Data fusion of electrical signals andelectroluminescence images forphotovoltaic degradation classification

Sammanfattning

The growing number of renewable energy distribution via photovoltaic panelsystems has increased exponentially at large scale over the recent decade. Areliable detection system of photovoltaic panels degradation is essential to theeconomic viability. The photovoltaic (PV) systems in a distributed energy farmsnetworks require reliable and automated fault detection mechanism to maintainefficiency and reduce maintenance costs. Usually, fault on photovoltaic such ascell cracks, hotspots, delamination, soiling, and degradation can significantlyimpact energy yield and long-term system stability of energy production.Fault diagnosis of PV remains difficult because many faults manifest differentlyacross thermal images, electroluminescence images, and electrical operatingdata. This thesis investigates an asynchronous late-fusion edge-AI frameworkfor PV fault detection using open-source data. The thesis project aims to useopen-source dataset of electrical signal and electroluminescence images.However, it is restricted using heterogeneous open-source datasets that are notmatching at the sample level.Several literature reviews, confirm that thermal imaging, EL inspection,current voltage (I–V curve) analysis, and lightweight embedded inference haveall matured significantly, yet lack of heterogeneous open-source datasets,annotation scarcity, and edge resource limitation remain main barriers. Thisthesis proposed developing a multimodal deep learning framework that fuseselectroluminescence (EL) imaging and current–voltage (IV) curve analysis usingthe OEDI PV-IV-EL dataset (Sandia National Laboratories, 616 paired measurements across 438 unique modules). The proposed asynchronous late-fusion used an EfficientNet-B0 backbone for spatial feature extraction from EL images, coupled with a multi-layer perceptron (MLP) encoder for IV scalarfeatures, with late-fusion concatenation which results in a 1408-dimensionaljoint representation. The EL images contribute complementary spatialinformation crack morphology, and the IV scalar features complements withelectrical signals, justifying the multimodal design.The proposed system achieves 84.78% overall accuracy and 87.01% balancedaccuracy with a Cohen's kappa of 0.7735 which implies a strong agreement.Critically, all 23 Severe-class degradation modules are correctly identified(recall = 1.000), the most safety-relevant outcome for field deployment. Themean per-module Pmp degradation rate is −0.883 %/year, with 43.4% ofmeasured modules exceeding the IEC 61215 warranty threshold of 0.8 %/year.