Due to the complexity and the diversity of the bearing fault signal, this article proposes a method of feature extraction and mode recognition applied on incipient fault signal of motor bearing. A method based on the Euclidean distance parameter is presented to extract the fundamental information of fault such as the intensity of failure. Subsequently, a method compounding k-folder cross validation algorithm and the improved probabilistic neural network dealing with the Euclidean distance parameter is proposed to discriminate fault pattern. The improved probabilistic neural network improves the accuracy of calculating the probability density function and diagnosis result. Both simulated and experimental fault signal is used to verify the efficiency and accuracy of the proposed method. In conclusion, it is shown that this detecting process can effectively identify the motor bearing faults and improve the recognition accuracy and computation efficiency.