Weighted quantile discrepancy-based deep domain adaptation network for intelligent fault diagnosis

Abstract

Numerous intelligent fault diagnosis models do not perform well owing to the lack of labeled data with fault information and domain shifts under different working conditions. In this study, a weighted quantile discrepancy (WQD) metric is proposed and integrated into a deep adversarial learning framework to develop a deep weighted quantile domain adaptation network (DWQDAN) to tackle the aforementioned fault diagnosis challenges. The WQD metric first introduces quantile theory to consider the influence of different quantiles on domain adaptation concerning fault diagnosis. It can integrate the complete distribution information to obtain more accurate distribution matching characteristics with less computational complexity compared with classical metrics. By incorporating the dynamic WQD and adversarial learning, the DWQDAN can effectively learn domain-invariant features to perform different domain adaptation tasks. Its utility is illustrated using empirical applications with the public CWRU and Paderborn bearing datasets. The results show that DWQDAN performs better than some state-of-the-art methods. Adjustable parameters, including the number of quantiles and corresponding weights, provide an effective method for practical fault diagnosis.