Journal of Computer Applications ›› 2025, Vol. 45 ›› Issue (4): 1042-1052.DOI: 10.11772/j.issn.1001-9081.2024040448

• Artificial intelligence • Previous Articles     Next Articles

Multi-branch multi-view based contextual contrastive representation learning method for time series

Guangju YANG1,2, Tianjian LUO1,2(), Kaijun WANG1,2, Siqi YANG1,2   

  1. 1.College of Computer and Cyber Security,Fujian Normal University,Fuzhou Fujian 350117,China
    2.Digital Fujian Internet-of-Thing Laboratory of Environmental Monitoring (Fujian Normal University),Fuzhou Fujian 350117,China
  • Received:2024-04-15 Revised:2024-06-25 Accepted:2024-06-28 Online:2025-04-08 Published:2025-04-10
  • Contact: Tianjian LUO
  • About author:YANG Guangju, born in 1999, M. S. candidate. His research interests include time-series analysis, pattern recognition.
    LUO Tianjian, born in 1990, Ph. D., associate professor. His research interests include pattern recognition, brain-computer interface, electroencephalograph signal processing.
    WANG Kaijun, born in 1965, Ph. D., associate professor. His research interests include intelligent learning and reasoning.
    YANG Siqi, born in 2000, M. S. candidate. Her research interests include brain-computer interface, pattern recognition.
  • Supported by:
    National Natural Science Foundation of China(62106049);Natural Science Foundation of Fujian Province(2022J01655)

多分支多视图的时间序列上下文对比表征学习方法

杨光局1,2, 罗天健1,2(), 王开军1,2, 杨思琪1,2   

  1. 1.福建师范大学 计算机与网络空间安全学院,福州 350117
    2.数字福建环境监测物联网实验室(福建师范大学),福州 350117
  • 通讯作者: 罗天健
  • 作者简介:杨光局(1999—),男,福建三明人,硕士研究生,主要研究方向:时间序列分析、模式识别
    罗天健(1990—),男,湖北黄冈人,副教授,博士,主要研究方向:模式识别、脑机接口、脑电信号分析
    王开军(1965—),男,浙江温州人,副教授,博士,主要研究方向:智能学习与推理
    杨思琪(2000—),女,湖南衡阳人,硕士研究生,主要研究方向:脑机接口、模式识别。
  • 基金资助:
    国家自然科学基金资助项目(62106049);福建省自然科学基金资助项目(2022J01655)

Abstract:

Time series data are applied in various industries. However, the lack of label information and the complex temporal-spectral variations pose challenges for learning representations of time series. Therefore, a Multi-Branch Multi-View contextual Contrastive Representation Learning (MBMVCRL) method for time series was proposed. Firstly, time series samples were enhanced from both time and frequency perspectives and then input into a multi-branch multi-view model to extract multi-perspective feature representations of the time series. Secondly, for contrastive representation learning, the contextual contrastive loss and cross-prediction loss were calculated on the basis of the feature representations from the two perspectives, and joint training was conducted to obtain the optimal feature representation. Finally, to validate the representation capability of the proposed method for time series, an Affine Nonnegative Collaborative Representation (ANCR) classifier was used for the down-stream classification tasks. Experimental results show that the proposed method improves the recognition accuracy by 5.15, 0.90, and 1.89 percentage points, respectively, compared to the mainstream Time-Series Temporal and Contextual Contrasting (TS-TCC) method in human action, epilepsy, and sleep state recognition tasks. Ablation experimental results demonstrate the importance of the multi-branch multi-view model, highlighting the proposed model’s characteristics of low parameter sensitivity, fast convergence, and good generalization across different time series applications.

Key words: time series, self-supervised learning, contrastive representation learning, multi-branch, multi-view

摘要:

时间序列数据在众多行业中拥有广泛应用,然而受限于标注信息的缺失和复杂的时频域多变性,针对时间序列的表征学习成为一项挑战性任务。因此,提出一种用于时间序列的多分支多视图的上下文对比表征学习(MBMVCRL)方法。首先,从时频这2个视角增强时间序列样本,并把结果分别输入多分支多视图模型,从而提取时间序列的多视角特征表达;其次,为进行对比表征学习,分别根据2个视角的特征表达,计算上下文对比误差和交叉预测误差,并联合训练以获取最优的特征表达;最后,为验证所提方法对时间序列的表征能力,采用仿射非负协同表征(ANCR)分类器进行下游的分类任务。实验结果表明,相较于主流的时间序列时序上下文对比学习(TS-TCC)方法,所提方法在人体动作、癫痫和睡眠状态识别任务上的识别准确率分别提升了5.15、0.90和1.89百分点。消融实验结果则表明了多分支多视图模型的重要性,强调了所提模型拥有的参数敏感性不高和收敛快的特点,可见所提模型在不同时间序列应用上具有不错的泛化性。

关键词: 时间序列, 自监督学习, 对比表征学习, 多分支, 多视图

CLC Number: