Abstract Numerous sensors have been deployed in different locations in components of wind turbines to continuously monitor the health status of the turbine system and accordingly, generate a large volume of operation data by the supervisory control and data acquisition (SCADA) system. Naturally, these sensory data are multivariate time series with high spatio-temporal correlations. It is still challenging to effectively model such correlations and then enable an accurate fault diagnosis. To this end, we proposed a new spatio-temporal fusion neural network (STFNN) for wind turbine fault diagnosis. Specifically, a multi-kernel fusion convolution neural network (MKFCNN) with multiple convolution kernels of different sizes is first designed to extract multi-scale spatial correlations among different variables. Then, we adopt the long short-term memory (LSTM) to further learn the temporal dependence of the learned spatial features. The proposed STFNN model provides an end-to-end fault diagnosis way, which can directly learn spatio-temporal dependency from the raw SCADA data and give the fault diagnosis result. The effectiveness and superiority of the proposed method are evaluated on a generic wind turbine benchmark simulation dataset and a SCADA dataset from a real wind farm. Both experimental results have indicated that the proposed method outperformed several compared methods.