This paper presents an integrated Structural Health Monitoring (SHM) strategy for flax fiber reinforced thermoplastic composites, combining Nonlinear Resonance Acoustic Spectroscopy (NLRAS), Acoustic Emission (AE), and data-driven damage identification based on machine learning. The proposed approach aims to provide both early damage detection and physical interpretation of damage mechanisms by coupling global nonlinear acoustic indicators with local acoustic emission activity, in line with recent SHM developments for composite materials (Bentahar and El Guerjouma, 2008; Allagui et al., 2023).

The experimental investigation was conducted under progressive three-point bending loading, with the material characterized at successive damage levels. NLRAS measurements were performed using integrated piezoelectric transducers to excite and monitor a resonance mode of the specimen under controlled fast and slow dynamics. The evolution of nonlinear acoustic parameters, including resonance frequency shift and hysteretic behavior, was analyzed as a function of excitation amplitude. Such nonlinear resonance-based indicators are known to be highly sensitive to damage-induced contact and hysteresis effects in heterogeneous materials (Bentahar and El Guerjouma, 2008; Mechri, et al 2019, Dolbachian et al., 2023). The results demonstrate a strong sensitivity of nonlinear resonance parameters to damage accumulation, with measurable stiffness degradation detected at very early stages, even when conventional acoustic emission activity remains limited.

In parallel, acoustic emission signals were continuously recorded during mechanical loading in order to capture local damage events. A large dataset of AE waveforms was processed using advanced signal processing techniques to extract temporal, spectral, and time–frequency features. An unsupervised machine learning framework was then applied, including feature selection, anomaly detection, and clustering. Based on multiple cluster validity criteria, Spectral Clustering was identified as the most suitable algorithm for the AE dataset, following recent advances in AE-based damage classification using machine learning approaches (Almeida, 2023). This approach enabled the identification of four distinct families of AE signals, which were physically interpreted and associated with the main damage mechanisms occurring in flax fiber composites: matrix cracking, fiber/matrix debonding, fiber pull-out, and fiber breakage. The temporal evolution of the identified AE clusters revealed a clear and sequential activation of damage mechanisms, consistent with previous observations on progressive damage in polymer-based composites (Bentahar and El Guerjouma, 2008; Allagui et al., 2023).

The Felicity Ratio was further employed to quantify the loss of elastic memory and to correlate AE activity with damage severity. A significant decrease in the Felicity Ratio was observed for precursor mechanisms, indicating early reactivation of micro-damage, whereas critical mechanisms remained close to the Kaiser effect until advanced damage stages, as commonly reported in AE-based SHM studies.

Overall, this work demonstrates that the combined use of nonlinear acoustics, acoustic emission, and machine learning constitutes a robust and highly sensitive SHM framework for composite structures. By linking global nonlinear indicators to local, physically interpretable damage mechanisms, the proposed approach offers significant potential for early damage detection, damage mechanism identification, and long-term monitoring of bio-based composite materials.

Ensuring the visual appearance quality of automotive interior buttons requires reliable inspection of painted surfaces. However, powder-coated finishes present complex reflectance behaviors, including directional glare and irregular highlight patterns, which often mask or resemble subtle defect features. To address this limitation, we explore a BRDF-based Photometric Stereo (PS) approach that accounts for complex reflectance characteristics when capturing fine geometric features of painted button surfaces. Leveraging BRDF-based modeling and multi-directional illumination, the PS method derives pixel-wise surface normals that are more robust to specular highlights and better represent true surface geometry. These surface normal maps are subsequently utilized as input to a deep learning model for detecting defective and anomalous surface regions. When integrated with AI-based anomaly detection, PS-derived geometric representations contributed to more reliable identification of fine surface irregularities, effectively enhancing the detectability of micro-defects that are visually indistinct in painted automotive components.

Environmental and operational variabilities (EOVs) are one of the primary factors limiting the reliability of data-driven Machine-Learning (ML) based Structural Health Monitoring (SHM), as changes in temperature or operating conditions often induce response fluctuations comparable to those caused by moderate structural damage. This issue is particularly evident in long-term monitoring of existing structures, where environmental cycles interact with fast- and slow-varying damage processes and may mask or mimic them.

