High-precision gas-liquid two-phase flow pattern identification method based on data conversion and multi-head attention mechanism

Abstract

The invention discloses a high-precision gas-liquid two-phase flow pattern identification method based on data conversion and a multi-head attention mechanism, and the method comprises the steps: collecting the conductivity fluctuation signals of five flow patterns, namely bubble flow, bubble plug flow, plug flow, plug split flow and stratified flow, according to an annular conductivity sensor after a gas-liquid two-phase flow dynamic experiment device is constructed; normalized data are converted into feature matrixes, feature differences between different flow patterns are highlighted, high-level features which do not exist in a time sequence can be revealed, a gas-liquid two-phase flow pattern recognition network is constructed, the feature matrixes of the different flow patterns are input into the recognition network for feature extraction, and the recognition accuracy of the gas-liquid two-phase flow pattern recognition network is improved. According to the multi-head self-attention mechanism of the window and the multi-head self-attention mechanism of the shift window, local and global features in the feature matrix can be captured, and then horizontal pipe gas-liquid two-phase flow pattern recognition is achieved. According to the method, the complex flow pattern characteristics of the gas-liquid two-phase flow of the horizontal pipe can be extracted more comprehensively, the accuracy of flow pattern recognition is improved, and the calculation cost is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of gas-liquid two-phase flow pattern recognition, and specifically relates to a high-precision gas-liquid two-phase flow pattern recognition method based on data conversion and multi-head attention mechanism, which is used to realize the recognition of gas-liquid two-phase flow pattern in a horizontal pipe. Background Art

[0002] At present, two-phase flow is widely used in industrial fields such as oil pipelines, chemical reactors, and aerospace. According to the different components, two-phase flow mainly includes gas-liquid two-phase flow, gas-solid two-phase flow, liquid-solid two-phase flow, and liquid-liquid two-phase flow. The accurate identification of two-phase flow pattern is the basis for obtaining flow parameters, and more directly affects the analysis results of two-phase flow resistance characteristics, heat transfer and mass transfer characteristics. Therefore, the research on two-phase flow pattern identification is of great significance to ensure the efficient and safe operation of industrial production systems.

[0003] In order to accurately identify the flow pattern, it is necessary to extract information that characterizes the two-phase flow. However, due to the inherent geometric complexity and variability of the flow structure, it is difficult to accurately characterize and extract the characteristic information. In recent years, commonly used signal processing methods such as power spectral density, wavelet transform, entropy theory and statistical methods have been used to extract the characteristic information of two-phase flow. However, most of these methods process the measurement signals based on static or statistical features, ignore the signal's dependence on time and take a lot of time to calculate. In addition, the flow pattern of two-phase flow can also be identified from image processing. For example, Shanthi et al. used a high-speed digital camera to obtain images of different flow patterns, and then extracted texture features and shape features from the preprocessed images for flow pattern identification. However, in general, most of the actual industrial processes are non-transparent pipes, and it is difficult for the camera to capture the flow view inside the pipe, so this method is limited in application. In order to overcome the above problems, people have proposed to convert the measurement signal into data. For example, WX Liu et al. used a recursive graph algorithm to convert the measurement signal of the conductivity sensor array, and distinguished three typical flow patterns of two-phase flow based on the differences in RP texture structure under different flow patterns. LF Zhang used a pseudo-image coding algorithm to convert the power spectrum into a two-dimensional grayscale image to visualize the characteristic differences of different flow patterns in the frequency domain. Compared with the other two methods, data conversion of complex measurement signals helps to reveal high-level features that are not available in the time series. Therefore, this method is a more intuitive high-dimensional feature representation method. Convolutional neural networks are usually used for feature extraction, because the convolution layer can discover some hidden features in the input features, and the subsequent layers can produce correct outputs. In addition, the multi-head attention mechanism can process information in parallel and capture features of different dimensions. The parallelism of multi-head attention can improve computational efficiency. Considering the advantages of data conversion and multi-head attention mechanism, a high-precision gas-liquid two-phase flow pattern recognition method based on data conversion and multi-head attention mechanism is proposed. The purpose is to more comprehensively extract the complex flow pattern characteristics of gas-liquid two-phase flow in horizontal pipes and improve the accuracy of flow pattern recognition. This method uses a ring-shaped four-electrode conductivity sensor to measure the gas-liquid two-phase flow rate, and then converts the measurement signal into a characteristic matrix. The characteristic matrix is ​​used as input to construct a gas-liquid two-phase flow pattern recognition network, extract the characteristics of the characteristic matrix, and establish a mapping relationship between the characteristic matrix and the flow pattern, and finally realize the recognition of the gas-liquid two-phase flow pattern in the horizontal pipe. There is no related report in this regard. Summary of the invention

[0004] The technical problem solved by the present invention is to provide a high-precision gas-liquid two-phase flow pattern recognition method based on data conversion and multi-head attention mechanism. Compared with the commonly used signal processing methods, the method proposed in the present invention shows superior performance in extracting the flow pattern characteristics of gas-liquid two-phase flow in horizontal pipes, effectively overcoming the problem of difficulty in fully capturing the nonlinear relationship of the two-phase flow pattern signals, and improving the accuracy of flow pattern recognition and reducing the computational cost.

[0005] The present invention adopts the following technical scheme to solve the above technical problems: a high-precision gas-liquid two-phase flow pattern recognition method based on data conversion and multi-head attention mechanism, and the specific process is: after the gas-liquid two-phase flow dynamic experimental device is built, the gas-liquid two-phase flow conductivity fluctuation signals of five flow patterns of bubbly flow, bubbly plug flow, plug flow, plug diversion and stratified flow are collected according to the annular conductivity sensor, the collected gas-liquid two-phase flow conductivity fluctuation signal data is enhanced and then normalized, and then the normalized data is converted into a feature matrix by polar coordinates, and then time series feature pre-extraction is performed, which helps to reveal high-level features that are not in the time series by highlighting the feature differences between different flow patterns, and constructs a gas-liquid two-phase flow pattern recognition network and initializes and trains the gas-liquid two-phase flow pattern recognition network, and inputs feature matrix samples of different flow patterns into the gas-liquid two-phase flow pattern recognition network for time series feature extraction, and its window multi-head self-attention mechanism and shift window multi-head self-attention mechanism can capture local and global features in the feature matrix and thus realize the recognition of the horizontal pipe gas-liquid two-phase flow pattern.

