Single-sensor impact force reconstruction in composite structures via a physics-guided multi-source data fusion network

Abstract

With the continuous advancement of structural health monitoring (SHM) systems for composite structures, accurately reconstructing time-varying impact forces from complex, noise-contaminated acceleration signals presents a critical computational challenge. To address the inherent limitations of conventional multi-impact inversion methods while maintaining real-time performance, we propose the Force Reconstruction with Transmissibility Enhanced Network (FRTE-Net). By integrating multimodal time–frequency domain physical information with attention mechanisms, FRTE-Net effectively aligns its computational workflow with the generalized transmissibility computational process. Experimental validation conducted on a composite flat plate and a foam-core blade using a single sensor demonstrates FRTE-Net’s substantial performance advantages. Specifically, the proposed method achieves a peak force prediction accuracy exceeding 88.91 %, outperforming CNN, DCNN, and DCSCNet by 19.96 %, 4.84 %, and 12.80 %, respectively, with peak position errors ranging from 0.61 to 29.81 ms. Given the richer input information used by FRTE-Net, these improvements should be interpreted as the overall benefit of the proposed multi-source physics-guided framework. The average Pearson’s correlation coefficients of 0.9059 and 0.8847 for the two test structures further affirm FRTE-Net’s efficacy in capturing the temporal profiles of repeated impacts at identical locations with high fidelity. Finally, comprehensive evaluation across varied sensor placements and artificially induced damage stages reveals the structural center as the optimal sensor location, maintaining peak accuracy above 88.48 % even under severe damage conditions. Collectively, these findings indicate that FRTE-Net provides a practical and computationally efficient approach for single-sensor impact force reconstruction in the tested composite structures.