Smart infrastructure systems require monitoring frameworks that can function autonomously, remain effective under varying operational conditions, and issue dependable warnings without requiring labeled damage information. This study presents an unsupervised structural health monitoring (SHM) framework for transportation infrastructure that integrates LLM-inspired representation learning with a gated multi-GPT fusion autoencoder for robust intact-only anomaly detection using networked strain measurements. In the proposed pipeline, raw multi-channel strain-gauge signals are first processed through an adversarial autoencoder (AAE) to extract compact and informative feature representations, which are then fed into the downstream model. Multiple GPT-based reconstruction experts analyze these segmented AAE-derived features in parallel, while a compact gating network adaptively combines their outputs. The proposed workflow supports calibration using only intact-condition data, adaptive threshold setting, and continuous monitoring under non-stationary loading scenarios. The framework is validated on a laboratory pedestrian bridge instrumented with multichannel strain gauges and exposed to walking-induced excitation, where it achieves high anomaly detection performance across two damage levels. Robustness studies under both white and colored noise, together with threshold sensitivity analysis, confirm stable class separability and reliable operational performance. Comparative evaluation against leading unsupervised anomaly detection methods further highlights improved robustness to correlated noise and a reduced number of false alarms, demonstrating a scalable sensing-to-decision solution for resilient SHM.

The safe and reliable operation of natural gas compressor units is crucial for ensuring a secure and stable gas supply. However, the interaction between natural gas and pipelines inevitably induces flow-induced vibrations, leading to long-term cyclic stress variations in the pipelines. This results in the accumulation of fatigue damage and the initiation of cracks in stress concentration areas, thereby threatening the health of the compressor unit system. In this context, this study proposes a deep learning framework for predicting flow-induced vibration and fatigue life in pipelines, which is based on numerical simulations and experimental measurements, enabling accurate prediction of pipeline vibration responses and fatigue states under different gas transmission conditions. This study first conducts modal analysis on numerical and experimental models, updating the finite element model using the first three natural frequencies to accurately reflect the dynamic characteristics of the experimental model. Subsequently, a deep learning approach based on a multi-layer perceptron is proposed for predicting pipeline flow-induced vibration responses. The time-frequency acceleration responses of the pipeline under different gas flow conditions are predicted via numerical simulation, and the corresponding strain fields at stress concentration locations are evaluated using a quantitative relationship with the predicted accelerations. The results demonstrate that the proposed deep learning framework, leveraging simulated and measured pipeline data, enables accurate prediction of flow-induced vibrations and fatigue life.

Structural Health Monitoring (SHM) aims to monitor in real-time the health state of engineering structures. For thin structures, Lamb Waves (LW) are particularly effective for SHM applications. A bonded piezoelectric transducer (PZT) generates LW in the form of a short tone burst, creating an initial wave packet (IWP) that propagates through the structure while interacting with boundaries, defects, and other features. These interactions leave precise yet complex signatures in the signals recorded by the sensors. In practice, however, extracting the specific signature associated with damage is often challenging for traditional SHM signal-processing methods, making reliable damage detection difficult.

To address these limitations, we employ a Deep Learning–based approach known as the Rank Reduction Autoencoder (RRAE). The RRAE is an autoencoder whose latent space is constrained to follow a low-rank SVD structure, ensuring that it captures only the most salient features of the measured signals. In this work, we extend the original concept by enforcing the SVD modes to be predictive of both damage location and severity. This is achieved through an additional neural network that takes the reduced latent representation produced by the RRAE and directly estimates the damage characteristics. The joint training of the RRAE and the feature-extraction MLP therefore yields, at convergence, a robust and powerful tool for damage detection.

The proposed technique is demonstrated on two case studies. The first, a more academic example, involves damage detection on a thin plate. The second is based on an experimental campaign conducted by CETIM and focuses on a 2-meter pipe segment, where the proposed architecture performs damage detection using measurements of acoustic wave emissions.

The liquid rocket engine is a complex system characterized by strong coupling among multiple components. The occurrence and progression of faults within it exhibit dynamic propagation characteristics in both temporal and spatial dimensions, posing severe challenges to the systematicity, early detection, and accuracy of diagnostic methods. To address the bottleneck of traditional methods in characterizing the spatiotemporal propagation patterns of faults, this paper proposes a Propagation-Aware Dynamic Graph Network (PADGN). This method abstracts the engine system into a dynamic graph model, where nodes represent key components or sensor measurement points, and edges characterize the physical connections and functional dependencies between them. The core of the model consists of three innovative modules: First, the Dynamic Propagation Graph Convolution Module adaptively learns the correlation strength between nodes under different operational conditions, accurately characterizing the real-time propagation paths of faults. Second, the Trend-Similarity Self-Attention Module analyzes the local evolutionary patterns of time series rather than instantaneous absolute values, effectively identifying early fault precursors with common causes and significantly improving early warning capability. Finally, the Condition-Adaptive Spectrum Enhancement Module employs the Discrete Cosine Transform for feature extraction and enhancement of the quasi-periodic operational signals of the engine, effectively suppressing signal distortion caused by transient processes and improving the robustness of state representation. Experimental results demonstrate that the PADGN model can not only achieve high-precision fault classification and detection but also locate fault sources by interpreting the network's attention weights, providing a novel, interpretable, and systematic intelligent diagnostic solution for liquid rocket engines.

Structural Health Monitoring (SHM) is essential for maintaining the reliability and longevity of civil infrastructure. Within the range of SHM techniques, autoencoder models have shown strong performance in identifying damage by capturing informative representations from high-dimensional data. This work examines the use of Convolutional Autoencoders (CAEs) for SHM, emphasizing their ability to detect structural deterioration when appropriately tuned. A novel detection scheme is introduced that integrates CAE retraining with an exclusion-based decision strategy. The method is assessed through two applications: (i) an aluminum plane frame tested in the laboratory across five incremental damage conditions under impulsive excitation, and (ii) the Z24 Bridge, a full-scale system influenced by environmental effects and controlled settlement scenarios. Key hyperparameters—latent dimension, number of convolutional filters, and minibatch size—are systematically optimized to enhance performance. All evaluations are carried out in the frequency domain, using Mahalanobis distance as the indicator of damage. The results show that the proposed approach is highly generalizable, suitable for diverse monitoring contexts, and stable across both training and validation phases.