[0006] Furthermore, the construction process of the gas-liquid two-phase flow dynamic experimental device is as follows: an annular conductivity sensor is used to measure the gas-liquid two-phase flow. The annular conductivity sensor is composed of four stainless steel annular electrodes E1, E2, M1, and M2 embedded in the inner wall of a plexiglass tube, wherein E1 and E2 are an excitation electrode pair, M1 and M2 are a measuring electrode pair, the ring spacing between M1 and M2 is 60 mm, the ring spacing between E1 and M1 is 70 mm, and the ring spacing between E2 and M2 is 70 mm. The experimental process adopts a current excitation and voltage measurement working mode. By passing a sinusoidal alternating current with a frequency of 20 kHz into E1 and grounding E2 at the same time, an electrical measurement sensitive field can be formed between the excitation electrode pairs.

[0007] Furthermore, the gas-liquid two-phase flow conductivity fluctuation signal acquisition process is as follows: the two-phase flow is experimentally measured using a ring conductivity sensor, the experimental acquisition frequency is 1kHz, the acquisition time is 20 seconds, that is, a set of measurement voltage signals obtained in one acquisition is 20,000 data points, and the measurement voltage is collected three times under the same flow state to obtain three sets of measurement voltage signals. The flow state is changed, and the acquisition operation is repeated to obtain the measurement voltage signals under multiple flow states, thereby forming a gas-liquid two-phase flow data set. The flow types in the data set include bubbly flow, bubbly plug flow, plug flow, plug flow and stratified flow. In order to enhance the feature representation ability of the model, sufficient training samples are required. The overlapping sampling method is used to process each set of measurement voltage signals in the gas-liquid two-phase flow data set respectively, thereby increasing the number of training samples. The single sampling length is set to 4000, and the moving step is set to 1000. The enhanced sample number calculation formula is: In the formula, Numis the number of samples after data augmentation, Indicates rounding down. S t is the number of data points of a set of measured voltage signals, S is a single sampling length, S step is the moving step length.

[0008] Furthermore, the specific process of data normalization and converting the normalized data into characteristic matrix data by polar coordinates is as follows: for a single sampling length Time series of gas-liquid two-phase flow Normalization is performed so that all data values ​​are limited to between 0 and 1, and the normalized data of the gas-liquid two-phase flow time series is obtained. , the normalized formula is:

[0009] In the formula, ( i =1,2,…, s ) represents the normalized value, min is the minimum value, and max is the maximum value; Will Converted to polar coordinates, the angle θ i and radius r i The calculation process is as follows: In the formula, arccos ( ) represents the arccosine function, t i is the timestamp; Normalize the time series to the interval [0,1] θ i The angle is 0~π / 2; 0< r i ≤1, normalized data In polar coordinates, they are represented as values ​​of different angles and radii within the unit circle.

[0010] Further definition, the specific process of time series feature pre-extraction is: in the normalized data After conversion to polar coordinates, the sum or difference of polar angles between each polar coordinate value is calculated to identify the time correlation in different time intervals, and two polar coordinate matrices are generated, where the cosine function of the sum of two angles is used to obtain Sum Matrix, using the sine function of the difference between the two angles Different Matrix, the specific calculation formula is:

[0011] In the formula, cos( ) represents the cosine function, sin( ) represents the sine function; After calculation, we get two sizes: Each element in the two matrices is linearly transformed and normalized to the range of [0-1]. The calculation expression of this step is:

[0012] In the formula, f ( x ij ) is the value after linear normalization, x ij Represents the value of any position in the matrix, min is the minimum value, max is the maximum value, Indicates rounding down, based on the segment mapping method f ( x ij ) values ​​are mapped to the three-dimensional space of X, Y, and Z; when When , the calculation method is:

[0013] when When , the calculation method is:

[0014] when When , the calculation method is:

[0015] The linearly transformed values ​​are then mapped to the X, Y, and Z spaces and combined into a three-dimensional array to construct and Three-dimensional data, get normalized data of The characteristic matrix is ​​the characteristic matrix sample set that constitutes flow pattern recognition. D 1; Get normalized data of The characteristic matrix is ​​the characteristic matrix sample set 2 that constitutes flow pattern recognition D 2. Set the feature matrix sample set D 1 and feature matrix sample set 2 D 2 is collectively referred to as the feature matrix sample set D .

[0016] Furthermore, the specific process of constructing the gas-liquid two-phase flow pattern recognition network is as follows: the input is the feature matrix sample set D, the output of the network is the recognition accuracy of five flow types. The structure of the recognition network is as follows: first, the feature matrix sample set 1 and the feature matrix sample set 2 are input into the convolution layer 1 in sequence, and the convolution kernel with a size of 4×4 and a step size of 4 is used to perform convolution operation on the feature matrix samples. The input feature dimension is set to 3 and the output feature dimension is 48; then four processing modules are constructed: processing module 1 includes convolution layer 2 and repeatedly stacked attention modules. Convolution layer 2 uses a convolution kernel with a size of 1×1 and a step size of 1. Processing module 2, processing module 3, and processing module 4 all include a pooling layer for downsampling and a repeatedly stacked attention module. Each repeatedly stacked attention module is composed of stacked attention modules. The repeatedly stacked attention modules in processing module 1, processing module 2, processing module 3, and processing module 4 are respectively composed of 2, 2, 6, and 2 attention modules stacked. Each attention module contains eight layers in total: its input layer is a normalization layer, the second layer is a multi-head self-attention mechanism for the window, and the third layer is Normalization layer, the fourth layer is a multi-layer perceptron, which consists of multiple fully connected layers, GELU activation functions and dropout layers. The fifth layer is a normalization layer, the sixth layer is a multi-head self-attention mechanism for displacement windows, the seventh layer is a normalization layer, and the eighth layer is a multi-layer perceptron. The output feature dimensions of each processing module are 96, 192, 384 and 768 respectively; finally, after a normalization layer, a global average pooling layer, a GELU activation function and a fully connected layer, the construction of the recognition network is completed. The window refers to the non-overlapping local area divided by the convolution of the input feature matrix. The multi-head self-attention mechanism of the window refers to the multi-head self-attention mechanism that only performs multi-head self-attention calculations within each window. The displacement window refers to the local area originally divided by the convolution moving in a certain direction, so that there are some overlapping parts between adjacent windows, thereby generating a new window. The multi-head self-attention mechanism of the displacement window refers to the multi-head self-attention mechanism that performs multi-head self-attention calculations within the new window; The specific calculation formula in the identification network is as follows: The calculation method of convolution layer 1 is:

