Load Classification Using Machine Learning on NanoEdge AI Studio for STM32 Microcontroller

Lezama J.; Romero H.; Sayhua A. · TEMSCON LATAM 2025 - Technology and Engineering Management Society Conference (2025) · DOI: 10.1109/temsconlatam65810.2025.11238594

Resumen

This paper introduces an embedded system designed for the accurate classification of domestic electrical loads through current signature analysis. The system leverages a machine learning model based on the SVM algorithm, implemented on an STM32 Nucleo U575ZI microcontroller using NanoEdge AI Studio. By acquiring current signals via the DMA module and employing Fast Fourier Transform for feature extraction, the system effectively distinguishes between seven distinct domestic loads: iron, vacuum cleaner, kettle, printer, fan, blender, and a 'no load' state. This solution achieves a remarkable 99.88% precision while maintaining exceptional resource efficiency, requiring only 2.8 KB of RAM and 2.9 KB of Flash memory. Furthermore, rapid classification is achieved with an execution time of just 0.7 ms, making it ideally suited for real-time applications. The fusion of lightweight machine learning and efficient signal processing demonstrates a practical and effective approach to intelligent energy monitoring in smart homes. These results underscore the potential of embedded systems for accurate and low-cost load classification, contributing to the advancement of edge computing by showcasing the feasibility of deploying sophisticated machine learning models on resource-constrained devices without sacrificing performance or accuracy.

Instituciones: Universidad Nacional Tecnológica de Lima Sur