A generative adversarial networks based methodology for imbalanced multidimensional time-series augmentation of complex electromechanical systems

Abstract

Multidimensional monitoring time-series of complex electromechanical systems (CESs) plays a foundational role in data-based state management, maintenance, and performance adjustment. However, it is still a challenging work to extract valuable and complete information due to the imbalanced data. To address this issue, a methodology called GAN4MTS (Generative Adversarial Networks for Multidimensional Time-Series) that generates synthetic data closely mimicking the characteristics of real data was proposed, thus directly tackling the problem of data imbalance. First, the uniqueness of multidimensional time series of CESs was analyzed to identify the requirements for data augmentation and to define the problem formulations. Second, the architecture and loss functions of GAN4MTS model were designed based on generative adversarial networks and three specific constraints. Finally, the effectiveness of the proposed work was validated through comparative analysis. Furthermore, the intrinsic mechanisms of data augmentation in enhancing the model capabilities were discussed. The proposed methodology serves as a comprehensive technical solution for data augmentation, enabling the generation of high-quality synthetic data that adheres to the constraints of multidimensional time series in CESs. Additionally, as an open architecture model, this work provides novel methods for time-series data augmentation, addressing the issues of low accuracy and limited model generalization in state classification, identification, and prediction tasks for CESs caused by the presence of highly imbalanced data.