Modeling and real-time prediction for complex welding process based on weld pool

A major assumption in developing intelligent robot in industrial fields is that the intelligence has to be from senior human workers. However, in many industrial applications, this assumption may not hold. For example, in welding process, a senior welder can continually choose proper weld parameters and tune weld performance based on their observations of the liquid weld pool. But a robot with better pool shape measurement system and quicker respond speed may outperform human welders. In such cases, intelligent welding agents, if developed successfully, would greatly reduce welding cost by avoiding much expensive human training efforts and improve the welding performance. In recent years, machine learning techniques have emerged as a new learning framework to address such problem. The theory of machine learning provides a normative account of how agents may predict their future action based on a model structure, in which the training and future data have to be in the same feature space and have the same distribution. In this paper, Gaussian Process Regression (GPR) is proposed to model real-time welding process. Gas tungsten arc welding experiments were performed and the experimental data are utilized to validate the proposed method. We train the model using 1284 pairs of data in one experiment and test this model on a challenging domain of 4 other experiments by 11,642 pairs of data, using the same algorithm, GPR architectures, and hyperparameters. Bayesian Optimization Algorithm (BOA) surrogated with GPR is proposed to provide robust prediction results. This paper opens a door for enabling robot with industrial skills, resulting in the first artificial welder that is capable to respond real-time observations and provide robust predictions.