Evaluation of machine learning algorithms for robustness in the classification of flat surfaced materials for varied roughness during machining

Abstract

This work evaluates the performance of various machine learning (ML) algorithms, including support vector machines (SVM), logistic regression, and decision trees, in classifying materials based on variations in surface roughness. The robustness of these classification algorithms is tested using a non-contact method based on machine vision for flat-surfaced materials with different roughness levels. The materials considered for training the classifiers include Aluminium (Al), Copper (Cu), Medium-Density Fibreboard (MDF), and Mild Steel (MS). Surface images of these materials, each with varying roughness, are acquired for feature extraction. Subsequently, three distinct datasets are curated from these features and used to train the algorithms to evaluate their robustness in material classification. Each algorithm is trained on one dataset and tested on the other two, with this process repeated across all datasets and classification models. The results indicate that the models performed inadequately on new datasets with differing roughness values. To address this limitation, all three datasets are merged into a single combined dataset. While the models did not perform satisfactorily on individual datasets, they achieved an overall accuracy of more than 90 % on the combined dataset. Among the algorithms, decision trees demonstrated superior performance, achieving a prediction accuracy of 99 % and recall, precision, and F1 scores exceeding 90 %, thereby showcasing their robustness compared to other methods.