XVIII IWRA World Water Congress Beijing China 2023
Sub-theme 2: Promoting Water Efficiency, Productivity and Services
Author(s): Ziyang Xua, Haixing Liu, Guangtao Fu, Yukai Zeng, Yunchen Li
Ziyang Xua, Haixing Liua,* , Guangtao Fub, Yukai Zenga, Yunchen Lia
a School of Hydraulic Engineering, Dalian University of Technology, Dalian, Liaoning, 116024, China
b Centre for Water Systems, University of Exeter, Exeter, EX4 4QF, UK
Poster: PDFAbstractIn our paper, we introduce MDMR_ISFFS for feature selection and evaluate its performance in conjunction with five classifiers (DT, RF, XGBoost, SVM, and MLP) for real water distribution system leak detection. Here are the key conclusions:
- High Accuracy: All five classification models, when combined with our feature selection method, achieve impressive leak detection accuracies ranging from approximately 94% to 98%.
- Key Features: We identify four key features essential for leak detection. These are F1 (Mean of frequency), F5 (Peak frequency), T1 (Mean), and T14 (Zero-crossing rate). Each of these features is selected by at least four classifiers.
- Feature Interaction Analysis: Using the SHAP method, we analyze the interaction mechanism of these key features. High values of F1, T14, and F5 positively influence the predicted leakage probability (indicated by positive SHAP values), while a high value of T1 has a negative impact (negative SHAP value). Additionally, F5 and F1 have the strongest interactions with features F1 and T14, respectively, and T14 exhibits the strongest interactions with features F5 and T1.
- Efficient Feature Selection: Our proposed feature selection method converges quickly to achieve high accuracy with a smaller number of features compared to other methods.