A Multimodal Driven Network to Inverse Localization of Random Impacts on Composite Structures using Strain Sensors

Impact localization is essential to structural health monitoring in special-purpose carbon fiber composite structures. However, both sparse sensor networks and limited feature signals are the main causes of difficulty in locating low-velocity impacts. To overcome this problem, a temporal-spectral-spatio multimodal attention network (TssAN) is proposed, combining multimodal feature generation and deep learning driven strategies. To generate a temporal-spectral-spatio multidimensional feature database, continuous wavelet transforms (CWT) spectrograms and the relative position matrix (RPM) images are extracted from the responses of the strain sensing array. These signals are prior filtered by using a boosted successive variational mode decomposition (SVMD). The extracted feature maps are adaptively fused by the agent attention mechanism to optimize a driving parameter set for impact localization model. The suggested method is verified, and the experimental result shows that the mean absolute errors (MAE) of TssAN are reduced by 1.24mm, 1.32mm, 1.66mm, 2.49mm, 2.04mm, and 1.05mm, respectively, compared with other selected typical methods. As a data-driven approach, this method improves impact localization accuracy by fusing multimodal data, effectively complementing purely physics-based models. The multimodal deep learning model provides a novel technical pathway for the impact localization field.