Cross-device domain adaptation method based on domain decoupling and class confusion minimization feature alignment
By using a feature alignment method that minimizes domain decoupling and class confusion, features of the source and target domains are separated and constrained, solving the problem of poor feature alignment in cross-device scenarios in existing technologies and achieving highly accurate fault diagnosis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHONGQING UNIV OF POSTS & TELECOMM
- Filing Date
- 2024-01-22
- Publication Date
- 2026-07-03
AI Technical Summary
Existing domain adaptation methods cannot effectively solve the problem of classifier accuracy in the target domain after feature alignment in cross-device scenarios, and the redundant information contained in the features affects the alignment effect.
We employ a feature alignment method based on domain decoupling and class confusion minimization. By using a feature extractor and decoder with convolutional channel separation, we separate the domain-specific features and domain-invariant features of the source and target domains. Furthermore, we constrain the model's learning behavior through sample reconstruction loss, conditional adversarial domain adaptation loss, and classification loss, thereby achieving feature decoupling and minimizing class confusion.
It improves the model's classification accuracy and feature representation capabilities in the target domain, and enhances the fault diagnosis effect of cross-device domain adaptation.
Smart Images
Figure CN118035783B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of domain adaptation fault diagnosis, specifically involving a cross-device domain adaptation method based on domain decoupling and class confusion minimization feature alignment. Background Technology
[0002] Rotating machinery plays a crucial role in modern industry. The complex and variable operating conditions make key components such as bearings, gears, and motors prone to failure, leading to equipment damage and even major accidents. Therefore, conducting fault diagnosis of key components in rotating machinery can improve the operational reliability of equipment and has significant engineering implications.
[0003] Deep learning-based intelligent fault diagnosis methods can reduce the reliance on expert knowledge in traditional diagnostic methods and achieve efficient end-to-end diagnosis as well as high-precision diagnosis of complex systems. Therefore, various deep learning methods are widely used in the intelligent diagnosis of key components of rotating machinery. However, traditional deep learning methods require a large number of labeled samples and satisfy the condition that the training and test data follow independent and identically distributed rules to support sufficient training of the supervised learning diagnostic model and high-precision fault diagnosis. This makes them unsuitable for real-world engineering scenarios where target equipment has a large amount of data that is completely unlabeled.
[0004] Domain adaptation methods utilize source domain data with abundant labels, aligning features extracted from the source and target domains through discrepancy metrics and adversarial training. This transfers the mapping established in the source domain to the target domain, enabling the creation of a high-precision fault diagnosis model for the target domain data. However, existing domain adaptation methods only constrain the consistency of source and target domain features in terms of distribution and statistical characteristics, neglecting the accuracy of the classifier's classification behavior in the target domain after feature alignment. This leads to significant confusion when training the model to classify target domain samples in cross-device scenarios. Furthermore, existing domain adaptation methods directly align features extracted by the feature extractor as domain-invariant features, resulting in features containing a large amount of redundant information that negatively impacts feature alignment performance in cross-device scenarios. Summary of the Invention
[0005] To address the aforementioned problems, this invention discloses a cross-device domain adaptation method based on domain decoupling and class confusion minimization feature alignment. The technical solution of this invention is as follows:
[0006] A cross-device domain adaptation method based on domain decoupling and class confusion minimization feature alignment includes the following steps:
[0007] Step 1: Collect labeled vibration signals from other devices and unlabeled vibration signals from the target device to construct labeled source domain and unlabeled target domain datasets. Use the labeled source domain dataset as the training set and divide the unlabeled target domain dataset into training set and test set.
[0008] Step 2: Construct a domain adaptation model, which includes a feature extractor that separates convolutional channels, a decoder that fuses convolutional channels, a classifier, and a domain discriminator;
[0009] Step 3: The model training process constrains the model's learning behavior through sample reconstruction loss, conditional adversarial domain adaptation loss, classification loss, and class confusion minimization loss. The model decouples features into domain-specific features and domain-invariant features to enhance the extraction effect of domain-invariant features, aligns domain-invariant features, and minimizes inter-class confusion in the target domain.
[0010] Step 4: Use the target domain test set data to test the trained model and obtain the fault diagnosis results.
