Sistema de monitoramento de focos de incêndio com rede neural embarcada e comunicação sem fio

Abstract

Environmental wildfires represent a growing threat, demanding early detection methods that overcome the cost and latency limitations of traditional solutions. This work proposes the development of an embedded system for real-time fire spot monitoring. The device is based on the ESP32 microcontroller integrated with the MLX90640 thermographic camera, utilizing Edge Al for local image processing. The methodology included the creation of a specific thermal dataset and the training of a highly optimized Convolutional Neural Network (CNN), consisting of only 5,921 parameters. To enable execution on constrained hardware, the model underwent an 8-bit integer quantization process via TensorFlow Lite, ensuring deterministic inferences and low Flash and RAM memory consumption. Training results demonstrated a classification accuracy of 91.38%, proving the system's effectiveness in distinguishing active heat sources from neutral environments. The system integrates local inference with Wi-Fi connectivity, sending alerts and processed thermal images to a remote monitoring server via the HTTP POST protocol. It is concluded that the proposed solution is technically viable and demonstrates the potential of TinyML to democratize access to environmental surveillance technologies and rapid disaster response.