TITLE:
Application of a Projection Filtering Enhanced Convolutional Network in Fault Diagnosis of Variable-Speed Rolling Bearings
AUTHORS:
Lijie Tang, Weiwen Tian
KEYWORDS:
Rolling Bearing, Convolutional Network, Projection Filtering, Fault Diagnosis
JOURNAL NAME:
Open Journal of Applied Sciences,
Vol.16 No.3,
March
24,
2026
ABSTRACT: Under variable-speed conditions in modern industry, rolling bearing vibration signals are highly non-stationary, and fault features are easily obscured by speed interference and noise. Traditional methods and existing models, lacking systematic interference suppression and targeted feature screening, often result in low feature discrimination and insufficient diagnostic accuracy, limiting industrial applications. To address this, a projection-filtering-based convolutional network (NDACN) is proposed for rolling bearing fault diagnosis. First, a projection matrix is constructed via non-redundant attribute projection (NAP) to effectively strip variable-speed interference, converting non-stationary features into approximately constant-speed stationary features for de-conditioning. Second, a dynamic filtering threshold (DFT) module is adopted, using gradient retention and dynamic threshold adjustment to adaptively select effective fault features, mitigating key information loss and gradient degradation caused by traditional hard thresholding. Then, a parameterized attention convolution enhancement mechanism is introduced to focus on critical fault bands and feature regions through adaptive weight allocation, strengthening core fault representation while suppressing irrelevant interference, and reducing information loss via gated residual fusion. Finally, validation on three variable-speed bearing datasets shows that the proposed method achieves superior diagnostic accuracy and stability under variable-speed conditions.