Wellhead water content metering method based on deep separation patch network model

Abstract

The invention discloses a wellhead water content metering method based on a deep separation patch network model, and the deep separation patch network model comprises a multi-scale patch embedding module, a first deep separation convolution module, a second deep separation convolution module, a self-adaptive channel weighted fusion module and a multi-layer perceptron. According to the wellhead moisture content metering method, a first deep separation convolution module and a second deep separation convolution module of a deep separation patch network model are combined to comprehensively extract local dynamic characteristics and global trends of normalized microwave phase voltage signals and normalized microwave amplitude voltage signals; and the accuracy and robustness of moisture content prediction are improved.

Description

Technical Field

[0001] The invention belongs to the technical field of intelligent measurement of oil fields, and in particular relates to a wellhead water content measurement method based on a deep separation patch network model. Background Art

[0002] Accurate measurement of wellhead water content is a key link in oilfield production management, which directly affects the efficiency of oil-water separation, production evaluation and production optimization. However, under complex working conditions such as low flow rate, high water content and nonlinear dynamic flow, the accuracy and stability of traditional measurement technologies face severe challenges. Among them, the conductivity method and capacitance method are based on the difference in conductivity and dielectric constant of oil-water mixed fluids for measurement. The conductivity method and capacitance method are prone to signal saturation under high water content conditions, and are sensitive to changes in temperature and flow rate, resulting in large measurement errors and unable to meet the measurement requirements in complex flow environments. The microwave measurement method uses the phase and amplitude changes of microwave signals when passing through the oil-water mixed fluid for measurement. It has the advantages of non-contact, high sensitivity and fast response. However, traditional microwave measurement methods usually use single frequency or fixed mode signals and cannot fully capture the multi-dimensional characteristics of the fluid. In addition, the traditional measurement method lacks measurement stability in nonlinear dynamic flow, low flow rate and complex environment, resulting in large errors.

[0003] The dynamic characteristics of oil-water two-phase flow are complex. The changes of microwave signals when passing through the oil-water mixed medium not only include local dynamic characteristics, but also show global trend characteristics. Local characteristics: microscopic disturbances, turbulence and instantaneous medium changes at the oil-water interface will cause short-term fluctuations in microwave signals. Global characteristics: water content shows an overall time series trend with the change of time, pressure and flow rate.

[0004] Traditional measurement techniques fail to fully consider the multi-scale characteristics of the signal and only capture part of the local or global information, resulting in an incomplete characterization of the moisture content characteristics. Summary of the invention

[0005] In view of the deficiencies in the prior art, the object of the present invention is to provide a deep separation patch network model (DSPNet).

[0006] Another object of the present invention is to provide a wellhead water content measurement method based on a deep separation patch network model. The wellhead water content measurement method of the present invention realizes high-precision prediction of the water content of oil-water mixed fluid.

[0007] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0008] A deep separation patch network model, comprising: a multi-scale patch embedding module, a first deep separation convolution module, a second deep separation convolution module, an adaptive channel weighted fusion module and a multi-layer perceptron;

[0009] The multi-scale patch embedding module is used to perform multi-scale patch embedding, linear transformation and position encoding in sequence. The multi-scale patch embedding is to divide the samples into N1 first-scale patches and N2 second-scale patches according to two different scales;

[0010] The samples include a normalized microwave phase voltage signal and a normalized microwave amplitude voltage signal. The normalized microwave phase voltage signal is obtained by normalizing the microwave phase voltage signal. The normalized microwave amplitude voltage signal is obtained by normalizing the microwave amplitude voltage signal. The microwave phase voltage signal and the microwave amplitude voltage signal are obtained by collecting the oil-water two-phase flow in the same time period.

[0011] The multi-scale patch embedding module projects N1 first-scale patches and N2 second-scale patches into a unified feature space through linear transformation, and obtains N1 linearly transformed first-scale patches and N2 linearly transformed second-scale patches;

[0012] The multi-scale patch embedding module performs position encoding on N1 linearly transformed first-scale patches and N2 linearly transformed second-scale patches and adds corresponding position information to obtain N1 first-scale patch features and N2 second-scale patch features. The N1 first-scale patch features form a vector array X1, and the N2 second-scale patch features form a vector array X2.

