Horizontal gas-liquid two-phase flow pattern discrimination system and method

Abstract

The invention discloses a horizontal gas-liquid two-phase flow pattern discrimination system and method, and belongs to the field of fluid measurement and flow pattern recognizing.The horizontal gas-liquid two-phase flow pattern discrimination system comprises a measurement pipeline, array sensors are arranged in the measurement pipeline, and the output ends of the array sensors are sequentially connected with a communication power source, a multi-path signal synchronous acquisition module and an upper computer processing module; and meanwhile, a detection network framework is constructed, and the collected data is analyzed by means of the network framework to realize flow pattern discrimination. According to the method, the problems of poor data stability, low precision, incapability of reflecting an internal mechanism and the like in an existing flow pattern identification method can be effectively solved.

Description

Technical Field

[0001] The invention relates to the field of horizontal gas-liquid two-phase flow fluid measurement and flow pattern identification, in particular to a horizontal gas-liquid two-phase flow flow pattern identification system and method. Background Art

[0002] As an unconventional oil and gas resource, shale gas is a hot topic in global energy research. It has low porosity and permeability and is a natural gas accumulation that mainly exists in an adsorbed or free state. The flow characteristics of gas-liquid two-phase flow are described by the flow parameters of the fluid in the pipeline. Different physical properties such as flow velocity, density, viscosity, etc. between the phases will form different flow patterns. In the past few decades, the flow pattern research of gas-liquid two-phase flow has mainly focused on two solutions: high-speed cameras and sensor technology. However, there are problems such as high equipment cost, complex data processing and poor real-time performance. In addition, high-speed cameras are sensitive to lighting conditions and environmental changes, and are easily disturbed, resulting in unstable flow pattern recognition results. However, traditional sensors are often affected by environmental factors when transmitting data, such as temperature and pressure changes, and frame drops during signal transmission, which will seriously affect the integrity and accuracy of the data.

[0003] Flow pattern represents the form or structure of fluid flow. The transformation of flow pattern and the change of its characteristics greatly affect the flow characteristics of two-phase flow. Accurate identification of flow pattern is of great significance to the study of gas-liquid two-phase flow. In the early stage of flow pattern identification, researchers identified the flow pattern of gas-liquid two-phase flow based on flow pattern diagrams of empirical methods. Flow pattern diagrams have the advantages of simplicity and intuitiveness, but they cannot reflect the mechanism of flow pattern conversion. With the further study of gas-liquid two-phase flow by researchers, the boundary of flow pattern transformation is given on the basis of mechanical analysis, and experimental data is used to correct it to improve the recognition accuracy. However, due to the complexity of gas-liquid two-phase flow, this method is usually complex and has low accuracy. At the same time, in the process of flow pattern signal acquisition, the spatiotemporal characteristics of gas-liquid flow and the similarity between channels are often ignored. The semantic consistency between the dynamic change characteristics of gas-liquid two-phase flow in the time dimension also needs to be paid attention to. Therefore, it is urgent to carry out new method and technology research on shale gas gas-liquid two-phase flow pattern identification. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a horizontal gas-liquid two-phase flow type discrimination system and method. In view of the problem that traditional sensor data transmission is easily affected by environmental interference, resulting in low flow type identification accuracy, the system can collect, image and process horizontal gas-liquid two-phase flow signals with high precision and stability, thereby realizing accurate discrimination of flow types.

[0005] To solve the above technical problems, the technical solution adopted by the present invention is: a horizontal gas-liquid two-phase flow type discrimination system, including a horizontally placed measuring pipeline, a measuring pipeline, an array sensor arranged inside, an array sensor, an output end of which is connected in sequence to a power supply, a multi-channel signal synchronous acquisition module, and a host computer processing module, a multi-channel signal synchronous acquisition module, and a plurality of input channels that can be synchronously converted for converting the collected data of the array sensor and inputting them into the host computer processing module, the host computer processing module, and a built-in detection network framework for processing the output data of the multi-channel signal synchronous acquisition module. The detection network framework includes N LSTM layers, 1 LSTM layer, a shared spatial attention module, a multi-channel self-attention module and a flow type classification module from the input to the output direction.

[0006] The further improvement of the technical solution of the present invention is that the structures of the various parts of the detection network framework are specifically as follows:

[0007] The shared spatial attention module includes a global average pooling layer, a feature partitioning layer, four deep one-dimensional convolutional layers with kernel sizes of 3, 5, 7, and 9, a concat operation, a grouping layer, a sigmoid function, and a multiplication operation with the original output from input to output.

