Prediction and Feature Importance of Earth Pressure in Shields Using Machine Learning Algorithms

To reduce subjectivity and uncertainty when maintaining suitable earth pressure in earth pressure shields that can prevent heave or collapse, many prediction models using machine learning algorithms were proposed, but little research into the effects of other parameters on earth pressure has been undertaken, and soil conditioning parameters are always ignored. To establish a model with thorough parameters and probe into influences of other parameters, multiple machine learning algorithms were attempted. Given the accuracy, diversity and functions, random forest (RF), LightGBM and Attention-back-propagation neural network (Attention-BPNN) were further analyzed. Then, two RF models were compared in this research, one with soil conditioning parameters and the other without. Meanwhile, a case study was utilized to verify the reliability of the model. Finally, the feature importance of three models was compared and the variation rules of the most four important features were discussed by controlling variates. The results showed that soil conditioning parameters delivered a significant reduction in the prediction error. The case study demonstrated that the proposed model can satisfy engineering requirements. More earth pressure should give priority to increasing propulsion pressure, advance rate, and reducing foam air flow, rotational speed of screw conveyor, and vice versa.