Leak-detection in water pipeline with Machine learning: A case study with Oras Intelligent Valve
Nguyen, Minh (2023)
Nguyen, Minh
2023
Master's Programme in Information Technology
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. Only for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2023-02-08
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202301271808
https://urn.fi/URN:NBN:fi:tuni-202301271808
Tiivistelmä
Leakage in pipelines is a rising concern due to its financial impact on water utility companies from the water losses. To this end, researchers proposed a multitude of techniques to detect leaks ranging from model-based, transient-based, to data-driven methods. This thesis aims to develop a machine-learning model to determine whether there are leaks in the pipeline. The empirical data for training the model was collected in the water laboratory of Oras’s factory, where a pilot pipeline was calibrated to simulate different leaking scenarios. Another aim was to prove that the advantage of the machine-learning approach can compensate for its disadvantages. From this, a comparison with the modeling approach can be drawn to argue whether it is more feasible to develop a machine-learning-based leak detection system in practice.
Most leak-detection studies in the past have been developed based on transient analysis, which compares the behavior of turbulent flow in the pipeline using a modeling tool and compares with the measured results from the monitored pipe system. This technique depends on extensive knowledge of the pipeline to approximate the fluid flow, and its accuracy is sensitive to measurement uncertainties of the modeling tool. On the contrary, the machine-learning approach does not require an in-depth understanding of the water distribution system (WDS). Instead, it utilizes the abundant availability of various data that can be collected from smart sensing devices.
In this work, an intelligent solenoid valve, with Silab’s EFR32MG22 controller as the processor, was utilized to collect pressure and flow rate measured from the pilot pipeline. Those measurements were transferred via Bluetooth SPP service to a desktop computer for filtering noise and feature-extraction. In model selection, an SVM model is chosen for the binary classification of leak vs. non-leak data. This model is optimized by implementing several pre-processing techniques, i.e., data scalers and transformers, and cross-validation of the model’s hyperparameters.
From comparing the results of models trained with data that was scaled using different transformers, which evaluate the accuracy, AUC, model’s fitting time, and prediction time, the power transform using Yeo-Johnson has the best performance (97% correct prediction on unseen data). The result of this thesis was also compared to previous related works, where we can observe that it has a competitive accuracy. Hence, this case study demonstrates a practical proof-of-concept where machine learning can be implemented for leak detection.
Most leak-detection studies in the past have been developed based on transient analysis, which compares the behavior of turbulent flow in the pipeline using a modeling tool and compares with the measured results from the monitored pipe system. This technique depends on extensive knowledge of the pipeline to approximate the fluid flow, and its accuracy is sensitive to measurement uncertainties of the modeling tool. On the contrary, the machine-learning approach does not require an in-depth understanding of the water distribution system (WDS). Instead, it utilizes the abundant availability of various data that can be collected from smart sensing devices.
In this work, an intelligent solenoid valve, with Silab’s EFR32MG22 controller as the processor, was utilized to collect pressure and flow rate measured from the pilot pipeline. Those measurements were transferred via Bluetooth SPP service to a desktop computer for filtering noise and feature-extraction. In model selection, an SVM model is chosen for the binary classification of leak vs. non-leak data. This model is optimized by implementing several pre-processing techniques, i.e., data scalers and transformers, and cross-validation of the model’s hyperparameters.
From comparing the results of models trained with data that was scaled using different transformers, which evaluate the accuracy, AUC, model’s fitting time, and prediction time, the power transform using Yeo-Johnson has the best performance (97% correct prediction on unseen data). The result of this thesis was also compared to previous related works, where we can observe that it has a competitive accuracy. Hence, this case study demonstrates a practical proof-of-concept where machine learning can be implemented for leak detection.