[0013] The input features of the first depth separation convolution module are vector array X1 and vector array X2. The first depth separation convolution module is used to perform depth separation convolution operations on vector array X1 and vector array X2 respectively. The output features of the first depth separation convolution module are first scale patch global fusion feature Y1 and second scale patch global fusion feature Y2.

[0014] The input features of the second deep separation convolution module are the first scale patch global fusion feature Y1 and the second scale patch global fusion feature Y2. The second deep separation convolution module is used to perform deep separation convolution operations on the first scale patch global fusion feature Y1 and the second scale patch global fusion feature Y2 respectively. The output features of the second deep separation convolution module are the first scale patch global fusion feature Y3 and the second scale patch global fusion feature Y4.

[0015] Each depthwise separable convolution operation includes: resize padding, depthwise convolution, first batch normalization, summation, pointwise convolution, and second batch normalization in sequence;

[0016] The adaptive channel weighted fusion module performs adaptive weighted fusion on the first-scale patch global fusion feature Y3 and the second-scale patch global fusion feature Y4 through the compression excitation network to obtain a multi-scale fusion feature. The dimension of the multi-scale fusion feature is (N1+N2)×D, where D is the dimension of the feature space.

[0017] The multi-layer perceptron is used to input multi-scale fusion features, and the multi-layer perceptron outputs the moisture content prediction value.

[0018] In the above technical solution, the length of the first scale patch is P1=50, and the length of the second scale patch is P2=250.

[0019] In the above technical solution, the number of first-scale patches N1=T / P1, the number of second-scale patches N2=T / P2, and T is the length of the sample.

[0020] In the above technical solution, the dimension of each first-scale patch after linear transformation is 1×D, and the dimension of each second-scale patch after linear transformation is 1×D, where D=512.

[0021] In the above technical solution, the dimension of the first-scale patch global fusion feature Y1 is N1×D, and the dimension of the second-scale patch global fusion feature Y2 is N2×D.

[0022] In the above technical solution, the dimension of the first-scale patch global fusion feature Y3 is N1×D, and the dimension of the second-scale patch global fusion feature Y4 is N2×D.

[0023] In the above technical scheme, the depth convolution used to perform depth separation convolution operation on the vector array X1 uses N1 convolution kernels, each of which has a size of 1×25; the depth convolution used to perform depth separation convolution operation on the vector array X2 uses N2 convolution kernels, each of which has a size of 1×125; the depth convolution used to perform depth separation convolution operation on the first-scale patch global fusion feature Y1 uses N1 convolution kernels, each of which has a size of 1×25; the depth convolution used to perform depth separation convolution operation on the second-scale patch global fusion feature Y2 uses N2 convolution kernels, each of which has a size of 1×125.

[0024] In the above technical solution, the summation used in the deep separation convolution operation on the vector array X1 is: adding each first-scale patch feature and the output of the first batch normalization corresponding to the first-scale patch feature to obtain the first-scale patch fusion feature X; the summation used in the deep separation convolution operation on the vector array X2 is: adding each second-scale patch feature and the output of the first batch normalization corresponding to the second-scale patch feature to obtain the second-scale patch fusion feature X.

[0025] In the above technical solution, the summation used in the deep separation convolution operation on the first-scale patch global fusion feature Y1 is: each first-scale patch global fusion feature Y1 is added to the first batch normalized output corresponding to the first-scale patch global fusion feature Y1 to obtain the first-scale patch fusion feature Y; the summation used in the deep separation convolution operation on the second-scale patch global fusion feature Y2 is: each second-scale patch global fusion feature Y2 is added to the first batch normalized output corresponding to the second-scale patch global fusion feature Y2 to obtain the second-scale patch fusion feature Y.

[0026] In the above technical scheme, the point-by-point convolution for performing depth separation convolution operation on the vector array X1 uses N1 convolution kernels, each of which has a size of 1×1; the point-by-point convolution for performing depth separation convolution operation on the vector array X2 uses N2 convolution kernels, each of which has a size of 1×1; the point-by-point convolution for performing depth separation convolution operation on the first-scale patch global fusion feature Y1 uses N1 convolution kernels, each of which has a size of 1×1; the point-by-point convolution for performing depth separation convolution operation on the second-scale patch global fusion feature Y2 uses N2 convolution kernels, each of which has a size of 1×1.

[0027] In the above technical solution, the multi-layer perceptron includes: an input layer, a discard layer, a hidden layer, a discard layer and a prediction layer connected in sequence, the input layer is used to input multi-scale fusion features, and the prediction layer is used to output moisture content prediction values.

[0028] In the above technical solution, the activation function of the multi-layer perceptron adopts ReLU.

[0029] In the above technical solution, the number of neurons in the input layer is 1024, the number of neurons in the hidden layer is 256, and the dropout rate of the dropout layer is set to 0.3.

[0030] In the above technical solution, the loss function value of the deep separation patch network model adopts the L1 loss function.

[0031] A wellhead water content measurement method based on a deep separation patch network model comprises: inputting a test set into a trained deep separation patch network model to obtain a water content prediction value.

[0032] The present invention has the following advantages due to the adoption of the above technical solution:

[0033] 1. The wellhead water content measurement method of the present invention combines the first deep separation convolution module and the second deep separation convolution module of the deep separation patch network model to comprehensively extract the local dynamic characteristics and global trends of the normalized microwave phase voltage signal and the normalized microwave amplitude voltage signal, thereby improving the accuracy and robustness of water content prediction.

[0034] 2. The adaptive channel weighted fusion module in the deep separation patch network model of the present invention optimizes the expression of features, making the present invention suitable for complex working conditions such as low flow rate and high water content.

[0035] 3. The multi-scale patch embedding module combined with the deep separation patch network model of the present invention enhances the modeling capability of dynamic changes in time series and is suitable for complex working conditions such as dynamic nonlinear flow, high water content and low flow velocity. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is an architecture diagram of the deep separation patch network model (DSPNet) of the present invention. DETAILED DESCRIPTION

[0037] A method for measuring water content at a wellhead based on a deep separation patch network model of the present invention is described in detail below in conjunction with the accompanying drawings and embodiments.

[0038] Example 1

[0039] like Figure 1 As shown, a deep separation patch network model (DSPNet) includes: a multi-scale patch embedding module, a first deep separation convolution module, a second deep separation convolution module, an adaptive channel weighted fusion module and a multi-layer perceptron (MLP);

[0040] The multi-scale patch embedding module is used to perform multi-scale patch embedding, linear transformation and position encoding in sequence. The multi-scale patch embedding is to divide the samples into N1 first-scale patches and N2 second-scale patches according to two different scales. The length of the first-scale patch is P1=50, the length of the second-scale patch is P2=250, the number of first-scale patches is N1=T / P1=60, the number of second-scale patches is N2=T / P2=12, the dimension of each sample is T×C, C=2 is the feature dimension, and the sample includes a normalized microwave phase voltage signal and a normalized microwave amplitude voltage signal. The normalized microwave phase voltage signal is obtained by normalizing the microwave phase voltage signal, and the normalized microwave amplitude voltage signal is obtained by normalizing the microwave amplitude voltage signal. The microwave phase voltage signal and the microwave amplitude voltage signal are obtained by collecting the oil-water two-phase flow at the wellhead in the same time period. Both the first-scale patch and the second-scale patch contain information of all feature dimensions.

[0041] The multi-scale patch embedding module projects N1 first-scale patches and N2 second-scale patches into a unified feature space through linear transformation, and obtains N1 linearly transformed first-scale patches and N2 linearly transformed second-scale patches. The dimension of each linearly transformed first-scale patch is 1×D, and the dimension of each linearly transformed second-scale patch is 1×D, where D=512, and D is the dimension of the feature space.