[0011] Further, step 1 is specifically as follows: Vibration signals from other devices and the target device are collected. After obtaining time-domain data, downsampling and sliding window operations are performed to construct a large number of time-domain samples containing rich information. The obtained samples are then processed using a max-min normalization method. Based on this, a labeled source domain training set is constructed. Unlabeled target domain training set With test set D test ,in and These represent single source domain samples and target domain samples, respectively. j The labels are the corresponding labels for the source domain samples.
[0012] Furthermore, in step 2, the domain adaptation model is constructed as follows:
[0013] The model includes a feature extractor that separates convolutional channels, a decoder that fuses convolutional channels, a classifier, and a domain discriminator. The feature extractor extracts and separates source domain samples. Source domain-specific features Domain-invariant features of the source domain and target domain samples Target domain-specific features Domain-invariant features of the target domain The decoder fuses domain-specific features with domain-invariant features and reconstructs the input sample. The domain discriminator aligns the domain-invariant features of the source and target domains. Based on this, the classifier identifies the fault type of the source domain's domain-invariant features, and thus identifies the fault type of the target domain's domain-invariant features aligned with the source domain features.
[0014] The encoder includes a feature extraction module and a feature branching module. The feature extraction module consists of several convolutional blocks, each containing several convolutional layers, instance normalization layers, non-linear activation function operations, and pooling layers. The feature branching module consists of one convolutional block and one branched convolutional block. This convolutional block outputs multi-channel features and then separates the features into three branches according to the channels, connecting them to the branch convolutional blocks. Each convolutional block contains several convolutional layers, batch normalization layers, and non-linear activation function operations. The branches are source domain-specific feature branches, domain-invariant feature branches, and target domain-specific feature branches. Source domain samples pass through the source domain-specific feature branches and the domain-invariant feature branches, while target domain samples pass through the target domain-invariant feature branches and the target domain-specific feature branches.
[0015] The decoder is divided into source domain decoder and target domain decoder with consistent structure. Each decoder consists of several deconvolution blocks and one fully connected layer. Domain-invariant features and domain-specific features are concatenated by channel and then input into the decoder. Each deconvolution block contains several upsampling operations, deconvolution layers, normalization layers and nonlinear activation function operations. The first few normalization layers are batch normalization, and the last few normalization layers are instance normalization.
[0016] The domain discriminator consists of one Dropout layer and several fully connected layers. The fully connected layers contain non-linear activation functions. The input consists of domain-invariant features of the source and target domains. The features are aligned through adversarial training.
[0017] The classifier consists of one Dropout layer and several fully connected layers. The fully connected layers contain non-linear activation functions. The input is domain-invariant features from the source domain, which are used to achieve pattern recognition of the input samples.
[0018] Furthermore, in step 3, the training process of the domain adaptation model is as follows:
[0019] The labeled source domain and unlabeled target domain training sets obtained in step 1 are used to adapt the model for training. The model training process is based on the sample reconstruction loss L. R Conditional Adaptive Loss L CA Classification loss L CLS Minimize loss L with class confusion MC The loss consists of four types.
[0020] Source domain samples With target domain samples After the feature extraction module of the encoder, the source domain-specific features are separated. Domain-invariant features of the source domain and the unique characteristics of the target domain Domain-invariant features of the target domain Source domain decoder fusion feature F S_CWith F S_I And decode to obtain the reconstructed source domain sample Target domain decoder fusion feature F T_C With F T_I And decode to obtain the reconstructed target domain sample The source domain reconstruction loss L is obtained through the mean square error. R_S Reconstruction loss L from the target domain R_T :
[0021]
[0022]
[0023] Obtain the complete reconstruction loss L R :
[0024]
[0025] The model achieves decoupling between domain-invariant and domain-specific features through feature separation-combination operations and minimizing the reconstruction loss. Based on this, a joint classifier G is used. c (·) and the domain discriminator G d (·) Calculate the conditional adversarial adaptation loss L CA :
[0026]
[0027] In the formula, g represents the classifier prediction result. To predict uncertainty, w(H(g)) = 1 + e -H(g) For entropy weights.
[0028] Based on feature alignment, the classifier can initially establish a mapping from features to patterns by completing the classification task on the labeled training set in the source domain. The classification loss L CLS Using cross-entropy loss:
[0029]
[0030] In the formula, y j,c With p i,c These are the sample's true label and the sample's label value and predicted probability in category c, respectively.
[0031] Calculating the loss by minimizing class confusion based on the prediction results of the class classifier on the target domain training set can further constrain the model to complete the classification task in the target domain. For the i-th sample in the target domain The probability that it belongs to the k-th category is:
[0032]
[0033] In the formula, T is the scaling factor, and Z is the result output by the classifier. For a sample Its uncertainty is calculated using information entropy:
[0034]
[0035] The importance probability of class confusion is obtained as follows:
[0036]
[0037] The loss for minimizing class confusion is calculated as follows:
[0038]
[0039]
[0040] Loss L of the domain adaptation model ALL The calculations based on all the above losses are as follows:
[0041] L ALL =L CLS +λ1L R +λ2L CA +λ3L MC (11)
[0042] In the formula, λ1, λ2 and λ3 are all hyperparameters that balance the various losses. By setting each parameter, all losses eventually converge to complete the model training process.
[0043] The beneficial technical effects of this invention are:
[0044] This invention, based on feature alignment achieved through conditional domain adaptation loss, considers the accuracy of the classifier's classification behavior on the target domain data, minimizing class confusion to further improve the model's classification accuracy in the target domain. Furthermore, by decoupling domain-specific features from domain-invariant fault features, it effectively enhances the ability to represent fault features across device conditions, further improving the domain-adaptive feature alignment effect and enabling cross-device domain-adaptive fault diagnosis. Attached Figure Description
[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention:
[0046] Figure 1 This is a flowchart of the steps of the cross-device domain adaptation method based on domain decoupling and class confusion minimization feature alignment of the present invention;
[0047] Figure 2 This is a schematic diagram of the domain adaptation model of the present invention; Detailed Implementation
[0048] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.
[0049] See Figure 1 As shown, this invention provides a semi-supervised rotating machinery fault diagnosis method based on limit marking of an adversarial decoupling autoencoder, which mainly includes the following steps:
[0050] Step 1: Collect vibration signals from other devices and the target device. After obtaining the time-domain data, perform downsampling and sliding window operations to construct a large number of time-domain samples containing rich information. Process the obtained samples using a max-min normalization method. Based on this, construct a labeled source domain training set. Unlabeled target domain training set With test set D test ,in and These represent single source domain samples and target domain samples, respectively. j The labels are the corresponding labels for the source domain samples.
[0051] Step 2: Construct a domain adaptation model, which includes a feature extractor that separates convolutional channels, a decoder that fuses convolutional channels, a classifier, and a domain discriminator. The feature extractor extracts and separates source domain samples. Source domain-specific features Domain-invariant features of the source domain and target domain samples Target domain-specific features Domain-invariant features of the target domain The decoder fuses domain-specific features with domain-invariant features and reconstructs the input sample. The domain discriminator aligns the domain-invariant features of the source and target domains. Based on this, the classifier identifies the fault type of the source domain's domain-invariant features, and thus identifies the fault type of the target domain's domain-invariant features aligned with the source domain features.
[0052] The encoder includes a feature extraction module and a feature branching module. The feature extraction module consists of several convolutional blocks, each containing several convolutional layers, instance normalization layers, non-linear activation function operations, and pooling layers. The feature branching module consists of one convolutional block and one branched convolutional block. This convolutional block outputs multi-channel features and then separates the features into three branches according to the channels, connecting them to the branch convolutional blocks. Each convolutional block contains several convolutional layers, batch normalization layers, and non-linear activation function operations. The branches are source domain-specific feature branches, domain-invariant feature branches, and target domain-specific feature branches. Source domain samples pass through the source domain-specific feature branches and the domain-invariant feature branches, while target domain samples pass through the target domain-invariant feature branches and the target domain-specific feature branches.
[0053] The decoder is divided into source domain decoder and target domain decoder with consistent structure. Each decoder consists of several deconvolution blocks and one fully connected layer. Domain-invariant features and domain-specific features are concatenated by channel and then input into the decoder. Each deconvolution block contains several upsampling operations, deconvolution layers, normalization layers and nonlinear activation function operations. The first few normalization layers are batch normalization, and the last few normalization layers are instance normalization.
[0054] The domain discriminator consists of one Dropout layer and several fully connected layers. The fully connected layers contain non-linear activation functions. The input consists of domain-invariant features of the source and target domains. The features are aligned through adversarial training.
[0055] The classifier consists of one Dropout layer and several fully connected layers. The fully connected layers contain non-linear activation functions. The input is domain-invariant features from the source domain, which are used to achieve pattern recognition of the input samples.
