Aiming at the existing problems of support vector machine (SVM), a new method based on Bayesian optimization and the establishment of the Morlet wavelet kernel SVM is proposed. The Bayesian optimization is used to improve the voting strategy of SVM. The Morlet wavelet kernel SVM is established to make it more appropriate for the classification of impact and nonlinear signals. An example of a rolling bearing proves that the proposed method is robustness and reliable.