Gas-liquid Two-phase Flow Pattern Recognition Based On Random Forest

Among the many parameters describing the gas-liquid two-phase flow,the flow pattern has important academic significance and industrial application value for the operation monitoring,process control and safety assurance of the two-phase flow system.Due to the complexity and randomness of two-phase flow,the flow pattern identification of gas-liquid two-phase flow has been a difficult problem that has not been solved well at home and abroad for a long time.This paper collects and analyzes the image signal and pressure difference signal of the horizontal pipe gas-liquid two-phase flow,extracts the characteristic quantity that effectively distinguishes the flow pattern and combines it with the recognition model to complete the analysis of laminar flow,wavy flow,bubbly flow,and plug-like flow.Recognition of the five flow patterns of flow and annular flow;the main research includes signal data preprocessing,feature extraction and selection,classifier training and classification algorithm execution.First of all,the image processing technology is applied to the flow pattern recognition of gas-liquid two-phase flow.For image recognition,it is necessary to transform the original image in high-dimensional space to low dimensional space,and filter out the features that have no contribution to image classification.After preprocessing the image signal of the two-phase flow,this paper extracts features from the three aspects of the gray distribution,shape feature and texture feature of the flow image based on the gray histogram,invariant moments and gray co-occurrence moments.Provide a strong basis for flow pattern identification.Secondly,for the non-linear and non-stationary signal such as the differential pressure fluctuation signal,the traditional HHT method has serious drawbacks such as difficult to determine the number of screenings,too rough trend function,and modal aliasing.An improved HHT method,namely complementary set is proposed.Empirical mode decomposition method(CEEMD),this paper applies it to the identification of gas-liquid two-phase flow pattern to decompose the differential pressure fluctuation signal of the two-phase flow.Two methods of empirical mode decomposition and complementary ensemble empirical mode decomposition are used to analyze the differential pressure signal.The advantages of CEEMD in the decomposition of differential pressure fluctuation signal have been compared and verified.The corrected method can not only solve the noise residue in EMD decomposition The defect can also change the phenomenon of over-decomposition of the inherent modal function of the signal.The CEEMD method has a very high resolution in the entire time and frequency range,and better characterizes the time-frequency characteristics of the differential pressure fluctuation signal.Based on CEEMD,the time-domain and frequency-domain features of the differential pressure signal are extracted and analyzed,and the time-domain feature parameters and IMF component energy of the differential pressure signal are obtained.The eigenvectors constituted can effectively identify five flow patterns.Finally,it is proposed to apply the random forest algorithm to the flow pattern recognition of gas-liquid two-phase flow.Select the appropriate algorithm as the decision tree node splitting algorithm.Based on the random forest algorithm,the image signal and pressure fluctuation signal of the two-phase flow are classified and identified,and the parameters of the random forest are optimized.Experiments have shown that the effective features extracted from the image signal and pressure difference signal of the two-phase flow can be well combined with the random forest algorithm in this paper.The model has good stability and generalization performance,and it can accurately identify the five flow patterns.The rate is higher.