Domain-invariant feature fusion networks for semi-supervised generalization fault diagnosis

Abstract

Machinery fault diagnosis based on deep learning methods is cost-effective to guarantee safety and reliability of mechanical systems. Due to the variability of machinery working condition and difficulty of data obtaining under different health states, it is desirable to enhance the generalization capability to unseen working conditions for the fault diagnosis models trained by available data sets under limited number of working conditions. Considering that labeling industrial data is also a laborious work, this paper proposes a novel semi-supervised domain generalization model, termed domain-invariant feature fusion networks (DIFFN) for intelligent fault diagnosis under unseen target working conditions. The main contributions are that, intra-domain-invariant features are considered to capture the intrinsic semantic information within the domain and are fused with inter-domain-invariant features to enhance the discrimination and generalization abilities in fault diagnosis. First, a domain-invariant representation learning method is established to learn the inter- and intra-domain-invariant features using two network branches and fuse them via a fusion module. Second, a mutual learning strategy is designed to enable the network branches and the fusion module to learn from each other, thereby improving the discrimination of the extracted features for accurate fault diagnosis. Lastly, a feature divergence maximization strategy is embedded between the two network branches to improve the generalization ability of the fault diagnosis model. Experiments on two bearing data sets demonstrate that the proposed model has better diagnostic accuracy and stability over state-of-the-art semi-supervised domain generalization methods, indicating its great potential for application in generalization fault diagnosis of machinery under unseen target working conditions.