A deep learning approach to classifying flow-induced vibration response regimes of an elliptical cylinder

Abstract

This study proposes a new approach that leverages deep learning to the study of flow-induced vibration (FIV), specifically to automate flow regime classification and to visualize the transitions between these regimes. Using previously obtained data on the amplitude response of an elastically mounted cylinder as a function of reduced velocity for a range of structural damping ratios, the time trace of the body displacement and fluid driving forces are first converted into a frequency-time representation using continuous wavelet transforms before being input to several pre-trained convolutional neural networks for feature extraction. When utilizing the outputs of each convolutional neural network for regime classification, we found that almost all the machine learning approaches had high cross-validation accuracy that was statistically insignificant from each other. The five best-performing classifiers were then used as an ensemble method, yielding a weighted accuracy of 99.1% on the test data. The FIV response regimes were further investigated by projecting the outputs of the pre-trained convolutional neural networks onto the first three modes identified with principal component analysis (PCA). The PCA plots indicated that, among all the models considered, Xception showed superior capability in delineating distinct locations for different FIV response regimes, based on the lock-in frequency and the presence of harmonics in the driving fluid forces. Moreover, the PCA plots also showed that increasing the structural damping ratio resulted in a diminished disparity in the dynamics of the identified FIV response regimes, leading to a less discernible separation between the regimes in the plots.