A Two‐Stage Deep Learning Method for Photovoltaic Fault Diagnosis via I‐V Curves and Vision Combination

ABSTRACT To address the limitations of fault diagnosis in photovoltaic (PV) systems, this paper proposes a two‐stage hierarchical diagnostic framework based on current‐voltage (I‐V) and computer vision (CV) combination, and cloud–edge collaborative task allocation. A hybrid 2D+1D convolutional neural network (CNN) first extracts joint features from I‐V curves for efficient string‐level screening, followed by an adapted ResNet for component‐level fault classification from drone imagery. Stage 1 performs string‐level fault detection, and Stage 2 performs component‐level fault type classification. Rigorous mathematical formulations address the ill‐posed mapping from string‐level signals to component‐level faults. Cloud–edge collaborative inference balances real‐time performance and resource efficiency. Experimental results demonstrate that the proposed framework achieves 96.7% overall end‐to‐end diagnosis accuracy, while reducing 71.4% drone flight coverage and 89.0% average diagnosis time compared to full‐coverage CV inspection, offering a scalable solution for intelligent PV operation and maintenance (O&M).