Distributional SuDaI-QGAN for Vibration-Based Pavement Anomaly Detection

This study investigates learning the conditional distribution of bicycle vibration responses as the basis for pavement anomaly screening. We propose a distributional extension of the SuDaI-QGAN framework (Distributional SuDaI-QGAN), which implements this formulation through adversarial learning combined with a Gaussian negative log-likelihood objective, so that the model predicts a conditional mean (μ) and an uncertainty scale (σ) rather than a single reconstruction target, enabling likelihood-based anomaly scoring. The formulation is realized in a compact hybrid quantum–classical architecture with 357 trainable parameters. Detection is assessed with recall as the primary screening criterion, balanced accuracy as a secondary summary, and specificity as the false-alarm indicator. On a real-world bicycle surface anomaly dataset evaluated under identical experimental protocols, the proposed model attains detection performance comparable to VQTransAE (433,732 parameters); a two-sided paired t test on seed-averaged segment-level anomaly counts detects no statistically significant difference (p = 0.4275). Parameter-matched classical controls trained under identical conditions reach comparable recall and F1 at similar or smaller parameter budgets and approach the proposed model’s specificity only at substantially larger hidden sizes. The present comparisons, therefore, support conditional-distribution learning as an effective basis for vibration-based screening, with the quantum–classical circuit as one compact implementation.