BP Neural Network–Based Analysis and Prediction of SHM Data of In-Service Concrete Bridge

Abstract

Structural health monitoring (SHM) has now been widely applied in various new bridges, especially in the service life, which makes it possible to evaluate the service performance of bridge structures in a more comprehensive and accurate way. However, there is also a need for effective analysis of large amounts of structural monitoring data and scientific support for bridge maintenance and management decisions. The aim of this paper is to propose an innovative BP neural network improvement method for more accurate analysis and prediction of massive monitoring data. Specifically, it includes taking the displacement and strain monitoring sample data of an active simply supported girder bridge as the research object and comparing the difference between the original BP neural network and the improved BP neural network in terms of prediction accuracy by analyzing and predicting these sample data, combined with the error analysis. In addition, this paper introduces the long short-term memory (LSTM) network, which has relatively mature applications in the field of bridge SHM, to conduct comparative experiments and systematically evaluate the prediction performance of the improved BP neural network. The results show that the monitoring data of in-service bridges belong to typical time series data and can be effectively analyzed by BP neural network method for the analysis and prediction and the error is small. Moreover, the improved BP neural network method can further reduce the errors in analyzing and predicting the monitoring data. The experimental results show that although the numerical differences of the results are small in most cases, the mean square error (MSE) of the improved BP neural network is 20.9719 and 26.4565 for the prediction of the strain data of 1# and 3# beams, respectively, which is significantly better than that of the LSTM, which is 23.5895 and 31.0791, and the mean absolute error (MAE) of the BP neural network is 3.7611 for the prediction of 5# beams, which is much lower than that of the LSTM, which is 5.2746. The MAE is 3.7611, which is much lower than 5.2746 of LSTM. These results side by side reflect that overall, the improved BP neural network outperforms LSTM. This study can be used for intelligent analysis and processing of bridge monitoring data. It can also provide reference for further analysis and evaluation of bridge structural service performance.