Rotating machine cross-working-condition fault diagnosis method based on dynamic distance polarization optimal transmission
By employing the dynamic distance polarization optimal transmission method, and utilizing the adaptive adjustment of intra-class and inter-class probability thresholds, combined with target domain pseudo-labels and mask matrices, the transmission uncertainty and misdiagnosis problems in cross-operating condition fault diagnosis are solved, achieving highly accurate fault diagnosis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NORTHWESTERN POLYTECHNICAL UNIV
- Filing Date
- 2026-02-06
- Publication Date
- 2026-05-12
AI Technical Summary
Existing fault diagnosis methods based on optimal transmission cannot guarantee the correctness of transmission and the effectiveness of diagnosis when operating conditions change, especially in fault diagnosis across operating conditions, which is prone to erroneous transmission and misdiagnosis.
The dynamic distance polarization optimal transmission method is adopted. By constructing a distance polarization regularizer, defining intra-class probability thresholds and inter-class probability thresholds, designing a penalized regularization term and gradient optimization mechanism, and combining target domain pseudo-labels and dynamic mask matrix, the polarization threshold is adaptively adjusted. The optimal transmission plan is obtained through supervised training by centroid mapping.
It reduces the uncertainty of cross-domain transmission and improves the accuracy and effectiveness of fault diagnosis, especially performing well under conditions with large differences in operating conditions.
Smart Images
Figure CN122020390A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of mechanical engineering technology, specifically to a method for fault diagnosis of rotating machinery under different operating conditions based on dynamic distance polarization optimal transmission. Background Technology
[0002] During the operation of rotating machinery, critical components such as rolling bearings inevitably fail. Fault diagnosis of these components is crucial. However, deep learning models face the problem of "domain offset" in engineering applications: that is, a model trained under one working condition cannot be directly deployed to another working condition lacking labeled data due to changes in the working condition.
[0003] Traditional fault diagnosis methods based on optimal transmission mainly diagnose faults through distance-regularized optimal transmission and coupled-regularized optimal transmission methods. The distance-regularized optimal transmission method indirectly constrains the transmission plan by adjusting the transmission cost, aiming to achieve distribution alignment between the source and target domains by minimizing the total transmission cost. The coupled-regularized optimal transmission method applies a regularization term to the transmission plan matrix, reducing computational complexity and achieving distribution alignment through the constraint of the transmission plan. Obviously, this fault diagnosis method based on optimal transmission has at least the following shortcomings: 1. The distance-regularized optimal transmission method lacks explicit modeling of the transmission plan. When the operating conditions differ significantly, it cannot guarantee the correctness of the transmission and is prone to erroneous transmission.
[0004] 2. The coupling regularization optimal transmission method leads to an overly dense transmission plan. This dense transmission makes the similarity judgment between specific sample pairs ambiguous, which greatly increases the uncertainty of the cross-domain transmission process. It also ignores the fault label information and cannot distinguish the transmission correspondence within and between classes, resulting in the incorrect matching of samples of different categories, causing misdiagnosis and failing to guarantee the effectiveness of fault diagnosis. Summary of the Invention
[0005] To address the aforementioned technical shortcomings, the present invention aims to provide a method for fault diagnosis of rotating machinery under various operating conditions based on dynamic distance polarization optimal transmission.
[0006] To solve the above technical problems, the present invention adopts the following technical solution: The present invention provides a method for fault diagnosis of rotating machinery across working conditions based on dynamic distance polarization optimal transmission, including: S1, constructing a distance polarization regularizer: based on the idea of large margin learning, by defining intra-class probability thresholds and inter-class probability thresholds, and designing a penalized regularization term and gradient optimization mechanism, the polarization direction of the optimal transmission plan is explicitly constrained.
[0007] S2. Dynamic Distance Polarization Regularization Strategy: The target domain pseudo-label is fused with the dynamic mask matrix to achieve adaptive adjustment of the polarization threshold, guiding the optimal transmission plan to the correct category plan direction polarization.
[0008] S3. Training of target classifier based on centroid mapping: Obtain the optimal transmission plan, and according to the optimal transmission plan, map the source domain labeled features to the target domain feature space to generate pseudo target features, and conduct supervised training of the target domain classifier.
