CN-121980502-A - Multi-mode federal cross-domain fault diagnosis method based on prototype comparison learning and application

CN121980502ACN 121980502 ACN121980502 ACN 121980502ACN-121980502-A

Abstract

The invention discloses a multi-mode federal cross-domain fault diagnosis method based on prototype comparison learning and application thereof, wherein the method comprises the following steps: each source client receives the initialized global model and the prototype and performs local model training by using the local data. In the training process, the source client side excavates potential essential association in the multi-source heterogeneous data through double-hierarchy prototype comparison, effectively breaks through a barrier of modal isomerism and realizes alignment of multi-modal fault characteristics. Then, local model optimization is performed by minimizing the total contrast loss and the classification loss, and local prototypes and local modality prototypes are generated. The central server aggregates the local prototypes and the local modal prototypes to generate global prototypes and global modal prototypes of comprehensive fault information, and aggregates the local models to construct global models with good generalization capability. The method solves the problems of modal isomerism and data distribution difference existing in multi-modal federal learning, and effectively improves the accuracy and generalization of global model cross-domain fault diagnosis.

Inventors

Wan Lanjun
SUN TAO
TAN HONGWEI
HONG DA
NI WEI
WU YUEZHONG
SHI BANGBING

Assignees

湖南工业大学

Dates

Publication Date: 20260505
Application Date: 20260120

Claims (10)

1. A multi-mode federal cross-domain fault diagnosis method based on prototype comparison learning is characterized by comprising the following steps: S1, collecting Multi-mode signal of machine under different working conditions Personal domain, data of different domains are stored to different domains A client; S2, the central server prototype the global Global modality prototype Global model Initializing and distributing the data to all source clients; S3, the first The individual source clients receive and utilize their local multimodal data Training a local model ; S3.1. Th The individual source client uses the feature extractor to extract the single mode feature The single-mode characteristics are fused by utilizing a multi-head attention mechanism, and fusion characteristics are generated : Wherein the method comprises the steps of Representing the splicing operation, , And Respectively represent the first Three projection matrices of the attention heads, Is a key matrix Dimension(s), To output a fusion matrix, Representing the number of attention heads; S3.2. Fusing features at the overall level Comparing with global prototype, calculating contrast loss : Wherein the method comprises the steps of Is a cosine similarity function, Is a temperature coefficient for adjusting the degree of sharpening of the similarity, Representing categories Global prototypes of (a); s3.3, enhancing the consistency of multi-mode fault characteristics through intra-mode prototype comparison and inter-mode prototype comparison at a mode level: (1) Intra-modal prototype comparison, i.e. unimodal features of the same modality Comparing with the global modal prototype to calculate the intra-modal contrast loss : Wherein the method comprises the steps of Representing a modality A corresponding global modal prototype, Representing a modality Lower category A corresponding global modality prototype; (2) Prototype comparison between modes, i.e. unimodal features in different modes Comparing with the global modal prototype to calculate the contrast loss between modes : Wherein the method comprises the steps of Representing excluding current modalities Is defined by a set of modalities; s3.4. Calculating the total contrast loss And classification loss And uses it to model local area Optimizing; The total contrast loss And classification loss The method comprises the following steps: Wherein the method comprises the steps of , And Is a weight parameter, Represent the first In the individual source client A true label of each sample, Represent the first In the individual source client Individual samples are predicted as categories Probability of (2); S4, extracting and fusing each modal feature by using the optimized local model, and clustering the single-modal features and the fused features to generate a local modal prototype Local prototypes : Wherein the method comprises the steps of Represent the first Mode in individual source clients The following is the first A set of features of each cluster, Representing the number of cluster centers, Represent the first Belonging to the first source client A set of local features for the individual clusters; s5, uploading the local prototype, the local modal prototype and the local model to a central server, and generating and updating a global prototype in a weighted average aggregation mode Global modality prototype Global model : Wherein the method comprises the steps of Represent the first In the individual source client Local prototypes corresponding to each category, Represent the first Mode in individual source clients Lower (th) Local modal prototypes corresponding to each category, Representing possession modalities The number of clients, Is an indication function; s6, judging whether a preset stopping condition is met, if not, returning to the S2, if so, stopping model training, and outputting a fault diagnosis model.
2. The method for diagnosing a multi-modal federal cross-domain fault based on prototype-versus-learning as claimed in claim 1, wherein in step S1, the method is divided into The data in the source domain are labeled data, and the data in the target domain are unlabeled data.
3. The method for diagnosing a multi-modal federal cross-domain fault based on prototype-versus-learning as claimed in claim 1, wherein in step S1, the different clients include different types and numbers of modalities, which are not identical and are shared Seed modality 。
4. The method for diagnosing a multi-modal federal cross-domain fault based on prototype-versus-learning as claimed in claim 1, wherein said multi-modal data in step S3 The method comprises the following steps: Wherein the method comprises the steps of Represent the first Modal set of individual source clients and 、 And Respectively represent the first In the individual source client Each comprises Multimodal sample of individual modalities and corresponding class labels, Represent the first Number of multi-modal samples in a single source client.
5. The method for diagnosing a multi-modal federal cross-domain fault based on prototype-versus-learning as claimed in claim 1, wherein the single-modal feature in step S3.1 is Expressed as: Wherein the method comprises the steps of Represent the first A set of modal characteristics of the multi-modal samples, Representing feature extraction operations and 、 Is the first In the individual source client Personal modality Is a sample of (a).
6. The multi-modal federal cross-domain fault diagnosis method based on prototype-versus-learning of claim 1, wherein the step S3.4 uses a random gradient descent algorithm to model locally And (5) optimizing.
7. The method for multi-modal federal cross-domain fault diagnosis based on prototype-versus-learning as claimed in claim 1, wherein in step S3.4, at the fifth step of The first source client The parameters of the local model in the local training of the wheel are updated as follows: Wherein the method comprises the steps of Represent learning rate, Represent the first Loss function of local model in individual source client.
8. The multi-modal federal cross-domain fault diagnosis method based on prototype-contrast learning according to claim 1, wherein in step S4, the single-modal features and the fusion features are clustered by using a K-Means algorithm.
9. The multi-mode federal cross-domain fault diagnosis method based on prototype-contrastive learning according to claim 1, wherein the preset stop condition in step S6 is a set maximum federal round.
10. The multi-mode federal cross-domain fault diagnosis method based on prototype-comparative learning according to claim 1, wherein the method is applied to fault diagnosis of bearings and rotating electrical machines.

