基于原型动态对齐的时间序列分类方法

Time Series Classification Based on Prototype Dynamic Alignment

摘要: 随着传感器技术、智能医疗和工业监测的发展，时间序列数据在脑电信号分析、设备状态识别和行为检测等领域中的应用日益广泛，其动态变化和时间错位特征给分类任务带来了较大挑战。传统深度学习方法通常依赖大量标注样本，在低资源场景下容易出现过拟合；而动态时间规整方法虽然能够处理时间错位问题，但难以通过训练过程持续优化模型表示。因此，本文提出一种基于原型动态对齐网络的时间序列分类方法。该方法首先利用特征编码模块提取时间序列的潜在表示，然后构建可学习类别原型以刻画不同类别的典型时序结构，并引入软动态时间对齐机制计算样本与类别原型之间的匹配关系。实验评估在多个公开UCR时间序列数据集上进行，结果表明，所提方法在低资源训练条件下具有较好的分类稳定性，并在训练样本较充分时仍保持一定竞争力，验证了模型的有效性。

Abstract: With the development of sensor technology, intelligent healthcare, and industrial monitoring, time series data have been widely used in electroencephalogram analysis, equipment state recognition, behavior detection, and other fields. Their dynamic variations and temporal misalignment bring considerable challenges to classification tasks. Traditional deep learning methods usually rely on a large number of labeled samples and are prone to overfitting in low-resource scenarios. Although dynamic time warping can handle temporal misalignment, it is difficult to continuously optimize model representations through training. Therefore, this paper proposes a time series classification method based on a prototype dynamic alignment network. The proposed method first uses a feature encoding module to extract latent representations of time series, then constructs learnable class prototypes to describe typical temporal structures of different categories, and introduces a soft dynamic time alignment mechanism to measure the matching relationship between samples and class prototypes. Experiments are conducted on multiple public UCR time series datasets. The results show that the proposed method achieves stable classification performance under low-resource training conditions and remains competitive when sufficient training samples are available, which verifies the effectiveness of the model.