[0056] Step 3: Train the model using the input domain and the training sets of labeled source domain and unlabeled target domain. The model training process is based on the sample reconstruction loss L. R Conditional Adaptive Loss L CA Classification loss L CLS Minimize loss L with class confusion MC The loss consists of four types.
[0057] Source domain samples With target domain samples After the feature extraction module of the encoder, the source domain-specific features are separated. Domain-invariant features of the source domain and the unique characteristics of the target domain Domain-invariant features of the target domain Source domain decoder fusion feature F S_C With F S_I And decode to obtain the reconstructed source domain sample Target domain decoder fusion feature FT_C With F T_I And decode to obtain the reconstructed target domain sample The source domain reconstruction loss L is obtained through the mean square error. R_S Reconstruction loss L from the target domain R_T :
[0058]
[0059]
[0060] Obtain the complete reconstruction loss L R :
[0061]
[0062] The model achieves decoupling between domain-invariant and domain-specific features through feature separation-combination operations and minimizing the reconstruction loss. Based on this, a joint classifier G is used. c (·) and the domain discriminator G d (·) Calculate the conditional adversarial adaptation loss L CA :
[0063]
[0064] In the formula, g represents the classifier prediction result. To predict uncertainty, w(H(g)) = 1 + e -H(g) For entropy weights.
[0065] Based on feature alignment, the classifier can initially establish a mapping from features to patterns by completing the classification task on the labeled training set in the source domain. The classification loss L CLS Using cross-entropy loss:
[0066]
[0067] In the formula, y j,c With p i,c These are the sample's true label and the sample's label value and predicted probability in category c, respectively.
[0068] Calculating the loss by minimizing class confusion based on the prediction results of the class classifier on the target domain training set can further constrain the model to complete the classification task in the target domain. For the i-th sample in the target domain The probability that it belongs to the k-th category is:
[0069]
[0070] In the formula, T is the scaling factor, and Z is the result output by the classifier. For a sample Its uncertainty is calculated using information entropy:
[0071]
[0072] The importance probability of class confusion is obtained as follows:
[0073]
[0074] The loss for minimizing class confusion is calculated as follows:
[0075]
[0076]
[0077] Loss L of the domain adaptation model ALL The calculations based on all the above losses are as follows:
[0078] L ALL =L CLS +λ1L R +λ2L CA +λ3L MC (11)
[0079] In the formula, λ1, λ2 and λ3 are all hyperparameters that balance the various losses. By setting each parameter, all losses eventually converge to complete the model training process.
[0080] Step 4: Use the target domain test set data to test the trained model and obtain the fault diagnosis results.
[0081] Example 1:
[0082] In Example 1, the following two bearing datasets were used for method verification:
[0083] Laboratory Dataset: This dataset was generated by a laboratory bearing health monitoring experimental platform, which consists of a drive motor, test bearings, accelerometers, force gauges, and a data acquisition system. The sampling frequency was 10kHz, and the motor speed was set to 896rpm. The bearing used in the experiment was an SKF6205, and faults were simulated on the bearing surface using wire cutting.
[0084] CWRU Dataset: This is a publicly available dataset generated by a laboratory bearing health monitoring experimental platform. The platform includes a 2-horsepower electric motor, torque sensor, power meter, electronic controller, and the bearing under test. The experiment used a drive-end bearing, with a sampling frequency of 12kHz and a motor speed set between 1797 and 1720 rpm. The bearing used was an SKF6205, and faults were simulated on its surface through electrical discharge machining (EDM).
[0085] Table 1 provides a detailed description of the dataset.
[0086] Table 1 provides a detailed description of each bearing dataset.
[0087]
[0088] The method of this embodiment for fault diagnosis of the target domain includes the following steps:
[0089] Step 1: Perform sliding window and downsampling operations on the time-domain data of the laboratory dataset and the CWRU dataset to construct a large number of time-domain samples containing rich information. Process the resulting samples using max-min normalization. The sliding window length for the laboratory dataset is 3072, and the sliding window length for the CWRU dataset is 2048. The sample length for both datasets is downsampled to 1024. Based on this, a labeled source domain training set, an unlabeled target domain training set, and a test set are constructed. Each training set contains 500 samples per class, and each test set contains 100 samples per class.
[0090] Step 2: Construct a domain adaptation model, which includes a feature extractor that separates convolutional channels, a decoder that fuses convolutional channels, a classifier, and a domain discriminator.
