Interface damage identification of composite steel bridge decks based on AE and deep learning techniques

An interface damage identification method is proposed for UHPFRC-steel cold composite plates with adhesive-bonded corrugated shear connectors, integrating acoustic emission (AE) monitoring and deep learning. Two full-scale segment specimens were designed and subjected to three-point bending tests. AE signals collected during loading reveal a strong correlation between signal activity and damage progression, where sudden increases in AE energy indicate local structural instability. Principal component analysis (PCA)-based dimensionality reduction and feature selection using Laplacian score and Normalized Mutual Information jointly improve clustering performance. With the optimized features, K-means+ + achieves accurate classification of typical damage modes. A PSO-optimized variational mode decomposition approach achieves 6.8–34.6 % improvement in SNR and 0.01–0.03 increase in correlation compared to WOA, GA, and AIA methods. Peak frequencies of AE signals related to interface damage are primarily distributed in the 100–200 kHz range. A convolutional neural network, NiVGG-Net, is developed for damage pattern recognition, achieving over 90 % accuracy and outperforming LeNet-5 in convergence speed and stability. This study provides valuable insights into interface damage mode identification of UHPFRC-steel cold composite plates.