[0008] The multi-channel self-attention module includes a global average pooling layer, a grouping layer, three 1×1 convolutional layers, a channel attention and multi-scale spatial information self-attention mechanism layer, a global average pooling layer, a sigmoid function and a multiplication operation with the output of the shared spatial attention module from input to output.

[0009] The flow classification module includes three one-dimensional convolution blocks with filter sizes of 128, 256 and 128, kernel sizes of 8, 6 and 3 respectively, and a Softmax classifier from input to output.

[0010] A method for distinguishing flow patterns of horizontal gas-liquid two-phase flow is provided. First, a flow pattern classification module is trained. The specific steps of distinguishing flow patterns using the trained flow pattern classification module are as follows:

[0011] Step 1: Obtain the response sequence of the gas-liquid two-phase flow of the cross section of the optical fiber probe of the array sensor, collect N time series frames, each frame has a frame length of M data points, and a total of N×M data can be obtained;

[0012] Step 2: Use a multi-channel acquisition module to obtain multi-channel response sequence signals of the array sensor probe and generate a multi-channel response curve;

[0013] Step 3: Transmit the multi-channel corresponding curves to the detection network framework for flow type discrimination;

[0014] Step 4: Input the curve information generated in the delay direction into N LSTM layers. Each LSTM layer corresponds to the measurement data of different channels. Each LSTM layer learns and fuses the information from a single channel.

[0015] Step 5: Integrate the spatial information. The outputs of the N LSTM layers are concatenated and input into a new LSTM layer, allowing interactive fusion of temporal information across different spatial locations. The final output size is (M, T), which indicates that the input features are fused M times in parallel to respond to the significant spatial multi-angle nature of gas-liquid flow.

[0016] Step 6: Input the output of step 5 into the spatial attention module to further explore the spatiotemporal characteristics of gas-liquid flow;

[0017] Step 7: Input the output of step 6 into the multi-channel self-attention module to combine the modulation space prior to calculate the similarity between channels;

[0018] Step 8: Input the output result of step 7 into the flow type classification module, process it through the full convolution layer, and finally perform flow type classification.

[0019] The further improvement of the technical solution of the present invention is that the process of learning and training the flow type classification module is as follows: establishing a rectangular coordinate system of the horizontal measurement pipeline cross section plane, determining the coordinates of each probe in the array probe sensor structure and the center of the cross section to be measured, setting the gas phase and liquid phase standards for each probe in the array structure, and initializing the value to the median of the gas phase and liquid phase output voltages;

[0020] At a point (r, θ, x) on the probe cross section, when the gas-liquid two-phase flow contacts the fiber optic probe alternately, the instantaneous local gas content at the point (r, θ, x) is:

[0021]

[0022] Where v g represents the gas phase, where v l represents the liquid phase, t represents the time, r represents the polar coordinate, θ represents the angle of the polar coordinate, and x represents the horizontal axis;

[0023] Various flow patterns are simulated in the measurement pipeline, and the network is used to learn and extract the gas content value output by the corresponding probe position under different flow patterns. The flow pattern classification module is trained using the set of gas content output values ​​corresponding to each probe, and the flow pattern classifier module parameters are optimized.

[0024] A further improvement of the technical solution of the present invention is that the experimental sequence collected in step 2 is expressed as:

[0025] Where T is the number of sampling points and n is the number of channels.

[0026] The further improvement of the technical solution of the present invention is that the specific steps of step 6 are as follows:

[0027] Step 6.1: Input the fused feature map into the shared spatial attention module for pooling layer processing;

[0028] Step 6.2: To understand the spatial distribution of changes and the relationship between changes before and after, divide the feature set into K independent sub-features of the same size The number of channels for each sub-feature is X j represents the jth sub-feature, where j∈[1,K], each sub-feature is independent, which facilitates efficient extraction of spatial information;

[0029] Step 6.3: In order to enrich information, enhance semantic consistency and minimize semantic gap, deep one-dimensional convolution with kernel sizes of 3, 5, 7 and 9 is applied to the sub-features;