[0042] The multi-scale patch embedding module will position encode the N1 first-scale patches after linear transformation and the N2 second-scale patches after linear transformation and add the corresponding position information to achieve patch enhancement, and obtain N1 first-scale patch features and N2 second-scale patch features. The N1 first-scale patch features form a vector array X1, and the N2 second-scale patch features form a vector array X2. The position encoding operation can enhance the sequential perception ability of the deep separation patch network model for time series. Both the first-scale patch features and the second-scale patch features retain the local dynamic change information and time sequence information of the sample.

[0043] The first depth separation convolution module and the second depth separation convolution module are connected in sequence (the first depth separation convolution module and the second depth separation convolution module refer to: Gong Z, Tang Y, Liang J. Patchmixer: A patch-mixing architecture for long-term time series forecasting [J]. arXiv preprint arXiv: 2310.00655, 2023.).

[0044] The input features of the first depth separation convolution module are vector array X1 and vector array X2. The first depth separation convolution module is used to extract local features of vector array X1 and vector array X2 respectively through depth separation convolution operations. The output features of the first depth separation convolution module are first-scale patch global fusion feature Y1 and second-scale patch global fusion feature Y2. The dimension of the first-scale patch global fusion feature Y1 is N1×D, and the dimension of the second-scale patch global fusion feature Y2 is N2×D.

[0045] The input features of the second deep separation convolution module are the first scale patch global fusion feature Y1 and the second scale patch global fusion feature Y2. The second deep separation convolution module is used to extract global features from the first scale patch global fusion feature Y1 and the second scale patch global fusion feature Y2 respectively through deep separation convolution operations. The output features of the second deep separation convolution module are the first scale patch global fusion feature Y3 and the second scale patch global fusion feature Y4. The dimension of the first scale patch global fusion feature Y3 is N1×D, and the dimension of the second scale patch global fusion feature Y4 is N2×D.

[0046] Each depthwise separation convolution operation includes: padding, depthwise convolution, first batch normalization, addition, pointwise convolution and second batch normalization performed in sequence; padding is used to adjust the dimension of its input so that the dimension of the input of padding is the same as the dimension of the output of the second batch normalization.

[0047] The depthwise convolution for depthwise separable convolution of vector array X1 uses N1 convolution kernels, each of size 1 × 25. The depthwise convolution for depthwise separable convolution of vector array X2 uses N2 convolution kernels, each of size 1 × 125. The depthwise convolution for depthwise separable convolution of first scale patch global fusion feature Y1 uses N1 convolution kernels, each of size 1 × 25. The depthwise convolution for depthwise separable convolution of second scale patch global fusion feature Y2 uses N2 convolution kernels, each of size 1 × 125.

[0048] The summation used in the deep separation convolution operation on the vector array X1 is as follows: each first-scale patch feature is summed with the first batch normalized output corresponding to the first-scale patch feature (used in the deep separation convolution operation on the vector array X1) to obtain the first-scale patch fusion feature X; the summation used in the deep separation convolution operation on the vector array X2 is as follows: each second-scale patch feature is summed with the first batch normalized output corresponding to the second-scale patch feature (used in the deep separation convolution operation on the vector array X2) to obtain the second-scale patch fusion feature x;

[0049] The summation used in the deep separation convolution operation on the first-scale patch global fusion feature Y1 is as follows: each first-scale patch global fusion feature Y1 is added to the first batch normalized output corresponding to the first-scale patch global fusion feature Y1 (used in the deep separation convolution operation on the first-scale patch global fusion feature Y1), and the first-scale patch fusion feature Y is obtained; the summation used in the deep separation convolution operation on the second-scale patch global fusion feature Y2 is as follows: each second-scale patch global fusion feature Y2 is added to the first batch normalized output corresponding to the second-scale patch global fusion feature Y2 (used in the deep separation convolution operation on the second-scale patch global fusion feature Y2), and the second-scale patch fusion feature Y is obtained;

