• deep learning
• fpga
• motor control
• neural network
• reinforcement learning

Artificial Intelligence (AI) and machine learning algorithms are spreading to a wide variety of applications across industry. This work focuses on the matter of electrical motor control. The objective was to design a neural-network-based controller that controls the rotational speed of a simulated DC motor. Five controller architectures were designed in the MATLAB/Simulink environment using different AI techniques. The performance of the proposed controllers was measured and compared with a traditional PI controller. The comparison was carried out using a custom benchmark simulation and a set of performance metrics. The best overall controller was based on a multilayer feed-forward neural network, trained using Reinforcement Learning (RL). The RL-based controller was then implemented on FPGA. A data representation analysis of the selected controller was performed first. Both floating-point and fixed-point approximations of the controller were studied and their performance measured. The final design was downloaded onto an FPGA. The HDL code for the design was generated with two different methods: using HDL Coder and DSP Builder. The usage of FPGA resources for the proposed controller was compared with the reference PI. Finally, the entire motor-controller system was emulated in real-time on FPGA.