An Original Machine Learning-Based Approach for the Online Monitoring of Refill Friction Stir Spot Welding: Weld Diagnostic and Tool State Prognostic

The process monitoring (PM) of refill friction stir spot welding (Refill FSSW) can play a substantial role in detecting various issues, especially defects in the spot being formed and the tool state degradation, which allows in time intervention to improve the welding process. Since Refill FSSW is somewhat an emergent technology, PM has received scarce attention. In this paper, the performance of PM using acoustic emission (AE) technique is studied for two purposes: detecting defects in weld while being formed and predicting the tool state. To do so, the common defects that can occur during the process were first intentionally created and monitored using AE. The corresponding collected data have served then as an input for two defect detection models. The first one is based on novelty detection and has shown an average classification performance. The second, which shows higher performance, uses multi-class classification algorithms. Concerning the tool state, a novel state index was developed to predict when the process must be stopped in order to clean the tool and avoid hence related weld defects and tool fracture.