Convolutional neural network based structural health monitoring for rocking bridge system by encoding time‐series into images

Abstract

Structural health monitoring of infrastructure especially bridges plays a vital role in post-earthquake recovery. Coupling emerging techniques in machine learning with structural health monitoring can provide unprecedented tools for damage detection and identification. This paper explores the use of time-series acceleration or displacement data collected from a shake-table experiment of a two-span bridge utilizing pretensioned rocking columns to predict the damage state of each bridge bent, where the major identified damage was the fracture of the longitudinal bars. To overcome the limitation of small data size collected during the shake-table test that hindered the use of artificial neural networks and recurrent neural networks, the time-series data were encoded into images using three methods Gramian angular summation field, Gramian angular difference field, and Markov transition field. Then, the encoded images were used as an input for convolutional neural network models. Three different data entries for the input layers were used including encoded images from recorded accelerations, drift ratios, and both. Two training/testing scenarios were proposed to test the efficacy of the convolutional neural networks. Convolutional neural network models trained on Markov transition field encoded images from acceleration performed with 100% accuracy during the training phase and more than 94% for the testing phase.