Failure Modes Detection of Nuclear Systems Using Machine Learning

Abstract

Early detection of the failure of a nuclear system is an important topic in nuclear energy. This paper proposes three machine learning methodologies to detect the failure modes (FM) of the Lead-Bismuth Eutectic eXperimental Accelerator Driven System (LBE-XADS) nuclear system after the first 10%, 50% and 90% time periods of the 3000 seconds mission time of the LBEXADS. The first methodology detects the FM of the LBE-XADS after the first 10% time period and consists of two Gaussian mixture-based (GM-based) classifiers. The second methodology detects the FM of the LBE-XADS after the first 50% time period and consists of a GM-based classifier and a neural network MLP1. The third methodology detects the failure mode of the LBE-XADS after the first 90% time period and consists of a GM-based classifier and a neural network MLP2. The three proposed methodologies outperformed the fuzzy similarity approach of the previous work.