Machine learning method for leak detection in water mains using acoustic data

Odeh, Francis Onuoha (2022) Machine learning method for leak detection in water mains using acoustic data. Masters thesis, Memorial University of Newfoundland.

[English] PDF - Accepted Version Available under License - The author retains copyright ownership and moral rights in this thesis. Neither the thesis nor substantial extracts from it may be printed or otherwise reproduced without the author's permission. Download (818kB)

Abstract

The global demand for potable water rises each year as a result of the continuous increase in the global population. Hence, there is a need to expand water distribution networks and maintain existing ones continually. However, the current state of the existing water distribution is a significant source of concern for many municipal water agencies since they are more prone to leakages attributable to age, corrosion, external loads, excessive water pressure, etc. These leakages, when they occur, usually go unnoticed for a while due to poor labor-intensive reactive leak detection techniques employed by many municipal water agencies. The delayed detection of any leakages often results in economic losses to water agencies, collateral damage to nearby infrastructures, and health issues due to the percolation of harmful substances through the broken water pipes. To this end, researchers have developed several leak detection methods such as infrared radiometric pipeline testing, fiber-optic detectors, acoustic emission detectors, and monitoring for fluid pressure, flow, and temperature. However, each technique has its limitations. The acoustic emission method stands out due to its higher sensitivity, leak detection accuracy, more extended detection range, and ease of use. Nonetheless, the results obtained when using the acoustic emission method in the field are still subject to human interpretation and, as such, pose challenges concerning the final decision to excavate for the leakages or not, especially at instances when the detection metrics are close to the leak thresholds. In such instances, the delay in decision-making or misinterpretation of the results due to human error could negatively impact early detection, remediation, and the volume of water loss. Consequently, as a case study, this research applied the machine learning technique to the real-life acoustic data obtained from the field investigations carried out by the city of Mount Pearl municipality to develop models for detecting leakages within the water distribution network. The developed models would analyze and interpret the results of the acoustic emission method, thereby tackling the challenges mentioned above. The Python programming language was employed for the data preprocessing and the development of binary classification models to identify leaky and non-leaky pipes within the network. Four classification models for leak detection prediction were built: the K Nearest Neighbor (KNN), Random Forest (RF), Artificial Neutral Network (ANN), and the Support Vector Machine (SVM). The performance evaluation showed that the RF model is the best model for leakage detection using acoustic data. The RF model has the lowest error rate, highest accuracy, precision, F1-score, specificity, recall, and Area Under the Receiver Operating Characteristic Curve (AUC-ROC).

Actions (login required)

View Item