Non-negative EMD manifold for feature extraction in machinery fault diagnosis

• Useful intrinsic mode functions (IMFs) are selected by correlation analysis from empirical mode decomposition (EMD).
• A new method is introduced to extract representative features from non-negative matrix factorization (NMF) results.
• Manifold learning is employed to refine the non-negative EMD features.
• Practical engineering cases verified the advantages of the proposed method in machinery fault diagnosis.

This paper proposes a novel non-negative empirical mode decomposition (EMD) manifold (NEM) method for feature extraction in machinery fault diagnosis. The NEM feature is extracted from the fault-related intrinsic mode functions (IMFs) by two main steps: non-negative EMD (NNE) feature construction and manifold refining. The first step employs non-negative matrix factorization (NMF) on IMFs selected by correlation analysis, and then extracts NNE features by optimization algorithms. The second step aims to further explore the intrinsic pattern of NNE features and remove redundant information to obtain more stable NEM features. The NEM feature is associated with the key information from massive vibration data, thereby exhibiting valuable properties for fault pattern recognition. The validity of NEM is confirmed by three engineering experiments including a gearbox case and two rolling-element bearing cases.