A method for identifying gas-liquid two-phase flow patterns based on DAS logging data

Abstract

The present invention provides a method for identifying the flow pattern of a gas-liquid two-phase flow based on DAS logging data, which belongs to the technical field of gas-liquid two-phase flow pattern identification and solves the problems of how to accurately extract characteristic information related to the flow pattern from complex acoustic signals and how to build an efficient identification model to process these data. The specific steps included in the present invention include S1 signal acquisition, S2 preprocessing and S3 flow pattern identification. The present invention utilizes distributed fiber acoustic sensing technology (DAS) to continuously collect downhole dynamic data, combines wavelet transform to extract time-frequency images, and uses a hybrid deep neural network (combining CNN and LSTM) to further enhance the ability to accurately extract and identify flow patterns from complex signals, thereby ensuring real-time and efficient flow pattern classification.

Description

Technical Field

[0001] The present invention belongs to the technical field of gas-liquid two-phase flow pattern identification, and in particular to a gas-liquid two-phase flow pattern identification method based on DAS well logging data. Background Art

[0002] In many modern industrial fields, such as petrochemicals, oil and gas storage and transportation, aerospace, and power engineering, dynamic monitoring of gas-liquid two-phase flow is of vital importance for ensuring production safety and improving economic benefits. However, existing flow pattern identification methods, such as flow pattern plate identification, direct measurement, and indirect measurement methods, all have certain limitations. Although the flow pattern plate identification method is based on experimental data, its scope of application and reliability are limited; the direct measurement method relies on manual observation, which is highly subjective and difficult to apply in actual production environments; the indirect measurement method relies on logging equipment to collect data for flow pattern identification, but is limited by instrument performance and extreme environmental conditions, and faces more challenges, especially in highly deviated wells, horizontal wells, and ultra-deep wells.

[0003] In recent years, the rapid development of distributed fiber optic sensing technology has provided new technical means for dynamic downhole monitoring of oil and gas wells. Distributed fiber acoustic sensing (DAS) technology, in particular, is capable of operating in high-temperature, high-pressure, and corrosive environments, enabling real-time monitoring of downhole fluid flow. Despite this, using DAS data for flow pattern identification still faces challenges, including accurately extracting flow pattern-related characteristic information from complex acoustic signals and building efficient identification models to process this data.

[0004] Furthermore, despite progress in flow pattern recognition using machine learning and deep learning techniques, existing methods such as support vector machines and artificial neural networks still exhibit insufficient recognition accuracy when processing high-dimensional, time-series data. In the field of deep learning, while convolutional neural networks (CNNs) and long short-term memory networks (LSTMs) offer new perspectives for feature extraction and time series analysis, effectively integrating these advanced technologies into DAS data analysis to achieve higher-precision and real-time flow pattern recognition remains a critical issue. Summary of the Invention

[0005] The purpose of the present invention is to address the above problems in the existing technology and propose a gas-liquid two-phase flow pattern identification method based on DAS logging data.

[0006] The object of the present invention can be achieved by the following technical solution: a method for identifying gas-liquid two-phase flow patterns based on DAS logging data, the identification method comprising the following steps:

[0007] S1 signal acquisition: The microstructured optical fibers are evenly arranged along the flow cross section of the wellbore, and the DAS acquisition system is used to continuously collect the DAS signals of the gas-liquid two-phase flow passing through the flow cross section of the wellbore through the microstructured optical fibers;

[0008] The design of microstructured optical fibers, such as photonic crystal fibers, allows them to have specific transmission properties, including a high degree of nonlinearity and tunable dispersion and modal properties. This makes microstructured fibers particularly useful in distributed acoustic sensing (DAS) systems because they can improve the sensing sensitivity and signal-to-noise ratio of acoustic events. In addition, these fibers can operate stably over a wide frequency band, enhancing the DAS system's ability to capture various acoustic fluctuations in complex downhole environments, thereby improving the accuracy and reliability of gas-liquid two-phase flow pattern identification.

[0009] S2 preprocessing: The DAS signal of the gas-liquid two-phase flow is converted using the continuous wavelet transform method to obtain the time-frequency image of the gas-liquid two-phase flow;

[0010] Continuous wavelet transform (CWT) is suitable for analyzing non-stationary signals (the statistical characteristics of the signal change with time). It can reveal the transient characteristics and spectral characteristics generated by different flow patterns and provide rich feature information for subsequent flow pattern identification.

