A bearing composite fault diagnosis method based on multi-label domain adaptive model

Abstract

The present invention provides a method for diagnosing composite bearing faults in an unknown target domain under variable operating conditions, including: constructing a fault feature extractor composed of a deep residual network based on a multi-layer domain adaptive method; inputting a preprocessed bearing vibration signal, performing distribution difference matching on the features extracted from the source domain and target domain data through multiple residual blocks, and obtaining transferable features; expressing composite faults as a combination of single faults through multi-label learning; using a binary association strategy to train corresponding binary classifiers for each type of single fault, separating the features of each single fault from the composite fault and diagnosing them separately. The present invention solves the problem that traditional diagnostic schemes rely on expert knowledge and are difficult to effectively decouple and identify composite faults, and realizes accurate diagnosis of composite bearing faults under variable operating conditions. At the same time, it gets rid of the dependence of existing methods on labeled data, and can accurately diagnose related but invisible target domains.

Description

Technical Field

[0001] The present invention relates to the fields of mechanical fault diagnosis, artificial intelligence and signal processing, and in particular to a mechanical compound fault diagnosis method based on a multi-label domain adaptive model. Background Art

[0002] In the field of rolling bearing fault diagnosis, deep learning algorithms can intelligently mine the essential features related to bearing faults from vibration signals, thus achieving good diagnostic results. However, most current research focuses on single-category bearing faults. In actual situations, bearing faults often appear in the form of composite faults composed of multiple single faults. Composite faults usually occur when multiple key parts of rotating machinery bearings are damaged at the same time, and are also the main cause of catastrophic accidents in rotating machinery. Compared with single faults, composite faults are more harmful to equipment and more difficult to diagnose. Therefore, the research on bearing composite fault diagnosis is of great significance and great challenges.

[0003] At present, there are several major problems that restrict the development of bearing composite fault diagnosis technology: 1) The bearing signal of composite fault usually contains multiple characteristic signals and strong aliasing noise, which is often nonlinear and interferes with each other. It is difficult to accurately decouple them in theory, which can easily lead to misidentification in the diagnosis process. 2) The composite fault signal has no obvious periodicity, the amplitude changes greatly, the peak at the fault characteristic frequency is not obvious, and the periodic component is almost completely covered by the noise. Traditional time domain and frequency domain signal processing methods are difficult to extract the characteristics of composite faults, so it is difficult to work on composite fault diagnosis problems. 3) Existing methods often regard composite faults as a new fault mode, while ignoring the potential relationship between them and single faults. Composite faults are actually a combination of single faults. Such methods are difficult to accurately classify them into multiple single faults. 4) When facing actual engineering applications, the effect of the model often declines significantly due to the different distribution of the target data set to be diagnosed and the training data set. Existing deep learning-based diagnosis methods often rely on a large amount of labeled data, but collecting labeled data of various composite faults in the target field is time-consuming, labor-intensive, and even impractical. This is the biggest engineering problem that restricts data-driven composite fault diagnosis.

[0004] In order to solve the problem that compound fault features are difficult to extract and identify, this paper introduces a multi-label learning (ML) method to express and classify compound faults, which not only gets rid of the dependence of traditional signal processing methods on professional technology and fault diagnosis experience knowledge, but also solves the problem that existing deep learning methods ignore the connection between compound faults and single faults and lack rigor and reliability. In order to solve the problem of compound fault diagnosis relying on labeled data, this paper attempts to use the domain adaptation (DA) method in transfer learning to share knowledge between different but related fields, so as to achieve accurate diagnosis of unlabeled target data. Summary of the invention

[0005] Technical problem to be solved: The purpose of the present invention is to propose a bearing composite fault diagnosis method based on a multi-label domain adaptive model, which does not require complex signal decoupling technology and expert prior knowledge, and can automatically extract and classify the various single fault features contained in the composite fault signal. It can also adapt to the data distribution differences caused by changes in working conditions, and still maintain accurate diagnostic effects when facing related but invisible target domains.

