A cross-domain diagnostic method based on dual-condition feature constraints
By employing a dual-condition feature constraint method, combined with cross-domain distribution perception minimizing class confusion loss and a dual-classifier architecture, the problem of insufficient robustness of a single classifier is solved, achieving higher accuracy and stability in cross-domain fault diagnosis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV OF POSTS & TELECOMM
- Filing Date
- 2026-05-08
- Publication Date
- 2026-07-10
AI Technical Summary
Existing cross-domain fault diagnosis methods lack robustness of a single classifier when facing complex cross-domain scenarios, and traditional methods ignore the spatial structural relationships of source domain samples, resulting in insufficient model generalization ability and diagnostic accuracy.
We adopt a dual-condition feature constraint-based approach, which integrates cross-domain distribution perception to minimize class confusion loss and a dual-classifier architecture. By utilizing the spatial structure information of the source domain class center to impose consistency constraints on the target domain, and combining dual-classifier collaborative classification, we achieve robustness and stability of domain-invariant features.
It significantly improves the model's generalization ability and diagnostic accuracy in cross-domain scenarios, effectively reduces the confusion between categories in the target domain, enhances the discriminative power of features and the reliability of predictions, and solves the local optimization problem of a single classifier.
Smart Images
Figure CN122365384A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of fault diagnosis and relates to a cross-domain diagnostic method based on dual-condition feature constraints. Background Technology
[0002] Rotating machinery, as core equipment in modern industrial systems, is widely used in critical fields such as aerospace, energy and power, and petrochemicals. Rolling bearings, as important components of rotating machinery, operate under complex and harsh environments for extended periods, making them the most vulnerable and prone-to-failure link in the system. Relevant statistics show that bearing failure is the leading cause of rotating machinery failures, accounting for approximately 40% of all equipment failures. Therefore, establishing efficient and accurate bearing fault diagnosis technology is of great significance for ensuring reliable equipment operation and preventing major accidents.
[0003] With the rapid development of deep learning technology, intelligent fault diagnosis methods based on vibration signals have become a research hotspot. However, traditional deep learning methods often assume that training and testing data come from the same distribution, an assumption that is difficult to meet in practical industrial applications. Due to differences in equipment operating environment, workload, sensor configuration, and other factors, data collected from different devices or the same device under different operating conditions often exhibit significant domain shifts, severely affecting the generalization performance of fault diagnosis models. In recent years, unsupervised domain adaptation (UDA) technology has provided an effective approach to solving cross-domain fault diagnosis problems.
[0004] Despite significant progress in cross-domain fault diagnosis, existing UDA methods still suffer from key shortcomings that limit their diagnostic performance. The robustness limitation of a single classifier: Existing methods generally employ a single classifier for fault identification, making them susceptible to domain shifts. The lack of diversity constraints and consistency verification mechanisms in single network structures often results in poor stability and generalization ability when facing complex cross-domain scenarios, hindering robust fault diagnosis performance. Furthermore, insufficient utilization of source domain structural information: Traditional minimal-class confusion loss functions rely solely on the distribution statistics of predicted probabilities, completely ignoring the spatial structural relationships and class center information of source domain samples. This probability-based equal weighting strategy leads to model overconfidence. Therefore, improving the accuracy of cross-domain diagnosis through methodological advancements is a technical problem that needs to be addressed by those in the field. Summary of the Invention
[0005] In view of this, the purpose of this invention is to provide a cross-domain diagnostic method based on dual-conditional feature constraints. This invention achieves a dual feature constraint mechanism by integrating a cross-domain distribution-aware minimization class confusion loss with a dual-classifier architecture. Specifically, the cross-domain distribution-aware minimization class confusion loss enhances the discriminative power of the target domain's class features while imposing structural consistency constraints on the target domain using the spatial structural information of the source domain's class centers, thus mitigating the overconfidence problem caused by entropy weights in traditional class confusion losses. The dual-classifier imposes consistency constraints on the source domain features through collaborative classification, ensuring the robustness of domain-invariant features. The synergistic effect of these two constraints effectively improves the model's generalization ability in cross-domain scenarios.
