Utilizing Various Deep Neural Network Technologies to Evaluate FC-CSRR Sensor Performance In Biomedical Applications

Abstract

This paper introduces a four-concentric-ring complementary split ring resonator (CSRR) sensor, designed on a Rogers 3210 substrate, targeting a frequency of 5.2 GHz. To optimise sensor performance, three deep neural network (DNN) models are developed and compared: Single-Step FrequencyMagnitude Model (S-FM), which consists of S-M (referring to only the magnitude output of the Single-Step Model) and S-F (referring to only the frequency output of the Single-Step Model), and Double-Step Frequency-Magnitude Model (D-F) and (D-M). The D-F model, featuring advanced architecture for frequency prediction, achieved the highest accuracy of 99.5% in testing with a training time of $\mathbf{1. 0 5}$ minutes. As for, the D-M model, offering the shortest training time of 0.74 minutes, maintained reasonable accuracies of $\mathbf{9 1. 1 \%}$ (testing), making it suitable for faster evaluations. While S-FM trained in one minute, model S_M shows a balanced performance of $\mathbf{9 1. 6 \%}$, and the S-F model shows lower accuracy 88.2% testing, fitting for less precision-critical applications. Our innovative proposed DNN methods imply that machine learning can be used to evaluate electromagnetic sensor performance for noninvasive blood glucose measurement.