Geotechnical Properties Inference Using MWD Data – A Probabilistic Machine Learning Approach-PPT Presentation

PPT Presentation

Pdf Summary

This presentation introduces a probabilistic machine-learning approach for inferring geotechnical properties from microtunneling measurement-while-drilling (MWD) data. Because conventional borehole-based methods provide localized information and existing approaches cannot directly derive soil properties from tunneling data, the study aims to enable continuous subsurface characterization with quantified uncertainty.<br /><br />The proposed method uses Bayesian sequential inference, implemented with Markov chain Monte Carlo and Hamiltonian Monte Carlo with the No-U-Turn Sampler (NUTS). Its performance is compared with Genetic Algorithm and Particle Swarm Optimization methods. A preprocessing workflow combines Savitzky–Golay filtering, seasonal-trend decomposition (STL), and Q-Q analysis to reduce sensor noise and address outliers.<br /><br />The approach is tested on three unlubricated microtunneling drives: two sections of the Alvarado Trunk Sewer project and the Morris Avenue tunnel. The presentation reports satisfactory inference results and compares predicted jacking forces and estimated soil properties across methods.<br /><br />The model currently assumes unlubricated drives and homogeneous soil layers. It is not directly applicable to lubricated segments, where inferred parameters may represent the combined behavior of soil and bentonite rather than native soil properties. The work could support safer and more efficient tunneling and contribute to sustainable infrastructure, sanitation, and urban development.

Keywords

probabilistic machine learning

geotechnical properties

microtunneling

measurement-while-drilling

Bayesian sequential inference

Markov chain Monte Carlo

Hamiltonian Monte Carlo

No-U-Turn Sampler

soil characterization

jacking forces