Improving Stability of Welding Model with ME-ELM

Abstract

Welding shape is important in evaluating welding quality, but accurate predictive model is hard to achieve, because welding is a complex nonlinear process, and the sampled data are inevitably contaminated. Extreme learning machine (ELM) is used to construct a single-hidden layer feedforward network (SLFN) in our study, for improving stability of welding model, M-estimation is combined with ELM and a new algorithm named ME-ELM is developed; researches indicate that it works more effective than BP and other variants of ELM in reducing influence, furthermore, it can improve the model’s anti-disturbance and robustness performance even if the data are seriously contaminated. Real TIG welding models are constructed with ME-ELM, by comparing with BP, multiple nonlinear regression (MNR), and linear regression (LR), conclusions can be gotten that ME-ELM can resist the interference effectively and has the highest accuracy in predicting the welding shape.