Date of Award

12-5-2018

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Systems Engineering

First Advisor

Mariofanna Milanova

Second Advisor

Haider Raad

Abstract

Nowadays, flexible devices have witnessed a great deal of interest in both industrial and academia. Consistently, flexible electronic systems require the integration of a conformal and flexible antenna operating in specific frequency bands to provide wireless connectivity which is highly demanded in today’s information-oriented society. Designing of microstrip antenna has become the focus of modern technologies due to their light weight, ease of fabrication, and ability to integrate electronics devices. Simple structure such has square patch, circular patch, or rectangular patch are easy to maintain using empirical equations which based on some experimental results. However, more complicated structures are hard to produce the design equation to synthesis the structure under consideration. Therefore, using EM simulators is a must to solve such type of structures. There are several well-known EM software used in both academia and industry to solve complicated design numerically such as High Frequency Structure Simulator (HFSS), CST Microwave Studio, Advanced Design System (ADS), etc. The use of such type of software require expensive license, large computing cluster, and time to solve the structure which varies from couple of minutes to several days in some cases. Due to the aforementioned drawbacks of using EM simulators, Artificial Neural Network (ANN) has become one of the candidate machine learning approaches for the antenna design for its capabilities in learning and generalization and ability to solve complex and non-linear problems. ANN is capable to implement and solve antenna and microwave devices without using EM simulator. ANN curve fitting and regression techniques are used to capture the trend of the data collected by EM simulators. The ANN sets the weight of the nodes and neural lines based on the assigned training algorithm and optimize the network based on the given dataset. The performance of the learning algorithm depends on the problem under consideration hence the complexity of the problem and the capability of the algorithm should be considered. The major leak on the previous research of using ANN for antenna design that there is no defined methodology to design ANN antenna models and it is ambiguate to the designer to how to collect valid dataset to train the network. In this dissertation, we present for the first time the design methodology of antennas using ANN. The methodology starts with how to collect valid dataset to train the model and how to prepare the dataset to avoid overfitting. In addition, applications of flexible antennas design using ANN is conducted for single band and Ultra-wide band antenna. The two designs have been simulated using EM tools and ANN model to compare the performance metric. Four prototypes are built using the ANN models associated with Levenberg Marquadt algorithm and measured the S-parameters and the radiation patterns. The simulated and measured results are in good agreement. Therefore, the methodology presented in this dissertation on how to collect valid dataset to build ANN model for flexible antennas designs is validated based on the measured data. This methodology could be used for future flexible antenna design without the problem of requiring large computing cluster, high cost of license, and undetermined simulation time.

Download

Share

COinS