A bearing fault diagnosis method based on sparse classification and multi-layer composite domain adaptation

Abstract

The present invention provides a bearing fault diagnosis method based on sparse classification of multi-layer composite domain adaptation, which belongs to the field of mechanical fault diagnosis. The method comprises collecting vibration data of rolling bearings with known faults and unknown faults under different working conditions, performing wavelet transform on the obtained source domain and target domain rolling bearing vibration data to obtain a two-dimensional time-frequency graph; constructing a Coral-MMD multi-layer composite domain adaptation label-consistent sparse classification network including a deep feature extraction module, a domain distribution difference module and a label-consistent sparse classification module; using the obtained source domain data of known fault states and a portion of target domain data of unknown fault states as input to train the network; finally, inputting the remaining data in the obtained target domain into the trained network, and obtaining the state label corresponding to the input data by the network's label-consistent sparse classification module to realize rolling bearing fault diagnosis under different working conditions. The present invention can realize rolling bearing fault diagnosis under variable working conditions and improve diagnostic accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical equipment maintenance, and in particular to a bearing fault diagnosis method based on sparse classification of multi-layer composite domain adaptation. Background Art

[0002] Rotating machinery is the most widely used machinery in industry, ranging from small items like automobiles, generators, and fans to large items like transport ships, aircraft carriers, and spacecraft. Rolling bearings are a key component of rotating mechanisms, and approximately 30% of mechanical failures are related to bearing damage. When a rolling bearing fails, the equipment shuts down and requires repair. Therefore, establishing an effective bearing condition monitoring and fault diagnosis system to ensure bearing health is crucial for modern industrial production. Due to the variability of rolling bearing operating conditions and the complexity of their working environment, the incidence of rolling bearing failures increases with age. Rolling bearing fault diagnosis technology is a comprehensive and comprehensive technology designed to predict and diagnose equipment failures, playing a crucial role in rolling bearing fault warning and identification. Bearing vibration signals are a key carrier of bearing condition information. Because the vibration signals caused by various faults vary, extracting valuable feature information from these complex vibration signals to assess equipment health has been a key research focus in both academia and industry.

[0003] Traditional fault diagnosis techniques rely on manual feature extraction and selection followed by a classifier. This model relies on expert experience and knowledge, significantly limiting its widespread adoption across industry. With the advent of the "big data cloud computing era," massive amounts of data are posing significant challenges to traditional fault diagnosis techniques. Against this backdrop, deep learning offers new promise in the field of rolling bearing fault diagnosis. Deep learning adaptively extracts valuable features from raw signals, eliminating the reliance on expert knowledge and experience for manual feature extraction and enabling self-learning diagnostic models. However, the research and application of deep learning in fault diagnosis is still in its early stages. In particular, the application of models from laboratory environments to industrial settings remains a challenge, primarily due to two key factors: 1. How to identify models with a broad range of applications or environments based on the characteristics of rolling bearing faults, thereby avoiding model diversity; and 2. How to improve the transferability of models from laboratory to industrial environments without compromising accuracy. Current deep learning approaches employed in this field suffer from the following challenges: 1. The model's feature extraction network is too shallow, resulting in poor adaptability of the model's feature extraction module. 2. Deep learning models often use SVM and Softmax classifiers, which have poor adaptability and affect the migration effect.

[0004] Based on the above problems, it is necessary to develop a bearing fault diagnosis method based on sparse classification of multi-layer composite domain adaptation. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a bearing fault diagnosis method based on sparse classification of multi-layer composite domain adaptation. By customizing a relatively deep feature extraction network to extract features, the network adaptability is made stronger. By adopting the LC-ksvd model to self-learn a dictionary and a classifier, the optimal classifier can be found according to different data distributions, the migration ability and adaptability of the model can be improved, and the rolling bearing fault diagnosis under variable working conditions can be realized, effectively improving the diagnostic accuracy of bearing faults under different working conditions.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0007] A bearing fault diagnosis method based on sparse classification of multi-layer composite domain adaptation includes the following steps:

[0008] S1, collecting vibration time domain signals of rolling bearings in different fault states under different operating conditions, filtering the collected time domain signals to remove noise in the time domain signals, performing wavelet transform on the denoised time domain signals, and obtaining a two-dimensional time-frequency graph corresponding to the time domain signals;

