Estimativa de parâmetros geotécnicos de resíduos sólidos urbanos utilizando algoritmos de aprendizado de máquina.

Abstract

The progressive expansion of sanitary landfills reinforces the need to develop geotechnical engineering projects that represent their mechanical behavior during its construction and maintenance. The realization of a complete program of tests to determination of the Mohr-Coulomb envelope, according to the deformation criterion, requires enough resources and time. The use of modeling and estimates is an alternative explored, with little use of machine learning. In the works of Falamaki and Shahin (2018) the Radial Basis Function (RBF) and Multilayer-Perceptron neural networks (MLP) showed good approximation results from the physical properties of the Solid Urban Waste (MSW), with a linear correlation coefficient of 0.97 and 0.89, respectively, for the effective cohesive intercept and friction angle. In the present work was carried out an approximation employing machine learning methods, requiring data collection from 61 triaxial trials from the literature resulting in 197 effective cohesive intercept and friction angle observations. the input data collected were the gravimetric composition; decomposition time; moisture content; dry specific weight; specific axial strain and maximum confining stress. You algorithms used were Extreme Learning Machine (ELM) neural networks; MLP and the Least-Squares Support Vector Regression (LS-SVR), running through the program GNU Octave. The best approximation result was obtained for the LS-SVR algorithm, being reached a coefficient of determination of 0.98 for the effective cohesive intercept and 0.97 for the effective friction angle. Among the neural networks, MLP obtained a performance superior to ELM. At the end, a representation of the sanitary landfill of Santo Tirso from the data of Gomes (2008), being performed an estimate of the envelopment of Mohr-Coulomb with neural networks and an analysis of slope stability of the profile of the embankment on the Slide. In the analyses, the differences obtained represented a discrepancy maximum of 1.8% between the calculated and the accepted safety factors.