Using photodiodes and supervised machine learning for automatic classification of weld defects in laser welding of thin foils copper-to-steel battery tabs

This paper has been designed to study whether photodiodes and supervised machine learning (ML) algorithms are sufficient to automatically classify weld defects caused by simultaneous variation of the part-to-part gap and laser power during remote laser welding (RLW) of thin foils, with applications in battery tabs. Photodiodes are used as the primary source of data and are collected in real-time during RLW of copper-to-steel thin foils in the lap joint. Experiments are carried out by the nLight Compact 3 kW fiber laser integrated with the Scout-200 2D scanner. The paper reviews and compares seven supervised ML algorithms (namely, k-nearest neighbors, decision tree, random forest, Naïve–Bayes, support vector machine, discriminant analysis, and discrete wavelet transform combined with the neural network) for automatic classification of weld defects. Up to 97% classification rate is obtained for scenarios with simultaneous variations of weld penetration depth and part-to-part gap. The main causes of misclassification are imputed to the interaction between welding parameters (part-to-part gap and laser power) and process instability at high part-to-part gap (high variation in the process not captured by the photodiodes). Arising opportunities for further development based on sensor fusion, integration with real-time multiphysical simulation, and semi-supervised ML are discussed throughout the paper.