Due to the non-stationary operating conditions and harsh working environment of wind turbines, the vibration pulse characteristics of wind turbine bearing faults are easily overwhelmed by random noise interference, which poses a challenge to accurately detect rolling bearing faults. In order to reduce the impact of random interference on subsequent feature extraction and algorithm complexity, an improved multi head self attention mechanism (IMHSA)-multi-scale convolutional network (MSCNN)-bidirectional long short-term memory network (BiLSTM) wind turbine bearing fault diagnosis method is proposed. Firstly, the IMHSA composed of periodic cavity self attention and local self attention enhances the features to reduce the impact of random interference and the time consumption during the feature enhancement process; Then, the MSCNN-BiLSTM network is used to extract spatial features and long-term dependency features from the fault signal; Finally, the fault diagnosis results of the fan bearings were output through the fully connected layer and Softmax layer, and the actual operating data of the rolling bearings on the experimental platform was used for numerical analysis. The effectiveness and superiority of the proposed method were verified by comparing it with other similar methods in the field.