Development of a hybrid damage identification system in a digital twin perspective for naval applications

In this thesis, the identification of structural damage in metallic 1D and 2D structures is addressed through the application of advanced Machine Learning (ML) techniques. Specifically, the research focuses on detecting and localizing stiffness reductions that signify localized damage. This problem is part of a more comprehensive research that aims at avoiding the risk of failure by implementing Structural Health Monitoring systems and Condition-based Maintenance strategies. This thesis focuses on developing methods for ship engineering applications, a field in which research is underway despite clear benefits. Moreover, given the growing interest in the scientific community towards Digital Twins (DT), a virtual tool that contains all the information of the asset it represents and that is continuously updated through the sensor network on the object, the developed method is directly integrated in a DT perspective. The present approach points to harnessing the power of Artificial Intelligence (AI) and ML as this technology has recently been proven to bring disruptive advancements in several fields. Thus, it is expected that this technology will help develop more precise and reliable tools for structural damage identification in the near future. Initially, the key aspects of ML-based damage identification were reviewed to highlight the strengths and challenges of these methods. A hybrid damage identification methodology that integrates concepts from both Physics-Based and Data-Driven approaches is then proposed to overcome their limitations. As a result, lessons from traditional modal curvature-based damage identification inform the feature engineering process and enhance the performance of ML algorithms. Moreover, this study employs a Transfer Learning framework to address the scarcity of real-world data by training the algorithms on a simulated population of damage cases. To ensure a reliable comparison between real and simulated data, Model Updating and noise characterization techniques were implemented by processing preliminary data obtained from experiments. The updated model constitutes the initial Digital Twin of the structure and is further updated with the damage identification results. In addition, a noise-reduction technique was introduced to mitigate the effects of measurement noise. The identification strategy is developed in two steps, exploiting the most suitable algorithm for each level of damage information. First, damage detection is achieved using a Histogram Score Novelty Detection (HSND) algorithm, and second damage localization is performed via a Regression Neural Network (RNN). Performing damage detection using a Novelty Detection approach results in training the algorithm only on intact data, thereby ensuring that any damage can be predicted. Moreover, the separation of detection and localization allows the use of a regression approach for localization, treating the entire structure as a continuous space, and therefore obtaining a tool capable of locating damage anywhere on the structure. The proposed framework was tested for two structural cases: a slender beam and a thin plate. For the slender beam, damage identification was validated using only simulated data from an analytical model. Nonetheless, information regarding the measurement noise was obtained from a previous experimental campaign. This first application is crucial for assessing the algorithms capabilities in identifying the damage. The optimal number of training samples was investigated to determine the best compromise between computational cost and prediction accuracy. The sensitivity of the method to both damage severity and noise levels was analyzed using the noise reduction technique. A comparison with the “embedded” physics-based index, in its stand-alone version, was also performed. The second application involved damage identification in a rectangular metallic plate, where a dedicated experimental campaign was conducted. To ensure the relevance of the training data, a structural optimization process was employed to align the finite element (FE) model with the real structure, which is a key point in the Transfer Learning framework. Experimentally identified noise, representative of real-world conditions, was then incorporated into the FE simulations prior to algorithm training. The methodology was first validated on additional FE-based test cases, followed by a comparison with the experimental data. Sensitivity analyses were also conducted to explore the impact of various factors, such as training sample size, damage severity and noise levels on the robustness of the methodology. This thesis concludes with a critical evaluation of the obtained results, followed by a discussion of future research directions to further enhance the proposed damage identification methodology.