Smartphone impact-sound inspection support for retaining walls: multi-hit aggregation and abstention under operator variability

抄録

Manual impact-sound inspection is attractive for retaining-wall screening because it can be applied to many blocks with simple field operation. However, when smartphone recording and hand striking are used, the measured acoustic features include both block-dependent response and operator-dependent excitation. This paper therefore examines a conservative inspection-support formulation rather than an automatic condition verdict. Using quality-gated smartphone impact sounds from 70 retaining-wall blocks with multi-operator comparisons, collected in a field measurement involving ten non-expert participants, we quantify operator variability, test whether a statically selected frequency band generalizes to unseen operators, and evaluate multi-hit aggregation and abstention in terms of cross-operator label disagreement. In the restricted maximum likelihood (REML) variance-component analysis, the estimated operator component was 34.1% for the low-frequency energy ratio (R_LF), 24.1% for the decay time (T20), 34.0% for the centroid frequency (fc), and 30.6% for the peak frequency (fpeak), values large enough to affect block-level ranking and thresholding. A frequency band that looked favorable on the selection data did not improve leave-one-operator-out generalization; its median ratio was 1.17, lower than 1.68 for the fixed 80-800 Hz reference band. Multi-hit aggregation reduced single-hit jitter, but did not remove operator-dependent label disagreements by itself. Introducing a hold state produced a risk-coverage tradeoff in terms of cross-operator disagreement: a proposed operation using 30-hit aggregation, a blended acoustic score, and hold reduced bootstrap-mean cross-operator high-/low-concern label disagreement from 20.9/70 blocks under baseline one-hit operation to 4.1/70 blocks, with coverage decreasing from 100% to 67%. The contribution is thus an evidence-linked inspection-support framework that makes uncertainty visible and suppresses overstatement from accessible acoustic observations.