Multi-branch spatial pyramid dynamic graph convolutional neural networks for solar defect detection

Abstract

The imperative for automating solar panel monitoring techniques has become increasingly apparent with the global expansion of photovoltaic usage and the continuous installation of large-scale photovoltaic systems. Manual or visual inspection, limited in its applicability, is insufficient to manage this growing demand. To address this, we propose a novel Multi-Branch Spatial Pyramid Dynamic Graph Convolutional Neural Network (MB SPDG-CNN) for automatic fault detection in solar photovoltaic panels. The proposed architecture utilizes two separate input branches for thermal and RGB images, effectively leveraging complementary information from both image types. This multi-branch design enables the model to extract multi-stage features through a spatial pyramid pooling layer, enhancing feature fusion and improving classification accuracy. Additionally, compared to single-branch systems, our approach prevents feature redundancy and loss of important contextual information by fusing features from different layers in a unified end-to-end manner. Extensive experiments show that the proposed MB SPDG-CNN significantly outperforms single-branch architectures and other existing methods, achieving a precision of 99.78 %, recall of 98.91 %, and F1-score of 99.78 %. The integration of both RGB and thermal features within a multi-branch setup resulted in a 10 % improvement in detection rates compared to single-branch models, demonstrating the effectiveness of our architecture in achieving robust and accurate defect detection.