Dependable Data-based Design of Embedded Model Predictive Control

Implementations of model predictive control on resource-limited embedded devices usually do not follow theoretical schemes which guarantee stability by design because of real-time and feasibility issues or conservativeness. However, a dependable design is essential for safety-critical systems. This work proposes a dependable model predictive control system architecture, which uses an offline computed set of states declared to be dependable. A method is introduced for computing a classification function of such sets based on binary trees and convex hulls. The approach is tailored for embedded hardware limited in computing capacity and memory. Finally, the approach is applied to a realistic model of the safety-critical magnet control system of the magnetic levitation vehicle Transrapid and deployed on a microcontroller.