Main body shape optimization of non-body-of-revolution underwater vehicles by using CNN and genetic algorithm

In this study, a three-dimensional main body shape of non-body-of-revolution underwater vehicle is established according to the shape of fish flat body, and the contour is controlled by three design variables. The Latin hypercube design is used to generate 500 samples. The grayscale image generated by the contour and the drag coefficient obtained by CFD simulation are used as the data set. The convolutional neural network(CNN) of the same structure is constructed to predict the drag coefficient and volume respectively, the curve that can reflect the characteristics of the three-dimensional model is manually generated as the input of the CNN model. In addition, 31 new samples were generated to verify the accuracy of CNN model, the error of the prediction results and the label is less than 3 %. Compared with the prediction results of the multi-layer perceptron and response surface methodology, the accuracy is very high. The mean square errors of the drag coefficient and volume predicted by CNN are 1.126 x 10-8 and 4.193 x 10-10, respectively. The mean square errors of the drag coefficient and volume predicted by MLP are 3.149 x 10-8 and 1.920 x 10-9, respectively. It can perfectly predict the drag coefficient and volume. Through genetic algorithm combined with CNN, the non-body-of-revolution shape with the lowest drag coefficient under specific volume is found. The drag coefficient is 0.01646, and its hydrodynamic data is better than the conventional body-of-revolution shape with the same volume. In this paper, the hydrodynamic coefficients of three-dimensional non-body-of-revolution is accurately predicted by using the contour image, combined with CNN and genetic algorithm, and the optimal solution of the shape is found, which provides a new idea for the optimization design of underwater vehicles.