Sweep Frequency Response Analysis (SFRA) plays a pivotal role in evaluating the health and performance of Transformers and Reactors. Traditionally, SFRA signature analysis has been subjective, introducing human bias and lacking a standardized interpretation methodology. This paper introduces an AI-driven objective analysis framework, employing techniques such as XML parsing, regex, linear interpolation, wavelet transform, distance algorithms, correlation analysis, and machine learning algorithms like Support Vector Machines (SVM), Random Forest with Ensemble Techniques.
The SFRA analysis pipeline begins with XML parsing to extract magnitude and phase data, addressing non-standardized data using regex for naming conventions. Linear interpolation ensures uniformity across different SFRA kits, while wavelet transform reduces noise for clearer interpretation. The Canberra distance algorithm evaluates signature correlation. Support Vector Machines (SVMs) and decision trees are trained on a dataset of 2670 SFRA signatures from 225 transformers and reactors, capturing frequency-domain attributes. SVM excels in handling complex, nonlinear relationships for robust fault analysis, accounting for varying test conditions. Ensemble methods like Random Forests and Gradient Boosting enhance fault analysis reliability by aggregating predictions from multiple decision trees.
The AI-driven objective analysis presented here showcases the effectiveness of integrating multiple techniques for the accurate interpretation and assessment of Transformers and Reactors. This comprehensive approach enhances diagnostic capabilities, facilitates proactive maintenance, and improves the overall reliability and efficiency of SFRA analysis in power transformers, ultimately saving time and cost for maintenance teams.
Sweep Frequency Response Analysis, Machine Learning, Support Vector Machine, Correlation Analysis, Wavelet Transform, Random Forest, Transformer, Reactor
