Ensemble machine learning for multiclass welding quality classification using sensor-derived signal features

This study proposes an ensemble machine-learning framework for multiclass MAG welding joint quality classification to reduce reliance on destructive inspection. Acoustic, electrical, and vibration signals were acquired using electric microphones, ADXL386 accelerometers, and current sensors. The experimental dataset, consisting of 105 original samples, was divided at the experiment level into training (80%) and testing (20%) sets to minimize data leakage and overfitting. The signals were processed using a genetic algorithm-optimized band-pass filter, denoised, and segmented into one-second windows, resulting in 1800 data segments. Features were then extracted from the time, frequency, and time-frequency domains and refined using delta-threshold mRMR feature selection. Class imbalance was addressed using oversampling, SMOTE, SMOTE-ENN, ENN, and undersampling. A stacked ensemble with SVM and Random Forest (RF) base learners and a multinomial logistic regression meta-learner captures both local and global decision structures. With Optuna-based hyperparameter tuning and stratified cross-validation, the proposed model achieves 94.4% accuracy, a 94.4% weighted F1-score, an 87.3% macro F1-score, an 87.1% balanced accuracy, and a 98.6% weighted ROC-AUC, outperforming standalone SVM, RF, SVM+XGBoost, and XGBoost classifiers. Confusion matrices and ROC-AUC curves further demonstrate the robustness and generalization capability of the proposed approach for imbalanced multiclass welding data.