Fault diagnosis method using supervised extended local tangent space alignment for dimension reduction

• Use multiple-domain features to construct high-dimensional fault sample.
• Propose a novel supervised manifold learning method for dimension reduction.
• Introduce iterative new sample embedding algorithm for new sample embedding.
• Verify the effectiveness of the proposed method by gearbox fault diagnosis.

A method of fault diagnosis that uses supervised extended local tangent space alignment (SE-LTSA) for dimension reduction is proposed to improve the effectiveness of fault diagnosis in machinery. Fault diagnosis is essentially a pattern recognition problem, and a key role in the process is feature extraction. First of all, multiple-domain fault features are extracted from vibration signals which comprehensively characterize the properties of the fault(s) in the machinery. Then, SE-LTSA is employed as the dimension reduction method to compress the multiple-domain fault features into low-dimensional eigenvectors. The proposed SE-LTSA method not only provides a good approximation of the nonlinear structure of the high-dimensional fault samples, but also maximizes the interclass dissimilarity. It achieves this by integrating class label information within the dimension reduction process. Finally, the low-dimensional eigenvectors are inputted to a classifier to recognize faults. The novel method was applied to diagnose the faults in a gearbox in order to verify its effectiveness. The experimental results indicate that dimension reduction using the proposed SE-LTSA method can reveal more sensitive fault features. The fewer, yet more sensitively detected, fault features significantly improves the accuracy of fault diagnosis.