Lamb waves have been long time used for damage detection due to their high sensitivity to interference in wave propagation and high damage coverage with a small number of sensor. However, due to the different wave speeds of multiple Symmetric and Antisymmetric modes, wave dispersion, and numerous reflections and losses introduced by the geometry, suh boundaries, reinforcements, and thickness changes. For the previousmentioned reasons, Lamb wave Structural Health Monitoring (SHM) techniques application is limited due to the difficult data processing required for an accurate damage location and characterization.

Even if it is possible to find a variety of methodologies and signal processing techniques to deal with damage detection, techniques based on Deep Learning (DL) has taken special relevance. Specifically, an Artificial Neural Network (ANN) model based on a Multi-Layer Perceptron (MLP) it is proposed to determine both the location and size of structural defects. In this case, the ANN input consists of specific characteristics extracted from the differential signal (damaged state vs undamaged state), such as the number of wave packets, maximum packet value, and amplitude, for a set of different locations, delamination areas and severities. This methodology proves to be highly effective, achieving a small localization error smaller than the size of the damage itself.

Even if the high potential of this DL technique, its benefits require a complex training, as large datasets are required, which are unfeasible to obtain solely from experimental setups, so an extensive application of simulations are required. The use of traditional Finite Element Methods (FEM) are not feasible, as they demand significant computational resources and time; other specific techniques such Spectral Element Mehod (SEM) of Physic Based ANN presents high difficulties to implement out-of-plane reinforcements or geometric changes.

This work applies the validated and computational effective Lamb wave simulation technique based on the Ray Tracing (RT) to simulate the damage response in a huge range of cases required to apply MLP ANN for SHM. RT technique enables precise modeling of wave reflection, refraction, damping, and mode conversion at structural discontinuities, and reduce the each simulation case to few minutes instead of several hours. It has been applied to a representative aerospace structures, such as UAV wing lower cover made of composite material. This structure presents change of thickness and a cobonded stiffeners. It is proposed to validate the technique with real Barely Visible Impact Damage (BVID) and high energy impacts in different locations to demonstrate the potential of the technique as well the RT training simulation method.

Health monitoring of bolted joints is crucial in guaranteeing structural integrity and safe operation in many engineering applications by enabling quick and accurate detection of bolt loosening, which is among the most common failure modes. Guided wave-based Structural Health Monitoring (GW-based SHM) systems are generally preferred for thin, plate-like structures given that they can identify minute structural issues and survey extensive areas effectively with a minimal number of transducers.

In this work, experiments are carried out using piezoelectric transducers mounted on an aluminium lap joint with two bolts to evaluate the Lamb wave propagation characteristics across the joint under various stages of loosening. Loosening is simulated by varying torque values on the bolts from 10 Nm down to 3 Nm, with a step size of 0.5 Nm, using a torque wrench. Three additional torque conditions, namely, a hand-tightened state, a state where the bolt is present but zero torque is applied, and a state where the bolt is completely absent, are also evaluated. Furthermore, seventeen different excitation frequencies spanning a wide range from 52 kHz to 260 kHz are employed for generalization. Investigation of the signals collected from the experiments reveals a clear reduction in the signal amplitude as the bolts are progressively loosened. This result is consistent with the reduction in the amount of wave energy transmitted across the joint under loosening due to the reduction in the contact area between the plates in the joint overlap region.

The central objective of this study is the development and evaluation of a novel unsupervised deep learning framework tailored for loosening detection, which centers on a comparative study of two hybrid autoencoder architectures: the CNN Encoder-GRU Decoder (C-G AE) and the GRU Encoder-CNN Decoder (G-C AE). Convolutional Neural Networks (CNNs) perform hierarchical feature extraction by capturing local spatial patterns and short-term dependencies. On the other hand, Gated Recurrent Units (GRUs) model the long-term dependencies and temporal dynamics. Combining both leverages the capability to extract localized features and their temporal evolution. This study enables a comparative evaluation of how effective the spatial-first versus temporal-first encoding strategies are in terms of latent representation of the time-series data, reconstruction fidelity, and the anomaly detection performance. Additionally, we will systematically investigate the influence of the ordering of the CNN and GRU components, as well as the impact of varying levels of noise and the volume of training data, on the overall anomaly detection performance.

