Improving Structural Monitoring Predictions Using Integrated Feature Selection and ML Optimization

With the rapid digitization of civil infrastructure systems, structural health monitoring (SHM) and predictive maintenance rely increasingly on high dimensional sensor data. However, handling datasets with hundreds of parameters such as strain, displacement, vibration, and environmental conditions poses challenges to model accuracy, computational efficiency, and interpretability. This study proposes a hybrid feature selection framework that integrates filter, wrapper, and embedded techniques, including Information Gain (IG), Recursive Feature Elimination (RFE), Random Forest (RF) importance ranking, and LASSO regularization. The objective is to identify the most informative variables while reducing redundant or noisy data. Using real world infrastructure monitoring data comprising over 200 parameters, several machine learning models—Random Forest Regression (RFR), XG Boost, Support Vector Regression (SVR), and Deep Neural Networks (DNN) were trained and evaluated under multiple feature selection scenarios. Experimental results show that hybrid feature selection improves RMSE by 12% on average and reduces training time by up to 30%. Ensemble models trained on hybrid selected features consistently outperformed models trained using all features. These findings highlight the effectiveness of combining engineering domain knowledge with machine learning based feature optimization in enhancing civil infrastructure predictive accuracy and computational efficiency.