Improved Transfer Component Analysis and It Application for Bearing Fault Diagnosis Across Diverse Domains

Abstract

In recent years, intelligent fault diagnosis models based on machine learning used for intelligent condition monitoring and diagnosis have achieved considerable success. However, in the current research, the diagnosis process is based on an assumption that the same feature distribution exists between training data and testing data. Regrettably, in real application, training data and testing data are often from diverse domains, the difference in feature distributions is often prevalent; in this case, the traditional diagnostic models lack adaptability. To address this issue, this work proposed a diagnosis framework based on domain adaptation. This framework is inspired by the domain adaptation ability of transfer learning, in that the model trained by the labeled data in source domain can be transferred to diagnose a new but similar target data. The domain adaptation algorithm transfer component analysis (TCA) and its improved algorithm- improved transfer component analysis (ITCA) are embedded into this framework, respectively, to verify its applicability. An experiment was conducted on the datasets of bearing to demonstrate the applicability and practicability of the proposed transfer framework. The results show that the proposed method presents high accuracy in the transfer task of bearing fault diagnosis under different conditions across diverse experimental positions and fault types.