Investigation into fatigue micro-crack identification of steel bridge decks based on acoustic emission detection technology

Abstract

With increasing traffic loads and extended bridge service life, fatigue damage in steel bridge decks has become a significant concern. Traditional detection methods often lack the accuracy and responsiveness needed for practical engineering applications. To address the non-stationary nature of acoustic emission (AE) signals during crack initiation and propagation, this study combines the K-singular value decomposition (K-SVD) dictionary learning algorithm with convolutional neural networks (CNN) to enhance AE signal processing and fatigue crack detection. The K-SVD algorithm functions as an adaptive filter, learning from AE signals in various damage states to remove background noise and retain critical structural characteristics. This processed AE data is then input into a CNN, where the improved signal clarity enables higher classification accuracy. Specifically, the integration of K-SVD with CNN achieved recognition accuracies of 93.64% and 92.56% for AE signals from damaged areas, and 95.32% and 94.27% for undamaged signals, on training and test sets, respectively. This approach demonstrates strong engineering potential by providing a scalable solution for real-time, accurate crack detection in bridge inspections. Though computationally intensive, K-SVD’s adaptive dictionary learning enhances CNN performance, making the combination viable with optimization strategies in practical settings. These results provide a theoretical foundation and practical guidance for improving fatigue crack detection in steel bridge decks, supporting future applications in automated bridge inspection.