The lecture was delivered by Dr. Hussein Momtaz, a faculty member in the department on Monday, 20/3/2023, in which he addressed the existence of many non-destructive testing techniques used in geophysical surveys to obtain information about the area buried under the surface of the earth. Ground penetrating radar is one such important tool that is used in a variety of applications, especially in detecting objects buried underground. The basic method of the device is to transmit electromagnetic waves and display the received signal in the form of data of either one-dimensional, two-dimensional or three-dimensional.
This study aims to use intelligent algorithms to identify the buried object through an image representing the scanning area. This is achieved using four stages. The first phase consists of building two practical models to evaluate and determine the characteristics of the radar device and the factors required to operate and determine the practical frequencies required in order to build a simulation model with high accuracy in system representation. The second stage is the creation of a simulation-based model to use the ground scanning radar device to detect buried objects, and the input data for the deep learning algorithm was obtained through the use of the simulation system, and this data is divided into four groups according to the type of buried material, the type of medium buried in it, and the depth and radius of the object. The first group is divided into four sections, the second into five, the third into three, and the fourth into three. In the third phase, twenty pre-trained deep learning networks were developed using learning transfer technology to automatically classify inputs and select the most suitable network for each of the groups. Hence finding the best network result from the average results of all networks for all groups, where the results showed that the best result with an accuracy rate of 84.16% and a loss ratio of 0.581.
The fourth phase of the project is the method of improving the results of networks from the previous phase and the results on these networks are improved using three types of optimizers. In addition to three types of change factors, the results of the confusion matrix and the training progress curve are used to determine the best performing network for each of the four output categories and then choose the result of the best network from the sum of the networks.