Scholarlink

<p>Time series data are widely encountered across scientific and industrial domains, but multivariate time series classification (MTSC) remains challenging because of the need to jointly model multidimensional feature interactions and complex temporal dependencies. To address these difficulties, MSCFormer, a hybrid neural architecture that integrates multiscale convolutional feature extraction with dual multihead self-attention for performing MTSC tasks, is proposed in this paper. The model introduces a multiscale feature module (MSFM) to capture fine-grained and coarse-grained local patterns and to perform channel fusion, yielding richer and more discriminative feature representations. In addition, MSCFormer adopts a dual multihead attention mechanism, where F-MHA enhances the featurewise dependency modeling process, and T-MHA focuses on long-range temporal relationships. MSCFormer is evaluated on 27 UEA benchmark datasets, achieving an average accuracy of 80.0%, an average ranking of 2.06, and a +4.1% improvement over the best baseline. Comprehensive ablation experiments and visualization analyses further verify the individual and complementary contributions of the MSFM, F-MHA, and T-MHA to the performance of MSCFormer. These results demonstrate that MSCFormer provides an effective and robust framework for multivariate time series classification tasks.</p>

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

MSCFormer: a multiscale convolutional transformer for multivariate time series classification

  • 期刊论文

  • 发表日期：

    2026年3月19日

查看全文
  • Jingchao Xie,
  • Mingxin Yang,
  • Yang Wu,
  • Rui Hou,
  • Wei Li,
  • Mianxiong Dong,
  • Kaoru Ota

摘要

Time series data are widely encountered across scientific and industrial domains, but multivariate time series classification (MTSC) remains challenging because of the need to jointly model multidimensional feature interactions and complex temporal dependencies. To address these difficulties, MSCFormer, a hybrid neural architecture that integrates multiscale convolutional feature extraction with dual multihead self-attention for performing MTSC tasks, is proposed in this paper. The model introduces a multiscale feature module (MSFM) to capture fine-grained and coarse-grained local patterns and to perform channel fusion, yielding richer and more discriminative feature representations. In addition, MSCFormer adopts a dual multihead attention mechanism, where F-MHA enhances the featurewise dependency modeling process, and T-MHA focuses on long-range temporal relationships. MSCFormer is evaluated on 27 UEA benchmark datasets, achieving an average accuracy of 80.0%, an average ranking of 2.06, and a +4.1% improvement over the best baseline. Comprehensive ablation experiments and visualization analyses further verify the individual and complementary contributions of the MSFM, F-MHA, and T-MHA to the performance of MSCFormer. These results demonstrate that MSCFormer provides an effective and robust framework for multivariate time series classification tasks.