[0006] Technical solution: A bearing composite fault diagnosis method based on a multi-label domain adaptive model, comprising the following steps:

[0007] Step 1: Collect and pre-process the vibration signals of the rotating machinery bearings under different working conditions, and convert them into frequency domain signals after Fast Fourier Transform (FFT);

[0008] Step 2: The labeled fault samples are used as the source domain data (Ds) for training the model, and the unlabeled samples under different working conditions are used as the target domain data (Dt) to be detected, so as to verify the effectiveness of the proposed method in diagnosing unlabeled target data;

[0009] Step 3: According to the characteristics of bearing fault diagnosis problem, a fault feature extractor with multiple residual block structures is constructed;

[0010] Step 4: Use Ds data to pre-train the network, save the network parameters of the pre-trained model to initialize the subsequent training model, and set the learning rate ξ and hyperparameter λ;

[0011] Step 5: Input the Ds and Dt samples into the pre-trained fault feature extractor, and extract transferable features by convolution layer by layer and measuring distribution differences;

[0012] Step 6: Based on the multi-kernel maximum mean difference (MK-MMD) combined with the subspace mapping of multiple kernel functions, the fault features are expressed more accurately, and the multi-layer domain adaptation strategy is used to match the distribution differences of the multi-layer fault features extracted from Ds and Dt;

[0013] Step 7: Use the binary relevance strategy in multi-label learning to build a multi-label classifier and perform multi-label transformation on the composite fault data: (x, y) → {(x, y j )|j=1,...,C}, where C represents the number of single fault categories;

[0014] Step 8: Convert the multi-label classification task into multiple binary classification problems, separate the single fault features from the composite fault, and train the corresponding binary classifiers for each type of single fault: h = (h 1 ,...,h C ):χ→{0,1} C , input the transferable features extracted by the feature extractor to obtain the classification result;

[0015] Step 9: Output the associated multi-label: And calculate the classification loss based on the binary cross entropy;

[0016] Step 10: Combine the classification loss with the distribution matching loss obtained by calculating the multi-layer MK-MMD to obtain the total loss function, and back-propagate it to update the weights, and use the Adam algorithm to accelerate model convergence and update model parameters;

[0017] Step 11: After the model training is completed, the diagnostic effect of the model is evaluated using the test set samples of Dt with labels.

[0018] Furthermore, the hyperparameter λ described in step 4 is a trade-off coefficient for adjusting the weight of the distribution matching loss.

[0019] Furthermore, the fault feature extractor described in step 5 is provided with 1 convolution layer for input, 1 ReLu activation layer and 4 residual units. In order to adapt to the size of the input bearing fault signal graph, the kernel size of the convolution layer in the network is set to 3×3.

[0020] Furthermore, the convolution operation in step 5 is to extract the fault features in the input Ds and Dt frequency domain images through the convolution calculation of the input layer, and output a 64-dimensional feature vector; the feature vector calculated and activated by the input layer is input into the residual block of the fault feature extractor, and F is used to represent the learned residual function. The calculation process of the features from the lth layer to the deeper Lth layer in the residual block is as follows:

[0021]

[0022] Furthermore, the MK-MMD described in step 6 measures the difference between corresponding distributions by calculating the distance between data of different distributions in the Reproducing Kernel Hilbert Space (RKHS). RKHS is a space composed of kernel functions k. For each RKHS, there is a unique k corresponding to it, and the kernel function directly affects the feature mapping ability of RKHS. The expression ability of a single kernel function is limited, and the MK-MMD method solves this problem well.

