Abstract
Partial discharge (PD) is a primary indicator of insulation degradation in power transformers. Timely identification of PD types is crucial for grid reliability. This paper presents a novel recognition algorithm, GLCMOP-SVM, which utilizes a self-adaptive Gray-Level Co-Occurrence Matrix of Optimal Parameters (GLCMOP) and a Support Vector Machine (SVM). By automating parameter selection through the Proportion of Off-Diagonal Elements (PODE), the method achieves high accuracy (94.6%) with minimal computational overhead, making it ideal for real-time monitoring.
Introduction
Power transformers are critical components of electrical infrastructure. Their failure often stems from insulation breakdown, preceded by Partial Discharge (PD)—localized electrical sparks. Engineers typically analyze these events using Phase-Resolved Partial Discharge (PRPD) maps. However, classifying the specific type of discharge (e.g., tip, surface, or air gap) is challenging due to the complexity of the patterns and the limitations of traditional feature extraction methods.
Standard Gray-Level Co-Occurrence Matrix (GLCM) techniques often require manual, subjective selection of parameters like the angle-offset, which can lead to inconsistent results. Meanwhile, deep learning approaches like Artificial Neural Networks (ANNs) demand extensive training data and significant computational resources, which are often unavailable for edge-based monitoring systems.
Methodology: The GLCMOP-SVM Approach
The proposed GLCMOP-SVM pipeline automates feature extraction and classification to improve both speed and accuracy.
1. Automated Parameter Selection (GLCMOP)
The core innovation is the Proportion of Off-Diagonal Elements (PODE). In a GLCM, off-diagonal elements represent transitions between different gray levels—essentially the boundaries of PD pulses against the background noise in a PRPD map. By calculating the PODE for various angles, the algorithm automatically selects the angle that maximizes this proportion, ensuring the most informative features are extracted without human intervention.
2. Feature Extraction and Dimensionality Reduction
The system initially extracts nine statistical features from the GLCMOP, including Energy, Entropy, and Correlation. To optimize for real-time performance, the authors applied a "distinguishability screening" process. By evaluating the intersection of probability distributions ($J_b$), they identified the six most discriminative features, reducing the feature space and improving the model's generalization ability.
3. Classification via SVM
A One-vs-Rest Support Vector Machine (OVR-SVM) was chosen for classification. SVMs are particularly effective for smaller datasets and offer robust performance in high-dimensional feature spaces compared to more complex neural networks.
Experimental Setup and Datasets
To validate the methodology, a laboratory test platform was constructed to simulate common transformer insulation defects.
The researchers collected 270 groups of PD data, evenly split between three types:
- Tip Discharge: Simulated using a needle-plate electrode.
- Air Discharge: Simulated using an epoxy resin board with an internal air gap.
- Surface Discharge: Simulated using a cylinder-plate electrode configuration.
Data was captured via ultra-high frequency (UHF) sensors and converted into $256 \times 256$ grayscale PRPD maps for analysis.
Results and Discussion
The GLCMOP-SVM method demonstrated superior performance compared to traditional GLCM-SVM, XGBoost, and ANN models:
- Recognition Accuracy: The proposed method achieved an overall accuracy of 94.6%, significantly higher than the ANN (87.1%) and XGBoost (83.3%).
- Computational Efficiency: Training was completed in just 2 seconds, five times faster than standard GLCM-SVM. The recognition time was also the lowest at 2.2 seconds.
- Resource Footprint: Memory consumption was remarkably low at 2.1%, compared to 7.9% for ANN, highlighting its suitability for integration into portable monitoring devices.
Conclusion
The GLCMOP-SVM algorithm provides a robust, efficient, and objective solution for transformer PD pattern recognition. By automating parameter selection and streamlining the feature space, it overcomes the limitations of manual feature engineering and the high resource demands of deep learning. Future work will focus on testing the algorithm against real-world field noise and expanding the classification library to include more complex fault types.
Summary updated