CN113313000B - An intelligent identification method of gas-liquid two-phase flow based on optical image - Google Patents

Abstract

The invention provides an optical image-based intelligent identification method for gas-liquid two-phase flow, which includes preparing training data sets and test data sets; constructing a full convolution network model FCN; identifying gas-liquid two-phase flow according to the trained FCN full convolution network model Bubbles in the flow: For the FCN fully convolutional network model after training, input a picture of the gas-liquid two-phase flow to be identified for the model, the bubbles in the picture can be almost accurately identified through the network, and the accuracy of bubble identification can be calculated and obtained. Spend. Introducing the FCN method based on deep supervised learning and data extraction to gas-liquid two-phase flow recognition, it can automatically extract information from the pixel level through multi-layer convolution operations to extract abstract semantic concepts, using upsampling layers and multi-scale The fusion technology is used to further optimize the results, so that the high-level subnets fuse the features of the low-level subnets many times, maintaining a very high resolution, thereby improving the accuracy of bubble identification.

Description

An intelligent identification method of gas-liquid two-phase flow based on optical image

technical field

The invention relates to an intelligent identification method of gas-liquid two-phase flow based on optical images, in particular to a gas-liquid two-phase flow identification method based on an improved full convolution network model FCN, especially for the problem of bubble identification in liquid, belonging to The field of gas-liquid two-phase flow identification and classification.

Background technique

In nature and industrial processes, multiphase flow problems are often involved, among which gas-liquid two-phase flow is the most common. Based on the complex hydrodynamic characteristics of bubbles in many fields such as chemical engineering, biopharmaceuticals, geophysics, and wastewater management, it is necessary to better understand the interaction between two-phase flows through experimental studies. Basic experimental research supports the development of various engineering fields by establishing closed experimental models. Furthermore, detailed experimental results can be used to compare and validate the accuracy of these models. In the experiments, it is of paramount importance to measure parameters such as bubble size and shape, velocity, interfacial area concentration, and porosity to assess accuracy and to develop models.

Experimental techniques focused on the characterization of bubble parameters can be divided into two broad categories: probe-based invasive (contact) methods and non-invasive (non-contact) methods. Non-invasive methods are fundamentally different from probe-based methods in that they do not interfere with the flow under study, avoid most of the disadvantages of invasive-based methods, and therefore generally exhibit higher spatial resolution. Typical non-invasive methods are the use of laser Doppler anemometers and image processing techniques. In order to determine the parameters of the gas-liquid two-phase flow, such as the average velocity of the continuous phase, the local gas concentration, the flow properties and their fluctuations in the dispersed phase, some model-based feature extraction methods are needed to extract information about the exact location of the bubbles in space and its size information. These features not only can accurately capture the characteristics of bubbles in two-phase flow, but also play an important role in subsequent researches in areas such as bubble tracking. Specifically, image processing methods are used to identify different bubbles in a two-phase flow and calculate the relevant parameters. Traditional image processing methods identify air bubbles in images through a series of filtering and manipulation steps, such as image type conversion, image filtering, image denoising and enhancement, image filling, etc. After applying these image processing steps, the geometric features of the image are visible from the final image. Finally, edge detection and further required parameter operations are performed through the discriminator algorithm, and the specific geometric features of the image are output. However, in the traditional image processing method, since the feature selection is subjective and artificial, the feature extraction is not comprehensive, so the intelligent selection of image features is a problem to be solved in this field.

In recent years, methods based on deep learning can intelligently process images and have been widely used in image recognition, classification and processing. For example, image processing methods based on Convolutional Neural Networks (CNN) are favored by developers due to their robustness and versatility. Some authors have thus proposed the use of deep learning for bubble identification in gas-liquid two-phase flows, which can provide similar or better results than classical image processing methods. Due to the high complexity of underwater bubble flow data, the robustness of the model is poor, and most industrial experiments require extremely high bubble identification accuracy. Identifying problems, introducing an improved deep learning model for precise bubble-focused identification has high application value in this field.

SUMMARY OF THE INVENTION

In view of the above-mentioned prior art, the technical problem to be solved by the present invention is to provide an optical image-based gas-liquid two-phase flow intelligent identification method for bubble identification. A model that can identify bubbles in the flow with a high degree of accuracy.

