Conference Agenda

Pavement condition forecasting remains limited by the reliance on costly structural measurements and curated tabular inputs, hindering scalability in large road networks. We investigate whether accurate multi-year forecasting can instead be achieved directly from raw surface geometry acquired in routine surveys. We present the first end-to-end pipeline forecasting annual Pavement Condition Index evolution from Laser Crack Measurement Systems (LCMS) data. Experiments on French national road network surveys demonstrate substantial improvements in spatial alignment and stable multi-year PCI forecasting using surface geometry alone, establishing LCMS data as a viable foundation for scalable pavement deterioration modeling.

Bridges, as critical components of transportation networks, demand reliable structural health monitoring (SHM) programs that enable quantitative assessment of their structural states and long-term performance under varying environmental and loading conditions. However, in many practical cases, vibration-based SHM projects, while effective and efficient, are short-term due to budget, accessibility, or operational constraints, leading to small vibration datasets that limit the reliability and detectability of structural damage. Such data scarcity hinders the use of machine learning models, which typically rely on abundant labelled data for robust pattern recognition and anomaly detection. To overcome this limitation, this study proposes an unsupervised deep learning methodology that integrates generative and discriminative models for enhanced damage detectability under small vibration data conditions. First, a generative neural network is employed to perform data augmentation by learning the underlying distribution of limited modal frequency samples of normal structural states and producing synthetic yet physically consistent features. Then, an unsupervised anomaly detection network, trained only on augmented healthy-state data, is utilized to identify damage-induced deviations. This combined generative–discriminative approach enables learning rich feature representations while preserving the unsupervised nature of the problem. The framework is validated using a modal frequency dataset of a cable-stayed bridge monitored within a short-term program. Despite limited modal frequency instances, the proposed method demonstrates the ability to enhance data diversity, improve class separability, and increase the sensitivity of damage indicators to structural damage.

Tower-radar computer vision (TRCV) represents an emerging application-oriented field of study. Here, image-type measurements are acquired from radar transceivers bound to the mast of wind power turbines and these radargrams subsequently get analyzed using data-driven algorithms. TRCV shows promise for reducing downtime and increasing safety of such renewable energy plants; accordingly, the respective monitoring systems should long-term perform real-time recognition of moving objects like rotor blades and their condition, disregard maintenance workers, identify birds and bats or unauthorized aerial vehicles, in order to trigger appropriate multi-faceted responses. This article focuses on the radar-only computer-vision task of classifying rotors without complementary costly instrumentation [1]. Progress in TRCV for blade monitoring remains, however, hindered by the limited publicly available data [2]. For this specialized task, recently [3] general-purpose feature extractors have proven robust to environmental and operational conditions and as a valuable building block in TRCV processing pipelines. Such large models, like OpenCLIP, have been pre-trained on internet-scale amounts of text-image pairs. They however (i) commonly still require additional data for transfer to dedicated use cases, as well as (ii) exhibit power demands or latencies incompatible with real-time resource-constrained application scenarios; models hence need to be compressed. In this paper, measured radargrams from field experiments are therefore complemented with the novel synthetic dataset SiWiRoRa as well as further open imagery. Here, several specialized image-type datasets are carefully compiled to span a context around the focal measured radargrams. Second, the main geometrical directions of this context landscape are explained by classical image statistics and with human perceptions collected from annotators. Third, large models are distilled towards lightweight capacity where it is found that both the synthetic dataset but also seemingly unrelated imagery can improve performance in blade classification. Accordingly, routes for enhancing the SiWiRoRa dataset are suggested and moreover implemented so as to further advance TRCV.

References

[1] Alipek, Sercan, et al. "Potential and Limitations of Anomaly Detection via Tower-Radar Monitoring of Wind Turbine Blades in Regular Operation with Convolutional Networks." EWSHM (2024).

[2] Mälzer, Moritz et al. "Radar-based structural monitoring of wind turbines blades: Field results from two operational wind turbines." IWSHM (2023).

[3] Kexel, Christian et al. "Mast-Bound and Too Curious: Overcoming Drift in Wind-Tower Radar for Blade Monitoring Using Pre-Training, Augmentation and Weight Consolidation Due to Correlated Conditions." LATAM-SHM (2026)

Structural health monitoring (SHM) of bridges leverages multiple sensing modalities for damage detection and localization; within this context, vibration-based approaches offer a practical and widely used source of information.
Building on this paradigm, this work proposes a conditional latent diffusion framework for spatial damage localization in masonry arch bridges using acceleration measurements.
Raw vibration signals are transformed into phase-aware time–frequency representations using the Short-Time Fourier Transform, preserving both magnitude and phase information.
Separate autoencoders are used to learn compact latent representations of spectrograms and spatial damage masks, while a supervised damage head structures the latent space according to damage severity.
A flow-matching latent diffusion model is then trained to learn the conditional distribution of damage masks given vibration-derived latent features.
During inference, the model generates spatial damage maps directly from measured acceleration data without requiring additional information about loading or environmental conditions.
The approach is evaluated on experimental data from two masonry arch bridges tested under controlled loading conditions.
Results demonstrate that the model captures meaningful relationships between vibration characteristics and spatial damage patterns, enabling physically consistent damage localization.
Performance analysis shows improved predictions for higher damage levels, while reduced accuracy in low-damage regimes reflects both dataset imbalance and limited separability in the learned latent space.
Overall, the proposed framework highlights the potential of combining phase-aware signal representations with multimodal latent diffusion for data-driven damage localization, while also emphasizing the importance of sufficiently diverse training data for robust performance under realistic monitoring conditions.