Conference Agenda

Digital Twin (DT) architectures that combine physics-based modelling with machine learning are redefining structural health monitoring (SHM) as a predictive, self-adaptive discipline. A persistent challenge lies in representing nonlinear and evolving dynamics of real-world systems while maintaining interpretability and computational tractability. This study introduces STRiDe - a Self-Tuning Reduced Intelligent Digital environment - which integrates modal reduction, adaptive operator learning, and kernel-based health inference within a hybrid framework for nonlinear SHM.

STRiDe is composed of three layers. The first defines a reduced physics-consistent core constructed from the dominant modal subspace of the structural system. This reduction preserves dynamic fidelity while enabling near-real-time simulation. The second module with adaptive identification and regression implements an energy-consistent adaptive operator that incrementally estimates effective stiffness and damping parameters through residual minimisation between measured and predicted responses. This mechanism allows the digital twin to self-adjust to nonlinearity and parameter drift without offline retraining. The third layer employs Kernel Principal Component Analysis (KPCA) to monitor latent-state deviations, producing Hotelling’s T²and Squared Prediction Error (SPE) indices as uncertainty-aware health indicators.

The methodology is evaluated on a four-degree-of-freedom (4-DOF) numerical system incorporating a cubic Duffing nonlinearity at the second DOF. A broadband base excitation (0.1-1.8 Hz) is applied, and acceleration responses are simulated using a fourth-order Runge–Kutta scheme. The recorded data are processed using Singular Spectrum Analysis (SSA) for denoising and feature enhancement prior to twin assimilation.

The results presented in Fig.1 confirm that STRiDe accurately reconstructs nonlinear behaviour while retaining computational efficiency. The nonlinear frequency-response (amplitude–phase) analysis reveals a reduction in response amplitude from 0.015 ms⁻² rms at 0.2 Hz to 0.002 ms⁻² rms at nearly 1.46 Hz, accompanied by a phase shift exceeding 1 rad, confirming stiffness-dependent hardening. The phase portrait and Poincaré map exhibit the emergence of an asymmetric limit cycle, characteristic of weakly nonlinear oscillation. The Short-Time Fourier Transform (STFT) contour map highlights persistent modal energy concentration near 0.3 Hz, corroborating the dominance of the first mode throughout the excitation sweep. During this transition, the adaptive STRiDe operator identifies gradual stiffness drift, evidenced by a 35% increase in SPE and a rise in T^2 from 7.4 to 8.1, both remaining within the 95% statistical confidence bounds. These indicators demonstrate reliable state sensitivity without false alarms.

STRiDe provides a robust, interpretable, and data-efficient foundation for hybrid Digital Twins operating under uncertain, nonlinear, and slowly evolving structural conditions. The framework exemplifies how adaptive operator learning, when embedded within a physics-informed digital environment, can sustain accuracy, interpretability, and resilience – key attributes for next-generation SHM systems in realistic monitoring environments.

Online structural health monitoring requires physical models to be updated in real time in presence of damage initiation and growth. Variants of the Kalman filter, such as the extended Kalman filter (EKF), can handle the nonlinearities associated with joint state–parameter estimation, enabling, for example, the tracking of stiffness degradation in a structure. However, the traditional EKF suffers from uncertainty in the initialization of key hyperparameters—namely, the process noise covariance and the measurement noise covariance. In this work, a plausibility index, proposed for model class selection, is adopted to determine the optimal forgetting factor for an innovation-based adaptive extended Kalman filter (AEKF). The AEKF automatically tunes the aforementioned hyperparameters by leveraging the forgetting factor, while the model class selection determines the posterior probability of a model, characterized by a given process noise covariance, by measuring its relative plausibility within a prescribed set of models. The plausibility index, termed conditional evidence, is computed using Laplace’s asymptotic expansion of a Bayesian probability integral. Under this approach, the best model is the one with the highest plausibility among all candidates; in this context, it corresponds to the model whose forgetting factor performs best within the user-specified set, i.e. returning also the best estimation results in the same set. The procedure is validated against the IASC–ASCE benchmark.

Moreover, although an AEKF with model class selection significantly improves the robustness and automation of SHM, it also increases the computational burden due to the multi-model evaluations required for selecting the optimal hyperparameter. As the dimension and complexity of the structure increases, the computational cost could become a bottleneck for real-time SHM. Considering the inherent parallelizability of the model class-selection method, we propose a solution to leverage the high computational power of GPUs to get an effective solution.

This paper presents the development and validation of a hybrid Structural Health Monitoring (SHM) strategy for wind turbines, integrating physics-based modeling with data-driven approaches through a Digital Twin (DT) framework. The methodology is demonstrated on a small-scale horizontal axis wind turbine, where a high-fidelity DT is constructed by combining a Finite Element (FE) model for modal characterization and a flexible Multi-Body (MB) model for system-level dynamic simulation. Experimental validation is performed through impact-based Experimental Modal Analysis (EMA) and operational measurements using accelerometers, strain gauges, and Fiber Bragg Gratings (FBGs). The FE model is iteratively calibrated against measured modal parameters, while the MB model outputs are validated against operational sensor data. This synergistic integration of validated mechanistic models with high-resolution experimental data establishes a robust foundation for real-time monitoring and predictive analytics. The proposed framework demonstrates the potential for extending DT capabilities toward advanced prognostics and recursive state estimation, supporting enhanced reliability and sustainability of wind turbine operations throughout their service life.

This abstract presents the outcomes of the AIRPASS project, a collaborative effort between Royal Netherlands Aerospace Centre (NLR) (an applied research organisation in the aerospace domain) and Bundeswehr Research Institute for Materials, Fuels and Lubricants (WIWeB) (an independent Bundeswehr research center which is responsible for the safety, technology and chemistry of materials and petroleum, oils and lubricants), which aimed to develop a novel inspection concept for aircraft structures by leveraging automation and digitalization. The project's core idea revolves around the central role of the digital twin, fed by non-destructive testing (NDT) data. Within AIRPASS, we successfully demonstrated an automated approach to link inspection data to the digital twin. Notably, NLR and WIWeB employed distinct methods to achieve this goal: NLR utilized visual properties of the inspected object to register the position of its contactless NDT sensor with respect to the digital twin, whereas WIWeB employed positioning devices that recognized reference points in 3D space, suitable for probe-like sensors. The method employed by NLR combines the simplicity of a Structure from Motion (SfM) system for camera positioning and orientation, with the geometric accuracy of high fidelity 3D scanning or CAD geometry for the projection of NDT measurements. The WIWeB’s system makes use of an optical tracking system, which enables precise locating of the sample, probe and Mixed Reality glasses. It features real-time feedback for the user by drawing the NDT results as a color-coded map into the texture of the 3D model and displaying it in the users field of view. A comparative analysis of the results from both approaches was also conducted, providing valuable insights into the effectiveness of each method. The AIRPASS project's findings pave the way for enhanced automation and digitalization in aircraft inspection processes.

Elastic guided waves are frequently used in Structural Health Monitoring (SHM) to detect, localize, and characterize damages in structures using in-situ transducers. Over the past decades, a number of data-driven techniques have been developed and validated in realistic applications. However, model-based techniques remain scarce, primarily because of the prohibitive computation times required to finely resolve such numerical problems. In this work, we present a structure-preserving model order reduction technique with a time-dependent basis that preserves the symplectic structure of the full-order equations. The basis is computed off-line from full-order snapshots by solving a Riemannian optimization problem. Numerical experiments are conducted on a Lamb wave propagation problem. The results demonstrate the potential of the proposed method to compress wave propagation data and to speed up computations in a multi-query parametric setting.