The object of the present invention is achieved in this way:

Step 1: Prepare training dataset and test dataset. For our existing gas-liquid two-phase flow video, use Python to extract the picture of each frame in the video, and use Labelme to label the picture of each frame. The label of each image X contains two types of pixels. The pixel corresponding to the background is 0, and the pixel corresponding to the bubble is 1. Here, the ratio of background pixels and bubble pixels may not be balanced. Then, each image and its corresponding label are constructed into a Dataset dataset, and the preprocessing operations of reading, decoding, normalization and normalization are performed, and 80% of the images and their corresponding labels are randomly used as training data. set, and the remaining 20% of the images and their corresponding labels are used as the test dataset.

为卷积层的每一种输出的特征图，E为损失函数Dice_loss。其具体更新方式为Step 2: Build a fully convolutional network model FCN, first use the VGG16 network module for transfer learning, use the convolutional base part in the VGG16 network, remove the fully connected layer, and use the pre-trained weights on the ImageNet dataset for training . Since semantic segmentation needs to classify each pixel on the image, the process of convolution kernel pooling is a "down-sampling" process, which makes the length and width of the image smaller and smaller, which requires the use of deconvolution upsampling. The method restores the final output to the size of the original image, and then uses Sigmoid to activate and output a value to achieve classification. In the network model FCN parameter update,

is the feature map of each output of the convolutional layer, and E is the loss function Dice_loss. Its specific update method is

where M _j represents the selected combination of input feature maps, k _ij is the convolution kernel used for the connection between the input i-th feature map and the output j-th feature map, and b _j is the bias corresponding to the j-th feature map setting, sensitivity

是

在做卷积时，与k_ij做卷积的每一个patch。(u,v) represents the element position in the sensitivity matrix,

Yes

When doing convolution, each patch that is convolved with k _ij .

Step 3: Train the constructed FCN network model, including

Step 3.1: Initialization parameters: batch_size of input training data, number of training iterations epoch, hyperparameter γ (learning rate), buffer_size (buffer size);

Step 3.2: Set the loss function Dice_loss: Consider one sample, then the loss function of the nth sample is

表示第n个样本的label tⁿ的第k维，

是第n个样本网络的输出(predict label)的第k维。Among them, c is the dimension of label, for classification problems, it means that these samples can be divided into c categories.

represents the kth dimension of the label t ⁿ of the nth sample,

is the kth dimension of the output (predict label) of the nth sample network.

是对m_t、n_t的校正，θ_t+1为更新后的参数，其具体更新方式为The optimizer uses the Adam optimization algorithm to update the weight parameters of the network, where m _t and n _t are the first and second moment estimates of the gradient, respectively,

is the correction of m _t and n _t , θ _t+1 is the updated parameter, and the specific update method is as follows

m _t =β ₁ m _t-1 +(1-β ₁ )g _t

Among them, the default values of η, β ₁ , β ₂ are η=0.001, β ₁ =0.9, β ₂ =0.999, ε=10 ⁻⁸ , β ₁ and β ₂ are numbers close to 1, ε is to prevent dividing by 0, _gt represents the gradient.

Step 3.3: Calculate and output the precision and recall of the network, and calculate the accuracy of bubble recognition separately. The specific expression of precision is

The specific expression of recall rate Recall is

Among them, TP is an example that the classifier considers to be a positive sample and is indeed a positive sample, FP is an example that the classifier considers to be a positive sample but is not actually a positive sample, and FN is an example that the classifier considers to be a negative sample but is not actually a negative sample.

Step 4: Identify the bubbles in the gas-liquid two-phase flow according to the trained FCN full convolution network model: For the FCN full convolution network model after training, input a picture of the gas-liquid two-phase flow to be identified for the model, Through the network, the bubbles in the picture can be almost accurately identified, and the accuracy of bubble identification can be calculated and obtained.

Compared with the prior art, the beneficial effects of the present invention are:

Aiming at the identification problem of bubbles in liquid based on optical images, the invention introduces an improved FCN full convolution network, and establishes an intelligent identification model of gas-liquid two-phase flow based on optical images. The advantages of this method are: (1) For the problem that it is difficult to label a large number of bubble training samples, we use Labelme for data intelligent labeling, which can remove the noise in the original image and divide the background and bubbles into two categories to facilitate network training. . (2) Using VGG16 for transfer learning, making full use of the similarity between models, improving the stability, generalizability and learning efficiency of the model. (3) Due to the uneven distribution of the two types of pixels in the acquired image data, the bubbles and the background, when the binary cross entropy loss function is used, the training results of the model will be more biased towards the type with more pixels in the image, so Dice_loss is used. training as a loss function. (4) The FCN method based on deep supervised learning and data extraction is introduced into the gas-liquid two-phase flow recognition, which can automatically extract information from the pixel level through multi-layer convolution operations to extract abstract semantic concepts. (5) Upsampling layers and multi-scale fusion techniques are used to further optimize the results, so that the high-level subnetworks fuse the features of the low-level subnetworks multiple times, maintaining extremely high resolution, thereby improving the accuracy of bubble recognition.

