and performs multiple discrete wavelet transforms on the signal using multiple groups of the same type of wavelet bases to realize the expansion of the network in the width direction; among them, the discrete wavelet transform layer realizes the decoupling of the signal frequency band, and combines the adaptive threshold residual shrinkage layer to dynamically adjust the frequency band threshold to accurately separate noise and effective features.

[0044] (3) The designed channel-frequency dual-domain collaborative attention network of the present invention significantly improves the robustness and discriminability of vibration signal feature extraction; in the channel dimension, the module uses channel attention to dynamically evaluate the signal-to-noise ratio of each sensor channel, adaptively suppresses the contribution of noise interference channels, and strengthens the weight allocation of high-value signal sources; in the frequency dimension, through frequency domain transformation and local convolution learning, it accurately focuses on the energy distribution of the fault-sensitive frequency band and enhances the time-frequency feature expression of impact components.

[0045] Additional aspects and advantages of the present invention will be given in part in the following description, become apparent in part from the following description, or be learned through the practice of the present invention. Brief Description of the Drawings

[0046] Figure 1 is a schematic flow chart of a method for processing vibration signals of power equipment based on an interpretable wavelet threshold network proposed in an embodiment of the present invention;

[0047] Figure 2 is a schematic structural diagram of a fault diagnosis model in an embodiment of the present invention;

[0048] Figure 3 is a schematic structural diagram of an adaptive threshold residual shrinkage layer in an embodiment of the present invention;

[0049] Figure 4 is a schematic structural diagram of a channel-frequency dual-domain collaborative attention network in an embodiment of the present invention;

[0050] Figure 5 is a schematic flow chart of the training process of a fault diagnosis model in an embodiment of the present invention;

[0051] Figure 6 is a graph of the change of the accuracy curve on the SEU dataset in an embodiment of the present invention;

[0052] Figure 7 is a graph of the change of the accuracy curve after adding noise data to the SEU dataset in an embodiment of the present invention;

[0053] Figure 8 is a schematic structural diagram of a system for processing vibration signals of power equipment based on an interpretable wavelet threshold network proposed in an embodiment of the present invention. Detailed Embodiment

[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0055] As Figure 1 shown, an embodiment of the present invention proposes a method for processing vibration signals of power equipment based on an interpretable wavelet threshold network. The method includes the following steps:

[0056] S10. Obtain vibration data of key components of power equipment;

[0057] It should be noted that in this embodiment, a vibration sensor is used to collect the time-domain vibration signals generated by key components of specific power equipment during operation.

[0058] S20. Input the vibration data into a pre-trained fault diagnosis model, where the fault diagnosis model includes a number of wavelet threshold networks connected in sequence. The number of wavelet threshold networks is the same as the number of discrete wavelet decomposition layers. The output of the last wavelet threshold network is connected to a channel-frequency dual-domain collaborative attention network, a global energy pooling layer, and a fault classification layer in sequence;

[0059] S30. Use each wavelet threshold network to detect the fault feature matrix of each frequency band channel of the vibration data;

[0060] S40. Use the channel-frequency dual-domain collaborative attention network to integrate the fault feature matrices of different channels, and perform weighted fusion on the fault feature matrices of different frequency bands to obtain fusion features;

[0061] S50. Use the global energy pooling layer to extract the average energy features of each frequency band channel from the fusion features as discriminative features;

[0062] S60. Use the fault classification layer to process the discriminative features and predict the fault type classification result.

[0063] In this embodiment, aiming at the core problems faced by the vibration signals of power equipment under complex working conditions, such as large non-stationary noise interference, difficult fault feature extraction, and insufficient model interpretability, it aims to break through the limitations of traditional noise reduction methods and construct an intelligent noise reduction diagnosis framework with both physical interpretability and deep feature mining ability as Figure 2 shown.

