In the fault diagnosis of aero-engine bearings, it is the key to diagnosis whether the bearing fault characteristics in the engine vibration signal can be accurately extracted. The early cracking and fatigue spalling of aero-engine bearings will produce instantaneous impulse in the vibration signal. However, due to the limitation of testing techniques and sensors installation, most of the key components of the rotating and transmission are difficult to measure directly in the installed state at present, usually by installing sensors on the casing for indirect monitoring. Therefore, in the early stage of the fault, the amplitude of the defect shock signal is significantly lower than the noise level and is masked by the background noise, which cannot be distinguished by conventional spectrum analysis. To this end, this paper proposes an early impulse fault feature extraction method for aero-engine bearings based on convolutional self-coding network. By analyzing the periodicity of the impulse components in the signal, using the translation-invariant learning characteristics of the convolutional self-encoding network, the periodic components in the signal are automatically captured, and the signal is decomposed into multiple feature components reconstructed by the convolution kernel to realize the signal characteristics. The self-learning of components, taking into account the characteristics of the kurtosis index describing the signal impulse component, uses the kurtosis index as the selection index of the optimal feature component, and then realizes the extraction of early impulse fault features. Finally, the effectiveness of the method is verified by simulation data and bearing data.