A data-driven model for steel bridge temperature behaviour based on deep learning technology and heat transfer analysis

Abstract

The implementation of machine learning techniques in structural health monitoring has garnered significant attention in recent years due to their performance in nonlinear modelling. Compared with the traditional artificial neural network (ANN) methods, deep learning (DL) technologies comprising multi-layer ANNs have become increasingly powerful for solving nonlinear problems over the past decade. Within the DL domain, recurrent neural networks are improved by long short-term memory (LSTM) networks, which can capture the long-term dependencies in sequential data prediction. This advancement in DL technology has significant implications for predicting complex structural behaviour under changing environmental conditions. As bridge structures will experience much higher temperatures due to climate change, LSTM models can help predict the nonlinear relationship between various ambient climatic conditions and the temperature behaviour of bridges based on the limited input, effectively leveraging their capability for time series analysis. Optimising the model hyperparameters of the neural network by the Bayesian optimization algorithm also improves the DL network. As the simple data-driven method cannot simulate the entire bridge temperature field, the nodal temperatures predicted by the DL model at limited locations will be combined with a suitable discrete numerical model for heat transfer analysis to further enhance the accuracy. The DL-based temperature simulation model is then applied to a steel bridge deck and verified by monitoring data.