WPFEformer : A Fault Diagnosis Model for Rotating Machinery Based on Multi-Domain Features and Residual Late Fusion

Abstract

Mechanical equipment fault diagnosis is critical for the operational safety and stability of industrial systems. Complex and variable working conditions often cause severe nonlinear signal degradation. Traditional single-domain models fail to capture comprehensive fault features under such conditions, leading to limited diagnostic performance. To tackle this issue, this paper proposes a novel fault diagnosis model integrating multi-domain features and a residual late fusion mechanism. Firstly, the proposed model explicitly captures multi-domain analytical features through dedicated fuzzy entropy and frequency feature extraction modules, instead of only extracting deep temporal features from vibration signals. Secondly, a residual late fusion module is designed to align and fuse these heterogeneous features via high-dimensional mapping and element-wise addition, facilitating effective feature refinement. Experimental results on the CWRU bearing and NUAA tool wear datasets verify that the proposed model achieves diagnostic accuracies of 99% and 81%, respectively. It outperforms single-domain methods, exhibiting superior feature representation capabilities and robustness in strong nonlinear fault diagnosis scenarios.