A collaborative adversarial framework: Distribution characteristics-guided alignment mechanism for fault diagnosis of machines considering domain shift

Abstract

Fault diagnosis of mechanical systems is essential to minimize damage and downtime in industrial fields. These systems frequently operate under varying and harsh conditions, leading to substantial changes in data distributions, commonly referred to as domain shift problem. This phenomenon presents a significant challenge for reliable fault diagnosis. Although many unsupervised domain adaptation methods effectively align data distributions, they often depend on target-domain pseudo labels. This dependency may lead to inaccurate diagnoses, particularly in the presence of abnormal samples. To address this limitation, a collaborative adversarial framework is proposed to exploit the intrinsic distribution characteristics of mechanical vibration signals to achieve distribution alignment. This framework introduces a two-level adversarial strategy to reduce distribution discrepancies. At the domain level, a novel Domain Alignment Loss (DAL) is designed to guide the adversarial game between the feature generator and the domain discriminator, thereby reducing marginal distribution discrepancies by considering both the amplitude and variability of vibration signals. At the class level, a new Class Alignment Loss (CAL) is proposed to steer the adversarial game between the feature generator and the two classifiers, using Gaussian Mixture Models (GMM) and Reproducing Kernel Hilbert Space (RKHS) to provide a more accurate measurement of conditional distribution discrepancies. Results on two datasets show that the proposed method achieves superior alignment capability and higher diagnostic accuracy compared to other state-of-the-art methods.