[0017] In the formula, Conv 1 is the output of convolutional layer 1, D k is the input of convolutional layer 1, W 1 represents the weight of convolutional layer 1, b 1 represents the bias of convolutional layer 1; The calculation method of convolution layer 2 is:

[0018] In the formula, Conv 2 is the output of convolutional layer 2, Conv 1 is the input of convolutional layer 2, W2 represents the weight of convolutional layer 2, b 2 represents the bias of convolutional layer 2; The calculation method of the normalization layer in the attention module is:

[0019] In the formula, LN ( x ) is the output of the normalization layer, x is the input of the normalization layer, γ represents the scale parameter, β represents the offset parameter, μ represents the mean of the input feature vector, σ represents the standard deviation of the input feature vector, ε is a very small number used to stabilize calculations; In the calculation process of the multi-head self-attention mechanism, the feature matrix is ​​first linearly projected onto qq In the space, the matrix is ​​divided into three parts, namely Q (query), K (key), V (value) feature matrix, the calculation of this step is expressed as:

[0020] In the formula, W Q , W K and W V For three weight matrices, the self-attention mechanism in each local window can be expressed as:

[0021] In the formula, K T express K The transpose of is the scaling factor, B is the relative position deviation of the window, and softmax is expressed as a normalized exponential function. Q , K , V Linear projection of feature matrix h This constitutes the multi-head self-attention mechanism MSA. The calculation of MSA is as follows:

[0022] In the formula, W o represents the weight parameter of linear projection, Concatenate represents the concatenation function, and the feature matrix size is , C To output feature dimensions, the computational complexity of MSA and WMSA is expressed as:

[0023] In the formula, M Represents the size of each window; The overall calculation process of the attention module is:

[0024] In the formula, MLP stands for multi-layer perceptron. LN is a normalization layer, WMSA represents the multi-head self-attention mechanism of the window, and SWMSA represents the multi-head self-attention mechanism of the displacement window. Output and input of the multi-head self-attention mechanism representing the window The residual connection of The output of the multi-head self-attention mechanism representing the shift window is The residual connection of and is the MLP output feature; The global average pooling layer is calculated as follows:

[0025] In the formula, GAP represents the output of the global average pooling layer, y kpq Indicates k Feature Matrix D i Located in ( p , q ) at the element, | D k | Indicates k The number of all elements of the feature matrix; The GELU activation function is calculated as follows:

[0026] In the formula, G Represents the GELU activation function layer input, and 0.044715 are the two adjustment coefficients of the GELU function. The hyperbolic tangent function tanh() is expressed as:

[0027] In the formula, e represents the natural logarithm, x The input variable represented by tanh(); The calculation method of the fully connected layer is as follows:

[0028] In the formula, FC represents the output of the global average pooling layer, W fc is the weight matrix of the fully connected layer, b fcrepresents the bias of the fully connected layer.

[0029] Furthermore, the specific process of initialization and training of the gas-liquid two-phase flow pattern recognition network is as follows: Network initialization usually requires initializing the weights and bias parameters of each layer of the network: for layer normalization, the weights are usually initialized to 1, and the biases are usually initialized to 0. For the fully connected layer, the weights and biases are usually initialized with small random values. The output of the network is a vector of length 5. V , which represents the probability that the input feature matrix sample belongs to a certain flow type category. The classification network uses the Adam optimizer, the initial learning rate is set to 0.0001, and the network loss function is the cross entropy function. The calculation method is as follows:

[0030] In the formula, Γ represents the number of samples, Λ represents the number of categories, y ρυ Indicates ρ Whether the sample belongs to the category υ The true label of ρ Belongs to category υ , then y ρυ =1, otherwise y ρυ =0, The model predicts ρ samples belong to the sample category υ The classification network performs back propagation according to the Adam optimizer to update the input layer weights, the weights from the intermediate layer to the output layer, the bias of the intermediate layer, and the bias of the output layer. The input feature matrix sample set is forward propagated through the convolution layer, pooling layer, global average pooling layer, activation function, and fully connected layer in the network structure, and the cross entropy loss between the recognized feature matrix flow type category and the real feature matrix flow type category is calculated. The gradient is calculated according to the loss function and back propagated through the network to update the network weights. The training cycle is set to 50 epochs. The evaluation index process of the gas-liquid two-phase flow pattern recognition network is as follows: The evaluation of the model uses Accuracy, Recall, Precision and F1 score as indicators to measure its performance. The formula is as follows:

[0031] In the formula, TP is the number of true positive examples, FN is the number of false negatives, FP is the number of false positives, TN is the number of true negative examples.

