A novel contrastive adversarial network for minor-class data augmentation: Applications to pipeline fault diagnosis

Abstract

In recent years, deep learning techniques have achieved significant progress in a variety of tasks. However, existing techniques might not perform well in the presence of data scarcity and imbalance, which is a problem commonly encountered in practice. In this paper, we propose a novel contrastive adversarial network (CAN) that aims to augment imperfect data with satisfactory quality and desired diversity. Specifically, we first propose a new distance metric, called class-aware mean discrepancy, to excavate features associated with operating conditions and generate data with improved compactness (within the same class) as well as enhanced discrimination (for different classes). Furthermore, a dynamic fault-semantic embedding scheme is developed to capture structural priors from real time-series data, which contributes to the comprehensive characterization of the context information of the generated data. Experimental results indicate that the proposed CAN outperforms some state-of-the-art augmentation approaches in terms of quality and diversity. Moreover, the proposed CAN is applied to the pipeline fault diagnosis problem with better diagnostic accuracy than that from the existing algorithms, which demonstrates the applicability of our research results in real-world scenarios.