[0064] Existing methods rely on fixed-threshold wavelet denoising, which is difficult to adapt to dynamic noise environments, resulting in the suppression of high-frequency weak fault features or the residue of low-frequency noise; at the same time, conventional deep learning models lack explicit modeling of the multi-band energy distribution, and the feature fusion mechanism is single, resulting in low utilization of fault-sensitive information. For this reason, in this embodiment, the decomposition of vibration data is realized through wavelet threshold networks, the differential weight fusion of multi-band features is further strengthened through the channel-frequency dual-domain collaborative attention network, and the mutation characteristics of band energy are quantified through the global energy pooling layer. Finally, an interpretable wavelet threshold network is constructed to enhance noise robustness and deeply extract fault-sensitive features, significantly improving the fault classification accuracy and the credibility of the diagnosis results.

[0065] Compared with the solution described in the patent application document with publication number CN116256174A, in this embodiment, the wavelet transform layer and the convolutional layer are alternately performed, dynamically fusing wavelet multi-scale decomposition and convolutional adaptive learning to realize the mining of complementary time-frequency information features. Finally, the fault feature matrices obtained by each wavelet threshold network are input into the designed channel-frequency dual-domain collaborative attention network, realizing a three-stage architecture of "wavelet decomposition-convolutional extraction-precise focusing", which not only achieves excellent results, but also solves the problems of computational redundancy and gradient conflict, realizes the lightweight of the model, and improves the efficiency.

[0066] As a further preferred technical solution, the wavelet threshold network includes a discrete wavelet transform layer, an adaptive threshold residual shrinkage layer, and a convolutional layer stacked in sequence, where:

[0067] The discrete wavelet transform layer is used to perform wavelet decomposition on the vibration data to obtain multiple wavelet coefficient frequency band channels;

[0068] The adaptive threshold residual shrinkage layer is used to adaptively estimate the thresholds of each wavelet coefficient frequency band channel;

[0069] The convolutional layer is used to locate and detect the fault feature matrices of each frequency band channel.

[0070] It should be noted that in this embodiment, a discrete wavelet transform layer is designed to decompose the vibration signal, and an adaptive threshold residual shrinkage layer is used to simulate the soft threshold denoising process; by alternately stacking the discrete wavelet layer, the adaptive threshold residual shrinkage layer, and the convolutional layer multiple times, the network is extended in the depth direction, and the signal is decomposed into different frequency bands, where the number of discrete wavelet layers is the number of discrete wavelet decomposition layers. In addition, similar to using multiple filters of the same size in the traditional CNN convolutional layer to extract multi-channel features of the input, multiple groups of wavelet bases of the same type are used to perform multiple discrete wavelet transforms on the signal to realize the expansion of the network in the width direction.

[0071] Further, the discrete wavelet transform layer uses a low-pass filter and a high-pass filter to decompose the vibration data; a Dropout layer is connected after the discrete wavelet transform layer.

[0072] Specifically, the discrete wavelet layer proposed in this embodiment is used to replace the original CNN hidden layer to autonomously learn the fault signal features, so that the vibration signal decomposes signals in different frequency bands at this layer, improving the model's learning ability for fault feature information. At the same time, this layer can enable the convolutional layer to locate and detect the fault information of each frequency band, making the model have good interpretability.

[0073] The discrete wavelet layer uses a low-pass filter h(n) and a high-pass filter g(n) to decompose the signal respectively. Based on the given wavelet basis function, a low-pass filter h(n) can be obtained, and g(n) is determined by h(n). Each channel in the input feature map (where M is the length of the feature map) is output after being filtered by a single discrete wavelet layer h(n) and g(n)

[0074] g(n) = (-1) n h(-n)

[0075] F = concat(h, g)

[0076] In the formula: h and g are the low-frequency approximation coefficient feature map and the high-frequency detail coefficient feature map respectively, and concat(·) represents feature concatenation along the channel dimension.

[0077] It should be noted that the low-pass and high-pass filter parameters in the discrete wavelet transform are globally fine-tuned and optimized through the gradient descent backpropagation algorithm.

