Deep reinforcement learning for optimizing control law parameters on a smart beam with distributed piezoelectric transducers

Abstract Recent research has focused on the design of adaptive structures with distributed transducers for efficient vibration mitigation. This study explored the optimization of a control strategy to enhance vibration mitigation in structures with multiple piezoelectric transducers. A key challenge lies in tuning control parameters while maintaining stability and robustness against environmental variations.
Artificial intelligence, particularly neural networks, offers a promising approach for handling complex multi-input, multi-output control problems. In this work, Deep Reinforcement Learning (DRL) is applied to autonomously tune Positive Position Feedback (PPF) controllers for a smart beam equipped with three pairs of collocated piezoelectric transducers. The training process prioritizes minimizing vibrations of the third bending mode, which is one of the most dominant. A pole-zero model, derived from experimental measurements, enables numerical learning process while preventing potential damage to the test bench.
A comparative analysis is conducted between the DRL strategy and a simplex optimization method, both using the same objective function. A statistical analysis evaluates the reliability and effectiveness of both approaches, highlighting their advantages and limitations in achieving robust control.