Imbalanced Chemical Process Fault Diagnosis Using Balancing GAN With Active Sample Selection

Current deep-learning-based fault diagnosis methods, though proven to be successful with sufficient fault data, cannot well address the challenges of sample availability in real-world industrial scenarios. To address this challenge, this article proposes an effective approach by exploiting data generation and sample selection techniques. Specifically, we first develop a balancing generative adversarial network (BAGAN)-based data generation technique to generate more discriminative fault samples by utilizing not only the fault samples, but also the normal samples. Second, a strategy is devised to select the samples generated by BAGAN, and on this basis, active learning (AL) is utilized to select the most informative samples. The stacked autoencoder (SAE)-based deep neural network (DNN) is used to classify the faults. The proposed method is evaluated by conducting computational experiments on the Tennessee Eastman (TE) dataset. The evaluation results demonstrate that the proposed BAGAN-based method with an active sample selection strategy achieves improved performance in imbalanced chemical fault diagnosis tasks.