Research On Foundation Pit Monitoring Data Filling And Deformation Prediction Based On Generative Adversarial Network

Foundation pit engineering occupies an important position in urban construction,but with the increase of its quantity and difficulty and the frequent occurrence of engineering accidents,the protection requirements of foundation pits and their surrounding buildings are also improved,so the design mode of foundation pit engineering also shifts from controlling strength and stability to controlling deformation.In order to effectively improve the early warning capability of foundation pit deformation and solve the problems of traditional prediction methods which consider few factors affecting deformation,difficulties in making full use of historical monitoring data,relatively low prediction accuracy and poor generalization ability.In this paper,four models such as VAE-CNN-LSTM model are built to predict the foundation pit deformation,and the prediction accuracy of each model is compared by taking the Jinan Yellow River Tunnel South Bank Foundation Pit Workhole Project as an example.The main research conclusions is as follows.(1)Summarize the historical monitoring data of the foundation pit and fill in the missing values in the data set.The 261 days of monitoring data from 102 monitoring points in the monitoring log of the foundation pit workings on the south bank of Jinan Yellow River Tunnel are summarized.Due to the influence of force majeure factors in each of the monitoring data acquisition as well as storage,some of the data have missing problems,and the incomplete data set is not conducive to the accuracy of the prediction results.Therefore,the comprehensive comparison of four interpolation networks,GAIN,P-GAIN,random forest and KNN,in the selected sample dataset filling effect.P-GAIN is based on GAIN network with the improvement and optimization of the initial interpolation method.The RMSE,MAE and MSE values of P-GAIN are 0.1308,0.0243 and 0.0171,respectively,indicating that the P-GAIN model has the highest accuracy among the four interpolation models,so the P-GAIN model is selected to complete the filling of missing values in the monitoring data.(2)Organize and analyze the monitoring data,and divide the training set and test set.Considering the indicators affecting the pit deformation and the characteristics of the monitoring data,the convergence value of the wellhead was determined as the prediction index,and the monitoring data of 101 monitoring points,such as the support pile displacement,support axial force and groundwater level,were taken as the characteristic indexes.And in order to test the prediction effect,80%of the monitoring data were divided as the training set and 20%as the test set,and the data were normalized to eliminate the difference in magnitude between the data.(3)Build the pit deformation prediction model and evaluate the prediction effect.The CNN model,LSTM model,combined CNN-LSTM model and combined VAE-CNN-LSTM model are built to predict the deformation of the working well pit.In the process of building,relevant parameters such as optimization algorithm,activation function,network regularization and learning rate in which the models can achieve optimal results are set,and MAE,MSE andR~2 are used as evaluation indexes.The results show that the prediction accuracy of the model is ranked as VAE-CNN-LSTM>CNN-LSTM>LSTM>CNN,which meanss the combined VAE-CNN-LSTM model has the best prediction effect.(4)Clear the best proportion of hidden variables.VAE can extract representative feature data from the sample dataset,output low-dimensional implied vectors,remove redundant information and values with weak correlation,and achieve condensed extraction of data information which can ensure that important features are not lost and improve the overall prediction accuracy of the combined models.As the percentage of compressed implied vectors relative to the total data affects the final prediction accuracy,the prediction results using data of 261 days and data of 150 days are compared at 20%,40%,60%and 80%,and the results show that the model has the highest accuracy at 60%,and the MAE,MSE andR~2 of the test set using 261 days of data with 60%of implied variables are 0.5622,0.5200 and 0.8921,respectively,which are higher than the accuracy using 150-day data,which further indicates that more valid historical monitoring data will improve the prediction accuracy.In summary,the P-GAIN network has extremely strong interpolation ability to achieve high-quality filling of missing data.The combined VAE-CNN-LSTM model with the implied vector at 60%can give full play to the advantages of each part of the model,fully exploit the correlation features between historical monitoring data and future trends,and obtain excellent prediction results,so as to achieve the purpose of reducing the risk of accidents,reducing safety hazards and effectively guiding construction.