Scholarlink

Multivariate Time Series Classification (MTSC) is crucial in fields such as Human Activity Recognition, Medical Diagnostics, and System Monitoring. Traditional classification methods often emphasize temporal dynamics or utilize similarity measures like dynamic time splitting (DTW) for data classification. However, the challenge lies in effectively capturing the spatio-temporal consistency of multivariate data, which is vital for accurate classification as system states evolve. This paper introduces CSTFormer, a new transformer-based model specifically designed for MTSC tasks. CSTFormer employs a dynamic adaptive time-frequency filter to address noise and capture global patterns, thereby enhancing the clarity and relevance of the extracted features. We also design the Dynamic Multiscale Temporal Attention Module and Dynamic Spatial Attention Module for CSTFormer that captures complex spatiotemporal dependencies by learning temporal and spatial contexts across channels and time dimensions. Furthermore, to integrate spatiotemporal contexts more coherently, this paper proposes a sophisticated fusion technique called Dynamic Cross-Distribution Fusion, which merges their distributions rather than simply concatenating them at the feature level. This method helps maintain the continuity of temporal and spatial features throughout the learning process, ensuring the model’s effective adaptation across various classification scenarios. Extensive evaluations of multiple datasets, including those of the UCR and UEA repositories, show that CSTFormer significantly outperforms existing models, with an average improvement in accuracy of 23.5% in challenging scenarios. Ablation studies confirm the effectiveness of key components of CSTFormer, highlighting its reliability and the critical role of its modules in achieving high classification accuracy.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

CSTFormer: Cross Spatial-Temporal Learning Transformer with Fourier Filtering Operators for Multivariate Time Series Classification

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Feiyue Qiu,
  • Wei Wu,
  • Liping Wang

摘要

Multivariate Time Series Classification (MTSC) is crucial in fields such as Human Activity Recognition, Medical Diagnostics, and System Monitoring. Traditional classification methods often emphasize temporal dynamics or utilize similarity measures like dynamic time splitting (DTW) for data classification. However, the challenge lies in effectively capturing the spatio-temporal consistency of multivariate data, which is vital for accurate classification as system states evolve. This paper introduces CSTFormer, a new transformer-based model specifically designed for MTSC tasks. CSTFormer employs a dynamic adaptive time-frequency filter to address noise and capture global patterns, thereby enhancing the clarity and relevance of the extracted features. We also design the Dynamic Multiscale Temporal Attention Module and Dynamic Spatial Attention Module for CSTFormer that captures complex spatiotemporal dependencies by learning temporal and spatial contexts across channels and time dimensions. Furthermore, to integrate spatiotemporal contexts more coherently, this paper proposes a sophisticated fusion technique called Dynamic Cross-Distribution Fusion, which merges their distributions rather than simply concatenating them at the feature level. This method helps maintain the continuity of temporal and spatial features throughout the learning process, ensuring the model’s effective adaptation across various classification scenarios. Extensive evaluations of multiple datasets, including those of the UCR and UEA repositories, show that CSTFormer significantly outperforms existing models, with an average improvement in accuracy of 23.5% in challenging scenarios. Ablation studies confirm the effectiveness of key components of CSTFormer, highlighting its reliability and the critical role of its modules in achieving high classification accuracy.