[0023] The kernel function k in MK-MMD is a positive definite function {k u}, represented by the weighted sum of multiple kernel functions:

[0024]

[0025] Let H k is the RKHS of the function f:X→R defined on X, φ:X→H k To map X to space H k The feature map on Ds is P(X S ) and Dt marginal distribution P(X T ) is as follows

[0026]

[0027] Furthermore, the multi-layer domain adaptation strategy described in step 6 is to save the feature maps output by Ds and Dt after each residual block in the feature extractor respectively; the loss function of the distribution matching part obtained by calculating MK-MMD on the features extracted from Ds and Dt in multiple layers is as follows:

[0028]

[0029] Where N l Represents the number of layers for calculating MK-MMD, and are the feature vectors of Ds and Dt extracted from the lth layer respectively, and They represent the number of samples in Ds and Dt respectively, and K represents the number of Gaussian kernels.

[0030] Furthermore, the multi-label conversion described in step 7 is as follows:

[0031] (x,y)→{(x,y j )|j=1,...,C}→{(x,[y1 ,y 2 ,...,y C ])} (5)

[0032] Where C represents the number of single fault categories. For the composite fault composed of the three common single faults of inner race fault (IF), outer race fault (OF) and ball fault (BF) in rolling bearings, C is 3. The label of the i-th fault can be expressed as a three-dimensional row vector y j =0 means the sample does not belong to this category, y j =1 indicates that the sample belongs to this category; each fault category corresponds to a position in the row vector, and the composite fault composed of IF and BF can be expressed as

[101] .

[0033] Furthermore, the task conversion described in step 8 is specifically implemented as follows: three neurons are set in the output layer of the multi-label classifier, corresponding to three single faults respectively; let the output of each neuron be s i , through the Sigmoid function, it is converted into the probability P of whether such a single fault exists i The specific expression is as follows:

[0034]

[0035] When the output probability is greater than 0.5, it is determined that a single fault of this category exists, and the predicted labels of each single fault are associated to obtain the predicted composite fault label.

[0036] Furthermore, the binary cross entropy classification loss expression in step 9 is as follows:

[0037]

[0038] In the formula, n represents the number of samples, y represents the true label, Represents the predicted label value.

[0039] Furthermore, the total loss function expression described in step 10 is as follows:

[0040]

[0041] Where N l represents the number of layers for calculating MK-MMD, K represents the number of Gaussian kernels, λ is the trade-off coefficient, φ:X→H k To map X to space H k The feature map on the , C represents the number of single fault categories, hj (·) is the mapping from the sample input space to the j-th single fault label space.

[0042] Beneficial effects:

[0043] (1) Traditional signal processing methods based on time domain and frequency domain are difficult to extract the characteristics of complex faults. The bearing signals of complex faults usually contain multiple characteristic signals and strong confusion noise, which are often nonlinear and interfere with each other. It is difficult to accurately decouple them in theory. Therefore, traditional signal processing methods are difficult to work for complex fault diagnosis problems. The present invention can automatically extract complex fault characteristics in a data-driven way and learn the mapping from input signals to fault categories, which significantly improves the diagnosis accuracy.

[0044] (2) Traditional solutions require expert knowledge of each type of composite fault characteristics as support, and can only effectively separate independent signal sources when the signal sources meet a series of statistical assumptions. Traditional methods require manual feature extraction and rely on professional knowledge and complex parameter adjustment processes, with high time and labor costs. The present invention introduces a multi-label learning method to express and classify composite faults, breaking away from the reliance of traditional signal processing methods on professional technology and fault diagnosis experience knowledge.

[0045] (3) Models constructed by existing methods are often only effective for specific problems and conditions and are difficult to adapt to changes in working conditions. The time and difficulty required for modeling increase with the complexity and functionality of the equipment. The present invention can adapt to changing working conditions and has better robustness.

[0046] (4) Existing deep learning-based composite fault diagnosis methods often regard composite faults as a new fault mode, while ignoring the potential relationship between composite faults and single faults. Composite faults are actually a combination of single faults. Such methods are difficult to accurately classify them into multiple single faults. This is a "pseudo" composite fault diagnosis that ignores the nature of composite fault characteristics. The present invention establishes the connection between composite faults and single faults, extracts various single fault features contained in the composite fault signal and classifies them, thereby improving the diagnostic accuracy of composite faults.