[0009] The time domain signal of the rolling bearing with known fault status collected under known operating conditions constitutes the source domain, and the vibration time domain signal of the rolling bearing with unknown fault status collected under other operating conditions constitutes the target domain;

[0010] S2, constructing a Coral-MMD multi-layer composite domain-adapted label-consistent sparse classification network; the Coral-MMD multi-layer composite domain-adapted label-consistent sparse classification network includes a deep feature extraction module, a domain distribution difference module and a label-consistent sparse classification module;

[0011] S3: Take the source domain data with known fault states obtained in S1 and a portion of target domain data with unknown fault states as input, and use the fault state corresponding to each two-dimensional time-frequency graph in the source domain data as the expected output of the label-consistent sparse classification module. Determine the state classification loss value and classification accuracy based on the fault state label and predicted label corresponding to each two-dimensional time-frequency graph.

[0012] The domain classification loss value is determined by calculating the maximum difference in the distribution means of the features of the source domain and the target domain and the distance between the feature covariances;

[0013] The parameters of the Coral-MMD multi-layer composite domain-adapted label-consistent sparse classification network constructed in S2 are iteratively adjusted multiple times by using the state classification loss value and the domain classification loss value, thereby obtaining a trained Coral-MMD multi-layer composite domain difference label-consistent sparse classification network model;

[0014] S4, input the remaining data in the target domain obtained in S1 into the trained Coral-MMD multi-layer composite domain adapted label consistent sparse classification network, and the label consistent sparse classification module obtains the state label corresponding to the input data to realize rolling bearing fault diagnosis under different working conditions.

[0015] A further improvement of the technical solution of the present invention is that: in step 2,

[0016] The deep feature extraction module is used to extract features from the two-dimensional time-frequency graphs of the source domain and the target domain.

[0017] The domain distribution difference module is used to measure the maximum difference in the distribution means of the extracted features of the source domain and the target domain and the distance between the feature covariances;

[0018] The label consistent sparse classification module is used to learn a dictionary with sparse discrimination capability and a linear classifier for the extracted features of the source domain, and classify the extracted features of the source domain and the target domain.

[0019] A further improvement of the technical solution of the present invention is that: the deep feature extraction module adopts a customized multi-layer convolution residual block: including a first convolution module, a second convolution module, a first pooling layer, a first residual module, a second residual module, a third residual module, a fourth residual module, and a second pooling layer;

[0020] The first convolution module and the second convolution module each include a convolution layer, a BN layer, and a ReLu layer; the number of output channels of the convolution layer is 64, the convolution kernel is 3*3, and the stride is 1;

[0021] The outputs of the first convolution module and the second convolution module are calculated as follows:

[0022]

[0023] Among them, y represents the feature tensor after the convolution layer operation, f represents the nonlinear activation function, x represents the input feature tensor, W represents the weight of the convolution kernel, and B represents the bias coefficient;

[0024] The first pooling layer is a 3*3 maximum pooling layer, and the step size of the first pooling layer is 2;

[0025] The first residual module includes three residual blocks with 64 output channels; the second residual module includes six residual blocks with 128 output channels; the third residual module includes six residual blocks with 256 output channels; the fourth residual module includes three residual blocks with 512 output channels;

[0026] Among them, each residual block consists of a 1*1 convolution layer with its own mapping, a 3*3 convolution layer and a 1*1 convolution layer; the feature representation of the output after the i-th residual block is:

[0027] X i+1 =H i (X i )+X i

[0028] Among them, X i represents the input features of the i-th residual block, X i+1 represents the output feature of the i-th residual block, H i Composite function for feature extraction of layer i, including BN, Relu, 1*1Conv2d, 3*3Conv2d;

[0029] The second pooling layer is an average pooling layer.

[0030] A further improvement of the technical solution of the present invention is that: the domain distribution difference module adopts a composite domain difference loss of Coral and MMD;

[0031] The calculation formula of the composite domain difference loss is as follows:

[0032] Loss = α*loss coral +β*loss MMD

[0033] Among them, loss coral Represents Coral distance loss, loss MMD Represents MMD distance loss, α and β represent the weight coefficients of the two distance losses respectively; Coral distance loss loss coral The calculation formula is:

[0034]

[0035] Among them, D s and D t Represents source domain and target domain data respectively; n s and n t Represents the number of samples in the source domain and target domain respectively; C s and C t Represent the data covariance matrix of the source domain and the target domain respectively; represents the Frobenius norm, and d represents the number of dimensions of the sample feature space;

[0036] MMD distance loss MMD The calculation formula is:

[0037]

[0038] in, and They represent the j-th source domain feature and the k-th target domain feature output by the i-th layer respectively; φ(·) represents the mapping function; ||·|| H represents the regenerated Hilbert space norm.

