TY - JOUR
T1 - Application of Machine Learning for Leak Localization in Water Supply Networks
AU - Yussif, Abdul Mugis
AU - Sadeghi, Haleh
AU - Zayed, Tarek
N1 - Funding Information:
This research was funded by the Innovation and Technology Fund [Innovation and Technology Support Programme (ITSP)] and the Water Supplies Department of Hong Kong under the grant number ITS/067/19FP.
Publisher Copyright:
© 2023 by the authors.
PY - 2023/4
Y1 - 2023/4
N2 - Water distribution networks (WDNs) in urban areas are predominantly underground for seamless freshwater transmission. As a result, monitoring their health is often complicated, requiring expensive equipment and methodologies. This study proposes a low-cost approach to locating leakages in WDNs in an urban setting, leveraging acoustic signal behavior and machine learning. An inexpensive noise logger was used to collect acoustic signals from the water mains. The signals underwent empirical mode decomposition, feature extraction, and denoising to separate pure leak signals from background noises. Two regression machine learning algorithms, support vector machines (SVM) and ensemble k-nearest neighbors (k-NN), were then employed to predict the leak’s location using the features as input. The SVM achieved a validation accuracy of 82.50%, while the k-NN achieved 83.75%. Since the study proposes using single noise loggers, classification k-NN and decision trees (DTs) were used to predict the leak’s direction. The k-NN performed better than the DT, with a validation accuracy of 97.50%, while the latter achieved 78.75%. The models are able to predict leak locations in water mains in urban settings, as the study was conducted in a similar setting.
AB - Water distribution networks (WDNs) in urban areas are predominantly underground for seamless freshwater transmission. As a result, monitoring their health is often complicated, requiring expensive equipment and methodologies. This study proposes a low-cost approach to locating leakages in WDNs in an urban setting, leveraging acoustic signal behavior and machine learning. An inexpensive noise logger was used to collect acoustic signals from the water mains. The signals underwent empirical mode decomposition, feature extraction, and denoising to separate pure leak signals from background noises. Two regression machine learning algorithms, support vector machines (SVM) and ensemble k-nearest neighbors (k-NN), were then employed to predict the leak’s location using the features as input. The SVM achieved a validation accuracy of 82.50%, while the k-NN achieved 83.75%. Since the study proposes using single noise loggers, classification k-NN and decision trees (DTs) were used to predict the leak’s direction. The k-NN performed better than the DT, with a validation accuracy of 97.50%, while the latter achieved 78.75%. The models are able to predict leak locations in water mains in urban settings, as the study was conducted in a similar setting.
KW - acoustic sensors
KW - leak localization
KW - machine learning
KW - noise loggers
KW - water distribution networks
UR - http://www.scopus.com/inward/record.url?scp=85156108737&partnerID=8YFLogxK
U2 - 10.3390/buildings13040849
DO - 10.3390/buildings13040849
M3 - Journal article
AN - SCOPUS:85156108737
SN - 2075-5309
VL - 13
JO - Buildings
JF - Buildings
IS - 4
M1 - 849
ER -