TCDformer: A Temporal-Channel Decoupled Transformer Framework for Multivariate Time Series Classification

Multivariate time series classification (MTSC) is an important research direction in data mining, providing fundamental support for data analysis and intelligent decision-making in complex dynamic systems. However, current methods often follow the temporal-channel mixing approach, where temporal and channel information are jointly processed in the feature space. This inevitably leads to feature confusion and information loss. Furthermore, the existence of distribution shift degrades the generalization capability of the model. To this end, we propose a Transformer-based framework for MTSC, named TCDformer, which improves classification effectiveness by decoupling the modeling of temporal and channel dependencies. TCDformer consists of two core modules: the Channel Independent Transformer Module (CITM) and the Patch Independent Transformer Module (PITM). CITM first captures the pure temporal patterns within each channel, while PITM further explores the inter-channel interactions within fine-grained temporal segments. Additionally, reversible instance normalization (RevIN) is introduced to mitigate the distribution shift in the feature extraction layers, thereby achieving robust performance. Experimental results on 25 UEA datasets show that TCDformer outperforms current competitive MTSC methods.