Abstract: The rolling element bearing works in a very complex environment. Early bearing fault features are relatively weak and easily to be overwhelmed by the large amount of noise. How to reduce the noise and extract the fault features accurately is a difficult problem. Novel noise reduction method is put forward based on dual tree complex wavelet transform (DTCWT) and maximum variance unfolding (MVU). The DTCWT is applied to rolling-bearing vibration signals to structure the high dimensional signal space. The MVU is used to extract the real signal subspace from high dimensional signal space and eliminate the noise interference of the noise subspace. The proposed method makes full advantage of the nonlinear feature extraction ability of the MVU algorithms and the perfect reconstruction and near shift invariance of the DTCWT. The simulation and the bearing engineering signal is carried out to inspect the noise reduction method, and the results demonstrate that DTCWT_MVU can effectively eliminate the noise component of the bearing signal and keep its characteristic waveform and improve signal to noise ratio (SNR) with the practical and generality.

Key words: dual-tree complex wavelet transform, fault diagnosis, maximum variance unfolding, noise reduction, rolling bearings