华南理工大学学报(自然科学版) ›› 2024, Vol. 52 ›› Issue (12): 119-126.doi: 10.12141/j.issn.1000-565X.240077

所属专题: 2024年流体动力与机电控制工程

• 流体动力与机电控制工程 • 上一篇    下一篇

气/水介质湍流噪声强度的BP神经网络预测模型

朱睿1,2(), 刘宇2, 梁钰迎2, 申传鹏2   

  1. 1.厦门大学 航空航天学院,福建 厦门 361102
    2.西藏民族大学 信息工程学院,陕西 咸阳 712082
  • 收稿日期:2024-02-15 出版日期:2024-12-25 发布日期:2024-06-14
  • 作者简介:朱睿(1980—),男,博士,副教授,主要从事实验/计算流体力学研究。E-mail: zhurui@xmu.edu.cn
  • 基金资助:
    西藏自治区重点研发计划项目(XZ202401ZY0102);厦门市自然科学基金资助项目(3502Z20227179);教育部人文社会科学规划基金资助项目(23XZJAZH001);西藏自治区自然科学基金重点资助项目(XZ202401ZR0055);福建省自然科学基金资助项目(2022J01058);中国航空科学基金资助项目(2023Z032068001);水声对抗技术重点实验室基金资助项目(JCKY2024207CH05);西藏民族大学基金资助项目(Y2024050)

BP Neural Network Prediction Model for Turbulent Noise Intensity in Gas/Water Medium

ZHU Rui1,2(), LIU Yu2, LIANG Yuying2, SHEN Chuanpeng2   

  1. 1.School of Aerospace Engineering,Xiamen University,Xiamen 361102,Fujian,China
    2.College of Information Engineering,Xizang Minzu University,Xianyang 712082,Shaanxi,China
  • Received:2024-02-15 Online:2024-12-25 Published:2024-06-14
  • Supported by:
    the Key Research and Development Program of Xizang Autonomous Region(XZ202401ZY0102);the Humanities and Social Sciences Program of the Ministry of Education(23XZJAZH001);the Key Natural Science Foundation of Xizang Autonomous Region(XZ202401ZR0055);the Natural Science Foundation of Fujian Province(2022J01058)

摘要:

针对常规气/水介质湍流噪声强度数值求解耗时长、效率低的问题,建立相似流动条件下气/水介质钝体/空腔湍流噪声强度的神经网络预测模型,以根据水下噪声强度高效预测同雷诺数下的气动噪声强度,为气/水介质湍流噪声强度的高效预测以及控制方法、测噪试验介质可替换性的研究提供技术支撑。开展粒子图像测速试验,测得开缝圆柱绕流速度以验证数值方法的有效性。采用大涡模拟方法构建气/水介质湍流噪声强度的数值模型,模型的平均速度计算误差小于2.25%,测试值与模拟值的斯特劳哈尔数误差仅为0.89%。由数值模型获得1 338条数据信息,用于构建训练样本数据集。基于关键流动参数构建BP(反向传播)神经网络来映射气/水介质湍流噪声关系,并采用Levenberg-Marquardt算法训练预测模型,该模型以Sigmoid函数作为激活函数,包含8个输入神经元、1个输出神经元以及单隐藏层。研究结果表明:所提出的BP神经网络预测模型可根据水下噪声强度预测同雷诺数下的气动噪声强度,最大预测误差小于6.21 dB,平均误差为0.44 dB,模型泛化能力良好,在测试集非规律点处的总声压级误差为0.27 dB;相同硬件环境下,数值求解用时约30 h,BP神经网络模型预测用时仅70 s,显著提升了计算效率。

关键词: 气/水介质, 湍流噪声, BP神经网络, 预测模型, 开缝圆柱, 粒子图像测速

Abstract:

Addressing the issues of long computation time and low efficiency in numerical solutions for turbulent noise intensity in conventional air/water medium, a neural network prediction model for bluff body/cavity turbulent noise intensity in air/water medium under similar flow conditions was established. This model efficiently predicts aerodynamic noise intensity at the same Reynolds number based on underwater noise intensity, providing technical support for the efficient prediction and control methods of turbulent noise intensity in air/water medium, as well as the research on the interchangeability of noise testing medium. Particle image velocimetry experiments were conducted to measure the flow velocity around open-slot cylinders, validating the effectiveness of the numerical methods. A numerical model for turbulent noise intensity in air/water medium was constructed using the large eddy simulation method, achieving an average velocity calculation error of less than 2.25% and a Strouhal number error of 0.89% between test and simulated values. The numerical model generated 1 338 data points, which were used to construct a training sample dataset. Then, a backpropagation (BP) neural network was built based on key flow parameters to map the relationship between turbulent noise in air/water medium. The Levenberg-Marquardt algorithm was employed to train the predictive model, with the Sigmoid function selected as the activation function. The network comprises 8 input neurons, 1 output neuron, and a single hidden layer. The results demonstrate that the proposed BP neural network prediction model can predict aerodynamic noise intensity at the same Reynolds number as underwater noise intensity, with a maximum prediction error of less than 6.21 dB and an average error of 0.44 dB; that the model exhibits good generalization ability, with an error of 0.27 dB at irregular points in the test set; and that, under comparable hardware conditions, the numerical solution method required approximately 30 hours for computation, while the BP neural network prediction model took only 70 seconds, significantly improving the computational efficiency.

Key words: gas/water medium, turbulent noise, BP neural network, prediction model, open-slot cylinder, particle image velocimetry

中图分类号: