Gas-liquid two-phase flow pattern identification and water holding rate prediction method based on microstrip antenna and deep neural network

Abstract

According to the gas-liquid two-phase flow pattern identification and water holding rate prediction method based on the microstrip antenna and the deep neural network, a measurement system adopts a microwave transmission method principle, the amplitude is attenuated when microwaves are transmitted in the gas-liquid two-phase flow, an amplitude attenuation signal can reflect the change of the dielectric property of the fluid, and the flow pattern identification and water holding rate prediction of the gas-liquid two-phase flow are realized. The microstrip antenna microwave sensor system comprises a measurement pipeline, a microstrip antenna microwave transmission sensor, a power divider, an amplitude phase detector, a data acquisition module and an upper computer. When microwaves are transmitted in the gas-liquid two-phase flow, the amplitude is attenuated, an amplitude attenuation signal can reflect the change of the dielectric property of the fluid, an intelligent model based on a deep neural network is provided based on amplitude attenuation data measured by a microwave sensor, and the identification of the flow pattern and the measurement of the water holding rate based on the flow pattern are realized.

Description

Technical Field

[0001] The present invention relates to a method for flow pattern recognition and water holdup prediction in a gas-liquid two-phase flow process. The method can more efficiently and fully utilize sensor data through a deep network, more accurately identify flow patterns, and predict liquid phase holdup than traditional methods. Background Art

[0002] Gas-liquid two-phase flow is widely used in industrial fields such as petroleum, chemical industry, and energy. The accurate prediction of its water holdup is of great significance for optimizing production and improving economic benefits. In actual industrial processes, the flow pattern structure of gas-liquid two-phase flow is complex, and the gas and liquid phases interact and act in the measurement pipeline, which leads to great challenges in establishing its water holdup model and accurately predicting its water holdup. A variety of measurement methods have been proposed for the measurement of gas-liquid two-phase flow, including acoustic method, optical method, conductivity method and microwave method. These methods have their own scope of application. The acoustic method has simple equipment and low cost, but the propagation characteristics of sound waves are affected by many factors, which brings difficulties to the measurement. The optical method can realize multi-parameter measurement, but it is expensive and easily affected by the color and transparency of the measured material. The conductivity method is based on the conductive properties of the medium for measurement. Its sensor electrode must be in contact with the fluid, and its sensitivity is limited by the measurement method. The microwave method has good application prospects in the measurement of gas-liquid two-phase flow due to its high sensitivity, fast response and non-invasiveness. The microwave transmission method measures the amplitude attenuation and phase shift of microwaves after penetrating the medium to measure parameters such as liquid phase holdup. Since the liquid content varies greatly under different flow patterns, and different liquid contents absorb microwave energy to different degrees, the amplitude attenuation of microwaves can reflect information about different flow patterns and different water holdups. By optimizing the design of a deep neural network model based on CNN-LSTM, the amplitude attenuation signal of the microwave sensor can be used to quickly and accurately predict the corresponding flow pattern, and then the corresponding water holdup prediction model is input according to different flow patterns, ultimately achieving accurate prediction of water holdup based on flow pattern. Summary of the invention

[0003] In order to solve the above problems, the technical solution adopted by the present invention is: a microwave sensor system based on a microstrip antenna, comprising:

[0004] Signal source: used to generate microwave signals;

[0005] Power divider: used for dividing the microwave signal generated by the signal source into two identical signals: signal I and signal II;

[0006] Microstrip antenna microwave transmission sensor: used to receive the signal I transmitted by the power divider, and transmit the signal I after amplitude attenuation and phase shift reflecting the change of dielectric properties of gas-liquid two-phase fluid;

[0007] Amplitude phase detector: used to receive the signal II transmitted by the power divider and the signal I after amplitude attenuation and phase shift transmitted by the microstrip antenna microwave transmission sensor, and demodulate the amplitude attenuation and phase shift of the microwave after passing through the fluid by comparing the signal I and the signal II;

[0008] Data acquisition module: used to collect the amplitude attenuation and phase shift signals of the microwaves transmitted by the amplitude phase detector after passing through the fluid.

[0009] Further: the microstrip antenna microwave transmission sensor comprises

[0010] Transmitter: used for receiving the signal I transmitted by the power divider;

[0011] Receiving end: used to receive the signal I after amplitude attenuation and phase shift transmitted by the receiving end of the microstrip antenna microwave transmission sensor, reflecting the change of dielectric properties of gas-liquid two-phase fluid;

[0012] The transmitting end and the receiving end are symmetrically arranged on the outer wall of a measuring pipe in which a gas-liquid two-phase fluid is transported.