[0011] S3 Flow Pattern Recognition: S3.1 Model Establishment: A hybrid deep neural network flow pattern recognition model is established using a convolutional neural network (CNN) and a long short-term memory (LSTM) recurrent neural network. The hybrid deep neural network flow pattern recognition model includes an input layer, a first convolutional layer, a first batch normalization layer, a first Relu activation layer, a first maximum pooling layer, a first Dropout layer, a second convolutional layer, a second batch normalization layer, a second Relu activation layer, a second maximum pooling layer, a second Dropout layer, a Flatten layer, an LSTM layer, a first fully connected layer, a third Dropout layer, a second fully connected layer, a fourth Dropout layer, a third fully connected layer, a Softmax layer, and a classification layer.

[0012] S3.2 Model Improvement: S3.2.1 Hybrid Deep Neural Network Model Training: Collect time-frequency image samples of gas-liquid two-phase flow under different flow patterns; input the time-frequency image samples into the hybrid deep neural network model; use the hybrid deep neural network model to extract flow pattern features of the time-frequency image samples, classify the extracted flow pattern features to identify the flow pattern corresponding to the time-frequency image samples; train and optimize the hybrid deep neural network model based on the classification results to obtain the final flow pattern recognition model;

[0013] The goal of building a hybrid model using a convolutional neural network (CNN) and a long short-term memory recurrent neural network (LSTM) is to combine the strengths of both to enhance the model's expressive power. CNN excels at processing image data, effectively extracting spatial features from time-frequency images and identifying local dependencies and texture information. LSTM, on the other hand, excels at processing sequential data, particularly data with temporal dependencies, and is capable of learning the temporal dynamics of time-frequency images. Combining CNN with LSTM allows for the simultaneous capture of complex spatial and temporal features when analyzing time-frequency images, leading to more accurate flow pattern classification.

[0014] S3.2.2 Embedding Algorithm: Embed the t-distributed random neighbor embedding algorithm into the flow pattern recognition model; the t-distributed random neighbor embedding algorithm is used to reduce the dimensionality of the extracted flow pattern features to visualize the flow pattern classification effect of the model;

[0015] t-SNE is designed to preserve local structure in high-dimensional data. For example, in identifying gas-liquid two-phase flow patterns, similar patterns often appear close together in high-dimensional space. t-SNE ensures that this local structure is preserved in the low-dimensional mapping by optimizing the probability distribution of similarity. This helps distinguish different types of flow patterns, as similar patterns tend to cluster together in low-dimensional space, while dissimilar patterns tend to spread out.

[0016] S3.3 Flow pattern identification: The time-frequency image of the gas-liquid two-phase flow in step S2 is input into the flow pattern identification model. The output result of the flow pattern identification model is the specific flow pattern of the gas-liquid two-phase flow.

[0017] Preferably, in step S2, the formula for converting the DAS signal using the wavelet transform method is as follows:

[0018]

[0019] Where CWT(a, b) is the wavelet coefficient, a is the scale factor, b is the translation factor, t is the time, and X(t) is the DAS signal measured at time t; is the wavelet basis function.

[0020] Preferably, in the hybrid deep neural network flow pattern recognition model in step S3.1, the first convolution layer is provided with 10 convolution kernels, and the size of each convolution kernel is 3×3;

[0021] The pooling kernel size of the first maximum pooling layer is 2×2 and the stride is 2;

[0022] The dropout rate of the first Dropout layer is 0.2;

[0023] The second convolutional layer has 24 convolution kernels, each of which is 5×5 in size;

[0024] The pooling kernel size of the second maximum pooling layer is 2×1 and the stride is 2;

[0025] The dropout rate of the second Dropout layer is 0.1;

[0026] The LSTM layer has 30 neurons;

[0027] The first fully connected layer has 64 output nodes;

[0028] The dropout rate of the third Dropout layer 3 is 0.1;

[0029] The second fully connected layer has 32 output nodes;

[0030] The dropout rate of the fourth Dropout layer is 0.1;

[0031] The third fully connected layer has 5 output nodes.

[0032] Convolutional layer and pooling layer: The designed convolutional layer and pooling layer help to extract local features in gas-liquid two-phase flow and reduce computational complexity by reducing feature dimensions.

[0033] Dropout layer: Dropout is used between different layers to prevent the model from overfitting and enhance the generalization ability of the model.

