A novel method of fault feature extraction for automotive transmission bearing based on combination of wavelet packet-auto regressive model spectrum and divergence calculation is presented. The vibration signals under six differen t wear conditions are decomposed with wavelet packet, the time series at each frequency band are reconstructed, and reconstructed signals are analyzed by AR spectrum estimation. At last, divergence is calculated between new bearing and the other five ones. The experimental result indicates that transmission bearing fault features, such as axial clearance, radial clearance or fatigue spalling, can be shown in wavelet packet AR Spectrum evidently. This method is efficient in automobile transmission bearing fault feature extraction.