A CNN-based lamb wave processing model for field monitoring of fatigue cracks in orthotropic steel bridge decks

Abstract

Orthotropic steel bridge decks (OSDs) are prone to fatigue cracking under vehicle cyclic loads, and hence, fatigue crack monitoring is important, especially for the identification of crack growth. To achieve the intelligent monitoring of fatigue cracks in OSDs, Lamb wave technology equipped with machine learning algorithms was proposed in this study. A convolutional neural network (CNN) model with three functional layers was constructed, where a convolution layer was used to eliminate ambient noise disruption from wave feature data, a feature pooling layer to enhance wave features, and a full connection layer to identify crack dimension. The optimal time and frequency parameters of the feature pooling layer were determined based on numerical simulation data and, therefore, could be applied in general situations. In comparisons through field monitoring with continuous wavelet transform (CWT) and continuous Fourier transform (CFT) algorithms, the wave features obtained by the CNN model are more regularized and less fluctuant. The fatigue crack length identification errors of the CNN model are within 1 mm, which are smaller than those of the other algorithms. The CNN model obtains the highest accuracy rates under various loss thresholds. The proposed model also demonstrates superior identification accuracy in comparison to other existing methods. Consequently, when combined with Lamb wave technology, the model can therefore be suitable to applied in the fatigue crack monitoring of real OSDs.