[0030] Step 6.4: By applying one-dimensional convolution to the feature decomposition W dimensions respectively, the limited receptive field can be solved and the information extraction process can be realized Represents the spatial structure information of the jth sub-feature obtained after the light convolution operation; K j Represents the convolution kernel applied to the i-th sub-feature; the decomposition process of sub-features

[0031] Step 6.5: Accurately calculate the spatial attention map of each feature, aggregate sub-features at different positions, and select L group normalization for processing. This processing can better distinguish the differences between sub-features than batch normalization. Group normalization allows each sub-feature to be normalized independently without using batch statistical noise, which can effectively reduce the mutual interference between sub-features and prevent attention distraction; the calculation of the output feature is as follows

[0032]

[0033] X s =Attn W ×X

[0034] σ(·) represents Sigmoid normalization, It means there are L groups of GN along the W dimension.

[0035] The further improvement of the technical solution of the present invention is that the specific steps of step 7 are as follows:

[0036] Step 7.1: Perform a pooling layer operation on the output and then generate the mapping function of Q, K, and V Represents the convolution operation; where Q, K, V∈RB×C×N;

[0037] Step 7.2: After passing through the pooling layer and the Sigmoid function, it is multiplied with the feature map of the previous stage and finally outputs the result.

[0038] Due to the adoption of the above technical scheme, the technical progress achieved by the present invention is: using array probe sensors to obtain the response sequence of horizontal gas-liquid two-phase flow, analyzing the stability results of different channels, and mining and enhancing the flow pattern recognition features of more effective information through different spatial positions and channels, providing information for subsequent research on flow pattern recognition methods. The present invention proposes a flow pattern discrimination method for horizontal gas-liquid two-phase flow based on array probe sensors and deep learning technology. By setting N+1 LSTM layers, the characteristics of time series data can be fully utilized, and the measurement data from different channels and positions can be effectively captured and fused, thereby improving the stability and reliability of data processing. In particular, the newly added LSTM layer enables the model to interactively fuse information across different spatial positions and time scales, thereby more accurately reflecting the flow characteristics of gas-liquid two-phase flow. By introducing a shared spatial attention module, the module can deeply mine and enhance the effective features of flow pattern recognition by performing operations such as global average pooling, feature partitioning, and deep one-dimensional convolution on the feature map. Through the shared spatial attention mechanism, the model can focus on important spatial areas and ignore irrelevant information, further improving the accuracy and robustness of flow pattern recognition. In addition, the introduction of the multi-channel self-attention module enables the model to calculate the similarity between channels and modulate the spatial prior information, so as to better understand the mutual influence and interaction between different channels. This mechanism enables the model to more comprehensively capture the dynamic change characteristics of flow patterns when processing multi-channel data, further improving the accuracy and generalization ability of flow pattern recognition.

[0039] It effectively solves the problems of poor data stability, low accuracy and inability to reflect the internal mechanism in existing flow pattern identification methods, and provides a new solution for the identification of horizontal gas-liquid two-phase flow patterns. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] 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 only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative work.

[0041] Figure 1 It is a structural schematic diagram of the gas-liquid two-phase flow pattern classification system of the present invention;

[0042] Figure 2It is a schematic diagram of the network framework structure of the gas-liquid two-phase flow type classification system of the present invention;

[0043] Figure 3 This is a flow chart of a method for classifying flow patterns of horizontal gas-liquid two-phase flow according to the present invention;

[0044] Figure 4 Schematic diagram of imaging different flow patterns in a cross section of an array optical fiber probe structure designed in the present invention;

[0045] Among them, 1. measuring pipeline, 2. array sensor, 3. connected power supply, 4. multi-channel signal synchronous acquisition module, 5. host computer processing module, 6. inlet, 7. outlet, 81. gas, 82. water, 83. gas-liquid interface, 91. bubbles. DETAILED DESCRIPTION

[0046] The present invention is further described in detail below in conjunction with embodiments:

