CN-122020390-A - Rotary machine cross-working condition fault diagnosis method for dynamic distance polarization optimal transmission

CN122020390ACN 122020390 ACN122020390 ACN 122020390ACN-122020390-A

Abstract

The invention discloses a rotating machinery cross-working condition fault diagnosis method for optimal transmission of dynamic distance polarization, which relates to the technical field of mechanical engineering, and comprises the steps of firstly defining an intra-class probability threshold and an inter-class probability threshold, designing a punishment regularization item and a gradient optimization mechanism, displaying the polarization direction of a constraint optimal transmission plan, fusing a target domain pseudo tag with a dynamic mask matrix, the method comprises the steps of realizing self-adaptive adjustment of a polarization threshold, guiding the optimal transmission plan to be polarized in the correct category plan direction, finally obtaining the optimal transmission plan, mapping the source domain tagged features to the target domain feature space according to the optimal transmission plan, generating pseudo-target features, performing supervision training on the target domain classifier, reducing uncertainty in the cross-domain transmission process, guaranteeing the transmission accuracy and guaranteeing the fault diagnosis effectiveness of the rotating mechanical equipment.

Inventors

YU LIANG
YAN FUCHENG
WANG RAN

Assignees

西北工业大学

Dates

Publication Date: 20260512
Application Date: 20260206

Claims (10)

