Abstract: There exist no accurate evaluation indexes in traditional identification methods of flow regime.Usually,researchers can only give ambiguous description of the characteristics of each working condition.This brings out great dependence on judgement of researchers and makes automatic and on line identification of flow regime impossible.In order to overcome the shortcomings of traditional identification methods,mineral oil,air and water are used as working fluids to simulate the crude oil,natural gas and water multiphase flow in horizontal transportation pipeline,a piezo-resistance differential pressure transducer is adopted to measure the differential pressure of the flow.Thus the differential pressure signals which reflect the characteristics of fluctuation of oil-gas-water multiphase flow are obtained.By using the reconstruction of pesudo-phase space algorithm (Grassberger-Procaccia algorithm),the correlation dimensions of the denoised differential pressure signals are obtained.Based on the correlation dimensions of different flow regimes and working conditions,the characteristic vectors of different flow regimes and different working conditions are constituted.Thus the quantification of characteristics of flow regime of oil-gas-water multiphase flow is realized.The characteristic vectors are then input into the Radius Basis Function (RBF) neural network.Through training with study samples,the trained RBF neural network can automatically identify the flow regime of test samples (differential pressure signals of unknown flow regimes) on line.Results showed that the method discussed in this paper had the merits such as high accuracy,fast response and no artificial intervention.It has a promising prospect for application in petroleum and nuclear industries.

Key words: fractal, correlation dimension, Radius Basis Function, neural network, multiphase flow, flow regime, flow regime identification

摘要： 传统的流型识别方法对流型特征没有一个量化的评价指标,只能由识别者采用模糊的语言描述每种工况的特征,在很大程度上依赖于每个识者的主观判断,很难实现流型的在线自动识别.为了克服传统流型识别方法的缺点,本文在实验室采用机油、空气和水作为工质来模拟现场的水平管内的油气水多相流流动,采用压阻式压差传感器对水平管内的油气水多相流流动的压差进行测量,得到反映油气水多相流流动波动特性的压差信号.运用分形理论中的重构相空间算法(即Grassberger-Procaccia算法),算出经过滤噪处理后的油气水多相流流动的压差信号序列的关联维数,以不同流型不同工况的关联维数为基础构成不同流型不同工况下的多相流的特征向量,从而量化了油气水多相流的特征.将特征向量输入径向基函数(Radius Basis Function,简称RBF)神经网络,通过学习样本的学习,即可用来对测试样本(未知流型的压差信号)进行自动在线的流型识别.结果显示,该识别方法具有高精度、快速及不需人工干预等优点,在石化、核能等行业必将有广阔的应用前景.

关键词: 分形, 关联维数, 径向基函数, 神经网络, 多相流, 流型, 流型识别