[0039] A further improvement of the technical solution of the present invention is that: the label-consistent sparse classification module obtains a dictionary with combined discrimination and reconstruction criteria and an optimal multi-class linear classifier through learning; the features obtained by the feature extraction module and their corresponding category labels are used as inputs of the label-consistent sparse classification module for learning, thereby obtaining a sparse discrimination dictionary and a linear classifier for classifying and identifying the input data.

[0040] Due to the adoption of the above technical solution, the technical advancements achieved by the present invention are:

[0041] 1. The present invention uses a customized multi-layer convolutional residual block network to extract fault signal features, making the obtained features more comprehensive and complete, overcoming the shortcomings of incomplete feature extraction of other networks.

[0042] 2. The present invention further uses a domain distribution difference module to calculate the data distribution differences under different working conditions, so that the feature extraction module extracts the common features of the data under different working conditions, which can ensure that the model training and testing can still maintain high accuracy under different working conditions, improve the robustness and robustness of the model, and make the model more applicable.

[0043] 3. The present invention adopts a label-consistent sparse classification model, which uses fewer feature signals to sparsely express the signal to be tested, thereby simplifying the classification task, reducing the complexity of the model, and further improving the classification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. Those skilled in the art can derive other drawings based on these drawings without inventive effort.

[0045] Figure 1A flowchart of a bearing fault diagnosis method based on sparse classification of multi-layer composite domain adaptation provided by the present invention;

[0046] Figure 2 This is a structural diagram of a label-consistent sparse classification network based on Coral-MMD multi-layer composite domain adaptation in an embodiment of the present invention. DETAILED DESCRIPTION

[0047] It should be noted that the terms "including" and "having" and any variations thereof in the specification and claims of the present invention and the above-mentioned drawings are intended to cover non-exclusive inclusions. For example, a process, method, system, product or apparatus comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or are inherent to these processes, methods, products or apparatuses.

[0048] The embodiment of the present application solves the problem in the prior art of "low bearing fault diagnosis accuracy and poor migration effect under different working conditions" by providing a bearing fault diagnosis method based on sparse classification of multi-layer composite domain adaptation. The general idea is: using a vibration acceleration sensor to collect vibration data of rolling bearings with known faults and unknown faults under different working conditions, and performing wavelet transform on the obtained source domain and target domain rolling bearing vibration data to obtain a two-dimensional time-frequency graph; constructing a Coral-MMD multi-layer composite domain adapted label-consistent sparse classification network including a deep feature extraction module, a domain distribution difference module and a label-consistent sparse classification module; using the obtained source domain data of known fault states and a part of the target domain data of unknown fault states as input to train the Coral-MMD multi-layer composite domain adapted label-consistent sparse classification network; finally, inputting the remaining data in the obtained target domain into the trained Coral-MMD multi-layer composite domain adapted label-consistent sparse classification network, and the network's label-consistent sparse classification module obtains the state label corresponding to the input data, thereby realizing rolling bearing fault diagnosis under different working conditions.

[0049] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments:

[0050] like Figure 1 As shown, a bearing fault diagnosis method based on sparse classification of multi-layer composite domain adaptation includes the following steps:

[0051] S1. First, use a vibration accelerometer to collect vibration data from rolling bearings with known and unknown faults under different operating conditions. In this implementation, the sampling frequency is set to 12kHz. The health status of the rolling bearing includes normal status, various rolling element faults, various inner race faults, and various outer race faults. A non-overlapping wavelet transform is then performed on the collected time domain signals to obtain a 28x28 two-dimensional wavelet transform image. Each wavelet transform image sample of the vibration data with known faults is labeled with the state code corresponding to the fault type, with 0 representing normal status and 1, 2, 3, ..., N representing the fault type state sequence.

