In this paper, a feed-forward back-propagation artificial neural network (BP-ANN) and analysis of variance (ANOVA) are applied to a hot metal extrusion process, establishing a black box model as well as analyzing the effects of relevant process parameters on required forging load, under different operating conditions. Some finite element simulation data on extruding ck-45 steel, adopted from a published research paper, were used to train the neural model employing Levenberg-Marquardt learning algorithm. Die angle (15°–75°), friction coefficient between billet-die material pair (0.4–0.8), punch velocity (168–203 mm/s), and billet temperature (1000°C–1260°C) were selected as the inputs, while the extrusion load (tone) was considered as the network’s output. Based on the results during modeling attempts, a 4-10-10-1 size neural network has been decided on as the appropriate architecture of the process model. Testing predictive accuracy of the developed model was also done using a new data set (8 data samples), which has not been used in the training phase. The comparative errors with respect to the desired FEM simulations are all in acceptable ranges (less than 12%) thereby the network’s generalization capabilities were confirmed. Having established the appropriate neural model, analysis of variance (ANOVA) technique was then applied to the original training data base to find and recognize the level of importance of each parameters and their possible dual interactions on the extrusion loading force within 95% of confidence interval (α = 0.05). Based on the obtained inferences, the best optimal combination of parametric settings which leads to the minimum required extruding load was then revealed and recommended. The optimally minimized extrusion force was then predicted by the trained network model. Neural network tool box (NNET) of the Matlab software and design of experiments module of Minitab software were employed as platforms to develop neural simulations and ANOVA technique, respectively. The overall results indicate the feasibility and effectiveness of the proposed approach in a real manufacturing environment and eliminate the need to carry out expensive as well as time consuming trial and error experimentations to reach to the optimum operating conditions.

This content is only available via PDF.
You do not currently have access to this content.