[0013] Furthermore: the receiving end and the transmitting end have the same structure;

[0014] The receiving end includes a dielectric board;

[0015] A radiation patch is arranged on the upper surface of the dielectric plate;

[0016] A feeding point is provided on the radiation patch;

[0017] Radiation gaps are provided on both sides of the radiation patch in the length direction;

[0018] The lower surface of the dielectric plate is a ground plane;

[0019] A coaxial connector is arranged below the radiation patch.

[0020] Furthermore: the length L1 of the radiation patch is 28.5 mm, and W1 is 38 mm.

[0021] A gas-liquid two-phase flow pattern recognition and water holdup prediction method based on a deep neural network comprises the following steps:

[0022] Obtain the original gas-liquid two-phase fluid amplitude attenuation time series data;

[0023] Preprocessing the original gas-liquid two-phase fluid amplitude attenuation time series data to obtain enhanced gas-liquid two-phase fluid amplitude attenuation time series data;

[0024] Based on the original gas-liquid two-phase fluid amplitude attenuation time series data, a gas-liquid two-phase fluid flow morphology recognition model is constructed;

[0025] Based on the trained liquid two-phase fluid flow pattern recognition model, it is identified that there are three flow patterns of gas-liquid two-phase flow, namely slug flow, mixed flow and bubbly flow.

[0026] Based on the enhanced gas-liquid two-phase fluid amplitude attenuation time series data, a water holdup prediction model for each flow pattern is constructed;

[0027] Based on the trained water holdup prediction model for each flow type, the water holdup corresponding to each flow type of slug flow, mixed flow and bubbly flow is finally predicted.

[0028] Furthermore: the preprocessing adopts a sliding segmentation method.

[0029] Further: the process of identifying the existence of three flow forms of gas-liquid two-phase flow, namely slug flow, mixed flow and bubbly flow, based on the trained gas-liquid two-phase fluid flow form recognition model is as follows:

[0030] The original gas-liquid two-phase fluid amplitude attenuation time series data is first processed by Z-Score standardization to eliminate the scale effect between the data, and then the standardized data is input into the 1DCNN. 1DCNN roughly extracts the features of the input sequence, and integrates the effective information of the input sequence into the features through 3 convolutional layers and 3 pooling layers. The data processed by the 1DCNN module is input into the LSTM layer. The LSTM layer perceives the long-term dependency of the feature sequence. The features extracted by CNN-LSTM are first integrated by the flattening layer and then input into the DENSE layer, and finally the specific flow type of the gas-liquid two-phase flow is output.

[0031] Further: the process of finally predicting the water holdup corresponding to each flow type of slug flow, mixed flow and bubbly flow based on the trained water holdup prediction model for each flow type is as follows:

[0032] The microwave amplitude attenuation sequence signal is first preprocessed, and then enters different CNN-LSTM frameworks to extract their respective features according to their flow patterns. The extracted features are processed by the flat layer and then enter the DENSE layer, and finally the corresponding water holdup is output.

[0033] Further: the flow pattern recognition model includes a CNN-LSTM model;

[0034] In the CNN-LSTM model, the number of convolution kernels in convolution layer I is 16, the convolution kernel size is 2×1, the activation function is RELU, the size of pooling layer I is 2×1, the number of convolution kernels in convolution layer II is 32, the convolution kernel size is 3×1, the activation function is RELU, the size of pooling layer II is 2×1, the number of convolution kernels in convolution layer III is 64, the convolution kernel size is 4×1, the activation function is RELU, and the size of pooling layer III is 2×1.

[0035] Further: the water holdup prediction model for each flow type includes a bubbly flow water holdup prediction model, a mixed flow water holdup prediction model and a slug flow water holdup prediction model;

[0036] In the CNN-LSTM model of the bubbly flow holdup prediction model, the number of convolution kernels in convolution layer I is 16, the convolution kernel size is 2×1, the activation function is RELU, the maximum pooling layer I size is 2×1, the number of convolution kernels in convolution layer II is 32, the convolution kernel size is 3×1, the activation function is RELU, the maximum pooling layer II size is 2×1, the number of convolution kernels in convolution layer III is 64, the convolution kernel size is 4×1, the activation function is RELU, the maximum pooling layer III size is 2×1, the number of convolution kernels in convolution layer IV is 128, the convolution kernel size is 5×1, the activation function is RELU, and the maximum pooling layer IV size is 2×1;

