Data-Driven Prediction of Experimental Hydrodynamic Data of the Manta Ray Robot Using Deep Learning Method

Abstract

To precisely control the manta ray robot and improve its swimming and turning speed, the hydrodynamic parameters corresponding to different motion control variables must be tested experimentally. In practice, too many input control parameters will bring thousands of groups of underwater experiments, posing challenges to the duration and operability of the engineering project. This study proposes a generative adversarial network model to reduce the experimental period by predicting the hydrodynamic experimental time-series data of forces and torques in the three-coordinate directions in a Cartesian coordinate system through different combinations of motion control parameters. The motion control parameters include the rotation amplitude, frequency, and phase difference of the four steering gears which drive the pectoral fins. We designed the prototype and experimental platform and obtained 150 sets of experimental data.To prevent overfitting, the size of the dataset was expanded to 2250 groups by slicing time series, and the subsequences of varying lengths were extended to the same length by LSTM. Finally, the GAN model is used to predict the hydrodynamic time series corresponding to the different motion parameters. The results show that the GAN model can accurately predict the input both from the validation set and the unlearned interpolated motion parameters. This study will save experimental time and cost and provide detailed hydrodynamic experimental data for the precise control of manta rays.