Snack food frying process input-output modeling and control through artificial neural networks.

摘要

Automatic control can avoid the overreaction of human operators and ensure the consistency of the product quality. The snack food frying process is a complex process with nonlinearity, multivariate interactions, and a long time-lag. The input-output modeling strategy of system identification is taken to utilize artificial neural networks to deal with these characteristics of the process.;A type of multilayer feedforward network with direct linear connections between input and output layers is introduced to evaluate the relative contributions of linear and nonlinear components in the process dynamics. Evaluation and analysis of this network along with the regular multilayer feedforward network on the process input-output data lead to the conclusion that neural networks can characterize the process well. Further, a procedure for neural process model identification is established and applied to identify SISO and MIMO neural process models based on the cross-validation of training and testing data with the neural model complexity. For the purpose of control, neural process one-step-ahead and multiple-step-ahead predictors are established. The neural process multiple-step-ahead predictions are performed through the external recurrent neural network which is trained by the algorithm of backpropagation through time. Based on the neural process one-step-ahead prediction model, a design algorithm of an internal model controller is developed with iterative inverses at each sampling instant using a modified version of Newton's method and gradient descent method. The simulated internal model process controllers are tuned using a procedure established with three integral error objective functions. Based on the neural process multiple-step-ahead prediction model, a design algorithm of a predictive controller is developed with the on-line optimization using an approximate conjugate direction method which is free from gradient and one-dimensional search calculations. The simulated predictive process control actions depend only on the calculations of the different values of the designated objective function.;This research is a comprehensive treatment in neural network process modeling and control in food processing engineering. The developed methodology is capable of handling problems in modeling and control for the given process. The outcome of this research is expected to extend to other similar processes in biological product processing.