This contribution investigates how explicitly conditioning latent feature extraction on measured environmental parameters can improve damage detectability under varying environmental conditions. The study proposes the integration of temperature information into an Autoencoder (AE) architecture through conditioning strategies that combine structural response data with an auxiliary environmental channel. Different conditioning strategies are explored, including concatenating temperature to either the input layer or the latent code, applying temperature-based scaling, and modulating intermediate activations to help disentangle environmental effects from damage-related changes.

The methodology was evaluated using a benchmark dataset, developed to simulate realistic SHM data under controlled conditions, explicitly incorporating temperature-driven stiffness variations, and fast-varying damage mechanisms. The benchmark provided full control over the ground truth reference and allowed systematic testing under the modelled combined influences. The analysis focused on fast damage scenarios, where latent features showed damage sensitivity in previous studies, and investigated whether temperature-conditioned latent spaces would yield improved anomaly detection performance classifiers. The study also examined the interpretability of conditioned latent spaces, their dependence on normalisation strategies, and the implications for transferring the approach to real monitoring data, where temperature is routinely measured and naturally aligned with vibration records. The results are expected to contribute to ongoing discussions on modelling evolving system dynamics, using both physical and data-driven information, and improving the reliability of long-term data-driven SHM under environmental variability.

Structural health monitoring (SHM) of small structures is necessary to ensure their long-term sustainability. However, the large number and scattered distribution of such structures raise a significant techno-economic challenge, which calls for optimization of the monitoring solutions to be deployed.

Monitoring these structures requires complementary data to better understand their behavior. Such information is typically obtained through the deployment of additional sensors measuring temperature, solar radiation, humidity, and wind speed. Nevertheless, this approach is difficult to balance for small or isolated structures, whose maintenance programs already face tight constraints to ensure durability. Optimizing the number of sensors to be installed is therefore crucial.

In this context, artificial neural networks can model complex relationships between time series through learning processes and thus represent an alternative to the deployment of numerous physical sensors. Instead of multiplying measurement points on a structure, neural networks can model various phenomena affecting it (gradients, internal temperature, surface heating, drying, wind effects) from a smaller set of input data. In this context, this study focuses on the use of publicly available datasets in France

This paper presents investigates the use of Temporal Convolutional Networks (TCN) for sensor data regression based on external and public data, the challenges faced with such approach and discusses their potential as a “ready-to-use” analysis tool. A case study using multiple type of sensors on concrete structures is presented.

Acoustic emission (AE) monitoring is a method of structural health
monitoring that relies on the detection of elastic waves generated by the release
of concentrated strain energy when damage is created in a structural material.
One of its strengths is that registered waveforms can, in theory, be used to
draw conclusions about the origin of the signals, allowing to estimate damage
type, location, and severity. One of the problems, however, is how to automate
the interpretation and reliably retain useful information from AE waveforms,
which are often thousands of samples long. In this work, a method of waveform
analysis in the frequency domain is presented that combines principal compo-
nent analysis and an autoencoder to reduce the dimensionality of the problem
to a pair of parameters, capturing the spectrum shape and the spectrum en-
ergy content. Tensile tests were carried out on composite coupon specimens
while monitoring AEs, and the progression of damage was tracked by X-ray
scanning them in two locations before and after tensile testing. Locations and
types of damage in the scans are in good agreement with the results of the
AE monitoring and analysis framework. The method is proposed as a tool for
automated interpretation of AE signals with the potential to be generalized to
other material, layup, and sensor setups.

The effective deployment of Artificial Intelligence (AI) in Structural Health Monitoring (SHM) is consistently hindered by the gap between idealized mathematical models and the field data, which is inherently heterogeneous, and asynchronous. While the industry has achieved high Technology Readiness Levels for isolated sensor hardware, it continues to suffer from low Integration Readiness Levels, forcing engineers into repetitive, manual data-wrangling tasks that preclude high-level diagnostic reasoning.

This paper introduces a Reference Framework that proposes an architectural logic designed to navigate these integration challenges through Large Language Models (LLMs). We present a design hypothesis centered on a hybrid, multi-layered workflow that utilizes LLMs cognitive. As a result, this framework enables a cognitive layer to semantically interpret user intent and delegate complex, multi-step tasks to a library of deterministic, verified algorithms.

To mitigate persistent integration risks and realize the potential of autonomous diagnostics, we propose that future SHM system designs should prioritize Pre-Processed Data Ingestion. By positioning the LLM as a bridge, this framework establishes a foundation for transitioning from manual data wrangling to conversational, actionable insights