Machine learning-based deep data mining and prediction of vortex-induced vibration of circular cylinders

By means of numerical, experimental and field measurements, an unprecedented amount of data regarding vortex-induced vibration (VIV) of circular cylinders has become accessible recently. In this study, machine learning (ML) models are developed to predict the transverse VIV amplitude Y* and amplitude-velocity relations. First of all, an ML database is established by aggregating reliable, published data. Then, three ML techniques are developed, including support vector regression with particle swarm optimization (PSO + SVR), extreme learning machine (ELM), and improved ELM combined with preprocessed least squares QR decomposition (PLSQR + ELM). Three crucial parameters, mass damping ratio & alpha;, Reynolds number (Re) and mass ratio m*, are extracted by the partial least square (PLS) method to characterize VIV. The results show that the prediction accuracy of all algorithms is acceptable in the absence of noise, with the ELM-type being the highest. When noise is taken into account, PLSQR + ELM is the most efficient and robust. A new linear relation between the VIV amplitude Y* and log10(& alpha; /Re) is proposed, which is demonstrated to be better than published ones.

Automated summary

By using numerical, experimental, and field measurements, a large amount of data on vortex-induced vibration (VIV) of circular cylinders has been obtained. Machine learning (ML) models are developed in this study to predict the transverse VIV amplitude Y* and amplitude-velocity relations. Three ML techniques, including support vector regression with particle swarm optimization, extreme learning machine, and improved extreme learning machine combined with preprocessed least squares QR decomposition, are employed. The results demonstrate acceptable prediction accuracy for all algorithms, with the ELM-type algorithm performing the best.