Conditional diffusion model improves bridge monitoring anomaly detection

A conditional diffusion model designed to identify abnormal data in bridge structural health monitoring achieved a precision of 0.971, recall of 0.980, F1-score of 0.975 and accuracy of 0.978 on a constructed test dataset from a long-span cable-stayed bridge. The result, reported in a 2026 research article, comes from experiments using 90 consecutive days of multi-sensor measurements, with most abnormal test windows injected synthetically and the remainder drawn from maintenance logs and expert inspection. The study concerns the quality and interpretation of monitoring signals rather than the diagnosis of structural damage, and its findings remain confined to the bridge, dataset and anomaly types evaluated.

Bridge monitoring systems continuously record responses such as acceleration, strain, displacement, deflection and temperature, producing dense time-series data that can support condition assessment and warning systems. Those records can also be degraded by sensor faults, transmission errors, missing observations, noise, drift and abrupt fluctuations. If abnormal readings are treated as evidence of structural deterioration, they may generate unnecessary inspections or maintenance alarms; conversely, poor-quality data may conceal genuine deterioration. The researchers therefore framed anomaly diagnosis as the task of separating condition-consistent normal behavior from data that violates expected temporal, cross-sensor and environmental patterns.

The proposed framework begins by dividing raw records into fixed-length windows and learning the distribution of normal bridge responses. Rather than requiring a large collection of labeled abnormalities, it trains mainly on normal or manually verified high-quality data. A forward diffusion process gradually adds Gaussian noise to each sequence, while a learned reverse process reconstructs a condition-consistent signal. The underlying objective is a conditional normal-state distribution, allowing the model to learn how temperature, operating context and sensor location shape normal behavior instead of assuming that every unusual fluctuation is anomalous.

Each input window contains measurements from multiple channels, and the model embeds both individual observations and sensor identities, including geometric information derived from sensor coordinates. A Transformer-based temporal encoder then captures long-range dependencies and relationships among sensors. The condition branch encodes temperature, sensor location and a time index in an 80-dimensional vector, which is fused with the temporal representation through adaptive normalization and concatenation. In this design, normal thermal deformation that appears coherently across channels can remain close to the reconstructed state, while missing values, drift or sudden changes that break expected relationships produce larger discrepancies.

The denoising backbone comprises four residual temporal blocks with one-dimensional convolutions, group normalization, SiLU activation, sinusoidal diffusion-step embeddings and condition-driven adaptive layer normalization. It uses a hidden dimension of 128 and was trained with 200 diffusion steps under a cosine noise schedule. During inference, a test feature sequence is partially diffused to step 80 before iterative conditional denoising reconstructs it. This starting point retains information from the input, allowing the resulting difference between measured and reconstructed signals to remain informative about abnormality rather than comparing every observation with a signal generated independently from pure noise.

Anomaly decisions combine two forms of reconstruction error: pointwise signal-level discrepancy and a difference measured in the learned temporal feature space. A validation-set grid search selected weights of 0.65 for the signal term and 0.35 for the feature term, indicating that raw reconstruction error remained the dominant measure while feature-space differences added sensitivity to gradual changes. The threshold was set from the mean and standard deviation of scores for normal validation windows, using a coefficient of three, rather than being chosen as a single fixed value for every operating period. The authors also normalized condition variables using normal training data and excluded anomaly labels from the condition branch.

Evaluation used 38,520 windows of 256 points each, advanced with a stride of 128 points. The records came from 18 synchronized channels sampled effectively at 1 hertz: six strain, four deflection or displacement, four acceleration and four temperature channels. After a 60%/20%/20% split, the test set contained 6,548 normal windows and 1,156 abnormal windows, giving an anomaly proportion of about 15%. Missing intervals, impulse noise, sudden jumps, drift and local trend deviations were injected into normal windows for roughly 70% of the abnormal cases. About 30% were identified as real abnormal segments through maintenance logs and expert inspection, so most of the evaluation set did not represent independently observed field faults.

The conditional diffusion method outperformed a statistical threshold approach, Bi-LSTM, CNN-LSTM, a variational autoencoder, a GAN, a Transformer reconstruction detector and an unconditional diffusion model trained and tested on the same data. Each experiment was repeated five times with different random seeds, the reported table contained mean performance, and the standard deviation of F1-score was below 0.006 for all deep models. A paired t-test against the strongest baseline produced a reported value of p<0.05. The unconditional diffusion comparison was especially relevant to the model’s central design: retaining the same denoising backbone while removing temperature, location and time-index conditions reduced performance, supporting the authors’ interpretation that explicit conditional guidance helps separate normal environmental variation from abnormal monitoring states.

On an NVIDIA RTX 3090 graphics processor with a batch size of one, the proposed method required 6.4 milliseconds per data window, compared with 5.9 milliseconds for unconditional diffusion and 0.84 milliseconds for the Transformer detector. The model was therefore substantially heavier than single-forward-pass alternatives, although its processing time was far shorter than the 128-second stride used in the study and was presented as feasible for front-end data-quality control in this windowed setting. The results are not evidence of deployment on a live bridge or across different bridge types, and the researchers identified efficiency, cross-bridge transfer and integration of structural physical mechanisms as future work. The underlying monitoring data are available from the authors upon reasonable request, while no code was reported as available.

Several limitations temper the headline performance figures. The study used one cable-stayed bridge and a monitoring period of 90 days, and most test anomalies were synthetic. Simulated perturbations can resemble specified fault patterns but cannot reproduce every source of field degradation, interaction among faults or behavior under changing sensor configurations. The constructed test set also had a stated anomaly ratio of about 15%, so its results do not directly establish performance in a real monitoring stream with a much lower prevalence of abnormal windows. Ablations nonetheless showed that each major component contributed on this dataset: removing condition input lowered F1-score from 0.975 to 0.952, replacing temporal multi-sensor extraction lowered it to 0.936, and using only signal-level error or a fixed threshold also reduced performance. The work therefore offers a technically detailed proof of concept for screening bridge-monitoring data, not proof that the model can diagnose structural damage or support maintenance decisions without further validation.

Subject of Research: Conditional diffusion models for anomaly diagnosis in bridge structural health monitoring data

Article Title: Research on the application of deep learning in bridge health monitoring data analysis and anomaly diagnosis

Article References: Feng, Z., Wang, H., Tang, G., & Shao, J. (2026). Research on the application of deep learning in bridge health monitoring data analysis and anomaly diagnosis. Discover Artificial Intelligence, 6(1), Article 1096. https://doi.org/10.1007/s44163-026-02173-4

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02173-4

Keywords: bridge health monitoring, diffusion models, anomaly detection, structural monitoring, sensor data, time-series analysis, conditioned denoising, data quality, deep learning, Research, application, deep