Machine learning applications in nuclear power plant piping inspection: A review of methods, data, and future trends

Abstract

Piping safety is an important concern in nuclear power plants, where an inaccurate prediction of pipe failure will lead to financial loss and fatal incidents. Traditional inspection methods are too laborious and costly, and require expert knowledge. Machine learning (ML) models can learn from the available data and be used for nuclear piping safety. A comprehensive literature search reveals that no existing review has focused exclusively on piping safety using ML in nuclear power plants (NPPs). Keeping this in view, efforts have been made to provide a comprehensive review of this topic. The review reveals the main research gaps by investigating the current studies conducted for the safety of NPP piping systems subjected to corrosion using ML. Various ML models and datasets have been reviewed in this review article. It was observed that convolutional neural network, support vector machine, and artificial neural network are the most widely developed models. Vibration-based datasets have been extensively utilized in ML applications for analyzing pipe degradation due to their effectiveness in capturing structural health changes and predicting failures, pipe degradation severity, pipe wall thinning rate, flow accelerated corrosion rate, and pipe elbow thinning rate. Despite the successful implementation of the ML models on available data, major limitations include data scarcity (due to cost and labor required to run the inspections or simulations) and limited scenarios (restricted operating conditions and geometries of piping). In addition, future recommendations such as applying synthetic minority oversampling technique, generative adversarial neural network, transfer learning, and physics-informed ML are discussed to improve the ML application for NPP piping systems.