Feature Extraction Using Sparse Kernel Non-Negative Matrix Factorization for Rolling Element Bearing Diagnosis.

Area of Science:

• Mechanical Engineering
• Signal Processing
• Machine Learning

Background:

• Bearing faults generate complex vibration signals, hindering traditional spectral analysis for early detection.
• Existing matrix analysis methods like SVD and RSVD lack interpretability for fault feature extraction.
• Non-negative Matrix Factorization (NMF) offers improved interpretability but can be further enhanced.

Purpose of the Study:

• To propose a novel feature extraction method for early bearing fault detection.
• To enhance the interpretability and effectiveness of Non-negative Matrix Factorization (NMF) for bearing diagnostics.
• To extract periodic impulse fault features from vibration signals with greater accuracy.

Main Methods:

• Developed a sparse Kernel Non-negative Matrix Factorization (KNMF) method.
• Applied L1 regularization to the NMF model for sparser spectral bases.
• Utilized a linear kernel function within KNMF for time-frequency decomposition.
• Extracted fault features from decomposed subspaces based on sparse frequency bands.

Main Results:

• The proposed sparse KNMF method demonstrated superior performance in extracting fault features.
• Achieved deeper extraction of fault features compared to conventional NMF and SVD.
• Experimental validation confirmed the high effectiveness of the method.

Conclusions:

• Sparse KNMF is a powerful tool for early bearing fault detection.
• The method provides enhanced interpretability and accuracy in identifying fault signatures.
• KNMF offers a significant advancement over existing matrix-based diagnostic techniques.