A priori information-guided generative adversarial network for data augmentation: application to pipeline fault diagnosis

Abstract

ABSTRACT With the rapid development of artificial intelligence, deep learning is considered a promising technique for intelligent fault diagnosis using large amounts of data in various industrial fields. Under such circumstances, imbalanced datasets in the real world will not only hinder the further development of classification models, but also degrade the performance of existing models. To overcome this limitation, this paper proposes a novel Mel-Frequency Cepstral Coefficent-based Generative Adversarial Network (MFCC-GAN) to augment the high-quality small class data. Specifically, the MFCC is first used to capture the time- and frequency-domain features of the real signals as a priori information, which is then fed into the generative model. The temporal structural features and energy features contained in a prior information can provide effective guidance for the process of the generative model to map the Gaussian distribution to the real-world distribution. Moreover, a contrastive loss is introduced to refine the discriminative features of the generated signals, aiming to improve the distinguishability among different health states. Experimental results show that the MFCC-GAN algorithm improves the quality and fidelity of the generated data compared to other state-of-the-art algorithms.