A machine learning approach for maximizing direct current power of nonlinear energy harvesting systems subjected to periodic impulse excitation

Abstract

Vibration energy harvesting is a thoroughly tested approach to support the growing number of Internet-of-Things devices by providing a local and sustainable electrical power resource. Yet, the discovery of best practices for deploying vibration energy harvesters in environments is limited by inability to decipher high-dimensional data associated with design and implementation details. The study aims to devise one such approach to conclusively identify conditions under which an impulse-excited nonlinear energy harvesting system delivers peak electrical power. To accomplish this goal, physics-based and machine learning algorithms are integrated to uncover the underlying relationships between the impulse-induced nonlinear dynamics and the result converted electrical energy. The accuracy of the predictions from a machine learning model are confirmed through experiments and against cross-validation simulation data. The parameters that result in snap-through, as determined through machine learning methods, are then probed through a first-principles model to confirm that optimal circuit design conditions agree with the principle of impedance matching. The findings guide attention to previously unseen nuances of exploiting nonlinear vibration energy harvesting systems in impulse excitation environments by revealing parameter sensitivities that inhibit or promote the realization of snap-through vibration for peak power generation. This research demonstrates a successful synthesis of machine learning algorithms and first principles models to guide attention to optimal designs of nonlinear energy harvesting systems subjected to periodic impulse excitation.