Multimodal feature fusion for detecting debonding in FRP-reinforced concrete interfaces using dual-branch CNN

The percussion method offers a viable solution for detecting debonding damage in fiber-reinforced polymer (FRP)-reinforced concrete interfaces. Leveraging the percussion method, a nondestructive technique, acoustic signals generated by light tapping on the FRP surface are analyzed to detect debonding. Traditional inspection methods often fail to identify internal structural defects, while our method introduces a dual-branch neural network that combines the strengths of gated recurrent units for temporal feature extraction and convolutional neural networks for spectral feature extraction. The innovation of this study lies in the integration of multimodal feature fusion and multi-head attention mechanisms, which allow the model to capture both temporal and spectral characteristics of the acoustic signals with greater precision. This dual-branch architecture, enhanced by attention mechanisms, significantly improves the accuracy and robustness of debonding detection compared to conventional techniques. Experimental results demonstrate the method’s potential in providing a scalable solution for real-time, noninvasive structural health monitoring of FRP-reinforced concrete structures.