FPGA Realization of a Neural Network based Motor Controller.

FPGAs are a hardware solution for applications that require low power usage and real time execution. This work focuses on motor control. We show the design and FPGA implementation of a motor controller based on a small Neural Network (NN) as alternative to a traditional Proportional Integral (PI) controller that was used as reference. Performance metrics are defined from a simulation of the target motor. We investigate different NNs, training, code generation methods and numerical precisions. The best-performing controller is a multilayer NN trained with Reinforced Learning. The controller is implemented and runs on an Intel MAX 10 FPGA. The methodology described in this work is easily applicable to more complex designs when FPGAs become the best implementation platform.