Multi-domain Consistent Augmentation via Diffusion Models for Imbalanced Time Series Classification

Class imbalance and the scarcity of training samples remain critical challenges in time series classification, severely hindering performance. To mitigate these issues, data augmentation techniques are utilized, with certain deep generative models demonstrating notable potential. Nonetheless, their efficacy is undermined by the limited diversity inherent in single-domain generation and the distributional shift of the synthesized data. Consequently, this paper proposes a Multi-Domain Consistent Augmentation (MDCA) framework, which leverages multi-domain diffusion models to tackle these challenges. MDCA augments trainsets in both time and time-frequency domains while also extracting multi-domain features to enhance discriminability. Furthermore, a Dynamic Discriminator Modulator (DDM) is proposed to adjust the significance of generated samples adaptively during the training of classification models, effectively addressing the issue of distribution shift. Comprehensive experiments conducted on UCR archive and industrial motor datasets demonstrate the effectiveness and generalization capabilities of MDCA.