1. A method for diagnosing a cross-working condition fault of a rotary machine for optimal transmission of dynamic distance polarization is characterized by comprising the following steps: S1, constructing a distance polarization regularizer, namely, based on a large marginal learning thought, displaying a polarization direction of a constraint optimal transmission plan by defining an intra-class probability threshold and an inter-class probability threshold and designing a punishment regularization item and a gradient optimization mechanism; S2, a dynamic distance polarization regularization strategy is that a target domain pseudo tag is fused with a dynamic mask matrix to realize self-adaptive adjustment of a polarization threshold value, and polarization of the optimal transmission plan direction in the correct category plan direction is guided; and S3, training the target classifier based on gravity center mapping, namely acquiring an optimal transmission plan, mapping the source domain tagged features to a target domain feature space according to the optimal transmission plan, generating pseudo target features, and performing supervision training on the target domain classifier.
2. The method for diagnosing a cross-working condition fault of a rotating machine for optimal transmission of dynamic distance polarization according to claim 1, wherein the method for constructing the distance polarization regularizer is characterized by comprising the following specific steps: Reducing the intra-class sample transmission probability and expanding the inter-class sample transmission probability through large marginal learning, extracting the source domain sample characteristics and the target domain sample characteristics through a shared characteristic extractor, calculating the similar distance between the source domain sample and the target domain sample, and defining an intra-class probability threshold and an inter-class probability threshold; and designing a punishment regularization term, defining punishment logic, simultaneously carrying out gradient deduction to obtain the polarization direction of the optimal transmission plan, and combining the distance polarization regularization term, the transmission cost and the entropy regularization term to construct an optimal transmission optimization target containing polarization constraint.
3. The method for diagnosing a cross-working condition fault of a rotating machine for optimal transmission of dynamic distance polarization according to claim 2, wherein the calculating of the similarity distance between a source domain sample and a target domain sample comprises the following steps: The expression for the similar distance is: in the following Representative Source Domain sample , A sample number representing the source domain, Representative target domain sample , A sample number representing the target field, Representing the characteristics of the source domain samples, The object represents the sample feature of the object domain, and , Representing a similar distance of the source domain sample from the target domain sample.
4. The method for diagnosing a cross-working condition fault of a rotating machine for optimal transmission of dynamic distance polarization according to claim 2, wherein the method is characterized by designing a penalty type regularization term and defining penalty logic, and comprises the following specific processes: The expression of the penalty regularization term is: in the following Representing a distance polarization regularization term, Representing the probability of a source domain sample being transmitted to a target domain sample, Representing source domain number An inter-class probability threshold for the class samples, Representing the number of the fault class, Representing source domain number Intra-class probability threshold for class samples when When the penalty regularization term outputs a non-zero penalty value.
5. The method for diagnosing a cross-working condition fault of a rotating machine for optimal transmission of dynamic distance polarization according to claim 2, wherein the step of deriving the gradient to obtain the polarization direction of the optimal transmission plan comprises the following steps: Definition of the definition Class sample threshold center, then In the following Representative of Class sample threshold center, regularizing term pairs The gradient formula of (2) is: in the following Representing regularized term pairs Is a gradient of (2); Will be And (3) with For comparison, if Then , Increasing, transmitting probability polarization into class, if Then , And (5) reducing the probability polarization transmitted to the classes.
6. The method for diagnosing a cross-working condition fault of a rotating machine for optimal transmission of dynamic distance polarization according to claim 2, wherein the construction of the optimal transmission optimization target containing polarization constraint comprises the following specific processes: ; In the middle of Representing an optimal transmission plan matrix, 、 Respectively representing the weight coefficient of the entropy regularization term and the weight coefficient of the distance polarization regularization term, Representing the term of regularization of the entropy, Representing a distance polarization regularization term, Representing a matrix of transmission costs and, Representing a transmission plan matrix.
7. The method for diagnosing a cross-working condition fault of a rotating machine for optimal transmission of dynamic distance polarization according to claim 1, wherein the dynamic distance polarization regularization strategy comprises the following specific steps: The method comprises the steps of obtaining target domain features through a shared feature extractor, calculating the cluster gravity centers of various faults in a target domain, calculating the similarity between a target domain sample and the cluster gravity centers of various faults through Euclidean distances, and distributing the target domain sample to a fault class with the highest similarity to obtain a target domain pseudo tag; Constructing an inter-class mask matrix and an intra-class mask matrix In the following Represents the mask matrix between classes, A real label representing a sample of the source domain, Representing the pseudo tag of the target domain, Representing the size of the training batch and, In the following Representing an intra-class mask matrix, and performing relaxation adjustment on an initial polarization threshold according to the inter-class mask matrix and the intra-class mask matrix; the adjusted initial polarization threshold value is fused into a regularization term to obtain a dynamic distance polarization regularizer, namely In the following Representing a dynamic distance polarization regularization term, and combining the dynamic distance polarization regularizer, the transmission cost and the entropy regularization term to construct an optimal transmission optimization target containing dynamic polarization constraint, namely Simultaneously optimized by gradient descent The transmission probability is directionally polarized.
8. The method for diagnosing a cross-operating condition fault of a rotating machine for optimal transmission of dynamic distance polarization according to claim 7, wherein the step of performing relaxation adjustment on an initial polarization threshold is as follows: in-class slack adjustment: in the following Representing the inter-class threshold after the relaxation adjustment, Representing an initial inter-class threshold matrix; Inter-class slack adjustment: in the following Representing the relaxed intra-class threshold value, Representing an initial intra-class threshold matrix.
9. The method for diagnosing a cross-operating condition fault of a rotating machine for optimal transmission of dynamic distance polarization according to claim 7, wherein said optimizing comprises The transmission probability is directionally polarized according to the total loss of the transmission probability, and the specific process is as follows: By passing through Acquiring a new threshold center And will And proceed with In contrast, if Then , Increasing, transmitting probability polarization into class, wherein Representing pairs of dynamic distance polarization regularization terms If the gradient of (1) Then , And (5) reducing the probability polarization transmitted to the classes.
10. The method for diagnosing a cross-working condition fault of a rotating machine for optimal transmission of dynamic distance polarization according to claim 1, wherein the training of the target classifier based on gravity center mapping comprises the following specific processes: acquiring source domain features and target domain features through a shared feature extractor, further solving an optimal transmission plan, and simultaneously mapping the source domain features to a target domain feature space to acquire pseudo-target features, namely In the following Representing the characteristics of the pseudo-object, Representative target domain sample Is characterized in that, Representing an element of the optimal transmission plan, Representing the total number of target domain samples; And matching the pseudo target features with the source domain real labels to obtain training data of the target classifier, performing supervision training on the target domain classifier, and learning fault class discrimination boundaries in the target domain feature space.

