Feature Extraction of Oil-Paper Insulation Raman Spectroscopy Based on Manifold Dimension Transformation

Transformers play a crucial role in power systems. In this respect, fault diagnosis and aging state assessment have garnered significant attention from researchers. Herein, accelerated thermal aging and Raman scattering experiments are conducted on oil-paper insulation samples to accurately detect aging states. The samples are categorized into different aging stages based on the polymerization degree of the insulating paper. Principal component analysis (PCA), multi-dimensional scale change method (MDS), and isometric mapping algorithm (Isomap) are employed to extract features from the Raman spectra. Subsequently, the XGBoost strong classifier, optimized through Bayesian hyperparameter optimization (BO-XGBoost), is utilized to distinguish between four and ten states among 175 groups of samples after feature extraction. The subsequent classification results of the three feature-extraction methods are compared. The results indicate that Isoamp, which pertains to the manifold dimension transformation, achieves the highest average discriminative accuracy after feature extraction. The discriminative accuracies for aging states four and ten are 97.0% and 95.1% respectively, demonstrating that Raman spectroscopy manifold dimension transformation enhances the distinctiveness between samples of different aging states in the feature-extraction process. The diagnostic model constructed with Isomap and BO-XGBoost enables accurate discrimination of the aging states of oil-paper insulation.

Automated summary

Transformers are essential in power systems. Researchers have focused on fault diagnosis and aging state assessment. This study conducted accelerated thermal aging and Raman scattering experiments on oil-paper insulation samples to detect aging states accurately. They categorized the samples into different aging stages based on the polymerization degree of the insulating paper. Principal component analysis (PCA), multi-dimensional scale change method (MDS), and isometric mapping algorithm (Isomap) were used to extract features from the Raman spectra. The XGBoost classifier, optimized through Bayesian hyperparameter optimization (BO-XGBoost), was employed to distinguish between four and ten states among 175 groups of samples after feature extraction. The results showed that Isomap achieved the highest average discriminative accuracy, with 97.0% and 95.1% for aging states four and ten respectively, indicating that Raman spectroscopy manifold dimension transformation enhances the distinctiveness between samples of different aging states. The diagnostic model constructed with Isomap and BO-XGBoost enables accurate discrimination of the aging states of oil-paper insulation.