A multi-physics informed antenna sensor model through the deep neural network regression

Abstract

A passive wireless strain sensing method using antenna sensors has significantly advanced structural health monitoring systems. Since the dimensions of antenna sensors are sensitive to their strain sensing performance and operating frequency, an iterative tuning process is required to achieve a final optimized design. Although multi-physics finite element simulation enables accurate estimation of antenna performance for each turning iteration, the simulation process requires high computational resources. Therefore, antenna tuning processes are recognized as obstacles to delay the final design process. In this study, we explore the potential of multi-physics informed models as an alternative approach for analyzing antenna sensors. Through deep neural networks, as a branch of the machine-learning algorithms, we formulate multi-physics informed models with six input parameters (antenna dimensions) and two outputs (resonance frequency and strain sensitivity). Twenty-two hundred high fidelity data sets are prepared by simulating multi-physics models: 1,600, 400, and 200 data sets are applied to deep neural network regression (DNNR) training, validating, and testing, respectively. From extensive data investigation, an optimized DNNR architecture is obtained to be two layers, with 16 neurons in each layer. Its training, validating, and testing values of mean square errors are 13.01, 44.22, 37.27, respectively. Finally, the proposed multi-physics informed model predicts the resonance frequency and strain sensitivity with errors of 0.1% and 0.07%, respectively. In addition, since the average computation speed for each tuning process is 0.007 seconds, the practical usefulness of the proposed method is also proven.