Power Maximization of Vibration Piezoelectric Energy harvester Using Artificial Neural Network as Surrogate Model

초록

Design optimization is a method to obtain optimum design object iteratively enhancing it. Iterative approach requires burdens on measuring performances such as the number of computing the value or gradient. So, surrogate model which is approximating the original ones is employed, and surrogate model-based design optimization can lower those burdens. Kriging surrogate model that is the one of the surrogate models is widely used since it shows excellent predicting accuracy for the model with nonlinearity [1, 2]. However, kriging uses a single parameter for one design variable when approximating the original model. So, when the original model consists of flat and nonlinear (or spiky) properties, accuracy of kriging decreases since single parameter is not able to approximate both flat and nonlinear domain simultaneously [3]. This problem makes it difficult to apply kriging to design optimization of piezoelectric energy harvester (PEH). PEH is a device to generating electricity by converting mechanical energy into electrical energy. Among several mechanical energy, vibration is widely used, and in order to maximize the power produced, the harvester is designed to resonate for this vibration [4]. In this case, the resonant harvester shows very low power in the overall design domain, but high power in a specific narrow area where resonance occurs. Thus, since kriging is inappropriate to replace this original model, artificial neural network (ANN) is employed for surrogate model in this research. Unlike the kriging that approximate the nonlinearity statistically by using single parameter, ANN provides nonlinearity by activation function and improves prediction accuracy by adjusting weights of information between layers in the network. This mechanism enables ANN to approximate both flat and nonlinear properties. Therefore, ANN can be used as an alternative surrogate model to kriging.