SF6 High-Voltage Circuit Breaker Contact Status Detection at Different Currents

Ze Guo, Linjing Li, Weimeng Han, Zixuan Guo2022 · DOI

TL;DR

This paper introduces a novel non-destructive method for detecting the contact status of SF6 high-voltage circuit breakers (HVCBs) by analyzing their vibration signals. By employing a deep neural network optimized with Bayesian optimization, the system achieves over 96% accuracy in identifying five distinct contact states across a wide range of operating currents, offering a promising advancement for predictive maintenance.

What Problem Does This Paper Address?

High-voltage circuit breakers (HVCBs) are critical components in power grids, and their reliable operation depends heavily on the condition of their contacts. Over time, these contacts can degrade due to mechanical stress and arcing, leading to increased contact resistance and potential failures. Current methods for assessing contact status are often destructive or require the equipment to be taken offline, making them impractical for continuous monitoring. Existing non-destructive testing (NDT) methods for in-service HVCBs are limited, especially when considering the varying operating currents. This paper aims to bridge this gap by developing an online, non-destructive method to accurately assess HVCB contact status under different current conditions using vibration analysis.

Methods

The core of this research lies in a multi-stage approach:

  • Vibration Signal Acquisition: A 40.5-kV SF6 HVCB prototype was used, equipped with an experimental platform featuring a high-current generator and acceleration sensors. Vibration signals were collected under various currents (1000 A to 3500 A) and for five distinct contact states (from good contact to extremely severe defects).
  • Feature Extraction: The collected vibration signals were preprocessed to remove noise and "zero drift." Fourier decomposition was then applied, and the amplitudes of signals at frequencies near integer multiples of 50 Hz (specifically, 50 Hz was chosen as the primary eigenvalue) were extracted to form an eigenvalue matrix. This matrix captures the mechanical vibration characteristics.
  • Deep Neural Network (DNN) for Classification: A DNN was designed to classify the five contact states. The network architecture included an input layer (receiving the eigenvalue matrix), hidden layers, and an output layer representing the classification results.
  • Bayesian Optimization (BO) for Hyperparameter Tuning: To overcome the limitations of manual hyperparameter tuning (which is time-consuming and often suboptimal), Bayesian optimization was employed. BO efficiently searches for optimal hyperparameters (learning rate, number of hidden layer nodes, number of hidden layers, and dropout rate) for the DNN, leading to a more robust and generalized model. The Adam optimizer was used for training the DNN.

Datasets

The study utilized experimental data collected from a 40.5-kV SF6 HVCB prototype. The dataset comprises vibration signals recorded under:

  • Five distinct contact states: Characterized by varying contact resistances (48 μΩ for good contact, 78 μΩ, 93 μΩ, 117 μΩ, and 150 μΩ for progressively poorer contact).
  • Six different current levels: Ranging from 1000 A to 3500 A, with increments of 500 A.

This resulted in 30 sets of experimental data, each with a 45-second vibration signal sampled at 51.2 kHz. The data was then divided into training (70%), validation (20%), and testing (10%) sets for model development and evaluation.

Key Insights and Ideas

  • Vibration as a Diagnostic Indicator: The research confirms that mechanical vibrations generated by electromagnetic forces within HVCBs are directly correlated with the contact status and operating current. The dominant vibration frequency observed was around 100 Hz, consistent with theoretical predictions of twice the power supply frequency.
  • Frequency-Based Feature Extraction is Crucial: While time-domain signals showed some variation, the frequency domain analysis, particularly focusing on specific frequency bands (like 50 Hz and its multiples), proved more effective in capturing the subtle changes indicative of contact degradation.
  • The Power of Hyperparameter Optimization: The study highlights the significant impact of hyperparameter optimization. Using Bayesian optimization dramatically improved recognition accuracy compared to fixed or manually tuned parameters, demonstrating its value for complex machine learning models in industrial diagnostics.
  • DNNs Excel at Complex Pattern Recognition: Deep neural networks, when properly trained and optimized, can effectively learn the intricate relationships between vibration signal features and HVCB contact states, even under varying current conditions.

Key Results

  • The proposed recognition model achieved an accuracy rate above 96% for identifying five different contact states of the HVCB across currents ranging from 1000 A to 3500 A.
  • In specific current ranges (e.g., 2000 A to 3000 A), the model demonstrated exceptionally high accuracy, with some instances reaching 100%.
  • The study successfully demonstrated that vibration signal amplitudes increase non-linearly with increasing current.
  • A noticeable increase in the 300-Hz signal amplitude was observed when the HVCB contact was in a poor state.

Limitations and Future Work

  • Limited to SF6 HVCBs: The current study focuses specifically on SF6 HVCBs. While SF6 is prevalent, future work could explore the applicability of this method to HVCBs using alternative quenching media.
  • Experimental Setup: The experiments were conducted on a prototype. Validation on a wider range of real-world HVCB models and under more diverse environmental conditions would strengthen the findings.
  • Feature Selection: While 50 Hz was chosen as a primary eigenvalue, further research could explore more sophisticated feature selection techniques or combinations of features to potentially enhance accuracy and robustness.
  • Real-time Implementation: The paper lays the groundwork for real-time monitoring. Future work could focus on optimizing the computational efficiency of the model for seamless integration into existing monitoring systems.
  • Broader Fault Diagnosis: This work specifically targets contact status. Expanding the diagnostic capabilities to include other potential HVCB faults (e.g., mechanical wear in other components, insulation issues) would be a valuable next step.

Summary updated

BibTeX
@misc{guo2022sf,
  author = {Ze Guo and Linjing Li and Weimeng Han and Zixuan Guo},
  title = {{SF6 High-Voltage Circuit Breaker Contact Status Detection at Different Currents}},
  year = {2022},
  doi = {10.3390/s22218490},
}
Figure