An instance-based weighted feature matching method for cross-domain fault diagnosis in an adaptive adversarial field

Abstract

An adversarial domain adaptive cross-domain fault diagnosis method based on instance-weighted feature matching is presented, belonging to the field of rotating machinery fault diagnosis. First, data is collected, a transfer task is constructed, and the data is partitioned to obtain source domain training samples, target domain training samples, and target domain test samples. Second, an adversarial domain adaptive network model based on instance-weighted feature matching is constructed. Third, the training process is implemented by minimizing a comprehensive loss function, ultimately obtaining the trained adversarial domain adaptive network model. Finally, fault diagnosis and performance evaluation are performed. This invention is the first to simultaneously consider the alignment of marginal and conditional distributions in adversarial domain adaptation, effectively solving the negative transfer problem caused by class misalignment and significantly improving the accuracy and robustness of cross-domain diagnosis. Furthermore, end-to-end reliable training accelerates model convergence and improves final performance. It is independent of specific network structures and has good versatility and promotional value.

Description

Technical Field

[0001] This invention belongs to the field of rotating machinery fault diagnosis and relates to an adversarial domain adaptive cross-domain fault diagnosis method based on instance-weighted feature matching. Background Technology

[0002] Rotating machinery typically operates for extended periods under high-speed, heavy-load, and high-power conditions, and rolling bearings, as core components, are prone to frequent failures. Therefore, there is an urgent need for a diagnostic method capable of effectively analyzing massive amounts of data and automatically providing accurate diagnoses. Deep learning's adaptability to massive data and its excellent feature learning capabilities have prompted researchers to propose intelligent fault diagnosis methods based on deep models. However, these methods require two conditions to be met in practical fault diagnosis: first, sufficient labeled samples are needed to train the deep model; second, the training and test data need to be independent and identically distributed. However, in real-world scenarios, these conditions are difficult to meet. On the one hand, the high cost of labeled samples results in only a large number of unlabeled samples or a small number of labeled samples. On the other hand, the data distribution difference between training and test data means that even if the model is sufficiently trained on the training data, it can still lead to poor performance on the test data.

[0003] Therefore, intelligent fault diagnosis based on transfer learning has gradually attracted the attention and research of scholars. Transfer learning applies knowledge or patterns learned in one domain or task to another related but different domain or problem. Domain adaptation, as an important method of transfer learning, focuses on solving the problem of inconsistent data distribution in unlabeled or partially labeled data. These methods can be divided into traditional shallow methods and deep model-based methods. Shallow domain adaptation methods typically reduce the differences in shallow features by minimizing the distance metric of cross-domain differences, thereby achieving task transfer from the source domain to the target domain. However, due to the limited representational power of shallow features, the transfer effect on different specific tasks is poor. Therefore, the feature learning capabilities of deep neural networks are needed to learn task-specific deep transferable features. Generally, deep neural networks can learn transfer feature representations and match cross-domain feature distributions through difference measurements; or they can assist the target domain task through data reconstruction, learning invariant feature representations to achieve two-domain transfer. The idea of ​​generative adversarial networks can also be used to learn domain-invariant features for domain adaptation by adversarially challenging a discriminator and using feature extractors to align source and target domain data in the data space or feature space. However, adversarial methods typically only consider global feature adaptation, ignoring the relationships between data categories and only aligning the marginal distributions of data in the source and target domains. Therefore, when the data in the two domains differ significantly, it can lead to misalignment of categories in the source and target domains, and may even result in negative transfer, severely impacting the effectiveness of domain adaptation.

[0004] Based on the above discussion, it is necessary to establish an adversarial cross-domain diagnostic method that can adapt to both edge distribution and conditional distribution in both the source and target domains for use in rotating machinery faults. Summary of the Invention

[0005] To address the shortcomings of existing adversarial domain adaptation methods in cross-domain fault diagnosis, this invention proposes an instance-weighted feature matching-based adversarial domain adaptation (IWFM-ADA) method. This method introduces multilinear mappings to fuse category information and employs an entropy-based instance weighting strategy to achieve joint distribution adaptation, thereby improving the model's fault classification accuracy in the target domain.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: An adversarial domain adaptive cross-domain fault diagnosis method based on instance-weighted feature matching includes the following steps: Step 1: Collect data, construct the transfer task, partition the data, and finally obtain source domain training samples, target domain training samples, and target domain test samples. Specifically: Step 1.1, Data Collection: The original vibration signals of the rotating machinery under two operating conditions are collected by an accelerometer. The vibration signal of the first operating condition is used as labeled source domain data, and the vibration signal of the second operating condition is used as unlabeled target domain data. Both the source domain data and the target domain data contain fault and normal samples.

[0007] Step 1.2, Migration Task Definition: Specify the labeled source domain data ,in For the first One source domain sample, For the corresponding actual fault category label, The total number of categories, The total number of samples in the source domain. Indicates the source domain sample index; specifies the unlabeled target domain data. ,in For the first One target domain sample, Indicates the total number of samples in the target domain. This represents the sample index of the target domain. The data distributions of the source and target domains are different.

[0008] Step 1.3 Data Splitting: Use all source domain samples as source domain training samples. Randomly divide the target domain data into target domain training samples (accounting for more than 50%) and target domain test samples.

[0009] Step 2: Construct an adversarial domain adaptive network model for instance-weighted feature matching.

[0010] The adversarial domain adaptive network model is an end-to-end deep learning network, comprising three core components: a feature extractor, a label classifier, and a domain discriminator. The feature extractor Composed of convolutional neural networks, used to extract... Extracting high-level feature representation ,in Feature extractor Network parameters.

[0011] The label classifier Composed of fully connected layers and Softmax layers, it represents high-level features. Substitute the label classifier To predict the probability distribution of fault categories ,in Represents a label classifier Network parameters.

[0012] The domain discriminator Composed of fully connected layers, used to discriminate high-level feature representations. Whether it comes from the source domain or the target domain.

[0013] Step 3: Train the adversarial domain adaptive network model constructed in Step 2. The training process is achieved by minimizing a comprehensive loss function, ultimately yielding the trained adversarial domain adaptive network model. Specifically: The loss function consists of three parts: a source domain classification loss function, an adversarial domain adaptation loss function, and an instance-weighted feature matching loss function. Specifically: Step 3.1, Source Domain Classification Loss Function as follows: (1) in, The source domain classification loss function; , These represent feature extractors. and label classifier Network parameters. This represents the total number of training samples in the source domain; This represents summing the loss over all samples in the source domain; Indicates the first One source domain training sample. Indicates the first The true fault category label corresponding to each source domain training sample. Indicates the feature extractor processes the sample The feature vectors output later. This represents the probability distribution of predicted fault categories output by the label classifier based on the input features. This represents the cross-entropy loss function, used to measure the difference between the predicted fault category probability distribution and the true fault category label.

[0014] Step 3.2, Adversarial Domain Adaptation Loss Function : Through the domain discriminator Adversarial training prompts feature extractors Learn domain-invariant features to align the edge distributions of the source and target domains.

[0015] (2) in, This represents the adversarial domain adaptation loss function. Domain discriminant Network parameters. Representative from the source domain or target domain The Training samples. Indicates sample domain tags, This indicates that the sample comes from the source domain. This indicates that it comes from the target domain. This represents the probability that the domain discriminator determines the source (source domain or target domain) of the input based on its features. This represents the binary cross-entropy loss function, used to measure the difference between the domain prediction result and the true domain label.

[0016] Step 3.3, Instance-weighted feature matching loss function : This step is the core innovation of this invention, used to achieve conditional distribution alignment, and includes two sub-steps: Step 3.3.1, Instance Weighting: To address the issue of unreliable pseudo-labels in the target domain during the initial stage of end-to-end training, weights are calculated for each source domain sample based on the entropy minimization criterion. The lower the entropy, the higher the certainty of the prediction, and the greater the weight.

[0017] (3) (4) in, Indicates the first The instance weights of each source domain sample. Represents the natural constant. Indicates the first Predicted probability distribution of each source domain sample Information entropy is used to measure the uncertainty of prediction. This indicates the total number of fault categories. This represents the summation over all fault categories. This indicates that the label classifier predicts the first... The source domain sample belongs to the first The probability of each category. It represents the natural logarithm.

[0018] Step 3.3.2, Joint Distribution Alignment: Through multilinear mapping (outer product operation) ) features With category prediction probability Fusion generates enhanced features that contain category information. Subsequently, based on the weighted maximum mean difference metric, the distance between the enhanced feature distributions of the source and target domains is calculated as the matching loss.

[0019] , (5) (6) The symbols in formulas (5) and (6) are explained as follows: This represents the source domain enhancement feature obtained through the outer product operation; This represents the enhanced features of the target domain obtained through the outer product operation; This represents the original source domain features extracted by the feature extractor; This represents the original features of the target domain extracted by the feature extractor; This represents the predicted probability distribution of the label classifier for samples in the source domain. This represents the predicted probability distribution of the label classifier for samples in the target domain. This represents the outer product operation, used to fuse feature and category information. This represents the feature matching loss function. The compromise parameter representing instance weights is used to adjust the degree of influence of the weights, and can be set to 1 in this invention. This represents the mapping function that maps features to the reproducing kernel Hilbert space. Indicates the first Enhanced features of each source domain sample. Indicates the first Enhanced features for each target domain sample. denoted as the norm in the regenerated kernel Hilbert space. Represents the regenerated nucleus Hilbert space.

[0020] Step 3.4, the total loss function of the network model is the weighted sum of the above three losses: (7) in, This represents the total loss function for model training. and These are two hyperparameters.

[0021] Step 3.5, the model training adopts an adversarial optimization strategy: firstly, by minimizing the total loss... Optimize feature extractor and label classifier To obtain the feature extractor and label classifier Network parameters and The process is as follows: (8) Then and Substitute ,get By maximizing Optimize the domain discriminator Obtain the domain discriminator Network parameters : (9) This results in a well-trained adversarial domain adaptive network model.

[0022] Step 4: Fault diagnosis and performance evaluation.

[0023] Input the target domain test samples into the adversarial domain adaptive network model trained in step 3.5, and then use the feature extractor. and label classifier The fault type prediction result for each test sample is obtained. The classification accuracy is calculated by comparing the prediction results with the true labels (for evaluation only) to quantify the model's cross-domain diagnostic performance.

[0024] The beneficial effects of this invention are as follows: Compared with the prior art, the present invention has the following significant advantages: (1) Joint distribution adaptation: For the first time, the alignment of marginal distribution and conditional distribution is considered simultaneously in the adversarial domain adaptation, which effectively solves the negative transfer problem caused by class misalignment and significantly improves the accuracy and robustness of cross-domain diagnosis.

[0025] (2) End-to-end reliable training: The innovative instance weighting mechanism uses information entropy to dynamically evaluate the transferability of source domain samples. In the early stage of training, high-deterministic samples are given high weights, which guides the adversarial domain adaptive network model to learn features that are easy to transfer first, which accelerates model convergence and improves final performance.

[0026] (3) Strong versatility: This method does not depend on a specific network structure and can be easily integrated into different deep feature extraction networks. It is suitable for various complex industrial scenarios, such as fault diagnosis between different loads and different devices, and has good versatility and promotion value. Attached Figure Description

[0027] Figure 1 This is a flowchart of an adversarial domain adaptive cross-domain fault diagnosis method based on instance-weighted feature matching.

[0028] Figure 2 The results are feature visualizations on the CWRU dataset A→B transfer task; (a) is the visualization of source domain features extracted by the CNN method; (b) is the visualization of target domain features extracted by the CNN method; (c) is the visualization of source domain features extracted by the DANN method; (d) is the visualization of target domain features extracted by the DANN method; (e) is the visualization of source domain features extracted by the CDAN method; (f) is the visualization of target domain features extracted by the CDAN method; (g) is the visualization of source domain features extracted by the IWFM-ADA method; and (h) is the visualization of target domain features extracted by the IWFM-ADA method. Detailed Implementation

[0029] To make the problem solved by the present invention, the method scheme adopted, and the effect achieved by the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the present invention and not for limiting the present invention. Furthermore, it should be noted that, for ease of description, only the parts related to the present invention are shown in the accompanying drawings, not all of them.

[0030] First: Data; This example uses CWRU bearing data to evaluate the algorithm's performance. The CWRU dataset includes 10 health states (normal, inner ring failure with 3 dimensions, outer ring failure with 3 dimensions, and rolling element failure with 3 dimensions).

[0031] Second: Implementation steps; Step 1: Collect data, construct the transfer task, partition the data, and finally obtain source domain training samples, target domain training samples, and target domain test samples. Specifically: Step 1.1, Data Collection: The original vibration signals of the rotating machinery drive end under two working conditions, 1HP and 2HP, are collected by an accelerometer. The vibration signal under the first working condition is used as labeled source domain data, and the vibration signal under the second working condition is used as unlabeled target domain data. Both the source domain data and the target domain data contain fault and normal samples.

[0032] Step 1.2, Migration Task Definition: Specify the labeled source domain data Let A be the denoted A, where For the first One source domain sample, These are the corresponding actual fault category labels, with 10 representing the total number of categories. The total number of samples in the source domain. Indicates the source domain sample index; specifies the unlabeled target domain data. Let B be the denoted B, where For the first One target domain sample, Indicates the total number of samples in the target domain. This represents the sample index of the target domain. The data distributions of the source and target domains are different. Construct a migration task A→B.

[0033] Step 1.3 Data Splitting: Use all source domain samples as source domain training samples. Randomly divide the target domain data into target domain training samples (70%) and target domain test samples (30%). Each class has 500 samples, with a sample length of 1200.

[0034] Step 2: Construct an adversarial domain adaptive network model for instance-weighted feature matching.

[0035] The adversarial domain adaptive network model is an end-to-end deep learning network, comprising three core components: a feature extractor, a label classifier, and a domain discriminator. The feature extractor Composed of convolutional neural networks, used to extract... Extracting high-level feature representation ,in Feature extractor Network parameters.

[0036] The label classifier Composed of fully connected layers and Softmax layers, it represents high-level features. Substitute the label classifier To predict the probability distribution of fault categories ,in Represents a label classifier Network parameters.

[0037] The domain discriminator Composed of fully connected layers, used to discriminate high-level feature representations. Whether it comes from the source domain or the target domain.

[0038] Step 3: Train the adversarial domain adaptive network model constructed in Step 2. The training process is achieved by minimizing a comprehensive loss function, ultimately yielding the trained adversarial domain adaptive network model. Specifically: The loss function consists of three parts: a source domain classification loss function, an adversarial domain adaptation loss function, and an instance-weighted feature matching loss function. Specifically: Step 3.1, Source Domain Classification Loss Function as follows: (1) in, The source domain classification loss function; , These represent feature extractors. and label classifier Network parameters. This represents the total number of training samples in the source domain; This represents summing the loss over all samples in the source domain; Indicates the first One source domain training sample. Indicates the first The true fault category label corresponding to each source domain training sample. Indicates the feature extractor processes the sample The feature vectors output later. This represents the probability distribution of predicted fault categories output by the label classifier based on the input features. This represents the cross-entropy loss function, used to measure the difference between the predicted fault category probability distribution and the true fault category label.

[0039] Step 3.2, Adversarial Domain Adaptation Loss Function : Through the domain discriminator Adversarial training prompts feature extractors Learn domain-invariant features to align the edge distributions of the source and target domains.

[0040] (2) in, This represents the adversarial domain adaptation loss function. Domain discriminant Network parameters. Representative from the source domain or target domain The Training samples. Indicates sample domain tags, This indicates that the sample comes from the source domain. This indicates that it comes from the target domain. This represents the probability that the domain discriminator determines the source (source domain or target domain) of the input based on its features. This represents the binary cross-entropy loss function, used to measure the difference between the domain prediction result and the true domain label.