Description

Rotary machine cross-working condition fault diagnosis method for dynamic distance polarization optimal transmission Technical Field The invention relates to the technical field of mechanical engineering, in particular to a rotating machinery cross-working condition fault diagnosis method for optimal transmission of dynamic distance polarization. Background In the running process of the rotary machine, key components such as a rolling bearing and the like inevitably fail, and the fault diagnosis of the components is very important, but a deep learning model faces the problem of domain offset in engineering application, namely a model trained in one working condition cannot be directly deployed to the other working condition lacking label data due to the change of the working condition. The conventional fault diagnosis method based on optimal transmission is mainly used for diagnosing faults through a distance regularization optimal transmission method and a coupling regularization optimal transmission method, the distance regularization optimal transmission method is used for indirectly restraining a transmission plan by adjusting transmission cost, the distribution alignment of a source domain and a target domain is realized by taking the minimum total transmission cost as a target, the coupling regularization optimal transmission method is used for applying regularization terms to a transmission plan matrix, the calculation complexity is reduced and the distribution alignment is realized through the restraint of the transmission plan, and obviously, the fault diagnosis method based on optimal transmission at least has the following defects that 1, the distance regularization optimal transmission method is used for lacking explicit modeling on the transmission plan, when working condition differences are obvious, the accuracy of transmission cannot be guaranteed, and error transmission is easy to generate. 2. The optimal transmission method of the coupling regularization can cause too dense transmission plans, and the dense transmission can cause the similarity judgment between specific sample pairs to be fuzzy, so that the uncertainty of the cross-domain transmission process is greatly increased, the label information of faults is ignored, the corresponding relation of the transmission in-class and the transmission between classes cannot be distinguished, samples of different classes are erroneously matched, misdiagnosis is caused, and the effectiveness of the fault diagnosis cannot be guaranteed. Disclosure of Invention Aiming at the technical defects, the invention aims to provide a cross-working condition fault diagnosis method for the rotating machinery with optimal transmission of dynamic distance polarization. The invention provides a rotating machinery cross-working condition fault diagnosis method for optimal transmission of dynamic distance polarization, which comprises the following steps of S1, constructing a distance polarization regularizer, and displaying a polarization direction of a constraint optimal transmission plan by defining an intra-class probability threshold and an inter-class probability threshold and designing a punishment regularization item and a gradient optimization mechanism based on a large marginal learning thought. S2, a dynamic distance polarization regularization strategy is that a target domain pseudo tag is fused with a dynamic mask matrix to realize self-adaptive adjustment of a polarization threshold value, and polarization of the optimal transmission plan direction in the correct category plan direction is guided. And S3, training the target classifier based on gravity center mapping, namely acquiring an optimal transmission plan, mapping the source domain tagged features to a target domain feature space according to the optimal transmission plan, generating pseudo target features, and performing supervision training on the target domain classifier. The invention has the beneficial effects that 1, the invention provides a rotating machinery cross-working condition fault diagnosis method for optimal transmission of dynamic distance polarization, firstly, based on a large margin learning thought, the polarization direction of a constraint optimal transmission plan is displayed by defining an intra-class probability threshold and an inter-class probability threshold and designing a punishment regularization item and a gradient optimization mechanism, secondly, a target domain pseudo tag and a dynamic mask matrix are fused to realize self-adaptive adjustment of the polarization threshold, the polarization of the optimal transmission plan in the correct class plan direction is guided, finally, an optimal transmission plan is obtained, and the source domain tag feature is mapped to a target domain feature space according to the optimal transmission plan to generate a pseudo target feature, and the target domain classifier is supervised and trained, so that the uncertain