Hybrid Quantum Convolutional Neural Network for Photovoltaic Defect Classification Using Electroluminescence Images

Abstract

As photovoltaic (PV) systems become more widespread, they need inspection methods that are both reliable and automated. Near-infrared electroluminescence imaging can reveal internal cell defects that visual inspection may miss. Many conventional CNN-based classifiers, however, use deep backbones with large numbers of trainable parameters and often require substantial labelled datasets. This has encouraged interest in more compact hybrid architectures. Here, we evaluate a Hybrid Quantum Convolutional Neural Network for classifying PV defects in electroluminescence images. The model replaces the first convolutional stage of a classical CNN with a quantum convolutional layer based on angle encoding, cyclic entanglement, parameterized rotations and Pauli-Z measurement. We tested it on a five-class subset of 409 images from the PVEL-AD dataset and compared it with CNN-based models including ResNet, DenseNet, GoogLeNet, AlexNet and CNN-LSTM. The subset was chosen to keep quantum-circuit simulation computationally manageable. With approximately 35K trainable parameters, the Hybrid QCNN reached 75.8% accuracy in this feasibility experiment. It used far fewer parameters than several classical models, but its deployment is constrained by the latency introduced by quantum-circuit simulation. The results support the feasibility of the approach, while further work is needed before it can support real-time monitoring.