Leak Detection and Classification of Water Pipeline Data Using LSTM Auto-Encoder with Xavier Initialization

Modern cities operate extensive water supply networks. In Korea, the water supply network coverage reaches 98.9%, but more than 10% of water is still leaking every year. Previously, leaks in water pipelines were detected using flow meters and pressure meters, but these methods could only detect leaks after they occurred, necessitating on-site investigations to accurately pinpoint the location of the leak. In the United States, aging water and sewage facilities incur economic losses of hundreds of billions of dollars every year. To address this issue, this paper proposes an LSTM (Long Short-Term Memory) Auto-Encoder model, enhanced with Xavier Initialization, that can detect leaks using water pipeline data and automatically identify the location and extent of the leakage. This model achieved 93.7% performance by applying the Xavier Initialization technique to the LSTM Auto-Encoder model and achieved relatively high accuracy compared to the model using only LSTM Auto-Encoder. The results of this paper can serve as an effective predictive tool for monitoring water pipeline leaks.