[0034] LSTM layer: The LSTM layer is integrated to capture the temporal dependencies in the signal, which is critical for time series data such as DAS data.

[0035] Fully connected layers and Softmax layers: These layers are used for final feature integration and flow type classification.

[0036] Preferably, in step S3.2.1, during the training of the hybrid deep neural network model, cross entropy is selected as the loss function. The specific formula of cross entropy as the loss function is as follows:

[0037]

[0038] Among them, h(z) is the true classification result of the time-frequency image sample z, and f(z) is the output classification result of the time-frequency image sample z.

[0039] Cross entropy is very sensitive to the accuracy of model predictions, especially when the predictions are wrong. If the model's predicted probability deviates from the true label, the cross entropy value will increase significantly, prompting the model to adjust in the right direction. Cross entropy can well handle the situation where the output layer uses the softmax function. It quantifies the inconsistency between the probability distribution of the model output and the expected distribution, helping the model improve its predictions for uncertain categories during training.

[0040] Preferably, in step S3.2, during the training of the hybrid deep neural network flow pattern recognition model, the network parameters are updated using a stochastic gradient descent optimization algorithm; an Adam optimizer is used, the learning rate is set to 0.001, the maximum number of iterations is set to 100, and the batch size is 32; the ReLU function is used as the activation function,

[0041]

[0042] The reasons for using the stochastic gradient descent optimization algorithm are as follows:

[0043] 1. High efficiency: Each iteration uses only one or a small batch of samples to calculate the gradient, which greatly reduces the demand for computing resources.

[0044] 2. Adaptable: It can handle large-scale data sets because it does not need to use all the data in every step.

[0045] 3. Escaping local minima: Due to its inherent noise nature, it helps the model escape from local minima and may find a more global minimum.

[0046] The Adam optimizer combines the advantages of momentum and RMSprop. It is an optimization algorithm with adaptive learning rate. It is particularly suitable for processing non-static and very large data sets. The reasons for choosing Adam include:

[0047] 1. Adaptive learning rate: Adam automatically adjusts the learning rate of each parameter based on the first-order moment estimate (mean) and second-order moment estimate (uncentered variance) of the parameter's most recent gradient.

[0048] 2. Less parameter tuning required: Compared with other optimization algorithms, Adam relies less on fine-tuning of hyperparameters.

[0049] 6. A method for identifying gas-liquid two-phase flow patterns based on DAS logging data, characterized in that, in step S3.2.2, a t-distributed random neighbor embedding algorithm is used to perform dimensionality reduction processing on the extracted flow pattern features. The specific process is as follows:

[0050] A1 defines the space containing flow pattern feature data as a high-dimensional space, and uses Gaussian probability distribution to represent the position information of data points in the high-dimensional space. The following formula is used to calculate the position of each data point x in the high-dimensional space. i and other data points x j The similarity p between j|i , similarity p j|i Used to reflect the relative position relationship between data points;

[0051]

[0052] Among them, σ i For data point X i The Gaussian mean square error centered on X k is the observation data point in high-dimensional space;

[0053] A2 initializes low-dimensional embedding: Map the relative position relationship of data points in high-dimensional space to low-dimensional space, use t distribution instead of Gaussian probability distribution in low-dimensional space to represent the position information of data points, and use the following formula to calculate the y value of each data point in low-dimensional space. i and other data points y j The similarity q between j|i ,

[0054]

[0055] Among them, y k is the observed data point in the low-dimensional space;

[0056] A3 Optimize low-dimensional embedding: Use KL divergence as the loss function C to measure and minimize p j|i and q j|i differences;

[0057]

[0058] Calculate the objective function C for each data point y in the low-dimensional space i The gradient of y is updated by the following formula i ,

[0059]

[0060] Through multiple iterations, gradually adjust y i to minimize the KL divergence until convergence.

[0061] Compared with the existing technology, this gas-liquid two-phase flow pattern identification method based on DAS logging data has the following beneficial effects:

[0062] 1. High Precision and Real-Time: By utilizing distributed fiber acoustic sensing (DAS) technology to continuously collect downhole dynamic data and combining it with wavelet transform to extract time-frequency images, this method can achieve high-precision identification of fluid flow patterns. The use of a hybrid deep neural network (combining CNN and LSTM) further enhances the ability to accurately extract and identify flow patterns from complex signals, ensuring real-time and efficient flow pattern classification.

