流场中的特征识别一直是流体力学中一个重要的问题。对于流场而言，无论是从加速流场计算的角度，还是从探究物理意义的角度，始终存在着对流场进行特征识别的需求。然而，当流场面临海量高维信息的时候，现有的方法在流场计算和数据后处理上存在一定的困难。

深度学习在图像等领域的应用中以其更高的效率和更优的精度备受关注，为流场识别提供了参考。在流体力学领域，可以通过以卷积神经网络为首的深度学习模型对流场进行特征识别和降维操作。识别出的特征不仅能够进行流场预测，也可以影响流场重构结果的精度。进一步， 深度强化学习能够对流场进行流动控制，并且可以借助流场特征模型提高流动控制的效率。

因此，本文的主要研究目的是基于卷积神经网络为代表的深度学习方法，对流场进行识别降维，并且通过对流场的预测、重构和流动控制过程将深度学习方法应用到实际问题中。主要内容分为以下几个方面：

1. 基于流场信息的特征识别与预测

基于卷积神经网络建立了深度学习特征识别模型。该模型能够通过尾流的非定常流动反映绕流体表面压力震荡。 研究的对象为圆柱绕流流动中旋涡的交替非定常流场速度场，通过识别其特征进而预测圆柱表面压力系数。同时，本文先后通过对卷积特征的可视和借助迁移学习的泛化，均印证了特征提取的准确性。最终，将该方法成功应用到更为复杂的带空化的水翼绕流中，通过空化流动下速度场及空化体积分数成功预测翼型表面压力系数。

2. 基于特征识别的流场重构与应用

在深度学习能够成功提取流场特征的前提下，借助自动编码器能够编码解码的网络特点，通过速度场的压缩降阶过程重构原始流场信息。该研究通过对二维流场速度场的编码与解码，能够得到流场特征并重构出原始流场。同时，通过对深度学习网络模型不同拓扑结构下重构流场的对比，得到重构结果与特征维度的关系。进一步，将上述流场重构过程应用在近壁湍流流场中，通过卷积自动编码器结构的重构能力和对特征维度的限制能够重构湍流中仅包含大尺度的流场，为近壁湍流建模提供基础。

3. 考虑流场特征识别的强化学习与流动控制

对流场有了上述认识的基础上，进一步基于强化学习实现对流场的流动控制。借助对双方柱绕流中前方柱位置的自适应优化过程，得到使尾流场流场震荡最低的流动状态。进一步，由于该强化学习的训练过程较为耗时，使用基于特征识别的预测模块替代该模型中流场仿真部分，能够成功对强化学习进行加速。

Feature detection in flow fields has always been an important issue in fluid mechanics. For flow fields, there is always a need for feature detection of flow fields, both from the perspective of accelerating flow field calculations and from the perspective of exploring the physical implications. However, when flow fields are confronted with massive amounts of high-dimensional information, the existing methods for flow field computation and data post-processing have certain limitations.

Deep learning has received much attention for its higher efficiency and better accuracy in applications such as images, providing a reference for flow field detection. In the field of fluid mechanics, flow fields can be characterised and dimensionality reduced by deep learning models led by convolutional neural networks. The identified features can not only predict the flow field, but also influence the flow field reconstruction results. At the same time, the extracted features can also be used as key information for flow field reconstruction, and the dimensionality of the features directly affects the accuracy of the flow field reconstruction results. Further, deep reinforcement learning can be used in the flow control of the flow field and can be used to influence the accuracy of the flow field reconstruction. and can be used to improve the efficiency of flow control with the help of flow field feature models.

Therefore, the main research objective of this paper is to identify and reduce the dimensionality of flow fields based on deep learning methods represented by convolutional neural networks, and apply them to practical problems through the prediction, reconstruction and flow control process of flow fields. The main components are divided into the following areas.

1. Feature detection and prediction based on flow field information

A deep learning feature recognition model based on convolutional neural networks has been developed. The model is able to reflect surface pressure oscillations around the fluid through an unsteady flow wake. The process is realised on the flow around a cylinder, and feature detection and extraction of the velocity field of the flow field is used to predict the pressure coefficients on the surface of the cylinder. The accuracy of the feature extraction is confirmed by visualisation of the convoluted features and generalisation with the aid of transfer learning. Finally, the method is successfully applied to a more complex hydrofoil bypass flow with cavitation, and the velocity field and cavitation volume fraction under cavitation flow are used to successfully predict the airfoil surface pressure coefficient.

2. Flow field reconstruction and application based on feature detection

On the premise that deep learning can successfully extract the flow field features, the original flow field information is reconstructed through the compression and downscaling process of the velocity field with the help of the network features of the autoencoder that can encode and decode. The study is able to obtain the flow field features and reconstruct the original flow field by encoding and decoding the velocity field of a two-dimensional flow field. At the same time, a comparison of the reconstructed flow field under different topologies of the deep learning network model is carried out to obtain the relationship between the reconstructed results and the dimensionality of the features. Further, the above flow field reconstruction process is applied to the near-wall turbulent flow field, and the reconstructive capability of the convolutional autoencoder structure and the restriction on the feature dimension can reconstruct the flow field that contains only large scale in the turbulent flow, providing a basis for the near-wall turbulent flow modelling.

3. Reinforcement learning and flow control considering flow field feature detection

Based on this understanding of the flow field, further flow control of the flow field is achieved based on reinforcement learning. By means of an adaptive optimization process for the position of the front column in the flow around two square cylinder flow, the flow state that minimises the oscillation of the flow field in the wake is obtained. Further, as the training process for this reinforcement learning is time consuming, the use of a prediction module based on feature recognition is used instead of the flow simulation part of the model to successfully accelerate the reinforcement learning.