Wasserstein distance based deep adversarial transfer learning for intelligent fault diagnosis with unlabeled or insufficient labeled data

Abstract

Intelligent fault diagnosis is one critical topic of maintenance solution for mechanical systems. Deep learning models, such as convolutional neural networks (CNNs), have been successfully applied to fault diagnosis tasks and achieved promising results. However, one is that two datasets (in source and target domains) of similar tasks are with different feature distributions because of different operational conditions; another one is that insufficient or unlabeled data in real industry applications (target domains) limit the adaptability of the source domain well-defined models. To solve the above problems, the concept of transfer learning should be adopted for domain adaptation, in the meantime, a network performs both supervised and unsupervised learning is required. Inspired by Wasserstein distance of optimal transport, in this paper, we propose a novel Wasserstein Distance-based Deep Transfer Learning (WD-DTL) network for both supervised and unsupervised fault diagnosis tasks. WD-DTL learns domain feature representations (generated by a CNN based feature extractor) and minimizes distributions between the source and target domains through an adversarial training process. The effectiveness of the proposed WD-DTL is verified through 16 different transfer tasks. Results show that WD-DTL achieves the highest diagnostic accuracies when compared to the existing Maximum Mean Discrepancy and CNN networks in almost all transfer tasks.