Data augmentation for forecasting industrial aging processes via conditional multimodal generative time-series models

Data augmentation has shown to be effective for improving generalization performance of deep neural networks, especially in the regime of high noise and scarce data. However, this approach has not been applied to industrial aging processes (IAP) forecasting, where observed data are multimodal time-series, and therefore existing augmentation methods are not suitable for data generation. In this paper, we propose Seq-MVAE, a generative architecture that can generate complex time-series data consisting of multiple heterogeneous modalities. Seq-MVAE is capable of conditional generation, i.e., Seq-MVAE learns the joint distribution across the modalities, and allows users to generate a part of the modalities that are coherent with the other (given) modalities. This enables not only missing value imputation but also conditional generation, which is known to be crucial for data augmentation. We evaluate the generative performance and other aspects of Seq-MVAE on an artificial dataset generated based designed to simulate an industrial aging process, and show the effectiveness of data augmentation by Seq-MVAE on a real-world dataset acquired from an industrial plant.