Abstract
Concrete structures are susceptible to damage over their lifespan and cracks are often the earliest noticeable form of damage. Crack width is an indicator of the severity of crack damage. Unaddressed cracks can develop to failures potentially resulting in injury, death and huge economic costs. Traditional methods such as visual inspection are subjective, tedious and prone to human error. Recently image-based methods have been employed, typically relying on image processing techniques for damage detection. These methods are also hindered by several limitations such complex manual parameter selection and restriction to measuring crack width in pixels.To overcome the drawbacks of existing methods, this project aimed to develop an advanced deep learning image-based method for crack detection and measurement. The developed method overcomes limitations of existing methods by eliminating the need for manual parameter selection, overcoming subjectivity, susceptibility to human error and tedious inspections. Furthermore, crack width was measured in millimetres using a novel laser calibration method. The project made the following contributions to knowledge. The scarcity of datasets for training deep learning models in crack detection is mitigated by creation and curation through a series of data acquisition experiments. Highly accurate deep learning models for crack classification and segmentation are developed, trained on curated datasets, which outperform traditional image processing methods. A novel method for measuring crack width in millimetres is proposed, providing more insightful damage assessment in practical applications. A framework for applying deep learning models in image-based crack inspection that aids their integration into current structural health monitoring standards and guidelines is developed.
The project not only demonstrates the potential of deep learning in enhancing the reliability and efficiency of structural health monitoring, but it also lays a solid foundation for future research to extend the developed principles and models to a broader range of damage types and materials.
| Date of Award | 2024 |
|---|---|
| Original language | English |
| Supervisor | Jiping Bai (Supervisor) & Ian Wilson (Supervisor) |