Bridge Damage Identification Using Time-Varying Filtering-Based Empirical Mode Decomposition and Pre-Trained Convolutional Neural Networks

Abstract

Highlights: What are the main findings? A novel bridge damage identification framework is proposed, combining TVFEMD for signal denoising and pre-trained CNNs for accurate damage classification. The study finds that ResNet-50 performs optimally in damage classification tasks, especially when processing TVFEMD-processed signals, with improved clustering and separability of features. What is the implication of the main finding? The proposed method improves the robustness of structural health monitoring systems in noisy environments, enhancing damage identification accuracy in real-world conditions. It offers a practical and scalable approach for intelligent structural health monitoring in real-world engineering applications. Structural damage identification provides a theoretical foundation for the operational safety and preventive maintenance of in-service bridges. However, practical bridge health monitoring faces challenges in poor signal quality, difficulties in feature extraction, and insufficient damage classification accuracy. This study presents a bridge damage identification framework integrating time-varying filtering-based empirical mode decomposition (TVFEMD) with pre-trained convolutional neural networks (CNNs). The proposed method enhances the key frequency-domain features of signals and suppresses the interference of non-stationary noise on model training through adaptive denoising and time–frequency reconstruction. TVFEMD was demonstrated in numerical simulation experiments to have a better performance than the traditional EMD in terms of frequency separation and modal purity. Furthermore, the performances of three pre-trained CNN models were compared in damage classification tasks. The results indicate that ResNet-50 has the best optimal performance compared with the other networks, particularly exhibiting better adaptability and recognition accuracy when processing TVFEMD-denoised signals. In addition, the principal component analysis visualization results demonstrate that TVFEMD significantly improves the clustering and separability of feature data, providing clearer class boundaries and reducing feature overlap.