改进BP神经网络在气液两相流流型识别中的应用

Application of improved BP neural network in identification of air-water two-phase flow patterns

采用统计理论与分形理论相结合的方法对水平管内空气 水两相流的压差波动信号进行分析 ,得到了压差均值、标准差、偏斜度、能量份额、盒维数、关联维数和Hurst指数 7个参数 ,并将上述参数作为流型的特征向量输入自适应学习率的改进BP神经网络 ,通过对训练样本的学习 ,网络可以实现对未知流型的客观识别 .仿真结果表明 :统计参数与分形参数相结合得到的流型特征向量可以很好地反映各流型之间的差异 ,网络识别率高达 93% ,并且改进后的BP网络具有收敛速度快、不易陷入局部极小的优点 .

The pressure difference fluctuation signals of the air-water two-phase flow in a horizontal pipe were analyzed by statistical and fractal theory. Seven parameters, i.e. the averaged value of the pressure difference, standard deviation, skewness, energy ratio, box dimension, correlation dimension and Hurst index were obtained and be used as the characteristic vector of an improved BP neural network with self-adapted learning ratio. Learning form training samples, the network could accomplish the objective identification of the unknown flow patterns. The simulated results showed that the flow pattern characteristic vector which was obtained by statistical and fractal parameters could reflect the difference between various flow patterns and the recognition possibility of the network could reached up to about 93 percent. Moreover, the improved BP neural network had the merits of fast convergence for simulation and avoidance of local minimum.