[0050] The point-wise convolution for depthwise separable convolution of vector array X1 uses N1 convolution kernels, each of size 1×1. The point-wise convolution for depthwise separable convolution of vector array X2 uses N2 convolution kernels, each of size 1×1. The point-wise convolution for depthwise separable convolution of first scale patch global fusion feature Y1 uses N1 convolution kernels, each of size 1×1. The point-wise convolution for depthwise separable convolution of second scale patch global fusion feature Y2 uses N2 convolution kernels, each of size 1×1.

[0051] The output of the second batch normalization for performing a deep separation convolution operation on the vector array X1 is the first-scale patch global fusion feature Y1, and the output of the second batch normalization for performing a deep separation convolution operation on the vector array X2 is the second-scale patch global fusion feature Y2; the output of the second batch normalization for performing a deep separation convolution operation on the first-scale patch global fusion feature Y1 is the first-scale patch global fusion feature Y3, and the output of the second batch normalization for performing a deep separation convolution operation on the second-scale patch global fusion feature Y2 is the second-scale patch global fusion feature Y4.

[0052] The adaptive channel weighted fusion module performs adaptive weighted fusion on the first-scale patch global fusion feature Y3 and the second-scale patch global fusion feature Y4 through a compressed excitation network (Mei Wenkai, Xiao Yingyuan. Click-through rate prediction model of attention factor decomposition machine based on compressed excitation network [J]. Journal of Tianjin University of Technology, 2024, 40(02): 7-12.) to optimize the feature expression and obtain multi-scale fusion features. The dimension of the multi-scale fusion features is (N1+N2)×D.

[0053] The multi-layer perceptron (MLP) is used to input multi-scale fusion features and capture their complex dynamic features. The multi-layer perceptron outputs the predicted value of water content. The multi-layer perceptron includes: an input layer, a discard layer, a hidden layer, a discard layer and a prediction layer connected in sequence. The input layer is used to input multi-scale fusion features and extract high-dimensional nonlinear relationships. The prediction layer is used to output the predicted value of water content of oil-water two-phase flow. The activation function of the multi-layer perceptron uses ReLU. The number of neurons in the input layer is 1024, the number of neurons in the hidden layer is 256, and the discard rate of the discard layer is set to 0.3. The discard layer is used to improve the generalization ability of the deep separation patch network model and reduce the risk of overfitting. The prediction layer is used to map the features output by the discard layer to the target space.

[0054] Example 2

[0055] The method to obtain the fluid data set is:

[0056] The microwave sensor is installed on the oil-water pipeline at the wellhead, so that the microwave sensor collects microwave phase voltage signals and microwave amplitude voltage signals. During the time period of collecting microwave phase voltage signals and microwave amplitude voltage signals, the oil-water two-phase flow at the wellhead is manually sampled multiple times, and artificial tests are performed to obtain multiple water content test values ​​of the oil-water two-phase flow, and the water content test values ​​are used as the water content label values ​​of the oil-water two-phase flow. Among them, the microwave sensor can use a sensor with corresponding functions. For example, the microwave sensor can use the double-helix microwave sensor in CN202311628606.5. In this embodiment, the microwave phase voltage signal and the microwave amplitude voltage signal are collected at a frequency of 1.2 GHz.

[0057] The sampling frequency of the microwave sensor is 1kHz. The microwave sensor performs 24-hour measurements in different time periods. The collected microwave phase voltage signals and microwave amplitude voltage signals are normalized and divided according to the length of T=3000. Each portion is a sample, forming a total of 28,800 samples. Multiple moisture content label values ​​are matched one by one with the samples to form a fluid data set (each sample in the fluid data set includes a normalized microwave phase voltage signal and a normalized microwave amplitude voltage signal).

