An on-line arc welding quality monitor and process control system

Abstract

This paper illustrates an on-line are welding quality monitor and process control system that combines two modified cerebellar model articulation controller (MCMAC) neural networks and a linear discriminant function (LDF) method to establish an on-line (1) prediction of the quality measurements, (2) quality classification measurement of the are welding process, and (3) corrective estimation of are welding process controllable variables system. The approach uses parallel multiple input state variables and a linear neighborhood sequential training (LNST) algorithm, which make MCMAC faster than conventional CMAC and back-propagation neural networks. It also produces a useful quality indicator for experts to predict weld quality, eliminate waste (rework and lost production time), and reduce production and maintenance cost. The comparison of conventional CMAC and MCMAC network shows the training efficiency of MCMAC, which based on CMAC parameter selection, training algorithm, training sample selection memory size, and convergence time. The proposed online quality monitor and process control system has been trained and tested with welding arc sound signals, which have shown satisfactory accuracy for welding quality classification and great potential for real-world process control applications.