Interpretable Water Leakage Detection Using Federated Prototype -Based Learning

Título: Interpretable Water Leakage Detection Using Federated Prototype -Based Learning

Autores: Diego Perdigão Sousa, Polycarpo Souza Neto, José Mairton Barros da Silva Júnior, Charles Casimiro Cavalcante & Carlo Fischione

Resumo: This work presents an interpretable and privacy-aware solution for leakage detection in water distribution networks using federated prototype-based learning. Real data from pumping stations in Stockholm expose a non-independent and identically distributed data scenario, where each client reflects distinct operational conditions. Despite data heterogeneity, the model achieves consistently high performance. Interpretability is achieved via Voronoi-based prototypes, while privacy emerges from client-specific decision boundaries. The approach shows that combining federated learning and prototype-based models enables scalable, explainable, and secure anomaly detection in critical infrastructure.

Palavras-chave: federated learning; leakage detection; prototype-based models; water distribution network.

Páginas: 8

Código DOI: 10.21528/CBIC2025-1189943

Artigo em PDF: CBIC_2025_paper1189943.pdf

Arquivo BibTeX:
CBIC_2025_1189943.bib