Oversampling for Imbalanced Time Series Classification Based on Generative Adversarial Networks

Sufficient training data for time series classification algorithms is critical. However, training data in many realworld applications exists a large imbalance between the majority class and the minority class inevitably to result in reducing the time series classification accuracy. Although a lot of existing solutions attempt to address the imbalanced learning problem, the synthetic samples based on the oversampling methods are short of diversity or variation to fail to accurately reveal the real minority class distribution. In order to create diverse artificial samples, a new oversampling method for imbalanced time series classification based on generative adversarial network (O-GAN) is proposed in this paper. An O-GAN application to imbalanced discharge voltage time series classification of lithium-ion cells is implemented. In addition, we selected 61177 cells with capacities between 2.06 Ah and 2.10 Ah as experimental samples. The samples are first preprocessed to obtain a labeled train set. For 1063 minority class samples, we use four oversampling methods to generate 1000 synthetic samples, respectively. Finally, the fully convolutional network (FCN) is considered as a discharge voltage time series classification algorithm to evaluate the oversampling methods. The results show that the proposed O-GAN method can effectively address the imbalanced time series classification problem.