The diagnosis of incipient fault of rolling bearing is the effective measure to realize safety production and to avoid major accident. By using high precision accelerometer to collect the vibration signals of bearing, the wavelet soft threshold noise reduction method was used to eliminate noise so as to enhance signal noise ration (SNR) of collected signal. Based on wavelet singularity detection technology, the extraction of initial fault characteristics submerged in noise background method was discussed, at the same time the shortcomings of traditional Fourier transform were pointed out. Research shows that the method is effective, the extraction of fault feature frequency and the fault characteristic frequency from theoretical calculation are basically the same. Research results provide a new way for incipient fault detection and diagnosis of rolling bearing.