[0078] Furthermore, considering that the Dropout layer is an effective method for preventing overfitting, and at the same time, due to the existence of the discrete wavelet layer, the model is more likely to obtain features from the signal, resulting in overfitting. Therefore, the method of using the Dropout layer after the discrete wavelet layer well solves this problem and further improves the generalization ability of the model.

[0079] It should be noted that deeply embedding the wavelet transform into the convolutional neural network combines the multi-scale frequency decomposition of the wavelet transform and the excellent adaptive feature learning ability of the convolutional layer. When deeply mining the time-frequency information alternately, the wavelet transform containing rich physical prior knowledge endows the convolutional neural network structure with interpretability. The two learn and promote each other, bringing an improvement in diagnostic performance in the collision and fusion of traditional methods and intelligent methods.

[0080] As a further preferred technical solution, the discrete wavelet transform layer designed in this embodiment is used to decompose the vibration signal, and then the adaptive threshold residual shrinkage layer is used to simulate the soft threshold denoising process. Wavelet soft threshold denoising is a classic method for denoising vibration signals. By thresholding the decomposition coefficients, near-zero features are converted to zero to reduce noise. The soft threshold process is as follows:

[0081]

[0082] In the formula: s, represents the magnitudes of the wavelet coefficients before and after applying the threshold, and λ is a non-negative threshold.

[0083] It should be noted that the key to the wavelet soft threshold denoising task is the setting of the threshold λ: if λ is too small, there will be more residual noise; if λ is too large, it is easy to misclassify weak and useful fault features as noise and remove them. Inspired by the channel attention mechanism, the ResidualShrinkage Building Unit (RSBU) embeds the threshold in the network structure and automatically determines it through model training, avoiding the cumbersome manual setting. Moreover, the RSBU dynamically generates channel-level adaptive thresholds through the Squeeze-and-Excitation (SE) module, replacing the traditional fixed threshold or global threshold method. Its core purpose is to adaptively adjust the soft threshold parameters according to the noise distribution characteristics of different channels in the input feature map, achieving more refined noise suppression and feature retention.

[0084] Therefore, in this embodiment, the soft threshold process is constructed as a soft threshold denoising activation layer, which is embedded behind each stationary wavelet packet convolutional layer to adaptively estimate the threshold of each decomposition coefficient in each layer and achieve the screening of noise or redundant information. Figure 3 The basic structure of the adaptive threshold residual shrinkage layer is shown:

[0085] The adaptive threshold residual shrinkage layer includes a convolutional layer Conv_1, a batch normalization layer BN_1, and a soft threshold learning branch network. The convolutional layer Conv_1 is connected to the batch normalization layer BN_1. The output of the batch normalization layer BN_1 is respectively connected to a soft threshold function and the soft threshold learning branch network. The output of the soft threshold learning branch network is connected to the soft threshold function. The output features of the soft threshold function are added element-wise to the input features of the convolutional layer Conv_1 to obtain channel-level adaptive thresholds;

[0086] An activation function is also connected after the batch normalization layer BN_1.

[0087] As a further preferred technical solution, the soft threshold learning branch network includes a global power pooling layer, a fully connected layer FC_1, a batch normalization layer BN_2, and a fully connected layer FC_2 connected in sequence. The output features of the fully connected layer FC_2 are added element-wise to the output features of the global power pooling layer and then used as the input of the soft threshold function;

[0088] Activation functions are connected after both the batch normalization layer BN_2 and the fully connected layer FC_2.

[0089] It should be noted that in view of the problem that it is difficult to determine the threshold in soft threshold denoising, this embodiment designs a soft threshold denoising activation layer based on the SE module and the residual block, realizes the adaptive estimation of the threshold, reduces the noise interference information, and effectively improves the noise robustness of the diagnostic model. According to the characteristics of mechanical failures, there are small intra-class differences and large inter-class differences in the band energy distribution of vibration signals of different fault types. Therefore, this embodiment adopts a global power pooling layer to extract the average energy information of the band channels as the basis for selecting the threshold.

