<title>Predicting the backside width of weld pool during pulsed GTAW process based on a neural network model</title>

Abstract

Modeling of welding process is the base of process control. Because welding process is a multivariable, strong coupling, time-varying and nonlinear system, traditional modeling methods are not suitable. In this paper, the dynamic neural network model for predicting backside width of pulsed GTAW weld pool by welding parameters and topside shape parameters was constructed. Orthogonal method was applied to design the sampling experiments. Experiments were carried on low carbon steel with 2mm thickness during pulsed gas tungsten arc butt-welding with gap. Based on self-developed vision sensor, double-side images of weld pool were captured simultaneously in a frame. By image processing the topside dimension and shape of weld pool, such as length, maximum width, gap width and the half-length ratio, and the backside dimension such as area, length and maximum width were calculated. Artificial neural network was applied to establish the model for predicting backside width of weld pool. The inputs of the model were the topside dimension, shape of weld pool and welding parameters such as pulse current, pulse duty ratio, and welding speed. The output of the model was the backside width of weld pool. The algorithm was the extended delta-bar-delta (EDD), and the learning ratio automatically determined by the algorithm. Threshold function was sigmoid function. The training cycle was selected to be 50000. The final EMS error of backside width was 5.2 percent. The simulation experiments were carried out to test the accuracy of the ANN model. From the results of the test, the output of ANN model can predict the backside width precisely.