Data-driven approach for intelligent tunnel dust concentration prediction

Abstract

The precise evaluation of tunnel dust concentration (TDC) stands as a primary concern within engineering practices. However, comprehensive and accurate prediction of TDC becomes increasingly challenging due to the expanding multi-source datasets and cumulative errors. This study employs innovative data-driven ensemble learning methodologies, specifically random forest (RF) and gradient boosting regression tree (GBRT), to delineate the intricate relationships between TDC values and various rock tunnel characteristics. Five diverse variables sourced from multiple origins are examined and employed as inputs within the database, while the TDC values derived from the dust concentration meter serve as the target outputs. Subsequently, a Bayesian optimization approach, the Tree-structured Parzen Estimator (TPE), is introduced to automatically ascertain the optimal hyper-parameters for the ensemble models. A comprehensive comparison is conducted between the two ensemble learning models and a singular machine learning algorithm, the classification and regression tree (CART), concerning predictive accuracy and resilience via 10-fold cross-validation (CV). The findings reveal the superior performance of the hybrid ensemble learning models over the individual ML models. Notably, the TPE-GBRT algorithm adeptly captures the measurement evolution, showcasing the lowest prediction errors and the highest correlation coefficient. These predicted outcomes significantly contribute to enhancing the engineering comprehension of the interrelation between rock tunnel parameters and TDC values.