Damage detection of structures based on wavelet analysis using improved AlexNet

Deep learning-based approaches have garnered a great deal of interest among different methods in structural health monitoring (SHM), whose primary objective is to assess structural health status. AlexNet is renowned for being a practical convolutional neural network architecture utilized to classify images. However, having numerous parameters, AlexNet gives rise to significant computational costs when applied to classify a small number of damage types occurred in civil engineering. Therefore, in this paper, the improved AlexNet with a lower computational cost and images generated by continuous wavelet transform (CWT) have been employed to classify multiple types of damages which is the pivotal characteristics of the proposed procedure. The performance of this approach was validated with responses obtained from three structures. The acceleration responses converted to time–frequency images were derived from undamaged and damaged states, such as scour, damage in connections, and stiffness reduction. Moreover, the efficiency in detecting the damage extent was investigated. The approach exhibited proficient performance in multi-classification, achieving high prediction accuracy rates of more than 99%, 97%, and 94% in a benchmark bridge, a frame model, and a real truss bridge, respectively. The sensitivity analysis results showed that the Bump mother wavelet had the best performance among studied mother wavelets; moreover, investigation of the number of input images showed that the higher the number of pictures, the more accurate the results, and prediction accuracy fluctuated by no more than approximately 5%. Finally, results demonstrated that the accuracy improves as the duration of each record increases, and the related decrease in accuracy is roughly 3%.