The models will be trained exclusively on minimally processed time-series signals representing the torqued conditions by minimizing the reconstruction error. The unsupervised learning paradigm helps in avoiding the enormous task of labeling all the possible damage scenarios. The performance of the trained models will then be assessed using unseen signals representing loosened conditions. By rigorously comparing the models' capacity to correctly detect anomalous signals, this work aims to identify the most efficient and reliable hybrid deep learning architecture for the monitoring of joint integrity.

Ultrasonic guided waves hold significant potential for non-intrusive monitoring of damage in composite structures, contingent on the efficacy of the onboard monitoring system to reliably acquire, process signals. By mapping the extracted signal features with parameterized damage metrics, it is possible to realize an automated framework for the assessment of structural integrity. While the fundamental S₀ and A₀ modes are known to be sensitive to damage and serve as digital damage identifiers when properly characterized, the specific ways in which damage influences the waveguide dispersion properties are not fully understood. To address this gap, it is vital to establish a generic, extendable and reproducible wave mode reconstruction methodology so that the fundamental ultrasonic guided wave modes can be isolated and subsequently investigated for damage signatures.

To introduce damage, a thin-walled symmetric carbon fiber reinforced composite panel was subjected to cyclic displacement-controlled compressive fatigue loading. Throughout the test, structural responses to monotonic toneburst sinusoidal excitations, applied over a user-defined frequency band, were acquired after every 25,000 cycles of fatigue. A physics-informed harmonic wave propagation model generated individual S₀ and A₀ mode realizations that were subsequently superimposed to produce accurate reconstructions of the acquired experimental signals. A regularized residual error function was formulated to account for discrepancies from measurement noise, unmodeled higher-order modes, and other sources of error. A probabilistic Bayesian joint parameter estimation approach was employed to minimize this error, calibrate the wave mode characteristics and quantify the uncertainties associated with the physics-informed, data-driven estimates.

The proposed methodology effectively captured progressive degradation in the calibrated parameters and quantified the uncertainties in estimation, revealing distinct directional and modal sensitivities to fatigue damage. This achievement underscores the efficacy and reliability of the calibrated ultrasonic guided wave modes as reliable damage identifiers, thereby taking a crucial step towards realizing a robust physics informed and data-driven structural health monitoring framework for safety critical engineering assets.

The continuous welded rail (CWR) has been widely adopted in modern railways for its ability to support high transport speeds and reduce maintenance compared with jointed tracks. However, CWR is susceptible to internal stress due to restrained of its thermal expansion and contraction along the rail’s axial direction. During hot conditions, rail expansion is constrained, generating significant axial forces that can lead to thermal buckling depending on track conditions. The magnitude and direction of these built-up stresses are governed by the rail temperature relative to the rail neutral temperature (RNT). Accurately estimating RNT, and thus the associated thermal forces, without baseline measurements remains a long-standing challenge in railway engineering. In this study, we present recent advancements in rail thermal force estimation using intrinsic local resonances in rails. The results demonstrate that the frequencies of these local resonances, specifically, the zero-group-velocity modes and cutoff-frequency resonances whose energy is confined near the excitation zone, are insensitive to the presence of rail supports yet highly sensitive to variations in rail temperature and axial load. This unique characteristic makes them highly promising for accurate and non-intrusive RNT estimation. An automatic impulse-dynamic testing system was developed and mounted on a mobile platform to extract rail local resonances from a fully instrumented site at the FAST loop in MxV Rail, where strain gauges and thermocouples provide ground-truth RNT and rail longitudinal force data. Two complementary experiments were conducted: a stationary test that continuously collected local resonance data at a fixed location with known RNT and rail forces, and an in-motion test using the mobile sensing platform with piezoelectric excitation and non-contact microphones to identify resonance modes along the rail. Probabilistic machine learning algorithms were developed to predict rail thermal forces at the instrumented site using two directly measurable inputs: rail temperature and the frequencies of rail local resonances. The performance of the proposed framework was evaluated by comparing the predicted thermal forces with the ground-truth data, demonstrating its capability for reliable and data-driven RNT estimation.