The vibration signal of the rolling bearing is usually nonstationary under complicated operating status and some typical fault features tend to be covered by other components, which brings great difficulty for the fault feature extraction. In the light of this problem, a new procedure based on the synchrosqueezed wavelet transform (SWT) is proposed for the feature extraction of the rolling bearing signal. The vibration signals of the rolling bearing are analyzed under various operating status and the signal feature space is extracted to reflect the operating conditions of rolling bearing. Second, the non-negative matrix factorization (NMF) is performed to simplify and optimize the signal feature space so as to extract the feature parameters for fault diagnosis and pattern recognition. Finally, the support vector machine is applied to classify the various vibration signals of the rolling bearing. The results indicate that the proposed method is superior to the traditional time domain feature extraction method in pattern recognition.