[0009] The beneficial effects of this invention are as follows: 1. This invention provides a method for cross-operating condition fault diagnosis of rotating machinery based on dynamic distance polarization optimal transmission. First, based on the idea of large margin learning, it defines intra-class probability thresholds and inter-class probability thresholds, and designs a penalized regularization term and gradient optimization mechanism to explicitly constrain the polarization direction of the optimal transmission plan. Second, it fuses the target domain pseudo-label with the dynamic mask matrix to achieve adaptive adjustment of the polarization threshold, guiding the optimal transmission plan to the correct category plan direction polarization. Finally, it obtains the optimal transmission plan, and based on the optimal transmission plan, maps the source domain labeled features to the target domain feature space to generate pseudo-target features. It also performs supervised training on the target domain classifier, reducing the uncertainty of the cross-domain transmission process, ensuring the correctness of the transmission, and ensuring the effectiveness of fault diagnosis.
[0010] This invention reduces the intra-class sample transmission probability and increases the inter-class sample transmission probability through large-margin learning. It extracts source domain sample features and target domain sample features through a shared feature extractor, calculates the similarity distance between source domain samples and target domain samples, defines intra-class probability thresholds and inter-class probability thresholds, designs a penalized regularization term and defines the penalty logic, and performs gradient derivation to obtain the polarization direction of the optimal transmission plan. By combining the distance polarization regularization term, transmission cost and entropy regularization term, an optimal transmission optimization objective containing polarization constraints is constructed, ensuring the accuracy of sample matching.
[0011] This invention acquires target domain features through a shared feature extractor and calculates the cluster centroids of various faults in the target domain. Simultaneously, it calculates the similarity between target domain samples and the cluster centroids of various faults using Euclidean distance, assigns the target domain samples to the fault category with the highest similarity, obtains target domain pseudo-labels, constructs inter-class and intra-class mask matrices, and relaxes and adjusts the initial polarization threshold based on these matrices. The adjusted initial polarization threshold is then incorporated into the regularization term to obtain a dynamic distance polarization regularizer. Furthermore, it optimizes the loss and directionally polarizes the transmission probability, accurately capturing the correspondence between intra- and inter-class faults and ensuring the correctness of fault diagnosis.
[0012] This invention obtains source domain features and target domain features through a shared feature extractor and solves the optimal transmission plan. At the same time, it maps the source domain features to the target domain feature space to obtain pseudo-target features. The pseudo-target features are paired with the real labels in the source domain to obtain training data for the target classifier. The target domain classifier is then trained under supervision to learn the fault category discrimination boundary in the target domain feature space, thereby increasing the accuracy of fault diagnosis. Attached Figure Description
[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0014] Figure 1 This is a schematic diagram of the implementation steps of the method of the present invention.
[0015] Figure 2 This is a flowchart of the algorithm in the method of the present invention.
[0016] Figure 3 This is the cross-domain fault diagnosis framework of the method of the present invention.
[0017] Figure 4 This refers to intra-class relaxation and inter-class relaxation in the method of this invention.
[0018] Figure 5 This is the experimental platform for the Soochow University bearing dataset of this invention.
[0019] Figure 6 This refers to the fault categories included in the Soochow University bearing dataset experiments of this invention.
[0020] Figure 7 This is the domain of the Soochow University bearing dataset experiment of this invention.
[0021] Figure 8 This invention demonstrates the robustness of the Soochow University bearing dataset experiment.
[0022] Figure 9 This is the experimental platform for the Paderborn University bearing dataset of this invention.
[0023] Figure 10 This refers to the fault categories included in the Paderborn University bearing dataset experiments of this invention.
[0024] Figure 11 This is the domain of the Paderborn University bearing dataset experiment of this invention. Detailed Implementation
[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0026] Please see Figure 1 , Figure 2 and Figure 3 As shown, the present invention provides a method for fault diagnosis of rotating machinery across operating conditions based on dynamic distance polarization optimal transmission, including: S1, constructing a distance polarization regularizer: based on the idea of large margin learning, by defining intra-class probability thresholds and inter-class probability thresholds, and designing a penalized regularization term and gradient optimization mechanism, the polarization direction of the optimal transmission plan is explicitly constrained.
[0027] In a specific embodiment, the construction of the distance polarization regularizer is carried out as follows: the intra-class sample transmission probability is reduced and the inter-class sample transmission probability is increased through large margin learning, and the source domain sample features and target domain sample features are extracted through a shared feature extractor. The similarity distance between the source domain sample and the target domain sample is calculated, and the intra-class probability threshold and the inter-class probability threshold are defined.
