A gas-liquid two-phase flow field real-time synchronous measurement method and system

Abstract

The application discloses a kind of gas-liquid two-phase flow field real-time synchronous measurement method and system, it is related to flow field measurement technical field.The method comprises: obtaining the high-speed image collected by high-speed camera;The high-speed camera uses setting parameter;The high-speed image is post-processed, and optimization image is obtained;The post-processing includes image denoising, gray segmentation, contour segmentation and roundness denoising;PWC-Net network is constructed;The optimization image is extracted using the PWC-Net network, and gas-liquid two-phase flow field is predicted according to the extracted feature, and the final fluid velocity field is obtained.The application can realize the fast, accurate, stable gas-liquid two-phase flow field velocity measurement.

Description

Technical Field

[0001] This invention relates to the field of flow field measurement technology, and in particular to a method and system for real-time synchronous measurement of gas-liquid two-phase flow fields. Background Technology

[0002] Particle image velocimetry (PIV) involves adding tracer particles to a flow field, illuminating the flow field with a laser sheet, capturing a sequence of particle images using a high-speed camera, and then calculating the particle displacement using image analysis algorithms to obtain the flow field velocity distribution. This technique has the advantages of non-contact and full-field measurement and is widely used in various fluid mechanics experiments and engineering fields.

[0003] Traditional PIV algorithms are mainly divided into two categories: cross-correlation algorithms and optical flow algorithms, each with its own characteristics in principle and application. Cross-correlation algorithms calculate the displacement of tracer particles across two frames of images to obtain the velocity field of the fluid. Optical flow algorithms in PIV are a velocity field measurement method based on pixel motion estimation in an image sequence. Unlike cross-correlation algorithms, optical flow algorithms directly estimate fluid motion by analyzing the grayscale changes of each pixel in the image. However, traditional PIV techniques suffer from low computational efficiency, poor real-time performance, and sensitivity to noise when dealing with complex flow fields. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for real-time synchronous measurement of gas-liquid two-phase flow fields, aiming to solve or improve at least one of the above-mentioned technical problems.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A method for real-time synchronous measurement of gas-liquid two-phase flow field, comprising:

[0007] Acquire high-speed images captured by a high-speed camera; the high-speed camera uses set parameters;

[0008] The high-speed image is post-processed to obtain an optimized image; the post-processing includes image denoising, grayscale segmentation, contour segmentation, and roundness denoising.

[0009] Construct a PWC-Net network;

[0010] The optimized image is feature extracted using the PWC-Net network, and the gas-liquid two-phase flow field is predicted based on the extracted features to obtain the final fluid velocity field.

[0011] Optionally, the setting parameters include at least resolution, exposure time, acquisition frame rate, and number of frames to be saved.

[0012] Optionally, the construction process of the PWC-Net network is as follows:

[0013] Obtain the training dataset; the training dataset includes training data and corresponding data labels;

[0014] A pre-trained network is constructed based on a CNN network; the CNN network includes convolutional layers, activation functions, pooling layers, and fully connected layers;

[0015] The training dataset is input into the pre-trained network, and training is performed with the goal of minimizing the loss between the network output and the data label. When the network meets the preset accuracy requirements, the trained network is identified as the PWC-Net network.

[0016] Optionally, the step of using the PWC-Net network to extract features from the optimized image and predicting the gas-liquid two-phase flow field based on the extracted features to obtain the final fluid velocity field specifically includes:

[0017] The optimized image is input into the PWC-Net network, and multi-scale feature extraction is performed on two adjacent frames to obtain feature maps containing semantic and detail information at different levels.

[0018] Based on the feature map, image deformation is performed, and the two frames of images are aligned to obtain the deformed feature map;

[0019] A similarity metric is obtained for the similarity between each position on the deformable feature map and different optical flow offsets, and a cost volume is constructed based on the similarity metric.

[0020] Based on the cost body, the gas-liquid two-phase flow field is predicted to obtain the final fluid velocity field.

[0021] Optionally, the step of performing image deformation based on the feature map to align two frames and obtain a deformed feature map specifically includes:

[0022] The deformation operation is represented as W(F1,w). Bilinear interpolation is used to deform the feature map F1 of the previous frame, aligning it with the feature map F2 of the next frame. In two-dimensional space, for the coordinates (x,y) on feature map F1, the deformed coordinates are (x+w). x ,y+w y ), where w x and w y These are the components of the optical flow w in the x and y directions, respectively, and their mathematical expressions are:

[0023] W(F1,w)(x,y)=∑ i,j F1(x+i,y+i)·max(0,1-|iw x |)·max(0,1-|jw y |)

[0024] Where i and j represent the interpolation offset.

