Structural health monitoring of steel moment frame buildings via sequence-based recurrent neural networks

Abstract

Signal-based damage detection has gained extensive attention in recent years due to its capability in improving the deficiencies of previous structural health monitoring methods. Deep learning, with high capabilities in feature learning, has emerged as a powerful tool for sequence classification. In this paper, sequence-based deep learning models using long short-term memory (LSTM) and gated recurrent units (GRU) networks are used to detect structural damages and damage locations applied to steel building structures. To propose an appropriate deep-learning method for structural health monitoring, sets of monitoring data from the IASC-ASCE benchmark building were used. The data was collected from 15 sensors to collect accelerations attached to a 4-story steel moment frame building. The data has been properly pre-processed, denoised, sliced, and normalized. K-fold cross-validation validation is performed. The networks are designed using various combinations of recurrent neural networks, such as LSTM and GRU. It is concluded that stacked multilayer bidirectional long short-term memory networks, with an accuracy of 98%, have a superior performance in detecting the presence and location of structural damage.