Jakub Lengiewicz

Abstract: There is growing evidence that piezoelectric energy harvesters (PEHs) not only generate electricity from vibration sources, but their voltage signals can also be utilised simultaneously to sense various contexts of interest. Little is known, however, about any potential tradeoff between its energy harvesting and sensing performances within the design space of the PEH device. The lack of such knowledge prevents the optimal design of dual-functional PEH devices that are expected to be used both as a power source and a sensing system. This paper presents a deep-learning-based framework to infer a passing vehicle’s speed from the generated voltage signal by a PEH installed under the bridge. The proposed methodology utilises AlexNet, one of the most effective Convolutional Neural Networks (CNNs) widely recognised for image classification. Representative images are generated from the voltage signal’s time-frequency analysis and fed into the AlexNet network to extract traffic-dependent information. Continuous Wavelet Transformation (CWT) is employed to create these images, which has shown promising results in our context. Moreover, the framework uses Transfer Learning, a technique that reduces the size of required databases. Our results lead us to propose a framework for jointly optimising energy harvesting and sensing for PEH devices. The multi-objective optimisation framework considers using the Kriging surrogate models and a genetic algorithm (GA), which improves the computation efficiency of the process. A comprehensive optimisation study case is conducted in a real-world setting.

Patricio Peralta-Braz is a Ph.D. candidate in the department of Civil Engineering at UNSW, working on modelling of piezoelectric energy harvesters.