Description of drawings

Fig. 1 is the FCN full convolutional network framework schematic diagram that the present invention uses VGG16 to carry out migration learning improved;

Fig. 2 is the data labeling that the present invention uses Labelme to carry out for gas-liquid two-phase flow data;

3(a) to 3(c) are the convergence results, precision and recall results of the present invention using the improved FCN model for gas-liquid two-phase flow data;

Fig. 4 is the individual bubble identification accuracy that the present invention uses the improved FCN model for gas-liquid two-phase flow data;

FIG. 5 is the identification result of the present invention using the improved FCN model for gas-liquid two-phase flow data.

Detailed ways

The present invention will be described in further detail below with reference to the accompanying drawings and specific embodiments.

Based on the full convolution network model FCN in the semantic segmentation algorithm, the invention proposes an optical image-based intelligent identification technology for gas-liquid two-phase flow for the image sequence data of gas-liquid two-phase flow. In terms of model structure, the classic network VGG16 is first introduced for transfer learning, the fully connected layer is removed, the convolutional base part in the network is retained, and the weights pre-trained on the ImageNet dataset are used for training. Next, the core ideas of semantic segmentation - deconvolution upsampling and skip structure are introduced. Since the network needs to classify each pixel on the image, the process of convolution kernel pooling is a "down-sampling" process, which makes the length and width of the image smaller and smaller, and "deconvolution" is the inverse process of convolution , so that the length and width of the image can be restored, and the high-resolution results are preserved; then a jump fusion structure is introduced to make up for the lost features of the previous convolutional layers and pooling layers. Combined with the jump-level structure of the results of different depth layers, the The results are upsampled and combined to optimize the output; the method ensures robustness and accuracy, inpainting the restored image. On the loss function, using the binary cross entropy loss function, the training results of the model will be more biased towards the class with more pixels in the image, so Dice_loss is used as the loss function for training. The above improvements in model structure and loss function can make the model converge faster and the data quality is higher. The improved FCN model is introduced to train the gas-liquid two-phase flow image sequence data set, which can obtain accurate identification of air bubbles and improve the accuracy of the gas-liquid two-phase flow identification model.

Example

A gas-liquid two-phase flow intelligent identification method based on an improved fully convolutional network FCN optical image proposed by the present invention includes the following steps:

Step 1: Use the video source of the gas-liquid two-phase flow as the original data, use Python to extract the picture of each frame in the video, and use Labelme to label the picture of each frame. The label of each image X contains two types of pixels, the pixel corresponding to the background is 0, and the pixel corresponding to the bubble is 1, where the ratio of background pixels and bubble pixels is not balanced. Then build each image and its corresponding label into a Dataset dataset, and perform preprocessing of reading, decoding, normalization and normalization, and use 80% of the images and their corresponding labels as the training dataset, and the rest 20% of the images and their corresponding labels are used as the test dataset.

Step 2: First, based on the fully convolutional network FCN model, we build and use the fully convolutional part of the VGG16 network module as the convolution base for transfer learning, remove the fully connected layer, and use the pre-trained image on the ImageNet dataset. weights for training. Since semantic segmentation needs to classify each pixel on the image, the process of convolution kernel pooling is a "down-sampling" process, which makes the length and width of the image smaller and smaller, and the deconvolution up-sampling method needs to be used. The final output is upsampled to the size of the original image, and then activated using Sigmoid and outputting a value to achieve classification. In the parameter update of the improved FCN network model, x _j is the feature map of each output of the convolutional layer, and E is the loss function Dice_loss.

Its specific update method is

where M _j represents the selected combination of input feature maps, k _ij is the convolution kernel used for the connection between the input i-th feature map and the output j-th feature map, and b _j is the bias corresponding to the j-th feature map setting, sensitivity

是

在做卷积时，与k_ij做卷积的每一个patch。(u,v) represents the element position in the sensitivity matrix,

Yes

When doing convolution, each patch that is convolved with k _ij .

