A comparative study for the estimation of dynamical drying behavior of Echinacea angustifolia: regression analysis and neural network

Abstract

Echinacea angustifolia is a root with curative properties and with immunological and pharmaceutical applications after drying. Studies of its drying kinetics have been limited in the past. This paper presents thin-layer drying kinetics of these roots and a comparative study between regression analysis and a multilayer feed-forward neural network to estimate their dynamic drying behavior. Experiments were performed at drying air temperatures of 15, 30 and 45 °C, with air flow velocities of 0.3, 0.7 and 1.1 m/s, and root sizes of 3 mm or less, between 3 and 6 mm, and 6 mm or more. The dependence of the drying rate constants on the drying air conditions and root sizes was modeled as an Arrhenius-type equation, using non-linear regression analysis. Four different mathematical models available in the literature were fitted to the experimental data. In addition, a three-layer feed-forward neural network was used to estimate the roots dynamic drying behavior. A 4-30-l structure provided the least errors. A back propagation algorithm was developed (using MATLAB) and applied to training and testing the network. Comparing the r, r 2, reduced Chi-square ( χ 2) and SSR values of the four models and the feed-forward neural network, it was concluded that the neural network represented the drying characteristics better than the mathematical models. Among the mathematical drying models considered, the modified Page model, was found to be more suitable for predicting drying of E. angustifolia, with the values of r = 0.997, r 2 = 0.993, χ 2 = 3.29E−4 and SSR = 0.089. However, these values were obtained as r = 0.999, r 2 = 0.999, χ 2 = 3.66E−5 and SSR = 0.011 for the feed-forward neural network based estimation. Thus, it was deduced that the estimation of moisture content of E. angustifolia could be better modeled by a neural network than by the mathematical models.