In the special working environment of ship, stator current analysis is used to diagnose the fault of ship induction motor bearings. Time-frequency analysis is applicable to the processing of unsteady signals, and Time-frequency Domain Feature Fusion (TDFF) is proposed to construct the original feature set based on the features in the time-frequency domain. Feature screening is carried out using BP neural network single feature diagnosis accuracy evaluation. The selected features are weighted and fused, and finally the features that contain complete fault information are sensitive to the fault state of induction motor bearings. Improved grey Wolf algorithm (IGWO) is applied to optimize BP neural network, and fault identification is carried out. The experimental results show that the proposed fusion feature TDFF can effectively improve the fault identification accuracy and realize the motor bearing fault diagnosis.