BC-GAN: A threshold-free framework for unsupervised early fault detection in rotating machinery

Abstract

Early fault detection (EFD) is a crucial component of proactive maintenance that can prevent expensive downtime, enhance safety, and optimize equipment performance and longevity. Existing limitations of the contemporary EFD approaches frequently include: (1) manually designed features relying on expert skills and knowledge, (2) the vaguely determined pre-set thresholds to distinguish between the faulty and healthy state of the testing object (however, the threshold value can vary notably from task to task), and (3) most methods, which have been originally designed for vibration data, cannot be generalised to other techniques, for example acoustic emission (AE) signals. To address these issues, a novel binary-classification generative adversarial network (BC-GAN) is designed for general EFD problems and applied to several specific datasets acquired from different run-to-failure tests of rotating components. Compared with conventional methods, unsupervised BC-GAN directly outputs the probability that an input belongs to either the “health” or “fault” state of the rotating machine without a priory threshold setting. Experimental results demonstrate the high versatility of the proposed network, which can be applied in various laboratory and industrial settings to both vibration and AE signals.