Low-strain impact testing of piles

Paper

Pdf Summary

The paper investigates using artificial intelligence to automate pile-length interpretation in low-strain impact testing. This widely used, relatively inexpensive method sends an acoustic wave through a pile after a hammer strike; interpreting the reflected signal typically requires an experienced engineer.<br /><br />The researchers trained machine-learning models to identify pile length from test signals. They compared decision-tree methods with deep learning and selected deep learning for the final model. Training and validation used a database of 590,000 test results collected by six teams over 20 years. After removing erroneous or extreme cases, about 80% of the data was used for training and 20% for validation. The signals were standardized and processed to resemble the displays viewed by human analysts. The model’s predictions were compared with expert-assessed lengths.<br /><br />On approximately 110,000 validation results, the model’s predictions correlated strongly with expert analyses (R = 0.9757), with a mean absolute percentage error of 1.6%. About 90% of predictions were within 2% of the expert estimate, and about 94% were within 5%. The model also outperformed two simple alternatives: using the planned pile length or selecting the signal’s deepest trough.<br /><br />The findings suggest that AI could support analysts, flag potential human errors, and help train less-experienced practitioners. However, the study assessed pile length only; it did not yet reliably identify structural anomalies or provide the full integrity classifications and textual analysis envisioned by the researchers. Differences between AI and expert estimates cannot always be attributed to the model, since the expert interpretation may also be uncertain. The authors emphasize that AI should not replace engineering judgment and call for further development, broader datasets, and evaluation across more testing agencies and experts.

Keywords

artificial intelligence

machine learning

deep learning

pile-length interpretation

low-strain impact testing

acoustic wave

signal processing

pile integrity

engineering judgment

model validation