AudioLS: an intelligent sorting method for drilled lotus seeds based on air jet impact acoustic signal and 1D-CNN

• Published in: Journal of food measurement & characterization Vol. 18; no. 8; pp. 6939 - 6955
• Main Authors: Lu, Ange, Yan, Zhenkun, Cui, Hao, Ma, Qiucheng
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.08.2024; Springer Nature B.V
• Subjects: Accuracy; Acoustics; Air jets; Artificial neural networks; Chemistry; Chemistry and Materials Science; Chemistry/Food Science; Classification; Data acquisition; Engineering; Food quality; Food Science; Neural networks; Parameters; Seeds; Signal classification; Signal quality; Deep learning; Sound classification; Lotus seed; Non-destructive measurement; Convolutional neural network
• DOI: 10.1007/s11694-024-02705-5
• ISSN: 2193-4126

The existence of defective drilled lotus seeds will lead to problems such as plumule residue and lotus seed appearance damage, which decreases the quality of lotus seed food products. Therefore, it is vital to sort drilled lotus seeds. Since the drilled holes on defective seeds are not coaxial with the axis of the lotus seeds, blowing an air jet along the axial direction towards the end face of defective seeds will generate a different acoustic response than that of qualified ones. Based on this characteristic, this study proposes an intelligent sorting method for drilled lotus seeds based on air jet impact acoustic signal and one-dimensional convolutional neural network (1D-CNN) acoustic classification. The method directly extracts features from 1D acoustic signals and achieves automatic classification through the constructed 1D-CNN. First, the sorting principle, acoustic signal data acquisition and preprocessing, and dataset preparation methods were introduced. Then, the effect of hyper-parameters, including the number of convolutional layers, convolution kernel size, learning rate, and training epochs, on the performance of the 1D-CNN model was investigated. On this basis, the parameters were optimized to form the final 1D-CNN model – AudioLS (Audio of Lotus Seed). The accuracy, detection time, and parameters achieved by AudioLS were 98.04%, 25.12 ms, and 0.79 M, respectively. Compared with five classic 2D-CNN models, i.e., Residual Network (ResNet) 50, Visual Geometry Group (VGG) 16, VGG19, DenseNet121, and Extreme Inception (Xception), AudioLS achieved better performance. The accuracy increased by 1.82%, 1.30%, 1.28%, 1.83%, and 2.05%, respectively, and the detection time was shortened by 16.77%, 2.71%, 7.85%, 28.11%, and 11.92%, respectively. The research results verify the effectiveness of the proposed intelligent sorting method.