[0047] (5) Existing deep learning-based diagnostic methods often rely on a large amount of labeled data, but collecting labeled data for various types of composite faults in the target domain is time-consuming, laborious, and even impractical. When applied to actual engineering projects, the model effect often declines significantly due to the different distribution of the target data set to be diagnosed and the training data set. The present invention can accurately diagnose composite bearing faults under different working conditions, and at the same time realizes the diagnosis of unlabeled target domain data, greatly improving the practicality of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a composite fault diagnosis framework diagram based on the multi-label multi-layer domain adaptive method of the present invention;

[0049] Figure 2 A specific implementation method for comparing single fault signals and compound fault signals;

[0050] Figure 3 Based on Figure 1 The composite fault diagnosis model shown is used to evaluate the diagnostic accuracy of the model using data collected from the test bench. DETAILED DESCRIPTION

[0051] The specific implementation of the present invention is further described in detail below in conjunction with the drawings and examples.

[0052] Aiming at the problem of bearing composite fault diagnosis, this paper proposes a fault diagnosis framework with the ability to learn cross-domain transferable features based on transfer learning and multi-label learning. The specific structure is as follows: Figure 1 shown.

[0053] The bearing composite fault diagnosis method based on the multi-label domain adaptive model includes the following steps:

[0054] Step 1: Collect and pre-process the vibration signals of the rotating machinery bearings under different working conditions, and convert them into frequency domain signals after Fast Fourier Transform (FFT);

[0055] Step 2: The preprocessed frequency domain images are set as the source domain (Ds) and the target domain (Dt) according to different working conditions. The labels of the target data are invisible during the training process and are only used to evaluate the model effect during the test process, so as to verify the effectiveness of the proposed method in diagnosing unlabeled target data.

[0056] Step 3: According to the characteristics of bearing fault diagnosis problem, a fault feature extractor with multiple residual block structures is constructed;

[0057] Step 4: Use Ds data to pre-train the network, save the network parameters of the pre-trained model to initialize the subsequent training model, and set the learning rate ξ and hyperparameter λ, where the hyperparameter λ is the trade-off coefficient for adjusting the distribution matching loss weight;

[0058] Step 5: Input the Ds and Dt samples into the pre-trained fault feature extractor, and extract the transferable features through the convolution operation; wherein, the fault feature extractor is provided with 1 convolution layer for input, 1 ReLu activation layer and 4 residual units. In order to adapt to the size of the input bearing fault signal map, the kernel size of the convolution layer in the network is set to 3×3; the convolution operation is to extract the fault features in the input Ds and Dt frequency domain maps through the convolution calculation of the input layer, and output a 64-dimensional feature vector; the feature vector calculated and activated by the input layer is input into the residual block of the fault feature extractor, and F is used to represent the learned residual function. The calculation process of the features from the lth layer to the deeper Lth layer in the residual block is as follows:

[0059]

[0060] Step 6: Based on MK-MMD (Multiple Kernel-Maximum Mean Discrepancy) combined with the subspace mapping of multiple kernel functions, the fault features are expressed more accurately, and the multi-layer domain adaptive strategy is used to match the distribution differences of the multi-layer fault features extracted from Ds and Dt; the MK-MMD described in step 6 measures the difference between corresponding distributions by calculating the distance between data with different distributions in the Reproducing Kernel Hilbert Space (RKHS). RKHS is a space composed of kernel functions k. For each RKHS, there is a unique corresponding k, and the kernel function directly affects the feature mapping ability of RKHS; the expression ability of a single kernel function is limited, and the MK-MMD method solves this problem well;

