Interpretable temporal GANs for industrial imbalanced multivariate time series simulation and classification

Abstract

In industrial application, minority classes of multivariate time series often represent the classes of interest, such as event or fault identification. However, machine learning research is lacking on industrial multivariate time series with limited amounts of labeled data on minority classes. In this work, we propose Interpretable Temporal Generative Adversarial Networks for industrial imbalanced multivariate time series simulation and classification, or IT-GANs. Within IT-GANs, the generator approaches the distribution of each class and generates simulated time series for the minority classes, by leveraging information learned from the majority class; while the classifier benefits from such data augmentation, resulting in more accurate prediction. Furthermore, IT-GANs provides result interpretability with variable association learned from each class, with which we can pinpoint how time series classes differ from each other. Experiments show that the proposed IT-GANs achieves high quality of simulated data, good accuracy on imbalanced classification with learned variable association as result interpretation.