[0028] It should be noted that large margin learning essentially constructs a discrimination margin by defining the lower bound of intra-class and inter-class transfer probabilities, and then uses regularization penalties to force intra-class sample transfer probabilities to polarize towards higher values and inter-class sample transfer probabilities to polarize towards lower values.
[0029] The high value is 1, and the low value is 0.
[0030] It should be noted that defining the intra-class probability threshold and the inter-class probability threshold means defining the intra-class probability threshold and the inter-class probability threshold separately for each fault category, that is... , In the formula Representing the Inter-class probability threshold for class samples Representing the Within-class probability threshold for class samples Represents small parameters, Represents the number of samples in the source domain. Represents the target domain Number of samples per class Represents the number of samples in the target domain. This represents the fault category number.
[0031] It should also be noted that the fault categories include inner ring faults, ball faults, and outer ring faults.
[0032] Among them, the number of samples in the source domain and the number of samples in the target domain The number of class samples and the number of target domain samples are both obtained from the database.
[0033] It should be explained that both the source domain and the target domain represent operating conditions, but the operating conditions represented by the source domain are different from those represented by the target domain.
[0034] Design a penalized regularization term and define the penalty logic. At the same time, perform gradient derivation to obtain the polarization direction of the optimal transmission plan. Combine the distance polarization regularization term, transmission cost and entropy regularization term to construct the optimal transmission optimization objective with polarization constraints.
[0035] The calculation of the similarity distance between the source domain sample and the target domain sample described above is as follows: The expression for the similarity distance is: In the formula Representative source domain samples , Represents the source domain sample number, Representative target domain sample , Represents the sample number of the target domain. Represents the characteristics of source domain samples. The target represents the characteristics of the sample in the target domain, and , This represents the similarity distance between the source domain sample and the target domain sample.
[0036] It should be noted that the target domain sample and source domain samples All data is retrieved from the database.
[0037] The above describes the design of a penalized regularization term and the definition of the penalty logic. The specific process is as follows: The expression for the penalized regularization term is: In the formula Represents the distance polarization regularization term. This represents the probability that a sample from the source domain will be transmitted to a sample from the target domain. Representing the source domain Inter-class probability threshold for class samples Represents the fault category number. Representing the source domain The intra-class probability threshold for class samples.
[0038] when When the penalty-type regularization term outputs a non-zero penalty value, the penalty-type regularization term will be used.
[0039] The specific process for obtaining the polarization direction of the optimal transmission plan through gradient derivation, as described above, is as follows: Definition If the class sample threshold center, then In the formula represent If the threshold center of the class sample is used, then the regularization term is applied to... The gradient formula is: In the formula Represents regularization terms The gradient.
[0040] Will and To make a comparison, if ,but , Increase, transfer probability polarization within the class, if ,but , Decrease the probability polarization of inter-class transmission.
[0041] The specific process for constructing the optimal transmission optimization objective with polarization constraints described above is as follows: ; In the formula Represents the optimal transmission plan matrix. , These represent the weighting coefficients of the entropy regularization term and the distance polarization regularization term, respectively. Represents the entropy regularization term. Represents the distance polarization regularization term. Represents the transmission cost matrix. This represents the transmission plan matrix.
[0042] It should be noted that the double random constraint set, the domain marginal distribution, and the target domain marginal distribution are all obtained from the database, and the weight coefficients of the entropy regularization term and the distance polarization regularization term are set by relevant staff.
[0043] It should also be noted that, .
[0044] S2. Dynamic Distance Polarization Regularization Strategy: The target domain pseudo-label is fused with the dynamic mask matrix to achieve adaptive adjustment of the polarization threshold, guiding the optimal transmission plan to the correct category plan direction polarization.
[0045] In a specific embodiment, the dynamic distance polarization regularization strategy is implemented as follows: target domain features are obtained through a shared feature extractor, and the cluster centroids of each fault category in the target domain are calculated. At the same time, the similarity between the target domain samples and the cluster centroids of each fault category is calculated through Euclidean distance. The target domain samples are then assigned to the fault category with the highest similarity to obtain the target domain pseudo-label.
[0046] It should be noted that, In the formula Represents the target domain Cluster centroids for each fault category Represents the initial pseudo-label of the target domain. Represents the characteristics of the target domain.
[0047] Construct the inter-class mask matrix and the intra-class mask matrix, then In the formula Represents the inter-class mask matrix. The true labels representing the source domain samples, Represents pseudo-tags for the target domain. Represents the training batch size. In the formula The intra-class mask matrix represents the initial polarization threshold that is relaxed based on the inter-class mask matrix and the intra-class mask matrix.
