Single-Sensor AI Spots Bridge Damage From Bumpy Vehicle Rides

Engineers have long dreamed of checking the health of thousands of aging bridges without blanketing them in expensive sensor networks. A new study from Chongqing University and partner institutions reports progress toward that goal: an unsupervised artificial intelligence framework that detects structural changes in a bridge using acceleration data from a single mid-span accelerometer, supplemented only by vehicle entry and exit times and travel speeds. The work, published in Frontiers in Built Environment, was tested on a full-scale laboratory bridge and on a calibrated computer model, and it points to a practical middle ground between dense instrumentation and no monitoring at all.

The central obstacle the researchers set out to solve is variability. When a vehicle rolls across a span, the resulting vibration depends on the vehicle’s weight, speed, axle spacing and trajectory, as well as traffic flow and environmental conditions. Each passage is therefore a unique excitation, and the reconstruction errors that conventional anomaly-detection systems rely on can swing wildly from one event to the next. That stochasticity can trigger false alarms even when a bridge is perfectly intact, which has limited the real-world use of reconstruction-based autoencoders on bridges carrying normal traffic.

The team’s answer combines two strategies that had mostly been pursued separately. First, instead of judging a bridge by a single vehicle crossing, they treat the frequency-domain reconstruction errors from many passages as an empirical probability distribution. The deviation between that distribution and a healthy reference distribution is quantified with the second-order Wasserstein distance, a statistical measure that needs no parametric assumptions about the data and remains well defined even when the two distributions do not overlap. Second, rather than feeding the model only a frequency spectrum, they fuse two complementary views of the same acceleration record: the spatially aligned time-domain waveform, which preserves transient and local response details, and the passage-level fast Fourier transform magnitude spectrum, which more stably reflects the structure’s global dynamic behavior.

The learning machine at the heart of the framework is a dual-branch Conv-LSTM-attention autoencoder trained exclusively on healthy-state data. Each branch processes its input through a one-dimensional convolutional network for local temporal modeling, a double-layer stacked long short-term memory network for long-term dependencies, and an additive attention mechanism that adaptively weights key temporal segments. The encoded feature vectors from the two branches are then concatenated and projected into a unified 128-dimensional latent space, from which two independent decoders reconstruct the original time-domain and frequency-domain signals simultaneously. Because the time-domain reconstruction loss backpropagates through the fusion module to both encoders, the frequency-domain reconstruction ends up incorporating information from both inputs, linking waveform learning directly to the errors used for damage assessment.

To prove the gain really comes from fusion rather than from a bigger network, the researchers ran a minimal ablation analysis. They compared their Fusion-F model against a Single-F model with identical encoder-decoder depth, layer types, total parameter count and optimization procedure, in which the time-domain branch was retained but its input was set to zero and its reconstruction loss given zero weight. Under minor structural damage, the Fusion-F model’s error distribution shifted noticeably rightward relative to its healthy baseline, while the Single-F model showed a more limited shift with substantial overlap persisting between healthy and damaged states. The fused model also produced a lower baseline distance under healthy conditions, reflecting a more stable reconstruction-error distribution, and a more pronounced increase in Wasserstein distance under damage.

The experimental validation took place in the Bridge and Rail Vibration Laboratory at Chongqing University, on a full-scale simply supported T-beam bridge roughly 25 meters long and 7.2 meters wide, built from three parallel prefabricated T-beams. Because the structure was newly constructed, it was considered healthy at the start. Damage was simulated without harming primary components by altering the guardrails: in damage condition 1, a 5-meter section weighing 2,143 kilograms near mid-span was cut away, and in damage condition 2, a total of 10 meters weighing 4,286 kilograms was removed. A mid-span accelerometer recorded vertical acceleration continuously at 100 hertz, with each sample covering one complete vehicle passage from entry to exit.

The traffic itself was deliberately varied to mimic real operating conditions. Three vehicle types, a pickup truck, a passenger car and a van with nominal weights near 2,120, 1,910 and 1,665 kilograms, crossed the bridge in single, successive and simultaneous configurations, with ballast blocks of 0.2 to 1.0 ton randomly added to vary the loads. Six hundred test runs were performed for each of the three bridge states, yielding 1,800 samples in total, with recorded speeds ranging from 4 to 45 kilometers per hour and averaging 24.9 kilometers per hour. Of the 200 healthy-state samples, 160 were used for training and 40 for validation, and no damaged-state data entered the model at any point.

The results showed a consistent pattern. At the individual event level, differences between healthy and damaged spectra remained subtle, underscoring why event-based indicators are unreliable under stochastic traffic. At the distribution level, however, the second-order Wasserstein distance increased monotonically as damage escalated from the 5-meter to the 10-meter guardrail removal, and the fused model separated baseline from altered states more clearly than the frequency-only model examined in the study. In the numerical extension, a finite element model of the same bridge, calibrated against the experimental specimen and loaded with simulated two-axle vehicles on Class A to C road roughness, tested localized stiffness reductions of 10, 20 and 30 percent in a four-element mid-span region. The damage index again rose monotonically with the degree of stiffness loss.

Location mattered in a physically meaningful way. When stiffness reductions of varying severity were introduced at the quarter-span, one-third span and mid-span positions, the index increased progressively from the quarter-span toward the mid-span, with the strongest response at mid-span, where bending moments and modal displacement amplitudes are highest. The authors are careful to note, however, that because damage locations were specified in advance, the comparison does not demonstrate localization of an unknown defect from measured response. A Monte Carlo sampling study also established a practical data requirement: the damage index fluctuates considerably below roughly 140 vehicle passages in the experiment and stabilizes beyond approximately 200 to 260 passages, offering a guideline for how long field monitoring must accumulate data.

The researchers frame the contribution cautiously. The framework supports the feasibility of detecting global bridge condition changes with one bridge-mounted accelerometer plus auxiliary passage information, but it does not establish damage localization or generalization to operational traffic. Its sensitivity to highly localized or incipient damage such as micro-cracks or early-stage corrosion remains to be fully established, and long-term environmental variability from temperature and humidity has not been systematically accounted for. Future work, the team writes, should validate the approach against more realistic deterioration mechanisms including distributed cracking, progressive corrosion and fatigue, develop compensation strategies to decouple operational variations from genuine damage indicators, and extend the method to sparse sensor arrays that could improve spatial localization for larger or geometrically complex bridges.

Subject of Research: Unsupervised detection of bridge structural changes from single-sensor vehicle-induced vibration data

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

Article References: Wang, Z.-L., Xing, J., Ding, H.-X., Tan, Z.-X., Fang, Y.-L., & Liu, Z.-Y. (2026). Distribution-based unsupervised detection of bridge damages under variable vehicle passages using time-frequency feature fusion. Frontiers in Built Environment, 12, Article 1937630. https://doi.org/10.3389/fbuil.2026.1937630

Image Credits: AI Generated

DOI: 10.3389/fbuil.2026.1937630

Keywords: bridge health monitoring, structural health monitoring, unsupervised learning, autoencoder, time-frequency fusion, Wasserstein distance, vehicle-bridge interaction, damage detection, single-sensor monitoring, reconstruction error, deep learning, civil engineering