Feature selection, artificial neural network prediction and experimental testing for predicting breakage rate of maize kernels based on mechanical properties

Abstract

AbstractThe kernel breakage in harvest process is a significant problem which causes massive grain losses. In order to solve this problem, a possible solution is to predict the breakage rate and give suitable harvest time thus reducing grain losses. This study focused on using mechanical properties to predict breakage rate. Compression tests, compression–relaxation tests, insertion–retraction and stretch tests were conducted and 20 parameters were derived. Threshing tests were conducted to get breakage rate. As these parameters might be related and it is not practical to use massive parameters for prediction, the feature selection using neighborhood component analysis (FSNCA) was used for picking important parameters. The variables were compacted to an acceptable level by the FSNCA while various aspects of mechanical properties were included. The FSNCA determined the hardness of floury endosperm, rupture energy in vertical direction (x‐axis), allowable stress in lateral direction (z‐axis), hysteresis in x‐axis and elastic modulus in z‐axis as predictors. The artificial neural network (ANN) was built for prediction. The built ANN model showed an agreement in prediction results and real breakage rate (R = 0.98). The prediction error was ±1.7%.Practical ApplicationsThe mechanical properties were illustrated as potential indexes which can be utilized for predicting breakage rate of maize kernels in threshing process. The selection method reduced the dimension of predictors, which means fewer tests are required for prediction. As a result, this research proposed a possible solution for the prediction of maize breakage in harvest. Moreover, the prediction model can give guidance for optimal harvest time and thus reduce breakage rate in harvest.