Weld-penetration-depth estimation using deep learning models and multisensor signals in Al/Cu laser overlap welding

Abstract

Al/Cu laser-welded overlap joints, in which weld-penetration depth significantly influences both joint strength and electrical conductivity, are widely applied in automotive battery cells. In this study, a unisensor convolutional neural network (CNN) model that predicts penetration depth using coaxial weld-pool images as input and multisensor CNN models that utilize additional photodiode signals are proposed. The penetration depth was estimated using an optical coherence tomography sensor. The coefficient of determination values for the unisensor and multisensor CNN models were between 0.982 and 0.985, and their mean absolute errors were between 0.0278 and 0.0302 mm. The short-term Fourier transform multisensor model presented the best performance in terms of prediction of penetration depth when applied to the photodiode signal. The proposed prediction models were validated using a gradually varying laser power experiment, which demonstrated the efficacy of this approach and its potential use in automotive applications. Keywords: Laser welding, Al/Cu overlap joint, Penetration-depth estimation, Image sensor, Photodiode, CNN, Deep learning.