Avoiding neural network fine tuning by using ensemble learning: application to ball-end milling operations

Abstract

Surface roughness plays a key role in the performance of machined components—specially dies and moulds—manufactured for the aerospace and automotive industries, among others. However, roughness can only be measured off-line after the part has been machined, when cutting conditions may no longer be adjusted to surface roughness requirements. A reliable surface roughness prediction application is presented in this paper. It is based on ensemble learning for vertical high-speed milling operations with ball-end mills for finishing operations on quenched steel 1.2344 (AISI H13) that are widely used in the manufacture of moulds and dies. The new approach was validated with an experimental dataset that includes geometrical tool factors, cutting conditions, dynamic factors and lubricant type. An intensive comparison with an artificial neural network approach for the same dataset is included, to reveal the improvements of the new technique over other well-established ones for this industrial problem. This comparison shows that ensemble learning can by-pass the time-consuming task of tuning neural network parameters and can also improve prediction model accuracy, both of which are features that could lead to greater use of these kinds of prediction models in real workshops. Finally, a methodology, based on this new approach, is presented, in order to illustrate how the prediction model can be used in workshops to optimize cutting conditions in terms of their surface quality and productivity.