Bearing fault diagnosis method and system based on multi-source fusion deep adversarial network

Abstract

The invention discloses a bearing fault diagnosis method and system based on a multi-source fusion deep adversarial network, and belongs to the technical field of fault diagnosis. According to the method, an improved normalized Hilbert-Huang transform time-frequency spectrum feature extraction model is introduced to capture local features of a fault data time-frequency domain, and pooling operation is adopted to compress two-dimensional time-frequency image features in fault data time-frequency information so as to highlight important features. And a splicing operation is adopted to obtain a multi-source signal fusion feature result so as to capture multi-scale fault features. An improved Wasserstein generative adversarial network model is adopted to obtain data distribution results of real samples of different fault categories, so that a comprehensive data distribution model can be constructed, and the generalization ability of a fault diagnosis system can be improved; and an enhanced deep residual shrinkage network model is adopted to solve the gradient disappearance problem in the deep network, and the shrinkage network is adopted to improve the sparsity of the weight of the deep network to improve the generalization of the model.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a bearing fault diagnosis method and system based on a multi-source fusion deep adversarial network. Background Art

[0002] In modern industrial production, rolling bearings are the core components of many mechanical equipment, and their operating status is directly related to the stability and reliability of the entire equipment. Once a bearing fails, it may lead to reduced equipment efficiency and production interruption, or even cause serious safety accidents, resulting in huge economic losses and casualties. Therefore, it is particularly important to diagnose rolling bearing faults.

[0003] In the fault detection of rolling bearings, traditional methods mainly rely on signal processing technologies such as spectrum analysis, time domain analysis and time-frequency analysis. These methods usually judge the health status of bearings by extracting the characteristics of vibration signals, such as root mean square value, peak factor, kurtosis, etc. In addition, methods such as Fourier transform, short-time Fourier transform and wavelet transform are often used to analyze the frequency domain and time-frequency domain characteristics of vibration signals. However, traditional methods rely on expert experience and manually designed feature extraction processes, which leads to many limitations in practical applications. For example, different working conditions and fault types may have a significant impact on the feature extraction results, resulting in unstable detection results; and manually designed features are difficult to fully capture the complex information in the signal, especially when facing equipment such as rolling bearings with complex working conditions and changeable operating environments. In addition, traditional methods are usually highly dependent on data and are difficult to adapt to large amounts of real-time data processing needs.

[0004] With the development of artificial intelligence and computer technology, deep learning has gradually become a research hotspot in the field of fault detection. Deep learning can automatically learn feature representation from data by constructing multi-layer neural networks, avoiding the complexity of manually designed features. In particular, deep learning has shown great potential in rolling bearing fault detection. Convolutional neural networks are an important architecture in deep learning and are widely used in image processing and signal processing. In rolling bearing fault detection, convolutional neural networks can directly process time-frequency images or vibration signals, extract local features in the signal through the convolution layer, and perform feature compression through the pooling layer to retain the most representative fault information. Studies have shown that fault detection methods based on convolutional neural networks can effectively identify various types of rolling bearing faults, including inner ring faults, outer ring faults, and rolling element faults. In addition, the hierarchical feature extraction mechanism of convolutional neural networks enables it to capture multi-scale features in vibration signals, thereby improving the accuracy of fault detection. Compared with traditional methods, convolutional neural networks can not only automatically learn more discriminative features from raw data, but also adapt to different working conditions and complex environments, showing high robustness. The wide application of deep learning technology relies on a large amount of fault data. In rolling bearing fault detection, data-driven methods use a large amount of normal operation data and fault data to train models, so that deep learning models can accurately identify various fault modes. However, due to the low probability of failure and the imbalance of actual data, it is challenging to obtain rolling bearing fault data. In order to solve the problem of insufficient data, researchers usually use data enhancement technology to expand the size of the data set, such as generating more training data by adding noise, data translation, mirroring, etc.

[0005] Therefore, due to the complex working conditions of rolling bearing failures and insufficient data, the existing traditional methods and artificial intelligence methods urgently need a new rolling bearing fault diagnosis method to further improve the accuracy of fault diagnosis. To this end, the present invention proposes a bearing fault diagnosis method and system based on a multi-source fusion deep adversarial network. Summary of the invention

[0006] The technical problem to be solved by the present invention is: how to further improve the accuracy of rolling bearing fault diagnosis. A bearing fault diagnosis method based on a multi-source fusion deep adversarial network is provided, and a fault diagnosis method based on a multi-source fusion deep adversarial network is introduced, so that a better fault identification effect can be obtained in the field of rolling bearing fault diagnosis, and the diagnostic performance is greatly improved, which can improve the accuracy of fault detection, reduce maintenance costs, and improve operational safety and reliability.

