Session: 04-01: Mechanics of Smart Structure Applications

Paper Number: 110216

110216 - Artificial Intelligence for Active Vibration Control Optimization on Smart Structures 

New metamaterials are currently being developed with the objective of energy harvesting [Sezer and Ko ̧c, 2021], structural damping improvement, and energy diffusion optimization in mechanical structures [Collet et al., 2011]. They mainly use piezoelectric transducers within a network. Many control methods exist for vibration mitigation such as PID, optimal control, or other nonlinear controllers. This study uses the classical Proportional Integrator Derivative (PID) control law [Borase et al., 2020] as a basis.

The present manuscript considers Artificial Intelligence (AI) algorithms as a method for the optimization of smart structures vibrations controllers with piezoelectric transducers. As a preliminary work in this optimization with AI, an automatic algorithm using Machine Learning (ML) approach intends to tune a Proportional/Derivative controller on an experimental structure. Thus,
two colocalized piezoelectric transducers are used to minimize vibrations on a cantilever beam. The control law is a feedback between the two piezoelectric transducers colocalized close to the clamped end. Since it does not need any database, Reinforcement Learning (RL) approach is used to obtain an optimal experimental controlled behavior. This method allows feeding the algorithm with the final objective without providing prior knowledge about the influence of each of the controller parameters. Indeed, it learns by an iterative process to reach the goal represented by the maximum value of a defined reward function [Sutton and Barto, 1992]. The input disturbance on the piezoelectric actuator is a band-limited white noise (10-1000Hz) and is unknown for the controller. The reward function depends on the displacement variation measured at the beam-free end using a laser. For the algorithm, the aim is to minimize this displacement. Due to time consumption, the RL algorithm trains offline on an estimated model of the experimental setup. Two measured transfer functions define the input of the AI: TFP between the piezoelectric actuator and sensor and TFL between the piezoelectric actuator and the laser. This study investigates Policy based algorithm with stochastic gradient descent: TRPO (True Region Policy Optimisation) Reinforcement Learning algorithm developed by Schulmann [Schulman et al., 2015] researches a possible solution in a domain close to a previous stable domain. It increases the safe domain of solution research at each learning step. With the accurate hyper-parameters setup, this method allows staying in the control stability region. The algorithm policy evolves until it reaches the goal of vibration minimization or until it achieves the number of steps defined within the hyper-parameters. Finally, a few training runs allow to obtain an average representative result and an overview of the method’s efficiency.

After AI training, the study compares control law tuning between Reinforcement Learning results and classical approaches. The experimental setup allows the characterization of the two tuned control laws’ performances.

References

[Borase et al., 2020] Borase, R. P., Maghade, D. K., Sondkar, S. Y., and Pawar, S. N. (2020). A review of PID control, tuning methods and applications. International Journal of Dynamics and Control, 9(2):818–827.

[Collet et al., 2011] Collet, M., Ouisse, M., Cunefare, K. A., Ruzzene, M., Beck, B., Airoldi, L., and Casadei, F. (2011). Vibroacoustic energy diffusion optimization in beams and plates by means of distributed shunted piezoelectric patches. pages 265–302.

[Schulman et al., 2015] Schulman, J., Levine, S., Abbeel, P., Jordan, M., and Moritz, P. (2015). Trust region policy optimization. In Bach, F. and Blei, D., editors, Proceedings of the 32nd International Conference on Machine Learning, volume 37 of Proceedings of Machine Learning Research, pages 1889–1897, Lille, France. PMLR.

[Sezer and Ko ̧c, 2021] Sezer, N. and Ko ̧c, M. (2021). A comprehensive review on the state-of-the-art of piezoelectric energy harvesting. Nano Energy, 80:105567.

[Sutton and Barto, 1992] Sutton, R. S. and Barto, A. G. (1992). Reinforcement Learning:An Introduction.

Presenting Author: Maryne Febvre LTDS, Univ Lyon, CNRS, Ecole Centrale de Lyon, ENTPE, UMR5513 & Univ Lyon, INSA Lyon, CNRS, LaMCoS, UMR5259

Presenting Author Biography: Maryne Febvre received an M.Sc in Mechanical Engineering with an orientation in Mechatronics and Systems from the National Institute of Applied Science in Lyon (INSA Lyon, France) and an M.Sc in Acoustic and Vibration in Ecole Centrale Lyon (France) both in 2021. Since 2021, she is doing a Ph.D. in Control on Smart Structures with piezoelectric transducers using Artificial Intelligence at INSA Lyon in LaMCoS (France) with Simon Chesne and Jonathan Rodriguez and at Ecole Centrale Lyon in LTDS (France) with Manuel Collet.