Predicting and preventing defects in friction stir welding through sensor fusion and machine learning

Friction stir welding (FSW) produces high-integrity joints without melting; however, small variations in tool condition and process parameters can rapidly introduce defects, making real-time monitoring essential for industrial production. This study presents an explainable sensor-fusion and machine-learning framework for in process defect prediction in FSW, aimed at reducing quality-related costs by preventing scrap and rework and decreasing dependence on time-consuming post-weld inspection. Aluminium 1100 plates were welded using a high-carbon steel tool, and process parameters were planned using a Box–Behnken response surface design. During welding, signals were acquired from acoustic emission, vibration (accelerometer), and airborne sound (microphone) sensors. Time-domain features were extracted and evaluated using Support Vector Machines, Decision Trees, Random Forests and Neural Networks. The selected-feature fusion model achieved 98.2% accuracy, 99.1% precision and 97.3% F1-score, outperforming the single-sensor model based on the accelerometer (96.79% accuracy). Explainable AI (SHAP and LIME) identified the dominant contributing attributes, improving interpretability and supporting actionable decisions on the shop floor. Overall, the proposed approach enables robust in-process quality assurance and early defect detection, helping manufacturers reduce scrap and rework, shorten inspection cycles, and improve throughput in Industry 4.0-ready FSW production.