Faults Diagnosis of Railway Bearing Based on FIR-wavelet Packet and LVQ Neural Network

Yang Jianwei, Dechen Yao, Xi Li, Limin Jia, Yong QinThe Open Automation and Control Systems Journal 2015 · DOI

Abstract

In this paper, we presented a way for railway bearing fault diagnosis with the use of FIR-wavelet packet and LVQ neural network. First, the original vibration signal of trains' rolling bearing is denoised based on FIR. Then, the signals after de-noised are preprocessed by wavelet packet and the wavelet packet energy eigenvector is reconstructed. Those kinds of wavelet packet energy eigenvectors are used to train LVQ neural network. Finally, the intelligent fault diagnosis is realized. The result shows that this approach is effective to distinguish this kind of rolling bearing faults. This method has important practical value.

Introduction

Railway bearings are critical mechanical components that operate under extreme stress and high speeds. Their failure can lead to catastrophic accidents, making reliable fault diagnosis essential for rail safety. Traditionally, engineers monitor these bearings using vibration analysis. However, real-world diagnostic signals are often obscured by massive mechanical noise from the train's operation. Furthermore, bearing faults typically generate transient, non-stationary signals that traditional Fourier-based analysis—which assumes signal consistency over time—struggles to capture accurately.

This paper proposes a robust diagnostic pipeline that combines Finite Impulse Response (FIR) filtering, Wavelet Packet (WP) decomposition, and Learning Vector Quantization (LVQ) neural networks to effectively "hear" specific faults within noisy environments.

Methodology

The proposed diagnostic system follows a three-stage process: denoising, feature extraction, and classification.

1. Signal Denoising via FIR Filtering

The raw vibration signal is first processed through a Finite Impulse Response (FIR) filter. FIR filters are selected for their inherent stability and linear phase response. This ensures that the timing of vibration peaks—which is critical for identifying the periodicity of faults—is preserved while high-frequency background noise is suppressed.

2. Feature Extraction via Wavelet Packet Analysis

Unlike standard Wavelet Transforms that primarily decompose low-frequency components, Wavelet Packet (WP) decomposition breaks down both low and high frequencies across multiple layers. In this study, a three-layer decomposition is used, resulting in eight distinct frequency bands.

Three-layer wavelet packet decomposition tree used for feature extraction.

By calculating the energy eigenvector for these eight bands, the system creates a unique "fingerprint" for different mechanical states. This distribution of energy across frequency sub-bands captures resonance modulation components that are often the earliest indicators of bearing fatigue.

3. Classification with LVQ Neural Networks

The 8-dimensional energy vectors are fed into a Learning Vector Quantization (LVQ) neural network. LVQ is a competitive learning algorithm that excels at pattern classification by mapping input vectors to specific "prototype" categories.

Structure of the Learning Vector Quantization (LVQ) neural network used for fault classification.

The network is designed to classify signals into four categories: Normal, Outer Ring Fault, Inner Ring Fault, and Rolling Body Fault. LVQ was chosen over standard Back-Propagation (BP) networks due to its efficiency in training and its intuitive approach to nearest-neighbor style classification.

Results and Discussion

The methodology was validated using experimental vibration data. The energy eigenvectors demonstrated clear, repeatable differences between fault types. For instance, "Normal" signals exhibited high energy in the first frequency band, while "Outer Ring Faults" showed significant energy shifts to the third and seventh bands.

In the testing phase, the LVQ network achieved 100% diagnostic accuracy, correctly identifying all test cases across the four mechanical states. This suggests that the energy distribution across wavelet packets is a highly sensitive and reliable indicator of bearing health.

Conclusion

The combination of FIR filtering, Wavelet Packet analysis, and LVQ neural networks provides a powerful framework for automated railway bearing diagnosis. While the initial results are promising, future work should focus on testing the model's robustness against the higher variance found in real-world operations, such as varying train speeds and compound faults. Nevertheless, the high precision of this energy-based fingerprinting method offers significant practical value for predictive maintenance in the railway industry.

Summary updated

Abstract

In this paper, we presented a way for railway bearing fault diagnosis with the use of FIR-wavelet packet and LVQ neural network. First, the original vibration signal of trains' rolling bearing is denoised based on FIR. Then, the signals after de-noised are preprocessed by wavelet packet and the wavelet packet energy eigenvector is reconstructed. Those kinds of wavelet packet energy eigenvectors are used to train LVQ neural network. Finally, the intelligent fault diagnosis is realized. The result shows that this approach is effective to distinguish this kind of rolling bearing faults. This method has important practical value.
BibTeX
@article{jianwei2015faults,
  author = {Yang Jianwei and Dechen Yao and Xi Li and Limin Jia and Yong Qin},
  title = {{Faults Diagnosis of Railway Bearing Based on FIR-wavelet Packet and LVQ Neural Network}},
  journal = {The Open Automation and Control Systems Journal},
  year = {2015},
  doi = {10.2174/1874444301507010303},
}
Figure