[0025] Optionally, the step of obtaining a similarity measure based on the similarity between each position on the deformed feature map and different optical flow offsets, and constructing a cost body based on the similarity measure, specifically includes:

[0026] For each position (x, y) and different optical flow offsets (δx, δy) on the deformed feature map, calculate the similarity metric:

[0027] C(x,y,δx,δy)=W(F1,w)(x,y)·F2(x+δx,y+δy)

[0028] Where W(F1,w) represents the deformation operation, w represents the optical flow, F1 represents the feature map of the previous frame, and F2 represents the feature map of the next frame.

[0029] The similarity measures of all positions and optical flow offsets are combined to form a cost volume C; the cost volume C is a four-dimensional tensor used to record the degree of similarity under different positions and optical flow assumptions.

[0030] Optionally, the step of predicting the gas-liquid two-phase flow field based on the cost volume to obtain the final fluid velocity field specifically includes:

[0031] Based on the cost volume, a coarse-to-fine strategy is adopted, and the prediction results are upsampled sequentially according to the resolution from low to high to obtain the final fluid velocity field.

[0032] The present invention also provides a real-time synchronous measurement system for gas-liquid two-phase flow fields, which applies the method described above and includes: a high-power laser source, a high-speed camera, a synchronous controller, an optical correction box, and a host computer;

[0033] Multiple high-speed cameras are arranged around the optical correction box, and the light emission end of the high-power laser light source is located on the top of the optical correction box. Each of the high-speed cameras and the high-power laser light source is connected to the synchronization controller, which is also connected to the host computer.

[0034] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0035] This invention discloses a method and system for real-time synchronous measurement of gas-liquid two-phase flow fields. The method includes acquiring high-speed images captured by a high-speed camera; setting parameters for the high-speed camera; post-processing the high-speed images to obtain optimized images; the post-processing includes image denoising, grayscale segmentation, contour segmentation, and roundness denoising; constructing a PWC-Net network; using the PWC-Net network to extract features from the optimized images; and predicting the gas-liquid two-phase flow field based on the extracted features to obtain the final fluid velocity field. This invention enables rapid, accurate, and stable measurement of gas-liquid two-phase flow field velocity. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a system architecture diagram in this embodiment;

[0038] Figure 2 This is a diagram of the PWC-Net structure in this embodiment;

[0039] Figure 3 Here are the structure diagrams of the convolution module and the deconvolution module in this embodiment; wherein, (a) is the structure diagram of the convolution module; and (b) is the structure diagram of the deconvolution module.

[0040] Figure 4 A flowchart illustrating the configuration procedure for the high-speed camera in this embodiment;

[0041] Figure 5 This is a flowchart of the particle image velocimetry procedure in this embodiment. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] The purpose of this invention is to provide a method and system for real-time synchronous measurement of gas-liquid two-phase flow fields, aiming to solve or improve at least one of the above-mentioned technical problems.

[0044] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0045] This invention provides a method for real-time synchronous measurement of gas-liquid two-phase flow fields, comprising:

[0046] Step 100: Acquire a high-speed image captured by a high-speed camera; the high-speed camera uses set parameters.

[0047] Step 200: Post-process the high-speed image to obtain an optimized image; the post-processing includes image denoising, grayscale segmentation, contour segmentation, and roundness denoising.

[0048] Step 300: Construct the PWC-Net network.

[0049] Step 400: Use the PWC-Net network to extract features from the optimized image, and predict the gas-liquid two-phase flow field based on the extracted features to obtain the final fluid velocity field.

[0050] As a specific implementation method, the following specific processing procedure is provided.

[0051] First, a solution is proposed to address the existing problems. The overall concept is as follows:

[0052] Algorithmically, this approach introduces deep learning optical flow networks into the field of particle image velocimetry (PIV). By constructing a deep learning framework and optical flow estimation network architecture, features can be directly extracted from particle image pairs to predict the velocity field, thereby improving the accuracy and robustness of the calculated velocity field. The network structure fully considers the characteristics of the PIV task, such as combining the sparsity of particle images, noise interference, and the continuity constraints of the flow field to construct the learning dataset and train the model, thus ensuring accuracy while being closer to the field of fluid mechanics. The system introduces optical flow models with different structures, such as traditional optical flow networks that directly use convolutional neural networks and convolutional optical flow networks combined with multi-scale feature fusion mechanisms and attention mechanisms, to provide computational schemes with different accuracy levels and computational costs, offering different solutions for different task requirements.

