Construction and Experimental Research on Leakage Sound Dataset of Urban Water Supply Pipeline

Waterpipe is prone to leakage, resulting in signifi-cant expenses for water-supplying companies. Hence, there is a rising interest in machine detection or monitoring to find leakage using different sensors including microphones. In deep learning community, the emergence of public datasets has done a profound influence in promoting the advancement in acoustic scene and event detection and classification. But as far as we know, there are no public datasets focusing on the normal and anomalous sound from operating pipeline with pressure. This paper presents a new dataset of sound of operating pipeline with pressure, which we call a sound dataset for recording operating pipeline with pressure (ROPP dataset), containing 4.7 hours of annotated normal and anomalous sounds together with 16.1 hours of interfered sound events which are common interference when recording in streets. Experiments with mature environmental sound classification systems showed that the dataset was suitable for developing deep models for water leakage sound recognition.