SFCC: Data Augmentation with Stratified Fourier Coefficients Combination for Time Series Classification

Deep neural networks (DNNs) have shown remarkable performance in time series classification tasks. However, the DNNs rely on a mass of data, which may not accumulate in real scenarios. As such, researchers investigate data augmentation methods to solve the scarcity of labeled data. Some of them, such as the rotation, that borrowed from the computer vision are not applicable due to the unique property of time series data. Besides, existing frequency-based methods applied for audio and speech recognition generate new samples by changing original frequency information, which may introduce unreasonable variation. In this paper, we propose a novel time series data augmentation method called Stratified Fourier Coefficients Combination (SFCC). SFCC retains and combines the original Fourier coefficients to augment the time series datasets. First, we transform data into the frequency domain using the discrete Fourier transform (DFT). To maintain the initial data distribution, we stratify the coefficients into several groups and then randomly select the groups to concatenate the coefficients. Finally, a new sample is generated through inverse DFT. The experiments demonstrate that the augmentation by SFCC can effectively improve the performance of the DNNs and achieve state-of-the-art results compared with the other 12 benchmarking methods.