While unmanned aerial vehicles (UAVs) with thermal cameras are established tools for inspecting photovoltaic (PV) plants, most systems rely on offline or cloud-based data processing, which introduces latency and data transmission costs. This study presents a novel contribution by developing and validating a fully autonomous inspection system that performs real-time, onboard fault detection using embedded computer vision.

The core innovation lies in the implementation of a lightweight convolutional neural network (CNN) on a low-cost ESP32 microcontroller, a highly resource-constrained environment. This enables the system to analyze thermal images captured by the UAV and identify anomalies—such as hotspots and burned cells—directly in the field, eliminating the need for external processing infrastructure. The CNN model was trained on a custom dataset of annotated thermographic images, using data augmentation and class balancing to ensure robust performance.

Preliminary results demonstrate that this embedded approach enables accurate, autonomous, and non-invasive thermal inspections without requiring system shutdowns. By integrating georeferenced aerial data acquisition with immediate onboard analysis, our system significantly reduces operational costs and diagnostic latency compared to traditional methods. This work adds to the existing knowledge by proving the viability of deploying computer vision models on edge devices for industrial inspection, offering a scalable, efficient, and accessible architecture that advances intelligent predictive maintenance in the renewable energy sector.