Abstract: The sinking state prediction and optimal sensor placement are conducive to ensuring safe and steady sinking of open caissons and reducing monitoring costs. Based on LightGBM, a framework in the field of machine learning, a sinking state prediction model of super large open caisson is established. By using the monitoring data of the stress sensors at the bottom of the open caisson, the sinking speed of the open caisson, the height difference in the transverse direction and the height difference along the bridge direction are accurately predicted. Through the analysis of the sensor importance, the optimal sensor placement scheme that can meet the sinking state prediction accuracy is determined. The proposed sinking state prediction model and the optimal sensor placement method were applied to the super large open caisson sinking project of the main tower of Changtai Yangtze River Bridge. The results show that the model has high accuracy in predicting the sinking state of the open caisson, and the R² of the three prediction indexes is greater than 0.94. The important sensors for predicting the sinking state are mainly concentrated in the outer circle of the open caisson and the area near the transverse and longitudinal axes. Under the condition of satisfying the prediction accuracy of sinking state, the optimal sensor placement scheme can reduce the number of sensors by 45.5%. When the numbers of sensors in the optimal sensor placement schemes are same, the proposed optimization scheme is better than the scheme based on the correlation analysis of characteristic variables in terms of overall accuracy of sinking state prediction.

Key words: super large open caisson, sinking state prediction, optimal sensor placement, LightGBM, machine learning, feature importance