Using synthetic data for pretraining partial discharge detection in overhead transmission lines.

Area of Science:

• Electrical Engineering
• Artificial Intelligence
• Power Systems

Background:

• Accurate partial discharge (PD) detection is vital for medium-voltage overhead transmission line maintenance and preventing outages.
• Challenges include limited labeled data and significant electromagnetic interference.
• Existing methods struggle with real-world complexities and data scarcity.

Purpose of the Study:

• To develop and evaluate a hybrid simulation-and-data-driven framework for PD detection.
• To investigate the effectiveness of using synthetically generated PD signals for pretraining deep neural networks.
• To compare the performance of different deep learning architectures (CNN, ViT, LSTM) for this task.

Main Methods:

• A synthetic data generation pipeline was created, varying PD parameters (rate, amplitude, contact, noise) and output formats (time-series, spectrograms).
• Deep neural networks (CNNs, ViT, LSTM) were pretrained on synthetic data and fine-tuned on limited real overhead-line measurements.
• A comprehensive ablation study analyzed model sensitivity to synthetic data parameters and architectural choices.

Main Results:

• CNN-based models significantly outperformed ViT and LSTM on spectrogram-based PD classification.
• Pretraining on synthetic data, especially datasets simulating higher PD activity, improved downstream performance on real data by 10-20% (MCC).
• Poorly aligned synthetic data can hinder generalization, emphasizing the need for accurate noise calibration and domain-aligned simulation.

Conclusions:

• Architectural choice is critical for effective PD detection in overhead lines.
• Well-designed synthetic data serves as a powerful tool for enhancing PD monitoring reliability and cost-effectiveness, particularly when real labeled data is scarce.
• The hybrid approach offers a practical solution for preemptive maintenance in power transmission systems.