Multimodel Neural Network for Live Classification of Water Pipe Leaks From Vibro-Acoustic Signals

Abstract

A novel neural network model able to identify the nature of leaks in a live urban underground water pipe grid is proposed in this article. Traditional leak detection methods are often limited to detecting the existence or otherwise of a possible leak in the signals, yet do not provide information about the actual nature of the leak. This is of utmost importance for utilities to be able to rank and schedule their repair and rehabilitation efforts. The proposed model, termed multiclass time-frequency convolutional neural network (M-TFCNN), uses a multimodel categorization approach, enabling the identification of various types of leak-related events, namely, pipe breaks, hydrant leaks, service valve leaks, private leaks, and others. The proposed neural network architecture consists of a feature extraction module, a convolutional base, and a classification head, which has been meticulously engineered through numerous experiments and iterations. The model was trained on a large-scale dataset of vibroacoustic signals collected from an operating water main network in a major Australian city. The data encompass a substantive signal collection of validated representations of different types of leaks gathered over a span of three and a half years. The model's performance is evaluated using traditional metrics such as accuracy, precision, recall, and F-1-score. The presented results show that the proposed model surpasses conventional leak detection methods, accurately identifying different types of water leaks and achieving accuracies of up to 98\%. Overall, the proposed neural network model represents a notable practical step forward in the field of water leak detection by subcategorizing leaks and has the potential to revolutionize the way industry practitioners manage larger water infrastructure.