Automatic image analysis of surface defects in hybrid welding of dissimilar joints for lifecycle enhancement

Surface defects generated during laser-arc hybrid welding of dissimilar Al-steel joints act as critical fatigue crack initiators, yet process qualification often relies on subjective inspection and trial-intensive parameter exploration. This study presents an end-to-end framework that converts in-line weld-surface imaging into quantitative, fatigue-relevant quality guidance for laser-leading hybrid welding of AA6061-T6/DP590. Distance-triggered imaging with controlled illumination, pixel-to-mm calibration, and flat-field normalization ensures stable acquisition. A U-Net segmentation model achieves robust performance with Intersection over Union (IoU) of 0.92/0.89/0.88 (train/validation/test) and test-set AUPRC of 0.91. Post-processing enables instance-level defect quantification, and multiple descriptors are integrated into a Defect Severity Index (DSI) exhibiting high repeatability (DSI CV < 2.5%) and strong agreement with expert measurements (N = 50 ROIs). Using DSI as the response variable, an XGBoost model trained on N = 45 process conditions predicts defect severity and, with SHAP-based interpretation, identifies a recommended low-defect operating window validated by confirmation experiments. Fatigue S-N behavior stratified by DSI thresholds (Low ≤ 0.20, Mid 0.20-0.60, High > 0.60) demonstrates systematically longer life at lower DSI, establishing a deployable accept/monitor/rework decision framework that links surface quality directly to lifecycle performance.