Modeling Selected Properties of Extruded Rice Flour and Rice Starch by Neural Networks and Statistics

Cereal Chem. 83(3):223–227 Rice flour and rice starch were single-screw extruded and selected product properties were determined. Neural network (NN) models were developed for prediction of individual product properties, which performed better than the regression models. Multiple input and multiple output (MIMO) models were developed to simultaneously predict five product properties or three product properties from three input parameters; they were extremely efficient in predictions with values of R > 0.95. All models were feedforward backpropagation NN with three-layered networks with logistic activation function for the hidden layer and the output layers. Also, model parameters were very similar except for the number of neurons in the hidden layer. MIMO models for predicting product properties from three input parameters had the same architecture and parameters for both rice starch and rice flour. Food extrusion process modeling has been a difficult task due to its complexities. Various research efforts intended to model the process have been more machineand product-specific. The whole process can be viewed as consisting of a set of input parameters such as raw material characteristics, moisture content, feed rate, screw speed, screw configuration, and barrel temperature; system parameters such as residence time, specific mechanical energy, and pressure build up; and product properties such as radial expansion, mechanical properties, and chemical properties. These parameters are interdependent. Many researchers have tried to relate input parameters and output parameters, mainly using regression models to fit the experimental data. Some research efforts have concentrated on using regression analysis to predict system parameters from input parameters. Two approaches have been followed in modeling extrusion operations: dynamic modeling and steady state modeling. Dynamic modeling (Levine et al 1986, 1987) describes the reaction of a process immediately after a perturbation (10–15 sec) and is particularly useful for control and automation, while steady state modeling describes the state of the process after a period long enough for machine stabilization. Between dynamic and steady state models lies the domain of long period (a few minutes) instabilities and metastable states (Roberts and Guy 1986, 1987) that probably can be explained by qualitative models. Most studies directed toward understanding transformations in extruders have been empirical in nature. The most widely used approach is response surface methodology. This approach allows one to establish mathematical relationships between input variables and product properties (Olkku and Vainionpaa 1980; Antila et al 1983; Frazier et al 1983; Olkku et al 1984; Fletcher et al 1985). These results are clearly productand machine-specific, and the conclusions are limited to the scope of the investigations. Other approaches have been proposed. Mueser et al (1987) and Mueser and Van Langerich (1984) proposed a system analytical model for extrusion cooking of starch. Their model distinguished between process and system parameters that influenced target product properties (output parameters). Process parameters are the operating conditions that can be controlled and manipulated directly. System parameters are the properties that are influenced by the process parameters and subsequently affect the product characteristics (target parameters). It is believed that there is an appropriate function to describe the relationship between process parameters and system parameters or between system parameters and target parameters. This approach allows one to compare results obtained on the basis of more meaningful independent variables by eliminating the effects of operating conditions, materials processed, and extruder layout and geometry. In addition, the information is particularly useful for the scale-up of extrusion processes. Building on that model, many researchers have studied the relationships between process and target parameters (Taranto et al 1975; Olkku and Vainionpaa 1980; Antila et al 1983; Frazier et al 1983; Launay and Lisch 1984; Olkku et al 1984; Owusu-Ansah et al 1984; Fletcher et al 1985; Bhattacharya and Hanna 1987; Chinnaswamy and Hanna 1988). Only a limited number of studies have been reported on modeling system parameters from process parameters (Yacu 1985; Tayeb et al 1988) or building correlations between system parameters and target parameters (Guy and Horne 1988; Kirby et al 1988; Mueser et al 1987). Though regression techniques are commonly used, difficulties arise when dealing with the complex characteristics of some systems. Regression is usually limited to linear and static systems, and conventional nonlinear regression algorithms are clumsy when handling systems like the extrusion process with multiple inputs and outputs. One limitation to traditional mathematical modeling is that the mathematical relationships describing each process of the system must be closely approximated to obtain good results. Limitations in information introduce error in model predictions. Alternative techniques such as neural networks (NN) can reduce this difficulty (Batchelor et al 1997). For nonlinear problems, NN are a promising alternative technique (Borggard and Thodberg 1992). NN learn from examples through iteration, without requiring a priori knowledge of relationships between variables under investigation (Linko et al 1992; Erikaineen et al 1994). The advantage of NN over a rule-based model is that, if the process under analysis changes, new examples can be added and the NN can be retrained. This is easier than determining new models or rules. Moreover, no statistical assumptions are made on the behavior of the data. NN are not known for precision; if precision is less important than speed, NN may be useful. NN models have performed well even with noisy, incomplete, or inconsistent data (Bochereau et al 1991). Linko et al (1992) used NN with output feedback and time delays for the 1 A contribution of the University of Nebraska Agricultural Research Division, Lincoln, NE 68583. Journal Series No.13823. This study was conducted at the Industrial Agricultural Products Center. 2 MGP Ingredients, Inc., Atchison, KS 66002. 3 University of Nebraska, Industrial Agricultural Products Center, 208 L.W. Chase Hall, Lincoln, NE 68583-0730. 4 Corresponding author. Phone: 1-402-472-1634. Fax: 1-402-472-6338. E-mail: [email protected] 5 University of Nebraska, Biological Systems Engineering Department, 215 L.W. Chase Hall, Lincoln, NE 68583-0730. 6 Asian Institute of Technology, Food Engineering and Bioprocess Technology, P.O. Box 4, Pathumthani, Thailand 12120. DOI: 10.1094 / CC-83-0223 © 2006 AACC International, Inc.