Neural network predictive process modeling: Applications to food processing

Jeffrey John Rattray, Purdue University

Abstract

Current methods to optimize food processing systems are based on known levels and control of significant input variables. However, in many cases, raw material characteristics change continuously due to outside factors that cannot be controlled. There is a need to predict optimal process settings based on current material properties and past experience. Neural networks, a developing area of artificial intelligence systems, are capable of approximating complex mathematical functions and generalizing from training data to situations that have not seen before. Split-input modeling, a new approach, creates neural network models using past performance data that predict optimal process settings directly from raw material parameters and desired (target) product output attributes. The overall goal of this research was to develop modeling systems that could be used to predict process control settings. The hypothesis was that split-input modeling, which rearranges the flow of data in process modeling, can be successfully applied to model and optimize food processing systems. Data were obtained from the Purdue Enology laboratory for a wine fermentation process. Split-input models for the wine fermentation process were developed with modified functions from the Matlab neural network toolbox. The development platform was a 300Mhz Pentium II computer running Windows NT 4.0 with 128 MB of RAM. The error distributions of these models were analyzed to identify nonpredictable process settings, which were removed. Methods for optimizing the geometry, initialization, and training of the neural network models were developed and applied. The final split-input models showed good predictive ability over the test data set, as measured by the sum of squared errors (SSE). Neural network results showed that yeast supplement concentration had only a minor effect on the wine fermentation process outputs, and that desired values for all other process settings could be predicted accurately. Predictions were shown to be 95% accurate for discrete variables, ±1.3 kg for sugar, ±2.65 liters for water, and ±0.089 g/liter for yeast substrate concentration.

Degree

Ph.D.

Advisors

Floros, Purdue University.

Subject Area

Electrical engineering|Computer science

Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server
.

Share

COinS