[0063] 2. Adaptability and versatility: The method of the present invention is applicable to various well types, including conventional wells, highly deviated wells, horizontal wells, and ultra-deep wells, effectively solving the application limitations and performance degradation problems encountered by traditional instruments in these special well types.

[0064] 3. Data processing optimization: The t-distributed stochastic neighbor embedding (t-SNE) algorithm is introduced to reduce the dimensionality of flow pattern features. This not only optimizes the data processing and classification process, but also provides an effective way to visualize and analyze data, enhancing the model's interpretability and user-friendliness.

[0065] 4. Model optimization and training efficiency: By adopting advanced machine learning techniques and algorithms (such as the cross-entropy loss function and the Adam optimizer), the method of the present invention can efficiently adjust parameters and quickly converge when training hybrid neural networks, thereby reducing computing resource consumption and improving the speed and accuracy of flow pattern recognition.

[0066] 5. Reduced operational complexity and costs: Utilizing DAS technology, this method can operate stably under extreme environmental conditions (such as high temperature and pressure, corrosive environments, and the presence of geomagnetic and geoelectric interference). This technology does not rely on electronic components, resulting in higher reliability and longer service life in harsh downhole environments. The DAS system is easy to install, and due to its unique physical and chemical properties, it has low maintenance costs and is simple to operate. This reduces the workload of field operators and reduces long-term operating costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 It is a graph after pretreatment under different flow patterns of the present invention;

[0068] Among them, (a) smooth stratified flow; (b) wavy stratified flow; (c) slug flow; (d) bubbly flow; (e) mist flow.

[0069] Figure 2 It is a flow pattern classification visualization diagram of the present invention;

[0070] Among them, (a) original data visualization; (b) image visualization after CWT; (c) image visualization after 100 rounds of CNN-LSTM model training.

[0071] Figure 3 This is a flow pattern classification result diagram of the flow pattern recognition model of the present invention. DETAILED DESCRIPTION

[0072] The following are specific embodiments of the present invention and the accompanying drawings to further describe the technical solutions of the present invention, but the present invention is not limited to these embodiments.

[0073] like Figures 1 to 3 As shown, a gas-liquid two-phase flow pattern identification method based on DAS logging data includes the following steps:

[0074] S1 signal acquisition: The microstructured optical fibers are evenly arranged along the flow cross section of the wellbore, and the DAS acquisition system is used to continuously collect the DAS signals of the gas-liquid two-phase flow passing through the flow cross section of the wellbore through the microstructured optical fibers;

[0075] The design of microstructured optical fibers, such as photonic crystal fibers, allows them to have specific transmission properties, including a high degree of nonlinearity and tunable dispersion and modal properties. This makes microstructured fibers particularly useful in distributed acoustic sensing (DAS) systems because they can improve the sensing sensitivity and signal-to-noise ratio of acoustic events. In addition, these fibers can operate stably over a wide frequency band, enhancing the DAS system's ability to capture various acoustic fluctuations in complex downhole environments, thereby improving the accuracy and reliability of gas-liquid two-phase flow pattern identification.

[0076] S2 preprocessing: The DAS signal of the gas-liquid two-phase flow is converted using the continuous wavelet transform method to obtain the time-frequency image of the gas-liquid two-phase flow;

[0077] Continuous wavelet transform (CWT) is suitable for analyzing non-stationary signals (the statistical characteristics of the signal change with time). It can reveal the transient characteristics and spectral characteristics generated by different flow patterns and provide rich feature information for subsequent flow pattern identification.

[0078] S3 Flow Pattern Recognition: S3.1 Model Establishment: A hybrid deep neural network flow pattern recognition model is established using a convolutional neural network (CNN) and a long short-term memory (LSTM) recurrent neural network. The hybrid deep neural network flow pattern recognition model includes an input layer, a first convolutional layer, a first batch normalization layer, a first Relu activation layer, a first maximum pooling layer, a first Dropout layer, a second convolutional layer, a second batch normalization layer, a second Relu activation layer, a second maximum pooling layer, a second Dropout layer, a Flatten layer, an LSTM layer, a first fully connected layer, a third Dropout layer, a second fully connected layer, a fourth Dropout layer, a third fully connected layer, a Softmax layer, and a classification layer.

