| Surcharge preloading is a common technique for soft foundation reinforcement.Its duration affects the reinforcement effect and project cost.To select a reasonable unloading time,it is necessary to predict the whole process of the soft foundation reinforcement project.When the bearing capacity of the soft foundation meets the construction requirements,it demands to unload the overload part to facilitate the subsequent construction of the project.However,because the traditional settlement prediction method requires the preliminary settlement monitoring data to conform to the existing curve trend.At the same time,the finite element software simulation prediction is limited by the selection of the constitutive model and soil parameters.There is a large gap between the numerical simulation predicted settlement and the measured settlement.The rapid development of artificial intelligence technology drives the rapid development of subsidence prediction methods based on machine learning in deformation and subsidence prediction.But the settlement prediction by machine learning usually only focuses on the relationship between settlement and time,which does not include the site construction situation,soil parameters and other selection-related characteristics in the settlement prediction model.Moreover,the basic machine learning prediction model has a poor generalization ability.To this end,this thesis built a settlement prediction database containing information on site construction,pile load,and soil parametersby by collecting a large number of practical projects,after that,the sparrow search algorithm was improved for settlement prediction adopt a multi-strategy.The main research contents and achievements are as follows:(1)The traditional settlement prediction method has a poor performance on soft foundation settlement prediction engineering.Since the measured settlement trend deviates from exponential curve,the exponential curve method is not suitable for the preloading of large overload projects.Asaoka’s method can not predict the development trend of settlement,but only predict the maximum.The hyperbolic method can only predict the development trend of settlement after the completion of loading.Besides,there is a large error between the unloading time selected by the prediction result and the actual unloading time of the project.(2)Based on a large number of soft foundation treatment engineering and geotechnical engineering theories,a database containing 20 characteristics was established for settlement prediction model development,which contains 105 settlement measuring points and 893 sets of settlement data.The statistical analysis of the database showed that the soft foundation treatment method was vacuum combined load preloading,the drainage method was plastic drainage plate,and the layout of the drainage plate was square.The settlement in the database is mainly concentrated within 2500mm.According to the correlation analysis,the input features of the machine learning model were determined to the mutual information score.(3)Three machine learning algorithms were used to train and predict the database.The coefficient of determination(R~2)for the backpropagation neural network and the Random Forest model exceeds 0.9 in the training set,while only the Gaussian SVM model’s R~2 exceeds 0.9.However,in the test set,the R~2 for all models-backpropagation neural network,Random Forest,and Gaussian SVM model-is below 0.9.This suggests that the basic machine learning model has a weak generalization ability.(4)To address the issues of improper initial population distribution and a tendency to get stuck in local optima,a logistic-tent chaotic map,an adaptive nonlinear inertia weight decay parameter,and Levy flight behavior were used to improve the SSA algorithm.Through a comprehensive analysis of convergence precision,stability,and convergence speed,it found that the improved Sparrow Search Algorithm(ISSA)effectively enhances the optimization performance of the algorithm.(5)After applying the improved sparrow search algorithm(ISSA)to multiple machine learning models,there were significant improvements.Among these models,the ISSA Random Forest(RF)model demonstrated excellent performance in the settlement prediction database and was selected as the settlement prediction model.When applied to practical engineering,the ISSA RF model improved the accuracy of settlement prediction compared to the exponential fitting method and the Asaoka’s method.Moreover,the ISSA-RF model can predict the settlement of any stacking site according to the construction plan and survey report before the stacking construction. |
