In promoting the healthy development of the wind energy industry, condition monitoring of wind turbines (WT) plays a crucial role. Existing data-driven condition monitoring methods primarily rely on time-series data, such as supervisory control and data acquisition (SCADA) and condition monitoring system data analysis, failing to effectively utilize the information contained in the WT textual data, such as design specifications, operation manuals, papers, patents, maintenance records, fault reports, etc. They have limitations in the analysis of fault transmission causality and the interpretability of analysis results. Therefore, this article proposes a WT condition monitoring method driven by the fusion of a knowledge graph and spatio-temporal graph neural network (KG-STGNN). This method first utilizes textual data combined with WT structural and other information to construct a WT operation and maintenance knowledge graph, forming a directed graph structure for WTs. Then, SCADA data are embedded into the graph structure to generate WT time-series graph data. A high-order graph attention network (HGAT) and a Transformer are used to establish a spatio-temporal graph, the spatial and temporal characteristics in the graph data are mined for condition monitoring. Historical healthy data of the WT are used to train the KG-STGNN model. Finally, the operating status of the WT is determined based on the information represented by nodes in the graph, and a monitoring strategy is constructed to determine fault warning times and interpret the condition monitoring results. Analysis of actual WT cases evaluates the effectiveness of the proposed method, with its performance outperforming traditional graph neural network models. Case studies on two WTs show that the proposed method exhibits excellent performance in condition monitoring, since it has the lowest false alarm rate and the earliest anomaly warning capability. Ablation experiments show that the graph structure constructed from the knowledge graph is critical for improving model performance. The proposed monitoring strategy eliminates more than 85% of false alarms and also provides good interpretability for the monitoring results.