Neural networks based prediction modelling of hybrid laser beam welding process parameters with sensitivity analysis

Abstract

Present paper attempted to model complex relationship between CO2 laser–MIG hybrid welding parameters and it has been completed using different algorithms of artificial neural networks (ANN). Input parameters for the study include laser power, welding speeds and wires feed rate and tensile strength of the joint is considered as output. A full factorial experimental dataset is used for the purpose. Variants of back propagation neural networks (BPNN) and Radial Basis Function Networks have been used as training algorithm. Altogether 65 different ANN architecture have been trained and tested using 6 different training algorithms to find out ANN with best prediction capability. 3-11-1 ANN architecture trained using BPNN with Bayesian regularization shows best prediction capability (mean square error 3.24E − 04) and considered as Best ANN. That ANN will be useful for determining required value of welding process parameters to yield a specific welding strength and suitable for online process monitoring and control. Finally, a sensitivity analysis has been conducted and it is found that, maximum welding strength can be obtained with low wire feed rate (4 m/min), low welding speed (2 m/min) and high laser power (3 kW).