AGDM: Adaptive Granularity and Dimension Decoupling for Multidimensional Time Series Classification

Multidimensional time series classification has been extensively studied and applied in recent years; however, it still encounters several challenges. Firstly, due to the inability to determine the appropriate granularity for feature ex-traction, existing research fails to adapt to the fluctuation differences of different datasets at varying scales. Additionally, existing methods excessively prioritize inter-dimensional correlations when extracting features from multi-dimensional time series, resulting in the inability to accurately extract feature representations centered around a single dimension. Simultaneously, uniformly encoding multi-dimensional data with the same timestamp may lead to too small receptive fields and information distortion. Finally, existing research neglects the multimodal characteristics of multidimensional time series data under different perspectives. In response to the above challenges, this article proposes the AGDM network. Specifically, we have designed an adaptive granularity method to ex-tract features of different granularities in time series, so that the global context and local information of different time series data can be captured in a generalized manner. Furthermore, the AGDM network decouples multi-dimensional time series, thereby focusing more on extracting feature representations with a single dimension as the core and considering the correlation between dimensions at the feature level. Finally, we enrich the representation of each dimension by fusing multi-modal information. And encode the multimodal features of each dimension as a whole. Experimental results on multiple real datasets demonstrate that the AGDM network exhibits superior accuracy and performance compared to existing research.