Automation of the Identification of Vibration Modes of Car Bodies

Abstract

[Abstract]: In this thesis, two methods of automating the identification of the first torsional modes are studied. First, the Modal Assurance Criterion (MAC) as a classical approach for mode identification, and second, a Random Forest Classifier as a data-driven alternative, which was trained and tested with the results from the MAC. These two methods aim at reducing the time consuming work of manual labeling the first torsional mode in car bodies, specifically in trim body structures. To evaluate the identification with the MAC, a mechanical eigenvalue problem is solved in ANSA/Epylisis and the resulting mode shapes are compared against a torsional reference to check similarity across the different modes. This comparison requires approximately 0.06 seconds, and the MAC value obtained for the identified first torsional mode is 0.85. The Random Forest Classifier is trained with a dataset of eigenvectors obtained from the modal solution of 90 mechanical problems, whose labels were validated using the MAC. The training is improved using cross-validation techniques: Stratified k-fold and Synthetic Minority Oversampling Technique (SMOTE) and the algorithm is assessed with standard classifier metrics: accuracy, precision, recall, and F1-Score. It is then tested through several typical Machine Learning procedures: hypersensitivity analysis, generalization test, etc. Results indicate a training + test time of 31.92 seconds, identification metrics of approximately 87% to 100%, depending on the test performed, and a minimum prediction time of around 0.0076 - 0.055 seconds per experiment.