Neural Network-Based Classification for Walnut State Using Microwave
Scattering Parameters
Abstract
Recent changes in lifestyle have a significant impact on dietary
preferences. Nuts (walnuts, hazelnuts or pistachio) becomes essential to
meet daily nutritional requirements. Furthermore, walnuts are widely
preferred in industry; however, critical food safety issues arise due to
aflatoxin contamination caused by the mixing of rancid and healthy
walnut kernels during the shelling process. Therefore, it is crucial
to classify walnuts based on their kernel state (rancid and healthy)
prior to shelling, in order to improve shelf life and time efficiency.
In this study, it was aimed to classify unshelled walnuts using
microwave propagation parameters: transmission coefficients (TCs) and
reflection coefficients (RCs). For the classification, the samples were
divided into three groups: glued-unshelled walnut (GSW),
healthy-unshelled walnut (HSW) and rancid walnut (RW). RCs and TCs from
these groups were measured between 7-12 GHz employing Vivaldi antenna.
Obtained TCs and RCs were used to predict the walnut state through
neural network. The accuracy was examined for TC and RC, separately,
based on three different conditions: (1) training dataset ratio, (2)
frequency range and (3) sample location. The minimum error rates for RW
and HSW using RCs were 7.33%, 16.92% and 7.92%, for training
ratio: 0.8, the frequency range: 11.5-12 GHz, sample placement: center
point, respectively.