四种迁移学习架构在跨仪器绝缘纸聚合度评估中的效果对比

Near-infrared spectroscopy (NIRS) has become an important nondestructive technique for assessing the degree of polymerization (DP) of insulating paper as an alternative to traditional chemical methods. As a typical data-driven chemometric approach, NIRS often suffers from amplitude shifts in spectral data collected across different instruments, which severely limit its generalizability and large-scale deployment in engineering applications. In this work, we construct a model transfer scenario involving four representative spectrometers to analyze inter-instrument spectral discrepancies. We systematically compare the transfer learning fine-tuning performance of four mainstream deep neural network architectures under different parameter configurations, investigate the effect of layer-freezing strategies on transfer performance, and identify the optimal network structure and hyperparameter combination within a predefined search space. The results show that significant spectral amplitude differences exist among instruments, leading to drastic degradation of predictive performance when a source-domain model is directly applied to the target domain, rendering themodel nearly ineffective. Freezing strategies improve fine-tuning performance by stabilizing the network; specifically, freezing the front-end feature extraction layers while fine-tuning the higher-level decision layers enhances transferability without compromising stability. Among the four tested architectures 一 MLP, 1D-CNN, EOT, and ResNet 一 EOT achieved the lowest error in the source domain but performed worse after fine-tuning in the target domain, whereas ResNet exhibited higher source-domain error than EOT but achieved better fine-tuning performance. This indicates that source-domain training error is not strongly correlated with transfer effectiveness. Overall, a three-branch ResNet network incorporating multi-scale Inception modules achieved the best target-domain performance, with an RMSE of 78. 5 and a MAPE of 8. 6% after fine-tuning, significantly outperforming the other models. These findings provide theoretical support for constructing NIRS modeling frameworks with cross-instrument generalizability.