Acoustic signal-based deep learning approach for smart sorting of pistachio nuts

Abstract

This study focused on developing a one-dimensional convolutional neural network (CNN) model for sorting pistachio nuts using acoustic emissions signals. The mentioned system detects open and closed-shell pistachio nuts by dropping them onto a steel plate and analyzing the acoustic signals. A total of 1600 pistachio nuts from two pistachio nut varieties were used. Proposed CNN-based model learned features directly from raw time-domain data. In addition, the performance of feature learning from frequency spectrum and combined time-frequency data also were tested. Moreover, the proposed CNN model results were compared with random forest (RF) and multilayer perceptron (MLP) neural networks with manual features extracting from the time-domain, frequency domain, and wavelet domain data. The overall accuracy (OAC) and mean squared error (MSE) of the CNN classifier with raw time data were 98.75 % and 0.01, respectively. The corresponding performance of RF with manual time domain features was OAC = 50 % and MSE = 1.56, and for MLP was OAC = 31.46 % and MSE = 2.57, respectively. The results demonstrate that the proposed deep learning method outperforms the other comparative methods. The proposed protocol consists of a smart, non-invasive, and reliable technique for the online pistachio nuts sorting systems.