Analysis of Deep Learning and Tree Based Machine Learning Models in Leak Detection using LeakDB Benchmark

Résumé fourni par la source

Physical leakage accounts for a substantial share of the treated water lost from distribution networks, reaching approximately 44% in Turkey, yet conventional detection approaches are limited either by the impracticality of manual inspection at scale or by the cost of dense sensor deployment. This study examines whether machine learning applied to the pressure and flow data already recorded by SCADA systems can detect leaks with sufficient accuracy for operational use, and which class of model is best suited to such data. A sliding window of 48 time steps, corresponding to one diurnal demand cycle, was applied to the LeakDB Hanoi benchmark, which comprises 1,000 randomized one-year scenarios across 66 pressure and flow channels. Each window was reduced to a 466-dimensional feature vector formed from seven per-channel statistics together with cyclic encodings of time of day and day of week. Four algorithms received this representation and were evaluated across six classification metrics: Random Forest, XGBoost, LSTM, and 1D-CNN. Random Forest performed best, achieving 90.00% accuracy, an AUC-ROC of 0.9507, 77.59% recall, and 93.71% specificity, followed by XGBoost (82.49% accuracy, AUC-ROC 0.8904). Both deep learning models discriminated poorly, with AUC-ROC values of 0.5426 for the LSTM and 0.5110 for the 1D-CNN. The comparable degradation of a recurrent and a convolutional architecture indicates that the limiting factor is the pre-aggregated feature representation rather than any individual network design. Existing SCADA infrastructure, combined with statistical feature extraction and a tree-based ensemble, therefore offers a low-cost route to operational leak detection.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.