Terahertz spectroscopic detection of EPDM insulation defects based on a CNN-XGBoost cascade architecture and SHAP explainability

Addressing the challenges of distinguishing weak terahertz (THz) spectral responses from nonlinear baseline drifts in ethylene propylene diene monomer insulation, as well as the interpretability limitations of black-box models, this study adopts and optimizes a transparent defect recognition framework based on a CNN-XGBoost cascade architecture. A comprehensive dataset consisting of 1600 high-quality spectral samples is established via high-precision raster scanning over 20 independent physical models with systematically controlled burial depth gradients ranging from 0.5 to 2.0 mm. Within this framework, a one-dimensional convolutional neural network (1D-CNN) is employed to extract hierarchical nonlinear feature representations, thereby effectively decoupling intrinsic defect signatures from thickness-induced spectral distortions. Subsequently, an XGBoost classifier is introduced to further refine the decision boundaries and enhance classification generalization. The proposed approach achieves a classification accuracy of 99.38% on the experimental dataset, demonstrating its superior predictive capability. Furthermore, SHAP attribution analysis is implemented to evaluate the model’s data-driven decision logic. The results indicate that, rather than relying on incidental statistical correlations, the model consistently anchors its predictions to physically meaningful diagnostic frequency bands, notably around 0.89, 0.60, and 0.30 THz. These optimal statistical discriminators successfully capture the macroscopic spectral differences among defect types, such as broadband attenuation and localized signal dips. This work provides an accurate and explainable data-driven methodological reference for high-precision nondestructive testing (NDT).