Improved Generative Adversarial Networks With Filtering Mechanism for Fault Data Augmentation

Few-shot fault diagnosis (i.e., fault diagnosis with few samples) is a challenging issue in practice because fault samples are scarce and difficult to obtain. Data augmentation based on generative adversarial networks (GANs) has proven to be an effective solution. However, it often encounters problems such as difficult model training and low quality of generated samples. In this article, an improved GAN with filtering mechanism is developed for fault data augmentation, which introduces the self-attention mechanism and instance normalization (IN) into the GAN structure and utilizes a filtering mechanism. The implementation process comprises two parts, sample generation and abnormal sample filtering. The self-attention mechanism and IN adopted in sample generation can make the generative model easy to train and have better quality of the generated samples. For abnormal sample filtering, the isolated forest (IF) algorithm is used for detecting low-quality generated samples. The effectiveness of the proposed fault data augmentation method is verified using two public datasets for fault diagnosis purposes, and the results show that the proposed method can have better performance over the state-of-art GAN-based data augmentation methods.

I. Introduction

Fault diagnosis is critically important to improving the safety of modern engineering systems by promptly detecting and isolating abnormal parts, preventing fault propagation and evolution, hence reducing the loss after unexpected failures occur [1]. Benefiting from advanced sensing and information technology, data-driven fault diagnosis techniques have been deeply studied and widely adopted in many areas [2], [3], [4], [5]. However, most existing data-driven fault diagnosis methods require enough fault samples to develop fault diagnosis models, while fault samples are indeed scarce and difficult to obtain in real-world applications. It poses great challenges to implement the state-of-the-art data-driven fault diagnosis techniques. In recent years, fault diagnosis with limited fault samples has been attracting increasing attention in both academia and industries [6], [7], [8].