Neural network modeling and optimization for spray-drying coconut milk using genetic algorithm and particle swarm optimization

Developing an accurate model of the spray drying coconut milk process is a complicated procedure. It involves application of engineering knowledge to describe the relationship between the processing conditions and the powder properties. The complexity of the both factors reduce white-box modelling accuracy in spray drying coconut milk modelling. As an alternate, neural network modelling with optimization technique is an alternative method that provide an accurate model of the system. The objective of this research is to develop and compare various ANN spray drying coconut milk models. Firstly, using MATLAB program, the ANN model is developed based on optimized topology and is then furthered optimized by genetic algorithm (GA) and particle swarm optimization (PSO) using MINITAB program. Using a rotational central composite design, the model development process is based on 20 experimental data consisting of inlet temperature (140°C-180°C), concentration of maltodextrin and sodium caseinate (0 %w/w – 10 %w/w), which are established as the input parameters. Moisture content (3.64%-5.1%), outlet temperature (76.5°C-104.5°C) and surface free fat percentage (0.35 mg/100g-34.51 mg/100g) are the output parameters for the neural network. Effect of spray drying parameter on the powder quality is further analyzed using response surface methodology (RSM) method. The ANN model topology is designed using selection from the best training algorithm, transfer function, number of training runs (1000-5000), number of hidden layers (1-3) and nodes (5-15). The ANN model is further improved using GA and PSO. Each algorithm has its own parameters and is further optimized using RSM. Firstly, minimizing all three responses of the coconut milk powder leads towards the spray drying of coconut milk at the inlet temperature of 140°C combined with the concentrations of maltodextrin and sodium caseinate at 8% and 5% (w/w) and was recommended as the condition for RSM optimization. Using statistical method of highest R2 value and lowest MSE value, the ANN most optimum topology model consists of K-fold cross validation implements the Levenberg-Marquart training algorithm with hyperbolic tangent sigmoid transfer function using 4500 times of training runs with optimal topology configuration of 3-8-2-3. Integration of global search algorithm into ANN model further improved the model performance. The optimized selected GA parameters values are at maximum population size (100), minimum crossover rate (0.2) and maximum mutation rate (1.0). The obtained PSO parameters chosen are recorded at optimum value of C1 (4.0), C2 (0) and number of particles (100). GA-ANN model outperformed ANN and PSO-ANN model as GA-ANN recorded the lowest MSE value and highest R2 value. In engineering application wise, all four models are tested against external datasets to prediction accuracy and generalization capacity of all models, leading towards cost and time reduction in model development. Using linear regression analysis and comparative error analysis (MSE, R2, SEP and MPE), GA-ANN model outperformed in all dependent variables and achieved lowest MSE, SEP and MPE values and highest R2 values. This showed that GA-ANN model has the best prediction model for the spray drying of coconut milk system.

Text JESSE LEE KAR MING - IR.pdf Download (1MB)

Actions (login required)

View Item