Abstract: Bearings are critical component in high-end applications, and high-precision detection of internal and near-surface defects are the key to ensure their high quality.In the ultrasonic defect detection of bearings, the small defect signals near the surface of the bearing were submerged in the surface or bottom echo signals, and there was a blind area of defect detection.To solve this problem, a method based on the Hilbert spectrum for identifying near-surface defects in bearing inner ring was proposed.The processed ultrasonic data were compared in time-frequency domain based on Hilbert-Huang transform and wavelet transform.With the convolutional neural network, cross-experiments were used to verify the reliability of the two methods in identifying near-surface defects.The results show that the method of using Hilbert transform is better than wavelet transform, and it can classify the small defects on the upper and lower surfaces at the same time.The method effectively improves the detection accuracy of the near-surface defects.An average accuracy of the recognition of 98.83% is achieved, which can provide a new method for the recognition of near-surface defects.