Underlying Deep Learning Networks Diagnosis Evaluation and Generative Adversarial Network Data Augmentation Based on a Benchmark Accident Dataset

Abstract Nuclear energy plays an important role in the global energy supply, particularly as a key source of low-carbon electricity, while ensuring the safe operation of nuclear power plants (NPPs) is crucial. Given the significant impact of human-caused errors on three serious nuclear accidents in history, artificial intelligence (AI) has been increasingly interested assisting operators in NPP accident diagnosis. However, the field has lacked a benchmarking effort for underlying deep learning networks based on public datasets for accident diagnosis, as well as research on time-series data augmentation efforts for accident data. Based on the first open-source NPP accident dataset NPPAD, which covers all kinds of accident data of pressurized water reactors and has been validated with Fukushima nuclear accident data for reliability, this study explored accident diagnostic techniques based on Multilayer Perceptions (MLPs), Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs) and Transformers, and data augmentation techniques based on conditional Generative Adversarial Networks (cGANs) and Time-series Generative Adversarial Networks (TimeGANs). Experimental results show Transformers have the best classification ability for accident types. In addition, the standardization of data has a large impact on the classification effect, and standardization in data preprocessing will make all models substantially more effective. For data augmentation, TimeGANs exhibit higher performance compared to cGANs in generating fake datasets and can better capture the dynamic features of time-series data. This paper provides a reference for the application of deep learning through the study of three parts: a comprehensive comparison of underlying deep learning networks, data standardization processing, and data augmentation.