Strain Prediction Analysis of Bridge Structure Based on Improved EHO-LSTM Neural Network Model

Monitoring structural strain is an essential aspect of bridge monitoring, and it is a critical indicator of bridge stability. In-depth analysis and state prediction of monitoring data can effectively reflect the development and change trend of the bridge structure state. In this paper, we propose an improved EHO-LSTM recurrent neural network model to predict strains of bridge structural health monitoring (SHM) by examining the correlation between the monitoring data and time. The model was applied to the structure health-monitoring project of a city rail transit cable-stayed bridge in Guangzhou, where we selected 1008 phase data from abdominal measurement points as the experimental data to improve the EHO-LSTM model. The improved model was then compared with traditional RNN, SSA, and ARIMA prediction models, after which we discussed the bridge SHM strain prediction systematically. According to the results, the predicted data of each model are in basic agreement with the measured data. The improved EHO-LSTM model performed best in terms of prediction accuracy, with R2 of 0.87, RMSE of 3.48, and MSE of 3.06. Furthermore, as the number of periods increases, the absolute errors of the RNN, SSA, and ARIMA prediction models gradually increase, while the improved EHO-LSTM model maintains stability, which reflects its advantage in long-term memory.