[0053] Secondly, the specific technical solution includes the design of both hardware and software components.

[0054] Hardware components:

[0055] Traditional particle image velocimetry systems primarily employ a high-power laser source, a high-speed camera, a synchronization controller, an optical correction box, and a host computer as their hardware foundation. Specifically, as follows... Figure 1As shown, its core design integrates a transparent pipe and a correction box into the observation engineering system to achieve high-precision flow field measurement. Specifically, a laser source is mounted at the top of the pipe to generate a uniform and stable laser sheet light source to illuminate the tracer particles in the flow field. Two cameras are placed on either side of the transparent pipe to synchronously capture particle motion data within the pipe. These cameras are connected to a synchronization controller via data cables, transmitting the acquired image data to a host computer for real-time processing and analysis. Through this design, the system can efficiently acquire flow field information and, combined with deep learning algorithms, achieve high-precision, real-time particle image velocimetry, providing reliable technical support for research in complex flow field environments.

[0056] The laser source is one of the core components of the PIV system, used to illuminate tracer particles in the flow field. A high-speed camera captures images of the tracer particles' motion in the flow field; its performance directly affects the temporal and spatial resolution of the PIV system. A synchronization controller coordinates the timing of the laser source and camera, ensuring precise synchronization between the laser pulses and the camera exposure time. An optical correction chamber improves the accuracy of the flow field changes within the test tube by reducing optical distortion caused by factors such as pipe curvature and differences in refractive index between the fluid and the medium within the pipe. The computer is the control center of the PIV system, used to control the operation of hardware devices, acquire image data, and perform subsequent processing. The data processing system is typically equipped with high-performance computing hardware and specialized PIV software for image analysis, velocity field calculation, and result visualization.

[0057] Software component:

[0058] The software design consists of three core projects: an image post-processing program, a particle image velocimetry program, and a high-speed camera configuration program. Each project is developed according to specific functional requirements, which together form a complete high-speed image acquisition and processing system.

[0059] Image post-processing programs are mainly used to correct and optimize acquired image data, ensuring that the data is accurate and usable. Its core functions include image denoising, grayscale segmentation, contour segmentation, and circularity denoising.

[0060] The particle image velocimetry program is a key computational module of the system, used to perform real-time or offline particle image velocimetry analysis on data acquired by the high-speed camera. This program can read and process real-time data signals, and also perform post-processing analysis on offline data, ensuring flexible application in various scenarios. For example... Figure 5 As shown, the program is divided into two main modules: real-time PIV and offline PIV. Real-time PIV measures data by reading camera data, while offline PIV supports reading video or image data. All this data enters the PIV velocity measurement stage, generating three results: the vector field, energy field, and vorticity field of the PIV, which are then saved.

[0061] The high-speed camera configuration program allows users to set parameters and control the camera to ensure accurate image acquisition during data collection. The program features camera on / off control and real-time image preview, facilitating quick adjustments to shooting angles and fields of view. Flexible settings for parameters such as resolution, exposure time, acquisition frame rate, and number of frames saved adapt to various experimental needs. During real-time acquisition, the program dynamically adjusts camera settings to guarantee image data quality. Figure 4 As shown, the camera control module includes functions such as shooting preview, high-speed acquisition, dark field calculation, camera formatting, and video export, ensuring that shooting and data processing are performed as needed. The parameter setting module allows adjustment of parameters such as exposure time, shooting frame rate, resolution, and number of frames saved; these settings directly affect the quality of the captured image and the amount of data.

[0062] Based on the above two parts, the working principle and overall processing flow of the solution are provided.

[0063] The traditional PIV algorithm works by scattering tracer particles in the flow field, illuminating them with pulsed lasers, and recording images of the particles at different times using a camera. The images are then analyzed using either cross-correlation or optical flow methods. Cross-correlation divides the image into blocks, calculates correlations to determine particle displacement, but suffers from an "averaging effect" that reduces spatial resolution. Optical flow estimates the motion vector field as a global function, achieving pixel-level resolution, but is sensitive to noise and imaging conditions. While traditional algorithms are widely used in flow field measurements, they suffer from complex parameter settings and low computational efficiency. With technological advancements, deep learning-based PIV velocimetry algorithms are emerging to overcome these shortcomings.

