Identification of Polar Substances in Transformer Insulation Oil Based on Multi-Strategy Data-Enhanced Terahertz Spectroscopy

Power transformers are core equipment in power grids, and polar substances in their insulating oil—such as furfural, methanol, water, and formic acid—serve as key biomarkers for assessing insulation condition. Traditional detection methods are time-consuming and operationally complex, making it difficult to meet the demand for rapid on-site testing. Terahertz spectroscopy, with its high sensitivity to polar molecules and non-destructive testing capabilities, shows great potential for assessing insulation oil aging. However, existing research has largely focused on the detection of single substances or overall condition assessment, and faces challenges such as limited spectral sample data and insufficient model generalization ability. In this study, a transmission-type terahertz time-domain spectroscopy detection platform was established, and insulating oil samples containing different volume concentrations of polar substances were prepared to obtain their absorption spectra. To address the challenge of training with a small sample size, we proposed a multi-strategy spectral data augmentation method that integrates Gaussian noise addition, baseline shifting and intensity scaling, and minor frequency-axis shifts, thereby expanding the trainable data volume to four times that of the original data. Based on this, we used principal component analysis to extract spectral features and established a support vector machine classification model for pattern recognition of the four polar substances mentioned above. The results show that the model without data augmentation achieved only 86.0% accuracy on the test set, indicating poor generalization ability; however, after applying data augmentation, the model’s recognition accuracy on the test set improved to 96.0%, with both recall and precision for each substance remaining above 90.0%, effectively overcoming the issue of overfitting. This study demonstrates that terahertz spectroscopy, combined with data augmentation and machine learning algorithms, enables rapid, high-precision, and non-destructive identification of polar substances in insulating oil, thereby offering a potential new technical pathway for transformer insulation condition assessment.