首页|Contrastive learning based self-supervised time-series analysis
Contrastive learning based self-supervised time-series analysis
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
NSTL
Elsevier
Deep learning architectures usually require large scale labeled datasets for achieving good performance on general classification tasks including computer vision and natural language processing. Recent techniques of self-supervised learning have opened up new a research frontier where deep learning architectures can learn general features from unlabeled data. The task of self-supervised learning is usually accomplished with some sort of data augmentation through which the deep neural networks can extract relevant information. This paper presents a novel approach for self-supervised learning based time-series analysis based on the SimCLR contrastive learning. We present novel data augmentation techniques, focusing especially on time-series data, and study their effect on the prediction task. We provide comparison with several fault classification approaches on benchmark Tennessee Eastman dataset and report an improvement to 81.43% in the final accuracy using our contrastive learning approach. Furthermore we report a new benchmark of 80.80%, 81.05% and 81.43% for self-supervised training on Tennesee Eastman where a classifier is only trained with 5%, 10% or 50% percent of the available training data. Hence, we can conclude that the contrastive approach is very successful in time-series problems also, and further suitable for usage with partially labeled time-series datasets.(c) 2022 Elsevier B.V. All rights reserved.
NSTL
Elsevier
