Multi-source manifold space domain adaptation with a full-thresholding residual network for machinery fault diagnosis

Abstract

Unsupervised multi-source domain adaptation, which has been intensively investigated in recent years, is promising in handling fault diagnosis tasks when no labeling is available on the target datasets. Most approaches aim to learn the domain-invariant features of all domains in common feature spaces. However, addressing practical scenarios in which data comes from multiple domains with large shifts remains challenging. Hence, a new multi-source manifold space domain adaptation method (MMSDA) with a full-thresholding residual network is proposed for machinery fault diagnosis, in which specific domain-invariant features of the source and target domains are learned. First, a full-thresholding residual convolutional neural network (FTRCNN) is designed to extract useful features from both source and target domains, which are then projected into a specific domain feature space. Then, the proposed manifold neighbor consistency (MNC) domain alignment algorithm maps the feature space to a manifold space, ensuring that the samples maintain local neighbor geometric relations. Additionally, multi-kernel maximum mean discrepancy is used to reduce the inter-domain differences. Thus, the specific domain-invariant features of each source and target domain pair in the manifold feature space are extracted. Finally, the domain-specific classifier consistency (DSCC) loss is designed to minimize the shifts in all classifiers. Through experiments on three benchmarks, the proposed method demonstrates promising results on popular rotating machinery datasets for fault diagnosis.