Prediction of interface width in overlap joint configuration for laser welding of aluminum alloy using sensors

Abstract

We present a method that can predict the interface width in an overlapping joint configuration for laser welding of Al alloys using sensors and a convolutional neural network (CNN)-based deep-learning model. The inputs for multi-input CNN-based deep-learning prediction models are spectral signals, represented by the light intensity measured by a spectrometer and dynamic images of the molten pool filmed by a charge-coupled device (CCD) camera. The interface width, used as learning data for modeling, was constructed as a database along with the process signal by cross-sectional analysis. In this study, we present results showing high accuracy in predicting the interface width in the overlap joint configuration for Al alloy laser welding. For predicting the interface width, five models are created and compared: a single CCD and spectrometer sensor algorithm, a multi-sensor algorithm with two input variables (CCD, spectrometer), a multi-sensor algorithm excluding the processing beam in the spectrometer data on the combination of Al 6014-T4 (top)/Al 6014-T4 (bottom), and a multi-sensor algorithm applied to the combination of Al 6014-T4 (top)/Al 5052-H32 (bottom). The multi-sensor algorithm with two input variables (CCD and spectrometer) on the same material combination showed the highest accuracy among the models.