A Machine Learning Model for Pile Settlement Prediction Using Majority Voting-Based Feature Selection

Pile foundations are deep foundations commonly employed in bridge construction, high-rise buildings, trains, and situations requiring high bearing capacity and minimal settlement. Accurate prediction of pile settlement is essential for ensuring the safety and stability of deep foundations, yet traditional methods like in-situ load tests are often costly and impractical. The cone penetration test (CPT) is one of the most frequent in-situ tests for pile analysis because, like a model pile, the measured cone resistance and sleeve fiction can be used to estimate pile unit toe and shaft resistances, respectively. In this paper, a machine learning (ML) framework for pile settlement prediction with a genetic algorithm (GA) majority voting (MV) feature selection (FS) strategy to enhance model performance is presented. Three tree-based algorithms, each with a unique approach for tree development and feature handling—categorical boosting (CB), light gradient boosting (LGB), and random forest (RF) are selected for this purpose. The dataset was compiled from fifty-six pile case histories in different countries have been compiled including static loading tests which include maintained load tests and constant rate of penetration tests, shaft, and toe resistances which comprise CPT and CPTu (undrained CPT) sounding, the pile geometric and mechanical properties, the loads applied from the load tests as the model inputs, and recorded settlement values for the piles from the tests as the model output to be predicted. The CB model, coupled with the GA-MV approach, achieved the best predictive accuracy, yielding an <I>R²</I> of 0.926 and RMSE of 5.92 mm upon testing, while feature importance analysis identifies applied load (P) and pile length (L) as key predictors of settlement. Also, an overall decrease of the RMSE by 11.19% was observed between the CB-GAMV model (5.92 mm) and the CB-All features model (6.68 mm), and 9.41% between the CB-GAMV model and the CB-GA model (6.54 mm) on the validation set.