University of Twente Student Theses


Using Machine Learning to detect voids in an underground pipeline using in-pipe Ground Penetrating Radar

Abhinaya, Antoinette (2021) Using Machine Learning to detect voids in an underground pipeline using in-pipe Ground Penetrating Radar.

[img] PDF
Abstract:The maintenance of underground pipelines and sewage systems is a significant challenge since they are underground. A sewer pipe installed underground is surrounded by compact sand. Due to water leakage, increase in groundwater level, or natural phenomenon, the concrete structure of the sewer might develop cracks. Due to the deterioration of the sewer surface, voids are formed in the soil around the pipe. When left unnoticed, these voids lead to a loss in the soil’s structural integrity, leading to the formation of a sinkhole. This eventually leads to the ground surface caving in,which causes a significant disturbance to the surroundings and needs time and workforce to fix. In order to avoid the formation of a sinkhole, regular inspection of void formation by non-invasive methods is required. This research is part of the Technology Innovation for Sewer Condition Assessment - Longdistance Information-system (TISCALI) project. A part of TISCALI researches the use of noninvasive methods from within the underground utilities. The most commonly used noninvasive method of the sewer inspection is the Ground Penetrating Radar (GPR). In this thesis, in-pipe GPR is used wherein the system is sent inside the pipe. This helps in detecting voids around the sewer pipeline. A high frequency of operation can be used due to the reduction in distance between the voids and the antenna. The voids and objects underground create hyperbolic reflections in the GPR radargrams. The objective of the thesis is to analyse and choose an object detection method that can be useful in void detection. The second objective is to create a model that can detect hyperbolic reflections and thereby detect the different compositions of voids. In addition to this, Image processing methods are applied to the Machine-Learning algorithms to check for improved accuracy. The GPR radargrams are simulated using the gprMax software. Voids of differing sizes and compositions are simulated around a sewer pipeline. Faster-RCNN, YOLOV2, and YOLOV3 are chosen as the three CNN methods for void detection. In comparing the detector performance of the three methods, YOLOV3 has the highest accuracy of 99.6%. In void detection, it was inferred thatWater voids have the highest accuracy in detection in all three methods due to many reflections in the radargrams. In addition to this, image processing was applied to the YOLOV3 method and was deemed useful when detecting hyperbolic reflections that were disturbed or in a heterogeneous sand environment. In conclusion, YOLOV3 is the best fit for void identification, and in addition to feature extraction by image processing, it is notable in detection. Therefore, a symbiotic relationship between the operator and the algorithm can be created by applying machine learning to current in-pipe GPR practices.
Item Type:Essay (Master)
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:52 mechanical engineering, 53 electrotechnology
Programme:Systems and Control MSc (60359)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page