Double-level adversarial domain adaptation network for intelligent fault diagnosis

Abstract

Deep neural networks have been widely studied in the field of mechanical fault diagnosis with the rapidity of intelligent manufacturing and industrial big data, however, attractive performance gains usually come from a premise that source training data and target test data have the same distribution. Unfortunately, this assumption is generally untenable in practice due to changeable working conditions and complex industrial environment. To address this issue, a double-level adversarial domain adaptation network (DL-ADAN) is presented for cross-domain fault diagnosis, which is able to bridge the divergences between the source and target domains. Specifically, the proposed diagnostic framework is composed of a feature extractor based on deep convolutional network, a domain discriminator and two label classifiers, which conducts two minimax adversarial games. In the first adversarial stream, the feature extractor and domain discriminator game with each other to achieve domain-level alignment from a global perspective. On the other line, the extractor and two classifiers are against each other to conduct class-level alignment, in which Wasserstein discrepancy is used to detect outlier target samples. As a result, the extractor can learn transferable discriminative features for accurate fault diagnosis. Extensive diagnostic experiments are constructed for performance analysis and several state of the art diagnostic methods are selected for comparative study. The comprehensive results demonstrate the effectiveness and superiority of the proposed method.