Tomac, Beau (2022) Development of an Improved Concrete Defect Object Detection Method for Edge AI Inference Technology. [USQ Project]
|
Text (Project)
Tomac_B_Nguyen_Redacted.pdf Download (6MB) | Preview |
Abstract
This research analyses the development of a neural network object detection model for use with advanced edge AI inference technology capable of detecting alkali silica reaction (ASR) defects on concrete surface utilising ‘You Only Look Once’ (YOLO) algorithm. The dissertation encompasses research for a novel digital tool intended to aid structural health inspection processes of ASR defects with the ultimate intent of using it on edge AI inference technology – i.e. NVIDIA Jetson Nano.
The aim of this research is to provide a novel tool capable of recognising ASR defects on concrete surface. The need is evident when examining critical structural failures occurring on large structures. The project builds on the transfer learning capabilities of YOLO and the ability to train detection of any common structural defect with a distinct pattern of occurrence. The existing neural network models focus on methods such as acoustic emission techniques or the data analysis from semi-quantitative microscopic petrographic tools.
The methodology focuses on identifying visual ASR criteria, the training dataset and the neural network model development. Specific ASR visual criteria are ascertained through literature review. The training dataset was procured through UniSQ, TMR and printed and online media. The model is subsequently trained in YOLOv2, YOLOv3 and YOLOv4 and the results evaluated, compared and the models improved. The key outcome is the development of an ASR defect detection model using a small training dataset capable of detecting concrete surface ASR defects with sound confidence.
Based on the results thus far the key conclusions are the sound performance of the neural network model YOLOv2 (mean average precision of 60.33%) trained on 433 images capable of detecting ASR defects and regular cracks. Further research can be done on improving model performance using a small dataset. Due to the time and resource constraints, the use of edge AI inference technology to implement the developed YOLO model onto the Jetson Nano will have to be done at a later stage. Further research can also be done on improved training and data augmentation techniques to create high accuracy models trained on small training datasets.
|
Statistics for this ePrint Item |
| Item Type: | USQ Project |
|---|---|
| Item Status: | Live Archive |
| Faculty/School / Institute/Centre: | Current – Faculty of Health, Engineering and Sciences - School of Engineering (1 Jan 2022 -) |
| Supervisors: | Nguyen, Andy; Brown, Jason |
| Qualification: | Bachelor of Engineering (Honours) (Civil) |
| Date Deposited: | 19 Jun 2023 01:39 |
| Last Modified: | 20 Jun 2023 01:05 |
| Uncontrolled Keywords: | Machine Learning, Alkali Silica Reaction Detection, Concrete |
| URI: | https://sear.unisq.edu.au/id/eprint/51864 |
Actions (login required)
 |
Archive Repository Staff Only |