[0037] In the CNN-LSTM model of the mixed flow holdup prediction model, the number of convolution kernels in convolution layer I is 16, the convolution kernel size is 2×1, the activation function is RELU, the number of convolution kernels in convolution layer II is 16, the convolution kernel size is 2×1, the activation function is RELU, and the average pooling layer size is 2×1, the number of convolution kernels in convolution layer III is 32, the convolution kernel size is 3×1, and the activation function is RELU, the number of convolution kernels in convolution layer IV is 32, the convolution kernel size is 3×1, the activation function is RELU, and the average pooling layer size is 2×1. The size of convolution layer II is 2×1, the number of convolution kernels of convolution layer V is 64, the convolution kernel size is 4×1, and the activation function is RELU. The number of convolution kernels of convolution layer VI is 64, the convolution kernel size is 4×1, and the activation function is RELU. The size of maximum pooling layer III is 2×1, the number of convolution kernels of convolution layer VII is 128, the convolution kernel size is 5×1, and the activation function is RELU. The number of convolution kernels of convolution layer VIII is 128, the convolution kernel size is 5×1, and the activation function is RELU. The size of maximum pooling layer IV is 2×1;

[0038] In the CNN-LSTM model of the slug flow water holding capacity prediction model, the number of convolution kernels in convolution layer I is 16, the convolution kernel size is 2×1, and the activation function is RELU. The size of the maximum pooling layer I is 2×1, the number of convolution kernels in convolution layer II is 32, the convolution kernel size is 3×1, and the activation function is RELU. The size of the maximum pooling layer II is 2×1, the number of convolution kernels in convolution layer III is 64, the convolution kernel size is 4×1, and the activation function is RELU. The size of the maximum pooling layer III is 2×1, the number of convolution kernels in convolution layer IV is 128, the convolution kernel size is 5×1, and the activation function is RELU. The size of the maximum pooling layer IV is 2×1, the number of convolution kernels in convolution layer V is 256, the convolution kernel size is 6×1, the activation function is RELU, and the size of the maximum pooling layer V is 2×1.

[0039] The present invention provides a method for identifying flow patterns and predicting water holdup of a gas-liquid two-phase flow based on a microstrip antenna and a deep neural network. Based on the amplitude attenuation data measured by a microwave sensor system, an intelligent model based on a deep neural network is proposed to realize the identification of flow patterns and the measurement of water holdup based on flow patterns. The microwave sensor adopts the principle of microwave transmission method. The microwave sensor measurement system includes a measuring pipeline, two identical and symmetrically installed microstrip antennas, a microwave signal generator, a power divider, an amplitude phase detector, a data acquisition system and a host computer. When microwaves are transmitted in a gas-liquid two-phase flow, the amplitude will attenuate, and the amplitude attenuation signal can reflect the changes in the dielectric properties of the fluid. Based on the amplitude attenuation data measured by the microwave sensor, an intelligent model based on a deep neural network is proposed to realize the identification of flow patterns and the measurement of water holdup based on flow patterns.

[0040] The present invention has the following advantages due to the adoption of the above technical solution:

[0041] (1) The measurement system of the present invention is based on the relevant principles of microwave transmission method and has many advantages such as high sensitivity, good stability, fast response speed, and non-contact measurement.

[0042] (2) The modeling method of the present invention is based on a deep neural network, which solves the problem of difficulty in establishing a water holdup model. The method has the ability to simultaneously identify flow patterns and predict water holdup online, has high measurement accuracy, and has a wide range of applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0044] Figure 1 This is the structure diagram of the microstrip antenna microwave sensor measurement system;

[0045] Figure 2 It is a microwave transmission sensor;

[0046] Figure 3 The receiving end structure of the microwave transmission sensor; (a) the configuration diagram of the rectangular microstrip antenna, (b) the side view of the rectangular microstrip antenna;

[0047] Figure 4 Flow pattern recognition model architecture diagram;

[0048] Figure 5 Flow pattern identification results;

[0049] Figure 6 Water holding rate prediction model architecture diagram;

[0050] Figure 7 Water holding rate prediction results.

[0051] Description of Figure Numbers:

[0052] 1. Radiation gap; 2. Feed point; 3. Radiation patch; 4. Ground plane; 5. Dielectric board; 6. Coaxial connector; 7. Measurement pipeline; 8. Transmitter; 9. Receiver; 10. Signal source; 11. Power divider; 12. Amplitude and phase detector; 13. Data acquisition module; 14. Host computer. DETAILED DESCRIPTION

[0053] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0054] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. The following description of at least one exemplary embodiment is actually only illustrative and is by no means intended to limit the present invention and its application or use. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] Figure 1 This is the structure diagram of the microstrip antenna microwave sensor measurement system;

[0056] A microwave sensor system based on a microstrip antenna adopts the principle of microwave transmission method. When microwaves are transmitted in a gas-liquid two-phase flow, the amplitude will be attenuated, and the amplitude attenuation signal can reflect the change of the dielectric properties of the fluid. The microstrip antenna microwave sensor system includes: a signal source 10, a measuring pipeline 7, a microstrip antenna microwave transmission sensor, a power divider 11, an amplitude phase detector 12, a data acquisition module 13 and a host computer 14;

