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

Area of Science:

• Geotechnical Engineering
• Machine Learning Applications
• Signal Processing

Background:

• Particle size distribution is crucial for soil behavior.
• Bender elements measure shear wave propagation in soils.
• Convolutional Neural Networks (CNNs) excel at pattern recognition in complex data.

Purpose of the Study:

• To investigate the integration of bender element signals and CNNs for predicting sand particle size distribution.
• To develop and optimize a CNN model for classifying different sand types based on bender element data.
• To assess the feasibility of real-time monitoring of sand particle size variations.

Main Methods:

• Utilized time-series signals from bender elements for four sand types (0.5-7 mm particle size).
• Applied a one-dimensional CNN with batch normalization and ReLU activation, optimized via Bayesian techniques.
• Conducted experiments under varying vertical stresses (10, 50, 150 kPa) and cutoff frequencies (10, 50, 100 kHz).

Main Results:

• Higher vertical stresses increased resonant frequencies and reduced shear wave arrival times.
• The CNN model successfully classified the four sand types across different stress and frequency conditions.
• The unique signal patterns of each sand type were effectively captured by the CNN algorithm.

Conclusions:

• Bender element signals combined with CNNs offer a viable method for predicting sand particle size distribution.
• The developed framework demonstrates potential for real-time monitoring of soil particle size.
• CNNs can effectively interpret complex signal patterns from bender elements for geotechnical applications.