[0006] To achieve the above objectives, the present invention provides the following technical solution:
[0007] A cross-domain diagnostic method based on dual-condition feature constraints includes the following steps:
[0008] S1: Use a sliding window approach to process long sequence data and construct time-domain signal samples of the source and target domains;
[0009] S2: Construct an adaptive network that includes a feature extractor, a decoder, a biconditional feature constraint module, and a domain discriminator;
[0010] S3: Use trained feature extractors and classifiers to identify the health status of unlabeled data in the target domain;
[0011] The feature extractor module of the cross-domain diagnostic method based on dual-condition feature constraints is as follows:
[0012] The feature extractor module in the cross-domain diagnostic method based on dual-conditional feature constraints consists of multiple cascaded convolutional layers, normalization layers, and activation functions. The input is vibration modal signal data from the source and target domains. The feature extractor with shared parameters extracts domain-invariant deep feature representations.
[0013] The dual-condition feature constraint module in the cross-domain diagnostic method based on dual-condition feature constraints is specifically as follows:
[0014] The dual-condition feature constraint module employs two different classifier structures. Its input consists of source domain-invariant features extracted by the feature extractor, used to construct two distribution boundary-constrained feature extractors to prevent overfitting to a specific decision boundary. The output is the source domain predicted label. Then, the target domain's domain-invariant features are simultaneously input into two independent classifiers, obtaining two sets of predicted probabilities. For each classifier's predicted probability output, the same cross-domain distribution-aware minimization class confusion loss calculation method is used: the spatial relationship between the target domain features and the source domain class centers is integrated into the confusion matrix calculation process. Simultaneously, an adaptive weighting mechanism based on prediction entropy is introduced to differentiate the importance of samples. Finally, a class confusion matrix fusing spatial structure information and prediction confidence is constructed, specifically:
[0015] (a) Calculate the class centers c of each category in the source domain using the domain-invariant features of the source domain and the corresponding labels. c Specifically:
[0016] ;
[0017] Among them, S c f represents the number of samples belonging to class j in the source domain. i s Let be the high-dimensional spatial features of the i-th source domain sample.
[0018] (b) Calculate spatial affinity and combine the predicted probabilities, specifically:
[0019] ;
[0020] in, Let represent the feature representation of the j-th target domain sample, d be the average distance, and ε be a small stability constant. For the j-th sample in the target domain, the probability that it belongs to class c is... It can be represented as
[0021] ;
[0022] Where T is the temperature correction coefficient, which is set to 3.0 here. The predicted probability of the classifier reflects the confidence of the model in classifying samples in the target domain, while spatial affinity reflects the geometric relationship between the sample and the feature center of the given class in the feature space. Furthermore, the predicted probability of the classifier and the spatial affinity based on the source domain class center are fused by geometric averaging. The calculation process can be expressed as follows:
[0023] ;
[0024] (c) Calculate the entropy weights, specifically:
[0025] ;
[0026] ;
[0027] in, W represents the uncertainty matrix of the classifier's prediction for a given sample. jj B represents the significance probability of the sample confusing the class; B is the batch size.
[0028] (d) Calculate the confusion matrix First, we obtain a preliminary form of class obfuscation. Specifically:
[0029] ;
[0030] because Defining a single batch based solely on a single batch can easily lead to class imbalance when the number of categories is large. Therefore, a class normalization method is used to obtain the desired results. :
[0031] ;
[0032] The decoder module in the cross-domain diagnostic method based on dual-condition feature constraints is specifically as follows:
[0033] The decoder modules in the cross-domain diagnostic method based on dual-conditional feature constraints are a source domain feature decoder and a target domain feature decoder. Taking the source and target domain feature outputs from the feature extractor as input, it employs a reconstruction architecture combining multi-level upsampling and transposed convolution. Through inverse operations, it achieves high-quality reconstruction from compressed features to the original signal. The decoder structure is mirror-symmetric to the feature extractor, using corresponding multi-level upsampling operations to progressively recover the signal length, and utilizing transposed convolution for feature mapping and dimensionality upsampling. Simultaneously, the normalization strategy is consistent with the feature extractor to ensure the stability of the feature distribution, achieving a complete mapping from abstract feature representation to concrete signal reconstruction.
[0034] The domain discriminator module in the cross-domain diagnostic method based on dual-condition feature constraints is specifically as follows:
[0035] Source and target domain alignment is achieved through a domain discriminator. The input is the source and target domain invariant features extracted by the feature extractor. The domain discriminator determines whether the domain invariant features come from the source or target domain, and aligns the distribution between the two domains from a global perspective.
