Leakage Detection in District Heating Systems Using UAV IR Images: Comparing Convolutional Neural Network and ML Classifiers

In this paper, we proposed a method to detect leakages automatically in underground pipes of district heating networks based on images, which are captured by an Unmanned Aerial Vehicle (UAV). The original datasets are captured in a 16 bits format and later converted into an 8 bit format using Dynamic Range Reduction (DRR). Leakages in district heating networks can occur due to unprofessional installation, lack of maintenance or end of service life, etc. We have addressed
issues of leakage detection using a deep learning based approach, Convolutional Neural Network (CNN), and 8 machine learning classifiers. The experiments are carried out on seven different datasets, which are acquired at seven different cities in Denmark. We performed our experiments on both 16 bits and 8 bits data.
For performance analysis, 6 datasets are used for training and the remaining dataset for testing. Our proposed deep learning CNN achieves an average accuracy of 0.886 and 0.884 for 16 bits and 8 bits, respectively. Machine learning classifiers such as Adaboost (AB), Random Forest (RF) etc provide relatively lower average accuracy. Adaboost required less computational resources, achieves average accuracies of 0.800 and 0.793 for 16 bits and 8 bits, respectively