Umeå University's logo

Since the 1960s, nuclear power has been produced in Sweden, and as early as 1999, a process of decommissioning and dismantling six of the country’s twelve active nuclear reactors was initiated. To address the challenge of managing nuclear waste, a method known as KBS-3 has been developed. This method involves encapsulating the nuclear waste in a canister consisting of a load-bearing insert for mechanical stability, and a copper shell to provide a corrosion barrier. The copper-clad canisters are then to be embedded in bentonite clay, 500 meters deep within the Swedish bedrock.

In a future construction of this final repository, approximately 6,000 canisters will be manufactured, filled with spent nuclear fuel, sealed, and buried in the bedrock. The sealing of the canisters using the KBS-3 method will be carried out through friction stir welding. Currently, non-destructive testing of the welding process is performed using three techniques: ultrasound, eddy current testing, and radiography. Today, these techniques are used to analyze how to avoid defective weld zones.

Ensuring that the weld areas are free from significant defects will be a critical aspect of a future canister production, as this will prevent the risk of nuclear waste leaking into the environment, which could pose a threat to human health and ecosystems. At present, the analysis of weld defects is conducted manually by inspecting one image at a time. For a future production setting, there is a need to streamline this manual inspection process through more efficient methods. This report presents an analysis of effective machine learning methods for defect detection as a solution to this problem.

Four focus areas for detecting weld defects in X-ray images have been explored. The focus areas have been multi-class and binary defect classification, anomaly detection, and real-time defect detection. Two model comparisons have been made in classification, where a simple CNN (Convolutional Neural Network) and a more complex CNN-based model structure have been investigated. These two models have in turn been compared in application on two classes (defective and non-defective), as well as four classes of weld defects (cracks, porosity, lack of penetration, and no defect). Furthermore, anomaly detection using an Autoencoder model, and real-time defect detection using a re-trained YOLO model has been explored. The Simple CNN model obtained a test accuracy of 85.02% and 89.07% for the 4-class and 2-class datasets respectively, and the more complex WelDeNet model obtained a test accuracy of 94.42% and 96.65% for the two datasets respectively. The Autoencoder model for anomaly detection obtained a test accuracy of 98.89% for the mixed 2-class test dataset, and the re-trained YOLO v8 Nano for defect detection obtained a mAP (mean Average Precision) of 0.995 for two classes, as well as an F1-Score of 99%.

This project presents a set of machine learning models that can obtain high accuracy test results, and therefore have good reasons to be considered and further developed in a production context in the future.