TL;DR
This paper introduces a novel WPT-SVM approach for diagnosing multiple faults in vertical Essential Service Water (SEC) pumps, achieving a high accuracy of 94%. By combining wavelet packet transform for feature extraction and support vector machines for classification, the method effectively distinguishes between normal operation and six common fault types, outperforming a WPT-ANN approach.
What Problem Does This Paper Address?
Nuclear power plants rely heavily on Essential Service Water (SEC) pumps to maintain critical cooling systems. The operational reliability of these pumps is paramount for safety. While fault diagnosis in rotating machinery is a well-established field, accurately identifying multiple simultaneous faults in SEC pumps, which have a distinct vertical structure, remains a challenge. Existing methods often focus on distinguishing between normal and single-fault conditions, and may not be optimized for the specific vibration characteristics of SEC pumps. This paper aims to bridge this gap by developing a robust multi-fault diagnostic model tailored for these pumps.
Methods
The core of this research lies in a two-pronged approach:
- Wavelet Packet Transform (WPT): To handle the non-stationary nature of vibration signals from SEC pumps, WPT is employed. This technique decomposes the signal into multiple frequency sub-bands, allowing for a more precise analysis of local time-frequency features. The paper utilizes a four-layer WPT to decompose the signal and then reconstructs a new signal using specific sub-bands (C1, C2, C8, C10, C11, and C15) that were found to best represent the energy distribution across different fault conditions.
- Support Vector Machine (SVM): The reconstructed signals are then processed to extract twelve frequency-domain features (e.g., spectrum mean, kurtosis, standard deviation). These features are fed into a Support Vector Machine (SVM) classifier. SVM is chosen for its effectiveness in handling high-dimensional data and its robust generalization capabilities, especially with limited sample sizes, which is common in fault diagnosis scenarios. The study specifically uses a "one-to-one" strategy for multi-class classification.
For comparison, the researchers also implemented a WPT-Artificial Neural Network (ANN) approach.
Datasets
The study utilized a custom-built experimental bench designed to replicate the operational conditions of SEC pumps. This bench was used to generate data for seven distinct states:
- Normal state
- Misalignment
- Blade damage
- Ring wear
- Base instability
- Shaft bending
- Mass eccentricity
A total of 160 training samples and 40 testing samples were collected for each of the seven states, resulting in a dataset of 1120 training samples and 280 testing samples. This dataset was specifically created for this research to address the unique characteristics of SEC pumps.
Key Insights and Ideas
- Energy Distribution Shifts with Faults: The research highlights that different fault conditions significantly alter the energy distribution within the vibration signals. Specifically, while normal operation shows energy concentrated in the low-frequency range, various faults cause this energy to shift towards the mid- and high-frequency regions. This observation is crucial for effective feature extraction.
- WPT's Power in Feature Extraction: The study demonstrates that WPT is highly effective in isolating and analyzing specific frequency bands that are most indicative of fault conditions. The selection of specific sub-bands (C1, C2, C8, C10, C11, and C15) for signal reconstruction proved to be optimal, showcasing the power of WPT in capturing the distinctive features of SEC pumps.
- SVM's Advantage in Multi-Fault Diagnosis: The comparison between WPT-SVM and WPT-ANN reveals that SVM, with its inherent ability to handle non-linear relationships and create clear decision boundaries, is particularly well-suited for multi-category fault classification, especially when dealing with limited data. This helps in preventing overfitting and achieving better generalization.
Key Results
The proposed WPT-SVM method achieved a superior average accuracy of 94% in diagnosing multiple faults in SEC pumps. This represents a 5% increase in accuracy compared to the WPT-ANN method, which achieved an average accuracy of 89%. The WPT-SVM approach also demonstrated greater stability and less variation across different cross-validation folds. Confusion matrices further illustrated that the WPT-SVM method had fewer misclassifications, particularly for the "blade breakage" and "misalignment" fault states, compared to the WPT-ANN method.
Limitations and Future Work
The authors acknowledge that while their WPT-SVM approach shows significant promise, the effectiveness of ANN models is highly dependent on meticulous network architecture design and parameter tuning. This suggests that further optimization of ANN architectures could potentially improve their performance. Future work could explore more advanced feature extraction techniques or investigate the application of deep learning models for even more sophisticated fault diagnosis. Additionally, while the study focused on specific fault types, real-world scenarios might involve more complex or combined fault conditions that could be explored in future research.
Summary updated