I have 4 class bearing data, the four types of data are unified first wavelet noise reduction, and then using wavelet packet energy feature extraction, with bp, finally met a question now is not direct energy denoising in bp classification accuracy to 90%, more than 60% after de-noising, through debugging threshold function, only when the rigrsure + "s" can reach 80%, but even if 1 layer decomposition, the final classification accuracy is lower than not de-noising, bother, every answer or discuss, thanks