CNN-Based Similar Microwave Reflection Signals for Improved Detectability and Intelligent Characterization of Internal Defects in Composite Materials

Abstract

Near-field microwave imaging shows considerable promise for non-destructive evaluation of internal defects in high silica/phenolic composites, which are commonly used as thermal protection systems (TPS) for rocket/missile solid motor nozzles and space re-entry vehicles. However, effectively identifying defect features using post-processing algorithms remains challenging. To address this challenge, this paper proposes a microwave defect characterization algorithm based on Convolutional Neural Networks (CNN). A defect dataset derived from reflection microwave signals, was manually compiled by detecting samples with critical defects. The CNN framework was utilized for precise classification of microwave signals, employing a classification encoding strategy to extract two-dimensional defect information and achieve automatic localization and imaging of defects. Multiple deep learning models were compared in both simulations and experiments, revealing that the proposed CNN exhibited significant advantages in feature extraction, enabling highly effective identification of internal defects even with a limited dataset. Compared with traditional algorithms, the detection accuracy of the proposed 1D-SENet has been improved by 53.35% and 50.66%, respectively, and can achieve detection of defects with a minimum size of Φ6mm. These validate the effectiveness of algorithm in intelligent and automated microwave characterization of delamination defects within composite materials.