Open Collections

Rayhana, Rakiba

University of British Columbia

Water transmission networks are a community's most expensive and critical urban assets. Management of these pipelines is becoming one of the significant concerns as some of the water pipelines are at the end of their service lifespan. Nowadays, robotic platforms equipped with cameras and a wheeled system are used to videotape the internal state of the water pipelines. The acquired videos then require certified human personnel to assess the conditions of water pipelines manually. However, manual assessment is often error-prone, challenging, and expensive. Additionally, while inspecting the pipeline, the robotic platform experiences difficulties to pass through the valves installed inside the pipeline. This disrupts the inspection process and adds labor to a human operator. Hence, this research aims to mitigate the aforementioned issues of the inspection process by developing automated video data analytical frameworks to enhance inspection operations. At first, an extensive literature review was conducted, showing that automated frameworks display better performance when developed using deep learning-based techniques. Therefore, a deep learning-based defect characterization framework and a user-friendly defect characterization interface are developed to simultaneously detect, classify and segment the defects of water pipelines. The framework is developed by employing three residual networks (ResNet), a channel-spatial attention mechanism (CSA), Mask Canny-Regional Convolution Neural Network (MaskC-RCNN), and an ensemble mechanism. The results from this experiment depicted that the proposed model improved the defect detection rate and saved time for human operators. Then, a deep learning-based valve detection framework is developed to detect the valves by combining MobileNet-160 and Feature Pyramid Network (FPN). The experimental results and analysis shows that the developed algorithm detected the valves efficiently and accurately. Lastly, a real-time and super-lightweight valve detection algorithm with an edge device (NVIDIA Jetson TX2) is developed by combining MobileNetv3 and YOLOv4. The results from the proposed approach show a satisfactory performance and pave the door for its integration with the robotic platform. The research outcomes of this thesis contribute to the inspection process of the condition assessment of the water pipelines. The proposed algorithms can significantly improve the inspection process and aid in the decision-making process of pipe replacement/renewal plans.

Text

2023-01-05

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/83594

Electrical Engineering

University of British Columbia

UBCO

http://creativecommons.org/licenses/by-nc-nd/4.0/