Terahertz recognition of composite material interfaces based on ResNet-BiLSTM

Abstract

This manuscript presents a framework that integrates a residual network (ResNet) with a bidirectional long short-term memory network (BiLSTM) for effective feature extraction and interface positioning in terahertz signal analysis. This innovative approach employs spatial and temporal feature extraction modules to process terahertz signals, enabling the automatic detection of interfaces and defects in both coating and bonded structure samples. The experimental results indicate that the interface positioning accuracy of the method aligns with the manual analysis outcomes for coating samples exceeding 80 µm in thickness. Furthermore, it exhibits a high degree of consistency with X-ray CT detection results when detecting defects in the adhesive layer, with an error margin ranging from 1.07 % to 5.39 %. Significantly, we have also conducted preliminary analyses on actual samples, further underscoring the robustness and generalizability of our model. This method offers an efficient automated tool for terahertz signal analysis and holds promise for industrial detection applications. Future research will aim to broaden the range of samples and application scenarios, thereby increasing the technology’s practicality in quality control and production line detection.