Bearing defect diagnosis based on semi-supervised kernel Local Fisher Discriminant Analysis using pseudo labels

Area of Science:

• Mechanical Engineering
• Data Science
• Machine Learning

Background:

• Information fusion in bearing defect diagnosis can lead to high-dimensional, redundant data, hindering classification performance.
• Supervised methods require extensive labeled data, which is costly and difficult to obtain for critical machinery.
• Extracting optimal features from limited labeled and abundant unlabeled data is a significant challenge.

Purpose of the Study:

• To propose a novel bearing defect diagnosis model for improved feature extraction and classification accuracy.
• To effectively utilize unlabeled data for regularizing supervised dimensionality reduction.
• To address the challenges of high-dimensionality, data redundancy, and limited labeled samples in machinery fault diagnosis.

Main Methods:

• Developed a semi-supervised kernel local Fisher Discriminant Analysis (SSKLFDA) model incorporating pseudo labels.
• Employed Density Peak Clustering to generate pseudo cluster labels for both labeled and unlabeled data.
• Introduced regularization strategies based on pseudo labels to optimize scatter matrices and utilized the kernel trick for non-linear feature extraction.

Main Results:

• The proposed SSKLFDA model effectively extracts optimal features for classification.
• Utilizing unlabeled data through pseudo labels enhances the regularization of supervised dimensionality reduction.
• Experimental results demonstrate superior classification performance compared to existing dimensionality reduction methods across various scenarios.

Conclusions:

• The SSKLFDA model offers a robust solution for bearing defect diagnosis, particularly with limited labeled data.
• The method effectively handles high-dimensional, redundant data and improves feature discriminability.
• The semi-supervised approach with pseudo labels significantly enhances fault diagnosis accuracy and reliability.