Abstract
Periodic modulation transients caused by localized defects in rolling bearings are key indicators for condition monitoring and fault diagnosis. However, under complex operating conditions and strong noise interference, vibration signals often become highly nonstationary, with weakened modulation characteristics and fault features easily masked by broadband noise. To address these challenges, a Periodic Modulation Convolution Sparse Self-Learning (PM-CSSL) method is proposed for adaptive enhancement and accurate extraction of fault features. First, symplectic geometry mode decomposition (SGMD) is employed to denoise the signal while preserving periodic impulse information. Then, a parameterized atomic prototype based on the Laplace wavelet is constructed, and a candidate atom dictionary is generated using Latin hypercube sampling (LHS). The optimal atom is adaptively selected via a maximum correlation coefficient criterion and further periodized to form physically interpretable modulation atoms. Next, frequency-domain mapping and an energy concentration criterion are used to identify the optimal modulation atom. A convolutional sparse representation model is then established in the frequency domain to enhance modulation features and reconstruct the signal. Finally, envelope demodulation is applied, and the extracted frequencies are compared with theoretical fault frequencies for accurate fault identification. Experimental results demonstrate that the proposed method effectively enhances fault-related features and improves diagnostic accuracy under noisy conditions.
Funding Information
  • National Natural Science Foundation of China (52274230)

This publication has 30 references indexed in Scilit: