Abstract
Rotating machinery fault diagnosis remains challenging in practical vibration measurement scenarios, where variable operating regimes, noise interference, and non-stationary degradation hinder reliable interpretation of measured vibration signals. To address these challenges, this paper proposes a prior-guided dynamic fusion network (PGDFNet), termed PGDFNet, for robust vibration signal-based fault diagnosis through collaborative modeling of temporal dynamics, time-frequency structures, and statistical priors. Specifically, a gated temporal Mamba branch is designed to enhance fault-sensitive impulsive responses and capture long-range temporal dependencies from raw vibration signals. Meanwhile, a spatial-frequency Mamba branch models local spectral textures, multi-scale time-frequency structures, and global spatial-frequency dependencies from continuous wavelet transform representations. In addition, a prior-guided dynamic fusion module incorporates statistical priors into the dual-branch fusion process to adaptively coordinate temporal and time-frequency features under noisy measurement conditions. By integrating these components, PGDFNet improves the robustness and discriminability of vibration-based fault representations. Experiments on the CWRU, PU, and JNU datasets demonstrate that the proposed method outperforms representative baselines in classification accuracy, Gaussian and complex-noise robustness, class-imbalance robustness, and cross-domain transfer capability. Ablation, interpretability, and efficiency-complexity analyses further verify the effectiveness of the proposed components and the favorable trade-off between diagnostic performance and computational overhead.
Funding Information
  • Youth Talent Support Program of Shaanxi Science and Technology Association (23JK0701)
  • China Postdoctoral Science Foundation (2024M754225)
  • National Science and Technology Major Project of China (2024ZD0711701-002)
  • Xi’an Science and Technology Planning Projects (20240103)

This publication has 64 references indexed in Scilit: