The unnoticeable defects of hotspots and cracks and delamination in PV systems reduce system effectiveness and increase costs through energy yield deterioration. The current fault detection standards including Logistic Regression and Decision Trees alongside Support Vector Machine (SVM) do not possess adequate means to detect sophisticated spatial relationships between solar panel components. With advanced precision levels CNNs have demonstrated poor capacity to detect architectural arrangements between solar cell parts. The research implements Convolutional Graph Neural Networks (ConvGNN) to detect PV defects by using solar cells as nodes in a graph structure which includes spatial relationship edges. The dual relationship identification ability of ConvGNNs enables outstanding defect discovery performance at both small-scale and large-scale levels. The model achieves 92.7% overall accuracy in detecting PV system defects through its training process with thermal imaging and electroluminescence imaging data beyond standard AI-based detection methods. The developed approach implements a graph-based learning method for PV system monitoring which delivers improved performance reliability for big PV installations. When employed for PV fault detection the innovative AI method ConvGNN delivers a revolutionary approach to fault identification which optimizes maintenance operation efficiency by producing better power conservation alongside operational efficiency outcomes.