[0007] The present invention solves the above technical problems through the following technical solutions, and the present invention comprises the following steps:

[0008] Step S1: Data collection and preprocessing

[0009] Acquire and pre-process the data measured by the sensors of the rolling bearings to construct the multi-source signal data of the faults;

[0010] Step S2: Multi-source signal feature fusion

[0011] The fault multi-source signal data is input into the improved normalized Hilbert-Huang transform time-frequency feature extraction model to obtain the time-frequency information of the fault data; the time-frequency information of the fault data is subjected to pooling operation and compression, dimensionality reduction and splicing and fusion operations to obtain the multi-source signal fusion feature result;

[0012] Step S3: Obtain the real sample data distribution

[0013] The multi-source signal fusion feature results are input into the improved Wasserstein generative adversarial network model to obtain the data distribution results of the real samples;

[0014] Step S4: Bearing fault diagnosis

[0015] The data distribution results of real samples are input into the enhanced deep residual shrinkage network model to obtain the rolling bearing fault diagnosis results.

[0016] Furthermore, in step S1, the specific processing process is as follows:

[0017] Step S11: acquiring data measured by sensors of rolling bearings, including bearing vibration signals, motor current signals, acceleration signals and rotation speed signals, and performing preprocessing of denoising and filtering operations to obtain preprocessed multi-source signal data.

[0018] Step S12: splicing and fusing the preprocessed multi-source signal data into a multi-source signal data matrix, and constructing fault multi-source signal data using principal component analysis.

[0019] Furthermore, in step S2, the specific processing process is as follows:

[0020] Step S21: Input the fault multi-source signal data into the improved normalized Hilbert-Huang transform time-frequency feature extraction model to obtain the fault data time-frequency information

[0021] Step S22: Time-frequency information of fault data Pooling operation is used to perform compression, dimension reduction and splicing fusion operations to obtain multi-source signal fusion feature results.

[0022] Furthermore, in the step S21, the improved normalized Hilbert-Huang transform time-frequency spectrum feature extraction model includes an empirical mode decomposition model, a Hilbert transform operation module and a normalized calculation operation module, and the specific processing process is as follows:

[0023] Step S211: Input the fault multi-source signal data into the empirical mode decomposition model to obtain i intrinsic mode functions c i (t), where t represents the time series index;

[0024] Step S212: The intrinsic mode function c i (t) Apply the Hilbert transform operation module to analyze and obtain the analytical signal z i (t):

[0025] z i (t) = c i (t)+jH{c i (t)}

[0026] Wherein, j represents the imaginary part of the complex number, H{·} represents the Hilbert transform;

[0027] And by the analytical signal z i (t) Calculate the instantaneous amplitude A(t) and instantaneous phase θ i (t);

[0028] Step S213: Set the instantaneous phase θ i (t) is normalized to obtain the normalized instantaneous frequency ω i (t):

[0029]

[0030] in, Instantaneous phase ω i (t) is the derivative with respect to time, π represents a constant value;

[0031] Step S214: Using the instantaneous frequency ω i (t) and instantaneous amplitude A(t), construct the time-frequency information of fault data

[0032] Furthermore, in step S214, the normalized instantaneous frequency ω i (t) as the vertical axis, time t as the horizontal axis, and the instantaneous amplitude A(t) as the color or grayscale value, the i-th intrinsic mode function c i (t) time-frequency spectrum; superimpose the time-frequency spectrum of the time-frequency spectrum of all intrinsic mode functions to obtain the time-frequency information of the fault data

[0033] Furthermore, in step S22, the specific processing process is as follows:

[0034] The time-frequency information of fault data The average pooling operation is used to compress the data and obtain the fault data sparse matrix P = {P1, P2, P2, ..., P N}, and use the interpolation operation I(·) to adjust the dimension of the fault data sparse matrix P1′=I(P k ), and then perform splicing and fusion operations by arranging them up and down to obtain the multi-source signal fusion feature results

[0035] Furthermore, in step S3, the specific processing process is as follows:

[0036] Step S31: Fusion of multi-source signal features Input into the original Wasserstein generative adversarial network model to obtain the loss function of the generator G And the loss function of the discriminator D in, is the generated fault sample data, x is the real fault sample data, P g is the generated fault data distribution, P r is the real fault data distribution, D(·) represents the discriminator, Represents the distribution of real fault data P r Sampling is performed, that is, x is from P r A sample randomly selected from

[0037] Denotes the distribution of generated fault data P g Sampling is performed, that is It is from P g A sample randomly selected from

[0038] Step S32: The generator G obtained above is used to establish a gradient penalty term L of the regularized discriminator D that satisfies the Lipschitz continuity constraint and is established using the gradient penalty term operation. GP :

[0039]

[0040] Among them, λ is the weight of the penalty term, represents the points sampled from the interpolation between the true fault data distribution and the generated fault data distribution, Represents the interpolation sample distribution Sampling is performed, that is is from A sample randomly selected from is the output of the discriminator D relative to The gradient of is the interpolation sample distribution, ||·||2 is the L2 norm;

[0041] Step S33: Substitute the loss function L of the above discriminator D D Combining the original loss and the gradient penalty term L GP , establish the loss function of the improved discriminator D The generator G and the discriminator D are optimized to obtain the improved Wasserstein generative adversarial network model, and then the multi-source signal fusion feature results are lose

[0042] Input the improved Wasserstein generative adversarial network model to obtain the data distribution results of real samples Among them, the optimization goal of the generator G is to minimize its loss function L G , the optimization goal of the discriminator D is to minimize And at the same time satisfy the Lipschitz constraint.

[0043] Furthermore, in step S32, the interpolation samples The calculation formula is as follows:

[0044]

[0045] Here, ε represents a random variable sampled from a uniform distribution U[0,1].

[0046] Furthermore, in step S4, the specific processing process is as follows:

[0047] Step S41: input the data distribution result of the real sample into the enhanced deep residual shrinkage network model, wherein the enhanced deep residual shrinkage network model includes a residual module, a shrinkage convolution attention module and an enhanced classifier; the residual module is used as the backbone network of the enhanced classifier, and the residual module outputs the feature result map y i :

[0048] y i =F(x tru ,{W i})+x tru

[0049] Among them, x tru Represents the data distribution result of the real sample, y i represents the output feature map of the ith residual module, F(·) represents the residual function, {W i} represents the weight parameter set of the inner layer of the residual module;

[0050] Step S42: Input the feature result map output by the residual module into the shrinkage convolution attention module, apply a global pooling operation to the input feature map, and obtain a new feature result map; input the two channel features of the new feature result map into a two-layer neural network formed by a multi-layer perceptron and a hidden layer, and obtain the output elements of the two channel features; perform product addition operations on the output elements of the two channel features, and pass the Sigmoid activation function to obtain feature maps of two dimensions, namely, space and channel; perform a full connection operation on the feature maps of two dimensions, namely, space and channel, and perform a shrinkage operation on the results of the processing of the channel information fully connected layer, and obtain the shrunken feature map channel vector:

[0051] z=W·x+b

[0052] in, represents the input channel vector, represents the weight matrix of the fully connected layer, represents the bias vector, represents the channel vector after contraction;

[0053] Perform channel-by-channel vector product on the feature map channel vector to obtain the attention mechanism result that assigns different weights to different channels of the input feature map;

[0054] Step S43: The above residual module and the compressed convolution block attention module constitute a deep residual contraction network basic model and obtain feature results, which are input into the enhanced classifier to obtain the rolling bearing fault recognition result.

[0055]

[0056] Where L represents the last layer index in the enhanced classifier, [x0,x1,…,x L ] represents the feature vector formed by connecting the outputs of all layers of the enhanced classifier, Pool(·) represents the pooling operation, W represents the weight matrix of the classifier layer, and σ(·) represents the function of the classifier.

[0057] The present invention also provides a bearing fault diagnosis system based on a multi-source fusion deep adversarial network. The rolling bearing fault diagnosis method described above includes:

[0058] The acquisition and preprocessing module is used to obtain and preprocess the data measured by the sensors of the rolling bearings to construct the multi-source signal data of the faults;

[0059] The feature fusion module is used to input the fault multi-source signal data into the improved normalized Hilbert-Huang transform time-frequency feature extraction model to obtain the time-frequency information of the fault data; the time-frequency information of the fault data is subjected to pooling operation and compression, dimensionality reduction and splicing and fusion operation to obtain the multi-source signal fusion feature result;

[0060] The data distribution acquisition module is used to input the multi-source signal fusion feature results into the improved Wasserstein generative adversarial network model to obtain the data distribution results of real samples;

[0061] The fault diagnosis module is used to input the data distribution results of real samples into the enhanced deep residual shrinkage network model to obtain the rolling bearing fault diagnosis results.

[0062] Compared with the prior art, the present invention has the following advantages:

[0063] 1. An improved normalized Hilbert-Huang transform time-frequency feature extraction model is introduced to capture the local features of the fault data in the time-frequency domain. At the same time, a pooling operation is used to compress the two-dimensional time-frequency image features in the time-frequency information of the fault data to highlight the important features. In addition, a splicing operation is used to obtain the multi-source signal fusion feature results to capture the multi-scale fault characteristics.

