An interpretable synthetic benchmark for machine learning in low-strain pile integrity testing

This is the authors' abstract. We don't add key points for this paper.

Low-strain pile integrity testing (LSPIT) is widely used for rapid pile screening, but reflectogram interpretation remains difficult because defect morphology, axial extent, signed relative cross-sectional change, attenuation, and wave propagation jointly shape the response. This study reformulates a synthetic archive of 90,900 reflectograms as a physically stratified machine-learning benchmark. Nine benchmark views (A0–A8) isolate neutral cases, axial-domain restrictions, severity thresholds, defect-length regimes, and sign regimes. Four objectives are considered: morphology classification, sign classification, absolute-severity regression, and signed-severity regression. In addition to the representation-specific baseline grid, a controlled five-fold grouped comparison was added in which logistic/ridge models, random forests, and a profile multilayer perceptron used the same 68-dimensional reflectogram-profile vector and identical train/validation/test groups. Under this matched condition, random forests achieved morphology accuracy of 0.810±0.010 and sign accuracy of 0.858±0.002; the corresponding profile-MLP values were 0.756±0.008 and 0.850±0.007. For signed severity, random forest and profile MLP obtained R² values of 0.626±0.008 and 0.622±0.014, with mean absolute errors of 10.82±0.10 and 12.22±0.14 percentage points. A sample-size- and target-matched A5–A6 experiment confirmed that short defects consistently facilitate morphology recognition, but showed that severity difficulty is non-monotonic and depends jointly on axial length, morphology, and limited identifiability under the profile representation. A fourfold tail-weighting sensitivity analysis reduced severe-loss miss rates but increased central-range error, indicating both loss-driven and representation-limited causes of tail compression. The study is explicitly limited to synthetic benchmark evaluation; it provides a transparent basis for future, independently grounded field or high-fidelity numerical validation rather than evidence of deployment readiness.