Description

Multi-mode federal cross-domain fault diagnosis method based on prototype comparison learning and application Technical Field The invention relates to the technical field of fault diagnosis, in particular to a multi-mode federal cross-domain fault diagnosis method based on prototype comparison learning and application thereof. Background Existing cross-domain fault diagnosis methods typically rely on a large number of high quality samples. However, in a practical industrial scenario, it is difficult for a single enterprise to provide a large number of high quality training samples. Meanwhile, local data from different enterprises cannot be collected to train a fault diagnosis model due to conflict of interests and privacy protection appeal among the enterprises. In recent years, the advent of federal learning breaks the situation where it is difficult to co-train fault diagnosis models using local data of different enterprises due to privacy protection. Federal transfer learning has the technical advantages of federal learning and transfer learning, and is widely focused in the field of mechanical fault diagnosis. However, the current research of the fault diagnosis method based on federal migration learning is mostly limited to single-mode data, and the robustness of complementary information provided by multi-mode data to the fault diagnosis model is improved. Compared with single-mode federal learning, multi-mode federal learning has obvious advantages in terms of ensuring complementarity of mining multi-mode features under the privacy of data and enhancing robustness of a model, but current fault diagnosis data based on multi-mode federal learning is limited by privacy, availability and the like, and is deficient. In addition, in practical industrial application, the collected multi-mode data often have isomerism due to the difference of monitoring equipment among different enterprises, so that the effectiveness and accuracy of multi-mode learning are reduced. Based on this, a modal heterogeneous federal learning privacy protection method based on a cross-modal prototype is disclosed in CN120180463 a. Firstly, a server sends an initialized global model to a corresponding multi-mode client and a corresponding single-mode client, and local model training is carried out by utilizing the global model. The multimodal client and the unimodal client then upload the generated local prototype and the optimized local model, respectively, to the server. And then, after receiving the local prototype and the local model sent by each client, the server respectively executes prototype aggregation and model aggregation operations. In the prototype aggregation process, a server firstly complements the missing modes in a single-mode client by utilizing multi-mode prototype knowledge, and then aggregates local multi-mode prototypes of all clients in a clustering mode. And finally, after each client receives the global prototype pair and the global model distributed by the server, updating the local model and starting a new round of training. The patent can complement the missing modes in the single-mode client while guaranteeing the data privacy, so that the problem of modal isomerism is effectively solved. However, in the method disclosed in this patent, it is required to ensure that the target client has fault data involved in training, but in the actual fault diagnosis scenario of the rotating machine, it is difficult to collect the target domain data covering all the working conditions in advance due to the complex and variable working conditions, and it is also difficult to provide a visible target domain to participate in training of the fault diagnosis model, which definitely causes a huge obstacle to cross-domain fault diagnosis. Therefore, how to train a global model with good generalization capability to effectively realize cross-domain fault diagnosis under invisible working conditions while guaranteeing data privacy is an important challenge to be solved in the face of a scene with modal heterogeneity and invisible target domain. Disclosure of Invention The invention aims to solve the main technical problem that the prior modal heterogeneous multi-modal federal learning cross-domain generalization performance is insufficient, and provides a multi-modal federal cross-domain fault diagnosis method based on prototype comparison learning. The aim of the invention is realized by the following technical scheme: A multi-mode federal cross-domain fault diagnosis method based on prototype comparison learning comprises the following steps: S1, collecting Multi-mode signal of machine under different working conditionsPersonal domain, data of different domains are stored to different domainsA client; S2, the central server prototype the global Global modality prototypeGlobal modelInitializing and distributing the data to all source clients; S3, the first The individual source clients receive and utilize their local multim