[0048] The adjusted initial polarization threshold is incorporated into the regularization term to obtain the dynamic distance polarization regularizer, i.e. In the formula This represents the dynamic distance polarization regularization term. By combining the dynamic distance polarization regularizer, transmission cost, and entropy regularization term, an optimal transmission optimization objective with dynamic polarization constraints is constructed. Simultaneously, optimization is achieved through gradient descent. The total loss will be used to directionally polarize the transmission probability.
[0049] Please see Figure 4 As shown above, the relaxation adjustment of the initial polarization threshold is specifically performed as follows: Intra-class relaxation adjustment: In the formula This represents the inter-class threshold after relaxation adjustment. This represents the initial inter-class threshold matrix.
[0050] Inter-class relaxation adjustment: In the formula This represents the intra-class threshold after relaxation adjustment. This represents the initial intra-class threshold matrix.
[0051] It should be noted that both the initial inter-class threshold matrix and the initial intra-class threshold matrix were obtained from the database.
[0052] In the above, the optimization The total loss is used to polarize the transmission probability in a specific way, as follows: pass Obtain the new threshold center and will With In comparison, if ,but , Increase, and propagation probability polarization within the class, where Represents the dynamic distance polarization regularization term pair The gradient, if ,but , Decrease the probability polarization of inter-class transmission.
[0053] S3. Training of target classifier based on centroid mapping: Obtain the optimal transmission plan, and according to the optimal transmission plan, map the source domain labeled features to the target domain feature space to generate pseudo target features, and conduct supervised training of the target domain classifier.
[0054] In a specific embodiment, the training of the target classifier based on centroid mapping involves the following steps: Source domain features and target domain features are obtained through a shared feature extractor, and then the optimal transport plan is solved. Simultaneously, the source domain features are mapped to the target domain feature space to obtain pseudo-target features, i.e. In the formula Represents the characteristics of false targets. Representative target domain sample Features Elements representing the optimal transmission plan. This represents the total number of samples in the target domain.
[0055] It should be noted that the process of solving the optimal transmission plan is as follows: It is solved using the GCG algorithm, through... Decompose the objective function to facilitate DDPOT optimization, where Represents a convex compact constraint set. , If the initial transmission coupling matrix is set to a uniform distribution, then In the formula Represents the initial transmission plan matrix. represent The solution is a full 1 matrix, and the gradient is integrated into the cost matrix. The entropy-regularized OT problem is solved by the Sinkhorn algorithm. Then, the difference between the current solution and the solution of the subproblem is calculated, and the optimal step size is found by line search. Finally, the solution is updated to obtain the optimal transport plan.
[0056] Pair pseudo-target features with real labels from the source domain to obtain training data for the target classifier. Then, supervise the training of the target domain classifier to learn the fault category discrimination boundary in the target domain feature space.
[0057] It should be noted that the training data for the target classifier is... .
[0058] It should also be noted that the specific process of supervising the training of the target domain classifier to learn the fault category discrimination boundary in the target domain feature space is as follows: the pseudo-target features are input into the target domain classifier, the fault category prediction probability is output, the loss function is designed using cross-entropy loss, and the target domain classification loss and the source domain classification loss are integrated into the total loss as the basis for model update. During the cross-domain training phase, the model parameters are updated by random gradient descent based on the gradient of the total loss to ensure that the target classifier is gradually optimized. When the training epoch reaches the preset maximum value, the training is stopped, the final model parameters are saved, and the effectiveness of the target classifier is verified by experiments after training is completed.
[0059] Example 1: The Soochow University (SCU) bearing dataset was used, and the experimental platform was as follows: Figure 5 As shown, the experiment includes 10 fault categories, such as... Figure 6 As shown, the experiment defined the operating conditions into four domains, A, B, C, and D, based on different load torques (0, 1000, 2000, 3000 Nm) and a fixed speed of 896 rpm. Figure 7 As shown, 12 cross-domain diagnostic tasks were performed on this dataset, such as A→B, A→C, A→D, C→A, etc. The average diagnostic accuracy of this invention reached 92.23%. In comparison, the accuracy of MK-MMD was 85.20%, JMMD was 87.65%, JDOT was 86.60%, CKBWD was 89.26%, and EROT was 86.81%. This invention performed best among all the comparison methods, especially in AD and CA tasks with large differences in operating conditions, where its robustness advantage was obvious. Figure 8 As shown.
