To address the general problem of low accuracy and long training period of electro hydraulic servo fault diagnosis prediction, and the limited sensing data currently available from launch vehicles, an electro hydraulic servo fault diagnosis method based on BP network optimized by genetic algorithm (GA) and LM (Levenberg-Marquardt) algorithm, GRU (Gated Recurrent Unit) network and RF (Random Forest) is designed. Using GA to determine the optimal initial value of BP network avoids the BP network from falling into local optimal solutions, and the LM algorithm can significantly accelerate the convergence speed of the BP network and improve the training efficiency; GRU network is a special recurrent neural network, which is often used to deal with time series related problems and has advantages in dealing with time-sensitive signals; RF is a highly flexible machine learning method, which is suitable for dealing with classification problems. In this paper, we first use these three neural networks to deal with the classification problem of typical faults of electro-hydraulic servo respectively, and then use the diagnosis results of these three models to vote and combine the threshold judgment to diagnose, and obtain a diagnosis method with short training period and high accuracy.