Identification of vacuum OLTC faults using improved multimodal PKO-SVM.

Area of Science:

• Electrical Engineering
• Machine Learning
• Condition Monitoring

Background:

• Reliable operation of transformers relies on accurate fault identification in vacuum on-load tap changers (OLTCs).
• Current diagnostic methods often use single data sources or separate optimization steps, limiting their effectiveness.
• This can lead to suboptimal configurations and missed complementary fault information.

Purpose of the Study:

• To develop a multimodal PKO-SVM framework for enhanced fault identification in vacuum OLTCs.
• To integrate vibration-acoustic feature-level fusion with joint feature selection and SVM hyperparameter optimization.
• To improve diagnostic accuracy and model compactness for condition-based maintenance.

Main Methods:

• A dataset of 250 samples from five operating conditions was created, extracting vibration and acoustic features.
• A PKO-SVM framework was employed for joint optimization of feature selection and SVM hyperparameters (C and γ).
• A unified fitness function balancing recognition performance (Macro-F1) and feature compactness was utilized.

Main Results:

• The PKO-SVM model with fused features achieved 92.53% accuracy and 0.9252 Macro-F1.
• The fused feature set improved average accuracy by 8.43% (vs. acoustic-only) and 2.96% (vs. vibration-only).
• The proposed framework demonstrated superior performance and interpretability compared to other machine learning models.

Conclusions:

• The multimodal PKO-SVM framework offers a reliable and compact approach for diagnosing vacuum OLTC faults.
• Joint optimization of feature selection and hyperparameters enhances diagnostic model performance.
• This method holds significant potential for online condition assessment of vacuum OLTCs.