[0036] The training steps for the cross-domain diagnostic method based on dual-condition feature constraints are as follows:
[0037] source domain and target domain samples The source domain invariant features were obtained respectively. Domain-invariant features of the target domain By inputting the source domain domain-invariant features into two classifiers C1 and C2 with different structures, and applying the classification loss function L... CLS Update the parameters of the two classifiers:
[0038] ;
[0039] in, , and Let n represent the label value of the i-th sample and the probabilities of its classifiers C1 and C2 outputs belonging to class c, respectively, where C is the number of classes. s The number of samples in the source domain is given. The target domain probability distribution is obtained by inputting the domain-invariant features of the target domain into two classifiers with different structures. The class confusion matrix of the two classifiers is calculated through class normalization, and then... (The sentence is incomplete and requires further context to translate accurately.) MCC Minimizing the confusion between each class and other classes encourages the model to produce more deterministic and consistent predictions in the target domain, thereby reducing inter-class confusion and improving classification performance in domain adaptation. Specifically:
[0040] ;
[0041] in, and Let represent the cross-domain distribution-aware confusion matrices, where the predicted probabilities from the outputs of the first and second classifiers are used as inputs. Then, a feature alignment loss is employed to minimize the distributional difference between source-domain invariant features and target-domain invariant features, which can be expressed by the following formula:
[0042] ;
[0043] Where, d i Let R represent the binary label (domain label) of the i-th sample, and R() represent gradient reversal. Source domain domain-invariant features. Domain-invariant features of the target domain These are used as inputs to the source domain decoder and the target domain decoder, respectively, and the output is a source domain reconstructed sample. Reconstructing samples from the target domain Loss through source domain reconstruction and target domain reconstruction loss Applying L2 norm constraints yields the total reconstruction loss. Its expression is as follows:
[0044] ;
[0045] ;
[0046] ;
[0047] in, Let L2 norm be represented. The final loss is obtained:
[0048] ;
[0049] The beneficial effects of this invention are as follows: This invention proposes a cross-domain diagnostic method based on dual-conditional feature constraints for fault diagnosis. Addressing the issue that traditional minimum-class confusion loss constructs a confusion matrix solely based on the predicted probability distribution of the target domain, completely ignoring the geometric structural information inherent in the feature space, and failing to effectively utilize prior knowledge of the source domain's categories to guide feature alignment in the target domain, this invention proposes a minimum-class confusion loss function based on cross-domain distribution awareness. This function integrates the spatial relationship between target domain features and source domain class centers into the calculation of the confusion matrix, while introducing an adaptive weighting mechanism based on prediction entropy to achieve differentiated processing of sample importance. Furthermore, addressing the problem that traditional single-classifier structures often struggle to effectively constrain feature extractors to extract domain-invariant features with high discriminative and generalizable capabilities, and cannot prevent feature extractors from overfitting to a specific decision boundary, thus severely limiting the accuracy and generalization ability of cross-domain diagnosis, this invention proposes a dual-conditional feature constraint module. This module, by combining cross-domain distribution-aware minimum-class confusion loss and a dual-classifier architecture, imposes differentiated constraints on source and target domain features, achieving more accurate learning of domain-invariant features. This method effectively enables cross-domain fault diagnosis, overcoming the challenge of unlabeled fault samples and providing a feasible solution for cross-domain fault diagnosis. Specific advantages are as follows:
[0050] 1. This invention quantifies the class affiliation tendency of samples by calculating the distance between the target domain sample and the source domain class center. Smaller distances indicate higher feature similarity between the target domain sample and the corresponding source domain class. Furthermore, the predicted probability and spatial affinity are geometrically averaged to form a cross-domain distribution-aware confusion matrix. An adaptive weighting mechanism based on prediction entropy is introduced to dynamically adjust the contribution of samples with different confidence levels. By minimizing this cross-domain distribution-aware class confusion loss, the model effectively reduces confusion between different classes in the target domain, prompting target domain samples to converge towards the correct source domain class center in the feature space, thereby achieving more accurate cross-domain class alignment and higher classification performance.
