EFIM: An Exploring Feature-Invariant Module for Domain Generalization

Deep learning suffers from performance degradation when evaluated with different distributions of training and testing data. Domain generalization (DG) enables the model to generalize to data distributions in unseen domains, i.e., to learn domain-invariant representations. In this paper, we learn domain invariant features in terms of internal and mutual invariance. Internal invariance means that we can capture the intrinsic semantics of the data, i.e., properties within a domain that are independent of other domains, and mutual invariance means that features can be learned in multiple domains. We design an Exploring Feature-Invariant Module (EFIM). EFIM employs a knowledge distillation framework to capture Fourier phase information as internally invariant features and a triplet loss metric for multi-source domain features as mutually invariant features. Extensive experiments conducted on the Case Western Reserve University rolling bearing (CWRU) dataset and SJTU dataset validate the effectiveness of the EFIM.