The thesis presents the use of machine learning methods for the classification of structural damage. Two methods for development of a digital twin are presented, both of which combine dynamic methods of treating the structure with machine learning methods. First, we present a combination of analytical model using modal analysis, where the generation of learning data takes place on a simple example. In the second part, we consider the case of a complex structure where numerical models are used. Here we applied the methods of dynamic substructuring, which allows us to assemble individual substructures into a whole structure. The generation of training data is performed using a variation of stiffness at the damage site. We perform supervised learning of machine learning methods on both learning databases. This is followed by the re-generation of a new test database to verify correct classification. Probabilities of correct classification are compared between the applied machine learning methods. The best performing machine learning methods are evaluated between each other. Finally, we comment on the applicability of the presented procedures and the use of machine learning for the purposes of machine damage classification.
|