[0052] S2, building a Coral-MMD multi-layer composite domain-adapted label-consistent sparse classification network;

[0053] The Coral-MMD multi-layer composite domain-adapted label-consistent sparse classification network includes a deep feature extraction module, a domain distribution difference module, and a label-consistent sparse classification module.

[0054] The deep feature extraction module is used to extract features from two-dimensional time-frequency graphs of a source domain and a target domain, wherein the source domain is composed of time domain signals of rolling bearings with known fault states collected under known operating conditions, and the target domain is composed of vibration time domain signals of rolling bearings with unknown fault states collected under other operating conditions. The feature extraction module includes a first convolution module, a second convolution module, a first pooling layer, a first residual module, a second residual module, a third residual module, a fourth residual module, and a second pooling layer;

[0055] The first convolution module includes a convolution layer, a BN layer, and a ReLu layer; the number of input channels of the convolution layer is 3, the number of output channels is 64, the convolution kernel is 3*3, and the stride is 1;

[0056] The second convolution module includes a convolution layer, a BN layer, and a ReLu layer; the output channel of the convolution layer is 64, the convolution kernel is 3*3, and the stride is 1;

[0057] The calculations of the convolutional layer and the ReLu layer are as follows:

[0058]

[0059] y represents the feature tensor after the convolution layer operation, f represents the nonlinear activation function, x represents the input feature tensor, W represents the weight of the convolution kernel, and B represents the bias coefficient. The BN layer is calculated as follows:

[0060]

[0061]

[0062]

[0063]

[0064] Among them, x k represents the kth feature tensor of the input, μ represents the mean of the input feature tensor, σ 2 represents the variance of the input feature tensor, Represents the output of the input feature tensor after mean variance batch normalization, y k Represents x k Output after batch normalization.

[0065] The first pooling layer is a 3*3 maximum pooling layer, and the step size of the first pooling layer is 2;

[0066] The first residual module includes three residual blocks with 64 output channels; the second residual module includes six residual blocks with 128 output channels; the third residual module includes six residual blocks with 256 output channels; the fourth residual module includes three residual blocks with 512 output channels;

[0067] Among them, each residual block consists of a 1*1 convolution layer with its own mapping, a 3*3 convolution layer and a 1*1 convolution layer; the output feature after the i-th residual block can be expressed as:

[0068] X i+1 =H i (X i )+X i

[0069] Among them, X i represents the input features of the i-th residual block, X i+1 represents the output feature of the i-th residual block, H i Composite function for feature extraction of layer i, including BN, Relu, 1*1Conv2d, 3*3Conv2d;

[0070] The second pooling layer is the average pooling layer.

[0071] The domain distribution difference module is used to calculate the maximum difference and second-order statistics (i.e., covariance) of the distribution mean of the signal features output by each residual block after the source domain and target domain data pass through the first residual module, the second residual module, the third residual module, and the fourth residual module of the feature extraction layer, so that the feature extraction module can extract the features of the common part of the source domain and the target domain, ignore the features of the difference part of the source domain and the target domain, and enable the model to diagnose the types of bearing faults running under different working conditions; the domain distribution difference module is composed of the Coral layer and the MMD layer to calculate the composite domain difference loss. The calculation formula of the composite domain difference loss is as follows:

[0072] loss = α * loss coral +β*loss MMD

[0073] The loss in the above formula coral and loss MMD Represent the Coral distance loss and MMD distance loss respectively; α and β represent the weight coefficients of the two distance losses respectively.

[0074]

[0075] Where D s and D t are source domain and target domain data respectively; n s and n t are the number of sample points in the source domain and the target domain respectively; C s and C t are the data covariance matrices of the source domain and the target domain respectively; ||·|| F 2 represents the square of the matrix Frobenius norm, and d represents the number of dimensions of the sample feature space.

[0076]

[0077] Where, and They represent the j-th source domain data and the k-th target domain data output by the i-th layer respectively; φ(·) represents the mapping, which maps the original data to the Shilbit space of the regenerated kernel; ||·|| H represents the regenerated Hilbert space norm.

