Abstract: To address the problem that improper selection of relevant parameters in extreme learning machine（ELM） affects its diagnostic results and accuracy, the selection of weights and biases of the ELM based on bold eagle search（BES） algorithm was optimized, and a wind turbine fault diagnosis model combined with vulture search algorithm optimization and extreme learning machine（BES-ELM） was constructed. The relevant SCADA data of a wind turbine in a wind farm in four states, including generator overheating（S1）, feeder failure（S2）, converter cooling system failure（S3） and normal（S4） was preprocessed and features were selected to form a fault sample set, with 80% of the sample set being the training set and 20% being the test set. These fault samples were classified using standard ELM, extreme learning machine model based on genetic optimization algorithm and particle swarm optimization algorithm, respectively. The results show that compared with the standard extreme learning machine, the model optimized by genetic algorithm and particle swarm algorithm, the diagnostic accuracy of the BES-ELM model reaches 98.75%, which effectively improves the accuracy of wind turbine fault diagnosis.