Rolling element bearings are crucial components in all kinds of rotating machinery. Its fault detection is of great importance, as it ensure the performance of the whole machine. Periodic transient impulses caused by bearing defects are usually submerged in strong background noise which poses a challenge for effective fault feature extraction. To detect bearing faults reliably, a new fault feature extraction method is presented. First, the adaptive maximum second-order cyclostationary blind deconvolution (ACYCBD) is utilized to recover bearing fault related impulses, while the optimal filter length is chosen based on the harmonic significance index (HSI) which quantifies the diagnostic information contained in a deconvoluted signal. Second, cross-correlation is calculated between the teager energy operator (TEO) and the envelope of the deconvoluted signal to further eliminate the irrelevant noise. Finally, fast fourier transform (FFT) is employed to acquire the cross-correlation spectrum and the fault features can be extracted successfully. The performance of the proposed method is verified on both simulation signals and experimental signals acquired from a test rig. The superior abilities of noise reduction and fault detection are shown clearly when compared with some traditional method.