Intelligent computation of moisture content in shrinkable biomaterials

A technique of intelligent computation of moisture content in shrinkable biomaterials from multiple predictors was developed. All measurable predictors were structured in three sets: biomaterial properties (volume, density, porosity, diffusivity); drying conditions (time, air temperature, humidity, velocity, pressure); and drying technologies. Two typical drying models were considered: time-dependent (thermodynamical) and time-independent (relational). The relationship between predictors and moisture content was established on the basis of multi-factorial linear regression (MLR) and neural networks (NN). Accuracy of statistical approximation was strongly dependent on drying model and chosen set of predictors. Time-independent models demonstrated better accuracy (MSE = 0.214) than time-dependent models (MSE = 0.254). Redundant predictors did not affect the accuracy and generalization ability of statistical models. Results of NN training and testing showed superior accuracy with respect to statistical models. NN worked perfectly well for any combination of non-correlated predictors, improving accuracy to MSE = 0.01. Elimination of redundant predictors further improved accuracy and generalization ability of NN models. The performance of both models was tested for drying of ginseng roots in the range of temperatures from 38 to 50 degrees C, sizes from 10 to 32 mm, and relative humidity from 12 to 40%. Due to the high accuracy and computational efficiency, NN can be used as online estimator of moisture content in drying process.