| Gas-liquid two-phase flow is a common phenomenon in nature and industrial production.Due to the mutual interaction between the two phases and the compressibility of gases,there are some difficulties in the study of its performance.Gas-liquid two-phase flow exhibits different flow patterns during the flow process due to changes in flow rate,gas-liquid ratio,pressure,etc.,and thus exhibits different flow characteristics,bringing theoretical and practical problems in various applications such as combustible ice mining,oil and gas gathering and transportation,and two-phase flow metering.This article focuses on gas-liquid two-phase flow in a vertical circular tube.We set up a testing experimental platform and use a high-speed camera to capture the flow state under different parameter conditions,and then use machine learning algorithms to extract and compare image features to quickly and accurately identify the flow patterns.For the selected flow patterns,we propose methods for processing contours,displacements,and velocities using digital image and video sequence algorithms.We validate the above flow pattern recognition and feature extraction methods using images of gas-liquid two-phase flow in a vertical circular tube.The main contents of this article include:1.We design and build a gas-liquid two-phase flow testing system using a vertical circular tube as the carrier and a high-speed camera to capture the flow state.This system can create five typical flow patterns of bubbly flow,slug flow,annular flow,dispersed flow,and string flow in the vertical circular tube.The shooting frequency is up to 3000 frames per second,and the gas phase velocity range is 0-1.5×103 ml/min,and the liquid phase velocity range is 0-3.5×103 ml/min.2.We propose a gas-liquid two-phase flow pattern recognition model based on Alex Net.By training the model with machine learning,optimizing parameters,and analyzing training and result validation,the improved model training accuracy can reach 88.80%.3.Based on the small sample characteristics of gas-liquid two-phase flow measurement data,we propose a recognition algorithm combining support vector machines.We establish a new model combining convolutional neural networks with transfer learning and support vector machines,using VGG16,Res Net50,and Inceptionv3 pre-trained networks for image feature extraction,and then training a new classifier with the feature data,using grid search algorithms for hyperparameter tuning.The results show that the VGG16-SVM hybrid model can achieve an accuracy of 98% in discriminating different flow patterns,with recognition accuracy of bubbly flow and slug flow approaching 99%.4.We propose a method for analyzing the characteristics of slug flow in a vertical circular tube based on digital image processing,video sequence moving object detection and tracking algorithms.This method reads each frame from a slug flow video,extracts foreground objects using a background subtractor(MOG2),and processes the foreground objects with binary and median filtering to extract object contours.We analyze the object contours with the Lucas-Kanade optical flow method,calculate the object displacement and velocity,and store the velocity data in an Excel file using the Pandas library,followed by further processing using data processing software.This work combines high-speed camera experimental testing and machine learning algorithmbased gas-liquid two-phase flow pattern recognition and flow characteristic recognition methods to achieve recognition of five different flow patterns in a vertical circular tube based on machine learning technology,and to extract the contours,velocities,and pixel displacements between adjacent frames of slugs in slug flow video sequences using image processing and video analysis technology.This work can provide technical support for engineering applications such as gas-liquid two-phase flow pattern recognition,feature understanding under different flow patterns,gas-liquid two-phase flow control,and accurate measurement. |
PDF Full Text Request