[0061] The kernel function k in MK-MMD is a positive definite function {k u}, represented by the weighted sum of multiple kernel functions:

[0062]

[0063] Let H k is the RKHS of the function f:X→R defined on X, φ:X→H k To map X to space H k The feature map on Ds is P(X S ) and Dt marginal distribution P(X T ) is as follows

[0064]

[0065] The multi-layer domain adaptation strategy is to save the feature maps output by Ds and Dt after each residual block in the feature extractor respectively; the loss function of the distribution matching part obtained by calculating MK-MMD on the features extracted from Ds and Dt in multiple layers is as follows:

[0066]

[0067] Where N l Represents the number of layers for calculating MK-MMD, and are the feature vectors of Ds and Dt extracted from the lth layer respectively, and They represent the number of samples in Ds and Dt respectively, and K represents the number of Gaussian kernels;

[0068] Step 7: Use the binary association strategy in multi-label learning to build a multi-label classifier and perform multi-label transformation on the composite fault data. The multi-label transformation is as follows:

[0069] (x,y)→{(x,y j )|j=1,...,C}→{(x,[y 1 ,y 2 ,...,y C ])} (5)

[0070] Where C represents the number of single fault categories. For the composite fault composed of the three common single faults of inner race fault (IF), outer race fault (OF) and ball fault (BF) in rolling bearings, C is 3. The label of the i-th fault can be expressed as a three-dimensional row vector y j =0 means the sample does not belong to this category, y j =1 indicates that the sample belongs to this category; each fault category corresponds to a position in the row vector, and the composite fault composed of IF and BF can be expressed as

[101] ;

[0071] Step 8: Convert the task into multiple binary classification problems. Each binary classifier predicts a single fault label. Input the Ds feature extracted by the feature extractor and train multiple binary classifiers: h = (h 1 ,...,h C ):χ→{0,1} C The specific implementation of the task conversion is as follows: three neurons are set in the output layer of the multi-label classifier, corresponding to three single faults respectively; let the output of each neuron be s i, through the Sigmoid function, it is converted into the probability P of whether such a single fault exists i ; The specific expression is as follows:

[0072]

[0073] When the output probability is greater than 0.5, it is determined that a single fault of this category exists, and the predicted labels of each single fault are associated to obtain the predicted composite fault label;

[0074] Step 9: Output the associated multi-label: The classification loss is calculated based on the binary cross entropy; the binary cross entropy classification loss expression is as follows:

[0075]

[0076] In the formula, n represents the number of samples, y represents the true label, Represents the predicted label value;

[0077] Step 10: The total loss function is obtained by combining the classification loss and the distribution matching loss obtained by calculating the multi-layer MK-MMD, and back-propagating the weights to update the weights. The Adam algorithm is used to accelerate the model convergence and update the model parameters. The total loss function expression is as follows:

[0078]

[0079] Where N l represents the number of layers for calculating MK-MMD, K represents the number of Gaussian kernels, λ is the trade-off coefficient, φ:X→H k To map X to space H k The feature map on the , C represents the number of single fault categories, h j (·) is the mapping from the sample input space to the j-th single fault label space;

[0080] Step 11: After the model training is completed, the diagnostic effect of the model is evaluated using the test set samples of Dt with labels.

[0081] The implementation examples are as follows:

[0082] The present invention is used to diagnose the bearing fault data collected by the laboratory's self-made bearing fault experimental platform. The fault signal is as follows: Figure 2 As shown, the bearing model used in the experiment is 6205 deep groove ball bearing, and the sampling frequency is 10kHz.

[0083] The collected bearing vibration signals under four different motor loads (0kN, 1kN, 2kN, 3kN) are constructed into four data sets: C1, C2, C3, and C4. Each load includes 8 types of bearings with different health conditions, and the specific settings are shown in Table 1. They are one normal state (Normal), three single faults (IF, OF, BF) and four compound faults (IB, IO, OB, IOB), and the fault size is 0.2mm.