[0058] The fluid dataset is divided into training set, validation set and test set in a ratio of 8:1:1. The training set is used to learn the parameters of the deep separation patch network model and fit the distribution characteristics of the training set as much as possible. The validation set is used to evaluate the performance of the deep separation patch network model on unseen data and adjust related hyperparameters. The test set is used to finally evaluate the actual prediction performance of the deep separation patch network model.

[0059] Example 3

[0060] A method for training and evaluating a deep separation patch network model, comprising:

[0061] The training set in Example 2 is input into the deep separation patch network model in Example 1 for training, and the loss function value is calculated according to the prediction result. The AdamW optimizer (Zhu H, Zhang Z, Cong W, et al. APOLLO: SGD-like Memory, AdamW-level Performance [J]. arXiv preprint arXiv: 2412.05270, 2024.) is used, and the weight decay strategy (Zhang G, Wang C, Xu B, et al. Three mechanisms of weight decay regularization [J]. arXiv preprint arXiv: 1810.12281, 2018.) is combined to perform 200 rounds of training on 32 samples of each batch of training sets. After each training round, the verification set is input into the deep separation patch network model for moisture content prediction and it is determined whether the loss function value decreases: if the loss function value of the verification set does not improve within 10 consecutive rounds, the early stopping mechanism is triggered to terminate the training, and the parameters with the best performance of the verification set are saved as the trained deep separation patch network model.

[0062] Among them, the loss function value is calculated using the L1 loss function (Chen Cong. Extreme learning machine with L1 regularization and pinball loss function [J]. Information Technology and Informatization, 2023, (03): 37-40.).

[0063] The initial learning rate is set to 0.0001; the learning rate scheduling adopts the cosine annealing scheduler (Tang Zhenchao, Wei Wei, Luo Weiran, et al. Remote sensing image semantic segmentation integrating cosine annealing and dilated convolution [J]. Journal of Remote Sensing, 2023, 27(11): 2579-2592.), and the learning rate decreases with the training rounds.

[0064] Example 4

[0065] A method for measuring water content at a wellhead based on a deep separation patch network model, comprising:

[0066] The test set in Example 2 is input into the deep separation patch network model trained in Example 3 to obtain the moisture content prediction value, and the three indicators of mean square error (MSE), mean absolute error (MAE) and mean absolute percentage error (MAPE) are calculated to comprehensively evaluate the overall performance of the deep separation patch network model.

[0067] Example 5

[0068] A method for measuring wellhead water content is basically the same as Example 4, with the only difference being that the multi-scale patch embedding module is replaced by a single-scale patch embedding module, which performs a single-scale patch embedding on the sample and divides the sample into multiple scale patches according to a length of 50 patches.

[0069] Example 6

[0070] A method for measuring wellhead water content is basically the same as Example 4, with the only difference being that the multi-scale patch embedding module is replaced by a single-scale patch embedding module, which performs single-scale patch embedding on the sample and divides the sample into multiple scale patches according to a patch length of 250.

[0071] Example 7

[0072] A method for measuring water content at a wellhead is basically the same as Example 4, with the only difference being that each depth separation convolution operation in the first depth separation convolution module is replaced by a conventional one-dimensional convolution layer with a convolution kernel size of 1×3.

[0073] Example 8

[0074] A method for measuring water content at a wellhead is basically the same as that of Example 4, except that each deep separation convolution operation in the first deep separation convolution module is replaced by global pooling.

[0075] Example 9

[0076] A method for measuring wellhead water content is basically the same as Example 4, with the only difference being that each deep separation convolution operation in the first deep separation convolution module is replaced by a conventional one-dimensional convolution layer with a convolution kernel size of 1×3, and each deep separation convolution operation in the second deep separation convolution module is replaced by global pooling.

[0077] Example 10

[0078] A method for measuring water content at a wellhead is basically the same as that of Example 4, except that the adaptive channel weighted fusion module is replaced by a linear weighted fusion mechanism.

