Estudos de correlações e comparações entre limites de liquidez de solos obtidos pelos métodos de Casagrande e cone

Abstract

This research analyzes and compares values of liquidity limits (LL) obtained by the Casagrande percussion method, LLc (hard and soft base apparatus), and by the British and Swedish cones, LLp, for different fine soils and a wide range of values. of LL. For the LLp estimates, regressions (linear and non-linear) were used between LLc and LLp and feedforward artificial neural networks (ANNs) (FNN) trained using the multilayer perceptron backpropagation (MLP) algorithm, with one or two hidden layers. Experimental values of LLc and LLp previously compiled from the literature (507 samples) were selected for statistical analysis and ANNs. The experimental results were divided into groups according to the base hardness of the Casagrande appliance and the type of cone used, and divided into subgroups to assess the influence of the LL interval. ANNs were trained with LLc, LLp, IP and SUCS classification of soils as input parameters and compared with networks of input parameters LLc and LLp (24 networks for each dataset and obtained as LLp output). Through the statistical analysis was possible to make the selection and treatment of data and eliminate outliers. Data from each group and its subgroups were submitted to regression analysis to establish linear and non-linear correlations, and to obtain the coefficient of determination (R²). Linear correlations were submitted to residual analysis and hypothesis tests to verify the normality of the model and the independence of the variables. The normality of the models was verified by the graphic analysis of the residual frequency histograms and the Normal Probability plot and by the Kolmogorov-Smirnov (KS), Shapiro-Wilk (SW) and Durbin-Watson (DW) tests. Linear and nonlinear correlations and ANNs were compared using statistical techniques that include the results obtained for the root mean square error (RMSE), mean absolute error (MAE), coefficient of determination (R²), minimum and maximum values, mean and standard deviation () of LLp estimates. Data analysis indicates that the proposed models result in very close values for the LLp prediction. Statistical tests showed that the linear correlations obtained in this research, despite the high correlation coefficients (R²>0,74), were not signifficant. The ANN results show that in addition to the variability of the geotechnical properties of the experimental results that make up the data sets, the number of samples used in the LLp prediction also influences the results. The trained ANNs have potential application for LLp estimates and represent an additional tool for conventional empirical regression methods.