Defect detection in MAG welding using multi-sensor Wavelet Scattering Features and Deep Autoencoders

This paper presents an automated defect detection system for Metal Active Gas (MAG) welding, designed to improve the speed and reliability of welding. Defects can significantly compromise welding joint quality. The proposed system uses data collected from microphones, vibration sensors, and photodiodes during welding. These signals are processed using wavelet scattering transforms to extract features that capture the underlying patterns of the welding process. The resulting scattering coefficients are organized into vectors and fed into a deep autoencoder trained exclusively on defect-free welded joints. Defects are detected based on mean squared error, with an optimal threshold selected to clearly separate defect-free (“Good”) joints from defective ones by maximizing the F1 score. Since the model has not seen defective data during training, it produces higher errors when processing flawed welded joints, allowing for effective unsupervised defect detection. Comparative evaluation with supervised models indicates that SVM achieves 95.9% accuracy, 95.9% F1 score, and 99.5% ROC AUC with SMOTE-based balancing, while 1D CNN achieves 96.9% accuracy, 96.9% F1 score, and 99.9% ROC AUC with SMOTE. In contrast, the deep autoencoder demonstrates superior performance with 98.1% accuracy, 98.3% F1 score, and 99.7% ROC AUC without requiring labeled defect data or class balancing. These results highlight the effectiveness of the proposed unsupervised framework as a robust and scalable solution for defect detection in highly imbalanced welding datasets.