基于ConvLSTM与LiteFlowNet架构的粒子图像测速方法

打开文本图片集

关键词：粒子图像测速；深度学习；注意力机制；卷积长短期记忆网络 中图分类号：TP391文献标识码：A doi：1O.37188/CJLCD.2025-0052 CSTR：32172.14.CJLCD.2025-0052

Abstract： In particle image velocimetry （PIV），neural network-based methods ten face challenges when hling high-speed or complex nonlinear flows. These challenges include rapid changes in particle positions，which lead to difficulties in tracking matching，limited feature scale extraction， insuficient ability to capture effective features. To address these issues，a novel flow field estimation dynamic particletracking enhancement model LiteFlowNet-CL is proposed，based on the combination ConvLSTM the LiteFlowNet architecture.The study firstly enhances the ability the LiteFlowNet model to identify represent complex flow patterns， then leverages the temporal modeling advantages the ConvLSTM network to effctively suppress tracking errors high-speed moving particles across different time steps，thereby significantly reducing the likelihood particle image feature tracking loss. To validate the effctiveness the proposed model，this paper conducted comparative performance tests ablation experiments by using simulated particle images.Experimental results show that the improved velocity field estimation model achieved a root mean square error O.1OO 4. Compared with the classical LiteFlowNet optical flow estimation model，the error was reduced by 10.52% ，whilea further error reduction 1.463% was observed when benchmarked against the widely adopted high-performance LiteFlowNet-en model in PIV domain.The proposed model was verified to effectively enhance the capability capturing complex flow field characteristicsin particle image velocimetry，with itserrorprecision meeting experimental requirements for turbulence analysis. This achievement was recognized as providing a new technical pathway for PIV algorithm optimization， its application value was confirmed in promoting the development fluid mechanics experimental measurement technologies toward higher spatiotemporal resolution. The research methodology implementation process were systematically described，with comprehensive quantitative comparisons presented to validate the performance improvements.

words： particle image velocimetry； deep leaming； atention mechanism； convolutional long short-termmemory networks

1引言

在现代流体力学和相关工程领域的研究中，普遍存在各种复杂流动现象，如湍流、旋流、燃烧流以及多相流等[1-3]，传统的流体测量方法难以提供准确且实时的流场信息，如何有效地获取流场的准确定量信息成为流体力学研究中的重要课题。（剩余17279字）