[0060] Example 2: The Paderborn University (PU) bearing dataset was used, and the experimental platform was as follows: Figure 9 As shown, this dataset is characterized by including real damage generated by accelerated life testing, which increases the difficulty of diagnosis. The experiment includes six fault categories, such as... Figure 10 As shown, the operating conditions are divided into four domains: E, F, G, and H, based on different rotational speeds (900, 1500 r / min), radial forces (400, 1000 N), and torques (0.1, 0.7 Nm). Figure 11As shown, 12 cross-domain diagnostic tasks were performed on this dataset, such as E→F, E→G, G→F, H→G, etc. The average diagnostic accuracy of this invention reached 89.44%, which is also the best among all the comparison methods. For comparison, the average accuracies of other methods are as follows: MK-MMD accuracy is 876.53%, JMMD accuracy is 872.28%, JDOT accuracy is 876.40%, CKBWD accuracy is 881.46%, and EROT accuracy is 876.94%. In the GF task with huge differences in working conditions, the accuracy of this method reached 96.31%, which is 15.17% higher than the second best CKBWD method.
[0061] This invention first utilizes the large margin learning concept, defining intra-class and inter-class probability thresholds, and designing penalized regularization terms and gradient optimization mechanisms to explicitly constrain the polarization direction of the optimal transmission plan. Secondly, it fuses the target domain pseudo-label with a dynamic mask matrix to adaptively adjust the polarization threshold, guiding the optimal transmission plan to correctly polarize the class plan direction. Finally, it obtains the optimal transmission plan and, based on it, maps the source domain labeled features to the target domain feature space, generating pseudo-target features. Supervised training of the target domain classifier is then performed, reducing the uncertainty of cross-domain transmission, ensuring the correctness of transmission, and guaranteeing the effectiveness of fault diagnosis.
[0062] The examples described in this invention are not limited to the specific embodiments listed above. The examples are merely illustrative to facilitate understanding of the invention and do not constitute a limitation on the scope of protection of this invention. Any modifications, equivalent substitutions, etc., made within the spirit and principles of this invention should be included within the scope of protection.
[0063] The above description is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined in this specification, they should all fall within the protection scope of the present invention.
Claims
1. A method for fault diagnosis of rotating machinery across operating conditions based on dynamic distance polarization optimal transmission, characterized in that, Includes the following steps: S1. Construct a distance polarization regularizer: Based on the idea of large margin learning, by defining intra-class probability thresholds and inter-class probability thresholds, and designing a penalized regularization term and gradient optimization mechanism, the polarization direction of the optimal transmission plan is explicitly constrained. S2. Dynamic distance polarization regularization strategy: The target domain pseudo-label is fused with the dynamic mask matrix to achieve adaptive adjustment of the polarization threshold and guide the optimal transmission plan to the correct category plan direction polarization. S3. Training of target classifier based on centroid mapping: Obtain the optimal transmission plan, and according to the optimal transmission plan, map the source domain labeled features to the target domain feature space to generate pseudo target features, and conduct supervised training of the target domain classifier.
2. The method for fault diagnosis of rotating machinery across operating conditions based on dynamic distance polarization optimal transmission according to claim 1, characterized in that, The specific process for constructing the distance polarization regularizer is as follows: By using large margin learning to reduce the intra-class sample transfer probability and increase the inter-class sample transfer probability, and by using a shared feature extractor to extract source domain sample features and target domain sample features, the similarity distance between source domain samples and target domain samples is calculated, and intra-class probability thresholds and inter-class probability thresholds are defined. Design a penalized regularization term and define the penalty logic. At the same time, perform gradient derivation to obtain the polarization direction of the optimal transmission plan. Combine the distance polarization regularization term, transmission cost and entropy regularization term to construct the optimal transmission optimization objective with polarization constraints.
3. The method for fault diagnosis of rotating machinery across operating conditions based on dynamic distance polarization optimal transmission according to claim 2, characterized in that, The specific process for calculating the similarity distance between the source domain sample and the target domain sample is as follows: The expression for similarity distance is: In the formula Representative source domain samples , Represents the source domain sample number, Representative target domain sample , Represents the sample number of the target domain. Represents the characteristics of source domain samples. The target represents the characteristics of the sample in the target domain, and , This represents the similarity distance between the source domain sample and the target domain sample.
