Sensorless PMSM control system based on multi-agent reinforcement learning and fractional order LADRC for rapid control prototyping

The improvement of the performance of Permanent Magnet Synchronous Motor (PMSM) control systems is a challenge from many points of view, including the fact that that its mathematical model should also consider the parametric variations of the main modelled elements. In this context, the present work proposes an improved control structure based on Linear Active Disturbance Rejection Control (LADRC) controllers, but maintaining the cascade control structure of the Field Oriented Control (FOC) control strategy. The LADRC controller typically consists of an Extended State Observer (ESO), a Tracking Differentiator (TD), and a Proportional Differentiator (PD) control law so that the closed-loop behaviour is equivalent to an integrator. Considering the peculiarities regarding the execution speed of the control loops in the cascade structure, but also the requirement that the algorithms implementing the observers run faster than the algorithms implementing the controller, the paper proposes to synthesize a Fractional Order (FO) LADRC in the outer control loop, which implies that both the control law and the ESO-type observer are of fractional type, namely FOPD and FOESO. This approach could improve the overall performance of the control system through the additional tuning parameters. A Computational Intelligence Optimization (CIO) algorithm, the Slime Mold Algorithm (SMA), is used to optimize the tuning parameters of the controllers and observers. Another proposed improvement consists in the use of a Multi-Agent Reinforcement Learning (MARL) that includes three Twin-Delayed Deep Deterministic (TD3) agents. These are built, trained, and validated in both Software-in-the-Loop (SIL) and Rapid-Control-Prototyping-Hardware-in-the-Loop-Simulated (RCP-HILS) implementations to provide correction signals that are added to the control signals in the cascaded control loops. The comparative demonstration of the proposed control system starts from a benchmark system for PMSM control based on classical Proportional Integrator (PI) controllers and goes through the implementation steps in the SIL and RCP-HILS stages. Thus, we obtain prototypes for real-time implementation using different sampling periods corresponding to the switching characteristics of the power modules in the Digital Signal Processor (DSP) hardware platforms used for real-time control. The target hardware used to obtain an RCP is the SpeedGoat Performance Real-Time Target Machine.