A mixed adversarial adaptation network for intelligent fault diagnosis

Behind the brilliance of the deep diagnosis models, the issue of distribution discrepancy between source training data and target test data is being gradually concerned for catering to more practical and urgent diagnostic requirements. Consequently, advanced domain adaptation algorithms have been introduced to the field of fault diagnosis to address this problem. However, in performing domain adaptation, most existing diagnosis methods only focus on the minimization of marginal distribution divergences and do not consider conditional distribution differences at the same time. In this paper, we present a mixed adversarial adaptation network (MAAN) based intelligent framework for cross-domain fault diagnosis of machinery. In this approach, differences in marginal distribution and conditional distribution are reduced together by the adversarial learning strategy, moreover, a simple adaptive factor is also endowed to dynamically weigh the relative importance of two distributions. Extensive experiments on two kinds of mechanical equipment, i.e. planetary gearbox and rolling bearing, are built to validate the proposed method. Empirical evidence demonstrates that the proposed model outperforms popular deep learning and deep domain adaptation diagnosis methods.

Automated summary

In this paper, a mixed adversarial adaptation network (MAAN) based intelligent framework for cross-domain fault diagnosis of machinery is presented. Differences in marginal distribution and conditional distribution are reduced together through an adversarial learning strategy with a simple adaptive factor to dynamically weigh the relative importance of two distributions. Experimental results show that the proposed model outperforms popular deep learning and deep domain adaptation diagnosis methods.