<p>Empirical wavelet transform (EWT) is a time-frequency analysis method of non-stationary signals developed in recent years. However, EWT relies too much on the number of initial decomposition modes and has a serious mode mixing. To address these problems, a new signal decomposition method named Translational Gaussian wavelet transform (TGWT) is proposed in this paper. Firstly, the envelope of the spectrum of the original signal is computed by triangular smoothing technique, and then the segmentation boundary of spectrum is obtained, which effectively avoids the dependence on the preset number of initial modes. Subsequently, the modes functions in each frequency band of the segmented spectrum signal are filtered and reconstructed by using translational Gaussian filter banks, which can effectively overcome the problem of mode mixing. Finally, the proposed TGWT method is applied to the simulated signal and vibration signals of rolling bearing with local faults and the results show that the TGWT has excellent ability for fault feature extraction.</p>