Flow Pattern Recognition Method For Gas-liquid Two-phase Flow Based On Image Processing And Wavelet Convolution Neural Network

Gas-liquid two-phase flow is widely used in industrial production and it has important research value in atomic energy engineering,petrochemical engineering,aerospace engineering and chemical engineering.The gas-liquid two-phase flow pattern identification technology has great significance to the operation monitoring,safety assurance and process control in the industrial production process.Because of the dynamics,complexity and randomness of two-phase flow,researchers have not been able to get a clear understanding of the dynamic characteristics of two-phase flow and its transformation regularity.At present,the low accuracy and low efficiency of gas-liquid two-phase flow pattern recognition technology have not been well solved.In this paper,the horizontal pipe gas-liquid two-phase flow image signals were collected by high-speed photography method,and the Wavelet convolution neural network is used for feature extraction and classification of image signals,so as to complete the recognition of four flow types: layered flow,sluice flow,annular flow and wave flow.This paper mainly studies image data preprocessing,image feature extraction,wavelet transform and convolutional neural network training.Firstly,image technology and feature extraction technology are applied to two-phase flow pattern recognition.Laplacian high-pass filter,mean filter and other techniques are used to preprocess the image,which makes the image smoother and improves the contrast,reduces the influence of the image noise brought by the experimental environment and equipment,and is conducive to the subsequent image feature extraction.Then,from the gray distribution and edge shape of the flow image,the directional gradient histogram and gray histogram are extracted to carry out features,providing global information in the image for subsequent experiments.Secondly,the pooling layer of traditional convolutional neural network has some problems,such as loss of feature information,limited size of receptive field and slow convergence speed.In this paper,two-dimensional discrete wavelet transform technology is applied to the pooling layer of convolutional neural network.Through two-dimensional discrete wavelet transformation technology can obtain the characteristic figure low frequency part as a thumbnail,the optimization reduces the pooling layer to deal with the amount of data,and retains the characteristics of the main characteristics of the information,to avoid the traditional way of pooling feature information loss problem while expanding the receptive field of pooling layer,improves the speed of the model convergence.Finally,it is found that the traditional model uses a single feature,but different features have complementarity.The complementarity between features can enrich the classification strategy of the model and reduce the probability of model overfitting.In this paper,global features such as gray histogram and gradient direction histogram are combined with local spatial features extracted by convolutional neural network as the input features of the convolutional neural network to improve the information complementarity between features.The experimental results show that the two optimizations are helpful to improve the accuracy of the recognition model,and the final recognition accuracy reaches 98%,which is generally higher than other comparison models.