| With the accelerating urbanization process,land resources are becoming increasingly tight,and in order to use land resources more effectively,the development and utilization of underground space has been accelerated,resulting in a large number of foundation pit engineering.The excavation of foundation pits will inevitably have an impact on the pits themselves and the surrounding environment,which is very likely to cause safety accidents if not controlled.With the application and development of automatic monitoring technology,the accuracy and comprehensiveness of monitoring data obtained from foundation pit monitoring are higher and the timeliness is stronger.Therefore,it is of great engineering significance to make full use of the large amount of monitoring data obtained during foundation pit construction to achieve foundation pit deformation prediction,to ensure the safety during construction,and to make early warning and countermeasures against risks in advance.In this paper,the foundation pit deformation prediction method fusing auto-encoder depth feature extraction and LSTM network is investigated and validated by engineering examples relying on the engineering of the working well on the south bank of the Yellow River Tunnel in Jinan.The main work and conclusions of the paper are as follows.（1）Organize and analyze the monitoring data of the foundation pit engineeringConsidering the characteristics of the monitoring scheme and monitoring data of the foundation pit engineering of the working well on the south bank of Yellow River Tunnel in Jinan,262 days of monitoring data were selected as the experimental data,and the key index of foundation pit deformation,the accumulated change value of wellhead convergence,was determined as the prediction index,the horizontal displacement of the top of the supporting pile wall,the deep horizontal displacement of the supporting pile wall,the horizontal displacement of the column,the vertical displacement of the column,the support axis force,the groundwater level,the internal force of the enclosing pile wall The monitoring data of 53 monitoring points are used as the characteristic indexes,meanwhile,the data pre-processing is carried out for the selected completed experimental data.（2）Building a foundation pit deformation prediction model by fusing auto-encoder feature extraction and LSTM networkThe principle of existing foundation pit deformation prediction models and their advantages and disadvantages in engineering applications were comprehensively analyzed.Based on the feature extraction capability of the auto-encoder and the advantages of the LSTM network in the prediction of temporal data,four kinds of foundation pit deformation prediction models,LSTM,SAE-LSTM,SDAE-LSTM and VAE-LSTM,were built,and the hyper parameter search for the prediction models was carried out,and the optimal The hyper parameter combinations are: the number of hidden layers of the auto-encoder is 3,and the number of neuron nodes is 48,16,8;the number of hidden layers of the LSTM network is 3,and the number of neuron nodes is 64,64,32;the optimization algorithm is Adam algorithm;the activation function is Re LU function;the Dropout retention rate is 0.5;and the training batch size is 16.（3）Engineering example analysis relying on engineering of the working well on the south bank of the Yellow River in JinanA comparative analysis of the prediction ability of univariate LSTM prediction models and multivariate LSTM prediction models shows that the prediction accuracy of the models can be greatly improved by fully considering the multiple characteristic factors affecting the foundation pit deformation.The experimental study of the four foundation pit deformation prediction models shows that the prediction accuracy of the four models is VAE-LSTM >SDAE-LSTM > SAE-LSTM > LSTM by comparing the predicted values of the cumulative change of wellhead convergence with the field monitoring data,which shows that the prediction effect and generalization ability of the foundation pit deformation prediction model can be greatly improved by the feature extraction ability of the self-encoder.The prediction effect and generalization ability of the pit deformation prediction model are greatly improved by the autoencoder.The prediction errors M SE and MAE of the VAE-LSTM model are only 0.0575 and0.1788,and the fit between the predicted and real values is 0.9904.In summary,it can be seen that the method combining the feature extraction ability of autoencoder and the ability of LSTM network to predict time-series data,especially the VAE-LSTM prediction model,can achieve accurate prediction of foundation pit deformation.It is verified that foundation pit deformation prediction model combining the depth feature extraction of auto-encoder and LSTM network proposed in this paper is feasible and reliable in application,and can be used to guide the construction of similar engineering to discover the safety hazards existing in foundation pit during the construction period in time so that corresponding measures can be taken quickly. |