[0032] Compared with the prior art, the present invention has the following advantages and beneficial effects: The present invention proposes a high-precision gas-liquid two-phase flow pattern recognition method based on data conversion and multi-head attention mechanism. The method converts the gas-liquid two-phase flow conductivity fluctuation signal into polar coordinates, performs feature pre-extraction on it to obtain a three-dimensional feature matrix, and can highlight the feature differences between different flow patterns. The present invention constructs a gas-liquid two-phase flow pattern recognition network to convert the feature matrix sample set into a D The input is input into the model for feature extraction. The multi-head self-attention mechanism of the window and the multi-head self-attention mechanism of the shifted window can capture the local and global features in the image, thereby realizing flow pattern recognition. The results show that the high-precision gas-liquid two-phase flow pattern recognition method based on data conversion and multi-head attention mechanism proposed in the present invention can effectively overcome the problem of difficulty in fully capturing the nonlinear relationship of the two-phase flow pattern signal, more comprehensively extract the complex flow pattern characteristics of the horizontal pipe gas-liquid two-phase flow, and effectively improve the accuracy of the horizontal pipe gas-liquid two-phase flow pattern recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a flowchart of the overall method principle of the present invention; Figure 2 It is a basic structural schematic diagram of the annular four-electrode conductivity sensor in the present invention; Figure 3 It is a basic structural schematic diagram of the horizontal pipeline gas-liquid two-phase flow dynamic experimental system of the present invention; Figure 4 It is a schematic diagram of the process of feature pre-extraction for gas-liquid two-phase flow time series; Figure 5 It is the characteristic matrix diagram of five flow types; Figure 6 A structural diagram of a gas-liquid two-phase flow pattern recognition network constructed in the present invention; Figure 7 It is the recognition accuracy of a single flow type when it is recognized under the three evaluation methods of Recall, Precision and F1 score; Figure 8 This is a confusion matrix diagram of the gas-liquid two-phase flow pattern recognition network test results in the present invention. DETAILED DESCRIPTION

[0034] In conjunction with the accompanying drawings, the present invention provides a high-precision gas-liquid two-phase flow pattern recognition method based on data conversion and multi-head attention mechanism, which is described in detail.

[0035] The high-precision gas-liquid two-phase flow pattern recognition method based on data conversion and multi-head attention mechanism described in the present invention shows superior performance in extracting flow pattern characteristics of gas-liquid two-phase flow in horizontal pipes, effectively overcoming the problem of difficulty in fully capturing the nonlinear relationship of two-phase flow pattern signals, and improving the accuracy of flow pattern recognition and reducing computational costs.

[0036] like Figure 1 As shown, it is a flowchart of a high-precision gas-liquid two-phase flow pattern recognition method based on data conversion and multi-head attention mechanism provided by the present invention, and the specific steps are as follows: Step 1: sensor structure design. The present invention adopts an annular conductivity sensor for measuring gas-liquid two-phase flow. The annular conductivity sensor is composed of four stainless steel annular electrodes E1, E2, M1, and M2 embedded in the inner wall of a plexiglass tube, wherein E1 and E2 are an excitation electrode pair, and M1 and M2 are a measuring electrode pair. The ring spacing between M1 and M2 is 60mm, the ring spacing between E1 and M1 is 70mm, and the ring spacing between E2 and M2 is 70mm. The experiment of the present invention adopts a current excitation and voltage measurement working mode. By passing an alternating current with a sinusoidal wave frequency of 20kHz into electrode E1 and grounding electrode E2 at the same time, an electrical measurement sensitive field can be formed between the excitation electrode pair.

[0037] Step 2: Collecting the conductivity fluctuation signal of gas-liquid two-phase flow. The horizontal pipeline gas-liquid two-phase dynamic detection device is mainly composed of a water tank, an air pump, a flow meter, a phase mixer, a ring conductivity sensor and a computer. In the experiment, the liquid phase and the gas phase are tap water (density 998kg / m 3 , the dynamic viscosity is 1.01×10 -3 Pa·s) and air (density 1.2kg / m 3 , the dynamic viscosity is 1.81×10 -5 In the dynamic experiment, in order to obtain more comprehensive data, the air flow rate was controlled at [0 m 3 / h ~362 m 3 / h], the water flow rate is controlled within [3 m 3 / h ~ 11m 3 / h] range. The method of adjusting the gas speed with a fixed water speed is adopted, that is, water with a fixed speed is first introduced into the horizontal pipe through the water tank, and then the gas phase is gradually increased through the air pump. The speed of air and water can be set by the regulating valve, so as to match different flow states. When the flow state of the two-phase flow is stable, the measured voltage data is recorded. The two-phase flow is experimentally measured using an annular conductivity sensor. The experimental acquisition frequency is 1kHz, and the acquisition time is 20 seconds, that is, a set of measured voltage signals obtained in one acquisition is 20,000 data points. Under the same flow state, three sets of measured voltage signals can be obtained by collecting the measured voltage three times. By changing the flow state and repeating the acquisition operation, the measured voltage signals under multiple flow states can be obtained, thereby forming a gas-liquid two-phase flow data set. The flow types in the data set include bubbly flow, bubble plug flow, plug flow, plug diversion and stratified flow.

[0038] Step 3: Data enhancement. In order to enhance the feature representation ability of the model, sufficient training samples are required. The overlapping sampling method is used to process each set of measured voltage signals in the gas-liquid two-phase flow data set separately, thereby increasing the number of training samples. The single sampling length is set to 4000, the moving step is set to 1000, and the formula for calculating the number of samples after enhancement is:

[0039] In the formula, Num is the number of samples after data augmentation, Indicates rounding down. S t is the number of data points of a set of measured voltage signals, S is a single sampling length, S step is the moving step length.

[0040] Step 4: Data normalization. For a single sampling length Time series of gas-liquid two-phase flow Normalization is performed so that all data values ​​are limited to between 0 and 1, and the normalized data of the gas-liquid two-phase flow time series is obtained. , the normalized formula is:

[0041] In the formula, Represents the normalized value, min is the minimum value, and max is the maximum value.

[0042] Step 5: Convert to polar coordinates. Angle and radius The transformation process is as follows:

[0043] In the formula, t i Is the timestamp.

[0044] When the time series is mapped to [0,1], θ i is mapped to [0,π / 2]. Therefore, as time increases, It can be expressed as polar coordinates at different angles and radii within the unit circle.

[0045] Step 6: Pre-extraction of time series features. After conversion to polar coordinates, the sum or difference of polar angles between each polar coordinate value is calculated to identify the time correlation in different time intervals, and two polar coordinate matrices can be generated. The cosine function of the sum of two angles is used to obtain Sum Matrix, using the sine function of the difference between the two angles Different Matrix, the specific calculation formula is:

[0046] Where cos( ) represents the cosine function and sin( ) represents the sine function.

[0047] After calculation, we can get two sizes of Each element in the two matrices is linearly transformed and normalized to the range of [0-1]. The calculation expression of this step is

[0048] In the formula, f ( x ij ) is the value after linear normalization, x ij Represents the value of any position in the matrix, min is the minimum value, max is the maximum value, Indicates rounding down. Based on the segment mapping method, f ( x ij ) values ​​are mapped to the X, Y, and Z three-dimensional space.

