The control technology of electromechanical actuator(EMA) and its fault diagnosis is one of the key problems of multi-electric aircraft study. The method based on deep learning is used to diagnose and isolate the classic faults of EMA. The simulation model of EMA is modeling according to the working principle and control law. Four typical faults of EMA are studied,including return channel jam, spall, motor fault, and position sensor fault. Sparse Auto Encoder (SAE) algorithm can perform adaptive extraction of sensor data, which preserves dimensionality reduction and compression while preserving important features. Bidirectional Long Short Term Memory (BiLSTM) neural network is used to effectively process the time series data, which considering both past and future data during the fault diagnosis process. The established EMA model is simulated to obtain normal and faulty data sets, which are used to train the network by SAE-BiLSTM algorithm, and then the trained network is used for online fault diagnosis. After the experiment, SAE-BiLSTM algorithm can well complete the EMA fault detection and isolation.