A novel structural damage identification method based on multi-sensor data fusion and multimodal neural networks

Structural Health Monitoring (SHM) ensures the safety of engineering structures, but nonlinear, non-stationary sensor data with high noise limits damage identification accuracy. This study proposes a multimodal fusion model integrating a Re-parameterized Large Kernel Network (RepLKNet) and a Bidirectional Gated Recurrent Unit (BiGRU) global attention mechanism. A recursive graph extracts nonlinear dynamic features while maintaining temporal dependencies. Multimodal fusion enhances feature extraction and classification accuracy under complex conditions. To mitigate noise, an adaptive joint denoising algorithm combining Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) and Adaptive Wavelet Packet Threshold Denoising (AWPTD) is designed. The method is validated through simulations on the IASC-ASCE benchmark structure and physical model tests of high-pile wharves, demonstrating superior classification performance and robustness in noisy environments. Compared to traditional approaches, the proposed framework improves adaptability and accuracy, offering theoretical and practical advancements in multimodal SHM.