Research On Improved LSTM Model And Its Application In Foundation Pit Deformation Monitoring Data Analysi

With the rapid and steady development of China’s economy,the construction of infrastructure has been promoted,leading to an increase in excavation projects in cities.Safety risks during excavation construction mainly include deformation and damage to the excavation structure and the impact on surrounding buildings/structures,which could result in serious consequences such as excavation collapse,collapse of surrounding buildings,and subway tunnel settlement.Therefore,it is necessary to construct an accurate prediction model for excavation deformation based on real-time monitoring data,which can reflect the actual deformation of excavation projects and predict their future trends accurately.In this paper,the deep excavation project of the directional Hebei subway station in Nanjing is taken as a case study.Based on the monitoring data of this excavation project,a combination neural network model based on genetic algorithm optimized convolutional neural network long short-term memory(GA-CNN-LSTM)is constructed to predict the future deformation trend of the excavation.The main research content and results are as follows:(1)The preprocessing methods for deformation monitoring data were expounded on in the article,and the principles and implementation processes of Genetic Algorithm(GA),Convolutional Neural Networks(CNN),and Long Short-Term Memory(LSTM)were analyzed and researched.(2)The engineering overview,surrounding environment,hydrogeological conditions,and foundation pit support structure of the directional Hebei subway station deep foundation pit in Nanjing were introduced in the article,and the specific monitoring contents,monitoring point layout,and other related details of the foundation pit project were elaborated on.(3)The GA-CNN-LSTM foundation pit deformation prediction model optimized by genetic algorithm was constructed in the article.The CNN network is utilized to extract the spatial features of the monitoring data,and the LSTM network was used to extract the temporal features of the monitoring data.An improved genetic algorithm was employed to optimize the number of LSTM layers and neurons,the number of fully connected layers and neurons in the model.Through experimentation,the optimal hyperparameter combination for the GA-CNNLSTM model is obtained:2 CNN layers,64 neurons per layer,2 hidden layers for both LSTM and fully connected layers,with 188,125,64,and 48 neurons,respectively.The optimization algorithm is Adam,the activation function was ReLU,the training batch size was 8,and the number of iterations is 100.(4)The impact of multi-feature LSTM prediction models on the accuracy of foundation pit deformation prediction was analyzed.The Pearson correlation coefficient method was used to analyze the relationship between multiple features,and strong correlation features were added to the single-feature LSTM prediction model to implement multi-feature LSTM prediction.LSTM network parameter optimization was conducted for different input features.Experimental analysis results showed that the multi-feature LSTM prediction model could better capture the details and features of deformation inside the foundation pit,improve the accuracy of the prediction model,compared to the single-feature LSTM prediction model.(5)Using monitoring data from the deep foundation pit construction of the directional Hebei subway station as experimental data,41 monitoring points including the deep-level horizontal displacement of retaining piles,top horizontal displacement of retaining piles,top vertical displacement of retaining piles,axial force of support,and surface settlement were taken as feature indicators,while the deep-level horizontal displacement of the foundation pit,top horizontal and vertical displacement of the piles,and cumulative changes in surface settlement were taken as prediction indicators.The pre-processed monitoring data was divided into a training set and a test set in an 8:2 ratio.The MAE,MAPE,RMSE,and R2 were selected as evaluation indicators for model prediction accuracy,and four models,including CNN,LSTM,CNN-LSTM,and GA-CNN-LSTM,were tested for their prediction accuracy.Experimental analysis results showed that GA-CNN-LSTM had higher prediction accuracy in predicting foundation pit deformation.