TL;DR
This paper introduces a novel multidimensional feature extraction method for hydroelectric units, combining traditional time-frequency analysis with advanced deep learning techniques. By fusing IMF energy moments and Stacked Denoising Autoencoder (SDAE) features, the method significantly improves the accuracy of identifying equipment operating states, paving the way for more robust fault diagnosis.
What Problem Does This Paper Address?
Hydroelectric units are critical for power generation, and their stable operation is paramount. However, monitoring their health is challenging due to complex operating conditions, including water-machine-electricity coupling, which introduce nonlinear and non-stationary characteristics into vibration signals. Traditional methods relying on single features (time-domain or frequency-domain) struggle to capture the full picture of the unit's state, leading to difficulties in accurate fault diagnosis and prediction. This paper aims to bridge this gap by developing a more comprehensive feature extraction approach.
Methods
The proposed method integrates several techniques:
- Time-Domain and Frequency-Domain Analysis: Standard statistical features like standard deviation, peak-to-peak value, and kurtosis are extracted, along with frequency-domain features such as mean frequency amplitude and frequency centroid value.
- Complementary Ensemble Empirical Mode Decomposition (CEEMD) and IMF Energy Moment: CEEMD is used to decompose the vibration signal into intrinsic mode functions (IMFs). The energy moment of each IMF is then calculated. This approach captures time-frequency characteristics and helps overcome issues like modal aliasing inherent in simpler EMD methods.
- Stacked Denoising Autoencoder (SDAE): SDAE, an unsupervised deep learning model, is employed to adaptively learn deep, abstract features from the vibration signals. This overcomes the limitations of shallow feature extraction that often relies on manual experience.
- Multidimensional Information Fusion: The extracted features from the above methods are combined into a single, multidimensional feature vector. This fused vector is then used for state recognition.
Datasets
The method was validated using axial vibration data from the No. 3 axial flow paddle unit of the SK power station in China. The dataset comprised 100 sets of normal sample data and 100 sets of fault samples, each with a sampling length of 4096.
Key Insights and Ideas
- Limitations of Traditional Features: The study clearly demonstrates that relying solely on traditional time-domain and frequency-domain features leads to significant inaccuracies and randomness in state identification due to overlaps and limited discriminative power.
- CEEMD's Advantage: CEEMD effectively decomposes complex signals and its IMF energy moments provide valuable insights into signal energy distribution over time, complementing traditional features.
- SDAE for Deep Feature Learning: SDAE proves adept at learning abstract, latent features that effectively represent equipment states, reducing data redundancy and improving generalization.
- Synergy of Fusion: The true power lies in the fusion of these diverse features. Combining time-domain, frequency-domain, IMF energy moments, and SDAE deep features creates a richer representation that captures multidimensional complementarity, leading to superior state recognition.
Key Results
The proposed multidimensional information fusion feature extraction method, when used with a Backpropagation Neural Network (BPNN), achieved a perfect 100% recognition accuracy for both normal and fault states. This represents a significant improvement over:
- Traditional time-domain and frequency-domain features alone (76.67% accuracy).
- SDAE features alone (96.67% accuracy).
The fusion method improved accuracy by 23.33% and 3.33% respectively compared to these individual approaches. Similar improvements were observed when using a Support Vector Machine (SVM) classifier, with the fused features achieving 91.67% accuracy compared to 83.3% for traditional features and 100% for SDAE features (though the paper's BPNN results are more definitive for the fused approach).
Limitations and Future Work
- Limitations: The paper acknowledges that while SDAE can extract deep features, its shallow structure might limit its expressive ability for extremely complex data. The study primarily focuses on axial vibration data.
- Future Work: The authors suggest integrating various signals, such as vibration, acoustics, and temperature, to comprehensively capture equipment operating states and further enhance the accuracy and robustness of fault diagnosis. This aligns with the growing trend of multi-modal sensing in industrial IoT.
Summary updated