Deep Learning for Infrared-Based Photovoltaic Module Fault Classification in Drone Inspection: A Critical Review of Datasets, Models and Evaluation Practice, with a Binary-Classification Case Study

Abstract

Drone-borne infrared (IR) thermography combined with deep learning is widely
used for automated photovoltaic (PV) inspection. However, reported accuracies of
approximately 85–94% on public benchmarks provide limited evidence of reliability
under field conditions. This article critically reviews IR-based PV fault
classification and detection around three questions: what is predicted (task
granularity), from what data (resolution, provenance, and class structure), and how
success is measured (evaluation protocol).