Resolution enhancement of microwave sensors using super-resolution generative adversarial network. (1st March 2023)

• Record Type: Journal Article
• Authors: Kazemi, Nazli; Musilek, Petr
• Is Part Of: Expert systems with applications. Volume 213:Part C(2023)
• Publication Date: 2023-03-01
• Subjects: SRGAN -- Microwave sensor -- Glucose -- Resolution -- Coupled CSRR; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2022.119252
• Languages: English
• ISSNs: 0957-4174

Abstract

Abstract: This article presents an approach to significantly improve the resolution of a highly-sensitive microwave planar sensor response with a super-resolution generative adversarial network (SRGAN). Three identical complementary split-ring resonators are coupled so that the sensitivity is doubled. This highly-sensitive resonator with a deep transmission zero at 4.7 GHz is deployed to measure minute variations of glucose in interstitial fluid. Measuring the sensor response with 1001 frequency-points allows differentiating 10 glucose samples within the range of 40–400 mg/dL. However, in practical readout systems with limited number of frequency-points (here 28), recognizing the deep zero in the S 21 response lacks precision. Sensor responses (magnitude vs. frequency and phase vs. frequency) are converted into equivalent 2D images (heatmaps: phase vs. frequency with colored pixels as amplitude) to be compatible as SRGAN input. As a result of 8-fold resolution enhancement using SRGAN, the classification accuracy is substantially improved from 62.1% to 93.3%. The proposed passive sensor followed by an SRGAN unit is shown to be practical as a wearable glucose monitoring sensor due to its high-sensitivity and high resolution features in a low-profile design. Highlights: Three complementary split-ring resonators coupled to double the sensitivity. Transmission zero at 4.7 GHz deployed to measure glucose in interstitial fluid. Sensor response is converted to an image for generative adversarial network input. The resolution of sensor response is enhanced by 8-fold. Glucose classification accuracy is substantially improved from 62.1% to 93.3%.