Deep learning-based detection of structural damage using time-series data

Abstract

Previously, it was nearly impossible to use raw time series sensory signals for structural health monitoring due to the inherent high dimensionality of measured data. However, recent developments in deep learning techniques have overcome the need of complex preprocessing in time series data. This study extends the applicability of four prominent deep learning algorithms: Multi-Layer Perceptron, Long Short Term Memory network, 1D Convolutional Neural Network, and Convolutional Neural Network to structural damage detection using raw data. Three structures ranging from relatively small structures to considerably large structures are extensively investigated, i.e., 1D continuous beam under random excitation, a 2D steel frame subjected to earthquake ground motion, and a 3D stayed-cable bridge under vehicular loads. In addition, a modulated workflow is designed to ease the switch of different DL algorithms and the fusion of data from sensors. The results provide a more insightful picture of the applicability of Deep Learning algorithms in performing structural damage detection via quantitative evaluations of detection accuracy, time complexity, and required data storage in multi-damage scenarios. Moreover, these results emphasize the high reliability of 2DCNN, as well as the good balance between accuracy and complexity of Long Short Term Memory and 1D Convolutional Neural Network.