Informatics Publishing Limited

Jitendra Khatti ¹, Kamaldeep Singh Grover ¹

Affiliations
1 Department of Department of Civil Engineering, Rajasthan Technical University, IN

Abstract

This technical article helps identify the optimum performance AI model for predicting compaction parameters of soil. A comparative study is mapped between regression analysis (RA), Gaussian process regression (GPR), decision tree (DT), support vector machine (SVM), and artificial neural networks (ANNs) approaches using 59 soil datasets. The soil dataset consists of soil properties such as gravel content, silt content, sand content, specific gravity, clay content, plasticity index, and liquid limit. The soil properties are used as input parameters to develop the AI model to predict soil optimum moisture content and maximum dry density. The RA, GPR, SVM, DT, and ANN models are designated as MLR_X, GPR_X, SVM_X, DT_X, ANN_X, where the X is OMC and MDD. The performance of MLR_OMC, GPR_OMC, SVM_OMC, DT_OMC, LMNN_OMC, and GDANN_OMC is 0.9714, 0.9867, 0.9689, 0.9832, 0.9435, and 0.9520, respectively. Similarly, the performance of MLR_MDD, GPR_MDD, SVM_MDD, DT_MDD, LMNN_MDD, and GDANN_MDD is 0.9512, 0.9854, 0.9482, 0.9199, 0.8679, and 0.9395, respectively. Based on the performance of AI models, the GPR_OMC and GPR_MDD models are identified as the optimum performance model to predict the soil maximum dry density (MDD) and optimum moisture content (OMC). The predicted OMC and MDD are compared with laboratory OMC and MDD, and it is found that the GPR_OMC and GPR_MDD model has the potential to predict soil compaction parameters.

Keywords

Regression Analysis, Gaussian Process Regression, Support Vector Machine, Artificial Neural Network, Compaction Parameters of Soil

Full Text

   

Jitendra Khatti ¹, Kamaldeep Singh Grover ¹

Affiliations
1 Department of Civil Engineering, Rajasthan Technical University, IN

Abstract

The present research introduces the best architectural relevance vector machine (RVM) model for predicting the compaction parameters of soil. The two types of RVM models, i.e., single kernel function-based (SRVM) and dual kernels (parallel) function-based (DRVM), have been constructed in this study. However, the RVM is a kernel function-based approach. Therefore, linear, gaussian, laplacian, and polynomial kernel functions have been implemented in these models. Each model has been optimized by each Genetic algorithm (GA) and particle swarm optimization (PSO) algorithm. For this purpose, 59 soil samples have been collected from the literature. The root mean square error (RMSE), mean absolute error (MAE), and correlation coefficient (R) statistical tools have been used to measure the performance and accuracy of models. From the overall analysis, models MC10 and MD12 have predicted OMC (RMSE = 0.8194%, R = 0.9956, MAE = 0.7920%) and MDD (RMSE = 0.1310g/cc, R = 0.9941, MAE =0.0008g/cc) better than other RVM models. It has also been observed that the DRVM model predicts the compaction parameters better than the SRVM models. The GA algorithm is robust in predicting OMC prediction, and the PSO algorithm is robust in MDD prediction. The score analysis also confirms the robustness of the dual kernel function based DRVM models for predicting OMC and MDD of soil. The sensitivity analysis demonstrates that compaction parameter prediction is strongly influenced by the specific gravity, liquid limit, and plasticity index of soil.

Keywords

Compaction Parameters, Hybrid Approach, Genetic Algorithm, Particle Swarm Optimization Algorithm, Relevance Vector Machine

Full Text