Laser welding via scanner is employed for foil-cap welding in tabless batteries, enabling high productivity and flexibility. However, when welding dissimilar materials, keyhole often becomes unstable due to differing thermal and physical properties. This instability results in uneven penetration depth, significantly increasing risk of jelly roll damage and weld defects. Therefore, in-situ process monitoring and precise control of penetration depth during welding are essential.

To achieve in-situ monitoring, commercial systems have been developed based on optical sensors, photodiodes, spectrometers, and optical coherence tomography. Fast welding speeds, not only reduce amount of acquired sensor data but also deteriorate its quality. Despite its importance, research on the diagnosis and prediction of welding quality based on artificial intelligence algorithms remains insufficient. In authors’ previous studies, spectrometers were shown to classify penetration depth at low welding speeds by detecting specific spectra emitted from metal vapor and plasma. However, it is difficult to monitor welding processes at speeds exceeding 20 m/min due to their limited sampling rate. Photodiodes offer a higher sampling rate, enabling temporal resolution of process dynamics, even though they provide compressed spectral information compared to spectrometers. Therefore, the combined use of multiple sensors is expected to compensate for limitations of each sensor, enhancing classification accuracy of penetration depth under high-speed welding conditions.

In this study, a CNN-based classification model was developed using photodiode and spectrometer data collected at welding speeds exceeding 20 m/min, optimized for training time and high accuracy. Furthermore, fine-tuning reduced error rates and enhanced robustness under diverse welding conditions.