[0041] Step 3.3, Instance-weighted feature matching loss function : This step is the core innovation of this invention, used to achieve conditional distribution alignment, and includes two sub-steps: Step 3.3.1, Instance Weighting: To address the issue of unreliable pseudo-labels in the target domain during the initial stage of end-to-end training, weights are calculated for each source domain sample based on the entropy minimization criterion. The lower the entropy, the higher the certainty of the prediction, and the greater the weight.

[0042] (3) (4) in, Indicates the first The instance weights of each source domain sample. Represents the natural constant. Indicates the first Predicted probability distribution of each source domain sample Information entropy is used to measure the uncertainty of prediction. This indicates the total number of fault categories. This represents the summation over all fault categories. This indicates that the label classifier predicts the first... The source domain sample belongs to the first The probability of each category. It represents the natural logarithm.

[0043] Step 3.3.2, Joint Distribution Alignment: Through multilinear mapping (outer product operation) ) features With category prediction probability Fusion generates enhanced features that contain category information. Subsequently, based on the weighted maximum mean difference metric, the distance between the enhanced feature distributions of the source and target domains is calculated as the matching loss.

[0044] , (5) (6) The symbols in formulas (5) and (6) are explained as follows: This represents the source domain enhancement feature obtained through the outer product operation; This represents the enhanced features of the target domain obtained through the outer product operation; This represents the original source domain features extracted by the feature extractor; This represents the original features of the target domain extracted by the feature extractor; This represents the predicted probability distribution of the label classifier for samples in the source domain. This represents the predicted probability distribution of the label classifier for samples in the target domain. This represents the outer product operation, used to fuse feature and category information. This represents the feature matching loss function. The compromise parameter representing instance weights is used to adjust the degree of influence of the weights, and can be set to 1 in this invention. This represents the mapping function that maps features to the reproducing kernel Hilbert space. Indicates the first Enhanced features of each source domain sample. Indicates the first Enhanced features for each target domain sample. denoted as the norm in the regenerated kernel Hilbert space. Represents the regenerated nucleus Hilbert space.

[0045] Step 3.4, the total loss function of the network model is the weighted sum of the above three losses: (7) in, This represents the total loss function for model training. and These are two hyperparameters.

[0046] Step 3.5, the model training adopts an adversarial optimization strategy: firstly, by minimizing the total loss... Optimize feature extractor and label classifier To obtain the feature extractor and label classifier Network parameters and The process is as follows: (8) Then and Substitute ,get By maximizing Optimize the domain discriminator Obtain the domain discriminator Network parameters : (9) This results in a well-trained adversarial domain adaptive network model.

[0047] Step 4: Fault diagnosis and performance evaluation.

[0048] Input the target domain test samples into the adversarial domain adaptive network model trained in step 3.5, and then use the feature extractor. and label classifier The fault type prediction result for each test sample is obtained. The classification accuracy is calculated by comparing the prediction results with the true labels (for evaluation only) to quantify the model's cross-domain diagnostic performance.

[0049] Third, parameter settings; The structural parameters of the adaptive network model for adversarial domains are shown in Table 1. The feature extractor uses a 4-layer one-dimensional convolutional neural network. An Adaptive Moment Estimation (Adam) optimizer is used with an initial learning rate of 0.001, a batch size of 64, and 100 training iterations. The hyperparameters in the total loss function (Equation 7) are set to... .

[0050] Table 1: Structural parameters of the adaptive network model in the adversarial domain

[0051] Fourth, comparative experiments and results analysis; Three comparison methods were set up to verify the superiority of the invention method: (1) Convolutional Neural Network (CNN), which consists of a feature extractor and a label classifier. (2) Domain Adversarial Neural Network (DANN), which uses a domain discriminator for adversarial training to learn domain-invariant features. (3) Conditional Domain Adversarial Networks (CDAN).

[0052] Fifth, the results of the case studies; Table 2: Average classification accuracy (%) of different methods on the CWRU dataset

[0053] The average classification accuracy on the two transfer tasks A→B and B→A is shown in Table 2. The results show that the method of the present invention (IWFM-ADA) achieves the highest or near-highest accuracy on all tasks, especially on tasks with large distribution differences, demonstrating its superior cross-domain adaptability.