[0064] The core idea of ​​the deep learning PIV algorithm is to use a convolutional neural network (CNN) to directly learn the mapping relationship of the velocity field from particle image pairs. Compared with traditional PIV methods, this algorithm avoids manually designing features and complex optimization processes, and can automatically extract motion information from images and generate high-precision velocity fields.

[0065] The core components of a CNN include convolutional layers, activation functions, pooling layers, and fully connected layers.

[0066] Convolutional layers extract local features from images using convolutional kernels (filters) to generate feature maps. Convolutional layers are the core component of Convolutional Neural Networks (CNNs) used to extract features from input data (such as images). Their core operation is convolution, which involves sliding a convolutional kernel (filter) across the input data and calculating a weighted sum of local regions. The following is the formula for a convolutional layer:

[0067]

[0068] Where O is the output feature, K is the convolution kernel, and I is the input feature.

[0069] Activation functions are key components in neural networks, used to introduce non-linearity, enabling neural networks to learn and represent complex patterns. Pooling layers reduce the dimensionality of feature maps, decreasing computational cost and enhancing feature robustness.

[0070] Fully connected layers map extracted features to the final output through fully connected outputs. A fully connected layer completely connects each neuron in the input data to each neuron in the output data, achieving a linear transformation through weight matrices and bias vectors, and is often combined with activation functions to introduce non-linearity. The following is the formula for a fully connected layer:

[0071] y = W·x + b

[0072] Where x is the input vector, W is the weight matrix, b is the matrix bias, and y is the output vector.

[0073] The workflow of a CNN typically includes input image, feature extraction, and classification output. The image is first processed through multiple convolutional and pooling layers to extract hierarchical features, which are then mapped to the classification result through fully connected layers. The advantages of CNNs lie in their local receptive field, parameter sharing, and hierarchical feature extraction, enabling them to efficiently capture local features of an image and reduce the number of model parameters.

[0074] PWC-Net is essentially based on a CNN architecture, but unlike traditional CNNs, it is specifically designed for optical flow estimation tasks, with its structure and functionality optimized around optical flow computation. The model works as follows:

[0075] 1. Feature Extraction

[0076] For the two input images I1 and I2, multi-scale features are extracted using a convolutional neural network. Let the feature extraction function of the l-th layer be f. l (·), then the extracted feature map Here, 'l' represents the pyramid level. These feature maps contain semantic and detailed information at different levels, providing a foundation for subsequent optical flow estimation.

[0077] 2. Image distortion

[0078] Based on the currently estimated optical flow w, the feature map F1 of the previous frame is deformed to better align it with the feature map F2 of the next frame. The deformation operation can be represented as W(F1, w), specifically implemented through bilinear interpolation. In two-dimensional space, for the coordinates (x, y) on feature map F1, the deformed coordinates are (x + w). x ,y+w y ), where w x and w yThese are the components of the optical flow w in the x and y directions, respectively. The mathematical expression is:

[0079] W(F1,w)(x,y)=∑ i,j F1(x+i,y+i)·max(0,1-|iw x |)·max(0,1-|jw y |)

[0080] 3. Cost volume

[0081] A cost volume is constructed to measure the similarity between the deformed feature maps W(F1,w) and F2. For each position (x,y) on the feature map and different optical flow offsets (δx,δy), a similarity metric is calculated. Common similarity metrics include the dot product.

[0082] C(x,y,δx,δy)=W(F1,w)(x,y)·F2(x+δx,y+δy)

[0083] Combining all similarity measures of position and optical flow offset forms the cost volume C, which is a four-dimensional tensor that records the degree of similarity under different positions and optical flow assumptions.

[0084] 4. Optical flow estimation and refinement

[0085] The cost volume C is processed by a convolutional neural network to predict the optical flow offset at each location, thus obtaining an initial optical flow estimate. To improve accuracy, a coarse-to-fine strategy is adopted. Initial estimations are performed at low-resolution pyramid levels, followed by upsampling and refining the results by incorporating information from high-resolution levels. At each scale, a loss function L (such as L1 or L2 loss) is used to supervise training and optimize model parameters, resulting in improved prediction of optical flow. As close as possible to real optical flow w gt Taking L1 loss as an example:

[0086] Where N is the number of samples, and w gt These are the predicted optical flow and the actual optical flow for the i-th sample, respectively. Finally, the optical flow estimation can be further refined using a context network to improve its accuracy and smoothness.

