Flow feature extraction models based on deep learning

Abstract

Extraction and recognition of the features of flow field is an important research area of fluid mechanics. However, the wake flow field of object immersed in fluid is complicated in the case of medium- and high-Reynolds number, thus it is difficult to extract and recognize the key features by using traditional physical models and mathematical methods. The continuous development of deep learning theory provides us with a new method of recognizing the complex flow features. A new method of extracting the features of the flow time history is proposed based on deep learning in this work. The accuracy of four deep learning model for feature recognition is studied. The results show that the proposed model can identify different characteristics of the wake time history and object shapes accurately. Some conclusions can be obtained below (i) The model based on convolutional layers has higher accuracy and is suitable for analyzing the features of flow time history data. (ii) The residual convolutional network, with a deeper structure and more complex inter-layer structure, has highest accuracy for feature recognition. (iii) The proposed method can extract and recognize the flow features from the perspective of physical quantities time history, which is a high-accuracy method, and it is an important new way to study the features of flow physical quantities.