[0054] Sixth, feature visualization analysis; To further demonstrate the effectiveness of this invention, t-distributed stochastic neighborhood embedding (t-SNE) was used to visualize the features learned by different methods in the A→B task. The results are as follows: Figure 2 As shown in the figure, the results reveal severe class confusion in the target domain features of the CNN method. While the DANN method aligns the edge distribution through adversarial training, the class centers remain misaligned. In contrast, the source and target domain features extracted by the method of this invention (IWFM-ADA) are not only aligned in overall distribution, but each fault category also forms a clear and aligned cluster, intuitively demonstrating its effective ability to achieve joint distribution adaptation. This invention successfully introduces an instance-weighted feature matching mechanism into the adversarial domain adaptive framework, effectively solving the joint distribution adaptation problem in cross-domain fault diagnosis. Validated on the CWRU bearing dataset, this method significantly outperforms existing mainstream methods, providing reliable technical support for intelligent operation and maintenance of equipment under varying operating conditions in industrial practice.

[0055] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A cross-domain adaptive fault diagnosis method based on instance-weighted feature matching, characterized in that, The adversarial domain adaptive cross-domain fault diagnosis method includes the following steps: Step 1: Collect data, construct the transfer task, divide the data, and finally obtain source domain training samples, target domain training samples, and target domain test samples; Step 2: Construct an adversarial domain adaptive network model for instance-weighted feature matching; The adversarial domain adaptive network model is an end-to-end deep learning network that includes three core components: a feature extractor, a label classifier, and a domain discriminator. Step 3: Train the adversarial domain adaptive network model constructed in Step 2. The training process is achieved by minimizing a comprehensive loss function, and finally the trained adversarial domain adaptive network model is obtained. The loss function consists of three parts: source domain classification loss function, adversarial domain adaptation loss function, and instance-weighted feature matching loss function. Step 4: Fault diagnosis and performance evaluation.

2. The adversarial domain adaptive cross-domain fault diagnosis method based on instance-weighted feature matching according to claim 1, characterized in that, Specifically, step 1 is as follows: Step 1.1, Data Collection: The original vibration signals of the rotating machinery drive end under two working conditions, 1HP and 2HP, are collected by an accelerometer. The vibration signal under the first working condition is used as labeled source domain data, and the vibration signal under the second working condition is used as unlabeled target domain data. Both the source domain data and the target domain data contain fault and normal samples. Step 1.2, Migration Task Definition: Specify the labeled source domain data Let A be the denoted A, where For the first One source domain sample, These are the corresponding actual fault category labels, with 10 representing the total number of categories. The total number of samples in the source domain. Indicates the source domain sample index; specifies the unlabeled target domain data. Let B be the denoted B, where For the first One target domain sample, Indicates the total number of samples in the target domain. This represents the sample index of the target domain. The data distributions of the source and target domains are different. Construct a migration task A→B. Step 1.3, Data partitioning: Use all source domain samples as source domain training samples; randomly partition the target domain data into target domain training samples and target domain test samples.

3. The adversarial domain adaptive cross-domain fault diagnosis method based on instance-weighted feature matching according to claim 2, characterized in that, In step 1.3, the number of training samples in the target domain accounts for more than 50%.

4. The adversarial domain adaptive cross-domain fault diagnosis method based on instance-weighted feature matching according to claim 2, characterized in that, Specifically, step 2 is as follows: The adversarial domain adaptive network model is an end-to-end deep learning network, comprising three core components: a feature extractor, a label classifier, and a domain discriminator. The feature extractor Composed of convolutional neural networks, used to extract... Extracting high-level feature representation ,in Feature extractor Network parameters. The label classifier Composed of fully connected layers and Softmax layers, it represents high-level features. Substitute the label classifier To predict the probability distribution of fault categories ,in Represents a label classifier Network parameters. The domain discriminator Composed of fully connected layers, used to discriminate high-level feature representations. Whether it comes from the source domain or the target domain.

5. The adversarial domain adaptive cross-domain fault diagnosis method based on instance-weighted feature matching according to claim 4, characterized in that, Specifically, step 3 is as follows: Step 3.1, Source Domain Classification Loss Function as follows: (1) in, The source domain classification loss function; , These represent feature extractors. and label classifier Network parameters; This represents the total number of training samples in the source domain; This represents summing the loss over all samples in the source domain; Indicates the first One source domain training sample; Indicates the first The true fault category label corresponding to each source domain training sample; Indicates the feature extractor processes the sample The feature vector output later; This represents the probability distribution of predicted fault categories output by the label classifier based on the input features; This represents the cross-entropy loss function, used to measure the difference between the predicted fault category probability distribution and the true fault category label; Step 3.2, Adversarial Domain Adaptation Loss Function : Through the domain discriminator Adversarial training prompts feature extractors Learn generative domain-invariant features to align the edge distributions of the source and target domains; (2) in, This represents the adversarial domain adaptation loss function; Domain discriminant Network parameters; Representative from the source domain or target domain The One training sample; Indicates sample domain tags, This indicates that the sample comes from the source domain. Indicates that it comes from the target domain; This represents the probability that the domain discriminator determines the source of the input based on its features; This represents the binary cross-entropy loss function, used to measure the difference between the domain prediction result and the true domain label; Step 3.3, Instance-weighted feature matching loss function ; Step 3.4, the total loss function of the network model is the weighted sum of the above three losses: (7) in, This represents the total loss function for model training. and These are two hyperparameters. Step 3.5, the model training adopts an adversarial optimization strategy: firstly, by minimizing the total loss... Optimize feature extractor and label classifier To obtain the feature extractor and label classifier Network parameters and The process is as follows: (8) Then and Substitute ,get By maximizing Optimize the domain discriminator Obtain the domain discriminator Network parameters : (9) This results in a well-trained adversarial domain adaptive network model.

6. The adversarial domain adaptive cross-domain fault diagnosis method based on instance-weighted feature matching according to claim 5, characterized in that, The feature matching loss function in step 3.3 is used to achieve conditional distribution alignment, specifically as follows: Step 3.3.1, Instance Weighting: To address the issue of unreliable pseudo-labels in the target domain during the initial stage of end-to-end training, weights are calculated for each source domain sample based on the entropy minimization criterion. The lower the entropy, the higher the certainty of the prediction, and the greater the weight. (3) (4) in, Indicates the first The instance weights of each source domain sample. Represents the natural constant. Indicates the first Predicted probability distribution of each source domain sample Information entropy is used to measure the uncertainty of prediction. This indicates the total number of fault categories. This represents the summation over all fault categories. This indicates that the label classifier predicts the first... The source domain sample belongs to the first The probability of each category. It represents the natural logarithm. Step 3.3.2, Joint Distribution Alignment: Through multilinear mapping (outer product operation) ) features With category prediction probability Fusion generates enhanced features that contain category information. Subsequently, based on the weighted maximum mean difference metric, the distance between the enhanced feature distributions of the source and target domains is calculated as the matching loss. ， (5) (6) The symbols in formulas (5) and (6) are explained as follows: This represents the source domain enhancement feature obtained through the outer product operation; This represents the enhanced features of the target domain obtained through the outer product operation; This represents the original source domain features extracted by the feature extractor; This represents the original features of the target domain extracted by the feature extractor; This represents the predicted probability distribution of the label classifier for samples in the source domain. This represents the predicted probability distribution of the label classifier for samples in the target domain. This represents the outer product operation, used to fuse feature and category information. This represents the feature matching loss function. The compromise parameter representing instance weights is used to adjust the degree of influence of the weights, and can be set to 1 in this invention. This represents the mapping function that maps features to the reproducing kernel Hilbert space. Indicates the first Enhanced features of each source domain sample. Indicates the first Enhanced features for each target domain sample. denoted as the norm in the regenerated kernel Hilbert space. This represents the regenerated nucleus Hilbert space.

7. The adversarial domain adaptive cross-domain fault diagnosis method based on instance-weighted feature matching according to claim 6, characterized in that, Step 4 specifically involves: Input the target domain test samples into the adversarial domain adaptive network model trained in step 3.5, and then use the feature extractor. and label classifier This yields the fault type prediction results for each test sample.

Images