Bearing defect detection and damage degree detection always are very important problems in the field of rotating machinery.Although the current research on vibration signals has achieved quite good results,the diagnostic effect still needs to be improved when it is difficult to install vibration sensors.Aiming at the sound generated by faulty bearings under strong background noise,a diagnosis method was proposed based on recursive Hilbert transform and 1D convolution neural network to extract abstract features and carry out pattern recognition.A global average pooling layer was introduced into the convolution neural network structure to speed up the operation of the network.Finally,the validity of the proposed method was verified by data sets and its superiority was verified by comparing with other commonly used classification methods.The results show that the proposed algorithm can not only accurately identify the damaged parts of bearings,but also accurately distinguish the damage degree of components.