[0057] Signal source 10: used to generate microwave signals;

[0058] A power divider 11 is used to divide the microwave signal generated by the signal source 10 into two identical signals: a signal I and a signal II;

[0059] Microstrip antenna microwave transmission sensor: used to receive the signal I transmitted by the power divider 11, and transmit the signal I after amplitude attenuation and phase shift reflecting the change of dielectric properties of gas-liquid two-phase fluid,

[0060] Amplitude phase detector 12: used to receive signal II transmitted by the power divider and signal I after amplitude attenuation and phase shift transmitted by the microstrip antenna microwave transmission sensor, and demodulate the amplitude attenuation and phase shift of microwaves after passing through the fluid by comparing signal I and signal II;

[0061] The amplitude phase detector 12 uses the AD8302 chip as an amplitude attenuation and phase shift detector, which can provide accurate amplitude attenuation and phase shift measurement. According to the AD8302 chip data sheet, for amplitude attenuation A (dB) and phase shift θ (degrees), the output of the circuit satisfies the following relationship:

[0062] V mag (V)=0.03V / dB×A(dB)+0.9V

[0063] V phs (V)=-0.010V / degree×(|0|-90)(degree)+0.9V

[0064] Data acquisition module 13: used to collect the amplitude attenuation and phase shift signals of the microwaves transmitted by the amplitude phase detector after passing through the fluid.

[0065] The host computer 14 is used to receive the amplitude attenuation and phase shift signals of the microwaves transmitted by the data acquisition module 13 after passing through the fluid, and realize the gas-liquid two-phase flow pattern recognition and water holdup prediction based on the deep neural network.

[0066] Figure 2 It is a microwave transmission sensor;

[0067] The microstrip antenna microwave transmission sensor comprises

[0068] Transmitter 8: used for receiving a signal I transmitted by the power divider;

[0069] Receiving end 9: used to receive the signal I after amplitude attenuation and phase shift transmitted by the receiving end of the microstrip antenna microwave transmission sensor, reflecting the change of dielectric properties of gas-liquid two-phase fluid;

[0070] The transmitting end 8 and the receiving end 9 are symmetrically arranged on the outer wall of the measuring pipe in which the gas-liquid two-phase fluid is transported. t 20mm, outer diameter D t It is 30mm.

[0071] Figure 3 The receiving end structure of the microwave transmission sensor; (a) the configuration diagram of the rectangular microstrip antenna, (b) the side view of the rectangular microstrip antenna;

[0072] The receiving end 9 has the same structure as the transmitting end 8;

[0073] The receiving end 9 includes a dielectric plate; the dielectric plate 5 is a FR4 dielectric plate, the length L of the FR4 dielectric plate is 57mm, the width W is 76mm, the thickness h and the relative dielectric constant ε r 1.6mm and 4.4 respectively.

[0074] A radiation patch is arranged on the upper surface of the dielectric plate 5 ; the length L1 of the radiation patch is 28.5 mm, and the width W1 is 38 mm.

[0075] A feeding point 2 is provided on the radiation patch 3;

[0076] Radiation gaps 1 are provided on both sides of the radiation patch 3 in the length direction;

[0077] The lower surface of the dielectric plate 5 is a ground plane 4;

[0078] A coaxial connector 6 is arranged below the radiation patch 3 .

[0079] A very thin rectangular radiating patch is placed on a rectangular substrate with a ground plane on the back side. The width and length of the patch are W1 and L1 respectively. The distance between the feed point and point O is L1 and the dimension of the patch along its length extends a distance ΔL at each end. The thickness, width, length and dielectric constant of the substrate are t, W, L and εr respectively. A coaxial probe is used to feed the microstrip antenna. The inner conductor of the coaxial cable is connected to the radiating patch while the outer conductor is connected to the ground plane. The dielectric board, feed point, radiating slot, ground plane, and coaxial connector are the parts that make up the rectangular microstrip antenna.

[0080] The microwaves are fed into the antenna of the transmitting end 8 by coaxial feeding. The amplitude of the microwaves emitted by the transmitting end antenna will be attenuated when transmitted in the gas-liquid two-phase flow. The signal is received by the receiving antenna. The amplitude attenuation signal can reflect the change of the dielectric properties of the fluid.

