Image-based machine learning approach for structural damage detection through wavelet transforms

Abstract

Structural integrity is essential for safety in infrastructure, as it can help prevent catastrophic failures and financial losses. The significance of vibration-based damage detection has grown substantially in fields such as civil and mechanical engineering. Concurrently, the advancements in computational capacities have facilitated the integration of machine learning into damage detection processes through post-processing algorithms. Nevertheless, these require extensive data from structure-affixed sensors, raising computational requirements. In an effort to address this challenge, we propose a novel approach utilizing a pre-trained convolutional neural network (CNN) based on images to identify and assess structural damage. This method involves employing wavelet transform and scalograms to convert numerical acceleration data into image data, preserving spatial and temporal information more effectively compared to conventional Fourier transform frequency analysis. Six acceleration data channels are collected from carefully chosen nodes on a mini bridge model and a corresponding finite element bridge model, to train the CNN. The efficiency of training is further enhanced by applying transfer machine learning through two pre-trained CNNs, namely Alexnet and Resnet. We evaluate our method using different damage scenarios, and both Alexnet and Resnet show prediction accuracies over 90%.