| Summary: | Data-driven methods based on samples from a supervisory control and data acquisition system have been widely applied in water-supply-network burst detection to save unexpected economic and labor costs. However, the class imbalance problem in actual on-site monitoring needs to be revised to improve the performance of data-driven methods. In this study, we proposed a domain adaptation method to generate minor-category samples (pipeline-burst samples in general) of arbitrary pipe networks utilizing theoretical hydraulic models. The proposed method transferred pipeline-burst data generated from a random water supply network with theoretical hydraulic models to an actual imbalanced dataset. Accordingly, we established a machine learning model exploring a mapping matrix between two domains for minority-category data transfer. The experimental validation first verified the effectiveness and reliability of the proposed method between two customized water supply networks in terms of their bust recognition accuracy, model parameter sensitivity and time efficiency. Then, an actual monitoring dataset from a working water supply network was used to prove the suitability and compatibility of the proposed method. A bust-point location method was also provided based on the detection results of pipeline-bursting events. The validations show the superiority of our proposed approach for the imbalance data problem in pipe burst detection.
|
|---|