[0047] like Figure 1 As shown in FIG. 1 , it is a schematic diagram of the structure of a horizontal gas-liquid two-phase flow pattern discrimination system, a horizontally placed measuring pipe 1, one end of the measuring pipe is an inlet 6 and the other end is an outlet 7. The measuring pipe 1 is filled with gas-liquid two-phase shale gas. The pipe is filled with array sensors 2 composed of optical fiber probes. The probes are arranged all over the pipe cross section, with one probe at the highest point and one probe at the lowest point. The closer to the center, the more probes there are. The arrangement of the array sensors 2 is as follows: Figure 4 The output end of the array sensor 2 is connected to a power supply 3, a multi-channel signal synchronous acquisition module 4 and a host computer processing module 5 in sequence. The multi-channel signal synchronous acquisition module 4 includes a plurality of input channels that can be synchronously converted to convert the data collected by the array sensor 2 and input them to the host computer processing module 5. The host computer processing module 5 has a built-in detection network framework for processing the output data of the multi-channel signal synchronous acquisition module 4. The detection network framework is shown in FIG. Figure 2 As shown in the figure, from input to output, it includes N LSTM layers, 1 LSTM layer, shared spatial attention module, multi-channel self-attention module and flow classification module. The shared spatial attention module includes a global average pooling layer, a feature partitioning layer, 4 deep one-dimensional convolutional layers with kernel sizes of 3, 5, 7 and 9, a concat operation, a grouping layer, a Sigmoid function and a multiplication operation with the original output.

[0048] The multi-channel self-attention module includes a global average pooling layer, a grouping layer, three 1×1 convolutional layers, a channel attention and multi-scale spatial information self-attention mechanism layer, a global average pooling layer, a sigmoid function and a multiplication operation with the output of the shared spatial attention module from input to output.

[0049] The flow classification module includes three one-dimensional convolution blocks with filter sizes of 128, 256 and 128, kernel sizes of 8, 6 and 3 respectively, and a Softmax classifier from input to output.

[0050] With the help of the above system, the flow pattern of two-phase flow can be distinguished, such as Figure 3 The figure shows a flow chart of a method for classifying flow patterns of horizontal gas-liquid two-phase flows. First, the flow pattern classification module is trained. The process of training the flow pattern classification module is as follows: establishing a rectangular coordinate system of the plane of the horizontal measurement pipeline section, determining the coordinates of each probe in the array probe sensor structure and the center of the section to be measured, setting the gas phase and liquid phase standards for each probe in the array structure, and initializing the value to the median of the gas phase and liquid phase output voltages;

[0051] At a point (r, θ, x) on the probe cross section, when the gas-liquid two-phase flow contacts the fiber optic probe alternately, the instantaneous local gas content at the point (r, θ, x) is:

[0052]

[0053] Where v g represents the gas phase, where v l represents the liquid phase, t represents the time, r represents the polar coordinate, θ represents the angle of the polar coordinate, and x represents the horizontal axis;

[0054] Simulate various flow patterns in the measuring pipe, such as Figure 4 As shown, by controlling the inlet and outlet of the measuring pipe, various flow types are simulated inside the measuring pipe while the probe collects data, including laminar flow, wavy flow, elastic flow, plug flow, bubbly flow and annular flow. Figure 4 The probe positions are composed in different flow scenarios, where 81 represents gas, 82 represents water, 83 represents the gas-liquid interface, and 91 represents bubbles. Based on these representations, the information of the horizontal section can be obtained, and then the gas content of each probe can be obtained. The network is used to learn and extract the gas content value output by the corresponding probe position under different flow patterns, and the flow pattern classification module is trained using the corresponding gas content output value set of each probe to optimize the flow pattern classifier module parameters. The trained flow pattern classification module can judge the flow pattern through the collected data.

[0055] The specific steps for flow type discrimination using the trained flow type classification module are as follows:

[0056] Step 1: Obtain the response sequence of the gas-liquid two-phase flow of the cross section of the optical fiber probe of the array sensor, collect N time series frames, each frame has a frame length of M data points, and a total of N×M data can be obtained;

[0057] Step 2: Use the multi-channel acquisition module to obtain the multi-channel response sequence signal of the array sensor probe and generate a multi-channel response curve; the acquired experimental sequence is expressed as:

[0058] Where T is the number of sampling points and n is the number of channels.

[0059] Step 3: Transmit the multi-channel corresponding curves to the detection network framework for flow type discrimination;

[0060] Step 4: Input the curve information generated in the delay direction into N LSTM layers. Each LSTM layer corresponds to the measurement data of different channels. Each LSTM layer learns and fuses the information from a single channel.

[0061] Step 5: Integrate the spatial information. The outputs of the N LSTM layers are concatenated and input into a new LSTM layer, allowing interactive fusion of temporal information across different spatial locations. The final output size is (M, T), which indicates that the input features are fused M times in parallel to respond to the significant spatial multi-angle nature of gas-liquid flow.

