Partial Discharge Data Enhancement and Pattern Recognition Method Based on a CAE-ACGAN and ResNet

Songyuan Li, Xiaopeng Wang, Yue Han, Junji Feng, Zhen Yin, Jiawang Yang, Weipeng Luo, Jun XieSymmetry 2024 · DOI

TL;DR

This paper tackles the challenge of accurately identifying partial discharge (PD) in electrical equipment, which is hampered by imbalanced datasets and complex deep learning models. They propose a novel method combining a Convolutional Autoencoder-Auxiliary Classifier Generative Adversarial Network (CAE-ACGAN) for data enhancement with a Residual Network (ResNet) for robust pattern recognition, achieving a significant accuracy improvement.

What Problem Does This Paper Address?

Partial discharge (PD) is a critical indicator of insulation degradation in electrical equipment, making its accurate detection and classification vital for preventing failures. Traditional machine learning methods like KNN and SVM struggle with the inherent class imbalance and complexity of PD data. While deep learning, particularly Convolutional Neural Networks (CNNs), offers better feature extraction, their performance is heavily reliant on large, balanced datasets. PD events are often rare and sporadic, leading to highly imbalanced datasets where minority classes are underrepresented. This imbalance severely degrades the accuracy of deep learning models, making reliable PD pattern recognition a significant challenge. Existing data enhancement techniques like image transformation, undersampling, and oversampling have limitations, such as data redundancy or loss of crucial information.

Methods

The core of this paper lies in a two-pronged approach: data enhancement and pattern recognition.

  • CAE-ACGAN for Data Enhancement: The authors introduce a Convolutional Autoencoder-Auxiliary Classifier Generative Adversarial Network (CAE-ACGAN). This model combines the strengths of convolutional autoencoders for feature extraction and reconstruction with the generative power of GANs. The convolutional autoencoder acts as the generator, replacing fully connected layers with convolutional layers for better spatial feature learning. Deep residual units are employed within the latent space to extract robust feature distributions. The CAE-ACGAN is trained to generate high-quality synthetic PD spectrograms, effectively addressing the issue of imbalanced datasets by creating more samples for underrepresented classes. The model incorporates an auxiliary classification loss to guide the generation process towards specific PD types.
  • ResNet for Pattern Recognition: A Residual Network (ResNet) is employed as the classifier. ResNets are known for their ability to train very deep networks without suffering from vanishing gradients, thanks to their residual blocks that allow for adaptive network depth. This enables the network to learn complex PD features effectively while mitigating issues like gradient disappearance and network degradation. The enhanced dataset generated by the CAE-ACGAN is then used to train this ResNet model.

Datasets

The study utilizes a dataset of partial discharge (PD) signals collected from various simulated insulation defects in transformers, including tip discharge, suspension discharge, bubble discharge, and surface discharge.

  • Training Set A: This initial dataset is described as unbalanced, with varying numbers of samples for each discharge type (e.g., 285 for point discharge, 75 for bubble discharge).
  • Training Sets B and C: These are augmented datasets created by the CAE-ACGAN. Set B is balanced with 300 samples per class, and Set C is further augmented with 400 samples per class.
  • Test Set: A consistent test set of 200 samples per class is used across all experiments.

Key Insights and Ideas

  • Synergy of CAE and GAN: The CAE-ACGAN effectively leverages the reconstruction capabilities of autoencoders and the generative power of GANs, leading to high-quality synthetic PD data that closely mimics real-world signals.
  • Residual Networks for Depth and Accuracy: ResNets' architecture, with its residual blocks, proves crucial for handling the complexity of PD patterns and achieving high classification accuracy, even with deeper networks.
  • Addressing Class Imbalance is Paramount: The paper strongly demonstrates that simply using deep learning models on imbalanced data leads to suboptimal performance, particularly for minority classes. Data enhancement is not just beneficial but essential for reliable PD recognition.
  • Quantitative Validation of Generation Quality: The use of SSIM and PSNR metrics provides a robust, quantitative assessment of the generated PD spectrograms, showing the CAE-ACGAN's superiority over other generative models.

Key Results

  • Data Enhancement Quality: The CAE-ACGAN model achieved high SSIM (0.92) and PSNR (45.88 dB) metrics, outperforming other generative models like VAE, GAN, WGAN-GP, and BEGAN.
  • Unbalanced Dataset Performance: Without data enhancement (Training Set A), the ResNet achieved a maximum accuracy of 90.75% (for umbrella type).
  • Enhanced Dataset Performance: After data enhancement with CAE-ACGAN (Training Set C), the proposed method (CAE-ACGAN + ResNet) achieved an impressive 98% accuracy, representing a 7.25% improvement over the pre-enhancement level.
  • Improved Minority Class Recognition: The enhancement significantly boosted the recognition accuracy for bubble discharge (by 14.5% to 21.5%) and surface discharge (by 4% to 11%), which were previously challenging classes.

Limitations and Future Work

  • Parameter Refinement: The authors acknowledge that there is scope for further enhancement by refining parameters and structural elements of the proposed method.
  • Engineering Implementation: Expediting the implementation of the method in practical engineering scenarios is identified as a future direction.
  • Generalizability: While the paper focuses on specific PD types, future work could explore the method's effectiveness on a wider range of PD phenomena and real-world field data.
  • Computational Cost: Although ResNet helps control computational pressure, the training of GANs can still be computationally intensive. Further optimization for faster training could be beneficial.

Summary updated

BibTeX
@article{li2024partial,
  author = {Songyuan Li and Xiaopeng Wang and Yue Han and Junji Feng and Zhen Yin and Jiawang Yang and Weipeng Luo and Jun Xie},
  title = {{Partial Discharge Data Enhancement and Pattern Recognition Method Based on a CAE-ACGAN and ResNet}},
  journal = {Symmetry},
  year = {2024},
  doi = {10.3390/sym17010055},
}
Figure