Drone inspection with machine learning for asset maintenance

• Main Author: Er, Tricia Yishan
• Other Authors: Law Wing-Keung, Adrian
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Engineering; Drone inspection; Machine learning; Asset maintenance
• Online Access: https://hdl.handle.net/10356/176790
• Institution: Nanyang Technological University

Inspection of niches in columbariums are important to ensure that the niches are in good condition. The current practice involves officers physically conducting visual inspections of the niches. However, this task is extremely repetitive and time consuming, causing it to be susceptible to human errors. This study seeks to investigate the use of drones and machine learning for niche inspections and evaluate the potential based on accuracy of the object detection model. First, a drone was selected and flown in laboratory conditions. Next, an object detection model was trained to identify the various classes identified. The classes included anomalies such as cracks and vandalism, which were represented by hand-drawn symbols. The object detection model was successfully trained using Roboflow after multiple rounds of trial and error using programs such as TensorFlow and Roboflow. Tests using the trained object detection model were carried out to ensure that the model was able to accurately identify the different classes. A confidence interval range of 67 - 99% was achieved using the trained model. Finally, this study also proposes to use a larger dataset to increase the confidence interval, as well as different robots, such as land-based robots, to be tested out to better determine which type of robot would be better in inspection of the niches.