Abstract
In view of the non-linear and non-stationary of the rolling element bearing fault signal, the method of mathematical morphology analysis is introduced into the rolling element bearing fault diagnosis. Multi-scale morphological transform is applied to the analysis of the bearing signals. To describe the complexity of pattern spectrum curves by using sample entropy, and its value as the input vector of the neural network is used to realize the fault pattern classification by using the back-propagation (BP) neural network. Experimental results show that this method is effective.
Introduction
Rolling element bearings are critical components in industrial machinery, yet they are frequent points of failure. When a bearing develops a defect, such as a pit or crack, it generates complex vibration signals. These signals are inherently non-linear and non-stationary, meaning their statistical properties fluctuate over time. Traditional signal processing tools, such as the Fourier Transform, often struggle to capture the nuances of these messy vibrations. This research proposes a robust diagnostic pipeline that treats vibration signals as geometric landscapes, utilizing mathematical morphology and sample entropy to extract meaningful "fingerprints" for automated fault classification.
Experimental Methodology
The researchers validated their approach using experimental data from the Key Laboratory of Advanced Manufacturing Technology at the Beijing University of Technology. The test rig utilized SKF 6307 bearings operating at a shaft rotation speed of 1496 rpm with a sampling frequency of 12.8 kHz.
The dataset included four distinct bearing states:
- Normal bearing (no defect)
- Outer race fault
- Inner race fault
- Rolling body (ball) fault
A total of 200 samples were collected, with 100 used for training the neural network (25 per category) and 100 reserved for testing and verification.
Signal Analysis and Feature Extraction
The proposed diagnostic pipeline consists of four primary stages:
- Multi-scale Mathematical Morphology: The vibration signal is analyzed as a geometric shape. Using a "linear structuring element" (a geometric probe), opening and closing operations are performed at varying scales to filter noise and highlight structural features at different resolutions.
- Pattern Spectrum (PS): This technique quantifies the amount of information removed from the signal at each scale, creating a curve that represents the signal's morphological distribution.
- Sample Entropy (SampEn): To compress the data for classification, the Sample Entropy of the pattern spectrum curve is calculated. This single value represents the complexity and regularity of the morphological changes.
- BP Neural Network: These entropy signatures are fed into a Back-Propagation (BP) neural network trained to map specific signatures to their corresponding fault types.
Results and Discussion
A key finding of the study was the shift in "scale signatures" between healthy and faulty bearings. As shown in the pattern spectra, normal bearings typically reached maximum amplitude at a scale of 2, whereas faulty bearings peaked at a scale of 4. This shift suggests that mechanical faults introduce larger-scale geometric irregularities into the vibration signal.
By using Sample Entropy to reduce complex multi-scale curves into a single feature vector, the researchers significantly reduced the computational burden on the neural network while retaining essential diagnostic information. The final classification results were highly accurate:
- Normal Condition: 96% accuracy.
- Rolling Body Fault: 96% accuracy.
- Inner/Outer Race Faults: 92% accuracy.
Conclusion
The combination of mathematical morphology and sample entropy provides a reliable and efficient method for diagnosing bearing health. While race faults proved slightly more challenging to distinguish than rolling body faults, the overall performance remains high. Future research will explore optimizing the shapes of structuring elements and testing the algorithm's performance in high-noise industrial environments, comparing these hand-crafted features against automated deep learning approaches.
Summary updated