Fault Feature Recovery With Wasserstein Generative Adversarial Imputation Network With Gradient Penalty for Rotating Machine Health Monitoring Under Signal Loss Condition
Wenyang Hu, Tianyang Wang, Fulei Chu
Abstract
Rotating machine health monitoring systems sometimes suffer from large segments of continuous missing data in practical applications, which may lead to incorrect health assessments and maintenance decisions. However, current signal loss recovery methods are sometimes unable to impute loss monitoring signals and directly recover fault features without an intact training dataset. Moreover, these methods often fail to accurately capture the data distribution and recover the fault features when processing monitoring signals with large segments of continuous missing data and relatively high noise pollution from the rotating machinery. In this study, a fault feature recovery strategy called the Wasserstein generative adversarial imputation network with gradient penalty (WGAIN-GP) is proposed. The generative adversarial training strategy can help capture the distribution of the original monitoring signal. The mask mechanism ensures that the fault feature recovery method can be performed in an unsupervised manner and directly imputes the original loss signal and recovers the fault features. Finally, the introduction of the Wasserstein distance loss function and the gradient penalty further ensures that the fault features can be restored with relatively high accuracy and a lower probability of mode collapse. Case studies were conducted using both simulations and rotating machine experiments to demonstrate the effectiveness of the proposed method.