DEVELOPING OF WATER LEAKAGE DISCRIMINATION MODEL USING RECURRENCE PLOT AND CONVOLUTIONAL NEURAL NETWORK

Abstract

 Maintenance of aging infrastructure is an important issue in Japan, which is facing a declining population. In particular, since water pipes that serve as lifelines are buried underground, it is impossible to directly check the state of deterioration, and it is extremely difficult to detect water leaks from underground water pipes at an early stage. In order to overcome these issues, the development of health monitoring technology for infrastructure has been promoted recently, and the application of maintenance methods using high-sensitivity sensors and AI technology has been developed. In this study, we focused on the difference of deterministic properties of time series data of water leak sound, and visualized sound information on a two-dimensional plane using recurrence plot (RP). Then, we develop a leak discrimination model based on a convolutional neural network (CNN) using the visualized images. As a result of comparing the accuracy of water leak discrimination with the CNN model, an average accuracy rate of more than 90% was obtained, confirming the effectiveness of the proposed method combining RP and CNN.