The 1998 Civaux 1 nuclear plant incident exemplifies high-cycle thermal fatigue (HCTF) failure, caused by cracking in a pipe elbow downstream of a fluid mixing tee-junction. Similar cases in Penly 2 and Cattenom 3 (2023) highlight the persistent risk HCTF poses to nuclear safety. Pipes exposed to thermal transients over 50 °C at 1–10 Hz are especially vulnerable, with austenitic stainless steel’s low thermal conductivity further complicating detection.

This study applies inverse thermal modelling (ITM) to monitor internal temperature changes in SS304 under rapid (>250 °C, <1 min) transients on a purpose-built thermal fatigue rig. By using ultrasound wave velocity variations, ITM enables full-thickness thermal profiling. Ultrasonic C-scan data reveal shear wave velocity and attenuation changes as the number of thermal cycles that the specimen are exposed to increases. A full set of experimental data on several specimen will be reported on. The ultrasonic data shows clear changes before cracking on the outside surface is observed.

Monitoring bridges under variable traffic loads is essential for safety and predictive maintenance, since cyclic stresses in metallic structures can initiate and propagate cracks, reducing durability. Fatigue is largely driven by stress-amplitude ranges, typically obtained by cycle counting from strain-gauge measurements near structural joints. While strain gauges provide accurate continuous data, they require wiring and power and may suffer from disconnections or measurement bias, which limits their suitability for long-term monitoring.

To address these constraints, SilMach developed a passive mechanical sensor (ChronoMEMS) that requires no power supply and detects strain amplitudes at the installation location. Unlike active sensors, it outputs aggregated “binned” data: counts of cycles whose amplitudes fall within predefined, often wide, intervals, without recording the full time history. With continuous strain-gauge signals, damage can be computed directly after post-processing (e.g., rainflow counting), whereas binned data require an indirect, probabilistic approach.

Stress amplitudes in civil structures are commonly modeled using parametric distributions (e.g., Weibull, log-normal, or mixtures). Once an amplitude distribution is specified, fatigue damage can be expressed through an integral over the density, but it depends on unknown parameters that must be estimated. This study proposes estimating these parameters using an Expectation–Maximization (EM) algorithm adapted to binned data, relying only on interval counts from the passive sensor.

A key issue is parameter identifiability, which depends on how much information the binned counts provide relative to the model’s number of parameters. For a two-parameter Weibull model, the three-bin counts from a single ChronoMEMS sensor are sufficient to estimate both parameters. In contrast, for a mixture of two Weibull distributions (five parameters), three counts are generally insufficient, leading to an identifiability problem and motivating either simpler models or additional information.

To address this limitation, we assessed damage estimation by approximating a two-Weibull mixture with a single Weibull distribution. The resulting damage estimates are satisfactory on both simulated and real data, except for one simulated scenario in which the two modes of the mixture are widely separated—a configuration that appears unlikely given the available real data but which can not be completely excluded. Handling such cases requires combining two sensors to obtain counts over six intervals. This richer binning enables the estimation of models that are more flexible than a single Weibull distribution while remaining more parsimonious than a full two-Weibull mixture (e.g., an exponential–Weibull mixture, or a Rayleigh–Weibull mixture). Candidate models can then be compared using information criteria such as Akaike information criterion (AIC) and bayesian information criterion (BIC) to select the best trade-off between goodness of fit and parsimony, thereby reducing the risk of bias due to model misspecification. Finally, after estimating damage, we construct a confidence interval using a bootstrap procedure.
This study is part of the SIRCAPASS project (monitoring of road infrastructures using passive sensors), which received financial support from Bpifrance (research call i-Démo, part of the France Relance plan).

Full-scale fatigue testing is a standard and essential procedure in the development of new air vehicles. In this process, a full-scale aircraft is used as a test article and subjected to fatigue loads representing the loads expected during its life. Dozens of hydraulic loading jacks are utilized to simulate ground and flight loads, with cabin pressure added as required.

The primary objective of a full-scale fatigue test is to verify that the airframe maintains structural integrity throughout its intended design life. Beyond this, the test provides critical validation of crack initiation and growth predictions made during the design phase of aircraft development. It also serves to validate non-destructive inspection (NDI) techniques that will later be applied during maintenance checks in service.

