Boosting Time-Series Domain Adaptation via A Time-Frequency Consensus Framework

Unsupervised Domain Adaptation (UDA) has proven to be effective in addressing the domain shift problem in computer vision. However, compared with visual applications, UDA for time series brings forth additional challenges. Potential domain shifts may have varying impacts on both time and frequency features, rendering conventional UDA methods less effective in this context. To address these challenges, we propose a Time-Frequency Consensus Domain Adaption (TFCDA) framework to enhance UDA methods for time-series data. TFCDA designs a frequency encoder, a trainable Time-Frequency Mapping (TFM), and a consensus loss, building upon conventional UDA methods to boost their performance. The TFM is trained on source domain data to learn the inherent time-frequency feature mapping, while the novel consensus loss ensures consistent feature transfer during UDA in the target domain, effectively reducing domain shifts in both time and frequency, and thus boosting overall performance. Experimental evaluations on four publicly available time-series datasets demonstrate TFCDA’s effectiveness in enhancing existing UDA methods for time-series data, highlighting its potential for real-world applications.