[0084] Table 1. Composite fault diagnosis dataset settings

[0085]

[0086] Different data sets are used as source domains and target domains respectively, and multiple groups of diagnostic experiments are carried out using the method of the present invention to reduce randomness. The experimental task settings are detailed in Table 2.

[0087] Table 2 Cross-domain fault diagnosis task settings

[0088]

[0089] Based on the method of the present invention and combined with the characteristics of specific diagnostic tasks, a Figure 1 The composite fault diagnosis model shown in the figure is used to evaluate the model effect using the data collected by the experimental bench. The specific experimental results are shown in Figure 3 shown.

[0090] The model can diagnose the presence or absence of various single faults through composite fault signals. The experimental results show that its average diagnostic accuracy is as high as 99.36%. The model calculates the MK-MMD distance for the extracted multi-layer features from low to high, achieving accurate matching of the distribution differences between the source domain and the target domain. Finally, the trained and optimized model can effectively extract transferable composite fault features, and can accurately diagnose the target domain data without labeled information when the working conditions change.

[0091] The known technologies in the art involved in the present invention are not described in detail.

[0092] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the technical principles of the present invention, and these improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A bearing composite fault diagnosis method based on a multi-label domain adaptive model, characterized by: The following steps are involved: Step 1: Collect and pre-process the vibration signals of the rotating machinery bearings under different working conditions, and convert them into frequency domain signals after fast Fourier transform; Step 2: Use labeled fault samples as source domain data for training the model, and unlabeled samples under different working conditions as target domain data to be detected, so as to verify the effectiveness of the proposed method in diagnosing unlabeled target data; Step 3: According to the characteristics of bearing fault diagnosis problem, a fault feature extractor with multiple residual block structures is constructed; Step 4: Use source domain data to pre-train the network and set the learning rate ξ and hyperparameter λ; Step 5: Input the source domain and target domain samples into the pre-trained fault feature extractor, extract transferable features through layer-by-layer convolution and measure distribution differences; Step 6: Based on the multi-core maximum mean difference and the subspace mapping of multiple kernel functions, the fault features are accurately expressed, and the multi-layer domain adaptation strategy is used to match the distribution differences of the multi-layer fault features extracted from the source domain and the target domain; Step 7: Use the binary association strategy in multi-label learning to build a multi-label classifier and perform multi-label transformation on the composite fault data: (x, y) → {(x, y j )j=1,...,C}, where C represents the number of single fault categories; Step 8: Convert the multi-label classification task into multiple binary classification problems, separate the single fault features from the composite fault, and train the corresponding binary classifiers for each type of single fault: h = (h1,...,h C ):χ→{0,1} C , input the transferable features extracted by the feature extractor to obtain the classification result; Step 9: Output the associated multi-label: And calculate the classification loss based on the binary cross entropy; Step 10: Combine the classification loss with the distribution matching loss obtained by calculating the maximum mean difference of multiple layers and multiple cores to obtain the total loss function, perform backpropagation to update the weights, and use the Adam algorithm to accelerate model convergence and update model parameters; Step 11: After the model training is completed, the diagnostic effect of the model is evaluated using the test set samples of the target domain.

2. The bearing composite fault diagnosis method based on a multi-label domain adaptive model according to claim 1 is characterized by: The hyperparameter λ described in step 4 is a trade-off coefficient for adjusting the weight of the distribution matching loss.

3. The bearing composite fault diagnosis method based on a multi-label domain adaptive model according to claim 1 is characterized in that: The fault feature extractor described in step 5 is provided with 1 convolution layer for input, 1 ReLu activation layer and 4 residual units. In order to adapt to the size of the input bearing fault signal map, the kernel size of the convolution layer in the network is set to 3×3.

