Dual-Reweighted Sparse Low-Rank Enhanced Dictionary Learning for Early Fault Diagnosis of Rotating Machinery

• Published in: IEEE transactions on instrumentation and measurement Vol. 74; pp. 1 - 14
• Main Authors: Cai, Gaigai, Liu, Zeyu, Yan, Yulong, Wang, Jiaxing, Hu, Yaoyang, Wang, Shibin
• Format: Journal Article
• Language: English
• Published: New York IEEE 2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Comparative studies; Dictionaries; Dictionary learning; early fault diagnosis of machinery; Fault diagnosis; fault prior knowledge; Feature extraction; Harmonic analysis; Harmonics; Interference; Iterative algorithms; Learning; Machine learning; Machinery; Noise; Optimization; Regularization; reweighted strategy; Rotating machinery; Singular value decomposition; sparse representation; Vibrations
• ISSN: 0018-9456; 1557-9662
• DOI: 10.1109/TIM.2025.3599700

The fault vibration signal of rotating machinery is usually a mixture of impulses, harmonics, and noise. How to effectively extract the repetitive impulse signals is a key issue in machinery fault diagnosis. However, in the early stage of fault, weak fault features are easily masked by complex noise and harmonic interference, making classic denoising algorithms ineffective. This article proposes a dual-reweighted sparse low-rank enhanced dictionary learning (Re2wSLDL) method. Specifically, the sparse regularization term based on the <inline-formula> <tex-math notation="LaTeX">l_{1} </tex-math></inline-formula>-norm and the low-rank regularization term based on the nuclear norm are integrated into the dictionary learning model, so that the dictionary representing the fault features is learned under the combined constraints of two types of prior knowledge. In order to maintain the amplitude of the extracted signal while ensuring a singular value distribution that is more consistent with the ideal state, a reweighted enhancement strategy is introduced for the sparse and low-rank regularization terms. Second, an iterative algorithm based on variable splitting and alternating optimization is developed. Meanwhile, we conduct an in-depth study on the selection of key parameters. Simulation analysis and two experimental cases of early fault diagnosis verify the suppression effect of Re2wSLDL on harmonic and noise interference and the superior extraction performance of fault impulse signals. Comparative studies show that Re2wSLDL can effectively diagnose early faults and outperforms K-singular value decomposition (K-SVD) and Adaptive Enhanced Sparse Period Group Lasso (AdaESPGL).