The fault signal characteristics of motor rolling bearings for more electric aircraft are easily masked by strong background noise. Directly using machine learning, deep learning, or other methods results in a lower accuracy in fault recognition. In this article, a Northern Goshawk algorithm using a fusion subtraction optimiser and Cauchy strategy (SCNGO) is proposed to optimise the number of white noise additions and amplitude weights in the improved full set empirical mode decomposition method based on adaptive noise (ICEEMDAN). Then, a multi‐scale convolutional neural network (MCNN) is used to extract the time–frequency domain features of the de‐noised signal and perform information fusion. Finally, the bidirectional long short‐term memory network (BiLSTM) was used to learn the faults' fusion features and complete the faults' recognition at different speeds. The research results show that the SCNGO‐ICEEMDAN and MCNN‐BiLSTMT model shows significant advantages in bearing fault recognition with an average recognition accuracy of 98.67% at various speeds.