Integrating Machine Learning with Pile Monitoring Data for Predicting Soil Resistance

Abstract Accurate prediction of soil resistance to driving is crucial for the structural integrity and cost efficiency of offshore platform construction. Traditional methods often rely on user-dependent processing and interpretation of pile monitoring data, introducing uncertainties, and necessitating conservative design assumptions. Adopting machine learning models based on in-situ monitoring data offers substantial benefits, including improved decision-making in platform design and construction. It provides engineers and clients with a robust tool for benchmarking judgments in offshore pile acceptance and future installation predictions. This paper presents an innovative, data-driven approach that integrates machine learning models to enhance the prediction of soil resistance during pile installation in offshore environments. Using a dataset compiled during pile installations in the Arabian Gulf, several supervised machine learning models, including k-Nearest Neighbors (kNN), CatBoost, and Histogram-based Gradient Boosting Regression Trees (HistGrad), were developed and evaluated. Soil resistances were determined using the Case Pile Wave Analysis Program (CAPWAP) software, providing a robust foundation for model training. A comprehensive feature engineering analysis identified an optimal set of input variables—diameter, penetration, observed blow count, and maximum energy recorded at the sensor level (EMX)—to enhance model accuracy. The machine learning models demonstrated high accuracy and reliability in predicting soil resistance, with Mean Absolute Percentage Error (MAPE) ranging from 10.9% to 14.8% and Coefficient of Determination (R²) between 78.7% and 89.6% during testing. Specifically, the proposed model using the selected input features achieved an R² of 89.6% and a MAPE of 10.9%. These results indicate that the data-driven approach significantly enhances the precision of soil resistance predictions, reducing reliance on conservative assumptions typically made in pile design and drivability analyses.