4. The method for fault diagnosis of rotating machinery across operating conditions based on dynamic distance polarization optimal transmission according to claim 2, characterized in that, The design of the penalized regularization term and the definition of the penalty logic are as follows: The expression for the penalized regularization term is: In the formula Represents the distance polarization regularization term. This represents the probability that a sample from the source domain will be transmitted to a sample from the target domain. Representing the source domain Inter-class probability threshold for class samples Represents the fault category number. Representing the source domain The intra-class probability threshold for class samples; when When the penalty-type regularization term outputs a non-zero penalty value, the penalty-type regularization term will be used.
5. The method for fault diagnosis of rotating machinery across operating conditions based on dynamic distance polarization optimal transmission according to claim 2, characterized in that, The specific process for obtaining the polarization direction of the optimal transmission plan through gradient derivation is as follows: definition If the class sample threshold center, then In the formula represent If the threshold center of the class sample is used, then the regularization term is applied to... The gradient formula is: In the formula Represents regularization terms The gradient; Will and To make a comparison, if ,but , Increase, transfer probability polarization within the class, if ,but , Decrease the probability polarization of inter-class transmission.
6. The method for fault diagnosis of rotating machinery across operating conditions based on dynamic distance polarization optimal transmission according to claim 2, characterized in that, The specific process for constructing the optimal transmission optimization objective that includes polarization constraints is as follows: ； In the formula Represents the optimal transmission plan matrix. , These represent the weighting coefficients of the entropy regularization term and the distance polarization regularization term, respectively. Represents the entropy regularization term. Represents the distance polarization regularization term. Represents the transmission cost matrix. This represents the transmission plan matrix.
7. The method for fault diagnosis of rotating machinery under different operating conditions based on dynamic distance polarization optimal transmission according to claim 1, characterized in that, The dynamic distance polarization regularization strategy is described in the following process: The target domain features are obtained by using a shared feature extractor, and the cluster centroids of various faults in the target domain are calculated. At the same time, the similarity between the target domain samples and the cluster centroids of various faults is calculated by Euclidean distance. The target domain samples are assigned to the fault category with the highest similarity to obtain the target domain pseudo label. Construct the inter-class mask matrix and the intra-class mask matrix, then In the formula Represents the inter-class mask matrix. The true labels representing the source domain samples, Represents pseudo-tags for the target domain. Represents the training batch size. In the formula The intra-class mask matrix represents the initial polarization threshold, which is then relaxed and adjusted based on the inter-class and intra-class mask matrices. The adjusted initial polarization threshold is incorporated into the regularization term to obtain the dynamic distance polarization regularizer, i.e. In the formula This represents the dynamic distance polarization regularization term. By combining the dynamic distance polarization regularizer, transmission cost, and entropy regularization term, an optimal transmission optimization objective with dynamic polarization constraints is constructed. Simultaneously, optimization is achieved through gradient descent. The total loss will be used to directionally polarize the transmission probability.
8. The method for fault diagnosis of rotating machinery under different operating conditions based on dynamic distance polarization optimal transmission according to claim 7, characterized in that, The relaxation adjustment of the initial polarization threshold is performed as follows: Intraclass relaxation adjustment: In the formula This represents the inter-class threshold after relaxation adjustment. Represents the initial inter-class threshold matrix; Inter-class relaxation adjustment: In the formula This represents the intra-class threshold after relaxation adjustment. This represents the initial intra-class threshold matrix.
9. The method for fault diagnosis of rotating machinery across operating conditions based on dynamic distance polarization optimal transmission according to claim 7, characterized in that, The optimization The total loss is used to polarize the transmission probability in a specific way, as follows: pass Obtain the new threshold center and will With In comparison, if ,but , Increase, intra-class transmission probability polarization, where Represents the dynamic distance polarization regularization term pair The gradient, if ,but , Decrease the probability polarization of inter-class transmission.
10. The method for fault diagnosis of rotating machinery under different operating conditions based on dynamic distance polarization optimal transmission according to claim 1, characterized in that, The specific process for training the target classifier based on centroid mapping is as follows: By acquiring source and target domain features through a shared feature extractor, the optimal transmission plan is solved. Simultaneously, the source domain features are mapped to the target domain feature space to obtain pseudo-target features. In the formula Represents the characteristics of false targets. Representative target domain sample Features Elements representing the optimal transmission plan. Represents the total number of samples in the target domain; Pair pseudo-target features with real labels from the source domain to obtain training data for the target classifier. Then, supervise the training of the target domain classifier to learn the fault category discrimination boundary in the target domain feature space.