COST-OPTIMIZED UNCERTAINTY-BASED DECISION FRAMEWORK FOR AUTOMATED PHOTOVOLTAIC DEFECT DETECTION USING MONTE CARLO DROPOUT

Abstract

Solar photovoltaic (PV) energy systems are increasingly contributing to the ex pectations of global renewable energy infrastructure. However, to ensure their long term operational reliability requires accurate and timely detection of faults and operational defects of manufacturing. Deep learning models such as convolutional neural networks (CNNs), have showcased high classification performance on electro luminescence (EL) of PV cells. However, a crucial gap still remains: these models generate only point predictions without any degree of confidence or reliability, which makes them unreliable for high-stakes real-world deployment. The purpose of this thesis is to describe a two-stage research pipeline designed to fill this void. During stage one, a baseline CNN model was created to classify defects in EL images using a binary scheme on the standardized ELPV benchmark data set containing 2,624 grayscale EL images. The accuracy rate obtained by this baseline CNN model was 72.84%, but the primary drawback lies in its failure to provide any measure of uncertainty, thereby making it unreliable for autonomous solar panel maintenance tasks. In the second phase, this limitation is directly addressed through a novel cost optimized, uncertainty-aware decision framework. Monte Carlo (MC) Dropout is applied to the same CNN architecture — enabling the same dropout layers used for regularization during training to serve as a Bayesian approximation mechanism during inference. By performing 100 stochastic forward passes per input, a per sample uncertainty estimate is computed. A two-zone decision framework then routes predictions: high-confidence samples are accepted automatically, while un certain samples are forwarded to low-cost manual expert review. The proposed framework is evaluated against a realistic operational cost model incorporating false positive costs ($100), false negative costs ($800), and manual inspection costs ($20). Experimental results on the ELPV dataset demonstrate an iv 88.3% reduction in total operational cost compared to full automation ($63,200 → $7,400), while achieving 100% accuracy on all automated decisions and eliminating all false positives and false negatives in the automated stream. Furthermore, MC Dropout uncertainty is confirmed as a statistically reliable error indicator, with incorrect predictions exhibiting 27.2% higher uncertainty than correct ones.