Design and optimization of low-frequency piezoelectric MEMS energy harvesters based on machine learning

Ehsani Chimeh, Hamidreza (2022) Design and optimization of low-frequency piezoelectric MEMS energy harvesters based on machine learning. Masters thesis, Memorial University of Newfoundland.

[English] PDF - Accepted Version Available under License - The author retains copyright ownership and moral rights in this thesis. Neither the thesis nor substantial extracts from it may be printed or otherwise reproduced without the author's permission. Download (2MB)

Abstract

Portable electronic applications are typically powered by batteries, which have limited lifespan and size constraints. Energy harvesting from parasitic vibrations using piezoelectricity is a demanding solution to improving the power supply efficiency of low-power portable devices and autonomous sensor networks. Vibrational energy harvesters with high operating frequency are not desirable considering the low-frequency characteristics of ambient vibrations (e.g., wind). The operation frequency range (known as bandwidth) is a key feature that should be improved under an unpredictable or uncontrollable condition of ambient vibrations. In this thesis, several piezoelectric MEMS energy harvesters have been developed to tackle these challenges. In order to facilitate the design process and determine the optimum physical dimensions, an artificial neural network is used to model the designs. In the first step, a sample dataset is created by numerical modeling to train a deep neural network. The validation results indicate that the trained DNN model can achieve around 90% estimation accuracy of device features, such as resonant frequency and harvested voltage. Next, this trained model is integrated with genetic algorithm as a performance estimator to optimize the geometry of the harvester to lower the resonant frequencies and improve the harvested voltage. With the intention of improving the accuracy of deep neural network, the transfer learning method is used for modeling another design. In this method, the DNN is firstly trained with a dataset created by the lump-parameter model; then, the trained network is transferred to a new deep model for another round of training with highly accurate FEM data samples to further reduce prediction error. It was shown that the new model can estimate the device features with more than 94% accuracy, which is 4% higher than the regular DNN. Finally, in the last design, our proposed AI-based methodology is utilized to estimate the mode shapes to enlarge the operational bandwidth specifically. The optimized energy harvesters have been fabricated through a standard micromachining process. Our measurement results confirm that the proposed AI-based methods can help reach the balanced summit of higher power density, lower resonant frequency, and larger bandwidth among the published works.

Actions (login required)

View Item