Photodiode signal patterns: unsupervised learning for laser weld defect analysis

Abstract

Laser welding, widely used in industries such as automotive and aerospace, requires precise monitoring to ensure defect-free welds, especially when joining dissimilar metallic thin foils. This study investigates the application of machine learning techniques for defect detection in laser welding using photodiode signal patterns. Supervised models, including Support Vector Machine (SVM), k-Nearest Neighbors (kNN), and Random Forest (RF), were employed to classify weld defects into sound welds (SW), lack of connection (LoC), and over-penetration (OP). SVM achieved the highest accuracy (95.2%) during training, while RF demonstrated superior generalization with 83% accuracy on validation data. The study also proposed an unsupervised learning method using a wavelet scattering one-dimensional convolutional autoencoder (1D-CAE) network for anomaly detection. The proposed network demonstrated its effectiveness in achieving accuracies of 93.3% and 87.5% on training and validation datasets, respectively. Furthermore, distinct signal patterns associated with SW, OP, and LoC were identified, highlighting the ability of photodiode signals to capture welding dynamics. These findings demonstrate the effectiveness of combining supervised and unsupervised methods for laser weld defect detection, paving the way for robust, real-time quality monitoring systems in manufacturing. The results indicated that unsupervised learning could offer significant advantages in identifying anomalies and reducing manufacturing costs.