Graph-based fault diagnosis for rotating machinery: Adaptive segmentation and structural feature integration
Rotating machinery, such as bearings and gearboxes, plays a critical role in industrial operations; however, faults in these components can lead to costly downtime and inefficient maintenance. Traditional fault diagnosis methods often rely on manual expertise or deep learning models that require hig...
Saved in:
| Main Author: | |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-09-01
|
| Series: | Results in Engineering |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S2590123025026350 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Rotating machinery, such as bearings and gearboxes, plays a critical role in industrial operations; however, faults in these components can lead to costly downtime and inefficient maintenance. Traditional fault diagnosis methods often rely on manual expertise or deep learning models that require high computational resources and lack transparency, limiting their practicality in real-world settings. To address these, this study introduces a novel graph-based framework for fault diagnosis that emphasizes interpretability, efficiency, and robustness. Vibration signals are adaptively segmented using entropy optimization to focus on informative fault-related bursts, and then transformed into time-frequency representations. Each segment is modeled as a graph, extracting structural features such as average shortest path length, modularity, and spectral gap. These are combined with statistical signal descriptors to train logistic regression, support vector machines, and random forests. Evaluated on the Case Western Reserve University (CWRU) bearing dataset across 0–3 HP loads and the Southeast University (SU) gearbox dataset under varied load conditions, the framework outperformed traditional and deep learning approaches. The random forest classifier achieved up to 99.9% accuracy on the CWRU and 100% on the SU datasets. Even with added noise (deviation of 0.5), it maintains 89.3% and 98.1% accuracy on the CWRU and SU datasets, respectively. In cross-domain scenarios, it delivered F1-scores of 97.73% for cross-load and 97.99% for cross-fault transfers, showcasing strong generalizability. Combining graph-theoretic analysis with statistical signal processing eliminates the need for resource-intensive models, making it ideal for industrial fault diagnosis and predictive maintenance applications. |
|---|---|
| ISSN: | 2590-1230 |