[0079] S3.2 Model Improvement: S3.2.1 Hybrid Deep Neural Network Model Training: Collect time-frequency image samples of gas-liquid two-phase flow under different flow patterns; input the time-frequency image samples into the hybrid deep neural network model; use the hybrid deep neural network model to extract flow pattern features of the time-frequency image samples, classify the extracted flow pattern features to identify the flow pattern corresponding to the time-frequency image samples; train and optimize the hybrid deep neural network model based on the classification results to obtain the final flow pattern recognition model;

[0080] The goal of building a hybrid model using a convolutional neural network (CNN) and a long short-term memory recurrent neural network (LSTM) is to combine the strengths of both to enhance the model's expressive power. CNN excels at processing image data, effectively extracting spatial features from time-frequency images and identifying local dependencies and texture information. LSTM, on the other hand, excels at processing sequential data, particularly data with temporal dependencies, and is capable of learning the temporal dynamics of time-frequency images. Combining CNN with LSTM allows for the simultaneous capture of complex spatial and temporal features when analyzing time-frequency images, leading to more accurate flow pattern classification.

[0081] S3.2.2 Embedding Algorithm: Embed the t-distributed random neighbor embedding algorithm into the flow pattern recognition model; the t-distributed random neighbor embedding algorithm is used to reduce the dimensionality of the extracted flow pattern features to visualize the flow pattern classification effect of the model;

[0082] t-SNE is designed to preserve local structure in high-dimensional data. For example, in identifying gas-liquid two-phase flow patterns, similar patterns often appear close together in high-dimensional space. t-SNE ensures that this local structure is preserved in the low-dimensional mapping by optimizing the probability distribution of similarity. This helps distinguish different types of flow patterns, as similar patterns tend to cluster together in low-dimensional space, while dissimilar patterns tend to spread out.

[0083] S3.3 Flow pattern identification: The time-frequency image of the gas-liquid two-phase flow in step S2 is input into the flow pattern identification model. The output result of the flow pattern identification model is the specific flow pattern of the gas-liquid two-phase flow.

[0084] Preferably, in step S2, the formula for converting the DAS signal using the wavelet transform method is as follows:

[0085]

[0086] Where CWT(a, b) is the wavelet coefficient, a is the scale factor, b is the translation factor, t is the time, and X(t) is the DAS signal measured at time t; is the wavelet basis function.

[0087] Preferably, in the hybrid deep neural network flow pattern recognition model in step S3.1, the first convolution layer is provided with 10 convolution kernels, and the size of each convolution kernel is 3×3;

[0088] The pooling kernel size of the first maximum pooling layer is 2×2 and the stride is 2;

[0089] The dropout rate of the first Dropout layer is 0.2;

[0090] The second convolutional layer has 24 convolution kernels, each of which is 5×5 in size;

[0091] The pooling kernel size of the second maximum pooling layer is 2×1 and the stride is 2;

[0092] The dropout rate of the second Dropout layer is 0.1;

[0093] The LSTM layer has 30 neurons;

[0094] The first fully connected layer has 64 output nodes;

[0095] The dropout rate of the third Dropout layer 3 is 0.1;

[0096] The second fully connected layer has 32 output nodes;

[0097] The dropout rate of the fourth Dropout layer is 0.1;

[0098] The third fully connected layer has 5 output nodes.

[0099] Convolutional layer and pooling layer: The designed convolutional layer and pooling layer help to extract local features in gas-liquid two-phase flow and reduce computational complexity by reducing feature dimensions.

[0100] Dropout layer: Dropout is used between different layers to prevent the model from overfitting and enhance the generalization ability of the model.

[0101] LSTM layer: The LSTM layer is integrated to capture the temporal dependencies in the signal, which is critical for time series data such as DAS data.

[0102] Fully connected layers and Softmax layers: These layers are used for final feature integration and flow type classification.

[0103] Preferably, in step S3.2.1, during the training of the hybrid deep neural network model, cross entropy is selected as the loss function. The specific formula of cross entropy as the loss function is as follows:

[0104]

[0105] Among them, h(z) is the true classification result of the time-frequency image sample z, and f(z) is the output classification result of the time-frequency image sample z.

