.png){kind=icon}
.png){kind=icon}
Return
An intelligent fault diagnosis method for rolling bearings integrating multi-level feature extraction and supervised geometry and statistics-preserving embedding
.png){kind=icon}
PRE
AI
W
X
Z
D
F
D
DOI：10.1016/j.sigpro.2026.110681.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In industrial condition monitoring systems, accurate and robust fault diagnosis of rotating machinery in complex environments remains a critical challenge. Existing deep learning models, such as StarNet, often suffer from insufficient feature representation, information loss during pooling, and inadequate loss functions that fail to capture intra-class and inter-class relationships. To address these limitations, this paper proposes a novel network model based on PGStarNet-MFE and SWD-GSE. Firstly, a novel multi-head feature extraction (MFE) module is constructed using a multi-scale convolutional architecture to achieve hierarchical feature representation of time-frequency maps, thereby significantly enhancing the network’s feature extraction capability. Secondly, a supervised Sliced Wasserstein Distance-Geometry and Statistics-preserving Manifold Embedding (SWD-GSE) dimensionality reduction algorithm is designed to replace the traditional average pooling layer. This algorithm introduces a geometric structure preservation mechanism during feature space reduction, retaining critical diagnostic information. Finally, a linearly weighted loss function (PGLF) based on an improved Greater Cane Rat Algorithm is proposed to optimize feature space distribution via correlation modeling. The proposed approach is evaluated on three bearing datasets, and its performance compared with alternatives. Experimental results show the framework consistently achieves superior diagnostic accuracy and robustness. This study offers a novel pathway for intelligent bearing diagnosis under complex conditions.
Keywords:
fault diagnosis
rolling bearings
multi-level feature extraction
supervised embedding
geometry and statistics preservation
Journal
IF:
3.6
Papers:
9.8K
Citations:
1.7W
Organization
No organization information available
Cited Papers