PSI - Issue 54

Vasiliki Panagiotopoulou et al. / Procedia Structural Integrity 54 (2024) 482–489 Vasiliki Panagiotopoulou/ Structural Integrity Procedia 00 (2023) 000 – 000

486

5

damaged slant shaft from the full-scale test rig replaces the undamaged one, and dynamic tests assess the acquisition system's sensitivity to detect changes in the structure's dynamic response due to impact damage. During idle phases, damage can be artificially increased to observe the structure's response at different crack propagation rates. The STR comprises the actual slant shaft, joints, and couplings from the BERC test bed. Its primary purpose is to capture vibration signals during the rotation of the slant shaft in both damaged and undamaged conditions. This data supplements information from the BERC and is used for testing and developing SHM systems and algorithms. It's worth noting that there's potential for additional error in signals acquired from the BERC due to the disassembly and reassembly of the slant shaft within the TRDL. In contrast, the STR goes beyond supporting the slant shaft during impact tests; it also collects data during these tests at the shooting range for both damaged and undamaged configurations. This approach results in less biased data acquisition.

Fig. 3 Simplified Test Rig and numerical model on the slant shaft in damage condition.

3.3. The Digital Twin model

The creation of a Digital Twin (DT) model for the TRDL helps us identifying opportunities and risks concerning the structural condition of the system. The FEM is key component of SAMAS 2 DT, as it can effectively replicate the real behavior of the structure. This replication is achieved through continuous self-updating based on data acquired from sensors mounted on the actual helicopter structure. A common challenge arises when attempting to build a high-fidelity FE model using the original blueprints of a structure (e.g., uncertainties, imperfections, limitations). Consequently, before the FE model can be utilized to generate reliable data for training Machine Learning Algorithms (MLAs), it necessitates calibration to accurately capture the damage-sensitive features detected by the monitoring system on the actual structure. FE model updating techniques aim to achieve this calibration by comparing the measured structural responses with those derived from baseline FE models, specifically tailored for the structure, and validated against its undamaged behavior. The structural responses recorded by the monitoring system are used iteratively to adjust certain calibration parameters, such as stiffness properties and boundary conditions, defined at the finite element level. This adjustment process aims to minimize a set of objective functions that quantify the differences between the computed and measured structural responses (C. Sbarufatti, 2016). The probabilistic values of these model parameters that minimize the objective functions provide insights into the presence, location and extent of the structural damage. Despite the significant advantages of high-fidelity models in DTs, their use can be challenging due to the computational demands and time constraints involved. Model reduction techniques are designed to alleviate these challenges by transforming large FE models into smaller matrices. These surrogate models aim to mimic the original model's behavior while retaining essential information about mass, stiffness, and mode shapes, particularly focusing on the most influential low-frequency response modes of the structure. Various surrogate modeling methods have been explored, mainly categorized as intrusive and non-intrusive approaches. Intrusive surrogate modeling requires direct access to the original model to apply matrix reduction techniques, whereas non-intrusive methods involve offline computation and storage of input-output pairs from the original model. This allows for more efficient real-time data-driven modeling. In the specific application discussed here, the surrogate model establishes a link between the system's state (identifying features of impact damage, such as stiffness and mass matrices) and acceleration observations, which serve as indicators of degradation. State-of-the art meta-modeling algorithms like Gaussian Process (GPs), Polynomial Chaos Expansions (PCE), and Neural