Abstract:
Nowadays, the need for subway tunnels has increased considerably with urbanization and population growth in order to
facilitate movements. In urban areas, subway tunnels are excavated in shallow depths under densely populated areas and
soft ground. Its associated hazards include poor ground conditions and surface settlement induced by tunneling. Various
sophisticated variables influence the settlement of the ground surface caused by tunneling. The shield machine's operational
parameters are critical due to the complexity of shield-soil interactions, tunnel geometry, and local geological parameters.
Since all elements appear to have some effect on tunneling-induced settlement, none stand out as particularly significant;
it might be challenging to identify the most important ones. This paper presents a new model of an artificial neural network
(ANN) based on the partial dependency approach (PDA) to optimize the lack of explainability of ANN models and evaluate
the sensitivity of the model response to tunneling parameters for the prediction of ground surface and subsurface settlement.
For this purpose, 239 and 104 points for monitoring surface and subsurface settlement, respectively, were obtained from
line Y, the west bond of Crossrail tunnels in London. The parameters of the ground surface, the trough, and the tunnel
boring machine (TBM) were used to categorize the 12 potential input parameters that could impact the maximum settlement
induced by tunneling. An ANN model and a standard statistical model of multiple linear regression (MLR) were also used
to show the capabilities of the ANN model based on PDA in displaying the parameter's interaction impact. Performance
indicators such as the correlation coefficient (R2
), root mean square error (RMSE), and t-test were generated to measure
the prediction performance of the described models. According to the results, geotechnical engineers in general practice
should attend closely to index properties to reduce the geotechnical risks related to tunneling-induced ground settlement.
The results revealed that the interaction of two parameters that have different effects on the target parameter could change
the overall impact of the entire model. Remarkably, the interaction between tunneling parameters was observed more
precisely in the subsurface zone than in the surface zone. The comparison results also indicated that the proposed PDA-
ANN model is more reliable than the ANN and MLR models in presenting the parameter interaction impact. It can be
further applied to establish multivariate models that consider multiple parameters in a single model, better capturing the
correlation among different parameters, leading to more realistic demand and reliable ground settlement assessments. This
study will benefit underground excavation projects; the experts could make recommendations on the criteria for settlement
control and controlling the tunneling parameters based on predicted results.