Prediction of Welding Spatter Generation Rate in Short-Circuit Transfer by Convolutional Neural Network

Abstract

In this study, a machine learning method is proposed to predict the weld spatter generation rate, regardless of the type of short-circuit waveform. Short-circuit waveform data are collected at a high sampling rate of 10?20 kHz, and then, compressed using a new preprocessing method to effectively process the data at a high sampling rate. To predict the spatter generation rate, the welding waveform is converted into an image using the proposed data preprocessing method, and the converted data are fed into the convolutional neural network (CNN). A parametric study on data augmentation and data resolution is conducted concurrently to enhance the prediction accuracy with limited amount of data.