TL;DR
This paper tackles the critical challenge of accurately identifying partial discharge (PD) faults in electrical switchgear, a key factor in preventing power system failures. The authors propose a novel approach that combines Principal Component Analysis (PCA) for feature reduction with a Back Propagation Neural Network (BPNN) optimized by an improved Mantis Search Algorithm (MSA). This enhanced method demonstrates superior recognition accuracy and efficiency compared to existing techniques.
What Problem Does This Paper Address?
Switchgear is a vital component of power distribution networks, and its reliable operation is paramount. Partial discharges (PDs) are a significant threat to the insulation integrity of switchgear, potentially leading to catastrophic equipment failure and power outages. Accurately identifying and classifying different types of PDs (e.g., metal particle discharge, suspended discharge, creeping discharge) is crucial for early detection of insulation defects and timely maintenance. Traditional PD pattern recognition (PDPR) methods often struggle with issues like sensitivity to initial parameter values, slow convergence, and the tendency to get stuck in local optima, especially with the inherent randomness of PD phenomena. This paper aims to overcome these limitations by developing a more robust and efficient PDPR system.
Methods
The core of this research lies in a multi-stage approach:
- Data Acquisition and Feature Extraction: Ultrasonic signals from simulated PD events are collected. These signals are then processed to generate Phase Resolved Partial Discharge (PRPD) spectrograms. From these spectrograms, 14 statistical feature parameters are extracted, capturing characteristics like skewness, kurtosis, local peaks, and cross-correlation coefficients.
- Dimensionality Reduction with PCA: To address the computational burden and potential redundancy of the 14 features, Principal Component Analysis (PCA) is employed. PCA reduces the feature space to a more manageable set of key principal components (specifically, 7 components were found to be sufficient while retaining most of the variance).
- BPNN Optimization with Improved Mantis Search Algorithm (MSA): A Back Propagation Neural Network (BPNN) is used as the classifier. However, standard BPNNs are sensitive to initial parameter values and can converge slowly. To mitigate this, the authors introduce an "improved Mantis Search Algorithm" (MSA). This improved MSA features:
- Adaptive Parameter Updates: The probabilities 'p' (exploration vs. exploitation) and 'Pc' (sexual cannibalism) are dynamically adjusted based on the evolution stage, promoting better exploration early on and faster convergence later.
- Boundary Handling: A novel mechanism is implemented to keep the mantis positions within the feasible search space, preventing the algorithm from jumping out of bounds and improving its efficiency.
- Model Training and Recognition: The reduced set of 7 principal components serves as input to the BPNN. The improved MSA is used to globally search for and optimize the initial weights and biases of the BPNN. The trained BPNN model then performs the partial discharge pattern recognition.
Datasets
The study utilizes a dataset generated from a custom-built PD defect simulation test device for switchgear. This device simulates three types of PD: metal particle discharge, suspended discharge, and creeping discharge. The paper does not mention the use of any publicly available benchmark datasets for PDPR.
Key Insights and Ideas
- PCA as a Feature Selector: The paper effectively demonstrates that PCA can significantly reduce the number of input features for the BPNN without a substantial loss in recognition accuracy. This is crucial for improving training speed and reducing computational load.
- MSA for BPNN Optimization: The core innovation is the application of an improved MSA to optimize BPNN parameters. This addresses the common BPNN limitations of sensitivity to initial values and local optima, leading to better performance.
- Adaptive Parameters in MSA: Dynamically adjusting the exploration/exploitation balance and sexual cannibalism probability in the MSA is key to its improved performance, allowing it to navigate the search space more effectively.
- Boundary Handling for Robustness: The boundary handling mechanism in the improved MSA is vital for ensuring that the optimization process remains within valid parameter ranges, preventing erratic behavior and improving convergence.
- Trade-off between Features and Accuracy: The study highlights that while using more features (14) might slightly increase training accuracy, using a reduced set (7) after PCA can achieve comparable or even better testing accuracy with significantly fewer network parameters and faster runtime.
Key Results
- Improved Recognition Accuracy: The proposed Improved MSA-BPNN achieved higher recognition accuracies compared to the original MSA-BPNN and other benchmark classifiers (Decision Tree, KNN, SVM).
- With 7 features, the Improved MSA-BPNN achieved test accuracies of 92.59% (metal particle), 64.39% (suspended), and 90.00% (creeping).
- With 14 features, the Improved MSA-BPNN achieved test accuracies of 94.44% (metal particle), 62.88% (suspended), and 90.00% (creeping).
- Reduced Complexity and Faster Training: Using 7 features (after PCA) resulted in a BPNN with 113 parameters, compared to 183 parameters when using 14 features. This leads to reduced computational load and faster training times.
- Faster Convergence: The improved MSA demonstrated faster convergence and higher efficiency in optimizing the BPNN compared to the original MSA, as evidenced by the lower RMSE values in Figure 8.
- Effectiveness of PCA: The paper shows that PCA effectively reduces the feature set while maintaining high recognition accuracy, particularly for the suspended discharge type.
Limitations and Future Work
- Limited Dataset: The study relies on a dataset generated from a simulated test device. While this allows for controlled experiments, it may not fully capture the complexity and variability of real-world PD signals.
- Focus on Ultrasonic Signals: The current algorithm is primarily designed for ultrasonic detection signals. The authors suggest that future work could involve applying the algorithm to recognize PD patterns using signals from other sensor types to further augment recognition precision.
- Generalizability: While the algorithm is presented for switchgear, its applicability to other electrical equipment with different insulation types or operating conditions would require further validation.
- Real-world Validation: The paper focuses on experimental results from a simulated setup. Future work should involve extensive validation on actual field data from operational switchgear.
Summary updated