Data interpolation and characteristic identification for particle segregation behavior and CNN-based dynamics correlation modeling

Abstract

Particle segregation behavior in a binary granular bed subject to vibration has been investigated. An algorithm based on Locally Weighted Scatterplot Smoothing (LoWeSS) was developed for trajectory reconstruction and motion characteristics extraction of segregated particles. The Kriging interpolation was introduced to address the problem of the sparse spatial distribution of segregation velocity data, and the K-means clustering algorithm was used and indicated that the discrete distribution of segregation velocity data at layers of different heights in the granular bed has regionalized shape characteristics, including circular, elliptic, fusiform, and mono-symmetric shapes. Segregation velocity correlates well to dimensionless amplitude (A_d) and frequency (f_d). When A_d ∈ [0.6, 0.7] and f_d ∈ [0.75, 1], the ascending velocity of segregated particles within the lower layer of the granular bed is relatively fast, and some of the large particles initially located at the higher layer will first fall as the packing structure reorganization and then start to segregate. In addition, a data preprocessing algorithm based on Local Spatiotemporal Correlation Interpolating (LoStCoI) is developed to repair granular temperature data. The depth-wise spatiotemporal residual convolutional neural networks (CNNs) with the Spatial Pyramid Pooling (SPP) module can well characterize the correlation between granular temperature and segregation velocity. The validation errors for both the regression and classification tasks are less than 0.1, and the comprehensive evaluation index also achieves 0.9. Specifically, when provided with a sufficient amount of training data, the evaluation metrics for the regression task on the validation dataset exceed 99 %, and those for the classification task even reach as high as 99.5 %.