4. The bearing composite fault diagnosis method based on a multi-label domain adaptive model according to claim 1 is characterized in that: The convolution operation in step 5 is to extract the fault features in the input source domain and target domain frequency domain images through the convolution calculation of the input layer, and output a 64-dimensional feature vector; the feature vector calculated and activated by the input layer is input into the residual block of the fault feature extractor, and F is used to represent the learned residual function. The calculation process of the features from the lth layer to the deeper Lth layer in the residual block is as follows:

5. The bearing composite fault diagnosis method based on a multi-label domain adaptive model according to claim 1 is characterized in that: The multi-core maximum mean difference described in step 6 measures the difference between corresponding distributions by calculating the distance between data of different distributions in the reproducing kernel Hilbert space. The reproducing kernel Hilbert space is a space composed of kernel functions k. Each reproducing kernel Hilbert space has a unique corresponding k, and the kernel function directly affects the feature mapping ability of the reproducing kernel Hilbert space. The expression ability of a single kernel function is limited, and the multi-core maximum mean difference method solves this problem well. The kernel function k in MK-MMD is a positive definite function {k u }, represented by the weighted sum of multiple kernel functions: Let H k is the reproducing kernel Hilbert space of the function f:X→R defined on X, φ:X→H k To map X to space H k The feature map on the source domain, then the marginal distribution P(X S ) and the marginal distribution P(X T ) are as follows:

6. The bearing composite fault diagnosis method based on a multi-label domain adaptive model according to claim 1 is characterized by: The multi-layer domain adaptation strategy described in step 6 is to save the feature maps output by the source domain and the target domain after each residual block in the feature extractor respectively; the loss function of the distribution matching part obtained by calculating the multi-core maximum mean difference of the features extracted from the source domain and the target domain in multiple layers is as follows: Where N l Represents the number of layers for calculating the maximum mean difference of multiple cores, and are the feature vectors of the source domain and the target domain extracted in the lth layer, respectively. and They represent the number of samples in the source domain and the target domain respectively, and K represents the number of Gaussian kernels.

7. The bearing composite fault diagnosis method based on a multi-label domain adaptive model according to claim 1 is characterized in that: The multi-label transformation described in step 7 is as follows: (x,y)→{(x,y j )j=1,...,C}→{(x,[y 1 ,and 2 ,...,and C ])} Where C represents the number of single fault categories. For a composite fault consisting of three common single faults in rolling bearings, namely inner ring fault, outer ring fault and rolling element fault, C is 3. The label of the i-th fault can be expressed as a three-dimensional row vector y j =0 means the sample does not belong to this category, y j =1 indicates that the sample belongs to this category; each fault category corresponds to a position in the row vector, and the composite fault composed of the inner race fault and the rolling element fault can be expressed as [101].

8. The bearing composite fault diagnosis method based on a multi-label domain adaptive model according to claim 1 is characterized in that: The specific implementation of the task transformation described in step 8 is as follows: three neurons are set in the output layer of the multi-label classifier, corresponding to the three single faults respectively; let the output of each neuron be s i , through the Sigmoid function, it is converted into the probability P of whether such a single fault exists i , the specific expression is as follows: When the output probability is greater than 0.5, it is determined that a single fault of this category exists, and the predicted labels of each single fault are associated to obtain the predicted composite fault label.

9. The bearing composite fault diagnosis method based on a multi-label domain adaptive model according to claim 1 is characterized in that: The binary cross entropy classification loss expression in step 9 is as follows: In the formula, n represents the number of samples, y represents the true label, Represents the predicted label value.

10. The bearing composite fault diagnosis method based on a multi-label domain adaptive model according to claim 1, characterized in that: The total loss function expression described in step 10 is as follows: Where N l represents the number of layers for calculating the maximum mean difference of multiple cores, K represents the number of Gaussian cores, λ is the trade-off coefficient, φ:X→H k To map X to space H k The feature map on the , C represents the number of single fault categories, h j (·) is the mapping from the sample input space to the j-th single fault label space.

Images