[0091] The encoder includes a feature extraction module and a feature branching module. The feature extraction module consists of four convolutional blocks, each containing several convolutional layers, instance normalization layers, non-linear activation function operations, and pooling layers. The feature branching module consists of one convolutional block and one branched convolutional block. After outputting multi-channel features, the convolutional block separates the features into three branches according to the channels and connects them to the branch convolutional blocks. Each convolutional block contains several convolutional layers, batch normalization layers, and non-linear activation function operations. The branches are source domain-specific feature branches, domain-invariant feature branches, and target domain-specific feature branches.
[0092] The decoder is divided into source domain decoder and target domain decoder with consistent structure. Each decoder consists of 4 deconvolution blocks and 1 fully connected layer. Domain-invariant features and domain-specific features are concatenated by channel and then input into the decoder. Each deconvolution block contains several upsampling operations, deconvolution layers, normalization layers and nonlinear activation function operations. The first few normalization layers are batch normalization, and the last few normalization layers are instance normalization.
[0093] The domain discriminator consists of one Dropout layer and three fully connected layers. The fully connected layers contain non-linear activation functions. The input consists of domain-invariant features of the source and target domains. The features are aligned through adversarial training.
[0094] The classifier consists of one Dropout layer and three fully connected layers. The fully connected layers contain non-linear activation functions. The input is source domain invariant features, which are used to achieve pattern recognition of input samples.
[0095] Step 3: Reconstruct the loss L from the samples R Conditional Adaptive Loss L CA Classification loss L CLS Minimize loss L with class confusion MC Four loss methods yield the domain adaptation model loss L ALL =L CLS +λ1L R +λ2L CA +λ3L MC All hyperparameters were set to 1 to balance the various losses. The optimizer was SGD, with the number of epochs, learning rate, and batch size set to 100, 0.05, and 500, respectively. The learning rate was set according to... The loss function decays, where p is the number of iterations. The desired model is obtained after all losses converge.
[0096] Step 4: Use the target domain test set data to test the trained model and obtain the fault diagnosis results.
[0097] To demonstrate the effectiveness of the proposed method, the mainstream Deep Adversarial Domain Adaptation (DANN) method was selected for experimental comparison. Cross-device domain adaptation tasks A1-B1, A1-B2, A2-B1, A2-B2, B1-A1, B1-A2, B2-A1, and B2-A2 were constructed, with the symbols before and after "-" representing source and target domain data, respectively. The experimental results of the proposed method and DANN are shown in Table 2. It can be seen that the accuracy of the proposed method is significantly higher than that of DANN in all tasks, with an average accuracy of 98.42%, while DANN's is only 33.68%. The experimental results show that the proposed method can achieve high-accuracy cross-device domain adaptive fault diagnosis and is far superior to the mainstream DANN method.
[0098] Table 2. Comparison of accuracy (%) of the proposed method and DANN for various cross-device diagnostic tasks.
[0099]
Claims
1. A cross-device domain adaptation method based on domain decoupling and class confusion minimization feature alignment, characterized in that, Includes the following steps: Step 1: Collect labeled vibration signals from other devices and unlabeled vibration signals from the target device to construct labeled source domain datasets and unlabeled target domain datasets. Use the labeled source domain dataset as the labeled source domain training set and divide the unlabeled target domain dataset into an unlabeled target domain training set and an unlabeled target domain test set. Step 2: Construct a domain adaptation model, which includes a feature extractor that separates convolutional channels, a decoder that fuses convolutional channels, a classifier, and a domain discriminator; Step 3: The model training process constrains the model's learning behavior through sample reconstruction loss, conditional adversarial domain adaptation loss, classification loss, and class confusion minimization loss. The model decouples features into domain-specific features and domain-invariant features to enhance the extraction effect of domain-invariant features, aligns domain-invariant features, and simultaneously minimizes inter-class confusion in the target domain. In Step 3, the training process of the domain adaptation model is as follows: The labeled source domain training set and the unlabeled target domain training set obtained in step 1 are used to train the model. The model training process is based on the sample reconstruction loss. Conditional Adaptive Loss Classification loss Minimize loss with class confusion The loss consists of four types; Labeled source domain training set training set with unlabeled target domain After the feature extraction module of the feature extractor, the source domain-specific features are separated. Domain-invariant features of the source domain and the unique characteristics of the target domain. Domain-invariant features of the target domain Source domain decoder fusion features and And decode to obtain the reconstructed source domain sample Target domain decoder fusion features and And decode to obtain the reconstructed target domain sample ; The source domain reconstruction loss is obtained through mean squared error. Reconstruction loss with target domain : Obtain the complete reconstruction loss : The model decouples domain-invariant features from domain-specific features through feature separation-combination