Design of artificial neural network for food safety and quality during thermal processing in a can

A back-propagation artificial neural network (ANN) to control thermal processing for food safety and quality was developed. Five inputs (can sin, initial temperature, thermal diffusivity, sensitivity indicator of microorganism, and sensitivity indicator of quality) were used to predict three optimal outputs (sterilization temperature, process time, and quality degradation of the process). The thermal processing of canned food was modelled using a finite difference method. The ISIM simulation language was used to numerically solve the two-dimensional heat conduction equations for a finite cylinder and first order kinetics equation which describe the thermal inactivation of microorganisms and quality changes.;A back-propagation network was used to train and test the data generated from the simulation. A measure of dependency as well statistical tests were used to reduce the number of inputs. The results of the study were compared to another type of ANN, i.e., radial basis function (RBF). The mean relative error (MRE) was 0.2% in predicting the optimal process temperatures, 3.9% in predicting the process time, and 1.5% in predicting the quality degradation. A 4-layer neural network with 10 units in each hidden layer and 30,000 learning runs was optimum for its performance. The ANN showed high MRE (<25%) in predicting the outputs variables when tested with RBF network. The back-propagation network showed good convergence compared to RBF network which made it a better choice in designing ANN for thermal process applications.