[0081] The microwave generated by the signal source of the microwave sensor system is divided into two identical signals through a power divider 11, one signal is connected to the transmitting end 8 of the microstrip antenna microwave sensor, and the other signal is connected to one end of the amplitude attenuation and amplitude phase detector 12. The microwave signal emitted by the antenna of the transmitting end 8 is received by the antenna of the receiving end after propagating in the medium. The antenna of the receiving end 9 is connected to the other end of the amplitude phase detector 12. The amplitude phase detector 12 obtains the amplitude attenuation and phase shift signals after the microwave propagates through the medium, and converts them into voltage signals, which are then collected through a data acquisition card. The amplitude attenuation of the microwave signal can reflect the changes in the dielectric properties of the fluid, which is used for the identification of flow patterns and the prediction of water holdup.

[0082] A gas-liquid two-phase flow pattern recognition and water holdup prediction method based on a deep neural network comprises the following steps:

[0083] S1: Obtaining the original gas-liquid two-phase fluid amplitude attenuation time series data;

[0084] S2: Preprocessing the original gas-liquid two-phase fluid amplitude attenuation time series data to obtain enhanced gas-liquid two-phase fluid amplitude attenuation time series data; the preprocessing adopts a sliding segmentation method.

[0085] S3: Based on the original gas-liquid two-phase fluid amplitude attenuation time series data, a gas-liquid two-phase fluid flow morphology recognition model is constructed;

[0086] S4: Based on the trained gas-liquid two-phase fluid flow pattern recognition model, it is recognized that there are three flow patterns of gas-liquid two-phase flow, namely, slug flow, mixed flow and bubbly flow;

[0087] S5: Based on the enhanced gas-liquid two-phase fluid amplitude attenuation time series data, a water holdup prediction model for each flow pattern is constructed;

[0088] S6: Based on the trained water holdup prediction model for each flow type, the water holdup corresponding to each flow type of slug flow, mixed flow and bubbly flow is finally predicted.

[0089] Steps S1 / S2 / S3 / S4 / S5 / S6 are executed in sequence;

[0090] Through end-to-end training, the deep network can realize automatic recognition of flow patterns at a lower cost, avoid mathematical modeling of complex systems, and replace traditional water holdup measurement methods with higher accuracy. The collected data is expanded using sliding segmentation to increase the number of samples, and the step size of sliding segmentation is set to 50.

[0091] The input sample length of the flow pattern recognition model is 2000×1 during 1s, and the labels are bubbly flow, slug flow, and mixed flow. The input sample length of the water holdup prediction model is 4000×1 during 2s, and the label is the corresponding water holdup. In order to speed up the convergence of the flow pattern recognition network model training, the samples are preprocessed by subtracting the mean and dividing by the variance.

[0092] In the CNN-LSTM model of the flow form recognition model, the number of convolution kernels in convolution layer I is 16, the convolution kernel size is 2×1, the activation function is RELU, the size of pooling layer I is 2×1, the number of convolution kernels in convolution layer II is 32, the convolution kernel size is 3×1, the activation function is RELU, the size of pooling layer II is 2×1, the number of convolution kernels in convolution layer III is 64, the convolution kernel size is 4×1, the activation function is RELU, and the size of pooling layer III is 2×1.

[0093] Different CNN-LSTM models were designed and optimized to identify flow patterns and predict water holding capacity, ultimately achieving the identification of flow patterns and accurate prediction of water holding capacity based on flow patterns.

[0094] During the flow of gas-liquid two-phase flow, due to the effect of water on microwaves, the microwave signal will experience amplitude attenuation and phase shift, and the value of amplitude attenuation depends on the water content. Gas-liquid two-phase flow will produce different flow structures during the flow process, and the prediction of water content, i.e., water holdup, is affected by the flow structure. There are three main flow patterns in gas-liquid two-phase flow, namely, slug flow, mixed flow, and bubbly flow. Neural networks provide nonlinear mapping between input and output variables and the relationship between these variables, becoming an emerging tool for automatic identification of flow types. When conducting dynamic experiments, the sampling frequency was set to 2kHz, and data of 30 seconds in length was collected for each experimental condition. Therefore, the data collected by the microstrip antenna microwave sensor measurement system each time was a long sequence of 60000×2, one column of which was amplitude attenuation and the other was phase shift. The input data of the neural network designed in this application uses the amplitude signal obtained by the microwave sensor in the dynamic experiment. The input sample length of the CNN-LSTM model for flow pattern recognition is 2000×1 during 1s, and the label is the specific flow type corresponding to the amplitude signal.