[0051] 2. This invention constructs a dual-condition feature constraint module by combining a dual-classifier structure with a spatially aware class confusion minimization loss. First, it employs cross-domain distributional awareness to minimize class confusion loss, fully mining the spatial topological information of the source domain class centers. Geometric consistency normalization is applied to the feature distribution of the target domain samples, guiding the target domain features to converge towards the corresponding class centers in the source domain. Simultaneously, by fusing spatial distance and prediction probability, it suppresses the overconfidence phenomenon caused by the entropy weighting strategy in the classic MCC loss, thereby enhancing the discriminative power and prediction reliability of the target domain features. Second, by constructing a heterogeneous dual-classifier collaborative system, cross-classifier consistency normalization is applied to common domain-invariant features. This mechanism utilizes the differences in feature processing paradigms of multilayer perceptrons with different structures to form a mutually verifying discrimination system, effectively enhancing the stability and generalization ability of feature encoding. Furthermore, the consistency requirement of dual-classifier decisions avoids the local optimum traps and decision boundary uncertainties common in single-classifier systems. The combination of these two approaches avoids the local optimization problems that may result from a single constraint and achieves global optima in cross-domain feature learning, significantly enhancing the model's adaptability and diagnostic performance in complex domain shift scenarios.
[0052] Figure caption
[0053] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following figures are provided for illustration:
[0054] Figure 1 This is a schematic diagram of the workflow framework for a cross-domain diagnostic method based on dual-condition feature constraints. Detailed Implementation
[0055] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.
[0056] This invention utilizes the Western Reserve University bearing dataset, the Hanoi University of Science and Technology bearing dataset, and bearing test bench data I and II built in the laboratory for verification. Detailed operating parameters and acquisition conditions for the four bearing datasets are shown in Table 1. Based on this, 12 cross-machine fault diagnosis tasks are constructed, specifically including: A→B, B→A, A→C, C→A, A→D, D→A, B→C, C→B, B→D, D→B, C→D, and D→C, where "→" indicates a fault diagnosis task from the source domain to the target domain.
[0057] Table 1. Description and operating condition data of the cross-device experimental dataset
[0058]
[0059] As attached Figure 1As shown, a bearing cross-domain diagnostic method based on dual-condition feature constraints includes the following steps:
[0060] S1: Use a sliding window method to process long sequence data and construct bearing time-domain signal samples in the source and target domains;
[0061] S2: Construct an adaptive network that includes a feature extractor, a decoder, a biconditional feature constraint module, and a domain discriminator;
[0062] S3: Use trained feature extractors and classifiers to identify the bearing health status of unlabeled data in the target domain;
[0063] Furthermore, step S2 specifically includes:
[0064] The feature extractor module of the cross-domain diagnostic method based on dual-condition feature constraints is as follows:
[0065] The feature extractor module in the cross-domain diagnostic method based on dual-conditional feature constraints consists of multiple cascaded convolutional layers, normalization layers, and activation functions. The input is bearing vibration modal signal data from the source and target domains. The feature extractor with shared parameters extracts domain-invariant deep feature representations.
[0066] The dual-condition feature constraint module in the cross-domain diagnostic method based on dual-condition feature constraints is specifically as follows:
[0067] The dual-condition feature constraint module employs two different classifier structures. Its input consists of source domain-invariant features extracted by the feature extractor, used to construct two distribution boundary-constrained feature extractors to prevent overfitting to a specific decision boundary. The output is the source domain predicted label. Then, the target domain's domain-invariant features are simultaneously input into two independent classifiers, obtaining two sets of predicted probabilities. For each classifier's predicted probability output, the same cross-domain distribution-aware minimization class confusion loss calculation method is used: the spatial relationship between the target domain features and the source domain class centers is integrated into the confusion matrix calculation process. Simultaneously, an adaptive weighting mechanism based on prediction entropy is introduced to differentiate the importance of samples. Finally, a class confusion matrix fusing spatial structure information and prediction confidence is constructed, specifically:
[0068] (a) Calculate the class centers c of each category in the source domain using the domain-invariant features of the source domain and the corresponding labels. c Specifically:
[0069] ;
[0070] Among them, S c f represents the number of samples belonging to class j in the source domain. i s Let be the high-dimensional spatial features of the i-th source domain sample.
