SC Logo
IEEE Logo
Logo
IEEE Logo

Balanced Adaptation Regularization Based Transfer Learning for Unsupervised Cross-Domain Fault Diagnosis

Published in : IEEE Sensors Journal (Volume: 22, Issue: 12, June 2022)
Authors : Qin Hu, Xiaosheng Si, Aisong Qin, Yunrong Lv, Mei Liu
DOI : https://doi.org/10.1109/JSEN.2022.3174396
Summary Contributed by:  Qin Hu (Author)

The condition monitoring and fault diagnosis system typically uses vibration sensors to identify and classify fault patterns of mechanical equipment. Effective utilization of sensor signals for accurate and prompt diagnosis of potential faults is essential for ensuring safe, stable, long-time, full load and high-quality operation.

In practice, due to the changing working conditions of mechanical equipment, the data under different working conditions no longer satisfy the same distribution. Training samples and test samples usually possess different data distributions, which inevitably causes the performance degradation of most existing diagnostic models. Therefore, in the cross-domain fault diagnosis field, the main challenge is to accurately diagnose unknown fault samples under different working conditions based on labeled fault samples under known working conditions.

Transfer learning, as a machine learning method, provides a solution to overcome the above problem by improving the reusability of historical information and achieving information transfer between similar domains.

The paper presents a novel transfer learning algorithm, referred to as balanced adaptation regularization-based transfer learning (BARTL), to perform the machinery cross-domain fault diagnosis.

First, enhanced multi-scale sample entropy (EMSE) as hand-crafted features extract fault information from original vibration signals for each working condition. Experimental results demonstrated that the extracted features are less affected by the change in working conditions and can also discriminate the fault patterns under each working condition. Based on the extracted features, the BARTL algorithm is used for cross-domain fault diagnosis.

BARTL aims to construct an adaptive classifier in a reproducing kernel Hilbert space to achieve the cross-domain fault diagnosis. It explores balanced distribution adaptation and balanced label propagation in a unified framework to learn the adaptive classifier. Balanced distribution adaptation is used to dynamically minimize the marginal and conditional distribution discrepancy between domains, and balanced label propagation is used to adaptively minimize the structural risk losses of source and target domains. An iterative procedure is employed to refine the parameter matrix of the adaptive classifier, which is repeated until convergence. Once the adaptive classifier is obtained, a final diagnosis can be performed.

BARTL involves four parameters: two balance factors, the regularization parameter, and a number of iterations. Among them, one balance factor controls the structural risk losses of the source and target domains, while the other balance factor controls the complexity of the adaptive classifier. Parameter sensitivity analysis of the BARTL is conducted.

The effectiveness and superiority of BARTL are demonstrated using two public rolling bearing datasets. Multiple transfer tasks are implemented for each dataset, with diagnosis results presented through confusion matrices. The paper successfully showcases the advantages of balanced adaptation regularization-based transfer learning. The parameter sensitivity analysis of BARTL revealed significant influences of the balance factor and regularization parameter on diagnosis accuracy. The diagnosis performance of the proposed method is better than that of several contemporary state-of-the-art transfer learning methods. Further research is required to fuse other regularization terms or penalty terms in the BARTL that may produce more accurate results and generalization ability for machinery fault diagnosis under variable working conditions.

A non-profit organization, IEEE is the world's largest technical professional organization dedicated to advancing technology for the benefit of humanity.
Copyright 2023 IEEE – All rights reserved. Use of this website signifies your agreement to the IEEE Terms and Conditions
This site is also available on your smartphone.