PSI - Issue 64

Hendrik Holzmann et al. / Procedia Structural Integrity 64 (2024) 1303–1310 Hendrik Holzmann / Structural Integrity Procedia 00 (2019) 000 – 000

1305

3

In the past decades several machine learning methods have found their way into application. In the context of SHM, machine learning has proven to be useful to distinguish between fault-free and faulty structures or components, as described by Zaparoli Cunha et al. (2023). In general, the detection and localization can be done with an unsupervised learning approach. However the identification of the type of damage and its severity can only be achieved in a supervised learning mode (Worden et al. 2007). Although unsupervised learning for detection and localization is possible, a supervised learning can be more useful, when the classification task is clear and since it requires less data than unsupervised learning. For supervised learning, a labeled database for training and testing is necessary. For many applications – as the one discussed in this work – this is an expensive task when the database consists of measurements. To overcome this problem, several works have investigated the use of synthetic measurement data derived from parametric physical models (Seventekidis et al. 2020). For the SHM of a system, features must be derived from the data to distinguish between different system states. The best selection of features for damage identification is application-specific (Chen et al. 2020). This paper presents an application-near simulation workflow for the detection and localization of void defects in sandwich panels with diameters of a few centimeters. For this, the influence of failures on the response to a dynamic excitation is exploited. The workflow uses an experimentally validated finite element model, feature engineering, and a neural network. It initiates with the development and validation (sensitivity-based model updating based on modal parameters) of a mechanical finite element model using experimental data. The model serves as the foundation for a simulation campaign, examining both fault-free and models with defects. The simulation process incorporates variations in material parameters based on experimentally obtained standard deviations, ensuring a robust representation of real-world uncertainties. The resultant simulation dataset is subjected to a feature engineering process, encompassing time and frequency domain features. This multi-domain feature extraction enhances the dataset's informativeness. The feature dataset is employed to train a neural network for two primary use cases: Firstly, to distinguish between faulty and fault-free models, thereby providing an accurate fault detection mechanism. The second use case involves the localization of faults within the model. The effectiveness of this approach is demonstrated using various testing datasets.

Nomenclature PUR

Polyurethane

PIR Polyisocyanurate SHM Structural health monitoring PCA Principal component analysis

2. Pattern recognition framework This section presents the framework for detection and localization of defects in sandwich panels. The used model, the simulation campaign, feature engineering, and machine learning approach are described in detail. The framework is visualized in Fig. 2.