A Comparative Classification Study on the Use of Acoustic Emission Signals for Surface Roughness Condition Monitoring in End Milling of Stainless Steel

Surface roughness quality has implications on the functionally, assembly, service life, and appearance of the machined product. Considering the complex nature of metal cutting processes, computer simulation models may not provide the needed accuracy to predict surface conditions under all cutting conditions. Therefore, machine learning (ML) techniques can provide more reliable predication models that are based on real time cutting process sensory data. The implementation of artificial intelligence (AI) techniques in the monitoring of manufacturing processes has been gaining momentum. The focus of this study is to predict the surface roughness using acoustic emissions (AE) signals during the dry end milling of stainless steel. AE sensors have been widely used to monitor the condition of structures and manufacturing processes. Furthermore, acoustic sensors are non-invasive and can be used at any location without disrupting or stopping the machining process. Features extracted from the AE signals are used as surface roughness quality indicators. These features include frequency bands averaged amplitudes, statistical quantities of the wavelet decompositions, raw signal RMS values, and crest factor. In this work, several machine learning algorithms are used to process the extracted AE features for surface roughness characterization. The total AE features are first processed for feature set reduction since many of the features are highly correlated. This is done using both supervised and unsupervised feature reduction and subset selection methods. The features extracted from supervised feature reduction methods are used to train three supervised classifiers — k-nearest neighbor (kNN) classifier, a radial-basis function support vector machine (RBF-SVM), and a random forest (RF) classifier. The reduced feature set from the unsupervised feature reduction methods are used as input to two unsupervised clustering methods — K-Means and DBSCAN. The classifier models are trained using multi-fold cross-validated mix of subsets of the reduced features. In this study we have used ten models using two-fold cross validation for training and validation for the supervised learning methods. The results of supervised classification are compared to unsupervised clustering and are reported for an average of the ten models (or ten runs with distinct initializations of the clustering algorithm), along with a detailed nonparametric testing to verify statistical significance in performance level between pairs of algorithms.