In this study, a graph neural network-based predictive maintenance (GNNPM) algorithm was developed to diagnose faults and recommend maintenance strategies for wind turbine gearboxes. This research addresses key limitations of existing approaches, such as the absence of system-level analysis and the limited ability to perform fault diagnosis and maintenance decision-making simultaneously. Using simulation-generated fault and degradation data, a maintenance knowledge graph was constructed to model relationships between fault states and maintenance actions. The proposed method employs a time-domain relational graph convolutional network (TRGCN) to analyze the time-series data and evaluate the system states before and after maintenance. The GNNPM algorithm realized outstanding results in multilabeled fault classification and maintenance action prediction tasks. The TRGCN recorded average accuracies of 94.06 % and 97.10 % for fault mode and degradation state classification. These results surpassed those of traditional graph neural network models. Furthermore, TRGCN achieved high prediction accuracy for maintenance actions: 93.4 % for inspections, 93.06 % for minor repairs, 92.59 % for major repairs, and 90.9 % for replacements. By integrating knowledge graphs with graph neural networks, this study overcomes major limitations in predictive maintenance. It also provides a reliable and explainable framework for maintenance planning. The proposed framework establishes a foundation for improving decision making and enhancing predictive maintenance in wind turbines and other complex systems.