Wavelet Scalogram–Based Hierarchical Deep Learning for Structural Damage Detection in Steel Truss Bridges

Abstract

The abstract should summarize the contents of the paper in short terms, i.e. 150-250 words. Real-time health monitoring of large civil engineer-ing structures is essential to ensure safety, serviceability, and to prevent cata-strophic failure. This study proposes a hierarchical deep learning (DL) model for noise-robust bridge damage detection from dynamic acceleration data pro-cessed through wavelet scalogram. The hierarchical DL model consists of a convolutional neural network (CNN) designed for feature extraction, followed by long short-term memory (LSTM) and gated recurrent unit (GRU) for captur-ing time-dependent features and long-term dependencies. A 2D steel truss bridge is modeled in finite element-based software SAP2000, and damage is in-troduced in the members of the bridge by reducing the flange and web thickness with varying severity. A moving train load is applied over the bridge, and dy-namic acceleration data for different damage scenarios are captured from dif-ferent nodes. A varying percentage of Gaussian random noise is added to the data to determine the robustness of the model. The noisy acceleration responses are then converted into scalograms using the continuous wavelet transform (CWT). Four different mother wavelets are compared here, where the complex Morlet has proved to be most effective for feature extraction and classification performance. The proposed Hierarchical DL model has achieved high testing accuracy of over 94%, 89%, and 75% for 1%, 5%, and 10% noise, respectively. The results suggest that the proposed hierarchical DL framework is noise resili-ent and efficient for real-time damage identification of truss bridges from wave-let scalograms generated from dynamic acceleration response.