Feature Extraction and Classification of Partial Discharge Signals in C₄F₇N-Based Gas Insulated Systems: A Time-Domain and Machine Learning Approach

In the quest for an environmentally friendly alternative to SF<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">6</sub> gas to reduce carbon emissions, the C<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">4</sub>F<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">7</sub>N gas mixture (C<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">4</sub>F<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">7</sub>N/O<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub>/CO<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub>) emerged as a good candidate due to its lower environmental hazards and good dielectric strength. However, these new gas mixtures bring transition-related challenges due to their altered partial discharge (PD) patterns and dynamics, which complicate reliable PD detection and classification in C4F7N-insulated systems. Manufacturing or assembly flaws in gas-insulated equipment can enhance local electric fields, triggering PDs. Identifying PD sources is critical for effective online monitoring of high-voltage equipment, enabling early detection of insulation degradation and preventing potential failures. Historically, phase resolved partial discharge patterns employed for defect identification. However, challenges arise when multiple PD sources active simultaneously, as their patterns may partially overlap, leading to misunderstanding in the identification process. In this research, a machine learning algorithm is developed to classify PD sources using statistical and probabilistic analysis of time-domain parameters of PD pulses. Several PD source topologies are designed including protrusion, metallic particles, and floating electrode arrangements. Various time-domain parameters of PD pulses, along with their peak amplitude, are measured. The model employs Weibull distribution analysis, statistical variables such as kurtosis and skewness, and machine learning approaches to effectively distinguish between various types of PDs. Achieving approximately 98 % accuracy demonstrates the model’s effectiveness in classifying simultaneous PD sources. Moreover, the study confirms that time-domain statistical feature analysis significantly improves the reliability and safety of electrical systems using C<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">4</sub>F<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">7</sub>N gas mixtures.