[0106] Cross entropy is very sensitive to the accuracy of model predictions, especially when the predictions are wrong. If the model's predicted probability deviates from the true label, the cross entropy value will increase significantly, prompting the model to adjust in the right direction. Cross entropy can well handle the situation where the output layer uses the softmax function. It quantifies the inconsistency between the probability distribution of the model output and the expected distribution, helping the model improve its predictions for uncertain categories during training.

[0107] Preferably, in step S3.2, during the training of the hybrid deep neural network flow pattern recognition model, the network parameters are updated using a stochastic gradient descent optimization algorithm; an Adam optimizer is used, the learning rate is set to 0.001, the maximum number of iterations is set to 100, and the batch size is 32; the ReLU function is used as the activation function,

[0108]

[0109] The reasons for using the stochastic gradient descent optimization algorithm are as follows:

[0110] 1. High efficiency: Each iteration uses only one or a small batch of samples to calculate the gradient, which greatly reduces the demand for computing resources.

[0111] 2. Adaptable: It can handle large-scale data sets because it does not need to use all the data in every step.

[0112] 3. Escaping local minima: Due to its inherent noise nature, it helps the model escape from local minima and may find a more global minimum.

[0113] The Adam optimizer combines the advantages of momentum and RMSprop. It is an optimization algorithm with adaptive learning rate. It is particularly suitable for processing non-static and very large data sets. The reasons for choosing Adam include:

[0114] 1. Adaptive learning rate: Adam automatically adjusts the learning rate of each parameter based on the first-order moment estimate (mean) and second-order moment estimate (uncentered variance) of the parameter's most recent gradient.

[0115] 2. Less parameter tuning required: Compared with other optimization algorithms, Adam relies less on fine-tuning of hyperparameters.

[0116] Preferably, in step S3.2.2, the t-distributed random neighbor embedding algorithm performs dimensionality reduction processing on the extracted flow pattern features. The specific process is as follows:

[0117] A1 defines the space containing flow pattern feature data as a high-dimensional space, and uses Gaussian probability distribution to represent the position information of data points in the high-dimensional space. The following formula is used to calculate the position of each data point x in the high-dimensional space. i and other data points x j The similarity p between j|i , similarity p j|i Used to reflect the relative position relationship between data points;

[0118]

[0119] Among them, σ i For data point Xi The Gaussian mean square error centered on x k is the observation data point in high-dimensional space;

[0120] A2 initializes low-dimensional embedding: Map the relative position relationship of data points in high-dimensional space to low-dimensional space, use t distribution instead of Gaussian probability distribution in low-dimensional space to represent the position information of data points, and use the following formula to calculate the y value of each data point in low-dimensional space. i and other data points y j The similarity q between j|i ,

[0121]

[0122] Among them, y k is the observed data point in the low-dimensional space;

[0123] A3 Optimize low-dimensional embedding: Use KL divergence as the loss function C to measure and minimize p j|i and q j|i There are differences;

[0124]

[0125] Calculate the objective function C for each data point y in the low-dimensional space i The gradient of y is updated by the following formula i ,

[0126]

[0127] Through multiple iterations, gradually adjust y i to minimize the KL divergence until convergence.

[0128] The specific embodiments described herein are merely illustrative of the spirit of the present invention. Persons skilled in the art may make various modifications, additions, or substitutions to the described specific embodiments without departing from the spirit of the present invention or exceeding the scope of the appended claims.

