Evaluating the effects of process parameters on maximum extrusion pressure using a new artificial neural network-based (ANN-based) partial-modeling technique

Metal extrusion process accounts for the production of the majority of industrial and domestic aluminum sections. A major limitation to the success of any extrusion operation is the capability of the particular extrusion press to meet the maximum pressure requirements for that operation. In the present work, the effects of industrial extrusion process parameters and their interactions on the resulting maximum extrusion pressure, of an industrially extruded aluminum alloy, have been studied using a newly devised ANN-based partial modeling technique. Two operating parameters (initial billet temperature and ram speed) and three geometrical parameters (extrusion ratio, profile average thickness, and number of die cavities) were investigated. The main objective for developing this modeling technique is to overcome the limitations of presently available statistical modeling tools, as foreseen by the modeling needs for a complex thermo-mechanical process such as extrusion. The main present limitations are accounting for non-linearity in the process behavior, incorporating interaction effects and a meaningful determination of the highly significant process parameters and/or interactions. These three features have been, collectively, incorporated into the present model by means of combining statistical analysis of variance into ANN and by using a partial sum of squares analysis, which we propose to call the “present factor analysis.” Normal linear regression has been also employed for comparison purposes. According to the present model, maximum extrusion pressure has shown various degree of non-linearity in behavior with respect to the different process parameters and their significant interactions. It has been found that variations in the maximum extrusion pressure are mainly a function of initial billet temperature and its interactions with other process parameters, especially the ram speed. The present ANN-based model has shown superior prediction capabilities compared to the linear model with a marginal overall prediction error value of ±2.5 %.