Neural network control insights and strategies in low-cost educational platforms: A pneumatic levitation case study

This study presents a Neural Network Direct Controller (NNDC) for a pneumatic levitation system, addressing the challenge of controlling nonlinear systems without requiring a plant model. Unlike common neural control strategies that optimize the parameters of conventional controllers, the NNDC is a model-free direct controller that continuously updates its output via online training. This architecture eliminates the need for predefined control laws and offers a novel approach within low-cost educational platforms. Performance was benchmarked against a Proportional-Integral (PI) controller tuned with Cohen-Coon and a Proportional-Integral-Derivative (PID) controller using Ziegler-Nichols. The NNDC adapts parameters via error backpropagation with a computational cost 9.8 % higher than PI but 13.9 % lower than PID, making it efficient for embedded systems. The NNDC achieved up to 58 % faster settling times than PID and 52 % over PI. Overshoot analysis showed that NNDC consistently outperformed PID, reducing overshoot from 43.3 % to 13 % at key setpoints. To evaluate robustness, a mass variation test increased the ball's weight from 3 g to 10 g. Additionally, a 150 mm by 5 mm longitudinal slot in the guide tube enabled controlled, repeatable disturbances by partial obstruction, validating the controller's adaptability. This open-source system enables students to explore advanced control concepts without costly equipment. Beyond education, the NNDC's adaptability and low computational cost make it suitable for IoT-enabled and edge-computing systems requiring real-time control and suggest further exploration with recurrent neural network architectures and integration into Programmable Logic Controllers (PLCs) to enhance industrial applicability.