Zhihui Men , Dao Gong , Kai Zhou , Yuejian Chen , Jinsong Zhou
{"title":"极端样本稀缺性下双数据辅助轴承故障诊断的无监督域自适应方法","authors":"Zhihui Men , Dao Gong , Kai Zhou , Yuejian Chen , Jinsong Zhou","doi":"10.1016/j.ymssp.2025.113359","DOIUrl":null,"url":null,"abstract":"<div><div>In small-sample bearing fault diagnosis, synthetic signals generated by simulation or generative models are commonly used to augment datasets. However, such signals often lack realistic noise and nonlinear characteristics, resulting in a domain gap between synthetic and real data. To address this, we propose an end-to-end method based on style transfer to generate twin data that better resembles real-world signals. First, a finite element model is built to derive the relationship between contact stiffness and radial force, and dynamic simulations are conducted using RecurDyn to obtain initial signals. Then, an Adaptive Style Transfer Network (AdasTNet) is employed to transfer the “style” of real signals to the simulated ones, enhancing their similarity in both time and frequency domains. The resulting twin data serves as the source domain, while the real data—without any labels—is treated as the target domain. We perform unsupervised domain adaptation using a CNN backbone combined with domain adversarial training and Maximum Mean Discrepancy (MMD) minimization. Experimental results show that the proposed method outperforms conventional GAN-based approaches in both accuracy and stability. Moreover, the model is lightweight and efficient, making it well-suited for real-world deployment in data-scarce scenarios.</div></div>","PeriodicalId":51124,"journal":{"name":"Mechanical Systems and Signal Processing","volume":"239 ","pages":"Article 113359"},"PeriodicalIF":8.9000,"publicationDate":"2025-09-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"Unsupervised domain adaptation method for bearing fault diagnosis assisted by twin data under extreme sample scarcity\",\"authors\":\"Zhihui Men , Dao Gong , Kai Zhou , Yuejian Chen , Jinsong Zhou\",\"doi\":\"10.1016/j.ymssp.2025.113359\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"<div><div>In small-sample bearing fault diagnosis, synthetic signals generated by simulation or generative models are commonly used to augment datasets. However, such signals often lack realistic noise and nonlinear characteristics, resulting in a domain gap between synthetic and real data. To address this, we propose an end-to-end method based on style transfer to generate twin data that better resembles real-world signals. First, a finite element model is built to derive the relationship between contact stiffness and radial force, and dynamic simulations are conducted using RecurDyn to obtain initial signals. Then, an Adaptive Style Transfer Network (AdasTNet) is employed to transfer the “style” of real signals to the simulated ones, enhancing their similarity in both time and frequency domains. The resulting twin data serves as the source domain, while the real data—without any labels—is treated as the target domain. We perform unsupervised domain adaptation using a CNN backbone combined with domain adversarial training and Maximum Mean Discrepancy (MMD) minimization. Experimental results show that the proposed method outperforms conventional GAN-based approaches in both accuracy and stability. Moreover, the model is lightweight and efficient, making it well-suited for real-world deployment in data-scarce scenarios.</div></div>\",\"PeriodicalId\":51124,\"journal\":{\"name\":\"Mechanical Systems and Signal Processing\",\"volume\":\"239 \",\"pages\":\"Article 113359\"},\"PeriodicalIF\":8.9000,\"publicationDate\":\"2025-09-16\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"Mechanical Systems and Signal Processing\",\"FirstCategoryId\":\"5\",\"ListUrlMain\":\"https://www.sciencedirect.com/science/article/pii/S088832702501060X\",\"RegionNum\":1,\"RegionCategory\":\"工程技术\",\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q1\",\"JCRName\":\"ENGINEERING, MECHANICAL\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"Mechanical Systems and Signal Processing","FirstCategoryId":"5","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S088832702501060X","RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"ENGINEERING, MECHANICAL","Score":null,"Total":0}
Unsupervised domain adaptation method for bearing fault diagnosis assisted by twin data under extreme sample scarcity
In small-sample bearing fault diagnosis, synthetic signals generated by simulation or generative models are commonly used to augment datasets. However, such signals often lack realistic noise and nonlinear characteristics, resulting in a domain gap between synthetic and real data. To address this, we propose an end-to-end method based on style transfer to generate twin data that better resembles real-world signals. First, a finite element model is built to derive the relationship between contact stiffness and radial force, and dynamic simulations are conducted using RecurDyn to obtain initial signals. Then, an Adaptive Style Transfer Network (AdasTNet) is employed to transfer the “style” of real signals to the simulated ones, enhancing their similarity in both time and frequency domains. The resulting twin data serves as the source domain, while the real data—without any labels—is treated as the target domain. We perform unsupervised domain adaptation using a CNN backbone combined with domain adversarial training and Maximum Mean Discrepancy (MMD) minimization. Experimental results show that the proposed method outperforms conventional GAN-based approaches in both accuracy and stability. Moreover, the model is lightweight and efficient, making it well-suited for real-world deployment in data-scarce scenarios.
期刊介绍:
Journal Name: Mechanical Systems and Signal Processing (MSSP)
Interdisciplinary Focus:
Mechanical, Aerospace, and Civil Engineering
Purpose:Reporting scientific advancements of the highest quality
Arising from new techniques in sensing, instrumentation, signal processing, modelling, and control of dynamic systems