New method of disc electrode wear monitoring in contact arc metal grinding using deep learning and signal clustering techniques

Abstract Contact arc metal grinding (CAMG) is a thermally driven underwater cutting method developed for decommissioning in nuclear facilities. In contrast to conventional mechanical cutting methods, CAMG employs a hydraulically actuated rotating disk electrode that is supplied with high electric current. As the electrode approaches the workpiece, an electric arc is ignited, generating localized melting that facilitates the controlled removal of material. Inherent radiation shielding, effectively prevents the spread of airborne contamination, and enhances overall safety of the operating personnel. Thus, CAMG offers a particularly effective solution for the dismantling of highly activated or hard to access components where traditional mechanical tools cannot be used. Monitoring electrode wear plays a vital role in CAMG. The condition of the electrode can be evaluated through the analysis of current and voltage signals recorded during operation. The present study introduces a novel deep learning based approach for monitoring electrode wear in CAMG using signals recorded from experimental CAMG trials. Convolutional neural networks (CNN) are used to extract representative features from the signals, followed by Principal Component Analysis and t-distributed Stochastic Neighbour Embedding to evaluate clustering behaviour. The results show that wear progression is reflected in distinctive temporal patterns. Worn electrodes exhibit fewer characteristic peaks and lower electrical power. CNN-derived latent representations form clearly separable clusters corresponding to early, intermediate, and advanced wear levels. These findings demonstrate that combining CNN-based feature extraction with unsupervised clustering provides a reliable basis for real-time electrode condition assessment in CAMG and supports progress toward automation in underwater dismantling operations.