Abstract：The de-nosing principle using wavelet was discussed and the de-nosing property was compared using general threshold strategy. Following them a per-level de-nosing method based on the analysis to the autocorrelation of the wavelet detailed coefficients is proposed. The maximum wavelet decomposing level is determined according to the minimum interest frequency. The step of the per-level de-nosing method is then given and its experiment research to vibration model is processed. The results proved that the method is better for de-nosing, which is more suitable to recover the interest weak signal from intense background noise. Last, its application in the engine vibration is processed and the results proved that it can inhibit the background noise effectively and recover the interest information. It provides a practical strategy for engine vibration pre-processing.