IoT-Based Intelligent Water Leak Detection System Using Transfer Learning: An Affordable Solution for Developing Countries

Abstract:

Water losses in distribution networks represent a major challenge for developing countries, particularly in Benin where they reach 30% of distributed volume. This research presents an intelligent and cost-effective leak detection system combining embedded systems, IoT (Internet of Things) and artificial intelligence. The methodology is based on a four-layer IoT architecture integrating the ADXL345 accelerometer, Arduino Uno and LoRaWAN (Long Range Wide Area Network) communication for vibro-acoustic signal acquisition. The innovation lies in applying transfer learning to adapt a CNN (Convolutional Neural Network) model trained on high-quality sensors (PCB33350) to consumer-grade MEMS (Micro-Electro-Mechanical Systems) sensors, addressing the scarcity of labeled data. The system integrates a React. js web platform with FastAPI backend. Results validate the approach’s feasibility with a prototype operating continuously and has collected 19 hours of data for model training. The model achieves 96.2% sensitivity with a 66.7% F1-Score, prioritizing detection through a conservative strategy. The platform integrates real-time monitoring, geographic location, spectral visualization and history. This research demonstrates that a solution using consumer-grade components can effectively detect leaks with acceptable performance, paving the way for modernization of water infrastructure adapted to the Beninese context.

DOI:

PDF: https://ceur-ws.org/Vol-4232/Paper2.pdf