基于深度学习的PIV流场图像修复技术

粒子图像测速技术（PIV）是空天动力装置研究中常用的流场测试方法。但对具有复杂流动特征的燃烧室，通过传统互相关算法处理得到的流场结果往往具有一定缺陷。本文将深度学习应用于PIV后处理中，以实现流场数据的异常检测和修复。在甲烷预混对冲火焰数据集上，将异常划分为两种类型，并搭建U-Net卷积神经网络架构。经过训练和优化，模型以较高置信水平识别两类异常并使用不同策略自适应修复，过滤噪声并保留原始正常数据。同时模型具有较好的可迁移性，可以为其它种类的流场数据修复提供参考。与POD迭代法和中值滤波相比，神经网络强大的非线性特征具有明显的优势，这种方法不仅修复率高，而且在不同工况下鲁棒性好。

Inpainting PIV Flow Fields with Deep Learning

Particle Image Velocimetry(PIV) is widely used to measure the flow fields in aerospace researches. However, it can hardly tackle complex flow fields in combustors, and defects could usually be found in the flow field after the post-processing, where traditional cross-correlation method is applied. The deep learning method is applied in the PIV post-processing to achieve detection and inpainting of abnormal flow field data. On the counterflow premixed methane flame dataset, the abnormal data is classified into two categories, which is fixed with a U-Net-based convolutional neural network model. After training and optimizing, the model can detect anomalies at a high confidence level and repair them with different strategies. The noise is filtered while the original normal data retains. Meanwhile, the model has good transferability, which might help other flow data inpainting researches. Owing to the strong nonlinearity of the model, the proposed method behaves with high accuracy and robustness, which shows huge advantages over traditional inpainting models such as POD iteration method and median filtering.