The caisson foundation is widely used in large bridge projects due to its high overall stiffness and strong bearing capacity. The key to successful sinking construction is to maintain a safe and stable sinking process. Accurately predicting the sinking rate and tilt degree of the caisson foundation during soil excavation is crucial for effective sinking control. During the sinking process

extensive real-time monitoring data of soil pressure at the cutting edge are collected. These data exhibit high dimensionality

and the underlying mechanisms linking soil pressure to sinking rate and tilt degree are complex

posing challenges for traditional analytical methods. Therefore

the extra trees algorithm from machine learning is employed to establish a sinking state prediction model. This model extracts temporal and spatial features from the monitoring data

captures the complex relationships between cutting edge soil pressure and sinking behavior

and enables intelligent prediction of sinking rate and tilt degree. The model was applied to the northern anchor caisson project of the Zhangjinggao Yangtze River Bridge

with model evaluation parameters calculated to verify prediction accuracy. Additionally

the extra trees algorithm was compared with other common machine learning methods

and the influence of model parameters on prediction accuracy was analyzed. Results show that the established model achieves high prediction accuracy

with R2 values consistently greater than 0.9 in engineering applications

meeting project requirements. The extra trees algorithm outperforms other machine learning methods

and prediction accuracy improves with an increased number of individual decision trees and greater maximum tree depth. These findings provide valuable reference for controlling sinking rates and tilt degrees in similar caisson foundation projects.