[0087] In the above description, the PWC-Net architecture diagram is as follows: Figure 2 As shown in Figure 3, the structure diagrams of the convolution module and deconvolution module in the algorithm are shown in Figure 3.

[0088] exist Figure 2The paper describes the main structure of PWC-Net, which achieves feature separation through convolutional layers. Then, each convolutional feature is upsampled and used in a deconvolutional layer to estimate the optical flow. Next, the upper-layer optical flow features are distorted and matched with the cost, then concatenated and deconvolved to obtain the estimated optical flow features. This process is repeated, and finally, a convolutional layer is used to further refine the optical flow and output the prediction.

[0089] exist Figure 3 The diagram describes how the convolutional module extracts features through two Conv 3×3 convolutional layers, introduces non-linearity using the ReLU activation function, and finally uses a Pool 2×2 pooling layer for downsampling to reduce the feature map size, decrease data dimensionality, and extract more abstract feature representations. Similarly, the deconvolutional module uses two Conv 3×3 convolutional layers and ReLU activation, followed by upsampling through an UpConv 2×2 deconvolutional layer to expand the feature map size and restore data resolution.

[0090] As a specific embodiment, to measure the experimental parameters of the gas-liquid two-phase flow field in a horizontal tube, melamine-formaldehyde resin particles with a diameter of 7 μm and better hydrophilicity, encapsulating rhodamine dye, were used as liquid-phase tracer particles (also known as fluorescent tracer particles). Dioctyl sebacate (DEHS) droplets with a diameter of 5-6 μm, processed by a high-pressure atomizer, were used as tracers in the gas phase.

[0091] Image Acquisition: A high-speed camera is mounted near the gas-liquid two-phase flow experimental setup. The camera's shooting angle and focal length are adjusted to ensure a complete and clear image of the movement region of the tracer particles in the flow field. During acquisition, the camera's exposure time and gain parameters can be adjusted in real time according to the dynamic changes in the flow field to obtain high-quality particle images.

[0092] Model Training: A large number of particle image sequences of gas-liquid two-phase flow fields under different operating conditions were collected and accurately labeled, including information such as the actual velocity vectors of the particles. The labeled image sequences were divided into training, validation, and test sets. The model was trained using the training set data, employing a stochastic gradient descent optimization algorithm. After a certain number of training iterations, the model was validated using the validation set data. The loss function value and accuracy were observed, and the model parameters were adjusted based on the validation results to prevent overfitting or underfitting. Once the model achieved satisfactory performance on the test set, training was stopped, and the trained model parameters were saved.

[0093] Image post-processing: After image denoising, grayscale segmentation, contour segmentation, and roundness denoising, the particle image data is saved and used for subsequent analysis.

[0094] Velocity field calculation: The acquired sequence of particle images of the gas-liquid two-phase flow field is input into a trained deep learning model. Based on the learned features and patterns, the model outputs the corresponding velocity field information. The calculation results can be converted into vector fields, energy fields, and vorticity gradient fields, and directly displayed through a visualization interface, allowing users to fully grasp the flow field characteristics.

[0095] Therefore, this embodiment has the following beneficial effects:

[0096] High precision: By learning and extracting features from a large number of complex gas-liquid two-phase flow field images through deep learning models, it can more accurately identify particle motion trajectories. Compared with traditional PIV technology, it significantly improves the accuracy of velocity field measurement and can accurately capture subtle velocity changes and complex flow pattern characteristics in the flow field.

[0097] High efficiency: The parallel computing capabilities and optimized algorithm structure of deep learning models enable velocity field calculations to be completed in a short time, meeting real-time measurement requirements and greatly improving measurement efficiency.

[0098] Strong robustness: Because the model has been trained with a large amount of diverse data, it has strong adaptability to disturbances such as noise, bubble deformation and liquid phase turbulence in the gas-liquid two-phase flow field. It can stably obtain accurate velocity field information under complex working conditions, thus improving the robustness of the system.

[0099] High versatility: This system is not only suitable for measuring gas-liquid two-phase flow fields, but with proper adjustment and training, it can also be widely used for measuring other types of flow fields, such as pure liquid flow fields and gas-solid two-phase flow fields. It has good versatility and can meet the flow field measurement needs of different fields.

