Fault diagnosis study for nuclear power plants under imbalanced fault sample datasets based on deep learning

The safety of nuclear power plants is crucial, and the situation awareness of the power plant is significantly important for the plant operation during which the fault diagnosis technology plays a key role for the plant state monitoring. Operation of the nuclear power plant is predominantly in a normal state, resulting in a scarcity of fault data. This scarcity leads to imbalanced training data, posing a challenge to the accuracy of fault diagnosis models. This study proposed a fault diagnosis framework that integrates Self-Attention Time-Series Generative Adversarial Networks (SATimeGAN) to augment the training datasets and Convolutional Neural Network-Bidirectional Long Short-Term Memory networks (CNN-BiLSTM) to diagnose the faults, aimed at addressing the data imbalance issue and improving the diagnostic accuracy. Experimental results demonstrate preferable performance of the model in nuclear power plant fault diagnosis, especially in handling imbalance sample fault datasets, compared to traditional CNN, LSTM, and BiLSTM models.