DEEP LEARNING ENABLED APPROACH FOR NONLINEAR RESPONSE MODELING OF PIEZOELECTRIC ENERGY HARVESTERS

Piezoelectric energy harvesting has been extensively investigated to provide a sustainable power supply for small-scale devices. Though significant progress has been made in the past two decades, accurate modeling of piezoelectric energy harvesters (PEHs) still remains difficult or usually involves parametric identification from experimental results. This paper explores and leverages deep learning techniques for modeling behaviors of nonlinear PEHs. To this end, we present a deep learning-based surrogate model for nonlinear response analysis of PEHs with high complexity and strong nonlinearity. The key concept is to establish a long short-term memory (LSTM) neural network to capture the dynamic properties of the piezoelectric energy harvesters from limited training data and infer the voltage and power given unseen excitations in a data-driven fashion without the need of conventional time-consuming numerical simulation or parametric identification. In addition, the input and output sequences are preprocessed into stacked structures to reduce the complexity and enhance the feature learning. The performance of the proposed approach was successfully demonstrated through numerical data of a monostable PEH of Duffing stiffness nonlinearity and piezoelectric material nonlinearity. The results indicate that the proposed framework is a promising, reliable and computationally efficient approach for nonlinear response modeling of complex piezoelectric energy harvesters.