Enhanced Gap Classification in Laser Welding of Butt Joints Using Transfer Learning and Lightweight CNNs

Abstract Robotized laser beam welding demands precise control, with accurate joint gap estimation essential for adaptive gap-bridging. Challenges such as smoke, surface irregularities, and spectral emissions hinder classification performance. This study applies deep learning models to enable real-time gap classification during autogenous butt joint welding using beam-shaping. Images captured from varying joint gaps allowed evaluation under dynamic conditions. Four models were trained and assessed, with the best model surpassing application accuracy requirements. The findings support the development of adaptive welding systems for consistent weld quality in noisy, fluctuating environments.