UNSUPERVISED NEURAL NETWORK APPROACH FOR IDENTIFYING ANOMALOUS EVENTS IN NUCLEAR POWER PLANTS

抄録

Machine learning is a branch of engineering that focuses on designing algorithms that can analyze and make predictions from data. It is a subfield of artificial intelligence that utilizes mathematical optimization methods for specific tasks.

Anomaly detection is an unsupervised technique in the field of machine learning. In the context of predictive maintenance, anomaly detection is used to identify unexpected behavior in systems so that preventive action can be taken to avoid potential failures. In this study a methodology based on anomaly detection by means autoencoder algorithm is used to predict in advance failure of structure, system, and components (SCCs). An autoencoder is a neural network that is used to learn efficient data representations in an unsupervised manner. The objective of an autoencoder is to transform the input data into a condensed, latent representation, and then use this representation to recreate the original input. The dataset is provided by PWR-2loop simulator under nominal condition. The dataset can include several sensors output (temperature, pressure, etc.). Further, NPP-simulator can implement anomalous event. The goal of proposed approach is to detect in advance abnormal event avoiding the cost of unplanned downtime and improving equipment reliability. The results show the soundness of approach, predicting anomalous pattern before reactor scram. The methodology is a powerful tool for anomaly detection in predictive maintenance. When used correctly, autoencoders can learn a robust representation of normal behavior that can be used to effectively detect anomalies.