Open-set recognition based on the combination of deep learning and hypothesis testing for detecting unknown nuclear faults

Most current fault diagnosis techniques for nuclear systems mainly rely on the closed-set assumption, which restricts the diagnosis model to select from a set of pre-established known fault classes. However, the nuclear system is a dynamic open system, and unknown faults that have never been seen can occur at any time. Therefore, it is very meaningful to design a diagnosis model that can recognize both known and unknown faults. This paper proposes a fault diagnosis method for open-set scenarios. Specifically, a modified loss function is used to train a convolutional neural network (CNN) to learn more compact feature representations of known classes. The features output by the last fully connected layer of the CNN are taken as the scores belonging to each known class, and a calibration model based on extreme value theory (EVT) is introduced to calibrate the scores. In addition, hypothesis testing is introduced for statistical inference. The threshold is determined according to the confidence level to distinguish the known faults from the unknown faults. Experiments conducted on two sets of nuclear system faults simulation data demonstrate that the proposed model not only identifies more unknown faults without compromising the accuracy of known fault classification but also selects more appropriate thresholds for different datasets, thereby enhancing the model's generalization capability. Furthermore, experiments under varying degrees of openness also prove that our model exhibits higher robustness across different levels of openness.