[0090] Moreover, this embodiment takes into account that the amplitude of the vibration signal may not be strictly symmetric. Therefore, each wavelet band channel is assigned two trainable and learnable positive and negative thresholds λ - and λ + , regards the threshold as a trainable parameter and embeds it into the network structure, and self-learns the optimal threshold from the data through the gradient descent backpropagation algorithm.

[0091] As a further preferred technical solution, the channel-frequency dual-domain collaborative attention network includes a channel attention module and a frequency attention module, wherein:

[0092] The channel attention module is used to perform a transpose operation on the received fault feature matrix and multiply it with the fault feature matrix, and then output it to the SoftMax function to obtain a feature matrix that only retains the channel dimension;

[0093] The frequency attention module is used to perform a frequency domain transformation and local convolution on the feature matrix that only retains the channel dimension and the fault feature matrix to obtain a fused feature.

[0094] Furthermore, as Figure 4 shown, the channel attention module plays a key role in integrating the information differences between multiple information sources. By deeply comparing and analyzing the information flows from different sources, it can accurately extract the complementary information and unique contributions between channels, and then realize the comprehensive integration and optimization of information. In order to focus more on the channel dimension and reduce the structural complexity, this module directly performs a transpose operation on the feature matrix (C, H), multiplies it with the original matrix, and then outputs a feature matrix of (C, C) through the SoftMax function. This step effectively removes the position dimension H and retains the information of the channel dimension C. Through this design, the channel attention module not only improves the efficiency and accuracy of information processing, but also enhances the generalization ability and robustness of the model.

[0095] Furthermore, as Figure 4 shown, the frequency attention module includes a global average pooling layer GAP, a convolutional layer Conv_2, and a convolutional layer Conv_3 connected in sequence;

[0096] The feature matrix that only retains the channel dimension and the fault feature matrix are reshaped to obtain the feature matrix F as the input of the global average pooling layer GAP. The feature matrix F is multiplied element-wise with the output features of the convolutional layer Conv_3 and then added element-wise to the feature matrix F to obtain the fused feature;

[0097] An activation function is connected after both the convolutional layer Conv_2 and the convolutional layer Conv_3.

[0098] Specifically, the channel-frequency dual-domain collaborative attention network designed in this embodiment adopts a serial running architecture. First, the information of multiple channels is integrated, and then the frequency attention mechanism is used to filter out the signal components useful for the diagnosis task. Compared with the attention module designed in the prior art, such as the patent application document with publication number CN117694902A, the channel-frequency dual-domain collaborative attention network designed in this embodiment has a smaller computational amount, realizes lightweight, and first discards some features through the passband attention mechanism, improves sparsity, and enhances the generalization of the model.

[0099] Among them, the core idea of the frequency attention module is to use the self-learning ability of CNN to filter out the signal components useful for the diagnosis task from the input feature matrix F. It replaces manual feature selection and the automatic valuable feature attention mechanism is similar to human visual attention. First, the frequency attention module introduces a global average pooling layer (GAP_2) to compress the global information of the input feature matrix F. Then, the frequency attention module uses a simple encoding and decoding mechanism to capture the importance of these channel signals. The encoding and decoding operations of the frequency attention module are completed by two convolutional layers respectively. The first convolutional layer Conv_2 uses the ReLU function to provide non-linear transformation capabilities. The second convolutional layer Conv_3 uses the Sigmoid function, which is mainly used to map the obtained feature vector to the range of 0 to 1, so as to generate the weight vector ω.

[0100] The formula for the frequency attention module to calculate the fused feature is expressed as:

[0101]

[0102] In the formula: represents element-wise multiplication, Conv is one-dimensional convolution, is the output fused feature.

