Detection of Separator Contact Defects Using Spectroscopic Signals and Deep Learning in Copper Laser Welding for Lithium-Ion Batteries

Separator contact during laser welding of lithium-ion battery cells can cause thermal damage and safety issues. This study investigated whether spectroscopic signals could be used to distinguish normal welds from separator contact defects and evaluated the effects of spectrometer configuration and deep learning models on defect classification performance. A spectrometer-based monitoring system was established, and the signal characteristics were analyzed according to fiber core diameter, slit width, optical configuration, and integration time. Based on the optimized spectrometer conditions, deep learning models including LSTM, GRU, CNN, TCN, and CNN+LSTM were trained using both characteristic wavelength ranges (350-450 and 650-750 nm) and the full spectral range (232-1110 nm). Models trained with the characteristic wavelength ranges achieved separator F1-scores of 0.87-0.88, whereas those trained with the full spectral range achieved F1-scores of 0.91-0.99. Among the evaluated models, LSTM showed the best performance with an F1-score of 0.99. These results demonstrate that a spectrometer configuration combined with deep learning provides an effective approach for the detection of separator contact during laser welding.