Data Augmentation for Time Series Classification with Deep Learning Models

Gautier Pialla, Maxime Devanne, Jonathan Weber, Lhassane Idoumghar, Germain Forestier

Abstract: Deep Learning models for time series classification are benchmarked on the UCR Archive. This archive contains 128 datasets. Unfortunately only 5 datasets contain more than 1000 training samples. For most deep learning models, this lead to over-fitting. One way to address this issue and improve the generalization of the models is data augmentation. Although it has been extensively studied and is widely used for images, fewer works have been done on time series. InceptionTime is an ensemble of 5 Inception classifiers and is still regarded as the state-of-the-art deep learning model for time series classification. However, most of the work on data augmentation were not done on the Inception classifier. In this paper we solve this issue by studying 4 different data augmentation methods through 4 experiments on the Inception model. We studied trainings with one or several augmentations at the same time and with or without generating new samples at each epoch. We also conducted experiments with ensembling and benchmarked our results on the UCR Archive. We showed that using a combination of both the scaling and window warping data augmentation methods, we can significantly improve the accuracy of Inception and InceptionTime models.