[0064] 2. The improved Wasserstein generative adversarial network model is used to obtain the data distribution results of real samples of different fault categories, which helps to build a comprehensive data distribution model and improve the generalization ability of the fault diagnosis system; and the enhanced deep residual shrinkage network model is used to solve the gradient vanishing problem in the deep network. At the same time, the shrinkage network is used to improve the sparsity of the deep network weights to improve the generalization of the model, further realize the recognition accuracy of rolling bearing faults, and obtain accurate diagnosis results. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a flow chart of a bearing fault diagnosis method based on a multi-source fusion deep adversarial network in Embodiment 1 of the present invention;

[0066] Figure 2 is a schematic diagram of the process of step S2 in the first embodiment of the present invention;

[0067] Figure 3 is a process schematic diagram of step S21 in the first embodiment of the present invention;

[0068] Figure 4 is a process schematic diagram of step S3 in the first embodiment of the present invention;

[0069] Figure 5 is a process schematic diagram of step S4 in the first embodiment of the present invention;

[0070] Figure 6 It is a structural diagram of a bearing fault diagnosis system based on a multi-source fusion deep adversarial network in Example 2 of the present invention. DETAILED DESCRIPTION

[0071] The following is a detailed description of an embodiment of the present invention. This embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation method and a specific operation process are given, but the protection scope of the present invention is not limited to the following embodiment.

[0072] Embodiment 1

[0073] like Figure 1 As shown, the bearing fault diagnosis method based on the multi-source fusion deep adversarial network provided by the embodiment of the present invention is used to diagnose rolling bearing faults, including the following steps:

[0074] Step S1: acquiring data measured by a sensor of a rolling bearing and performing preprocessing to obtain fault multi-source signal data;

[0075] Step S2: Input the fault multi-source signal data into the improved normalized Hilbert-Huang transform time-frequency feature extraction model to obtain the time-frequency information of the fault data; perform pooling operation and compression dimension reduction and splicing fusion operation on the time-frequency information of the fault data to obtain the multi-source signal fusion feature result to capture the multi-scale fault characteristics;

[0076] Step S3: input the multi-source signal fusion feature results into the improved Wasserstein generative adversarial network model to obtain the data distribution results of the real samples;

[0077] Step S4: Input the data distribution results of the real samples into the enhanced deep residual shrinkage network model to obtain the rolling bearing fault diagnosis results.

[0078] The fault diagnosis method based on a multi-source fusion deep adversarial network provided by the present invention is based on data measured by sensors of rolling bearings. By collecting bearing vibration signals, motor current signals, acceleration signals and speed signal data from rolling bearings, fault diagnosis considers four types of rolling bearing health states, including normal, inner ring fault, outer ring fault and bearing cage fault; the collected signals are denoised and filtered to eliminate environmental noise and irrelevant signal components, and the principal component analysis method is used to construct fault multi-source signal data; an improved normalized Hilbert-Huang transform time-frequency feature extraction model is introduced to capture the local characteristics of the fault data in the time-frequency domain, and a pooling operation is used to compress the two-dimensional time-frequency image features in the time-frequency information of the fault data to highlight the most important features, and a splicing operation is used to obtain multi-source signal fusion feature results to capture multi-scale fault characteristics; an improved Wasserstein generative adversarial network model is used to obtain the data distribution results of real samples of different fault categories, and a comprehensive data distribution model is constructed; and an enhanced deep residual shrinkage network model is used to solve the gradient vanishing problem in the deep network, and the shrinkage network part is used to promote the sparsity of the deep network weights to improve the generalization and interpretability of the model, so as to improve the recognition accuracy of rolling bearing faults.

[0079] like Figure 2 As shown, in this embodiment, the specific processing process of step S2 is as follows:

[0080] Step S21: Input the fault multi-source signal data into the improved normalized Hilbert-Huang transform time-frequency feature extraction model to obtain the fault data time-frequency information

[0081] Step S22: Time-frequency information of fault data Pooling operation is used to perform compression, dimension reduction and splicing fusion operations to obtain multi-source signal fusion feature results.

[0082] Among them, Figure 3 As shown, the improved normalized Hilbert-Huang transform time-frequency spectrum feature extraction model of this embodiment includes an empirical mode decomposition model, a Hilbert transform operation module and a normalized calculation operation module.

[0083] More specifically, in this embodiment, the specific processing process of step S21 is as follows:

[0084] Step S211: Input the fault multi-source signal data into the empirical mode decomposition model to obtain i intrinsic mode functions c i (t), where t represents the time series index;

[0085] Step S212: The intrinsic mode function ci (t) Apply the Hilbert transform operation module to analyze, and according to the following formula (1), the analytical signal z is obtained i (t), and the analytical signal z i (t) Calculate the instantaneous amplitude A(t) and instantaneous phase θ i (t);

[0086] z i (t) = c i (t)+jH{c i (t)} (1)

[0087] Wherein, j represents the imaginary part of the complex number, H{·} represents the Hilbert transform;

[0088] Step S213: Set the instantaneous phase θ i (t), according to the following formula (2), the normalization operation is performed to obtain the normalized instantaneous frequency ω i (t);

[0089]

[0090] in, Instantaneous phase ω i (t) is the derivative with respect to time, π represents a constant value;

[0091] Step S214: Using the instantaneous frequency ω i (t) and instantaneous amplitude A(t), construct the time-frequency information of fault data

[0092] The fault multi-source signal data is input into the empirical mode decomposition model, and i intrinsic mode functions c are screened out according to the frequency range or energy standard. i (t); the normalized instantaneous frequency ω i (t) as the vertical axis, time t as the horizontal axis, and the instantaneous amplitude A(t) as the color or grayscale value, the i-th intrinsic mode function c i The time-frequency spectrum of the time-frequency spectrum of all the filtered intrinsic mode functions is superimposed to obtain the time-frequency information of the fault data.

[0093] More specifically, in this embodiment, the specific processing process of step S22 is as follows:

[0094] The time-frequency information of fault data The average pooling operation is used to compress the data and obtain the fault data sparse matrix P = {P1, P2, P2, ..., P N}, and use the interpolation operation I(·) to adjust the dimension of the fault data sparse matrix P1′=I(Pk ), and then perform splicing and fusion operations by arranging them up and down to obtain the multi-source signal fusion feature results

[0095] like Figure 4 As shown, in this embodiment, the specific processing process of step S3 is as follows:

[0096] Step S31: Fusion of multi-source signal features Input into the original Wasserstein generative adversarial network model to obtain the loss function of the generator G And the loss function of the discriminator D in, is the generated fault sample data, x is the real fault sample data, P g is the generated fault data distribution, P r is the real fault data distribution, D(·) represents the discriminator, Represents the distribution of real fault data P r Sampling is performed, that is, x is from P r A sample randomly selected from

[0097] Denotes the distribution of generated fault data P g Sampling is performed, that is It is from P g A sample randomly selected from

[0098] Step S32: The generator G obtained above is used to establish the gradient penalty term L of the regularized discriminator D that satisfies the Lipschitz continuity constraint according to the following formula (3), and the gradient penalty term operation is used to establish the gradient penalty term L GP ;

[0099]

[0100] Among them, λ is the weight of the penalty term, represents the points sampled from the interpolation between the true fault data distribution and the generated fault data distribution, Represents the interpolation sample distribution Sampling is performed, that is is from A sample randomly selected from is the output of the discriminator D relative to The gradient of is the interpolation sample distribution, ||·||2 is the L2 norm; and, according to the following formula (4), the interpolation sample can be calculated

[0101]

[0102] Where ε represents a random variable sampled from a uniform distribution U[0,1];

[0103] Step S33: Substitute the loss function L of the above discriminator D D Combining the original loss and the gradient penalty term L GP , establish the loss function of the improved discriminator D The generator G and the discriminator D are optimized to obtain the improved Wasserstein generative adversarial network model, and then the multi-source signal fusion feature results are Input the improved Wasserstein generative adversarial network model to obtain the data distribution results of real samples Among them, the optimization goal of the generator G is to minimize its loss function L G , the optimization goal of the discriminator D is to minimize And at the same time satisfy the Lipschitz constraint.

[0104] It should be noted that, in this embodiment, the generator G is responsible for generating fault feature vectors, and the discriminator D is responsible for distinguishing between real fault features and generated features; the Wasserstein distance is used to measure the distance between two probability distributions to improve training stability. In order to improve the training stability of the improved Wasserstein generative adversarial network and the quality of generated samples, a gradient penalty is introduced; the discriminator and the generator are trained alternately, and the two-time-scale update rule (TTUR) strategy is used, that is, the discriminator D is updated more frequently in the training cycle to balance the learning speeds of the two; the learning rates of the generator G and the discriminator D are set to 0.0001 and 0.0003, respectively.

