Aimed at the problems that fault diagnosis is low to accuracy caused by extracting fault features difficultly, and vibration signals of inter-shaft bearing on aero-engine are susceptible to noise interference at present, a fault diagnosis method is proposed for aero-engine inter-shaft bearing based on the improved residual attention network and bidirectional long short-term memory neural network(BiLSTM). Firstly, taking the original vibration signal as a model input, the local spatial features are extracted from the raw data by utilizing one-dimensional wide convolution, and the high-frequency noise is suppressed. And then, a residual network in combination with the improved channel attention is utilized for enhancing model attention to important features and reducing model com putational complexity, and the processed features are input into BiLSTM to further extract temporal correlation features. Finally, the features are input into the Softmax layer for fault classification. The experimental validation is conducted by using the Harbin Institute of Technology Aeroengine intershaft bearing dataset, and the results show that the proposed model can maintain the diagnostic accuracy of 98.64% even in the high noise environment with the signal-to-noise ratio of -4 dB, is prior to the other comparative models, and has the better ability to extract features and resist noise.