[0062] Step 6: Input the output of step 5 into the spatial attention module to further explore the spatiotemporal characteristics of gas-liquid flow;

[0063] Step 6.1: Input the fused feature map into the shared spatial attention module for pooling layer processing;

[0064] Step 6.2: To understand the spatial distribution of changes and the relationship between changes before and after, divide the feature set into K independent sub-features of the same size The number of channels for each sub-feature is X j represents the jth sub-feature, where j∈[1,K], each sub-feature is independent, which facilitates efficient extraction of spatial information;

[0065] Step 6.3: In order to enrich information, enhance semantic consistency and minimize semantic gap, deep one-dimensional convolution with kernel sizes of 3, 5, 7 and 9 is applied to the sub-features;

[0066] Step 6.4: By applying one-dimensional convolution to the feature decomposition W dimensions respectively, the limited receptive field can be solved and the information extraction process can be realized Represents the spatial structure information of the jth sub-feature obtained after the light convolution operation; K j Represents the convolution kernel applied to the i-th sub-feature; the decomposition process of sub-features

[0067] Step 6.5: Accurately calculate the spatial attention map of each feature, aggregate sub-features at different positions, and select L group normalization for processing. This processing can better distinguish the differences between sub-features than batch normalization. Group normalization allows each sub-feature to be normalized independently without using batch statistical noise, which can effectively reduce the mutual interference between sub-features and prevent attention distraction; the calculation of the output feature is as follows

[0068]

[0069] X s =Attn W ×X

[0070] σ(·) represents Sigmoid normalization, It means there are L groups of GN along the W dimension.

[0071] Step 7: Input the output of step 6 into the multi-channel self-attention module to combine the modulation space prior to calculate the similarity between channels;

[0072] Step 7.1: Perform a pooling layer operation on the output and then generate the mapping function of Q, K, and V Represents the convolution operation; where Q, K, V∈RB×C×N;

[0073] Step 7.2: After passing through the pooling layer and the Sigmoid function, it is multiplied with the feature map of the previous stage and finally outputs the result.

[0074] Step 8: Input the output result of step 7 into the flow type classification module, process it through the full convolution layer, and finally perform flow type classification.

[0075] The embodiments described above are merely descriptions of preferred implementation modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A horizontal gas-liquid two-phase flow pattern discrimination system, comprising a horizontally placed measuring pipe (1), wherein the measuring pipe (1) is filled with array sensors (2), and the output end of the array sensor (2) is sequentially connected to a power supply (3), a multi-channel signal synchronous acquisition module (4) and a host computer processing module (5), characterized in that: The multi-channel signal synchronous acquisition module (4) comprises a plurality of synchronously convertible input channels for converting the data collected by the array sensor (2) and inputting the converted data into the upper computer processing module (5); the upper computer processing module (5) has a built-in detection network framework for processing the output data of the multi-channel signal synchronous acquisition module (4); the detection network framework comprises N LSTM layers, 1 LSTM layer, a shared spatial attention module, a multi-channel self-attention module and a flow type classification module in order from the input to the output direction.

2. A horizontal gas-liquid two-phase flow pattern discrimination system according to claim 1, characterized in that: The structures of each part of the detection network framework are as follows: The shared spatial attention module includes a global average pooling layer, a feature partitioning layer, four deep one-dimensional convolutional layers with kernel sizes of 3, 5, 7, and 9, a concat operation, a grouping layer, a sigmoid function, and a multiplication operation with the original output from input to output. The multi-channel self-attention module includes a global average pooling layer, a grouping layer, three 1×1 convolutional layers, a channel attention and multi-scale spatial information self-attention mechanism layer, a global average pooling layer, a sigmoid function and a multiplication operation with the output of the shared spatial attention module from input to output. The flow classification module includes three one-dimensional convolution blocks with filter sizes of 128, 256 and 128, kernel sizes of 8, 6 and 3 respectively, and a Softmax classifier from input to output.

