Abstract
Surface roughness generation is influenced by many complex and interrelated factors. Moreover, in real industrial conditions many different milling tools have to be used to create a final product. Hence, the acquisition of experimental data used to set up artificial intelligence models of individual tools is a complicated task.
The aim of this paper is to present a new strategy to improve the artificial intelligence models used to predict surface roughness with small datasets. The new strategy proposed in this paper relies on dimensionality reduction to minimise the number of experiments required to train the models. Unlike in most other approaches, the dimensionality reduction is not applied to set an a priori determined dimension. In the proposed approach, the scale of dimensionality reduction is controlled by the quality of roughness prediction models created with the transformed data.
This strategy has been tested on high torque milling operations using multilayer perceptrons i.e. the most frequently used artificial intelligence models for this task. Experiments were conducted to obtain data to train the models. Finally, a comparison was made of the models' performance with and without dimensionality reduction based on Principal Component Analysis, which confirmed the merits of the proposed technique.