ScholarWorks@숙명여자대학교

초록

Recently, as time series classification using deep learning has been actively studied, securing large amounts of data is becoming more important. However, securing a large amount of time series data in which labels exist is often difficult. Therefore, in this paper, we proposed an effective data augmentation and contrastive learning-based pre-training technique to improve the performance of time series classification models. The proposed time series data augmentation technique can create new data of different lengths while maintaining the measurement interval and characteristics of the original time series data. In addition, the proposed pre-training technique for time series classification models can improve the performance of the time series classification model. Using the time series data generated by the proposed augmentation technique, it can pre-train the time series classification model to distinguish similar and dissimilar time series data. As a result of applying the proposed time series data augmentation and the pre-training technique to user activity recognition model, the accuracy of the model was improved by up to 18%p.

키워드