Improvements in Modelling of Complex Manufacturing Processes Using Classification Techniques

The improvement of certain manufacturing processes often involves the challenge of how to optimize complex and multivariable processes under industrial conditions. Moreover, many of these processes can be treated as regression or classification problems. Although their outputs are in the form of continuous variables, industrial requirements define their discretization in compliance with ISO 4288:1996 Standard. Laser polishing of steel components is an interesting example of such a problem, especially its application to finishing operations in the die and mould industry. The aim of this work is the identification of the most accurate classifier-based method for surface roughness prediction of laser polished components in compliance with the aforementioned industrial standard. Several data mining methods are tested for this task: ensembles of decision trees, classification via regression, and fine-tuned SVMs. These methods are also tested by using variants that take into account the ordinal nature of the class that has to be predicted. Finally, all these methods and variants are applied over different transformations of the dataset. The results of these methods show no significant differences in accuracy, meaning that a simple decision tree can be used for prediction purposes.