A Stable Adaptive Adversarial Network With Exponential Adversarial Strategy for Bearing Fault Diagnosis

Abstract

Given an industry’s development, fault diagnosis has received significant attention. Owing to complex and changeable working conditions and lacking marked fault data for fault diagnosis, domain adaptation has become a new solution. However, a variable operating environment will cause changes in data distributions, which complicates domain adaptation. Resultantly, a challenge exists in measuring a data distribution and combining it with domain adaptation. Thus, this article proposes a method of dynamically adapting a marginal distribution and a conditional distribution, including a new adaptive factor, which can use distance metrics and exponential functions to stably adapt to different data distributions in source and target domains. By dynamically adjusting the importance of the marginal and conditional distributions in a model, the proposed model can achieve excellent diagnostic results. Compared with a fixed-scale model without an adaptive factor that only has high diagnostic results for some working conditions, the proposed model has stable and accurate diagnosis results, whether it is facing different speeds and different loads. In addition, experiments are conducted to verify the effectiveness and usability of the proposed model.