基于互信息加权与时序嵌入的坝体裂缝预测模型OA

To address the nonlinear associations and multi-scale temporal dependencies in multi-source heterogeneous data from dam crack monitoring,this paper proposed a hybrid prediction model(M-GEX)based on mutual information weighting and temporal embedding fusion.Firstly,the model quantified and weighted input features using mutual information,then exteacted multi-dimensional statistical features reflecting short-term dynamics through sliding windows,and generated low-dimensional temporal embeddings containing long-term dependencies using a GRU encoder.These features were concatenated and fed into an XGBoost regression model for prediction,with R2 and ERMS constituting the evaluation framework.In prediction tasks using three sets of monitoring data from Meishan Reservoir,the model achieved R2 values exceeding 0.87 and ERMS values below 1.84 across all datasets,demonstrating high accuracy in trend fitting and peak-valley prediction.M-GEX effectively integrates short-term statistical features with long-term times-series embeddings,enhancing both prediction accuracy and model interpretability.This approach provides an efficient and scalable technical solution for dam safety monitoring and early warning systems.