[0095] After all experimental results were split into a 2000×1 structure, a total of 660 samples were generated. To ensure that the model made balanced use of the experimental data of the three flow types, the ratio of the training set and the validation set for each flow type was fixed. A total of 528 samples were used for network training, and the remaining 132 samples were used for validation. Figure 4 The CNN-LSTM framework for flow pattern recognition is demonstrated. The amplitude signal with a length of 2000 is first normalized by Z-Score to eliminate the scale effect between data, and then the normalized data is input to 1DCNN, which roughly extracts features from the input sequence and integrates the effective information of the input sequence into the features through 3 convolutional layers and 3 pooling layers. The data processed by the 1DCNN module is input to the LSTM layer, which perceives the long-term dependency of the feature sequence. However, the features extracted by CNN-LSTM cannot be directly connected to the neural network for classification. The features must first be integrated through the flat layer and then input to the DENSE layer, and finally the specific flow pattern of the gas-liquid two-phase flow is output.

[0096] In the network training of the flow pattern recognition model, the cross entropy loss function and Adam optimizer were used, the learning rate was set to 0.0001, the batch size was set to 64, and a total of 10 training cycles were performed. When the model performance obtained based on the training set is optimal, it is evaluated on the validation set for final comparison. Repeat the operation 20 times according to the above configuration,

[0097] Figure 5 The confusion matrices of the flow pattern identification model with the lowest and highest accuracy are shown. The results show that the flow pattern identification model has high accuracy and reliability. Different two-phase flow patterns have different dynamic characteristics, so the microwave sensor signals will have certain differences. Designing different water holdup prediction models according to different flow patterns is conducive to improving the utilization of data and increasing the accuracy of water holdup prediction.

[0098] The training of the water holding rate prediction model requires a large number of samples. In order to increase the number of samples, the sliding segmentation method is used to achieve data expansion. The step size of the sliding segmentation is set to 50. After data expansion, 1121 random samples can be generated for each water holding rate label, of which 1008 samples are used for network training and the remaining 113 samples are used for network testing.

[0099] Figure 6 The CNN-LSTM architecture designed for three different flow types: bubbly flow, mixed flow and slug flow is demonstrated.

[0100] The microwave amplitude attenuation sequence signal with a length of 4000 is first preprocessed by Z-Score, and then enters different CNN-LSTM frameworks to extract their respective features according to their flow patterns. The extracted features are processed by the flat layer and then enter the DENSE layer, and finally the corresponding water holdup is output.

[0101] In the CNN-LSTM model of the bubbly flow water holdup prediction model, the number of convolution kernels in convolution layer I is 16, the convolution kernel size is 2×1, the activation function is RELU, the maximum pooling layer I size is 2×1, the number of convolution kernels in convolution layer II is 32, the convolution kernel size is 3×1, the activation function is RELU, the maximum pooling layer II size is 2×1, the number of convolution kernels in convolution layer III is 64, the convolution kernel size is 4×1, the activation function is RELU, the maximum pooling layer III size is 2×1, the number of convolution kernels in convolution layer IV is 128, the convolution kernel size is 5×1, the activation function is RELU, and the maximum pooling layer IV size is 2×1.

[0102] In the CNN-LSTM model of the mixed flow holdup prediction model, the number of convolution kernels in convolution layer I is 16, the convolution kernel size is 2×1, the activation function is RELU, the number of convolution kernels in convolution layer II is 16, the convolution kernel size is 2×1, the activation function is RELU, and the average pooling layer size is 2×1, the number of convolution kernels in convolution layer III is 32, the convolution kernel size is 3×1, and the activation function is RELU, the number of convolution kernels in convolution layer IV is 32, the convolution kernel size is 3×1, the activation function is RELU, and the average pooling layer size is 2×1. The size of convolution layer II is 2×1, the number of convolution kernels of convolution layer V is 64, the convolution kernel size is 4×1, and the activation function is RELU. The number of convolution kernels of convolution layer VI is 64, the convolution kernel size is 4×1, and the activation function is RELU. The size of maximum pooling layer III is 2×1, the number of convolution kernels of convolution layer VII is 128, the convolution kernel size is 5×1, and the activation function is RELU. The number of convolution kernels of convolution layer VIII is 128, the convolution kernel size is 5×1, and the activation function is RELU. The size of maximum pooling layer IV is 2×1.

[0103] In the CNN-LSTM model of the slug flow water holding capacity prediction model, the number of convolution kernels in convolution layer I is 16, the convolution kernel size is 2×1, and the activation function is RELU. The size of the maximum pooling layer I is 2×1, the number of convolution kernels in convolution layer II is 32, the convolution kernel size is 3×1, and the activation function is RELU. The size of the maximum pooling layer II is 2×1, the number of convolution kernels in convolution layer III is 64, the convolution kernel size is 4×1, and the activation function is RELU. The size of the maximum pooling layer III is 2×1, the number of convolution kernels in convolution layer IV is 128, the convolution kernel size is 5×1, and the activation function is RELU. The size of the maximum pooling layer IV is 2×1, the number of convolution kernels in convolution layer V is 256, the convolution kernel size is 6×1, the activation function is RELU, and the size of the maximum pooling layer V is 2×1.

