Bearing Fault Diagnosis Based on Statistical Locally Linear Embedding

Area of Science:

• Engineering
• Data Science
• Signal Processing

Background:

• Machinery fault diagnosis relies on pattern recognition from complex signals.
• High-dimensional signal data often resides on nonlinear manifolds, complicating feature extraction and dimensionality reduction.
• Improving recognition performance is critical for effective fault detection.

Purpose of the Study:

• To propose a novel fault diagnosis approach using Statistical Locally Linear Embedding (S-LLE).
• To leverage fault class label information for enhanced manifold learning.
• To improve feature extraction, dimensionality reduction, and classification accuracy in machinery fault diagnosis.

Main Methods:

• Feature extraction from vibration signals using time-domain, frequency-domain, and Empirical Mode Decomposition (EMD).
• Application of the proposed Statistical Locally Linear Embedding (S-LLE) algorithm for dimensionality reduction.
• Classification and fault diagnosis in the reduced low-dimensional feature space.

Main Results:

• The S-LLE algorithm effectively translates complex signal modes into a salient low-dimensional feature space.
• The proposed approach demonstrates superior performance compared to Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), and standard Locally Linear Embedding (LLE).
• Validation using rolling bearing fault signals shows significant improvement in classification performance.

Conclusions:

• The S-LLE based fault diagnosis approach offers a robust method for pattern recognition in machinery diagnostics.
• This technique significantly enhances classification accuracy and outperforms traditional methods.
• The approach facilitates easier and faster pattern classification and fault diagnosis.