Machine learning-based pattern recognition of Bender element signals for predicting sand particle-size

This study explores the potential of integrating bender element signals with a convolutional neural network (CNN) to predict the particle size distribution of relatively uniform sand. A one-dimensional CNN analyzed time-series signals from bender elements across four sand types with particle sizes ranging from 0.5 to approximately 7 mm, under vertical stresses of 10, 50, and 150 kPa in three different cutoff frequencies (10, 50, and 100 kHz). The CNN architecture included convolutional layers augmented with batch normalization and ReLU activation functions, optimized through Bayesian techniques to enhance prediction accuracy. Experimental results demonstrated that higher stresses increased resonant frequencies and reduced arrival times of shear waves, with minor dependencies on soil type. Nevertheless, the developed CNN model well classified the four sand types at a given vertical stress and cutoff frequency, implying that the unique pattern of each sand type can be satisfactorily captured by the CNN algorithm. Overall, the framework shown in this study demonstrates that the bender element (or pattern of receiving shear wave signals) with the CNN model can be used in monitoring real-time variation of sand particle size.