Multi-sensor estimation algorithm for the friction coefficient and the braking torque of carbon braking systems for racing motorcycles

This paper investigates new methods for estimating the braking torque generated by carbon brakes mounted on MotoGPTM class motorcycles. A physical model of the brake was originally implemented in order to calculate the torque applied to the motorcycle front wheel using the measured brake fluid pressure and the braking system geometry as inputs. The results obtained with this method are currently limited by the usage of a constant friction coefficient in the equations of the model. Since it is extremely hard to estimate the friction coefficient using a classical model-based approach, it was deemed more convenient to develop a machine learning algorithm capable of identifying the friction coefficient in any operating condition of the brake. To this purpose, different machine learning models were evaluated and compared, paying particular attention to the dataset generation and to how the labels used as targets during training were calculated. An algorithm combining a Nonlinear Autoregressive Exogenous model (NARX) and an Artificial Neural Network (ANN) was finally selected and trained. The developed algorithm was successfully tested on several datasets collected during experimental tests on different tracks, providing satisfactory results in terms of braking torque estimation and friction coefficient identification.