[0103] It should be noted that the designed channel-frequency dual-domain collaborative attention module is used to integrate the feature information of different channels to enhance the expression ability of the feature map. By deeply mining the time-frequency information in each channel through this module, the correlation between multiple channels is fully considered, realizing the in-depth integration and efficient utilization of information. By serially running the channel attention module and the frequency attention module and outputting the results, the obtained result is regarded as the channel-frequency dual-domain collaborative attention module. The designed channel-frequency dual-domain collaborative attention mechanism significantly improves the robustness and discriminability of vibration signal feature extraction. In the channel dimension, the module dynamically evaluates the signal-to-noise ratio of each sensor channel using channel attention, adaptively suppresses the contribution of noise interference channels, and strengthens the weight allocation of high-value signal sources; in the frequency dimension, through frequency domain transformation and local convolution learning, it accurately focuses on the energy distribution of the fault-sensitive frequency band and enhances the time-frequency feature expression of impact components. The dual-domain attention realizes the progressive optimization of "channel cleaning → frequency band enhancement" in a serial manner, not only solving the balance problem of noise suppression and feature retention in traditional methods, but also the physical interpretability of the module provides key support for reliable decision-making in industrial scenarios.

[0104] In view of the problem of unclear characteristic signals in this embodiment, a channel-frequency dual-domain collaborative attention module is designed to integrate the information between multiple channels for complementarity and focus on the frequency band where the fault characteristics are located for weighting, not only obtaining more comprehensive information, but also enhancing the characteristics of vibration signals.

[0105] As a further preferred technical solution, the formula for the global energy pooling layer to extract the average energy feature of each frequency band channel from the fused features is:

[0106]

[0107] In the formula: represents the i-th channel feature map output by the channel-frequency dual-domain serial attention module, N is the length of the feature map, and y i is the average energy feature.

[0108] It should be noted that the network decomposes the original input signal into multiple wavelet coefficient frequency band channels through continuous stacking of the stationary wavelet packet convolutional layer and the soft threshold denoising activation layer. In order to further extract discriminative features helpful for fault identification from each frequency band, considering that the wavelet packet band energy is a common fault statistical index in traditional signal processing, a global energy pooling layer is designed to extract the average energy feature of each frequency band channel for simulating the mining of wavelet band energy information.

[0109] As a further preferred technical solution, the fault classification layer adopts the fully connected layer FC_3, and a Softmax activation function is connected after the fully connected layer FC_3.

[0110] It should be noted that this embodiment uses a fully connected layer and a Softmax function to classify fault types, and passes the flat vector obtained by the global energy pooling layer through global compression operation to the classifier network. The output probability q corresponding to the kth node category k for:

[0111]

[0112] Where: K is the total number of categories, z k is the output value of the kth node in the fully connected layer, W k 、b k are the weight and bias corresponding to the kth node in the fully connected layer respectively.

[0113] It should be noted that this embodiment specifically sets up 6 layers of alternating discrete wavelet layers, adaptive threshold residual shrinkage layers, and convolutional layers to deeply mine the useful time-frequency fault features hidden in the input signal, and has powerful feature extraction capabilities; then, the channel-frequency dual-domain collaborative attention module is used to integrate information between multiple channels, and focus on the energy distribution of fault-sensitive frequency bands to enhance the expression of fault features; finally, the fully connected layer and Softmax function are used to classify the fault type.

[0114] Furthermore, if Figure 5 As shown, the training process of the fault diagnosis model adopted in this embodiment is:

[0115] (1) Dataset acquisition

[0116] Depending on the background tasks of fault diagnosis, vibration sensors are used to collect time-domain vibration data generated by key components of specific power equipment during operation to obtain vibration data in different health states. By accumulating data over a period of time, the adequacy of fault data is ensured, and the data is classified and labeled according to the corresponding health status of the components.

[0117] (2) Data preprocessing and division:

[0118] The collected vibration data is truncated into multiple fixed time step sequences using sliding window segmentation enhancement technology to expand the sample size. Then, all samples are used to construct fault training and test sets according to a certain ratio. To improve the convergence speed of the network during training, each sample is normalized by Z-score, calculated as follows:

[0119]

[0120] Where: They are samples x i The mean and variance of is the sample x iThe normalized result, where N is the sample length.

[0121] In addition, when truncating the samples, set the data length of each sample to be greater than the number of sampled data points generated by the rotating component in one revolution to include as much complete fault impact information as possible.

