ABSTRACT

This study investigated the modelling of soft sensors for the prediction of turmeric thermal properties using data-driven methodology. The study examined the effect of drying time, drying temperature and air velocity during turmeric drying using exhaustive search technique; estimated the thermal properties of dried turmeric rhizome using existing empirical relations; developed soft sensors using Artificial Neural Network (ANN), Regression Tree (RT), Support Vector Machine (SVM), Gaussian Process Regression (GPR) method for the prediction of the thermal properties; and statistically compared the goodness of the models and select a model with better prediction. Proximate composition analysis was conducted for each of the dried samples of the turmeric to determine the nutritional composition. The soft computing methods were deployed in estimating specific heat, thermal conductivity, and thermal diffusivity of the dried turmeric using four input variables time, temperature, air velocity, and relative humidity individually and collectively. Two hundred and ninety-five (295) data set out of the three hundred data set obtained from the experiment, were used to develop, train and test the models using five-fold cross-validation with five (5) of the remaining data set aside and used for independent validation of the predictive model result. The average nutritional composition of the dried turmeric rhizomes were crude fibre (2.9%), crude protein of 4.22, and carbohydrate of 33.56%. Other nutrients include nitrogen 4.22%, ash 1.6%, and fat 2.9%, with a moisture content of 4.4% and 40.4% dry matter. The result of the model indicated that the square exponential of the GPR models has the best convergence for specific heat with the combination of all the input variables. Quadratic SVM have the best prediction for thermal conductivity with the combination of all input variable. Matern S/2 with all inputs is the model with the best estimation of specific heat, having an MSE of 0.000164 and R² of 1. Quadratic SVM with all inputs best estimate the thermal conductivity with R² of 0.98 and MSE of 0.0000864. Fine Gaussian SVM is the model with the best estimate for ther9mal diffusivity having using the input variables of Time, Air velocity and temperature having MSE and R² values of 0.00037461 and 0.09, respectively. The study concluded that ANN has the best prediction for thermal properties for a single input, whereas, for all input variants, the models differ in their estimation capabilities.