Clutch Judder Classification and Prediction: A Multivariate Statistical Analysis Based on Torque Signals

Abstract

Judder is the term used in the automotive industry to describe the longitudinal oscillation in a vehicle during its clutch system engagement. Past research has shown that judder can be explained using a behavior of slip speed and temperature captured by the clutch torque. This paper proposes and implements an innovative learning system for better characterization of the judder phenomenon. It is based on a multivariate data-driven analysis from torque signals. Our experimental results have been carried out using the following main resources: dry clutch system, passenger car, test bench, and six different organic facing materials. The multivariate statistical analysis implemented has allowed the development of a computationally efficient and highly accurate learning model to discriminate the torque signals from different facings, using few features and a regularized version of a standard linear classifier. Given this multivariate framework and calculating the correlation pairwisely to a known gold material, it has also been possible to predict judder problem in the vehicle based on a standard test bench in laboratory. We believe that the findings of this paper might reduce significantly the time of development and the cost of testing new friction materials for allowing judder-free performance on vehicles.