GBL-MSS: a graph broad learning method for rolling bearing fault diagnosis under strong noise interference

Abstract

Broad learning system replaces depth stacking with breadth expansion to achieve rapid training and incremental learning, but the inherent randomness of feature mapping under noisy data can lead to unstable performance. To address the aforementioned issues, a graph broad learning based on multisegment smooth strategy (GBL-MSS) method is proposed. By creating a multipeak term, GBL-MSS establishes a smooth error mechanism that simultaneously balances the fitting error with model complexity, effectively mitigates the impact of noise. Furthermore, the multipeak term enables the model to adapt more effectively to nonlinear data distributions and provides superior optimization guidance through flexible error processing. Additionally, graph embedding technology is integrated into GBL-MSS to enhance the extraction of intrinsic data relationships and structural attributes, and obtain strongly correlated feature information. Experimental results demonstrate that GBL-MSS exhibits significant advantages in terms of accuracy, precision, recall, F₁-score, and Kappa.