27 March 2026 in Measurement Science and Technology
Existing bearing fault diagnosis methods often perform poorly when confronted with the complex industrial challenges of small samples, strong noise, and variable operating conditions. This paper proposes a diagnostic approach that integrates signal processing techniques with an improved capsule network (CN). This approach first designs a novel wavelet hybrid threshold function to overcome the inherent limitations of traditional threshold-based noise reduction. This function is combined with successive variational mode decomposition, which can adaptively decompose signals, forming a synergistic noise reduction scheme that effectively enhances fault features under strong noise conditions. Building upon this foundation, the improved CN diagnostic model multi-scale selective kernel (MSSK)-CapsNet is constructed. By introducing the proposed MSSKNet to comprehensively capture multi-range fault features, and integrating grouped convolutions and nonlinear transformation modules into the primary capsule layer, the model aims to reduce computational complexity while enhancing its deep nonlinear representation capability for complex fault patterns. Validation results from the CWRU dataset and a self-built experimental platform dataset demonstrate that this method maintains optimal diagnostic performance across various complex operating conditions. Its accuracy and generalization capabilities surpass multiple mainstream approaches, proving its superior fault feature recognition and diagnostic capabilities.