[0105] like Figure 5 As shown, in this embodiment, the specific processing process of step S4 is as follows:

[0106] Step S41: Input the data distribution results of the real samples into the enhanced deep residual shrinkage network model, where Figure 3 As shown in FIG. 1 , the enhanced deep residual shrinkage network model includes a residual module, a shrinkage convolution attention module and an enhanced classifier; the residual module is used as the backbone network of the classifier, wherein the residual module, according to the following formula (5), obtains the residual module output feature result graph y i ;

[0107] y i =F(x tru ,{W i})+x tru (5)

[0108] Among them, xtru Represents the data distribution result of the real sample, y i represents the output feature map of the ith residual module, F(·) represents the residual function, {W i} represents the weight parameter set of the inner layer of the residual module;

[0109] Step S42: input the feature result map output by the residual module into the shrinkage convolution attention module, apply a global pooling operation to the input feature map, and obtain a new feature result map; input the two channel features of the new feature result map into a two-layer neural network formed by a multi-layer perceptron and a hidden layer, and obtain the output elements of the two channel features; perform product addition operations on the output elements of the two channel features, and use a Sigmoid activation function to obtain feature maps of two dimensions, namely, space and channel; perform a full connection operation on the feature maps of two dimensions, namely, space and channel, and perform a shrinkage operation on the results of the channel information fully connected layer processing, and obtain a shrunken feature map channel vector according to the following formula (6); perform a channel-by-channel vector product on the feature map channel vector, and obtain the attention mechanism result in which different channels of the input feature map are assigned different weights;

[0110] z=W·x+b (6)

[0111] in, represents the input channel vector, represents the weight matrix of the fully connected layer, represents the bias vector, Represents the channel vector after contraction.

[0112] Step S43: The above residual module and the compressed convolution block attention module constitute a deep residual shrinkage network basic model to obtain feature results, and input them into the enhanced classifier. According to the following formula (7), the rolling bearing fault recognition result is obtained.

[0113]

[0114] Where L represents the last layer index in the enhanced classifier, [x0,x1,...,x L ] represents the feature vector formed by connecting the outputs of all layers of the enhanced classifier, Pool(·) represents the pooling operation, W represents the weight matrix of the classifier layer, and σ(·) represents the function of the classifier.

[0115] In this embodiment, during the training process of the enhanced deep residual shrinkage network model, the batch size and the number of iterations for each class are set to 32 and 2000 respectively; the shrinkage convolution attention module mainly includes two independent attention functions: channel attention and spatial attention. In the channel attention submodule, the input features of each channel are first compressed by global average pooling and global maximum pooling operations to obtain the features of the entire spatial area; secondly, the above features are input into a shared multi-layer perceptron network to obtain deeper representation features; finally, the two representation features from the multi-layer perceptron are merged and activated to obtain the final channel attention vector; the residual module is used as the backbone network of the classifier, and then the shrinkage convolution attention module is received to form a basic module. Considering the richness of feature extraction, the present invention is designed to use 4 basic modules to form an enhanced deep residual shrinkage network model.

[0116] Embodiment 2

[0117] like Figure 6 As shown, this embodiment provides a bearing fault diagnosis system based on a multi-source fusion deep adversarial network, which diagnoses rolling bearing faults based on the bearing fault diagnosis method in embodiment one, including:

[0118] The fault multi-source signal data module 31 is used to obtain and pre-process the data measured by the sensor of the rolling bearing to obtain the fault multi-source signal data;

[0119] The fault diagnosis module 32 based on the deep generative adversarial network of multi-source signal fusion is used to first input the multi-source signal data of the fault into the improved normalized Hilbert-Huang transform time-frequency feature extraction model to obtain the time-frequency information of the fault data; apply pooling operation to the time-frequency information of the fault data and perform compression, dimensionality reduction and splicing and fusion operations to obtain the multi-source signal fusion feature results; then input the multi-source signal fusion feature results into the improved Wasserstein generative adversarial network model to obtain the data distribution results of the real samples; finally, input the data distribution results of the real samples into the enhanced deep residual shrinkage network model to obtain the rolling bearing fault diagnosis results.

[0120] This embodiment also provides an electronic device (computer, server, smart phone, network device, etc.), including a memory and a processor, wherein the memory is used to store a computer program executable by the processor, and when the processor executes the computer program, the bearing fault diagnosis method in the above-mentioned embodiment 1 is implemented.

[0121] In summary, the bearing fault diagnosis method and system based on the multi-source fusion deep adversarial network of the above embodiment introduces an improved normalized Hilbert-Huang transform time-frequency feature extraction model to capture the local characteristics of the fault data in the time-frequency domain, and simultaneously uses a pooling operation to compress the two-dimensional time-frequency image features in the time-frequency information of the fault data to highlight the most important features, and uses a splicing operation to obtain multi-source signal fusion feature results to capture multi-scale fault characteristics. The improved Wasserstein generative adversarial network model is used to obtain the data distribution results of real samples of different fault categories, which helps to build a comprehensive data distribution model and improve the generalization ability of the fault diagnosis system; and an enhanced deep residual shrinkage network model is used to solve the gradient vanishing problem in the deep network, and the shrinkage network part is used to promote the sparsity of the deep network weights, further realizing the recognition accuracy of rolling bearing faults and obtaining accurate diagnosis results.

