Meta-Learned Uncertainty-Gated Physics-Guided Multimodal Deep Learning for Robust Photovoltaic Defect Detection and Maintenance Decision Support

Abstract

 Photovoltaic (PV) defect diagnosis under real-world operating conditions is difficult due to the lack of available sensor modalities, their corruption, or its variability in signal quality. The traditional multimodal fusion approaches typically require all the sensors to be available, while fixed-weight physics-informed models impose possibly unreliable physical priors when essential modalities are missing. To overcome these challenges, in this study, a robust PV defect detection, power loss estimation, open-set defect recognition and maintenance decision support framework for PV systems is proposed based on a multimodal deep learning approach which incorporates uncertainty gating and physics guidance, termed MetaUG-PhyMOC-PVNet. The framework consists of a multimodal fusion backbone based on a transformer architecture, which handles variable-modality inputs, and combines electroluminescence images, infrared thermography, RGB images, and current–voltage characteristics. A light weight gating network is used as an estimation of the reliability of physics informed loss using modality-availability masks and modality-specific quality indicators. A bilevel meta-reweighting strategy is used in which the multimodal backbone is trained in the inner loop and the gating parameters are learned in the outer loop on a high quality validation set, which optimizes the gating mechanism. A single-diode model is developed that is differentiable, thus establishing physical consistency among the electroluminescence intensity and local shunt resistance. Also the learned gating weight is used to control the confidence of an evidential open-set recognition head, which allows the uncertain or unseen defects to be passed for human inspection. Experimental results show that the proposed framework can obtain an mAP of 0.86 under the missing-modality condition, whereas the fixed-weight physics-informed framework and generative imputation can only reach mAPs of 0.62 and 0.78 respectively. It also decreases the prediction error of power loss to RMSE (7.1 W) and MAPE (4.3%), and the recognition accuracy to 84.3% for open set and AUROC to 0.91. The results show that adaptive physics gating enhances robustness to partially and/or corrupted sensor data, and enables a reliable prioritization of PV maintenance and human-in-the-loop decision making.