文献类型：会议论文

英文题名：Self-Supervised Time Series Classification Method Based on FRFT Cross-Channel Fusion

作者：Sun, ZhongHeng[1];Wang, Yiyang[1];Wang, Meihui[1];Jiang, Culling[1];Wan, Yougjing[1]

机构：[1]East China Univ Sci & Technol, Sch Informat Sci & Engn, Shanghai 200237, Peoples R China

会议论文集：2025 International Joint Conference on Neural Networks-IJCNN

会议日期：JUN 30-JUL 05, 2025

会议地点：Rome, ITALY

语种：英文

外文关键词：self-supervised learning; time series classification; FRFT; contrastive learning; cross-channel fusion

摘要：Recently, self-supervised representation learning has been widely applied to various time series tasks (e.g., electric device classification). However, building models for large-scale time series classification remains challenging due to the high cost of labeling time series data, which requires specialized expertise. Additionally, many datasets consist solely of unlabeled data or contain only a small number of labeled samples. Existing solutions for traditional time series tasks are not directly applicable to modern complex temporal problems due to two unique characteristics: (i) long-term temporal dependencies and (ii) complex cross-channel interactions. To address these challenges, we propose TC-FrC, a self-supervised time series classification model that leverages seasonal-trend decomposition for data augmentation and incorporates contrastive losses based on both temporal features and the Fractional Fourier Transform (FRFT). Specifically, we: (i) design a sparse attention mechanism within the temporal contrastive module to enhance feature extraction and improve robustness in long-term time series data, (ii) introduce a seasonal-trend decomposition approach to mitigate inter-class feature confusion, and (iii) develop a cross-channel FRFT-based feature fusion module, which transforms contextual features from the time domain to the fractional domain. We extensively evaluate TC-FrC on seven publicly available datasets, including HAR, Sleep-EDF, and Epilepsy, among others. Experimental results demonstrate that our method outperforms state-of-the-art baselines across various evaluation metrics.