[0122] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A bearing fault diagnosis method based on a multi-source fusion deep adversarial network, characterized in that: The following steps are involved: Step S1: Data collection and preprocessing Acquire and pre-process the data measured by the sensors of the rolling bearings to construct the multi-source signal data of the faults; Step S2: Multi-source signal feature fusion Input the fault multi-source signal data into the improved normalized Hilbert-Huang transform time-frequency feature extraction model to obtain the fault data time-frequency information; Pooling operation is adopted for the time-frequency information of fault data, and compression, dimension reduction and splicing and fusion operations are performed to obtain the multi-source signal fusion feature results; Step S3: Obtain the real sample data distribution The multi-source signal fusion feature results are input into the improved Wasserstein generative adversarial network model to obtain the data distribution results of the real samples; Step S4: Bearing fault diagnosis The data distribution results of real samples are input into the enhanced deep residual shrinkage network model to obtain the rolling bearing fault diagnosis results.

2. The bearing fault diagnosis method based on multi-source fusion deep adversarial network according to claim 1 is characterized in that: In step S1, the specific processing process is as follows: Step S11: acquiring data measured by sensors of rolling bearings, including bearing vibration signals, motor current signals, acceleration signals and rotation speed signals, and performing preprocessing of denoising and filtering operations to obtain preprocessed multi-source signal data. Step S12: splicing and fusing the preprocessed multi-source signal data into a multi-source signal data matrix, and constructing fault multi-source signal data using principal component analysis.

3. The bearing fault diagnosis method based on multi-source fusion deep adversarial network according to claim 1 is characterized in that: In step S2, the specific processing process is as follows: Step S21: Input the fault multi-source signal data into the improved normalized Hilbert-Huang transform time-frequency feature extraction model to obtain the fault data time-frequency information Step S22: Time-frequency information of fault data Pooling operation is used to perform compression, dimension reduction and splicing fusion operations to obtain multi-source signal fusion feature results.

4. The bearing fault diagnosis method based on multi-source fusion deep adversarial network according to claim 3 is characterized in that: In the step S21, the improved normalized Hilbert-Huang transform time-frequency spectrum feature extraction model includes an empirical mode decomposition model, a Hilbert transform operation module and a normalized calculation operation module, and the specific processing process is as follows: Step S211: Input the fault multi-source signal data into the empirical mode decomposition model to obtain i intrinsic mode functions c i (t), where t represents the time series index; Step S212: The intrinsic mode function c i (t) Apply the Hilbert transform operation module to analyze and obtain the analytical signal z i (t): z i (t)=c i (t)+jH{c i (t)} Wherein, j represents the imaginary part of the complex number, H{·} represents the Hilbert transform; And by the analytical signal z i (t) Calculate the instantaneous amplitude A(t) and instantaneous phase θ i (t); Step S213: Set the instantaneous phase θ i (t) is normalized to obtain the normalized instantaneous frequency ω i (t): in, Instantaneous phase ω i (t) is the derivative with respect to time, π represents a constant value; Step S214: Using the instantaneous frequency ω i (t) and instantaneous amplitude A(t), construct the time-frequency information of fault data 5. The bearing fault diagnosis method based on multi-source fusion deep adversarial network according to claim 4 is characterized in that: In step S214, the normalized instantaneous frequency ω i (t) as the vertical axis, time t as the horizontal axis, and the instantaneous amplitude A(t) as the color or grayscale value, the i-th intrinsic mode function c i (t) time-frequency spectrum; superimpose the time-frequency spectrum of the time-frequency spectrum of all intrinsic mode functions to obtain the time-frequency information of the fault data 6. The bearing fault diagnosis method based on multi-source fusion deep adversarial network according to claim 3 is characterized in that: In step S22, the specific processing process is as follows: The time-frequency information of fault data The average pooling operation is used to compress the data and obtain the fault data sparse matrix P = {P1, P2, P2, ..., P N }, and use the interpolation operation I(·) to adjust the dimension of the fault data sparse matrix P1′=I(P k ), and then perform splicing and fusion operations by arranging them up and down to obtain the multi-source signal fusion feature results 7. The bearing fault diagnosis method based on multi-source fusion deep adversarial network according to claim 5 is characterized in that: In step S3, the specific processing process is as follows: Step S31: Fusion of multi-source signal features Input into the original Wasserstein generative adversarial network model to obtain the loss function of the generator G And the loss function of the discriminator D in, is the generated fault sample data, x is the real fault sample data, P g is the generated fault data distribution, P r is the real fault data distribution, D(·) represents the discriminator, Represents the distribution of real fault data P r Sampling is performed, that is, x is from P r A sample randomly selected from Denotes the distribution of generated fault data P g Sampling is performed, that is It is from P g A sample randomly selected from Step S32: The generator G obtained above is used to establish a gradient penalty term L of the regularized discriminator D that satisfies the Lipschitz continuity constraint and is established using the gradient penalty term operation. GP : Among them, λ is the weight of the penalty term, represents the points sampled from the interpolation between the true fault data distribution and the generated fault data distribution, Represents the interpolation sample distribution Sampling is performed, that is is from A sample randomly selected from is the output of the discriminator D relative to The gradient of is the interpolation sample distribution, ||·||2 is the L2 norm; Step S33: Substitute the loss function L of the above discriminator D D Combining the original loss and the gradient penalty term L GP , establish the loss function of the improved discriminator D The generator G and the discriminator D are optimized to obtain the improved Wasserstein generative adversarial network model, and then the multi-source signal fusion feature results are Input the improved Wasserstein generative adversarial network model to obtain the data distribution results of real samples Among them, the optimization goal of the generator G is to minimize its loss function L G , the optimization goal of the discriminator D is to minimize And at the same time satisfy the Lipschitz constraint.

