Feature Selection for Tool Wear Diagnosis Using Soft Computing Techniques

This paper examines feature selection methods in the context of milling machine tool wear diagnosis. Given raw sensor signals acquired during experiments, a pool of features was created through calculation by several feature extraction methods. Five techniques for selecting the most discriminating features were employed. These techniques included decision trees, neuralfuzzy methods, scatter matrix, and a crosscorrelation method. We used a diagnostic neural network to evaluate the five different feature selection schemes by comparing their classification rate and test errors.