[0079] The model is evaluated using three indicators: mean square error (MSE), mean absolute error (MAE), and mean absolute percentage error (MAPE). The results are as follows:

[0080] Table 1

[0081]

[0082] It can be seen from Table 1 that Example 4 can capture both local dynamic changes and global correlation features, significantly improving model performance. The patch length of Example 5 is too short to extract global features, affecting model performance. The patch length of Example 6 is too long, although it enhances global features, but its ability to capture local dynamic changes is insufficient.

[0083] Table 2

[0084]

[0085] As can be seen from Table 2, the model performance of Example 7 has declined significantly, and the first deep separation convolution module of the deep separation patch network model of the present invention has a stronger ability to extract local dynamic features. The model performance of Example 8 has also declined significantly, and the second deep separation convolution module of the deep separation patch network model of the present invention is irreplaceable in terms of modeling capabilities across time scales. The model performance of Example 9 is the worst. The first deep separation convolution module and the second deep separation convolution module of Example 4 can effectively capture local and global dynamic features, and are the key to improving the performance of the deep separation patch network model.

[0086] Table 3

[0087]

[0088] As can be seen from Table 3, the linear weighting mechanism with fixed weights adopted in Example 10 has significantly reduced performance. The dynamic weighting mechanism can automatically optimize the weight distribution according to the feature importance. The adaptive channel weighting module of Example 4 can dynamically enhance the phase features that are sensitive to changes in water content, thereby improving the prediction performance.

[0089] The present invention is described above by way of example. It should be noted that, without departing from the core of the present invention, any simple deformation, modification or other equivalent replacement that can be made by those skilled in the art without inventive effort falls within the protection scope of the present invention.

Claims

1. A deep separation patch network model, characterized in that: include: Multi-scale patch embedding module, first depth separation convolution module, second depth separation convolution module, adaptive channel weighted fusion module and multi-layer perceptron; The multi-scale patch embedding module is used to perform multi-scale patch embedding, linear transformation and position encoding in sequence. The multi-scale patch embedding is to divide the samples into N1 first-scale patches and N2 second-scale patches according to two different scales; The samples include a normalized microwave phase voltage signal and a normalized microwave amplitude voltage signal. The normalized microwave phase voltage signal is obtained by normalizing the microwave phase voltage signal. The normalized microwave amplitude voltage signal is obtained by normalizing the microwave amplitude voltage signal. The microwave phase voltage signal and the microwave amplitude voltage signal are obtained by collecting the oil-water two-phase flow in the same time period. The multi-scale patch embedding module projects N1 first-scale patches and N2 second-scale patches into a unified feature space through linear transformation, and obtains N1 linearly transformed first-scale patches and N2 linearly transformed second-scale patches; The multi-scale patch embedding module performs position encoding on N1 linearly transformed first-scale patches and N2 linearly transformed second-scale patches and adds corresponding position information to obtain N1 first-scale patch features and N2 second-scale patch features. The N1 first-scale patch features form a vector array X1, and the N2 second-scale patch features form a vector array X2. The input features of the first depth separation convolution module are vector array X1 and vector array X2. The first depth separation convolution module is used to perform depth separation convolution operations on vector array X1 and vector array X2 respectively. The output features of the first depth separation convolution module are first scale patch global fusion feature Y1 and second scale patch global fusion feature Y2. The input features of the second deep separation convolution module are the first scale patch global fusion feature Y1 and the second scale patch global fusion feature Y2. The second deep separation convolution module is used to perform deep separation convolution operations on the first scale patch global fusion feature Y1 and the second scale patch global fusion feature Y2 respectively. The output features of the second deep separation convolution module are the first scale patch global fusion feature Y3 and the second scale patch global fusion feature Y4. Each depthwise separable convolution operation includes: resize padding, depthwise convolution, first batch normalization, summation, pointwise convolution, and second batch normalization in sequence; The adaptive channel weighted fusion module performs adaptive weighted fusion on the first-scale patch global fusion feature Y3 and the second-scale patch global fusion feature Y4 through the compression excitation network to obtain a multi-scale fusion feature. The dimension of the multi-scale fusion feature is (N1+N2)×D, where D is the dimension of the feature space. The multi-layer perceptron is used to input multi-scale fusion features, and the multi-layer perceptron outputs the moisture content prediction value.

