A Time Series Data Augmentation Method Based on Dynamic Time Warping

Abstract

Deep learning has become a hot research topic in the field of time series analysis and data mining. Training models often requires balanced and large data sets, but on the one hand, the number of different types of data in time series data sets is often extremely imbalanced, on the other hand, some time series data are often difficult to collect. Therefore, it is necessary to augment the training data before training the deep learning model. SMOTE is a data augmentation method widely used in preprocessing imbalanced data sets, but the classical SMOTE method does not satisfy the characteristics of time series when processing time series, so it is not effective when applied to time series data sets. To address this point, we propose an oversampling data augmentation method based on dynamic time warping—DTW-SMOTE. For the possible phase shifts of time series with the same characteristics, this method uses dynamic time warping to obtain more reasonable similarity between different time series, which improves the classical SMOTE method and achieves data augmentation for time series data. We conducted comparison experiments using two models with different architectures: ResNet and LSTM. The experimental results show that the performance of the above models is significantly improved after training with the DTW-SMOTE method pre-processed dataset.