[0078] The label consistent sparse classification module is used to classify the features of the source domain output by the feature extraction module and calculate the classification accuracy in the training phase;

[0079] Label-consistent sparse classification module: It learns a dictionary with a combined discriminant and reconstruction criterion and an optimal multi-class linear classifier;

[0080]

[0081] Where Y represents the input sample data, D represents the overcomplete dictionary, W represents the classifier parameters, A represents the linear transformation matrix, X represents the sparse coding, Q represents the discriminant sparse coding matrix, and H represents the category label of the input signal Y; the first term in the above formula is the reconstruction error, the second term is the discriminant sparse coding error, and the third term is the classification error.

[0082] See Table 1, which shows the backbone network structure of the Coral-MMD multi-layer composite domain-adapted label-consistent sparse classification network provided by an embodiment of the present invention.

[0083] Table 1

[0084]

[0085]

[0086] S3, taking the source domain data of known fault states obtained in S1 and a portion of target domain data of unknown fault states as input, and taking the fault state corresponding to each two-dimensional time-frequency graph in the source domain data as the expected output of the label-consistent sparse classification module, determines the state classification loss value and classification accuracy through the fault state label and predicted label corresponding to each two-dimensional time-frequency graph; determines the domain classification loss value by calculating the maximum difference in the distribution mean of the features of the source domain and the target domain and the distance between the feature covariances; and iteratively adjusts the parameters of the Coral-MMD multi-layer composite domain-adapted label-consistent sparse classification network constructed in S2 multiple times using the state classification loss value and domain classification loss value. Thus, a trained Coral-MMD multi-layer composite domain difference label-consistent sparse classification network is obtained, and finally the parameters of the network and each module of the trained network are saved.

[0087] S4, rolling bearing fault diagnosis:

[0088] The remaining data in the target domain obtained in step S1 is input into the trained Coral-MMD multi-layer composite domain-adapted label-consistent sparse classification network, and then input into the Coral-MMD multi-layer composite domain difference label-consistent sparse classification network saved in step S3. The label-consistent sparse classification module of the network obtains the state label corresponding to the input data to realize rolling bearing fault diagnosis under different working conditions.

[0089] Table 2 shows the bearing fault data collected in the laboratory. The motor speeds are 800 r / min, 2000 r / min, and 3200 r / min, respectively. The load is zero. The sampling frequency of the vibration data is 12.8 kHz. The data of the working conditions with a motor speed of 800 r / min and a speed of 3200 r / min are used as the source domain and target domain data sets, respectively. After calculating the average accuracy of the test set, the ResNet50 network method, the label consistent sparse classification method, and the method of the present invention are tested. The results of each method are compared as follows:

[0090] Table 2

[0091] Bearing fault diagnosis methods Accuracy Recall F1 value ResNet50 network method 85.3％ 83.9％ 84.3％ Label Consistent Sparse Classification (LC-KSVD) method 86.3％ 86.1％ 85.2％ Method of the present invention 96.6％ 96.1％ 95.7％

[0092] As can be seen from Table 2, the accuracy of the method of the present invention reaches 96.6%, the recall rate reaches 96.1%, and the F1 value reaches 95.7%, which are all much higher than the values ​​​​achievable by the "ResNet50 network method" and the "label consistent sparse classification method". This shows that the bearing fault diagnosis method based on sparse classification of multi-layer composite domain adaptation provided by the present invention has high accuracy.

[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A bearing fault diagnosis method based on sparse classification of multi-layer composite domain adaptation, characterized by: The following steps are involved: S1, collecting vibration time domain signals of rolling bearings in different fault states under different operating conditions, filtering the collected time domain signals to remove noise in the time domain signals, performing wavelet transform on the denoised time domain signals, and obtaining a two-dimensional time-frequency graph corresponding to the time domain signals; The time domain signal of the rolling bearing with known fault status collected under known operating conditions constitutes the source domain, and the vibration time domain signal of the rolling bearing with unknown fault status collected under other operating conditions constitutes the target domain; S2, constructing a Coral-MMD multi-layer composite domain-adapted label-consistent sparse classification network; the Coral-MMD multi-layer composite domain-adapted label-consistent sparse classification network includes a deep feature extraction module, a domain distribution difference module and a label-consistent sparse classification module; S3: Take the source domain data with known fault states obtained in S1 and a portion of target domain data with unknown fault states as input, and use the fault state corresponding to each two-dimensional time-frequency graph in the source domain data as the expected output of the label-consistent sparse classification module. Determine the state classification loss value and classification accuracy based on the fault state label and predicted label corresponding to each two-dimensional time-frequency graph. The domain classification loss value is determined by calculating the maximum difference in the distribution means of the features of the source domain and the target domain and the distance between the feature covariances; The parameters of the Coral-MMD multi-layer composite domain-adapted label-consistent sparse classification network constructed in S2 are iteratively adjusted multiple times by using the state classification loss value and the domain classification loss value, thereby obtaining a trained Coral-MMD multi-layer composite domain difference label-consistent sparse classification network model; S4, input the remaining data in the target domain obtained in S1 into the trained Coral-MMD multi-layer composite domain adapted label consistent sparse classification network, and the label consistent sparse classification module obtains the state label corresponding to the input data to realize rolling bearing fault diagnosis under different working conditions.

