Convolutional Neural Network Based Damage Detection of IASC-ASCE Benchmark by Encoding Time-series into Images

Abstract

Civil engineering structures are essential to the infrastructure and directly impact people's lives and safety. Notably, many structures are in worrisome working conditions, with damage deterioration and occasionally collapse. Combining deep learning with structural health monitoring can provide unprecedented tools for structural damage detection and identification. This paper explores the use of acceleration features to predict the damage state of a structure using time-series acceleration data collected from the Phase I ASCE Structural Health Monitoring (SHM) benchmark model, where the primary identified damage is the removal of diagonal braces and beam connections. In this study, to overcome the limitations of using neural networks hampered by the small amount of data collected in shaker tests, three methods were used to encode time-series acceleration data as images, i.e., two-dimensional numerical matrix (2D Matrix), Short-Time Fourier Transform (STFT), and Wavelet Transform (WT). The encoded images were then used as input to the Convolutional Neural Networks (CNN) model. Two training/test scenarios are used to show the efficiency of the proposed method of structural damage recognition. The results show that the proposed method can predict structural damage and the WT combined with the CNN is the best approach regarding accuracy and efficiency.