Predicting load settlement response of pile foundation using deep neural network

Abstract

The primary objective of this study is to explore the capabilities of Deep Neural Network (DNN) models in predicting the load settlement behavior of pile foundations. As the safety and stability of a structure heavily rely on the performance of the piles, it is crucial to monitor their load settlement behavior. Conventionally, static load tests are conducted at construction sites to achieve this. However, static load tests come with various complexities, such as being expensive, time-consuming, and sometimes leading to destructive outcomes. The study aims to address these challenges by seeking a more efficient, accurate, and reliable technique to capture the complete response of pile load- settlement behavior. By utilizing DNN models, the research aims to provide an alternative approach that can potentially streamline the prediction process, ultimately enhancing the understanding and management of the pile foundation behavior in construction projects. In order to predict the load settlement behavior of pile foundations, this study employed the development of various Deep Neural Network (DNN) models, including MLP, LSTM, Bi-LSTM, 1D CNN, and TabNet. A dataset of approximately 712 load- settlement data points was gathered from 42 full-scale load test data sets, encompassing relevant information about pile characteristics and soil profiles. The selected input parameters for model development included pile geometry, stiffness, applied load, settlement, loading-unloading cycles, and the SPT (Standard Penetration Test) profile of the specific location. To facilitate model training and evaluation, the dataset was initially divided into training and testing sets. Several techniques, such as batch normalization, dropout, and principal component analysis (PCA), were applied to eliminate unnecessary dimensions, reduce the impact of noise, and handle outliers effectively within the dataset. By employing these techniques, the DNN models were better equipped to process and interpret the input data accurately, aiming to improve the accuracy and reliability of load settlement behavior predictions for pile foundations. Besides DNN models, the finite element method (FEM) has been used to carry out the simulation of static pile load tests. This method has the advantage over traditional analysis techniques as more realistic test conditions can be taken into account and displacements and stresses within the soil body and pile are coupled, thus representing more realistic pile-soil interaction behavior with more realistic assumptions. The commercial finite element program Plaxis-3D was used for this simulation purpose. The layered soil profile was modeled using the hardening soil model, while the pile was modeled as an embedded beam element using the elastic model. The selection of soil models and corresponding parameters was ensured by comparing the results with the laboratory test results and SPT correlations. However, the obtained prediction results from both models were compared with field test results, and eventually performances were assessed using statistical performance indicators like MSE, RMSE, MAE, and correlation coefficients. The findings demonstrated quite satisfactory performances of the DNN models and FEM to forecast both of the two loading-unloading cycles that are commonly generated during field static load tests. It is also evident that Plaxis-3D shows higher accuracy except when the TabNet model is employed with PCA. TabNet with PCA was concluded to be the optimized model (R2=0.92) for prediction of load settlement response. Hence, it can be recommended that using the learned simple model input parameters, it is possible to predict the load-settlement behavior within a satisfactory range by applying the proposed DNN models. This will ultimately reduce time and costs by the optimization of test plans.