Penetration depth monitoring in laser welding via end-to-end deep learning with a spatial–intensity representation of optical coherence tomography

This is the authors' abstract. We don't add key points for this paper.

In lithium-ion battery manufacturing, laser welding is a critical technique for busbar connection, and penetration depth directly determines the electrical reliability and mechanical strength of battery modules. Consequently, stable online monitoring of penetration depth is essential for quality control in busbar laser welding. Optical coherence tomography (OCT), as a non-contact measurement technique based on coherent ranging, has attracted increasing attention in recent years for penetration depth monitoring. However, OCT monitoring data acquired during laser welding often contain pronounced noise with complex origins. Conventional filtering methods typically rely on manual parameter tuning and mainly consider the density distribution of the measurement points. To address the above issues, this study develops an end-to-end deep learning approach that integrates the spatial distribution and signal intensity information of OCT data, which preserves detailed keyhole fluctuation information and, to some extent, reflects the reflection and scattering behavior of the measurement beam within the keyhole. Based on the conversion of raw OCT point cloud data into two-dimensional images, the weld region in laser welding OCT images is identified and segmented by a segmentation neural network, thereby enabling accurate in situ penetration depth monitoring. Experimental results show that the proposed method achieves a mean absolute error (MAE) of 0.067 mm and is capable of capturing the continuous fluctuations of penetration depth in detail, demonstrating its accuracy and applicability.