White-box regression (elastic net) modeling of earth pressure balance shield machine advance rate

This paper addresses the elastic net modeling of earth pressure balance shield machine (EPBM) advance rate for which there is no published physical model. Elastic net polynomial regression is employed to model the advance rate using data collected during the excavation of Sound Transit Northgate Link tunnels (N125) in Seattle, Washington, USA. The EPBM data was partitioned based on the geological profile to examine the influence of the soil type through which the EPBM was tunneling. The feature set for regression analyses included net thrust force, cutterhead rotation speed, conditioning foam flow rate, cutterhead torque, screw conveyor torques, as well as depths below ground surface and groundwater table. Third order polynomial models were able to model over 75% of the advance rate response on independent test data with normalized RMSE less than 20%. Using stratified sampling, only 10% of the EPBM data is required for model training to achieve these levels of accuracy and efficacy. Advance rate models were found to be considerably different across soil units indicating the soil type plays a significant role in EPBM response. The most influential parameters varied across soil types. Conditioning foam flow rate was the most important parameter in three of five soil units, while net thrust force and screw conveyor torque were the most influential features in two of five soil units. Partial dependence and individual conditional expectation analysis revealed that advance rate is positively related (increasing advance rate with increasing parameter value) and/or negatively related (decreasing advance rate with increasing parameter value) to varying degrees as a function of parameter value, all of which is strongly soil dependent.