Modelling the relationship between process parameters and mechanical properties using Bayesian neural networks for powder metal parts

A neural network based system is presented in this paper for modelling mech anical behaviour of powder metal parts as a function of processing conditio ns. The neural network selection is made using a Bayesian framework, which enables prediction of mechanical properties to be made, indicating a level of confidence in the result. The system gives good prediction accuracy for a number of commercially available ferrous powder materials; the performanc e for two different powder grades is reported. In order to select process p arameters that meet the required mechanical properties for the part, a prot otype process 'advisor' is developed using these neural network models. Thr ee different neural networks are trained to predict tensile strength, elong ation and hardness for ferrous powder grades, and are used in the process ' advisor' to recommend suitable process parameters.