Accurate clearance monitoring is critical to the safe and stable operation of wind turbines. However， existing clearance-monitoring systems are often costly and suffer from limited reliability. To address these challenges， this study proposes a deep-learning-based surrogate-model workflow for clearance prediction. High-fidelity simulations of a custom turbine are first performed using OpenFAST to generate a high-quality dataset that mitigates the high noise and low accuracy inherent in conventional measurement approaches. Feature engineering is then applied to reduce model training complexity. For the classification and regression-tree （CART） model， we introduced wind-speed-segmented training and a sample-weighting scheme that emphasizes large-deformation cases， thereby jointly improving overall predictive accuracy and peak-value fidelity. Results show that the modified CART model achieves， on average， a 7.51% reduction in RMSE； the proportion of positive prediction errors exceeding 0.5m and 1.0m are reduced by 17.8% and 35.4%， respectively. On a 100-minute test covering cut-in to cut-out wind speeds， the Segmented-Weighted model attains a maximum positive peak prediction error of 0.402m， an overall NRMSE of 1.49%， and proportions of positive errors greater than 0.5m and 1.0m of 10.2% and 1.69%， respectively. The proposed workflow provides an efficient， reliable， and cost-effective approach for wind-turbine clearance monitoring.