[0071] (b) Calculate spatial affinity and combine the predicted probabilities, specifically:
[0072] ;
[0073] in, Let represent the feature representation of the j-th target domain sample, d be the average distance, and ε be a small stability constant. For the j-th sample in the target domain, the probability that it belongs to class c is... It can be represented as
[0074] ;
[0075] Where T is the temperature correction coefficient, which is set to 3.0 here. The predicted probability of the classifier reflects the confidence of the model in classifying samples in the target domain, while spatial affinity reflects the geometric relationship between the sample and the feature center of the given class in the feature space. Furthermore, the predicted probability of the classifier and the spatial affinity based on the source domain class center are fused by geometric averaging. The calculation process can be expressed as follows:
[0076] ;
[0077] (c) Calculate the entropy weights, specifically:
[0078] ;
[0079] ;
[0080] in, W represents the uncertainty matrix of the classifier's prediction for a given sample. jj B represents the significance probability of the sample confusing the class; B is the batch size.
[0081] (d) Calculate the confusion matrix First, we obtain a preliminary form of class obfuscation. Specifically:
[0082] ;
[0083] because Defining a single batch based solely on a single batch can easily lead to class imbalance when the number of categories is large. Therefore, a class normalization method is used to obtain the desired results. :
[0084] ;
[0085] The decoder module in the cross-domain diagnostic method based on dual-condition feature constraints is specifically as follows:
[0086] The decoder modules in the cross-domain diagnostic method based on dual-conditional feature constraints are a source domain feature decoder and a target domain feature decoder. Taking the source and target domain feature outputs from the feature extractor as input, it employs a reconstruction architecture combining multi-level upsampling and transposed convolution. Through inverse operations, it achieves high-quality reconstruction from compressed features to the original signal. The decoder structure is mirror-symmetric to the feature extractor, using corresponding multi-level upsampling operations to progressively recover the signal length, and utilizing transposed convolution for feature mapping and dimensionality upsampling. Simultaneously, the normalization strategy is consistent with the feature extractor to ensure the stability of the feature distribution, achieving a complete mapping from abstract feature representation to concrete signal reconstruction.
[0087] The domain discriminator module in the cross-domain diagnostic method based on dual-condition feature constraints is specifically as follows:
[0088] Source and target domain alignment is achieved through a domain discriminator. The input is the source and target domain invariant features extracted by the feature extractor. The domain discriminator determines whether the domain invariant features come from the source or target domain, and aligns the distribution between the two domains from a global perspective.
[0089] Furthermore, step S3 specifically includes:
[0090] source domain and target domain samples The source domain invariant features were obtained respectively. Domain-invariant features of the target domain By inputting the source domain domain-invariant features into two classifiers C1 and C2 with different structures, and applying the classification loss function L... CLS Update the parameters of the two classifiers:
[0091] ;
[0092] in, , and Let n represent the label value of the i-th sample and the probabilities of its classifiers C1 and C2 outputs belonging to class c, respectively, where C is the number of classes. s The number of samples in the source domain is given. The target domain probability distribution is obtained by inputting the domain-invariant features of the target domain into two classifiers with different structures. The class confusion matrix of the two classifiers is calculated through class normalization, and then... (The sentence is incomplete and requires further context to translate accurately.) MCC Minimizing the confusion between each class and other classes encourages the model to produce more deterministic and consistent predictions in the target domain, thereby reducing inter-class confusion and improving classification performance in domain adaptation. Specifically:
[0093] ;
[0094] in, and Let represent the cross-domain distribution-aware confusion matrices, where the predicted probabilities from the outputs of the first and second classifiers are used as inputs. Then, a feature alignment loss is employed to minimize the distributional difference between source-domain invariant features and target-domain invariant features, which can be expressed by the following formula:
[0095] ;
[0096] Where, d i Let R represent the binary label (domain label) of the i-th sample, and R() represent gradient reversal. Source domain domain-invariant features. Domain-invariant features of the target domain These are used as inputs to the source domain decoder and the target domain decoder, respectively, and the output is a source domain reconstructed sample. Reconstructing samples from the target domain Loss through source domain reconstruction and target domain reconstruction loss Applying L2 norm constraints yields the total reconstruction loss. Its expression is as follows:
[0097] ;
[0098] ;
[0099] ;
[0100] in, Let L2 norm be represented. The final loss is obtained:
[0101] ;
[0102] Labeled source domain sample data and unlabeled target domain sample data are imported together into a cross-domain diagnostic method based on dual-condition feature constraints. By using the trained feature extractor and dual classifier structure, the bearing health status of the unlabeled target domain data is identified, and the final test results are obtained.
