Improving AI Imaging Models for a Multimodal Photovoltaic Fault Detection System

Photovoltaic (PV) systems in Smart Cities are affected by soiling, partial shading, and thermal anomalies, which reduce energy yield. This study presents the development of visual and infrared thermography (IRT) AI models for the improvement of an existing multimodal artificial intelligence platform for the detection of faults in PV modules. The IRT subsystem processes UAV-acquired images for hotspot detection at both the array and module levels. The RGB subsystem classifies modules into four categories: clean, shadowed, and two levels of soiling. The datasets include 1526 IRT array images, 549 IRT module images, and 1418 RGB images, with 6715 annotated panels. Models were trained and validated using supervised deep learning methods. The IRT module detector achieved an mAP@50 of 99.5%, with a precision of 99.1% and a recall of 99.3%. Array-level detection reached a mAP@50 of 99.1%, a precision of 96.6%, and a recall of 96.4%. The RGB subsystem obtained a mAP@50 of 98.1%, with a precision of 95.8% and a recall of 96.8%. The results contribute to the development of new solutions for PV technology in smart cities.