Tiny Reservoir Computing for Extreme Learning of Motor Control

The electromechanical system of a typical electric machine controller, usually composed of a motor, logic, and power electronics, represents a complex non-linear system that needs to adapt itself continuously over time. Factors as component aging and thermal derating tend to modify the motor controller system’s behavior and, consequently, its performances. Therefore algorithms based on Artificial Intelligence are expected to improve the motor’s drive due to Neural Networks’ ability to approximate such complex non-linear systems. In particular, because of the time-varying nature of signals in a controller, Recurrent Neural Networks (RNN) are required to elaborate time series. This paper proposes a novel control technique which extends the Field Oriented Control (FOC) algorithm using a method belonging to Extreme Learning Machine (ELM) category. In particular, it introduces a specifically developed Neural Networks (NN), named Semi Binary Deep Echo State Networks (SB-DESN), to achieve good control accuracy with a low complexity, allowing to deploy the proposed control into cheap micro-controllers, which are computing units characterized by severe memory, computational and power consumption constraints. The peculiar property of SB-DESN, such as fixed and predictable training complexity, makes the proposed technique well suited for adoption into Reinforcement Learning on micro- controllers to ensure the continuous over the time adaptability of the control. Beyond the use of SB-DESN for speed, torque, and magnetic flux control, another novel element introduced by this work is the complexity optimization based on sparsity properties of inter-Reservoir matrices. The effectiveness of the explored optimizations has been proved, reducing the memory footprint up to 370% and the computational complexity up to 200%. This corresponds to a SB-DESN, with only 2.5kB of memory footprint and an inference time of 20us on STM32H7, which allows its deployment into mass market FOC solution.