Energetic challenge for an embedded system based on ESP32: case of a Moroccan remote water meter reading

The management of water resources in developing nations, particularly in water-stressed countries like Morocco, is severely hindered by the reliance on manual water meter reading. This traditional approach leads to billing inaccuracies, high operational costs, and delayed leak detection, exacerbating the issue of Non-Revenue Water. While the deployment of fully integrated smart water meters offers a technical solution, the capital expenditure required for mass infrastructure replacement is prohibitive for local municipalities. To bridge this gap, this paper presents the design, implementation, and empirical energy profiling of an ultra-low-power, ESP32-S3-based Internet of Things (IoT) retrofit system capable of digitizing existing mechanical water meters via camera-based Convolutional Neural Network (CNN) digit recognition. To overcome the fundamental energy bottlenecks associated with visual edge AI, we propose an asymmetric master/slave computing architecture. The battery-operated slave nodes, mounted directly on the meters, are equipped with OV2640 cameras and utilize aggressive deep-sleep power management, drawing an idle current of merely 15 µA. These nodes are awoken via a Real-Time Clock (RTC) to execute a highly optimized measurement cycle—wake-up, image capture, preprocessing, and local transmission—in under 10 seconds before immediately returning to sleep. The centralized master unit aggregates data from up to 500 local slave nodes and handles the energy-intensive cellular cloud connectivity. Extensive field testing across twelve distinct mechanical meter types currently deployed in Morocco demonstrated a highly robust digit recognition accuracy of 98.70%. More critically, our empirical energy profiling reveals that this optimized ESP32 architecture reduces power consumption by 99.2% when compared to standard Raspberry Pi-based edge implementations. This efficiency extends the operational autonomy of a standard 3000 mAh battery from a mere 46 hours to over 150 days. By proving that low-cost microcontrollers can successfully execute visual AI workloads within strict power budgets, this retrofit solution provides resource-constrained utilities with a financially viable pathway to real-time water monitoring. Published via ChairConf (chairconf.com) — A Nexus Academia platform (nexus-academia.com)