Supervised Learning for Abrupt Change Detection in a Driven Eccentric Wheel

Abstract

Event detection is often a predominant challenge in processing non-stationary signals. In engineering mechanics, events may result from non-smoothness in the form of loss of contact, impact, or the onset of sliding-friction. An interesting example of such a mechanical system is a wheel whose center of mass does not coincide with its geometric center. An eccentric wheel may evolve in three distinct phases: roll without slip, roll with slip, and hop. Therefore, this paper seeks to explore and compare supervised learning methods for phase identification (i.e., roll, slip, and hop) in simulated data from a driven eccentric wheel. The mechanics of a torque driven wheel on a flat surface are derived through an augmented Lagrangian formulation and Coulomb friction is adopted to model transverse contact forces. To accommodate for non-smoothness, the system is broken down in complementary sub-problems and the simulation is conducted using event-based methods. The simulated data is then used to train a Naive Bayes classifier, a Support Vector Machine (SVM), and an Extreme Gradient Boosting (XGBoost) classifier. Lastly, the methods as well as their performance, merits, and drawbacks are discussed in detail.