SERS-based quantitative detection of furfural in transformer oil using an improved residual network and transfer...

Accurate quantification of furfural concentration in transformer oil is crucial for assessing the aging status of insulating paper; however, conventional quantitative models face significant challenges in generalizing due to interference from complex oil matrices. To address this, we propose a quantitative analytical framework integrating an improved residual network (ResNet) with transfer learning. Initially, a deep convolutional generative adversarial network (DCGAN) was introduced to augment the Raman spectral dataset of furfural, generating high-fidelity spectra (peak signal-to-noise ratio, PSNR = 46.02; structural similarity index, SSIM = 0.98) to mitigate the risk of overfitting due to small sample sizes. Subsequently, an improved ResNet model was developed by simplifying network structure and embedding Dropout regularization, significantly enhancing feature extraction capability and robustness (correlation coefficient, R² = 0.9921). To accommodate spectral distribution differences among multi-source oil samples, a transfer learning framework (Transfer-ResNet) was constructed by freezing lower network layers and fine-tuning deeper residual blocks, thus achieving cross-domain generalization (R² = 0.9914). Furthermore, Grad-CAM++ interpretability analysis revealed the model's multi-frequency response mechanisms, identifying key molecular features such as C-H/C-C vibrations and CN stretching. This proposed method provides a novel strategy for accurate, interpretable quantitative SERS detection of furfural in transformer oil, demonstrating significant potential for transformer condition assessment.