Parallel multi-scale pooling dynamic domain adaptation discriminative network: a method for fault diagnosis of critical components in gearboxes

Abstract In the context of varying working conditions, there exist significant differences in feature distributions between the source domain and the target domain. Current feature extraction networks struggle to comprehensively capture both global and local features. Furthermore, most existing domain adaptation methods primarily focus on a simplistic combination of marginal distribution alignment (MDA) and conditional distribution alignment (CDA), which limits their effectiveness in enhancing domain confusion capabilities and lacks dynamic adaptive strategies. This poses challenges for cross-domain fault diagnosis of critical gearbox components. To address these challenges, we propose a novel parallel multi-scale pooling dynamic domain adaptation discriminative network (MSPD-NET). Specifically, a new parallel multi-scale pooling module is designed to use different pooling models and pooling scales to achieve multi-level feature fusion, thereby extracting richer global and local high-level features. In order to better measure the distribution distance and further improve the domain confusion ability, a new distribution difference metric (DDM) is proposed. In addition, to solve the problem of error accumulation caused by fixed-weight MDA and CDA training, DDM uses inter-domain and intra-class Kullback-Leibler (KL) divergence to re-evaluate the contribution of MDA and CDA to domain differences, dynamically adjust their contribution weights during training, reduce ratio errors in training iterations, and improve intra-class compactness and inter-class separability. In addition, in order to enhance the discrimination ability of the MSPD-NET network, reduce fuzzy decision boundaries, and improve classification performance, I-Softmax is used to classify the target domain. Cross-domain fault diagnosis of key gearbox components is performed on two public bearing datasets and a private gearbox dataset. The experimental results show that the proposed method is superior to the currently popular domain adaptation method.