In order to automatically recognize different scales of bearing faults, a method of pattern recognition based on the multiwavelet packet sample entropy method and BP neural network is put forth. First, the vibration signals of rolling bearings with five different scaled outer race defects are decomposed into three layers using the GHM multiwavelet packet. The signal sample entropy of 16 decomposed frequency bands are then used as the neural network′s input vector, so that the complete information of the multiwavelet packet decomposition can be thoroughly utilized. Based on the learning and training of a three-layer BP neural network, and in comparison with the dB10 wavelet packet, it can be concluded that the convergence speed and identification accuracy of the multiwavelet packet sample entropy method is much better than those of the traditional wavelet neural network classification. As a result, the multiwavelet packet sample entropy method is effective in automatically recognizing different scales of bearing faults.