Title: Drone Footage Analysis for Quality Prediction and Automatic Leakage Detection in District Heating Systems
Supervisors
Principal supervisor: Prof. Søren Forchhammer
Co-supervisor: Associate Professor Michael Stübert Berger
Evaluation Board
Associate Professor Henrik Lehrmann Christiansen (Chairman), DTU, Fotonik, Denmark
Associate Professor Anthony Larcher, University of Le Mans, France
Professor Athanassios Skodras, University of Patras, Greece
Master of the Ceremony
Assoc. Prof. Lars Staalhagen, DTU Fotonik
Abstract:
Drone Infrared (IR) image processing for energy inspection is the main focus of this thesis. District heating companies produce and supply heat in the form of hot water or steam through underground pipes. Leakages in the pipes are a growing problem, which can occur for many reasons, e.g., due to improper installation, end of service life, and even newer pipes can degenerate with times. The leakages in the pipes can lead to unnecessary cost, and can even be a threat to humans. Therefore monitoring district heating networks is great interest for the power supply company.
Thermal sensors are capable of inspecting potential leakages in the pipes since they capture the temperature dierences; consequently, potential leakages in the pipes appears as hotspots. Therefore, the fast and e‑cient way of inspecting the district heating networks is to use a thermal sensor embedded on a UAV. The proposed detection process is based on the UAV IR images.
In the setup, the Infrared (IR) image is captured by UAV, and then the IR image is compressed on-board with H.264. Then, the stream is transmitted to the onground computer for navigation and detection purposes. Thus the quality of the obtained IR image should be good enough to detect the leakages in the pipes. Due to hardware limitations, the quality estimation is to be done on-ground in a No reference (NR) manner. This thesis focuses on two dierent perspectives, i.e. a low complexity NR Video Quality Assessment or Image Quality Assessment (VQA/IQA) method, and an automatic leakage detection in pipes of district heating networks.
In this thesis, A NR VQA/IQA is developed. For low complexity and fast feedback, a Bitstream Based (BB) approach was chosen for quality assessment. In this work, a set of BB features are computed by analyzing H.264 bit stream. Optionally Pixel-Based (PB) features are also estimated from the decoded sequences of the stream. Then a feature selection method is applied, and then the selected features are mapped using Support Vector Regression (SVR) to predict Full Reference (FR) objective metrics. It should be noted here; the original UAV IR sequences are captured by a 16-bit format. Therefore, for quality prediction, the dynamic range is reduced by a Dynamic Range Reduction (DRR) operator rst. Three low complexity DRR operators are investigated in the study to explore how DRR operators inuence the performance of quality predictions.
For NR VQA/IQA, the performance is evaluated by Spearman Rank Order Correlation Coe‑cient (SROCC) for dierent objective metrics with an accuracy of up to 0.99. The proposed models are compared with state of the art methods, which proves the validity of the model.
For automatic leakage detection in the pipes, in total, twelve datasets captured by an UAV at dierent cities in Denmark are analyzed. At rst, around 13.40 million image patches were extracted (out of 243082 images) using a region extraction algorithm, on-ground. Subsequently, a set of eight conventional Machine Learning (ML) classiers as well as a deep learning-based Convolutional Neural network (CNN) are applied on the patches to classify whether or not it is a leakage. For conventional ML classiers, it is required to extract features for classication. In our case, dense SIFT and dense SURF features are extracted from each of the image patches, and then these features are fed to eight ML classiers. For CNN, the input patches are directly fed into the network for classication. For evaluation, eleven sequences are used for training and the remaining set for testing which is performed on all splits.
We found that approximately 10-50% overlap from image to image during the inspection. Therefore the same leakages have appeared in multiples images, which we marked as one unique ID. Actual leakage detection is our main target. The proposed detection method gave an actual caught rate of 98.6%. The proposed model is compared with state of the art method, and the model achieved a 3.4% higher caught rate.