Keynote Lecture: Assessing the Potential of CFA Piling Rig Telemetry Data for Stratigraphic Profiling and Optimised Pile Installation-Paper

Paper

Pdf Summary

This study investigates whether machine-learning analysis of Continuous Flight Auger (CFA) piling rig telemetry can identify the transition from superficial soils to bedrock during installation. Accurate, real-time rock detection could reduce overly conservative pile lengths while maintaining safe rock sockets, lowering material use, costs, drilling time, and carbon emissions.<br /><br />Researchers analysed telemetry from approximately 1,900 piles across two UK projects, using measurements such as torque, crowd force, penetration resistance, and rotary drive pressure. They manually labelled rock boundaries and tested supervised and unsupervised models. Random Forest performed best, classifying soil and rock with 98.4% accuracy. K-means clustering achieved 92.6% accuracy and, although less precise, offers a way to explore large historical datasets without extensive manual labelling. Mutual-information feature selection identified influential telemetry variables and helped reduce model complexity.<br /><br />Predicted depth-by-depth classifications were converted into pile-specific rock boundaries. A rule requiring four consecutive readings produced a practical balance between detecting rock too early—which could compromise socket adequacy—and drilling unnecessarily far into rock. Spatial checks against neighbouring piles were proposed to flag anomalous predictions for review. Processing was fast enough for real-time use.<br /><br />Retrospective application suggested potential reductions of 5.5% in concrete volume and 8.12% in drilling time, with estimated savings of about 150 tonnes of CO₂e. However, performance declined when models were applied across projects, highlighting the influence of different rigs, operators, and ground conditions. The study was limited to two sandstone sites from one contractor; labels were partly subjective, and live deployment and pile capacity were not directly validated.<br /><br />The authors recommend site-specific supervised models for live projects and unsupervised methods for historical-data analysis. Broader testing and collaboration to share datasets could improve generalisability and support safer, more efficient CFA pile design and installation.

Keywords

machine learning

CFA piling

telemetry

bedrock detection

rock boundary

Random Forest

k-means clustering

pile installation

geotechnical engineering

carbon emissions