[0100] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0101] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for real-time synchronous measurement of gas-liquid two-phase flow field, characterized in that, include: Acquire high-speed images captured by a high-speed camera; the high-speed camera uses set parameters; The high-speed image is post-processed to obtain an optimized image; The post-processing includes image denoising, grayscale segmentation, contour segmentation, and roundness denoising. Construct a PWC-Net network; The optimized image is feature extracted using the PWC-Net network, and the gas-liquid two-phase flow field is predicted based on the extracted features to obtain the final fluid velocity field. The construction process of the PWC-Net network is as follows: Obtain the training dataset; the training dataset includes training data and corresponding data labels; A pre-trained network is constructed based on a CNN network; the CNN network includes convolutional layers, activation functions, pooling layers, and fully connected layers; The training dataset is input into the pre-trained network, and training is performed with the goal of minimizing the loss between the network output and the data label. When the network meets the preset accuracy requirements, the trained network is determined as the PWC-Net network. The optimized image is feature extracted using the PWC-Net network, and the gas-liquid two-phase flow field is predicted based on the extracted features to obtain the final fluid velocity field, specifically including: The optimized image is input into the PWC-Net network, and multi-scale feature extraction is performed on two adjacent frames to obtain feature maps containing semantic and detail information at different levels. Based on the feature map, image deformation is performed, and the two frames of images are aligned to obtain the deformed feature map; A similarity metric is obtained for the similarity between each position on the deformable feature map and different optical flow offsets, and a cost volume is constructed based on the similarity metric. Based on the cost body, the gas-liquid two-phase flow field is predicted to obtain the final fluid velocity field.

2. The method for real-time synchronous measurement of gas-liquid two-phase flow field according to claim 1, characterized in that, The settings parameters include at least resolution, exposure time, acquisition frame rate, and number of frames to be saved.

3. The method for real-time synchronous measurement of gas-liquid two-phase flow field according to claim 1, characterized in that, The step of performing image deformation based on the feature map, aligning two frames of images to obtain a deformed feature map, specifically includes: The deformation operation is represented as W(F1,w). Bilinear interpolation is used to deform the feature map F1 of the previous frame, aligning it with the feature map F2 of the next frame. In two-dimensional space, for the coordinates (x,y) on feature map F1, the deformed coordinates are (x+w). x ,y+w y ), where w x and w y These are the components of the optical flow w in the x and y directions, respectively, and their mathematical expressions are: W(F1,w)(x,y)=∑ i,j F1(x+i,y+i)·max(0,1-|i-w x |)·max(0,1-|j-w y |) Where i and j represent the interpolation offset.

4. The method for real-time synchronous measurement of gas-liquid two-phase flow field according to claim 1, characterized in that, The similarity between each position on the deformed feature map and different optical flow offsets is used to obtain a similarity metric, and a cost body is constructed based on the similarity metric, specifically including: For each position (x, y) and different optical flow offsets (δx, δy) on the deformed feature map, calculate the similarity metric: C(x,y,δx,δy)=W(F1,w)(x,y)·F2(x+δx,y+δy) Where W(F1,w) represents the deformation operation, w represents the optical flow, F1 represents the feature map of the previous frame, and F2 represents the feature map of the next frame. The similarity measures of all positions and optical flow offsets are combined to form a cost volume C; the cost volume C is a four-dimensional tensor used to record the degree of similarity under different positions and optical flow assumptions.

5. The method for real-time synchronous measurement of gas-liquid two-phase flow field according to claim 1, characterized in that, The prediction of the gas-liquid two-phase flow field based on the cost volume to obtain the final fluid velocity field specifically includes: Based on the cost volume, a coarse-to-fine strategy is adopted, and the prediction results are upsampled sequentially according to the resolution from low to high to obtain the final fluid velocity field.

6. A real-time synchronous measurement system for gas-liquid two-phase flow fields, using the method as described in any one of claims 1-5, characterized in that, include: High-power laser source, high-speed camera, synchronization controller, optical correction box and host computer; Multiple high-speed cameras are arranged around the optical correction box, and the light emission end of the high-power laser light source is located on the top of the optical correction box. Each of the high-speed cameras and the high-power laser light source is connected to the synchronization controller, which is also connected to the host computer.

Images