Manifold Learning Techniques for Machinery Fault Diagnosis

Manifold learning has emerged as a powerful framework for diagnosing mechanical faults by revealing intrinsic low-dimensional structures hidden within high-dimensional sensor data. Machinery fault signals—typically vibration, acoustic or electrical measurements—often lie on complex nonlinear manifolds shaped by underlying system dynamics. Traditional linear methods can fail to separate fault conditions when data distributions are nonlinear or corrupted by noise and outliers. Manifold learning techniques such as Local Linear Embedding, Laplacian Eigenmaps, Isomap, t-SNE and UMAP preserve local and/or global neighbourhood relationships to extract salient features. Kernel extensions and graph-based formulations further enhance robustness by incorporating nonlinear mappings or weighting schemes that suppress spurious artefacts. These approaches yield compact feature representations that improve classification of bearing, gearbox, rotor and valve faults, support real-time monitoring and reduce false alarms in industrial settings.

Research from Nature Portfolio

Recent studies have introduced an L1-norm graph embedding framework that restricts the influence of outliers on the learned manifold. By redefining graph embedding in an L1 space and employing a maximisation strategy, this approach generates low-dimensional mappings that remain stable even when data are contaminated by large-magnitude noise. Quantitative experiments demonstrate that the resulting representations capture more reliable substructures, leading to improved classification accuracy across diverse signal datasets.

Manifold Learning Techniques for Machinery Fault Diagnosis publication trend

The graph below shows the total number of articles in manifold learning techniques for machinery fault diagnosis across all publications each year (not limited to Nature Index journals).

Technical terms

Manifold learning: Set of nonlinear dimension reduction techniques that uncover low-dimensional structures in high-dimensional data.

Local Linear Embedding (LLE): Algorithm preserving local neighbourhood geometry by reconstructing each data point as a weighted combination of its neighbours.

Kernel principal component analysis (KPCA): Extension of PCA using kernel functions to capture nonlinear relationships.

UMAP (Uniform Manifold Approximation and Projection): Graph-based method preserving both local and global data structure through topological analysis.

Kernel Neighbourhood Preserving Embedding (KNPE): Kernel-based embedding that maintains local manifold structures by constructing a neighbourhood graph in feature space.

References

1. Improved Graph Embedding for Robust Recognition with outliers. Scientific Reports (2018).
2. Modified Local Linear Embedding Algorithm for Rolling Element Bearing Fault Diagnosis. Applied Sciences (2017).
3. Dimensionality reduction method of rotor fault data set based on KPCA-KNPE. Journal of Physics Conference Series (2022).
4. Analysis of UMAP, the method for reducing the dimensionality of initial data in machine learning for the purpose of failure prediction in a motive power service. Dependability (2022).