[0122] (3) Model training:

[0123] Set hyperparameters such as the number of iterations, batch size, and learning rate. Input the training set into the fault diagnosis model in small batches for training. Calculate the cross-entropy loss function value through forward propagation, and then use the error backpropagation algorithm to optimize the model parameters. After multiple rounds of training, save the network model with the optimal parameters.

[0124] According to Figure 2 the constructed fault diagnosis network shown, initialize the discrete wavelet layer with the filter bank corresponding to the discrete wavelet basis db16, and use random initialization for the other layers. The model uses the Adam adaptive optimizer to optimize the parameters. At the same time, set reasonable hyperparameters such as the number of training iterations, learning rate, and batch training size. Input the training set samples into the diagnostic network in small batches. After obtaining the predicted values through forward propagation, calculate the cross-entropy loss function value by comparing with the actual label values of the samples. The cross-entropy loss function is defined as follows:

[0125]

[0126] In the formula: p k ∈{0,1} represents the true label of the sample in the k-th category, and q k represents the predicted probability that the sample belongs to the k-th category. After calculating the cross-entropy error, use the gradient descent backpropagation algorithm to tune the network parameters. After multiple iterations, the network is fully trained, and save the optimal interpretable wavelet threshold network model parameters after reaching the preset total number of iterations.

[0127] (4) Model verification and application:

[0128] Input the test set samples into the trained fault diagnosis model, only perform the forward propagation calculation process, predict and output the fault status results of the unknown category samples, and verify the effectiveness of the model through means such as the change curve graphs of the accuracy rate, average accuracy rate, and confusion matrix during model training and testing. After the test results meet the expected effect, the fault diagnosis model can be further deployed in the practical application of fault diagnosis for key components of power equipment.

[0129] Furthermore, to further illustrate the present invention, it will be described in detail below with examples:

[0130] The hardware platform for the method network model experiment in this embodiment is configured with a 13th Gen Intel(R) Core(TM) i9-13900HX processor and an NVIDIA GeForce RTX 4060 Laptop GPU graphics card; the software platform is configured with the Python 3.9 programming language and the Pytorch 1.9.1 deep learning framework.

[0131] Taking the gearbox, a key component of power equipment, as an example, the method in this embodiment was verified for effectiveness on the Southeast University (SEU) gearbox dataset. This dataset includes a bearing dataset and a gear dataset collected from a power transmission simulation test bench. In the experiment, 9 kinds of state vibration data measured in the x-axis direction of the planetary gearbox under the speed-load configuration of 20Hz - 0V were selected, with a sampling frequency of 5120Hz. The detailed description is shown in Table 1.

[0132] Table 1 Description of the Southeast University Gearbox Dataset

[0133]

[0134] The total number of samples in the dataset is 2500 for each. Each state category includes 250 data samples, the sample length is 1024, the ratio of the training set to the test set is 7:3, and the test accuracy is the average of 5 repeated experiments. Figure 6 Figure 1 shows the accuracy curve change diagram of the method in this embodiment on the Southeast University dataset, where the proportion of the training set samples is 30%. From Figure 6 it can be seen that at the initial stage of iteration, the accuracy of the model on the training set and the test set increases rapidly. After a period of training, the accuracy curve and the loss curve gradually tend to be stable, further verifying the diagnostic performance of the model.

[0135] To simulate the noise interference of rotating machinery in a real industrial scenario, under the condition that the proportion of the training set is 30%, Gaussian white noise is added to the original samples to obtain noise data at different signal-to-noise ratios. The experimental results are shown in Table 2. It can be found that the classification accuracy of the model gradually increases as the noise intensity decreases (i.e., the higher the signal-to-noise ratio). The average accuracy is still above 80% in the case of strong noise with a signal-to-noise ratio of -4dB. In addition, from Figure 6 the accuracy curve change diagram, it can be seen that the proposed model maintains a stable training process under noise interference, verifying the anti-noise performance of the model.

