Redes neurais artificiais aplicadas no estudo das pressões em silos esbeltos cilíndricos metálicos: uma abordagem sobre confiabilidade estrutural

Abstract

Pressure prediction and structural assessment in silos is a topic of great interest in research studies. It can be considered that the degree of uncertainty in the calculations has had repercussions on the large number of failures and occurrences of collapses, in this type of structures in Brazil, and in the world. The objective of the present work was to develop a model for predicting pressures in comparison with experimental data. To obtain the experimental data, the pilot silo proposed by Pieper and Schutz was used, providing the basis for the DIN 1055 (1987) standard. In the silo, smooth plate walls and a flat bottom were used, varying the height/diameter ratio between 4, 6 and 8. Data values of horizontal pressures and friction of the product (maize) with the wall along the silo were observed and collected, for filling and discharge pressures. The results of the pressures obtained experimentally were inserted in an algorithm using Artificial Neural Networks (ANN) of the multilayer Perceptron type. The experimental pressures obtained on the silo walls, during the filling and discharge stages, were compared with the values generated by the ANN, as well as the values calculated by the AS 3774 (1996) and EN 1991-4 (2006) Standards. The ANN estimate was obtained with a confidence level of 90%, demonstrating the feasibility of its use in the predictions of pressures in silos. The calculated statistical indicators show the Mean Square Error values of horizontal and friction pressures: 0.074 and 0 067 respectively for filling and 0.071 and 0.072 respectively for discharge. For dimensions with lower values, the lowest values of horizontal pressures and friction were obtained. At elevation 1.25 m, the values 0.56 kPa and 1.12 kPa for horizontal pressures and 0.18 kPa and 0.41 kPa for friction pressures, being the filling and discharge values, respectively. Such values show significance in the actual data collected in the pilot silo. In comparing the methods tested, the ANN showed the best description of the predictability of pressure data.