Fault diagnosis based on conditional generative adversarial networks in nuclear power plants

• Fault diagnosis in nuclear power plants suffers from small sample problem. • A lightweight conditional generative adversarial network model is proposed. • Three fault datasets are selected and built for method verification. • The proposed method can generate multi-class high-quality fake samples. • The augmented datasets can improve diagnosis performance of many classifiers. Fault diagnosis techniques can detect abnormal states of equipment or systems, give warning information timely, help to optimize the maintenance schedule, reduce unplanned shutdowns, which play an important role in the safe and economic operation of nuclear power plants (NPPs). However, NPPs operate in normal state most of the time with few fault samples, forming a small sample fault diagnosis problem. In this paper, a fault diagnosis method based on lightweight conditional generative adversarial networks (CGAN) is investigated. 3 case studies indicate that the proposed method can generate high quality multi-class fault samples in small sample scenarios and significantly improve the fault diagnosis performance of several popular fault classifiers, like multi-layer perceptron, convolutional neural network, deep belief network and denoise auto-encoder. Experiment results also show that the proposed method has good diagnosis performance for both rotating machinery and NPP system fault datasets.