Application Oriented Input Design for Mass Estimation of Heavy-Duty Autonomous Mining Trucks

Sammanfattning

This thesis formulates and solves an application oriented input design problem, aimed at improving mass estimation in autonomous heavy-duty vehicles in an industrial setting. Application oriented input design is a method within system identification in which the input signal is designed to minimize the cost of the identification experiment while ensuring, with high probability, that the resulting estimation performance satisfies an application-specific degradation threshold. In this thesis, the input signal is defined as the vehicle's propulsion force. Accurate mass estimation is critical for achieving precise motion control, enhancing operational efficiency, and maintaining safety. The formulated input design problem results in a concave optimization problem. To address this, the original problem is decomposed into two constrained sub-problems: the first identifies input signal characteristics that minimize estimation variance, and the second determines the minimum experiment duration required to meet application-defined constraints. These problems are first studied analytically, excluding vehicle dynamics and velocity constraints, and then solved numerically while incorporating all relevant constraints, including a simplified dynamic model relating requested propulsion force to traction force. The optimal input signals derived from the optimization problems are subsequently analyzed and validated through both simulation and real-world experiments. Analytical results indicate that the optimal inputs resemble stochastic square waves alternating between upper and lower force bounds. Numerical results further suggest that these inputs exhibit minimal switching between the bounds, and that the resulting vehicle velocity remains close to zero. Simulation results demonstrate improved mass estimation performance compared to conventional driving. In real-world experiments, however, a bias introduced by unmodeled dynamics affected the outcome, favoring the conventional driving pattern in lower payload scenarios when the bias was not compensated for. After compensating for the bias introduced by wheel inertia, the driving pattern derived from the application oriented input design problem performed better. An important conclusion is that introducing perturbations in the longitudinal dynamics, whether through intentional variations in the propulsion force or gear shifts, positively contributes to mass estimation performance, both in terms of accuracy and the time required to reach a specified accuracy interval.