Ensuring the integrity of the primary circuit in nuclear power plants is crucial considering the extreme pressures and temperatures while operating Pressurized Water Reactors (PWR). Non-Destructive Testing (NDT) on such harsh environments is a challenging and complex scenario. Automated assistance on acquisition and analysis systems can importantly contribute as supplementary safety barrier by providing real-time alarms for potential existence of defects.
In this paper we present the application of Artificial Intelligence in Visual Testing (VT) of Bottom Mounted Nozzles (BMN) of the Reactor Pressure Vessel (RPV). The method that we apply is based on Object Detection using Convolutional Neural Networks (CNN) combined with the Transfer Learning technique in order to limit the necessary training time of the model and the use of Data Augmentation methods for reducing the size of the learning data set. The proposed CNN demonstrates great performances for automatic surface defect detection (cracks) in highly noisy environments with variating illumination conditions. These performances combined with accurate localization and characterization of the defects confirms the interest of advanced CNNs against traditional imaging processing methods for NDT applications. In this study, the results of a comparative blind-test between Human VT analysts are also presented.