Physics-Informed Machine Learning Digital Twin of Electro-Mechanical Brakes for Clamp-Force and Friction Estimation

2026-01-0806

To be published on 09/14/2026

Authors
Abstract
Content
As the industry is rapidly moving towards dry-by-wire architectures, accurate Electro-Mechanical Brake (EMB) models are increasingly critical for safety-relevant force control, diagnostics, and consistent performance over temperature and lining wear. However, widely used EMB identification methods often rely on linearized models or 'characteristic curve' calibrations that struggle to generalize across operating regions where friction, hysteresis, and compliance are strongly non-linear and temperature sensitive. Prior EMB force-estimation literatures addressed pad contact detection, hysteresis and thermal compensated clamp-force estimation, yet practical accuracy gaps remain during transients and under changing friction signatures. This work proposes a digital twin based on Physics-Informed Neural Networks (PINNs) that preserves the EMB's governing dynamics while learning both (i) physically meaningful parameters- such as effective system inertia, damping, stiffness etc., and (ii) a residual non-linear friction term constrained as a function of actuator motion states and temperature. This hybrid formulation maintains the physical structure of actuator-caliper dynamics, while the learned residual captures un-modelled effects such as temperature dependent friction and hysteresis that are poorly represented in linear baseline models. The model is trained using EMB test-stand measurements collected under step, ramp, chirp and sine-sweep excitations across multiple temperatures, using time-aligned force commands and measured signals such as motor torque/current, actuator positon/velocity, and pad-force from force sensor. Validation is performed in both time and frequency domains, comparing force-tracking and actuator-displacement transients as well as bandwidth, resonant features, and phase-behavior. Results show reduced pad-force prediction error and lower frequency-response phase mismatch compared to linear baseline. The resulting digital twin enables sensor-less force estimation, friction compensation design, and health monitoring.
Meta TagsDetails
Citation
Rai, P. and Gadhvi, T., "Physics-Informed Machine Learning Digital Twin of Electro-Mechanical Brakes for Clamp-Force and Friction Estimation," Brake Colloquium & Exhibition - 44th Annual, Palm Desert, California, United States, September 20, 2026, .
Additional Details
Publisher
Published
To be published on Sep 14, 2026
Product Code
2026-01-0806
Content Type
Technical Paper
Language
English