摘要

Parametric atom decomposition provides the chance of adaptation for the non-stationary signals decomposition. However, the parametric atoms, whether Fourier atoms or wavelet atoms preset artificially, are regular. These atoms do not have enough match degree to the practical signals which are non-stationary and irregular. Therefore, the coefficients of the parametric atom decomposition may be dense and their physical interpretation is difficult to determine or ambiguous. Extraction of impact features is the key problem for the fault diagnosis of rolling element bearings. According to the cyclo-stationary nature of the fault signal, authors extract basic atoms from the analyzed signal instead of artificially giving the regular basis, hence the match degree is improved and the coefficients of the decomposition have more explicit physical meaning. Compared with wavelet packet atom decomposition, the non-parametric atom decomposition captures the integrity of the impact and represents the fault signal more sparsely and has more intelligible interpretability. While compared with the spectral correlation density function, the non-parametric atom decomposition better emphasizes the main information of the fault and the frequency resolution of the novel method is irrelevant to the length of the analyzed signal.