8. The bearing fault diagnosis method based on multi-source fusion deep adversarial network according to claim 7 is characterized in that: In step S32, the interpolation samples The calculation formula is as follows: Here, ε represents a random variable sampled from a uniform distribution U[0,1].

9. The bearing fault diagnosis method based on multi-source fusion deep adversarial network according to claim 7 is characterized in that: In step S4, the specific processing process is as follows: Step S41: input the data distribution result of the real sample into the enhanced deep residual shrinkage network model, wherein the enhanced deep residual shrinkage network model includes a residual module, a shrinkage convolution attention module and an enhanced classifier; the residual module is used as the backbone network of the enhanced classifier, and the residual module outputs the feature result map y i : y i =F(x tru ,{W i })+x tru Among them, x tru Represents the data distribution results of the real sample, y i represents the output feature map of the ith residual module, F(·) represents the residual function, {W i } represents the weight parameter set of the inner layer of the residual module; Step S42: Input the feature result map output by the residual module into the shrinkage convolution attention module, apply a global pooling operation to the input feature map, and obtain a new feature result map; input the two channel features of the new feature result map into a two-layer neural network formed by a multi-layer perceptron and a hidden layer, and obtain the output elements of the two channel features; perform product addition operations on the output elements of the two channel features, and pass the Sigmoid activation function to obtain feature maps of two dimensions, namely, space and channel; perform a full connection operation on the feature maps of two dimensions, namely, space and channel, and perform a shrinkage operation on the results of the processing of the channel information fully connected layer, and obtain the shrunken feature map channel vector: z=W·x+b in, represents the input channel vector, represents the weight matrix of the fully connected layer, represents the bias vector, represents the channel vector after contraction; Perform channel-by-channel vector product on the feature map channel vector to obtain the attention mechanism result that assigns different weights to different channels of the input feature map; Step S43: The above residual module and the compressed convolution block attention module constitute a deep residual contraction network basic model and obtain feature results, which are input into the enhanced classifier to obtain the rolling bearing fault recognition result. Where L represents the last layer index in the enhanced classifier, [x0,x1,...,x L ] represents the feature vector formed by connecting the outputs of all layers of the enhanced classifier, Pool(·) represents the pooling operation, W represents the weight matrix of the classifier layer, and σ(·) represents the function of the classifier.

10. Bearing fault diagnosis system based on multi-source fusion deep adversarial network, characterized in that: The rolling bearing fault is diagnosed based on the bearing fault diagnosis method according to any one of claims 1 to 9, comprising: The acquisition and preprocessing module is used to obtain and preprocess the data measured by the sensors of the rolling bearings to construct the multi-source signal data of the faults; The feature fusion module is used to input the fault multi-source signal data into the improved normalized Hilbert-Huang transform time-frequency feature extraction model to obtain the time-frequency information of the fault data; the time-frequency information of the fault data is subjected to pooling operation and compression, dimensionality reduction and splicing and fusion operation to obtain the multi-source signal fusion feature result; The data distribution acquisition module is used to input the multi-source signal fusion feature results into the improved Wasserstein generative adversarial network model to obtain the data distribution results of real samples; The fault diagnosis module is used to input the data distribution results of real samples into the enhanced deep residual shrinkage network model to obtain the rolling bearing fault diagnosis results.

Images