[0104] In the network training of the water holding rate model, the MSE loss function and Adam optimizer were used, the learning rate was set to 0.001, the batch size was set to 64, and a total of 100 training cycles were performed. In order to intuitively evaluate the prediction results of water holding rate, we introduced three commonly used regression analysis indicators, namely, mean absolute error (MAE), mean absolute percentage error (MAPE), and correlation coefficient (R2). The prediction results are shown in Figure 2. Figure 7 As shown, Figure 7 (a) shows the weighted average results of the water holdup prediction models designed according to three different flow patterns. Figure 7 (b) shows the resulting metrics of inputting different flow patterns into the same water holdup prediction model.

[0105] The results show that designing different CNN-LSTM models according to different flow patterns can predict water holdup more accurately and make better use of data.

[0106] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A microwave sensor system based on a microstrip antenna, characterized in that: include: Signal source: used to generate microwave signals; Power divider: used for dividing the microwave signal generated by the signal source into two identical signals: signal I and signal II; Microstrip antenna microwave transmission sensor: used to receive the signal I transmitted by the power divider, and transmit the signal I after amplitude attenuation and phase shift reflecting the change of dielectric properties of gas-liquid two-phase fluid; Amplitude phase detector: used for receiving the signal II transmitted by the power divider and the signal I after amplitude attenuation and phase shift transmitted by the microstrip antenna microwave transmission sensor, and demodulating the amplitude attenuation and phase shift of the microwave after passing through the fluid by comparing the signal I and the signal II; Data acquisition module: used to collect the amplitude attenuation and phase shift signals of the microwaves transmitted by the amplitude phase detector after passing through the fluid.

2. A microstrip antenna microwave sensor system according to claim 1, characterized in that: The microstrip antenna microwave transmission sensor comprises Transmitter: used for receiving the signal I transmitted by the power divider; Receiving end: used to receive the signal I after amplitude attenuation and phase shift transmitted by the receiving end of the microstrip antenna microwave transmission sensor, reflecting the change of dielectric properties of gas-liquid two-phase fluid; The transmitting end and the receiving end are symmetrically arranged on the outer wall of a measuring pipe in which a gas-liquid two-phase fluid is transported.

3. A microstrip antenna microwave sensor system according to claim 2, characterized in that: The receiving end and the transmitting end have the same structure; The receiving end includes a dielectric board; A radiation patch is arranged on the upper surface of the dielectric plate; A feeding point is provided on the radiation patch; Radiation gaps are provided on both sides of the radiation patch in the length direction; The lower surface of the dielectric plate is a ground plane; A coaxial connector is arranged below the radiation patch.

4. The microstrip antenna microwave sensor system according to claim 3, characterized in that: The length L1 of the radiation patch is 28.5 mm, and W1 is 38 mm.

5. A gas-liquid two-phase flow pattern recognition and water holdup prediction method based on a deep neural network, characterized in that: The following steps are involved: Obtain the original gas-liquid two-phase fluid amplitude attenuation time series data; Preprocessing the original gas-liquid two-phase fluid amplitude attenuation time series data to obtain enhanced gas-liquid two-phase fluid amplitude attenuation time series data; Based on the original gas-liquid two-phase fluid amplitude attenuation time series data, a gas-liquid two-phase fluid flow morphology recognition model is constructed; Based on the trained gas-liquid two-phase fluid flow pattern recognition model, it is identified that there are three flow patterns of gas-liquid two-phase flow, namely slug flow, mixed flow and bubbly flow. Based on the enhanced gas-liquid two-phase fluid amplitude attenuation time series data, a water holdup prediction model for each flow pattern is constructed; Based on the trained water holdup prediction model for each flow type, the water holdup corresponding to each flow type of slug flow, mixed flow and bubbly flow is finally predicted.

6. The method for gas-liquid two-phase flow pattern recognition and water holdup prediction based on a deep neural network according to claim 1, characterized in that: The preprocessing adopts a sliding segmentation method.

