Gas-liquid Two-phase Flow Pattern Identification Based On Noisy Independent Component Analysis

Gas-liquid two-phase flow phenomenon exists widely in industry, agriculture and daily life. Because flow pattern not only affects the pressure loss and heat transfer characteristics of gas-liquid two-phase flow greatly, but also influences the measurement accuracy of flow parameters and the motion characteristics of two-phase flow system. So the research of gas-liquid two-phase flow pattern identification has academic importance and industrial application value. The interaction between gas phase and liquid phase increases the difficulty of flow pattern identification. Blind source separation technique can provide independent signals which represent characteristics of different flow patterns more effectively. So the accuracy of two-phase flow pattern identification can be improved using this technique.Considering measured signals corrupted by noise, the noisy Independent component analysis technique is used to process differential pressure fluctuation signals of gas-liquid two-phase flow. A noisy independent component analysis algorithm based on invasive weed optimization is proposed, aiming at the problems of poor stability for traditional Independent component analysis algorithms. The algorithm adopts negentropy with robust performance as the objective function based on analyzing the advantages and disadvantages of various independence criterions. It can also eliminate noise effects by introducing bias removal technique. Meanwhile, considering the influence of nonlinear function on estimating negentropy, an invasive weed optimization algorithm with excellent global optimization performance is introduced to optimize the objective function. Then optimal separation matrix and independent signals are obtained. Simulation experiments and test signals both verify the validity of noisy ICA algorithm based on invasive weed optimization. In contrast to FastICA algorithm and Fast NoisyICA algorithm, the proposed method can achieve better separation results. The method is adopted to separate the differential pressure fluctuation signals and the separated signals which represent characteristics of different flow patterns are obtained. After analyzing the time-domain characteristics of differential pressure signals, mean of measured differential pressure signals, skewness and kurtosis of separated signals are regarded as characteristic parameters to identify flow patterns.Empirical mode decomposition and wavelet decomposition are used, respectively, to decompose the differential pressure signals of gas-liquid two-phase flow into various characteristic scales. The energy of each component is regarded as characteristic parameters to identify flow patterns. Support vector machine models are trained by the flow pattern characteristics parameters extracted by IWO-NICA algorithm, Fast ICA algorithm, empirical mode decomposition and wavelet decomposition respectively. After training, gas-liquid two-phase flow pattern recognition results are obtained. Experimental results show that compared with other three methods, combination of IWO-NICA and support vector machine can achieve the best identification results, and the accuracy rates of some flow patterns are above 90%.