基于测点聚类的POD-BPNN风压重构方法  

A wind-pressure reconstruction method of POD-BPNN using measurement-point clustering

作　　者：杜晓庆[1] 沈祥宇 董浩天 陈统岳 Du Xiaoqing;Shen Xiangyu;Dong Haotian;Chen Tongyue(Shanghai University,Shanghai 200444,China)

机构地区：[1]上海大学,上海200444

出　　处：《土木工程学报》2024年第9期11-21,共11页China Civil Engineering Journal

基　　金：国家自然科学基金(51978392,52008239)。

摘　　要：文章提出本征正交分解(POD)与聚类分析结合的结构表面风压测点分类与关键测点布置方法,基于少量测点的风压数据,通过POD与误差反向传播神经网络(BPNN)方法实现方柱结构表面风压时程的重构。机器学习数据集为多风向角均匀来流下单方柱测压风洞试验得到的测点风压时程。将44个测点的风压时程数据POD降维,并采用K-means++聚类分析得到方柱周向轮廓系数分布,并基于轮廓系数的多风向角平均值,得到12、16、20和24个关键测点的轴对称布置方案。以关键测点的风压时程数据为训练集,采用POD-BPNN方法重构方柱表面其余测点所在位置的风压时程,并将风压时程及其统计值同试验结果对比。从12~20测点方案,风压重构精度逐步提升;20测点和24测点方案的重构风压差异较小,二者都能较好地重构方柱表面风压分布,仅在0°风向角方柱脉动风压误差偏大。A method of classifying measurement points and arranging key measurement points on the surface of a structure is proposed by combining Proper Orthogonal Decomposition(POD)and clustering analysis.Based on the wind pressure data from a small number of measurement points,the wind pressure time series on the surface of a square column are reconstructed using POD and Back Propagation Neural Network(BPNN).The machine learning data set is the pressure time series obtained from the wind tunnel test of uniform flow past a single square cylinder at various incidences.The time series data of 44 pressure measurement points are reduced in dimensionality by POD.K-means++clustering analysis is used to obtain the circumferential distribution of Silhouette score on the square column.Based on the multi-incidence averages of the Silhouette score,the axisymmetric arrangements of 12,16,20,and 24 key measurement points are obtained.The pressure time series of the key measurement points are used as the training set,and the pressure time series of the remaining measurement points are reconstructed by the POD-BPNN method.The reconstructed pressure time series and its statistical values are compared with the experimental results.The accuracy of wind pressure reconstruction is gradually improved when the number of measurement points increases from 12 to 20.The difference in the reconstructed pressure between the arrangement schemes of 20 and 24 measurement points is small,both schemes can reconstruct the wind pressure distribution on the surface of the square column well,and only the error of fluctuating wind pressure on the square column at 0°incidence angle is relatively large.

关 键 词：风压时程重构 聚类分析 本征正交分解 误差反向传播神经网络 风压测点布置 

分 类 号：TU312.1[建筑科学—结构工程]