[0049] when When , the calculation method is

[0050] when When , the calculation method is:

[0051] when When , the calculation method is: .

[0052] The linearly transformed values ​​are then mapped to the X, Y, and Z spaces and combined into a three-dimensional array. and Three-dimensional data, get normalized data of The characteristic matrix can constitute the characteristic matrix sample set of flow pattern recognition. D 1; Get normalized data of The characteristic matrix can constitute the characteristic matrix sample set 2 of flow pattern recognition D 2. Set the feature matrix sample set D 1 and feature matrix sample set 2 D 2 is collectively referred to as the feature matrix sample set D .

[0053] Step 7: Construct a gas-liquid two-phase flow pattern recognition network. Input is the feature matrix sample set D . The output of the network is the recognition accuracy of the five flow types. The structure of the recognition network is as follows: first, the feature matrix sample set 1 and the feature matrix sample set 2 are input into the convolution layer 1 in sequence, and the convolution kernel with a size of 4×4 and a step size of 4 is used to perform convolution operations on the feature matrix samples. The input feature dimension is set to 3 and the output feature dimension is 48; then, four processing modules are constructed: processing module 1 includes convolution layer 2 and repeatedly stacked attention modules. Convolution layer 2 uses a convolution kernel with a size of 1×1 and a step size of 1. Processing module 2, processing module 3, and processing module 4 all include a pooling layer for downsampling and a repeatedly stacked attention module. Each repeatedly stacked attention module is composed of stacked attention modules. The repeatedly stacked attention modules in processing module 1, processing module 2, processing module 3, and processing module 4 are respectively composed of 2, 2, 6, and 2 stacked attention modules. Each attention module contains eight layers: its input layer is a normalization layer, the second layer is a multi-head self-attention mechanism for the window, the third layer is a normalization layer, and the fourth layer is a multi-layer perceptron, which consists of multiple fully connected layers, GELU activation functions, and dropout layers. The fifth layer is a normalization layer, the sixth layer is a multi-head self-attention mechanism for the displacement window, the seventh layer is a normalization layer, and the eighth layer is a multi-layer perceptron. The output feature dimensions of each processing module are 96, 192, 384, and 768 respectively; finally, after a normalization layer, a global average pooling layer, a GELU activation function, and a fully connected layer, the recognition network can be constructed. The window refers to the non-overlapping local area divided by the convolution of the input feature matrix, and the multi-head self-attention mechanism of the window refers to the multi-head self-attention mechanism that only performs multi-head self-attention calculations within each window. The displacement window refers to the local area originally divided by the convolution moving in a certain direction, so that there is a partial overlap between adjacent windows, thus generating a new window. The multi-head self-attention mechanism of the displacement window refers to the multi-head self-attention mechanism performing multi-head self-attention calculations inside the new window.

[0054] In the identification network, the specific calculation formula is as follows: The calculation method of convolution layer 1 is:

[0055] In the formula, Conv 1 is the output of convolutional layer 1, D k is the input of convolutional layer 1, W 1 represents the weight of convolutional layer 1, b 1 represents the bias of convolutional layer 1.

[0056] The calculation method of convolution layer 2 is:

[0057] In the formula, Conv 2 is the output of convolutional layer 2, Conv 1 is the input of convolutional layer 2, W 2 represents the weight of convolutional layer 2, b 2 represents the bias of convolutional layer 2.

[0058] In the attention module, the calculation method of the normalization layer is:

[0059] In the formula, LN ( x ) is the output of the normalization layer, x is the input of the normalization layer, γ represents the scale parameter, β represents the offset parameter, μ represents the mean of the input feature vector, σ represents the standard deviation of the input feature vector, ε is a very small number used to stabilize calculations.

[0060] In the calculation process of the multi-head self-attention mechanism, the feature matrix is ​​first linearly projected onto qq In the space, the matrix is ​​divided into three parts, namely Q (query), K (key), V (value) feature matrix. The calculation of this step can be expressed as:

[0061] In the formula, W Q , W K and W V are three weight matrices. The self-attention mechanism in each local window can be expressed as:

[0062] In the formula, K T express K The transpose of is the scaling factor, Bis the relative position deviation of the window, and softmax is expressed as a normalized exponential function. Q , K , V Linear projection of feature matrix h This constitutes the multi-head self-attention mechanism MSA. The calculation of MSA is as follows:

[0063] In the formula, W o represents the weight parameter of linear projection, and Concatenate represents the concatenation function. The feature matrix size is ℎ×𝑤 , C To output feature dimensions, the computational complexity of MSA and WMSA is expressed as:

[0064] In the formula, M Represents the size of each window.

[0065] The overall calculation process of the attention module is:

[0066] In the formula, MLP stands for multi-layer perceptron. LN is a normalization layer, WMSA represents the multi-head self-attention mechanism of the window, and SWMSA represents the multi-head self-attention mechanism of the displacement window. Output and input of the multi-head self-attention mechanism representing the window The residual connection of The output of the multi-head self-attention mechanism representing the shift window is The residual connection of and is the MLP output feature.

[0067] The global average pooling layer is calculated as follows:

[0068] In the formula, GAP represents the output of the global average pooling layer, y kpq Indicates k Feature Matrix D i Located in ( p , q ) at the element, | D k | Indicates k The number of all elements in the feature matrix.

[0069] The GELU activation function is calculated as follows:

[0070] In the formula, G Represents the GELU activation function layer input, and 0.044715 are the two adjustment coefficients of the GELU function. The hyperbolic tangent function tanh() is expressed as:

[0071] In the formula, e represents the natural logarithm, x Represented as the input variable of tanh().

[0072] The calculation method of the fully connected layer is as follows:

[0073] In the formula, FC represents the output of the global average pooling layer, W fc is the weight matrix of the fully connected layer, b fc represents the bias of the fully connected layer.

