Unsupervised Machine Learnig approach for NPP LTO Program

Recent advancements in data-driven analysis methods, represented by those in artificial intelligence and machine learning, are improving the NPP performance ranging from the anomaly detection to the automated operational control of its complex systems. Indeed, the application of these methods can significantly improve the ability to operate safely NPP also in the long-term. In this framework, it is worthy to note that more than 67% of the reactors in operation must face ageing as they are more than 30 years old. This paper focuses on unsupervised Machine Learning (ML) and artificial neural network (ANN) approaches for anomaly detection of SCCs of NPPs. These methods, based on Mahalanobis distance and autoencoder neural networks respectively, are described including tasks of data analysis, monitoring, prognostics etc. Both ML and ANN were tested on anomaly pattern that deviates from nominal/normal plant conditions. LTO condition is also considered. To the aim of this study, the dataset is provided by a digital twin of primary pipe under inner temperature of 300C and internal pressure of 15.5 MPa. Finally, the two approaches are compared for performance assessment. The findings suggest that the implemented methodology is able to predict the pipe failure. The transition from time-based maintenance to predictive maintenance demonstrates to support in a profitable way NPP operation and LTO program allowing also to increase the value of nuclear reactor assets by potentially precluding serious consequences due to faults and failures of plant components.