LAUSR

Abstract Data augmentation is one of the necessary steps in the process of automated data-driven fault detection and diagnosis (FDD) for chillers, while real-world operational training samples are usually imbalanced. Faulty data samples are usually more difficult for collection than normal operation data. Existing works show that the generative adversarial networks (GAN) are useful generating synthetic faulty data samples to enrich the training dataset. However, it remains a problem for the automated FDD applications to select high-quality synthetic faulty samples generated by GAN. The FDD accuracy becomes unstable when the quality of synthetic fault data samples cannot be controlled entirely. In this study, we proposed to use the classic definition of anomaly detection to select high-quality synthetic fault data samples with the generative adversarial networks. Two anomaly detection methods were investigated, including the traditional variational auto-encoder (VAE) and the GANomaly. Through a series of experiments, it is justified that, with a small amount of real fault data, the proposed GAN-based chiller FDD framework with GANomaly achieves the highest FDD accuracy than all compared methods.