Distribution-based unsupervised detection of bridge damages under variable vehicle passages using time-frequency feature fusion

Unsupervised detection of bridge damages from vehicle-induced vibration is complicated by passage-to-passage variability and scarce damage labels. This study proposes a distribution-based framework using acceleration from one mid-span accelerometer, with vehicle entry/exit times and speed additionally required for event extraction and spatial alignment. For each passage, a dual-branch Conv–LSTM–attention autoencoder trained only on healthy data jointly encodes the spatially aligned time-domain response and the passage-level FFT magnitude spectrum. This fusion combines transient/local response information with global spectral characteristics. Frequency-domain reconstruction errors from multiple passages are treated as an empirical distribution, whose deviation from a healthy reference is quantified by the second-order Wasserstein distance. The framework is evaluated on a full-scale laboratory bridge under controlled but variable vehicle passages, with condition changes introduced by guardrail removal, and on a calibrated vehicle–bridge finite-element model with localized stiffness reductions. Compared with the frequency-only model examined herein, the fused model provides clearer separation between baseline and altered states. For the investigated cases, the index increases monotonically with imposed change severity, shows location-dependent sensitivity, and becomes less variable as passages accumulate, with approximate stabilization at 200–260 passages. The results support the feasibility of detecting global bridge condition changes using one bridge-mounted accelerometer plus auxiliary passage information, but do not establish damage localization or generalization to operational traffic.