The existing supervisory control and data acquisition (SCADA) system continuously collects data from wind turbines (WTs), which provides a basis for condition monitoring (CM) of WTs. However, due to the complexity and high dimension and nonlinearity of data, effective modeling of spatial-temporal correlations among the data still becomes a primary challenge. In this paper, we propose a novel CM approach based on the multidirectional spatial-temporal feature aggregation networks (MSTFAN) to accurately evaluate the performance and hence diagnose the faults of the turbines. Firstly, the data collected from various sensors are formulated into graph-structured data at each timestamp. Spatial-temporal network constructed by combing a graph attention network (GAT) and a temporal convolutional network (TCN) is used to extract spatial-temporal features of the data. Then, a bi-directional long short-term memory (BiLSTM) neural network is adopted to further study long-term spatial-temporal dependency of the extracted features. Finally, the condition score is obtained and the streaming peaks over threshold (SPOT) is applied to determine the abnormal threshold for early warning of the fault occurrence. Experiments on datasets from real-world wind farms demonstrate that the proposed approach can detect the early abnormal situation of the WTs, and outperform other established methods.