Optimizing Takagi-Sugeno Fuzzy Models For Improving Leak Detection in Water Distribution Networks

Leakage detection in water distribution networks (WDNs) is critical for reducing water loss and ensuring operational efficiency. While machine learning methods are often applied, they can lack interpretability. Takagi-Sugeno (Tsk) fuzzy systems offer a balance between accuracy and interpretability but are prone to overfitting and incur high computational costs, especially with large datasets. To address these issues, we explore various optimization and regularization techniques to improve Tsk performance. The models were trained on a large-scale benchmark dataset containing 1000 leak scenarios, each a year-long time series at half-hour intervals, totaling over 17 million data points. A systematic preprocessing pipeline was applied, including time-series segmentation, mutual information-based feature selection, and class imbalance handling. Alongside the baseline training of Adaptive Neuro-Fuzzy Inference Systems (ANFIS) using gradient descent (GD), we also trained Tsk models using stochastic gradient descent (SGD) and mini-batch gradient descent (MBGD). State-of-the-art regularization techniques such as uniform regularization and rule dropout were also incorporated to prevent overfitting. Key results show that SGD and MBGD models outperformed GD models in leak detection rates and achieved significantly lower false alarm rates than traditional machine learning models. These findings underscore the potential of fuzzy systems for effective leak detection, provided that appropriate learning and regularization techniques are employed.