Uncertainty Quantification Based Detection and Diagnosis for Unexpected Accident in Nuclear Energy Systems

A nuclear accident can have severe consequences for human society. Recently, deep learning detection and diagnosis technology has been widely employed in the nuclear field, thanks to advancements in artificial intelligence technology. However, training a deep learning (DL) model requires a significant number of labeled samples, and it is not possible to assess unexpected accidents. To address this issue, this study proposes an intelligent accident detection approach based on uncertainty. Initially, a probabilistic deep learning network is utilized to obtain diagnostic findings. Next, both known and unexpected accident samples are used to provide quantitative measures of uncertainty. The unexpected accident samples are then identified, and a threshold is established using the statistical outlier detection method to compute the overall detection rate. The operator is provided with the results for human intervention. Furthermore, the advanced nuclear system's integrated simulation experiment platform offers the opportunity to study with sample data from simulated accidents. The results of these experiments provide support for the effectiveness of the suggested probabilistic-based intelligent detection strategy.