7. The method for gas-liquid two-phase flow pattern recognition and water holdup prediction based on a deep neural network according to claim 1, characterized in that: The process of identifying the three flow forms of gas-liquid two-phase flow, namely, slug flow, mixed flow and bubbly flow, based on the trained gas-liquid two-phase fluid flow form recognition model is as follows: The original gas-liquid two-phase fluid amplitude attenuation time series data is first processed by Z-Score standardization to eliminate the scale effect between the data, and then the standardized data is input into the 1DCNN. 1DCNN roughly extracts the features of the input sequence, and integrates the effective information of the input sequence into the features through 3 convolutional layers and 3 pooling layers. The data processed by the 1DCNN module is input into the LSTM layer. The LSTM layer perceives the long-term dependency of the feature sequence. The features extracted by CNN-LSTM are first integrated by the flattening layer and then input into the DENSE layer, and finally the specific flow type of the gas-liquid two-phase flow is output.

8. The method for gas-liquid two-phase flow pattern recognition and water holdup prediction based on a deep neural network according to claim 1, characterized in that: The process of finally predicting the water holdup corresponding to each flow type of slug flow, mixed flow and bubbly flow based on the trained water holdup prediction model for each flow type is as follows: The microwave amplitude attenuation sequence signal is first preprocessed, and then enters different CNN-LSTM frameworks to extract their respective features according to their flow patterns. The extracted features are processed by the flat layer and then enter the DENSE layer, and finally the corresponding water holdup is output.

9. The method for gas-liquid two-phase flow pattern recognition and water holdup prediction based on a deep neural network according to claim 1, characterized in that: The flow form recognition model includes a CNN-LSTM model; In the CNN-LSTM model, the number of convolution kernels in convolution layer I is 16, the convolution kernel size is 2×1, the activation function is RELU, the size of pooling layer I is 2×1, the number of convolution kernels in convolution layer II is 32, the convolution kernel size is 3×1, the activation function is RELU, the size of pooling layer II is 2×1, the number of convolution kernels in convolution layer III is 64, the convolution kernel size is 4×1, the activation function is RELU, and the size of pooling layer III is 2×1.

10. The method for gas-liquid two-phase flow pattern recognition and water holdup prediction based on a deep neural network according to claim 1, characterized in that: The water holdup prediction model for each flow type includes a bubbly flow water holdup prediction model, a mixed flow water holdup prediction model and a slug flow water holdup prediction model; In the CNN-LSTM model of the bubbly flow holdup prediction model, the number of convolution kernels in convolution layer I is 16, the convolution kernel size is 2×1, the activation function is RELU, the maximum pooling layer I size is 2×1, the number of convolution kernels in convolution layer II is 32, the convolution kernel size is 3×1, the activation function is RELU, the maximum pooling layer II size is 2×1, the number of convolution kernels in convolution layer III is 64, the convolution kernel size is 4×1, the activation function is RELU, the maximum pooling layer III size is 2×1, the number of convolution kernels in convolution layer IV is 128, the convolution kernel size is 5×1, the activation function is RELU, and the maximum pooling layer IV size is 2×1; In the CNN-LSTM model of the mixed flow holdup prediction model, the number of convolution kernels in convolution layer I is 16, the convolution kernel size is 2×1, the activation function is RELU, the number of convolution kernels in convolution layer II is 16, the convolution kernel size is 2×1, the activation function is RELU, and the average pooling layer size is 2×1, the number of convolution kernels in convolution layer III is 32, the convolution kernel size is 3×1, and the activation function is RELU, the number of convolution kernels in convolution layer IV is 32, the convolution kernel size is 3×1, the activation function is RELU, and the average pooling layer size is 2×1. The size of convolution layer II is 2×1, the number of convolution kernels of convolution layer V is 64, the convolution kernel size is 4×1, and the activation function is RELU. The number of convolution kernels of convolution layer VI is 64, the convolution kernel size is 4×1, and the activation function is RELU. The size of maximum pooling layer III is 2×1, the number of convolution kernels of convolution layer VII is 128, the convolution kernel size is 5×1, and the activation function is RELU. The number of convolution kernels of convolution layer VIII is 128, the convolution kernel size is 5×1, and the activation function is RELU. The size of maximum pooling layer IV is 2×1; In the CNN-LSTM model of the slug flow water holding capacity prediction model, the number of convolution kernels in convolution layer I is 16, the convolution kernel size is 2×1, and the activation function is RELU. The size of the maximum pooling layer I is 2×1, the number of convolution kernels in convolution layer II is 32, the convolution kernel size is 3×1, and the activation function is RELU. The size of the maximum pooling layer II is 2×1, the number of convolution kernels in convolution layer III is 64, the convolution kernel size is 4×1, and the activation function is RELU. The size of the maximum pooling layer III is 2×1, the number of convolution kernels in convolution layer IV is 128, the convolution kernel size is 5×1, and the activation function is RELU. The size of the maximum pooling layer IV is 2×1, the number of convolution kernels in convolution layer V is 256, the convolution kernel size is 6×1, the activation function is RELU, and the size of the maximum pooling layer V is 2×1.

Images