[0103] To verify the effectiveness of the method of the present invention, the following schemes were selected for comparison: (1) retaining the dual classifier structure, canceling the cross-domain distribution perception to minimize the class confusion loss of the single classifier C1 or C2, and named as Method 1 and Method 2 respectively; (2) Method 3: retaining the dual classifier structure, but not applying the cross-domain distribution perception to minimize the class confusion loss of the target domain invariant features; (3) Method 4: only a single classifier structure, and applying the cross-domain distribution perception to minimize the class confusion loss of the target domain invariant features.
[0104] The experimental results are shown in Table 2. The accuracy of the method of the present invention has been improved in different cross-device experiments.
[0105] Table 2. Results of cross-device experiments
[0106]
[0107] According to the cross-device experimental results in Table 2: 1) The accuracies of Method 1 and Method 2 are 90.01% and 90.65% respectively, both significantly lower than the complete method, verifying the necessity of applying cross-domain distribution-aware class confusion loss in collaboration with dual classifiers. This indicates that introducing cross-domain distribution-aware constraints in only a single classifier cannot fully leverage the advantages of dual-condition feature constraints; only when both classifiers are simultaneously guided by cross-domain distribution-aware loss can optimal domain-invariant feature learning be achieved. 2) The accuracy of Method 3 drops significantly to 82.82%, forming the largest performance gap with the complete method, which fully verifies the key role of the cross-domain distribution-aware class confusion loss mechanism. This result shows that although traditional dual classifiers lacking spatial geometric constraints have certain complementarity, they cannot effectively solve the problem of target domain feature distribution offset and are difficult to achieve accurate cross-domain alignment. 3) The accuracy of Method 4 is 86.99%, which is better than Method 3 but still lower than the complete method, demonstrating the unique value of the dual classifier architecture. This indicates that even under the guidance of cross-domain distributed perceptual loss, the feature representation capability and decision robustness of a single classifier are still limited, while the synergistic effect of heterogeneous dual classifiers can provide a more stable and discriminative feature learning mechanism.
[0108] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made to it in form and detail without departing from the scope defined by the claims of the present invention.
Claims
1. A cross-domain diagnostic method based on dual-condition feature constraints, characterized in that: The method includes the following steps: S1: Use a sliding window approach to process long sequence data and construct time-domain signal samples of the source and target domains; S2: Construct an adaptive network that includes a feature extractor, a decoder, a biconditional feature constraint module, and a domain discriminator; S3: Use trained feature extractors and classifiers to identify the health status of unlabeled data in the target domain.
2. The cross-domain diagnostic method based on dual-condition feature constraints as described in claim 1, characterized in that: The feature extractor module in the cross-domain diagnostic method based on dual-condition feature constraints is specifically as follows: The feature extractor module in the cross-domain diagnostic method based on dual-conditional feature constraints consists of multiple cascaded convolutional layers, normalization layers, and activation functions. The input is vibration modal signal data from the source and target domains. The feature extractor with shared parameters extracts domain-invariant deep feature representations.
3. The cross-domain diagnostic method based on dual-condition feature constraints as described in claim 1, characterized in that: The dual-condition feature constraint module in the cross-domain diagnostic method based on dual-condition feature constraints is specifically as follows: The dual-condition feature constraint module employs two different classifier structures. Its input consists of source domain-invariant features extracted by the feature extractor, used to construct two distribution boundary-constrained feature extractors to prevent overfitting to a specific decision boundary. The output is the source domain predicted label. Then, the target domain's domain-invariant features are simultaneously input into two independent classifiers, obtaining two sets of predicted probabilities. For each classifier's predicted probability output, the same cross-domain distribution-aware minimization class confusion loss calculation method is used: the spatial relationship between the target domain features and the source domain class centers is integrated into the confusion matrix calculation process. Simultaneously, an adaptive weighting mechanism based on prediction entropy is introduced to differentiate the importance of samples. Finally, a class confusion matrix fusing spatial structure information and prediction confidence is constructed, specifically: (a) Calculate the class centers c of each category in the source domain using the domain-invariant features of the source domain and the corresponding labels. c Specifically: ； Among them, S c f represents the number of samples belonging to class j in the source domain. i s The high-dimensional spatial features of the i-th source domain sample; (b) Calculate spatial affinity and combine the predicted probabilities, specifically: ； in, Let represent the feature representation of the j-th target domain sample, d be the average distance, and ε be a small