[0074] Step 8: Initialization and training of the gas-liquid two-phase flow pattern recognition network. Network initialization usually requires initializing the weights and bias parameters of each layer of the network: for layer normalization, the weights are usually initialized to 1, and the biases are usually initialized to 0. For the fully connected layer, the weights and biases are usually initialized with small random values. The output of the network is a vector of length 5 V , which represents the probability that the input feature matrix sample belongs to a certain flow type category. The classification network uses the Adam optimizer, and the initial learning rate is set to 0.0001. The network loss function is the cross entropy function, and the calculation method is as follows:

[0075] In the formula, Γ represents the number of samples, Λ represents the number of categories, y ρυ Indicates ρ Whether the sample belongs to the category υ The true label of ρ Belongs to category υ , then y ρυ =1, otherwise y ρυ =0. The model predicts ρ samples belong to the sample category υThe classification network performs back propagation according to the Adam optimizer to update the input layer weights, the weights from the intermediate layer to the output layer, the bias of the intermediate layer, and the bias of the output layer. The input feature matrix sample set is forward propagated through the convolution layer, pooling layer, global average pooling layer, activation function, and fully connected layer in the network structure, and the cross entropy loss between the recognized feature matrix flow type category and the true feature matrix flow type category is calculated. The gradient is calculated according to the loss function and back propagated through the network to update the network weights. The training cycle is set to 50 epochs.

[0076] Step 9: Evaluation index of gas-liquid two-phase flow pattern recognition network. The evaluation of this model uses Accuracy, Recall, Precision and F1 score as indicators to measure its performance. The formula is as follows:

[0077] In the formula, TP is the number of true positive examples, FN is the number of false negatives, FP is the number of false positives, TN is the number of true negative examples.

[0078] Finally, the characteristic matrix samples of the gas-liquid two-phase flow are input into the trained and tested gas-liquid two-phase flow pattern recognition network to identify five flow patterns: bubbly flow, bubbly plug flow, plug flow, plug flow and stratified flow.

[0079] like Figure 2 The basic structure of the sensor is shown in FIG. 1 . The present invention adopts an annular conductivity sensor for measuring gas-liquid two-phase flow. The annular conductivity sensor is composed of four stainless steel annular electrodes E1, E2, M1, and M2 embedded in the inner wall of a plexiglass tube, wherein E1 and E2 are an excitation electrode pair, and M1 and M2 are a measuring electrode pair. The distance between the rings of M1 and M2 is 60 mm, the distance between the rings of E1 and M1 is 70 mm, and the distance between the rings of E2 and M2 is 70 mm.

[0080] like Figure 3 The figure shows the basic structure of the horizontal pipeline gas-liquid two-phase flow dynamic experimental system. The basic structure of the gas-liquid two-phase flow dynamic experimental system mainly consists of a ring electrode sensor, a water tank, an air pump, a mixer, a horizontal pipe test unit and a data acquisition system.

[0081] like Figure 4 As shown, this is a schematic diagram of the process of feature pre-extraction of gas-liquid two-phase flow time series. Figure 4 (a) is a schematic diagram of data normalization. Figure 4 (b) is a schematic diagram of the transformation of normalized data into polar coordinate system. Figure 4 (c) is generated and The result diagram of the feature matrix. Figure 6 Shown is the structural diagram of the gas-liquid two-phase flow pattern identification network. Figure 5 It is the characteristic matrix diagram of five flow types.

[0082] like Figure 7 As shown in the figure, it is the recognition accuracy of a single flow type under the three evaluation methods of Recall, Precision and F1 score. It can be seen from the figure that in the radar chart, the Markov transition field, the recursive graph, and The four data conversion methods are represented by different colors. The evaluation results of different flow patterns are connected to form polygons. The closer the polygon is to the black polygon, the higher the recognition rate. Different feature matrices are input into the gas-liquid two-phase flow pattern recognition network. The Recall, Precision and F1 score accuracy rates of bubbly flow, plug flow and stratified flow tend to 100%. and The evaluation results of the plug flow and plug split flow under the condition of are better than those of the Markov transition field and the recursive graph. and The data conversion method has a higher accuracy in flow pattern identification.

[0083] like Figure 8 As shown in Figure 1, the confusion matrix diagram of the gas-liquid two-phase flow pattern recognition network test results. Figure 8 As can be seen in (a), the input is When the characteristic matrix is ​​used, the accuracy rates of bubbly flow, bubbly plug flow, plug flow, plug flow and stratified flow are 100%, 97.8%, 100%, 96.7% and 100% respectively. Figure 8 As can be seen in (b), the input is When the characteristic matrix is ​​used, the corresponding accuracy values ​​of the five flow types are 100%, 98.9%, 100%, 97.5% and 100% respectively. The results show that the recognition of bubbly flow, slug flow and stratified flow is relatively accurate. In contrast, due to the transition flow, the accuracy of identifying bubbly flow and slug flow is slightly lower.

[0084] As shown in Table 1, the gas-liquid two-phase flow pattern recognition network in the present invention and the other four models are identified by the characteristic matrix sample set D 1 and D 2 Results of flow pattern recognition. As can be seen from the table, the recognition accuracy of the gas-liquid two-phase flow pattern recognition network proposed in this invention is better than that of the four networks VGG-16, GoogLeNe, DenseNet, and Vision Transformer (ViT). The inputs are the feature matrix samples D 1 and feature matrix sample D2, the flow pattern recognition rates displayed by the gas-liquid two-phase flow pattern recognition network proposed in the present invention are 98.6% and 99.1% respectively.

[0085]

[0086] It can be seen that the feature pre-extraction method proposed in the present invention is better than the Markov transition field and recursive graph. and The feature matrix can extract more effective flow pattern features. The gas-liquid two-phase flow pattern recognition network constructed by the present invention has a higher recognition accuracy than the four networks of VGG-16, GoogLeNe, DenseNet, and Vision Transformer (ViT). This also further confirms that the method proposed by the present invention can more comprehensively extract the complex flow pattern features of the gas-liquid two-phase flow in the horizontal pipe and improve the accuracy of flow pattern recognition.

Claims

1. A high-precision gas-liquid two-phase flow pattern recognition method based on data conversion and multi-head attention mechanism, characterized in that The specific process is as follows: after the gas-liquid two-phase flow dynamic experimental device is built, the gas-liquid two-phase flow conductivity fluctuation signals of five flow types, namely bubbly flow, bubbly plug flow, plug flow, plug diversion and stratified flow, are collected according to the annular conductivity sensor. The collected gas-liquid two-phase flow conductivity fluctuation signal data is enhanced and then normalized. The normalized data is converted into a feature matrix by polar coordinates, and then time series feature pre-extraction is performed. By highlighting the feature differences between different flow types, it is helpful to reveal high-level features that are not in the time series. A gas-liquid two-phase flow pattern recognition network is constructed and initialized and trained. The feature matrix samples of different flow types are input into the gas-liquid two-phase flow pattern recognition network for time series feature extraction. The multi-head self-attention mechanism of the window and the multi-head self-attention mechanism of the shifted window can capture the local and global features in the feature matrix and thus realize the recognition of the horizontal tube gas-liquid two-phase flow pattern.

2. The high-precision gas-liquid two-phase flow pattern recognition method based on data conversion and multi-head attention mechanism according to claim 1 is characterized in that The construction process of the gas-liquid two-phase flow dynamic experimental device is as follows: an annular conductivity sensor is used to measure the gas-liquid two-phase flow. The annular conductivity sensor consists of four stainless steel annular electrodes E1, E2, M1, and M2 embedded in the inner wall of a plexiglass tube, where E1 and E2 are an excitation electrode pair, M1 and M2 are a measuring electrode pair, the ring spacing between M1 and M2 is 60mm, the ring spacing between E1 and M1 is 70mm, and the ring spacing between E2 and M2 is 70mm. The experimental process adopts a current excitation and voltage measurement working mode. By passing a sinusoidal alternating current with a frequency of 20kHz into E1 and grounding E2 at the same time, an electrical measurement sensitive field can be formed between the excitation electrode pairs.

3. The high-precision gas-liquid two-phase flow pattern recognition method based on data conversion and multi-head attention mechanism according to claim 1 is characterized in that The acquisition process of the conductivity fluctuation signal of the gas-liquid two-phase flow is as follows: the two-phase flow is experimentally measured using a ring conductivity sensor. The experimental acquisition frequency is 1kHz and the acquisition time is 20 seconds, that is, a set of measured voltage signals obtained in one acquisition is 20,000 data points. The measured voltage is collected three times under the same flow state to obtain three sets of measured voltage signals. The flow state is changed and the acquisition operation is repeated to obtain the measured voltage signals under multiple flow states, thereby forming a gas-liquid two-phase flow data set. The flow types in the data set include bubbly flow, bubbly plug flow, plug flow, plug flow and stratified flow. In order to enhance the feature representation ability of the model, sufficient training samples are required. The overlapping sampling method is used to process each set of measured voltage signals in the gas-liquid two-phase flow data set respectively, thereby increasing the number of training samples. The single sampling length is set to 4000 and the moving step is set to 1000. The calculation formula for the number of samples after enhancement is: In the formula, Num is the number of samples after data augmentation, Indicates rounding down. S t is the number of data points of a set of measured voltage signals, S is a single sampling length, S step is the moving step length.

4. The high-precision gas-liquid two-phase flow pattern recognition method based on data conversion and multi-head attention mechanism according to claim 1 is characterized in that The specific process of data normalization and converting normalized data into characteristic matrix data by polar coordinates is as follows: Time series of gas-liquid two-phase flow Normalization is performed so that all data values ​​are limited to between 0 and 1, and the normalized data of the gas-liquid two-phase flow time series is obtained. , the normalized formula is: In the formula, Represents the normalized value, min is the minimum value, and max is the maximum value; Will Converted to polar coordinates, the angle θ i and radius r i The calculation process is as follows: In the formula, arccos ( ) represents the arccosine function, t i is the timestamp; Normalize the time series to the interval [0,1] θ i The angle is 0~π / 2; 0< r i ≤1, normalized data In polar coordinates, they are represented as values ​​of different angles and radii within the unit circle.

5. The high-precision gas-liquid two-phase flow pattern recognition method based on data conversion and multi-head attention mechanism according to claim 1 is characterized in that The specific process of time series feature pre-extraction is as follows: After conversion to polar coordinates, the sum or difference of polar angles between each polar coordinate value is calculated to identify the time correlation in different time intervals, and two polar coordinate matrices are generated, where the cosine function of the sum of two angles is used to obtain Sum Matrix, using the sine function of the difference between the two angles Different Matrix, the specific calculation formula is: In the formula, cos( ) represents the cosine function, sin( ) represents the sine function; After calculation, we get two sizes: s × s Each element in the two matrices is linearly transformed and normalized to the range of [0-1]. The calculation expression of this step is: In the formula, f ( x ij ) is the value after linear normalization, x ij Represents the value of any position in the matrix, min is the minimum value, max is the maximum value, Indicates rounding down, based on the segment mapping method f ( x ij ) values ​​are mapped to the three-dimensional space of X, Y, and Z; when f ( x ij )<0.3, the calculation method is: when When , the calculation method is when When , the calculation method is: The linearly transformed values ​​are then mapped to the X, Y, and Z spaces and combined into a three-dimensional array to construct and Three-dimensional data, get normalized data of The characteristic matrix is ​​the characteristic matrix sample set that constitutes flow pattern recognition. D 1; Get normalized data of The characteristic matrix is ​​the characteristic matrix sample set 2 that constitutes flow pattern recognition D 2. Set the feature matrix sample set D 1 and feature matrix sample set 2 D 2 is collectively referred to as the feature matrix sample set D .

