Multi-Manifold Learning Fault Diagnosis Method Based on Adaptive Domain Selection and Maximum Manifold Edge

Area of Science:

• Mechanical Engineering
• Signal Processing
• Machine Learning

Background:

• Rotating machinery generates nonlinear and non-stationary vibration signals.
• Existing feature sets often contain redundant information, complicating fault diagnosis.
• Traditional manifold learning methods can be sensitive to domain selection and struggle with non-uniform feature distributions.

Purpose of the Study:

• To develop a high-dimensional feature reduction method for rotating machinery fault diagnosis.
• To address the limitations of traditional manifold learning in handling complex vibration data.
• To improve the accuracy of fault identification in rotating machinery.

Main Methods:

• Adaptive neighborhood size selection based on sample density.
• Adaptive construction of between-manifold and within-manifold graphs using class information.
• Calculation of divergence matrices and edge distances for manifold graphs.
• Feature fusion reduction by maximizing edge distance and minimizing within-class differences.

Main Results:

• The proposed multi-manifold learning algorithm demonstrated superior performance on theoretical and real-world rotating machinery fault datasets.
• Achieved higher fault identification accuracy compared to conventional manifold learning methods.
• Effectively reduced feature redundancy while preserving crucial fault information.

Conclusions:

• The proposed adaptive multi-manifold learning method is effective for high-dimensional feature reduction in rotating machinery fault diagnosis.
• This approach offers improved accuracy and robustness over traditional methods.
• It provides a promising direction for intelligent fault diagnosis systems.