Project no. 407
Detecting Leakages and Anomalies in District Heating
Systems Using Thermal Drone Footage
F. Villebro, S. Forchhammer, and C. Mantel
DTU Fotonik, Technical University of Denmark
District heating is an efficient and popular way of providing heat to urban households. Heat
is generated at a centralized district heating plant after which it is transferred to the end user
through underground pipelines in the form of hot water or steam. Bad insulation or leakages
in the underground pipe system are hard to detect and causes economic and environmental
losses. Pinpointing where these anomalies appear is a tremendous task as pipelines can
cover vast areas. One solution is to equip pipelines with sensors, however depending on the
type there can be substantial initial investment and installation can be cumbersome. This
project aims to solve this problem in a non-intrusive way by using thermal drone footage and
machine learning in order to detect these anomalies.
Bad insulation or leakages in district heating pipelines causes their surroundings to heat up
and thus they are visible with a thermal camera. In combination with a aerial drone large
areas of the distribution pipelines can be investigated. Adding machine learning algorithms
on top of it can hopefully result in full automation with minimal human intervention.
In particular this project focuses on the data analysis part of the proposed solution. Data
gathering have been carried out by Drone Systems which have supplied their thermal drone
footage for data analysis. The supplied data consists of 640x512 resolution thermal images
taken at a height of around 100 meters above ground along with GPS data of each flight.
Two different models have been constructed. First of all a model for anomaly detection
without any prior knowledge about pipe location was implemented by investigating each
image of a flight individually and extracting hot areas for which features where calculated. A
principal component analysis was carried out and results fed into multiple different classifier
algorithms such as AdaBoost, Random Forest, Support Vector Machines and etc.
The second model includes prior knowledge of pipe location and combines images from one
flight to form a orthophoto. Overlaying a pipe mask hot areas are extracted and features
calculated which are fed into the same classifier algorithms.
Results are still work in progress, however they look somewhat promising.