Health Indicator Construction Method of Bearings Based on Wasserstein Dual-Domain Adversarial Networks Under Normal Data Only

Abstract

Rolling bearings are the most critical parts of rotating machinery and their damage is the leading cause of system failures. To ensure the reliability of the system, it demands to construct a health indicator (HI) to assess the state of degradation. However, existing HI construction methods (HICMs) have two limitations. First, the integration of well-designed features relies heavily on the experience of domain expert knowledge. Second, the construction of intelligent HI relies too much on life-cycle data. To cope with these limitations, this article proposed an HICM–Wasserstein dual-domain adversarial networks (WD-DAN), namely HICM-WD-DAN, which can extract generalized features with only normal data during the training. The dual-domain restriction of regularization promotes the generated signals approach to normal samples, making the constructed HI more robust and accurate. Moreover, to balance the weights of dual-domain parts automatically, an independent weighting structure is introduced. Finally, considering the actual degradation state of the system, the modified monotonicity and trendability indexes are proposed to evaluate the performance of HI. The effectiveness of HICM-WD-DAN is verified by bearings’ life-cycle data, and the results show that the constructed HI can represent the irreversible degradation process of bearings accurately and monotonously.