关键词: 多维时间序列表示学习, 自监督学习, 重构, 对比学习

Abstract: The spatiotemporal coupling features of a single multivariate time series(MTS) sample and the invariant discriminative features between the MTS samples can not be simultaneously captured by most of the existing self-supervised representation learning models for MTS. To this end, the MTS representation learning model combining reconstruction with contrasting ( MTS-RC ) is proposed. First, the MTS samples are augmented through binary noise masks, and then the local context features of the MTS and its augmented samples and long-term time-dependent relationship of local context features are respectively extracted by dilated convolution and transformer encoder. Finally, the spatiotemporal coupling characteristics of a single MTS sample are captured by reconstructing the mask value of augmented samples, and the MTS invariant discriminative characteristics are extracted between samples by contrastive learning. Experiments on classification on multiple MTS public datasets show that the classification accuracy of the classifier trained by the representation vector obtained from the MTS-RC outperforms typical baseline models. The results of ablation experiments show that the quality of MTS representation vector can be improved by the combined design of reconstruction and contrasting.

Key words: multivariate time series representation learning, self-supervised learning, reconstruction, contrastive learning