High-Efficiency Generative Adversarial Network Model for Chemical Process Fault Diagnosis

Abstract

Fault diagnosis techniques are essential to ensure the long-term reliability of industrial process systems. Current deep learning methods mainly rely on a large quantity of training data. Generative Adversarial Network (GAN) model has started to be utilized for diagnostic problems with small sample size and data imbalance in recent years, but the fault diagnosis performance heavily depends on the experience of the model builder. In this work, a high-efficiency GAN model (HGAN) is proposed for chemical process fault diagnosis. HGAN integrates the advantages of Wasserstein GAN and Auxiliary Classifier GAN to promote the generating model training stability and the discriminative model training efficiency with Bayesian optimization. Experiments on the benchmark Tennessee Eastman process under the circumstance of small samples show that the presented model can achieve a satisfactory fault diagnosis accuracy without the assistance of a redundant deep neural network classifier and with reduced effort in model tuning.