Self-supervised deep contrastive and auto-regressive domain adaptation for time-series based on channel recalibration

Time-series based unsupervised domain adaptation (UDA) techniques have been widely adopted to the applications of intelligent systems, such as sleep staging, fault diagnosis, and human activity recognition. However, recently methods have overlooked the importance of temporal feature representations and the distribution discrepancies across domains, which deteriorated UDA performance. To address these challenges, we proposed a novel Self-supervised Deep Contrastive and Auto-regressive Domain Adaptation (SDCADA) model for crossdomain time-series classification. Specifically, the cross-domain mixup preprocessing strategy is applied to reduce sample-level distribution discrepancy, then we proposed to introduce the channel recalibration module for adaptively selecting discriminative representations. Afterwards, the auto-regressive discriminator and teacher model are proposed to reduce the distribution discrepancies of feature representations. Finally, a total of six losses, including contrastive and adversarial learning, are weighted and jointly optimized to train the SDCADA model. The proposed SDCADA model has been systematically experimented on four cross-domain time-series benchmarked datasets, and its classification performance surpasses several recently proposed state-of-the-art models. Moreover, it effectively captures discriminative and comprehensive cross-domain time-series feature representations with parameter insensitivity.