operations and minimizing reconstruction loss; based on this, a joint classifier is used. Domain discriminator Calculate the conditional adversarial domain adaptation loss : In the formula, g represents the classifier prediction result. To predict uncertainty, Entropy weights; Based on feature alignment, the classifier can initially establish a mapping from features to patterns by completing a classification task on a labeled source domain training set, and the classification loss... Using cross-entropy loss: In the formula, Let c be the true label value of the j-th source domain sample in the labeled source domain training set. Let be the predicted probability of the j-th source domain sample in the labeled source domain training set in class c; Calculating the loss by minimizing class confusion based on the classifier's prediction results on the unlabeled target domain training set can further constrain the model to complete the classification task in the target domain; for the i-th sample in the target domain The probability that it belongs to the k-th category is: In the formula, T is the scaling factor, and Z is the result output by the classifier; for a sample Its uncertainty is calculated using information entropy: Obtain the importance probability of class confusion discrimination: The loss for minimizing class confusion is calculated as follows: Loss of the domain adaptation model The losses are calculated based on the above: In the formula, , and These are all hyperparameters that balance the various losses. By setting each parameter, all losses eventually converge to complete the model training process. Step 4: Test the trained model using unlabeled target domain test set data to obtain fault diagnosis results.
2. The cross-device domain adaptation method based on domain decoupling and class confusion minimization feature alignment according to claim 1, characterized in that: Step 1 is specifically as follows: Vibration signals from other devices and the target device are collected. After obtaining time-domain data, downsampling and sliding window operations are performed to construct a large number of time-domain samples containing rich information. The obtained samples are then processed using max-min normalization. Based on this, a labeled source domain training set is constructed. Unlabeled target domain training set test set with unlabeled target domain ,in and These are single source domain samples and target domain samples, respectively. The labels are the corresponding labels for the source domain samples.
3. The cross-device domain adaptation method based on domain decoupling and class confusion minimization feature alignment according to claim 1, characterized in that: In step 2, the domain adaptation model is constructed as follows: The labeled source domain training set is extracted and separated using a feature extractor. Source domain-specific features of source domain samples Domain-invariant features of the source domain and unlabeled target domain training set Target domain-specific features of the target domain samples Domain-invariant features of the target domain ; The decoder fuses domain-specific features and domain-invariant features to reconstruct the input sample; the domain discriminator aligns the domain-invariant features of the source and target domains. Based on this, the classifier determines the fault type of the source domain invariant features, thereby realizing the fault type determination of the target domain invariant features aligned with the source domain features; The feature extractor includes a feature extraction module and a feature branching module; The feature extraction module consists of several convolutional blocks, each containing several convolutional layers, several instance normalization layers, several non-linear activation function operations, and several pooling layers. The feature branching module consists of one convolutional block and one branch convolutional block. After the convolutional blocks of the feature extraction module output multi-channel features, the features are separated into three branches according to the channels and connected to the branch convolutional blocks. Each branch convolutional block of the feature branching module contains several convolutional layers, several batch normalization layers, and several non-linear activation function operations. The branches are source domain-specific feature branches, domain-invariant feature branches, and target domain-specific feature branches. Source domain samples in the labeled source domain training set are processed through the source domain-specific feature branches and the domain-invariant feature branches, while target domain samples in the unlabeled target domain training set are processed through the domain-invariant feature branches and the target domain-specific feature branches. The decoder is divided into source domain decoder and target domain decoder with consistent structure. Each decoder consists of several deconvolution blocks and one fully connected layer. Domain-invariant features and domain-specific features are concatenated by channel and then input into the decoder. Each deconvolution block contains several upsampling operations, several deconvolution layers, several normalization layers and several nonlinear activation function operations. The first few normalization layers are batch normalization, and the last few normalization layers are instance normalization. The domain discriminator consists of one Dropout layer and several fully connected layers. The fully connected layers contain non-linear activation functions. The input consists of domain-invariant features of the source domain and domain-invariant features of the target domain. The features are aligned through adversarial training. The classifier consists of one Dropout layer and several fully connected layers. The fully connected layers contain non-linear activation functions. The input is domain-invariant features from the source domain, which are used to achieve pattern recognition of the input samples.