In order to enhance the utilization value of the scarce fault samples of rolling bearing and aimed at the characteristic that support vector machines (SVM) are sensitive to noises, a new method of bearing running state recognition based on wavelet threshold denoising and SVM was proposed. The present gathered samples of bearings vibration signals were denoised with wavelet threshold denoising method, and the corresponding denoised samples were gotten. On this basis, the SVM model was preliminarily established by combining with parameter optimization of SVM. Then, the samples which were classified incorrectly were denoised afresh and the SVM model was reconstructed until the penalty factor and the accuracy of cross validation fixed preconcerted requirements, so as to realize the establishment of optimal SVM model and the identification of bearing running state. However, the disadvantages of the traditional softthresholding and hardthresholding functions of their own restricted the effects of signal denoising and feature extraction, then the adjustability of denoising processing could be realized. Firstly an improved thresholding function was presented, and the advantages of this function were analyed by the MATLAB simulation experiments. The final diagnosis example of rolling bearing indicates that the introduction of the improved threshold denoising method can effectively enhance utilization rate of sample data, the generalization capability and antinoise property of SVM and the reliability of the intelligent diagnosis of rolling bearing.