REAL-TIME DEFECT DIAGNOSIS OF PHOTOVOLTAIC MODULES BASED ON CONVOLUTIONAL NEURAL NETWORKS AND NANO-OPTICAL SURFACE CHARACTERIZATION

The consistency and power effectiveness of photovoltaic (PV) parts are highly dependent on the timely detection and diagnosis of surface and structural defects. Challenges, such as limited diversity, environmental sensitivity, hardware constraints, and difficulties in achieving real-time, large-scale deployment, affect accurate defect detection. The objective is to establish a real-time defect diagnosis framework for PV modules by integrating convolutional neural networks (CNNs) with nano-optical surface characterization (NOSC) for high-resolution feature extraction and rapid defect identification. Nano-optical surface data were collected from PV modules, preprocessed using spectral normalization (SN), non-local means denoising (NLMD), and adaptive contrast enhancement (ACE), and subsequently transformed into spatial–spectral feature representations through a CNN. The proposed model employs a mosquito flying optimized attention-based conditional variational autoencoder (MFO-ACoVAE) for PV defect classification, where the ACoVAE learns latent defect-specific representations through attention, and the MFO tunes model parameters to enhance convergence and classification accuracy. The trained CNN with NOSC [Formula: see text] MFO-ACoVAE model was implemented in Python and deployed on an embedded GPU, achieving real-time inference in 0.28[Formula: see text]s. The proposed MF-ACoVAE achieved 98.7% accuracy, 98.7% recall, 98.5% precision, 92.45% F1-score, a classification accuracy of 98.4%, and ROC-AUC values [Formula: see text], showing superior robustness and efficiency in PV defect detection. The proposed system enables accurate, real-time, non-destructive PV defect detection, supporting predictive maintenance and enhancing module reliability.