Year: 2026
Type: article
Abstract: Single-modal terahertz nondestructive testing faces inherent bottlenecks in quantifying thin concealed air gaps inside layered composite materials widely used in aerospace and advanced manufacturing. The reflection mode suffers from waveform superposition and interference, which restricts its quantitative accuracy. The transmission mode has a higher accuracy upper limit but cannot realize independent defect localization. To tackle these complementary limitations, this work proposes a machine learning-assisted dual-modal geometric correction method.The Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) algorithm is first applied to preprocess reflection-mode terahertz signals. For the transmission mode, a Light Gradient Boosting Machine (LightGBM) model built on four-domain fused features is constructed for air gap thickness prediction, and the SHapley Additive exPlanations (SHAP) method is introduced to interpret the model's internal decision mechanism. Parameter calibration is finally achieved via dual-modal data fusion guided by geometric constraints.Transmissionmode thickness prediction reaches a coefficient of determination (R²) above 0.999. After dualmodal correction, the relative error of air gap thickness approaches zero, and the maximum absolute error of gap-substrate distance is only 0.02104 mm, with both localization and quantification performance significantly enhanced.This method can provide a valuable technical reference for quantitative nondestructive detection of air gap defects in layered composite materials.
Source: Physica Scripta
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W7225620847