Unsupervised Domain Adaptation for Machine Fault Diagnosis via Subspace Interpolation Using Deep Transforms

• 全文链接
• 请求原文

作   者：

Kriti Kumar;  Angshul Majumdar;  A. Anil Kumar;  M. Girish Chandra; 

作者机构：

Indraprastha Institute of Information Technology;  TCS Research; 

关键词：

Adaptation models;  transform learning (TL);  Transforms;  machine fault diagnosis;  Inspection;  Task analysis;  unsupervised learning;  Fault diagnosis;  Training;  Deep models;  Interpolation;  subspace modeling;  domain adaptation (DA); 

期刊名称：

IEEE sensors journal

i s s n：

1530-437X

年卷期：

2024 年 24 卷 13 期

页   码：

20959-20969

页   码：

摘   要：

Machine condition monitoring plays a vital role in the realm of Industry 4.0 manufacturing. Data-driven techniques have become increasingly popular for machine inspection, ensuring reliable operation and minimal downtime by learning complex relationships from data. However, these techniques assume that the train (source) and test (target) data follow similar distribution, which is often not true in real-life scenarios. Domain discrepancies may arise between the source and target data due to various factors, such as data collected from the same machine but at different working conditions, sensor placements, and so on, or data collected from similar (related) but different machines. Moreover, the availability of limited data poses an additional challenge. To address these issues, this article introduces a novel feature augmentation approach employing deep transform learning (DTL)-based subspace interpolation for unsupervised domain adaptation (UDA). By learning deep transforms to model the source and target domains and interpolating intermediate domains, domain-invariant features are generated for cross-domain classification. Unlike dictionary-based subspace interpolation methods, the proposed DTL method provides improved performance with reduced complexity due to the inherent benefits of transform learning (TL). Experimental results on different bearing fault datasets demonstrate the superior performance of this method, even with limited data. An accuracy improvement of $approx 5$ % (or more in some cases) is reported over the best-performing competing techniques for most adaptation cases, considering the challenging adaptation between different but related machines, which is of vital importance in real industrial applications.