Investigation Of Dimensionality Reduction Methods For Chatter Detection With SVM

Abstract

Chatter vibrations significantly affect the quality and efficiency of machining operations. Machine learning algorithms for intelligent chatter detection are viable options when there is sensor data available at the machine tool level. Most machine learning methods require a feature engineering phase where the most valuable data should be extracted and prepared as input for a machine learning classifier. The selection of the proper dimensionality reduction method at this early stage enhances the performance of the classifier. This study aims to investigate the effectiveness of several dimensionality reduction methods when using Support Vector Machine (SVM) as a classifier. Vibration signals collected during slot milling are binary labeled as stable (0) and chatter (1). Signals were reshaped to 0.5second segments and 0.1-second segments. Ten-dimensional (10D) statistical time-domain features extracted from signals were reduced to three-dimensional (3D) feature space with Principal Component Analysis (PCA), t-Distributed Stochastic Neighbor Embedding (t-SNE), and Autoencoder (AE) dimensionality reduction methods. Signals were classified by SVM classification with various training distributions. The effectiveness of different dimensionality reduction techniques and different training distributions were compared for chatter detection. Furthermore, it was observed that dimensionally reduced features were classified quicker and more accurately than statistical time-domain features.