Research On Prediction Method Of Foundation Pit Deformation Based On Spatio-temporal Correlation Under Deep Learning Framework

In view of the fact that pit deformation is random and fuzzy because of the many influencing factors,and the current situation that the traditional foundation pit deformation prediction methods are not highly accurate and have insufficient generalization ability,this paper proposes a spatio-temporal correlated pit deformation predict model coupled with convolutional neural network and long short-term memory network under the deep learning framework,taking pit monitoring data as input.Capturing pit deformation trends and predicting typical deformation indicators at key locations of the foundation pit,aiming to improve the accuracy of pit deformation prediction,and relying on Xiamen Xiang’an intercity railway foundation pit construct project to carry out engineering validation analysis of the research conducted.The main work and conclusions are as follows.（1）Analyze the current research status of existing pit deformation predict models at home and abroad and collation.Based on 286 days of monitoring data of Xiamen Xiang’an intercity railway foundation pit construction project,the accumulated change in surface settlement is determined as the prediction index,and the monitoring data of 33 monitoring points of 8 factors,including surface settlement,vertical displacement of enclosure wall,horizontal displacement of enclosure wall,deep horizontal displacement of enclosure wall,column settlement,groundwater level,support axis force and internal force of enclosure wall,are determined as the characteristic index.Analyze the influencing factors of foundation pit deformation,according to the frequency and magnitude of data changes,select the area where the key position of the pit deformation is located to study,proposing an integrated approach to missing data and anomalous data,and normalize the data afterwards to remove the effect of dimensionality,and split the data into a training set and a test set.（2）For pit deformation have temporal and spatial correlation,proposed CNN-LSTM pit deformation predict model with spatio-temporal correlation in a deep learning framework.The model integrates convolutional neural network and long short-term memory network to extract the spatial features of pit deformation with convolutional neural network to capture the temporal features of pit deformation with long short-term memory network,and finally output the prediction results of the typical deformation index of surface settlement of pit deformation with fully connected layers.The choice of model parameters has a great influence on the prediction effect of the model.In this study,the model parameters were optimized after several trials in modelling,the final number of layers of the CNN and LSTM network is set to 2,and the neuron nodes are set to 64,the optimization algorithm was Adam,the regularization rounding rate was0.2,the learning rate was 0.001,the training batch size was 24,and the number of epochs was set to 100.（3）Set up a controlled experiment with monitoring data measured at the site of Xiamen Xiang’an intercity railway pit construction project as input,set up a time-series based MLP model and an LSTM model,as well as a CNN model based on the null domain,compare with the prediction results of the CNN-LSTM spatio-temporal correlation model studied in this paper.The MAE,MSE and MAPE error evaluation metrics are introduced to test the accuracy of the models,analyze and compare the predicted and true values of the different models,to verify and evaluate the prediction effectiveness of the CNN-LSTM convolutional long-short spatiotemporal correlation model.The experimental results show that the MAE,MSE and MAPE metrics of the CNN-LSTM model are only 0.155,0.076 and 0.63,and the mean value of the absolute error between the predicted and true values of the model reaches 0.041,which proves that the prediction model considering both temporal and spatial features has better prediction effect than the model considering only temporal or spatial single features.The spatio-temporal correlation pit deformation prediction method proposed in this paper can effectively improve the accuracy and precision of pit deformation prediction.（4）Consider that there is an order problem between the extraction of spatial features and temporal features by neural networks switch the order of combining convolutional neural networks and long short-term memory networks and re-modeled.The LSTM-CNN long shortterm convolutional neural network model was constructed and compared with the CNN-LSTM convolutional long short-term memory network model to analyze the effect of the order of network combination on the accuracy of the model.The MAE,MSE and MAPE metrics of the trained LSTM-CNN model were 0.204,0.096 and 0.899 respectively,and the mean value of the absolute error between the predicted and true values of the model was 0.127.The comparison results showed that the prediction results of the LSTM-CNN model were inferior to those of the CNN-LSTM model,which showed that the network model with spatial feature extraction first had better prediction better.The comprehensive analysis shows that the CNNLSTM model has the best prediction effect and is highly applicable in the field of foundation pit deformation predict.