3. A method for distinguishing flow patterns of horizontal gas-liquid two-phase flow, implemented based on the distinguishing system according to claim 1 or 2, characterized in that: First, the flow type classification module is trained and the specific steps of using the trained flow type classification module to distinguish flow types are as follows: Step 1: Obtain the response sequence of the gas-liquid two-phase flow of the cross section of the optical fiber probe of the array sensor, collect N time series frames, each frame has a frame length of M data points, and a total of N×M data can be obtained; Step 2: Use a multi-channel acquisition module to obtain multi-channel response sequence signals of the array sensor probe and generate a multi-channel response curve; Step 3: Transmit the multi-channel corresponding curves to the detection network framework for flow type discrimination; Step 4: Input the curve information generated in the delay direction into N LSTM layers. Each LSTM layer corresponds to the measurement data of different channels. Each LSTM layer learns and fuses the information from a single channel. Step 5: Integrate the spatial information. The outputs of the N LSTM layers are concatenated and input into a new LSTM layer, allowing interactive fusion of temporal information across different spatial locations. The final output size is (M, T), which indicates that the input features are fused M times in parallel to respond to the significant spatial multi-angle nature of gas-liquid flow. Step 6: Input the output of step 5 into the spatial attention module to further explore the spatiotemporal characteristics of gas-liquid flow; Step 7: Input the output of step 6 into the multi-channel self-attention module to combine the modulation space prior to calculate the similarity between channels; Step 8: Input the output result of step 7 into the flow type classification module, process it through the full convolution layer, and finally perform flow type classification.

4. A method for distinguishing flow patterns of horizontal gas-liquid two-phase flow according to claim 3, characterized in that: The process of learning and training the flow type classification module is as follows: establish a rectangular coordinate system for the horizontal measurement pipeline cross section plane, determine the coordinates of each probe in the array probe sensor structure and the center of the cross section to be measured, set the gas phase and liquid phase standards for each probe in the array structure, and initialize the value to the median value of the gas phase and liquid phase output voltage; At a point (r, θ, x) on the probe cross section, when the gas-liquid two-phase flow contacts the fiber optic probe alternately, the instantaneous local gas content at the point (r, θ, x) is: Where v g represents the gas phase, where v l represents the liquid phase, t represents the time, r represents the polar coordinate, θ represents the angle of the polar coordinate, and x represents the horizontal axis; Various flow patterns are simulated in the measurement pipeline, and the network is used to learn and extract the gas content value output by the corresponding probe position under different flow patterns. The flow pattern classification module is trained using the set of gas content output values ​​corresponding to each probe, and the flow pattern classifier module parameters are optimized.

5. A method for distinguishing flow patterns of horizontal gas-liquid two-phase flow according to claim 3, characterized in that: The experimental sequence collected in step 2 is expressed as: Where T is the number of sampling points and n is the number of channels.

6. A method for distinguishing flow patterns of horizontal gas-liquid two-phase flow according to claim 3, characterized in that: Step 6 The specific steps are as follows: Step 6.1: Input the fused feature map into the shared spatial attention module for pooling layer processing; Step 6.2: To understand the spatial distribution of changes and the relationship between changes before and after, divide the feature set into K independent sub-features of the same size The number of channels for each sub-feature is X j represents the jth sub-feature, where j∈[1,K], each sub-feature is independent, which facilitates efficient extraction of spatial information; Step 6.3: In order to enrich information, enhance semantic consistency and minimize semantic gap, deep one-dimensional convolution with kernel sizes of 3, 5, 7 and 9 is applied to the sub-features; Step 6.4: By applying one-dimensional convolution to the feature decomposition W dimensions respectively, the limited receptive field can be solved and the information extraction process can be realized Represents the spatial structure information of the jth sub-feature obtained after the light convolution operation; K j Represents the convolution kernel applied to the i-th sub-feature; the decomposition process of sub-features Step 6.5: Accurately calculate the spatial attention map of each feature, aggregate sub-features at different positions, and select L group normalization for processing. This processing can better distinguish the differences between sub-features than batch normalization. Group normalization allows each sub-feature to be normalized independently without using batch statistical noise, which can effectively reduce the mutual interference between sub-features and prevent attention distraction; The output features are calculated as follows X s =Attn W ×X σ(·) represents Sigmoid normalization, It means there are L groups of GN along the W dimension.

7. A method for distinguishing flow patterns of horizontal gas-liquid two-phase flow according to claim 3, characterized in that: Step 7 The specific steps are as follows: Step 7.1: Perform a pooling layer operation on the output and then generate the mapping function of Q, K, and V Represents the convolution operation; where Q, K, V∈RB×C×N; Step 7.2: After passing through the pooling layer and the Sigmoid function, it is multiplied with the feature map of the previous stage and finally outputs the result.

Images