A Novel Transformers Fault Diagnosis Method Based on Probabilistic Neural Network and Bio-Inspired Optimizer

Lingyu Tao, Xiaohui Yang, Yichen Zhou, Li Yang2021 · DOI

TL;DR

This paper introduces a novel fault diagnosis method for power transformers by combining a Probabilistic Neural Network (PNN) with an improved bio-inspired optimization algorithm (ISSA). This approach effectively optimizes the PNN's smoothing factor, leading to significantly improved accuracy and robustness in identifying transformer faults compared to existing methods.

What Problem Does This Paper Address?

Power transformers are critical components in electrical grids, and their failure can lead to substantial economic losses and power outages. Traditional fault diagnosis methods, particularly those relying on Dissolved Gas Analysis (DGA), often struggle with diagnostic accuracy and stability, especially in complex scenarios. While artificial intelligence methods have shown promise, there's a continuous need for more reliable and accurate fault diagnosis systems that can handle the intricacies of transformer operational data. This paper aims to bridge this gap by developing a more robust and accurate fault diagnosis model for oil-immersed power transformers.

Methods

The core of this work lies in the integration of two key components:

  • Probabilistic Neural Network (PNN): PNNs are chosen as the base classifier due to their strong fault tolerance, fast training, and ability to converge to a Bayesian classifier with sufficient data. They are particularly well-suited for multi-classification problems.
  • Improved Salp Swarm Algorithm (ISSA): The PNN's performance is highly sensitive to its smoothing factor ($\sigma$). To optimize this parameter, the authors propose ISSA, an enhanced version of the Salp Swarm Algorithm (SSA). ISSA incorporates the Sine Cosine Algorithm (SCA) to improve exploration capabilities and a disruption operator (Dop) to increase population diversity, thereby enhancing the algorithm's ability to avoid local optima and achieve faster convergence.

The ISSA algorithm is used to find the optimal smoothing factor for the PNN, creating the ISSA-PNN model. This model is then trained and tested using pre-processed DGA data.

Datasets

The study utilizes dissolved gas analysis (DGA) data collected from oil-immersed power transformers. This data was sourced from Jiangxi Province Power Supply Company and transformer factories. After pre-processing and filtering using the IEC three-ratio method, 555 valid characteristic gas data samples were obtained, categorized into four fault types:

  • Low Temperature Overheating (LT) (<150 °C)
  • Low Temperature Overheating (LT) (150–300 °C)
  • Partial Discharge (PD)
  • Arc Discharge (AD)

A random selection of 80% of these samples (444 sets) was used for training, and the remaining 20% (111 sets) were used for testing.

Key Insights and Ideas

  • The critical role of the PNN's smoothing factor: The paper highlights that the PNN's classification performance is highly dependent on the smoothing factor ($\sigma$). An inappropriate value can lead to rapid convergence and significantly reduced accuracy.
  • Synergy of bio-inspired optimization and neural networks: The combination of ISSA and PNN demonstrates a powerful approach to tackle complex classification problems. ISSA's ability to efficiently search for optimal parameters complements PNN's classification strengths.
  • ISSA's enhanced exploration and exploitation: The introduction of SCA and the disruption operator in ISSA significantly improves its ability to explore the search space and avoid getting stuck in local optima, leading to better parameter optimization for the PNN.
  • Robustness against overfitting: The ISSA-PNN model shows strong robustness, performing well on test data without significant degradation from training performance, indicating it doesn't easily fall into overfitting.

Key Results

The ISSA-PNN model demonstrated superior performance across various metrics:

  • Average Accuracy: Achieved an impressive average accuracy of 99.65%, outperforming other optimized PNN methods (SSA-PNN: 97.37%, MVO-PNN: 97.02%, BA-PNN: 96.52%, SOA-PNN: 95.80%, PSO-PNN: 94.49%) and the traditional PNN (86.70%).
  • Specific Fault Accuracy: Showed high accuracy for all fault types, including 100% accuracy for LT (150-300 °C), PD, and AD faults.
  • F1-Score: The ISSA-PNN model achieved the highest Marco F1-score of 99.18%, indicating excellent precision and recall.
  • MSE: The Mean Squared Error (MSE) of training for ISSA-PNN was 0.00000, and the MSE of testing was 0.08108, both being the lowest among the compared methods.
  • Convergence: The ISSA algorithm demonstrated rapid and efficient convergence, quickly finding the global optimum and showing better stability and anti-interference compared to other optimization algorithms.

Limitations and Future Work

  • Limitations: While the ISSA-PNN model shows excellent performance, it was noted that for LT (<150 °C) faults, its accuracy (98.59%) was slightly inferior to BA-BP and GA-BP (both 99.06%). However, the overall performance across all fault types and the average accuracy still favored ISSA-PNN.
  • Future Work: The paper does not explicitly detail future work directions. However, the success of this hybrid approach suggests potential for its application to other complex diagnostic problems or the exploration of further enhancements to the ISSA algorithm or PNN architecture. The authors also mention the need to meet practical engineering requirements, implying that further validation in real-world industrial settings could be a future step.

Summary updated

BibTeX
@misc{tao2021novel,
  author = {Lingyu Tao and Xiaohui Yang and Yichen Zhou and Li Yang},
  title = {{A Novel Transformers Fault Diagnosis Method Based on Probabilistic Neural Network and Bio-Inspired Optimizer}},
  year = {2021},
  doi = {10.3390/s21113623},
}
Figure