Real-time Estimation of Brake Temperature using Physics-Informed Machine Learning Model

The current work presents a novel approach to estimate brake surface temperature in real time to aid in brake wear prognostics. Brake prognostics involve estimating the brake pad wear in real-time that helps in its predictive maintenance. Brakes are a safety critical system for vehicles, therefore, demands accurate and robust pad wear estimation, ensuring vehicle safety. However, it involves several challenges. The estimation of pad wear is fundamentally a two-stage process: the first stage involves the accurate prediction of brake pad surface temperature, while the second stage utilizes this thermal history to calculate cumulative material wear. A significant challenge in estimating brake pad wear without expensive sensor is that it is sensitive to the surface temperature prediction; any error in the thermal model propagates and compounds in the wear prediction stage. To identify surface temperature, traditional physical sensors are often cost-prohibitive or prone to failure in the harsh thermal and mechanical environments of the wheel end, necessitating a robust virtual sensing solution that can capture complex, non-linear heat transfer dynamics. The current work addresses the above challenge of identifying temperature dynamics using a physics-informed machine learning (PhyML) approach. We employ Symbolic Regression, a data-driven method that discovers the underlying mathematical equations of a system by searching for the optimal functional relationship between variables. Symbolic regression provides an interpretable model that generalizes across most automotive platforms, providing a transparent and analytically tractable alternative to traditional "black box" models. For generating the temperature dataset, two test vehicles were equipped with thermal sensors and were undergone different braking scenarios. Results show good prediction accuracy from the model in estimating brake surface temperature at different braking conditions, from mild to hard braking. With mild braking, the SR model showed 95% accuracy in temperature prediction, whereas for harsh braking, it was around 89%. The resulting high-fidelity temperature estimation significantly reduces the error rate in the brake pad wear prediction, enabling reliable, sensor-less cloud based brake health monitoring and enhancing the efficiency of software-defined vehicle maintenance.