Scholarlink

<p>To address the limitations of traditional multivariate time series classification (MTSC) methods in effectively integrating global context modeling with local feature extraction, this study proposes a dual-branch feature fusion GSTT model. The local feature extraction module (GAT-Shapelets) constructs a topological graph structure of temporal subsequences using the graph attention network (GAT) to derive Shapelets, focusing on capturing critical local features. The global feature extraction module (TSDRA-Transformer) employs a hierarchical convolutional architecture combined with a Time Series Dimensionality Reduction Attention (TSDRA) mechanism-based Transformer, which reduces the computational complexity of attention while capturing long-term dependencies across time steps. Through parallel computing optimization and distributed training frameworks, the GSTT model achieved efficient processing of large-scale data, significantly improving computational efficiency. Experimental evaluations on 30 datasets from the UEA multivariate time series benchmark demonstrate that the GSTT significantly outperforms baseline methods in classification accuracy. Ablation studies further validate the complementary synergy of local and global feature representations in the dual-branch architecture, underscoring the model’s robustness and efficacy. The experiments were conducted on a high-performance GPU cluster, utilizing parallel computing to accelerate model training and inference, demonstrating the superior performance of GSTT in a supercomputing environment.</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

A multivariate time series data classification method based on dual-branch structure GSTT

  • 期刊论文

  • 发表日期：

    2025年9月5日

查看全文
  • Qifan Zhang,
  • Houqun Yang,
  • Yizhen Wang,
  • Jianqiang Huang

摘要

To address the limitations of traditional multivariate time series classification (MTSC) methods in effectively integrating global context modeling with local feature extraction, this study proposes a dual-branch feature fusion GSTT model. The local feature extraction module (GAT-Shapelets) constructs a topological graph structure of temporal subsequences using the graph attention network (GAT) to derive Shapelets, focusing on capturing critical local features. The global feature extraction module (TSDRA-Transformer) employs a hierarchical convolutional architecture combined with a Time Series Dimensionality Reduction Attention (TSDRA) mechanism-based Transformer, which reduces the computational complexity of attention while capturing long-term dependencies across time steps. Through parallel computing optimization and distributed training frameworks, the GSTT model achieved efficient processing of large-scale data, significantly improving computational efficiency. Experimental evaluations on 30 datasets from the UEA multivariate time series benchmark demonstrate that the GSTT significantly outperforms baseline methods in classification accuracy. Ablation studies further validate the complementary synergy of local and global feature representations in the dual-branch architecture, underscoring the model’s robustness and efficacy. The experiments were conducted on a high-performance GPU cluster, utilizing parallel computing to accelerate model training and inference, demonstrating the superior performance of GSTT in a supercomputing environment.