Taking advantage of approximate shift-invariance, a method for combining the three high-pass wavelet components in a special rule is proposed to obtain a wavelet component signal with high time resolution. Results show t hat the method eliminates the influence of down sampling in wavelet decomposition on the accuracy of signal analysis and conquers the frequency aliasing. A fault diagnosis method based on overcomplete rational discrete wavelet tran sform(ORDWT) is proposed and applied to the early fault diagnosis of a roller bearing. Comparative results indicate that the method can effectively extract early fault feature than dyadic discrete wav elet transform.