IMPROVING THE GENERALIZATION PERFORMANCE OF WATER LEAKAGE DETECTION MODEL BY NOISE REDUCTION USING PSEUDO SOUNDS

Abstract

 This study conducted an experiment focusing on improving the water leakage detection model's generalization performance (the capability to recognize sounds as leakage or not) through the use of AI. Convolutional neural networks were used to create a leak detection model using actual leak sounds collected at 10 sites for each of ductile cast iron pipes (metal pipes) and rigid polyvinyl chloride pipes (non-metal pipes). Recurrence plots were used to develop the model and apply to their testing. In the data preprocessing, band pass filtering and noise reduction process, utilizing actual field pseudo sound data such as transformer and sewer flow sounds, were applied. The generalization performance is then evaluated if it has improved. Through the experiment, it was made clear that the accuracy increased for the low accuracy points by applying both processes. Furthermore, it was confirmed that this pre-processing approach, in addition to the leak detection model, is superior to a leak surveyor's examination of the acoustic data.