artificial neural networks: classification of sensorial

ARTIFICIAL NEURAL NETWORKS: CLASSIFICATION OF SENSORIAL ACCEPTIBILITY
BASED ON INSTURMENTAL MEASUREMENTS
CARVALHO, N. B.(1); MINIM, V. P. R.(1); ROCHA, R. A.(1); DELLA LUCIA, S. M.(2); MINIM, L.
A.(1)
(1) Departamento de Tecnologia de Alimentos, Universidade Federal de Viçosa, Campus
Universitário, CEP 36570-000, Viçosa, MG, Brazil. (2) Departamento de Engenharia Rural,
Centro de Ciências Agrárias – Universidade Federal do Espírito Santo, Alto Universitário, Caixa
postal 16, CEP 29500-000, Alegre, ES, Brazil. E-mail: [email protected]
Understanding the behavior and desires of consumers is a key tool for the success of the food
industry. However, because research and consumer trials are expensive and time consuming
procedures, it is desired to utilize instrumental measurements that may be related to sensory
acceptability, and thus can be used routinely in industry. Therefore, the objective of this study
was to classify the sensory acceptability of light cheesecurds made with different combinations
of fat and water, based on instrumental texture measurements using artificial neural networks
(ANN). Scores were used for acceptance of nine light cheesecurds formulations, which were
evaluated using the nine-point hedonic scale filled out by one hundred consumers of
cheesecurds. A back-propagation ANN was developed containing eight neurons in the input
layer corresponding to the variables related to instrumental measurements of the initial stress,
viscosity, tangent δ, firmness, gumminess, cohesiveness, chewiness and elasticity, and two
neurons in the output layer, corresponding to the Accepted (scores 6 to 9) and Rejected classes
(1 to 5). An ANN consisting of 8-3-3-2 neurons showed excellent performance in classification
of the two acceptance categories, presenting a high classification rate of cheesecurds in the
Accepted (99%) and Rejected categories (97%) for data validation, as well as a low validation
error (0.1022). Thus, it appears that the ANN presented great potential for modeling sensory
acceptance in function of the instrumental measurements of the product, with advantages such
as accuracy, speed, excellent capacity for generalization and simplicity, which makes it a
promising alternative for industrial applications.
Financial support: CNPq, FAPEMIG.