Claims

1. A gas-liquid two-phase flow pattern identification method based on DAS logging data, characterized in that: The identification method comprises the following steps: S1 signal acquisition: The microstructured optical fibers are evenly arranged along the flow cross section of the wellbore, and the DAS acquisition system is used to continuously collect the DAS signals of the gas-liquid two-phase flow passing through the flow cross section of the wellbore through the microstructured optical fibers; S2 preprocessing: The DAS signal of the gas-liquid two-phase flow is converted using the continuous wavelet transform method to obtain the time-frequency image of the gas-liquid two-phase flow; S3 Flow Pattern Recognition: S3.1 Model Establishment: A hybrid deep neural network flow pattern recognition model is established using a convolutional neural network (CNN) and a long short-term memory (LSTM) recurrent neural network. The hybrid deep neural network flow pattern recognition model includes an input layer, a first convolutional layer, a first batch normalization layer, a first Relu activation layer, a first maximum pooling layer, a first Dropout layer, a second convolutional layer, a second batch normalization layer, a second Relu activation layer, a second maximum pooling layer, a second Dropout layer, a Flatten layer, an LSTM layer, a first fully connected layer, a third Dropout layer, a second fully connected layer, a fourth Dropout layer, a third fully connected layer, a Softmax layer, and a classification layer. S3.2 Model Improvement: S3.2.1 Hybrid Deep Neural Network Model Training: Collect time-frequency image samples of gas-liquid two-phase flow under different flow patterns; input the time-frequency image samples into the hybrid deep neural network model; use the hybrid deep neural network model to extract flow pattern features of the time-frequency image samples, classify the extracted flow pattern features to identify the flow pattern corresponding to the time-frequency image samples; train and optimize the hybrid deep neural network model based on the classification results to obtain the final flow pattern recognition model; S3.2.2 Embedding Algorithm: Embed the t-distributed random neighbor embedding algorithm into the flow pattern recognition model; the t-distributed random neighbor embedding algorithm is used to reduce the dimensionality of the extracted flow pattern features to visualize the flow pattern classification effect of the model; S3.3 Flow pattern identification: The time-frequency image of the gas-liquid two-phase flow in step S2 is input into the flow pattern identification model. The output result of the flow pattern identification model is the specific flow pattern of the gas-liquid two-phase flow; In step S3.2.1, during the training of the hybrid deep neural network model, cross entropy is selected as the loss function. The specific formula of cross entropy as the loss function is as follows: Among them, h(z) is the true classification result of the time-frequency image sample z, and f(z) is the output classification result of the time-frequency image sample z; In step S3.2, during the training of the hybrid deep neural network flow pattern recognition model, the network parameters are updated using the stochastic gradient descent optimization algorithm; the Adam optimizer is used, the learning rate is set to 0.001, the maximum number of iterations is set to 100, and the batch size is 32; the ReLU function is used as the activation function, 2. The gas-liquid two-phase flow pattern identification method based on DAS logging data according to claim 1, characterized in that: In step S2, the DAS signal is converted using the wavelet transform method according to the following formula: Where CWT(a, b) is the wavelet coefficient, a is the scale factor, b is the translation factor, t is the time, and X(t) is the DAS signal measured at time t; is the wavelet basis function.

3. The gas-liquid two-phase flow pattern identification method based on DAS logging data according to claim 1, characterized in that: In the hybrid deep neural network flow pattern recognition model in step S3.1, the first convolutional layer is equipped with 10 convolution kernels, each of which is 3×3 in size; The pooling kernel size of the first maximum pooling layer is 2×2 and the stride is 2; The dropout rate of the first Dropout layer is 0.2; The second convolutional layer has 24 convolution kernels, each of which is 5×5 in size; The pooling kernel size of the second maximum pooling layer is 2×1 and the stride is 2; The dropout rate of the second Dropout layer is 0.1; The LSTM layer has 30 neurons; The first fully connected layer has 64 output nodes; The dropout rate of the third Dropout layer 3 is 0.1; The second fully connected layer has 32 output nodes; The dropout rate of the fourth Dropout layer is 0.1; The third fully connected layer has 5 output nodes.

4. The method for identifying gas-liquid two-phase flow patterns based on DAS logging data according to claim 1, characterized in that: In step S3.2.2, the t-distributed random neighbor embedding algorithm performs dimensionality reduction on the extracted flow pattern features. The specific process is as follows: A1 defines the space containing flow pattern feature data as a high-dimensional space, and uses Gaussian probability distribution to represent the position information of data points in the high-dimensional space. The following formula is used to calculate the position of each data point x in the high-dimensional space. i and other data points x j The similarity p between j|i , similarity p j|i Used to reflect the relative position relationship between data points; Among them, σ i For data point x i The Gaussian mean square error centered on x k is the observation data point in high-dimensional space; A2 initializes low-dimensional embedding: Map the relative position relationship of data points in high-dimensional space to low-dimensional space, use t distribution instead of Gaussian probability distribution in low-dimensional space to represent the position information of data points, and use the following formula to calculate the y value of each data point in low-dimensional space. i and other data points y j The similarity q between j|i , Among them, y k is the observed data point in the low-dimensional space; A3 Optimize low-dimensional embedding: Use KL divergence as the loss function C to measure and minimize p j|i and q j|i differences; Calculate the objective function C for each data point y in the low-dimensional space i The gradient of y is updated by the following formula i , Through multiple iterations, gradually adjust y i to minimize the KL divergence until convergence.

Images