A novel bearing feature extraction technique via adaptive VMD and nonlinear-dynamics feature-space mapping under strong background noise

In practical applications, bearing fault signatures are often masked by heavy ambient noise. To improve the sensitivity of fault features under strong background noise, this paper proposes a novel nonlinear-dynamics feature-space mapping (NDFSM) strategy to enhance feature separability and noise robustness. By integrating this strategy with a robust parameter-adaptive variational mode decomposition (VMD), a robust bearing feature extraction algorithm is developed. In the signal preprocessing stage, a hybrid whale optimization algorithm-harris hawks optimization (WOA-HHO) approach is proposed to adaptively optimize the VMD parameters. This process is guided by a novel composite fitness function constructed from envelope entropy and envelope spectrum kurtosis. Subsequently, the proposed NDFSM strategy is employed to effectively extract latent fault characteristics from the optimal intrinsic mode function (IMF). A nine-dimensional complexity vector is mapped into a 2D spatial domain to generate distinguishable 2D feature maps, which explicitly enhances feature separability for the subsequent complexity pattern recognition network (CPR-Net). Finally, these generated feature maps are fed into the network to recognize different bearing health states. Experimental verification on the Case Western Reserve University (CWRU) dataset and a rotor-planetary gear test bench demonstrates that the proposed algorithm achieves an average diagnostic accuracy of 100% and 98.92% (under −10 dB noise), respectively. These results comprehensively confirm its superiority and reliability in harsh environments.