stability constant; for the j-th sample in the target domain, the probability that it belongs to class c. It can be represented as ； Where T is the temperature correction coefficient, which is set to 3.0 here; the predicted probability of the classifier reflects the confidence of the model in classifying samples in the target domain, while the spatial affinity reflects the geometric relationship between the sample and the feature center of the given class in the feature space; further, the predicted probability of the classifier and the spatial affinity based on the source domain class center are fused by geometric averaging, and the calculation process can be expressed as follows: ； (c) Calculate the entropy weights, specifically: ； ； in, W represents the uncertainty matrix of the classifier's prediction for a given sample. jj B represents the significant probability of class confusion for the sample; B is the batch size. (d) Calculate the confusion matrix First, we obtain a preliminary form of class obfuscation. Specifically: ； because Defining a single batch based solely on a single batch can easily lead to class imbalance when the number of categories is large. Therefore, a class normalization method is used to obtain the desired results. : 。 4. The cross-domain diagnostic method based on dual-condition feature constraints as described in claim 1, characterized in that: The decoder module in the cross-domain diagnostic method based on dual-condition feature constraints is specifically as follows: The decoder module in the cross-domain diagnostic method based on dual-conditional feature constraints includes a source domain feature decoder and a target domain feature decoder. They take the source domain and target domain features output by the feature extractor as input and adopt a reconstruction architecture that combines multi-level upsampling and transposed convolution. Through inverse operation, they achieve high-quality reconstruction from compressed features to the original signal. The decoder structure forms a mirror symmetry with the feature extractor. It uses multi-level upsampling to gradually recover the signal length and completes feature mapping and dimensionality increase through transposed convolution. At the same time, it maintains a normalization strategy consistent with the feature extractor to ensure the stability of the feature distribution, thereby achieving a complete mapping from abstract feature representation to specific signals.
5. The cross-domain diagnostic method based on dual-condition feature constraints as described in claim 1, characterized in that: The domain discriminator module in the cross-domain diagnostic method based on dual-condition feature constraints is specifically as follows: Source and target domain alignment is achieved through a domain discriminator. The input is the source and target domain invariant features extracted by the feature extractor. The domain discriminator determines whether the domain invariant features come from the source or target domain, and aligns the distribution between the two domains from a global perspective.
6. The cross-domain diagnostic method based on dual-condition feature constraints as described in claim 1, characterized in that: The training steps for the cross-domain diagnostic method based on dual-condition feature constraints are as follows: source domain and target domain samples The source domain invariant features were obtained respectively. Domain-invariant features of the target domain By inputting the source domain domain-invariant features into two classifiers C1 and C2 with different structures, and applying the classification loss function L... CLS Update the parameters of the two classifiers: ； in, , and Let n represent the label value of the i-th sample and the probabilities of its classifiers C1 and C2 outputs belonging to class c, respectively, where C is the number of classes. s The number of samples in the source domain; by inputting the domain-invariant features of the target domain into two classifiers with different structures, the probability distribution of the target domain is obtained. The class confusion matrix of the two classifiers is calculated using the class normalization method, and then... MCC Minimizing the confusion between each class and other classes encourages the model to produce more deterministic and consistent predictions in the target domain, thereby reducing inter-class confusion and improving the classification performance of domain adaptation; specifically: ； in, and Let represent the cross-domain distribution-aware confusion matrix, where the predicted probabilities output by the first and second classifiers are used as inputs. Then, a feature alignment loss is employed to minimize the distributional difference between source-domain invariant features and target-domain invariant features, which can be expressed by the following formula: ； Where, d i Let R represent the binary label (domain label) of the i-th sample, and let R() represent gradient reversal; source domain domain-invariant features Domain-invariant features of the target domain These are used as inputs to the source domain decoder and the target domain decoder, respectively, and the output is a source domain reconstructed sample. Reconstructing samples from the target domain Loss through source domain reconstruction and target domain reconstruction loss Applying L2 norm constraints yields the total reconstruction loss. Its expression is as follows: ； ； ； in, Represent the L2 norm; obtain the final loss: ； By training the model to convergence, we can effectively extract domain-invariant features and successfully transfer cross-domain knowledge, enabling the model to achieve classification accuracy in the target domain that is similar to that in the source domain, thereby solving the problem of distribution differences in domain adaptation.