Tucker Decomposition-Based Feature Selection and SSA-Optimized Multi-Kernel SVM for Transformer Fault Diagnosis.

Area of Science:

• Electrical Engineering
• Data Science
• Artificial Intelligence

Background:

• Power transformer fault diagnosis is crucial for grid reliability.
• Conventional dissolved gas analysis (DGA) methods struggle with feature representation and high-dimensional data.
• Existing techniques often lack accuracy in complex fault classification scenarios.

Purpose of the Study:

• To develop an intelligent diagnostic framework for accurate power transformer fault classification.
• To overcome limitations in feature engineering and dimensionality reduction for DGA data.
• To enhance the performance of fault diagnosis systems through optimized machine learning models.

Main Methods:

• Systematic feature engineering expanded DGA data from 5D to 12D using IEC 60599 ratios and statistical descriptors.
• Tensor decomposition (Tucker) reduced feature dimensionality from 12 to 7, preserving 95% discriminative information.
• Sparrow Search Algorithm (SSA) optimized a multi-kernel Support Vector Machine (MKSVM) with RBF, polynomial, and sigmoid kernels.

Main Results:

• The proposed framework achieved a 98.33% classification accuracy for seven fault categories.
• Demonstrated superior performance compared to Kernel PCA, deep learning, and ensemble methods.
• Successfully retained significant discriminative information during dimensionality reduction.

Conclusions:

• The integrated framework offers a reliable and accurate solution for transformer condition monitoring.
• The synergistic approach of feature engineering, tensor decomposition, and SSA-optimized MKSVM enhances diagnostic capabilities.
• This advancement contributes to improved power system reliability and operational safety.