2. The bearing fault diagnosis method based on sparse classification of multi-layer composite domain adaptation according to claim 1 is characterized by: In step 2, The deep feature extraction module is used to extract features from the two-dimensional time-frequency graphs of the source domain and the target domain. The domain distribution difference module is used to measure the maximum difference in the distribution means of the extracted features of the source domain and the target domain and the distance between the feature covariances; The label consistent sparse classification module is used to learn a dictionary with sparse discrimination capability and a linear classifier for the extracted features of the source domain, and classify the extracted features of the source domain and the target domain.

3. The bearing fault diagnosis method based on sparse classification of multi-layer composite domain adaptation according to claim 2 is characterized by: The deep feature extraction module adopts a customized multi-layer convolution residual block: including a first convolution module, a second convolution module, a first pooling layer, a first residual module, a second residual module, a third residual module, a fourth residual module, and a second pooling layer; The first convolution module and the second convolution module each include a convolution layer, a BN layer, and a ReLu layer; the number of output channels of the convolution layer is 64, the convolution kernel is 3*3, and the stride is 1; The outputs of the first convolution module and the second convolution module are calculated as follows: Among them, y represents the feature tensor after the convolution layer operation, f represents the nonlinear activation function, x represents the input feature tensor, W represents the weight of the convolution kernel, and B represents the bias coefficient; The first pooling layer is a 3*3 maximum pooling layer, and the step size of the first pooling layer is 2; The first residual module includes three residual blocks with 64 output channels; the second residual module includes six residual blocks with 128 output channels; the third residual module includes six residual blocks with 256 output channels; the fourth residual module includes three residual blocks with 512 output channels; Among them, each residual block consists of a 1*1 convolution layer with its own mapping, a 3*3 convolution layer and a 1*1 convolution layer; the feature representation of the output after the i-th residual block is: X i+1 =H i (X i )+X i Among them, X i represents the input features of the i-th residual block, X i+1 represents the output feature of the i-th residual block, H i Composite function for feature extraction of layer i, including BN, Relu, 1*1Conv2d, 3*3Conv2d; The second pooling layer is an average pooling layer.

4. The bearing fault diagnosis method based on sparse classification of multi-layer composite domain adaptation according to claim 2 is characterized by: The domain distribution difference module adopts the composite domain difference loss of Coral and MMD; The calculation formula of the composite domain difference loss is as follows: Loss=α*loss coral +β*loss MMD Among them, loss coral Represents Coral distance loss, loss MMD Represents MMD distance loss, α and β represent the weight coefficients of the two distance losses respectively; Coral distance loss loss coral The calculation formula is: Among them, D s and D t Represents source domain and target domain data respectively; n s and n t Represents the number of samples in the source domain and target domain respectively; C s and C t Represent the data covariance matrix of the source domain and the target domain respectively; represents the Frobenius norm, and d represents the number of dimensions of the sample feature space; MMD distance loss MMD The calculation formula is: in, and They represent the j-th source domain feature and the k-th target domain feature output by the i-th layer respectively; φ(·) represents the mapping function; ||·|| H represents the regenerated Hilbert space norm.

5. The bearing fault diagnosis method based on sparse classification of multi-layer composite domain adaptation according to claim 2 is characterized by: The label-consistent sparse classification module obtains a dictionary with combined discrimination and reconstruction criteria and an optimal multi-class linear classifier through learning; the features obtained by the feature extraction module and their corresponding category labels are used as inputs of the label-consistent sparse classification module for learning, thereby obtaining a sparse discriminant dictionary and a linear classifier for classifying and identifying the input data.

Images