[0136] Table 2 Diagnostic Accuracy at Different Signal-to-Noise Ratios

[0137]

[0138] In addition, as Figure 8As shown in the figure, another embodiment of the present invention also proposes a power equipment vibration signal processing system based on an interpretable wavelet threshold network, including:

[0139] A data acquisition module 10, configured to acquire vibration data of key components of a power equipment, and output the vibration data to a fault diagnosis module 20;

[0140] A pre-trained fault diagnosis model is deployed inside the fault diagnosis module 20. The fault diagnosis model includes a number of wavelet threshold networks connected in sequence. The number of wavelet threshold networks is the same as the number of discrete wavelet decomposition layers. The output of the last wavelet threshold network is connected to a channel-frequency dual-domain collaborative attention network, a global energy pooling layer, and a fault classification layer in sequence;

[0141] Each wavelet threshold network is used to detect the fault feature matrix of each frequency band channel of the vibration data;

[0142] The channel-frequency dual-domain collaborative attention network is used to integrate the fault feature matrices of different channels, and perform weighted fusion on the fault feature matrices of different frequency bands to obtain a fusion feature;

[0143] The global energy pooling layer is used to extract the average energy feature of each frequency band channel from the fusion feature as a discriminative feature;

[0144] The fault classification layer is used to process the discriminative feature and predict the fault type classification result.

[0145] It should be noted that for other embodiments or specific diagnosis implementation methods of the fault diagnosis model adopted in the power equipment vibration signal processing system based on the interpretable wavelet threshold network of the present invention, reference can be made to the above method embodiments, and details are not described herein again.

[0146] Note that the logic and / or steps represented in the flowchart or described otherwise herein, for example, can be considered as a definite sequence list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can fetch and execute instructions from the instruction execution system, apparatus, or device), or used in combination with these instruction execution systems, apparatuses, or devices. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of the computer-readable medium include the following: an electrical connection part with one or more wirings (electronic device), a portable computer disk cartridge (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpreting, or otherwise processing it as appropriate, and then storing it in a computer memory.

[0147] It should be understood that various parts of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application specific integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), etc.

[0148] In the description of this specification, the description referring to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples", etc. means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0149] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0150] Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for processing vibration signals of power equipment based on an interpretable wavelet threshold network, characterized in that Including: Obtaining vibration data of key components of power equipment; Inputting the vibration data into a pre-trained fault diagnosis model, where the fault diagnosis model includes several wavelet threshold networks connected in sequence, the number of wavelet threshold networks is the same as the number of discrete wavelet decomposition layers, and the output of the last wavelet threshold network is sequentially connected to a channel-frequency dual-domain collaborative attention network, a global energy pooling layer, and a fault classification layer; Detecting the fault feature matrix of each frequency band channel of the vibration data by using each wavelet threshold network; Integrating the fault feature matrices of different channels by using the channel-frequency dual-domain collaborative attention network, and performing weighted fusion on the fault feature matrices of different frequency bands to obtain fusion features; Extracting the average energy feature of each frequency band channel from the fusion features as discriminative features by using the global energy pooling layer; Processing the discriminative features by using the fault classification layer to predict the fault type classification result.

2. The method for processing vibration signals of power equipment based on an interpretable wavelet threshold network according to claim 1, characterized in that, The wavelet threshold network includes a discrete wavelet transform layer, an adaptive threshold residual shrinkage layer, and a convolutional layer stacked in sequence, where: Performing wavelet decomposition on the vibration data by using the discrete wavelet transform layer to obtain multiple wavelet coefficient frequency band channels; Adaptive estimating the threshold of each wavelet coefficient frequency band channel by using the adaptive threshold residual shrinkage layer; Locating and detecting the fault feature matrix of each frequency band channel by using the convolutional layer.

3. The method for processing vibration signals of power equipment based on an interpretable wavelet threshold network according to claim 2, wherein The discrete wavelet transform layer decomposes the vibration data by using a low-pass filter and a high-pass filter; A Dropout layer is connected after the discrete wavelet transform layer.

