Determination of the Amount of Grain in Silos With Deep Learning Methods Based on Radar Spectrogram Data

• Published in: IEEE transactions on instrumentation and measurement 2021, Vol.70, p.1-10
• Main Authors: Duysak, Huseyin, Ozkaya, Umut, Yigit, Enes
• Format: Article
• Language: English
• Subjects: Artificial neural networks; Backscattering; Classification; Continuous wave radar; Convolution; Convolutional neural network (CNN); Correlation coefficients; Datasets; deep learning; Feature extraction; grain quantity measurement; Grain silos; Grain storage; Laser radar; Machine learning; Measurement by laser beam; Parameters; Radar data; Radar equipment; Radar measurement; Radar measurements; radar-level measurement; Signal processing; Spectrograms; Stacks; Surface treatment; Ultrasonic variables measurement
• ISSN: 0018-9456; 1557-9662
• DOI: 10.1109/TIM.2021.3085939

Since the grain is a crucial food source, the determination of the quantity of stored grain in silos is inevitable in terms of commercial and correct inventory planning. In this study, a convolutional neural network (CNN) is developed to determine grain quantity using the spectrograms of the radar backscattering data. The radar backscattering signals of different amounts of grain for different grain surface condition types are collected using a stepped frequency continuous-wave radar system. In the scaled model silo, a total of 5681 measurements are carried out for grain stacks with different surface patterns and different weights (0-20 kg). Then, the dataset is constituted by using the spectrograms of these radar measurements. Randomly selected 4261 data corresponding to 75% of the dataset are used for training and the remaining 1420 data are used for testing. The proposed method is compared with pretrained CNN. Accuracy of the methods is given with metric parameters for both classification and regression. The classification task results of the proposed method are obtained as 98.45% accuracy, 98.15% sensitivity, 99.07% specitivity, 98.77% precision, 98.45% F1-Score, and 97.62% Matthews correlation coefficient. The regression task results are calculated as 0.3228 mean absolute error, 0.5150 mean absolute percentage error (MAPE), 0.9649 mean squared error, and 0.9823 root-mean-squared error. The proposed method is also compared with previous studies in the literature (with 3.29 MAPE) and its superiority is demonstrated with metric parameters. The results point out that, if CNN is properly modeled and trained, the combination of CNN and proper signal processing can provide effective results in the quantity measurement applications of the grain stacks.