An informative frequency band extraction method with refined frequency band division and Gaussian-regularized kurtosis for bearing fault diagnosis

Extracting the informative frequency band is a challenging task in bearing fault diagnosis, as fault-induced resonances are often unknown and heavily masked by strong interference from other rotating components and background noise. In this study, a new method with refined frequency band division and Gaussian-regularized kurtosis is proposed to extract the informative frequency band for bearing fault diagnosis. Specifically, an enhanced dual-tree complex wavelet packet transform is introduced to achieve finer frequency-band resolution while retaining all the advantages of the original transform. Meanwhile, a more noise-robust indicator, termed Gaussian-regularized kurtosis, is defined to quantify fault information within the divided frequency bands. The analysis results for simulated and several real-world bearing signals across diverse cases verify the effectiveness of the method. Compared with several advanced methods, the proposed method demonstrates clear advantages in extracting the informative frequency band and exhibits significant potential for diagnosing bearing faults under strong interference.