Step 3: Calculate the loss function. If the binary cross-entropy loss function BinaryCrossentropy is used, when the number of bubble pixels in each image in the dataset is too small and the background pixels are too many, the training results will be biased towards the background category, so that the network cannot perform well on the bubbles. identify. Therefore, here we use the loss function Dice Loss for medical image segmentation, whose expression is

表示第n个样本的label tⁿ的第k维，

是第n个样本网络输出(predict label)的第k维。In the above formula, c is the dimension of label. For classification problems, it means that these samples can be divided into c categories.

represents the kth dimension of the label t ⁿ of the nth sample,

is the kth dimension of the nth sample network output (predict label).

According to the above formula, if the model is fully fitted, the value of the loss function will be infinitely close to 0. During the training process, it can be judged whether the model has converged according to the value of the loss function. If the value of the Dice Loss loss function no longer decreases, the network model Convergence, the training is complete.

Step 4: Train the improved FCN fully convolutional network model. As shown in Figure 1, first, the input network of the gas-liquid two-phase flow picture and the picture marked with Labelme is extracted by convolution, and the length, width and channel of the picture are changed by pooling; The network performs feature fusion with the previous low-level subnet and restores the final output to the size of the original image, and combines these results to optimize the output while ensuring robustness and accuracy; then calculate the loss function of the network and optimize it according to Adam The algorithm updates the weight parameters, and iteratively trains the parameters of the network according to the above method, until the loss function value no longer decreases or remains stable, the network model converges, and the training is completed. Input the gas-liquid two-phase flow to be recognized pictures, and use the network model to generate semantic The result after segmentation.

Step 5: After the training is completed, input a picture of the gas-liquid two-phase flow to be recognized, in which each picture X contains two types of pixels, the pixel corresponding to the background is 0, the pixel corresponding to the bubble is 1, and the background pixel and The proportion of bubble pixels is not balanced. The network can almost accurately identify the bubbles in the picture, automatically generate the results of semantic segmentation, and calculate the final accuracy of bubble recognition separately.

Combined with the specific parameter example, the data in this example comes from a two-phase flow experiment performed by simulating a closed loop. By adjusting the pipeline valve, the gas-liquid two-phase flow can be controlled to rise vertically and fall vertically. For the vertical descending two-phase flow, the water pump The deionized water in the driving water tank enters the bubble generator and the vertical test section respectively through the filter and the rear, and returns to the water tank after being separated from the gas and water by the gas-water separator. The integrated air compressor generates compressed air and stores it in a compressed air tank with a volume of 300L. The compressed air enters the bubble generator through the solenoid valve, and then enters the experimental section and the gas-water separator, and finally discharges into the atmosphere. The total length of the experimental section is about 3.7m, which is composed of plexiglass tube sections of different lengths with an inner diameter of 50.8mm and the connection of the acquisition window. The flat design enables high-speed photography. By adjusting the relative positions of the plexiglass tube section and the acquisition window, two-phase flow parameters can be measured at different positions along the flow axis. The two-phase flow picture data set is obtained by intercepting each frame of the video. We use it as the training set for network model training, and normalize and standardize the data before training.

Analysis of the bubble identification results of the two-phase flow experiment simulating a closed loop:

The experimental data set is 9351 gas-liquid two-phase flow pictures intercepted by the video source obtained from the simulated closed-loop two-phase flow experiment. We randomly selected 80% of them as the training set and the remaining 20% as the test set. According to the improved FCN network constructed The model and the training method are trained. Table 1 shows the results obtained by using the model when we configure the network parameters, namely the number of training sets and test sets, loss function value, precision, recall rate, Test the accuracy of bubble recognition on 9351 images, these 4 indicators that measure the fit of the network. The lower the loss function value, the better the classification effect. When the accuracy is higher, it means that the classifier considers it to be a positive class, and the part that is indeed a positive class accounts for a higher proportion of all classifiers that are considered to be a positive class. When the recall rate is higher, it means that the classifier thinks that it is a positive class, and the proportion of the part that is indeed a positive class in all positive classes is higher. Figure 3(a) to Figure 3(c) are the convergence value curve, precision curve, and recall rate curve of the model. It can be seen that with the increase of the number of iterations, the convergence and stability of the model are significantly improved. improvement.

Based on the improved fully convolutional network FCN model, 7481 gas-liquid two-phase flow pictures are used as the training data set, and 1870 gas-liquid two-phase flow pictures are used as the test data set. We input the training data set and test data set into the network in training. Figure 4 is the final accuracy curve of the calculated individual bubbles identified in the network, from which it can be seen that the accuracy can reach 98.37%, and the computational complexity is low and the amount of computation is small. Figure 5 shows the result of the network identifying bubbles in the gas-liquid two-phase flow picture. From the test results and the picture prediction results, we can all draw the conclusion that the network has extremely high performance for gas-liquid two-phase flow identification, especially for bubble identification. Accuracy.

Table 1 The final output result of the training set and test set data of the present invention

Claims (1)

1. An intelligent identification method of gas-liquid two-phase flow based on optical images is characterized by comprising the following steps:

step 1: preparing a training data set and a testing data set: extracting a picture of each frame in a gas-liquid two-phase flow video by using Python, performing data labeling on the picture of each frame by using Labelme, wherein a label of each picture X comprises two pixel types, a pixel corresponding to a background is 0, a pixel corresponding to a bubble is 1, then constructing each picture and a label corresponding to the picture into a Dataset, performing reading, decoding, normalization and standardization preprocessing operations, randomly taking 80% of the pictures and the labels corresponding to the pictures as training datasets, and taking the rest 20% of the pictures and the labels corresponding to the pictures as testing datasets;

step 2: constructing a full convolution network model FCN, firstly carrying out migration learning by using a VGG16 network module, using a convolution base part in a VGG16 network, removing a full connection layer, using weights pre-trained on an ImageNet data set for training, wherein each pixel on an image needs to be classified due to semantic segmentation, and a process of convolution kernel pooling is a process of down-sampling, so that the length and the width of the image are smaller and smaller, the finally obtained output needs to be restored to the size of an original image by using a method of deconvolution up-sampling, then using Sigmoid for activation and outputting a value, thereby realizing classification, in updating parameters of the network model FCN,

for each output profile of the convolutional layer, E is the loss function Dice _ loss, which is updated specifically in the manner of

Wherein M is_jInput feature map combination, k, representing selection_ijIs a volume for connecting the input ith characteristic diagram and the output jth characteristic diagramBuild-up of nuclei, b_jIs the bias and sensitivity corresponding to the j-th characteristic diagram

(u, v) represents the element positions in the sensitivity matrix,

is that

When making convolution, with k_ijMaking each patch of the convolution;

and step 3: training the constructed FCN network model specifically comprises

Step 3.1: initializing parameters: inputting batch _ size of training data, training iteration times epoch, hyper-parameter gamma learning rate and buffer _ size buffer area size;

step 3.2: setting a loss function Dice _ loss: considering a sample, the loss function of the nth sample is

Where c is the dimension of label, for the classification problem, it means that these samples can be classified into c classes,

label t representing the nth sampleⁿThe (c) th dimension of (a),

is the kth dimension of the output (predict label) of the nth sample network;

the optimizer updates the weight parameters of the network using an Adam optimization algorithm, where m_t、n_tRespectively a first moment estimate and a second moment estimate of the gradient,

is to m_t、n_tCorrection of theta_t+1The updated parameters are specifically updated in the following way

m_t＝β₁m_t-1+(1-β₁)g_t

Wherein, eta, beta₁，β₂Default is η ═ 0.001, β₁＝0.9，β₂＝0.999，ε＝10^-8，β₁And beta₂Are all numbers close to 1, ε is to prevent division by 0, g_tRepresents a gradient;

step 3.3: calculating and outputting Precision, Recall, of the network, embodied as Precision, and calculating the accuracy of bubble identification separately

The Recall rate Recall is specifically expressed as

Where TP is an example where the classifier considers positive samples and is indeed positive samples, FP is an example where the classifier considers positive samples but is not actually positive samples, and FN is an example where the classifier considers negative samples but is not actually negative samples;

and 4, step 4: identifying bubbles in the gas-liquid two-phase flow according to the trained FCN full convolution network model: for the FCN full convolution network model after training, inputting a picture to be identified of gas-liquid two-phase flow for the model, identifying bubbles in the picture almost accurately through a network, and calculating and obtaining the accuracy of bubble identification.

Images