The doctoral thesis focuses on the problem of generating training datasets for identifying structural damage using machine learning. It turns out that, experimentally, it is often difficult to capture a dataset that encompasses a wide range of different damage states. On the other hand, numerical models describe the real structural state with limited accuracy, and high computational cost also poses a significant constraint. Within the scope of the doctoral work, a methodology based on dynamic substructuring is proposed. This approach enables the creation of hybrid models that reflect the real state of the structure while allowing the simulation of various damage scenarios. Dynamic substructuring facilitates the updating of numerical models with experimental data, as well as the coupling and decoupling of different types of models (experimental, numerical, and analytical), enabling the formation of representative hybrid training datasets. In the first part of the research, a hybrid training dataset was created for identifying damage in the joints of a laboratory structure, which often represent critical points of the system. The second part focuses on monitoring the condition of more complex joints, where dynamic substructuring is used to generate a dataset transferable across different assemblies. Experimental testing of the proposed approaches resulted in the successful creation of training datasets, which were then used to train an algorithm for localizing and identifying damage in the laboratory structure, as well as an algorithm capable of evaluating the quality of a riveted joint based on its dynamic response.
|