4. The method for processing vibration signals of power equipment based on an interpretable wavelet threshold network according to claim 2, wherein, The adaptive threshold residual shrinkage layer includes a convolutional layer Conv_1, a batch normalization layer BN_1, and a soft threshold learning branch network. The convolutional layer Conv_1 is connected to the batch normalization layer BN_1. The output of the batch normalization layer BN_1 is respectively connected to a soft threshold function and the soft threshold learning branch network. The output of the soft threshold learning branch network is connected to the soft threshold function. The output feature of the soft threshold function is added element-wise to the input feature of the convolutional layer Conv_1 to obtain a channel-level adaptive threshold; An activation function is also connected after the batch normalization layer BN_1.

5. The method for processing vibration signals of power equipment based on an interpretable wavelet threshold network according to claim 4, characterized in that, The soft threshold learning branch network includes a global power pooling layer, a fully connected layer FC_1, a batch normalization layer BN_2, and a fully connected layer FC_2 connected in sequence. The output feature of the fully connected layer FC_2 is added element-wise to the output feature of the global power pooling layer and used as the input of the soft threshold function; Activation functions are connected after both the batch normalization layer BN_2 and the fully connected layer FC_2.

6. The method for processing vibration signals of power equipment based on an interpretable wavelet threshold network according to claim 1, characterized in that The channel-frequency dual-domain collaborative attention network includes a channel attention module and a frequency attention module, where: The channel attention module is used to perform a transpose operation on the received fault feature matrix, multiply it with the fault feature matrix, and then output it to the SoftMax function to obtain a feature matrix that only retains the channel dimension; The frequency attention module is used to perform frequency domain transformation and local convolution on the feature matrix that only retains the channel dimension and the fault feature matrix to obtain fusion features.

7. The method for processing vibration signals of power equipment based on an interpretable wavelet threshold network according to claim 6, characterized in that, The frequency attention module includes a global average pooling layer GAP, a convolutional layer Conv_2, and a convolutional layer Conv_3 connected in sequence; The feature matrix that only retains the channel dimension and the fault feature matrix are reshaped to obtain the feature matrix F as the input of the global average pooling layer GAP. The feature matrix F is multiplied element-wise with the output features of the convolutional layer Conv_3 and then added element-wise to the feature matrix F to obtain the fused feature; An activation function is connected after both the convolutional layer Conv_2 and the convolutional layer Conv_3.

8. The method for processing vibration signals of power equipment based on an interpretable wavelet threshold network according to claim 1, wherein, The formula for the global energy pooling layer to extract the average energy features of each frequency band channel from the fused feature is expressed as: In the formula: represents the i-th channel feature map output by the channel-frequency dual-domain serial attention module, N is the length of the feature map, and y i is the average energy feature.

9. The method for processing vibration signals of power equipment based on an interpretable wavelet threshold network according to any one of claims 1-8, characterized in that, The fault classification layer uses the fully connected layer FC_3, and a Softmax activation function is connected after the fully connected layer FC_3.

10. A power equipment vibration signal processing system based on an interpretable wavelet threshold network, characterized in that Including: A data acquisition module for acquiring vibration data of key components of power equipment and outputting the vibration data to the fault diagnosis module; A pre-trained fault diagnosis model is deployed inside the fault diagnosis module. The fault diagnosis model includes a number of wavelet threshold networks connected in sequence. The number of wavelet threshold networks is the same as the number of discrete wavelet decomposition layers. The output of the last wavelet threshold network is connected to a channel-frequency dual-domain collaborative attention network, a global energy pooling layer, and a fault classification layer in sequence; Each wavelet threshold network is used to detect the fault feature matrix of each frequency band channel of the vibration data; The channel-frequency dual-domain collaborative attention network is used to integrate the fault feature matrices of different channels and perform weighted fusion on the fault feature matrices of different frequency bands to obtain the fused feature; The global energy pooling layer is used to extract the average energy features of each frequency band channel from the fused feature as discriminative features; The fault classification layer is used to process the discriminative features and predict the fault type classification result.

Images