• deep learning
• fault detection
• thermography
• photovoltaic solar panels
• yolov3
• convolutional neural network
• object detection

In the last decade, the photovoltaic market has grown impressively and, with it, the size of the solar plants. Monitoring large systems is expensive and also hard for the operators. Due to these reasons, this thesis proposes an approach that allows the identification of photovoltaic panels' anomalies by analyzing images captured with Unmanned Aerial Vehicle (UAV). The model is a Multi-Stage architecture composed of an anomaly detector, a panel detector, and a False Alarm Filter removing hotspot detections corresponding to false positives outside the detected panel area. In this way, the model, tested on real cases using thermographic and visible images, is more robust than a single anomalies detector.