2. According to the deep separation patch network model of claim 1, it is characterized in that The summation in the deep separation convolution operation on the vector array X1 is: summing each first-scale patch feature and the output of the first batch normalization corresponding to the first-scale patch feature to obtain the first-scale patch fusion feature X; The summation used in the deep separation convolution operation on the vector array X2 is: adding each second-scale patch feature and the output of the first batch normalization corresponding to the second-scale patch feature to obtain the second-scale patch fusion feature X.

3. According to the deep separation patch network model of claim 1, it is characterized in that The summation used in the deep separation convolution operation on the first-scale patch global fusion feature Y1 is: adding each first-scale patch global fusion feature Y1 and the first batch normalized output corresponding to the first-scale patch global fusion feature Y1 to obtain the first-scale patch fusion feature Y; the summation used in the deep separation convolution operation on the second-scale patch global fusion feature Y2 is: adding each second-scale patch global fusion feature Y2 and the first batch normalized output corresponding to the second-scale patch global fusion feature Y2 to obtain the second-scale patch fusion feature Y.

4. According to the deep separation patch network model of claim 1, it is characterized in that The point-by-point convolution for depthwise separation convolution of vector array X1 uses N1 convolution kernels, each of which is 1×1 in size; The point-by-point convolution for depth-separable convolution of the vector array X2 uses N2 convolution kernels, each with a size of 1×1; the point-by-point convolution for depth-separable convolution of the first-scale patch global fusion feature Y1 uses N1 convolution kernels, each with a size of 1×1; the point-by-point convolution for depth-separable convolution of the second-scale patch global fusion feature Y2 uses N2 convolution kernels, each with a size of 1×1.

5. According to the deep separation patch network model of claim 1, it is characterized in that The multi-layer perceptron includes: an input layer, a discard layer, a hidden layer, a discard layer and a prediction layer which are connected in sequence. The input layer is used to input multi-scale fusion features, and the prediction layer is used to output a moisture content prediction value.

6. The deep separation patch network model according to claim 1, characterized in that: The activation function of the multi-layer perceptron is ReLU.

7. The deep separation patch network model according to claim 1, characterized in that: The number of neurons in the input layer is 1024, the number of neurons in the hidden layer is 256, and the dropout rate of the dropout layer is set to 0.

3.

8. The deep separation patch network model according to claim 1, characterized in that: The loss function value of the deep separation patch network model adopts the L1 loss function.

9. The deep separation patch network model according to claim 1, characterized in that: The length of the first scale patch is P1=50, the length of the second scale patch is P2=250, the number of first scale patches N1=T / P1, the number of second scale patches N2=T / P2, T is the length of the sample, the dimension of each first scale patch after linear transformation is 1×D, the dimension of each second scale patch after linear transformation is 1×D, D=512; The dimension of the first-scale patch global fusion feature Y1 is N1×D, and the dimension of the second-scale patch global fusion feature Y2 is N2×D; The dimension of the first-scale patch global fusion feature Y3 is N1×D, and the dimension of the second-scale patch global fusion feature Y4 is N2×D; The depthwise convolution for performing a depthwise separation convolution operation on the vector array X1 uses N1 convolution kernels, each of which has a size of 1×25; The depthwise convolution for performing depthwise separation convolution on the vector array X2 uses N2 convolution kernels, each with a size of 1×125; the depthwise convolution for performing depthwise separation convolution on the first-scale patch global fusion feature Y1 uses N1 convolution kernels, each with a size of 1×25; the depthwise convolution for performing depthwise separation convolution on the second-scale patch global fusion feature Y2 uses N2 convolution kernels, each with a size of 1×125.

10. A method for measuring water content at a wellhead based on a deep separation patch network model, comprising: The test set is input into the trained deep separation patch network model described in any one of claims 1 to 9 to obtain a moisture content prediction value.

Images