Deep-learning-based alarm system for accident diagnosis and reactor state classification with probability value

The reactor protection system (RPS) in a research reactor is a well-known conventional setpoint-based protection system. The RPS performs protective actions with the generation of alarms when the measurement values exceed the setpoints. The RPS has disadvantages in that alarms are not generated before the measurement values exceed the setpoints; they are generated at the time of protection actions are performed. In addition, each alarm has a direct relation with signals, not accidents, so it is difficult to identify the accident type quickly. Thus, new methods are required to diagnose and classify accidents. We propose a deep-learning-based alarm system. The proposed alarm system is modeled with convolutional and fully connected neural networks. The proposed scheme is designed from safety analysis in the safety analysis report. We prepare various datasets and scenarios for training and test. The results show that the proposed alarm system provides fast diagnosis alarms with probability values.