Such full-scale tests are time-consuming, typically last several years and are often completed after the first aircraft are delivered to the customers. When fatigue cracking is detected early in the test, design modifications can be introduced into serial production at relatively low cost. However, retrofitting aircraft already delivered to the customers is highly expensive, as it often requires complicated maintenance operation and substantial downtime. Such repairs, including aircraft downtime, can cost tens of thousands of dollars per airplane, and when multiplied across an entire fleet, the expense can reach millions.

The economic incentive for early crack detection during full-scale fatigue tests is therefore clear. Traditionally, two methods are employed to detect cracks during full-scale fatigue test:

1. Scheduled non-destructive inspections at predefined airframe locations.
2. Strain gauge monitoring, where variations in strain distribution are used as indirect indicators of crack initiation.

While these methods are valuable, they are limited in their ability to efficiently process the vast amounts of data generated during long fatigue campaigns. This is where machine learning, and particularly Principal Component Analysis (PCA), becomes effective.

PCA is capable of reducing high-dimensional strain gauge datasets into a smaller set of uncorrelated principal components, allowing efficient data compression while preserving the dominant structural response patterns. Subtle deviations from these baseline patterns, such as those caused by load redistribution due to crack initiation and propagation, can then be identified as anomalies. Unlike simple threshold-based monitoring, PCA captures correlations across multiple sensors simultaneously, providing a more sensitive and robust early-warning mechanism.

The objective of this paper is therefore to propose a PCA-driven approach for anomaly detection in full-scale fatigue testing. It will be demonstrated that PCA can efficiently handle large datasets, highlight abnormal structural behavior, and enable earlier and more reliable detection of cracks (see Figure 1).

By integrating PCA into fatigue testing workflows, the aviation industry can reduce the risk of late design modifications, minimize costly retrofits, and ultimately save hundreds of thousands of dollars per test campaign.

The monitoring of progressive fatigue damage in composite structures is critical for ensuring structural integrity and reliability, particularly in the absence of precise knowledge regarding material properties and stochastic loading conditions and damage evolution mechanisms. In this study a novel, fully data-driven Functional Multiple Model (F-MM) framework is introduced for the detection, characterization, and estimation of fatigue damage severity in composite structures based on random vibration response signals.

The foundation of the F-MM framework lies in the representation of random vibration signals using Functionally Pooled Multiple Models, with the Functionally Pooling component enabling the explicit modeling of the structural dynamics across varying fatigue loading cycles and the Multiple Model component capturing uncertainties related to the population of structures, material variability, boundary conditions, and stochastic excitation. By jointly accounting for variability across both the population and the fatigue progression, the framework offers a comprehensive characterization of the evolving global dynamics behavior associated with progressive damage.

The experimental validation of the framework is based on a set of composite thermoplastic coupons subject to controlled Tension–Tension fatigue under interrupted loading conditions. At each interruption, the coupons are subjected to random vibration testing with broadband acoustic excitation accompanied by non-destructive verification C-scans. The vibration response signals are then provided to the F-MM framework to identify and track the evolution of the structural dynamics throughout their fatigue life, as well as for monitoring their current condition in terms of damage detection, discrete Fatigue Stage classification, and precise damage severity estimation.

The results of the study demonstrate that the postulated framework achieves: (a) A monotonic mapping of the changes in the structural dynamics with fatigue damage progression, establishing a robust basis for diagnostic inference; (b) Excellent damage detection even for the minor damage associated with fatigue cycles; and (c) very good classification of the current damage state to one of the selected discrete FSs; and, for the first time in the literature, (d) precise estimation of fatigue damage severity throughout the coupons’ lifespans. The framework also exhibits strong generalization capabilities across coupons, underscoring its potential for population-based monitoring.

Overall, the study highlights the potential of the F-MM framework as a unified, purely data-driven, monitoring tool for a population of composite structures operating under uncertainty. Its ability to integrate stochastic dynamic modeling with population-level variability warrants effective monitoring with the potential of real-time and in-operation application without reliance on deterministic models, material properties, or prior fatigue history. The findings pave the way for the scalable deployment of the framework for structural health monitoring applications in composite components across various sectors.