Adaptive weak fault detection of rotating machinery via feature mode decomposition and feedforward-feedback stochastic resonance

Abstract

Structural health monitoring of critical rotating machinery is vital for preventing catastrophic failures, yet extracting incipient fault signatures remains challenging due to heavy background noise and complex transfer paths. To address this, a novel blind fault feature enhancement framework is proposed, integrating feature mode decomposition (FMD) with an improved feedforward-feedback Gaussian multi-stable stochastic resonance (SR) model. Unlike traditional denoising, this method actively utilizes the stochastic resonance mechanism to transfer noise energy into fault features, facilitated by a feedforward-feedback loop that enhances system robustness. First, FMD adaptively decouples impulsive fault transients from periodic interferences. Subsequently, a quantum genetic algorithm, guided by a newly constructed blind detection index, optimizes the potential well parameters without requiring prior fault knowledge. The method is rigorously validated on both a standard rolling bearing dataset and a complex wind turbine planetary gearbox dataset. Results demonstrate that it successfully extracts weak failure frequencies from highly coupled vibration signals, showing superior signal-to-noise ratio improvement compared with the conventional coupled Gaussian tri-stable SR model.