Chatter Detection and Identification Based on Mode Decompositions

Abstract

To meet the demand for industrial production of calendering or rolling, it is necessary to detect the onset of chatter before chatter marks appear on the workpiece. Therefore, mode decomposition techniques (empirical, bivariate empirical, and variational) combined with machine learning (ML) are used to detect impending failures. Signals from acceleration sensors are decomposed into a discrete number of modes, isolating the high-frequency oscillations by identifying local minima and maxima. Feature sets (peak to peak, standard deviation, etc.) of true intrinsic mode functions are extracted for training an ML model to detect the vibration states followed by the prediction of chatter marks. This innovative prediction model based on mode decompositions and ML shows its feasibility for early chatter identification.