6. The high-precision gas-liquid two-phase flow pattern recognition method based on data conversion and multi-head attention mechanism according to claim 1 is characterized in that The specific process of constructing the gas-liquid two-phase flow pattern recognition network is as follows: the input is the feature matrix sample set D , the output of the network is the recognition accuracy of five flow types. The structure of the recognition network is as follows: first, the feature matrix sample set 1 and the feature matrix sample set 2 are input into the convolution layer 1 in sequence, and the convolution kernel with a size of 4×4 and a step size of 4 is used to perform convolution operation on the feature matrix samples. The input feature dimension is set to 3 and the output feature dimension is 48; then four processing modules are constructed: processing module 1 includes convolution layer 2 and repeatedly stacked attention modules. Convolution layer 2 uses a convolution kernel with a size of 1×1 and a step size of 1. Processing module 2, processing module 3, and processing module 4 all include a pooling layer for downsampling and a repeatedly stacked attention module. Each repeatedly stacked attention module is composed of stacked attention modules. The repeatedly stacked attention modules in processing module 1, processing module 2, processing module 3, and processing module 4 are respectively composed of 2, 2, 6, and 2 attention modules stacked. Each attention module contains eight layers in total: its input layer is a normalization layer, the second layer is a multi-head self-attention mechanism for the window, and the third layer is Normalization layer, the fourth layer is a multi-layer perceptron, which consists of multiple fully connected layers, GELU activation functions and dropout layers. The fifth layer is a normalization layer, the sixth layer is a multi-head self-attention mechanism for displacement windows, the seventh layer is a normalization layer, and the eighth layer is a multi-layer perceptron. The output feature dimensions of each processing module are 96, 192, 384 and 768 respectively; finally, after a normalization layer, a global average pooling layer, a GELU activation function and a fully connected layer, the construction of the recognition network is completed. The window refers to the non-overlapping local area divided by the convolution of the input feature matrix. The multi-head self-attention mechanism of the window refers to the multi-head self-attention mechanism that only performs multi-head self-attention calculations within each window. The displacement window refers to the local area originally divided by the convolution moving in a certain direction, so that there are some overlapping parts between adjacent windows, thereby generating a new window. The multi-head self-attention mechanism of the displacement window refers to the multi-head self-attention mechanism that performs multi-head self-attention calculations within the new window; The specific calculation formula in the identification network is as follows: The calculation method of convolution layer 1 is: In the formula, Conv 1 is the output of convolutional layer 1, D k is the input of convolutional layer 1, W 1 represents the weight of convolutional layer 1, b 1 represents the bias of convolutional layer 1; The calculation method of convolution layer 2 is: In the formula, Conv 2 is the output of convolutional layer 2, Conv 1 is the input of convolutional layer 2, W 2 represents the weight of convolutional layer 2, b 2 represents the bias of convolutional layer 2; The calculation method of the normalization layer in the attention module is: In the formula, LN ( x ) is the output of the normalization layer, x is the input of the normalization layer, γ represents the scale parameter, β represents the offset parameter, μ represents the mean of the input feature vector, σ represents the standard deviation of the input feature vector, ε is a very small number used to stabilize calculations; In the calculation process of the multi-head self-attention mechanism, the feature matrix is ​​first linearly projected onto qq In the space, the matrix is ​​divided into three parts, namely , , The feature matrix is ​​calculated as follows: In the formula, W Q , W K and W V For three weight matrices, the self-attention mechanism in each local window can be expressed as: In the formula, K T express K The transpose of is the scaling factor, B is the relative position deviation of the window, and softmax is expressed as a normalized exponential function. Q , K , V Linear projection of feature matrix h This constitutes the multi-head self-attention mechanism MSA. The calculation of MSA is as follows: In the formula, W o represents the weight parameter of linear projection, Concatenate represents the concatenation function, and the feature matrix size is , C To output feature dimensions, the computational complexity of MSA and WMSA is expressed as: In the formula, M Represents the size of each window; The overall calculation process of the attention module is: In the formula, MLP stands for multi-layer perceptron. LN is a normalization layer, WMSA represents the multi-head self-attention mechanism of the window, and SWMSA represents the multi-head self-attention mechanism of the displacement window. Output and input of the multi-head self-attention mechanism representing the window The residual connection of The output of the multi-head self-attention mechanism representing the shift window is The residual connection of and is the MLP output feature; The global average pooling layer is calculated as follows: In the formula, GAP represents the output of the global average pooling layer, y kpq Indicates k Feature Matrix D i Located in ( p , q ) at the element, | D k | Indicates k The number of all elements of the feature matrix; The calculation method of GELU activation function is as follows: In the formula, G Represents the GELU activation function layer input, and 0.044715 are the two adjustment coefficients of the GELU function. The hyperbolic tangent function tanh() is expressed as: In the formula, e represents the natural logarithm, x The input variable represented by tanh(); The calculation method of the fully connected layer is as follows: In the formula, FC represents the output of the global average pooling layer, W fc is the weight matrix of the fully connected layer, b fc represents the bias of the fully connected layer.

7. The high-precision gas-liquid two-phase flow pattern recognition method based on data conversion and multi-head attention mechanism according to claim 1 is characterized in that The specific process of initialization and training of the gas-liquid two-phase flow pattern recognition network is as follows: Network initialization usually requires initializing the weights and bias parameters of each layer of the network: For layer normalization, the weights are usually initialized to 1, and the biases are usually initialized to 0. For the fully connected layer, the weights and biases are usually initialized with small random values. The output of the network is a vector of length 5. V , which represents the probability that the input feature matrix sample belongs to a certain flow type category. The classification network uses the Adam optimizer, the initial learning rate is set to 0.0001, and the network loss function is the cross entropy function. The calculation method is as follows: In the formula, Γ represents the number of samples, Λ represents the number of categories, y ρυ Indicates ρ Whether the sample belongs to the category υ The true label of ρ Belongs to category υ , then y ρυ =1, otherwise y ρυ =0, The model predicts ρ samples belong to the sample category υ The classification network performs back propagation according to the Adam optimizer to update the input layer weights, the weights from the intermediate layer to the output layer, the bias of the intermediate layer, and the bias of the output layer. The input feature matrix sample set is forward propagated through the convolution layer, pooling layer, global average pooling layer, activation function, and fully connected layer in the network structure, and the cross entropy loss between the recognized feature matrix flow type category and the real feature matrix flow type category is calculated. The gradient is calculated according to the loss function and back propagated through the network to update the network weights. The training cycle is set to 50 epochs. The evaluation index process of the gas-liquid two-phase flow pattern recognition network is as follows: The evaluation of the model uses Accuracy, Recall, Precision and F1 score as indicators to measure its performance. The formula is as follows: In the formula, TP is the number of true positive examples, FN is the number of false negatives, FP is the number of false positives, TN is the number of true negative examples.

Images