Empirical Mode Decomposition(EMD) decomposition is a critical step in Hilbert-Huang Transform(HHT), accompanied by overshoot and endpoint effect. The Genetic Algorithm(GA) is used to optimize and select the unknown parameters including the penalty function C and default parameters of Gaussian kernel of Support Vector Machines(SVM). GA-SVM is applied to extend signals to deal with endpoint effect, and cubic Hermite polynomial interpolation is adopted for envelope fitting. In order to extract the early stage fault frequency features of mechanical equipment, wavelet packet noise reduction pretreatment is performed, combined with the extraction experiment of bearing fault feature frequency by using improved HHT transform. The experimental results show that the proposed method can improve the accuracy of fault frequency extraction.