Intelligent real-time monitoring method for oil well liquid production state

Abstract

The application discloses an intelligent real-time monitoring method for oil well liquid production state, comprising the following steps: S1, obtaining a to-be-predicted sample; S2, using a trained OilWellNet network to predict the to-be-predicted sample, and obtaining a prediction result and a water cut prediction value of the to-be-predicted sample. The OilWellNet network comprises a dynamometer card working condition judgment network and a multi-working condition water cut prediction network. The dynamometer card working condition judgment network comprises a convolution layer module, a pooling layer and a full connection layer connected in sequence. The multi-working condition water cut prediction network comprises a feature coding layer, a multi-scale pooling layer, a dynamic division layer, a feature fusion layer and a parameter prediction layer. The intelligent real-time monitoring method for oil well liquid production state can make real-time response to changes under different working conditions through the trained OilWellNet network, thereby optimizing management and decision-making in the oil well production process and accurately predicting the water cut of the oil well.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of oil and gas production engineering and intelligent monitoring, and specifically relates to an intelligent real-time monitoring method for liquid production state of an oil well. BACKGROUND

[0002] At present, the monitoring of the liquid production state of an oil well mainly relies on manual inspection and conventional sensor monitoring. The manual inspection method usually makes a state judgment by periodically recording and analyzing the production parameters of the oil well, such as pressure, temperature and liquid level, but it has the problems of slow data updating, difficulty in meeting real-time requirements, and large data acquisition error due to reliance on manual experience. On the other hand, the conventional sensor monitoring technology usually uses a single type of sensor, such as an electrical conductivity sensor or a densimeter, to monitor the liquid production state, but since a single type of sensor can only provide a single physical parameter, it is difficult to fully characterize the complex production state of the oil well, and the detection accuracy is low under complex working conditions such as high water cut and heavy oil.

[0003] Although the traditional technology can achieve the monitoring of the liquid production state to some extent, it has significant limitations:

[0004] 1. The production state of the oil well is complex and variable, especially under special working conditions such as high water cut, gas invasion and loss, and the traditional method is difficult to accurately distinguish different working conditions.

[0005] 2. In high water cut wells, due to the great change of water-oil ratio, the traditional model cannot capture the water-oil distribution dynamics in real time, resulting in high prediction error of water cut.

[0006] 3. The traditional monitoring system mostly uses offline analysis method, and cannot realize real-time monitoring and early warning function.

[0007] 4. The existing technology usually only processes a single signal, lacks multi-modal data fusion capability, and is difficult to use the synergistic effect of multi-source data to improve the accuracy and robustness of monitoring.

[0008] In recent years, with the rapid development of artificial intelligence technology and sensor technology, intelligent monitoring technology has gradually become the research focus of oilfield production optimization. The successful application of deep learning algorithm in image classification and time series analysis provides a new idea for the intelligent diagnosis of oil well production state, and the multi-modal data fusion technology can integrate the data of different sensors and mine the potential laws in complex production environment. However, the existing intelligent monitoring technology still has the problems of insufficient fine classification of working conditions, lag of data processing and model inference behind production dynamics, and insufficient applicability, especially under extreme working conditions such as high water cut and heavy oil thermal recovery, its performance still needs to be optimized and improved. SUMMARY

[0009] In view of the deficiencies of the prior art, the present application aims to provide an OilWellNet network.

[0010] Another object of the present application is to provide an intelligent real-time monitoring method for oil well liquid production state, which is based on the trained OilWellNet network, combines the indicator diagram image and the standardized microwave difference frequency signal, realizes intelligent identification of the oil well working state and water cut prediction, and thus improves the intelligence and accuracy of the oil well liquid production monitoring.

[0011] The object of the present application is achieved by the following technical solutions.

[0012] An OilWellNet network comprises an indicator diagram working condition judgment network and a multi-condition water cut prediction network, each sample input to the OilWellNet network comprises a pump indicator diagram image collected at T time and a standardized microwave difference frequency signal in a T-t~T time period, the standardized microwave difference frequency signal is obtained by standardizing the microwave difference frequency signal; T is the collection time, and t is in the range of 0.5min to 1min.

[0013] The indicator diagram working condition judgment network comprises a convolution layer module, a pooling layer and a full connection layer connected in sequence, the convolution layer module is used for inputting the pump indicator diagram image of the sample and extracting different scale local features of the pump indicator diagram image, and the convolution layer module outputs a feature X1.

[0014] The pooling layer is used for inputting the feature X1 and outputting a feature X2.

[0015] The full connection layer inputs the feature X2 and outputs a prediction result, the prediction result is used for predicting the working condition of the oil well, and the working condition is sufficient liquid supply, insufficient liquid supply, heavy oil thermal recovery, gas influence or loss influence.

[0016] The multi-condition water cut prediction network is a sufficient liquid supply water cut prediction network, an insufficient liquid supply water cut prediction network, a heavy oil thermal recovery water cut prediction network, a gas influence water cut prediction network or a loss influence water cut prediction network.

[0017] The multi-condition water cut prediction network is selected according to the prediction result of the indicator diagram working condition judgment network: when the prediction result is sufficient liquid supply, the multi-condition water cut prediction network is the sufficient liquid supply water cut prediction network; when the prediction result is insufficient liquid supply, the multi-condition water cut prediction network is the insufficient liquid supply water cut prediction network; when the prediction result is heavy oil thermal recovery, the multi-condition water cut prediction network is the heavy oil thermal recovery water cut prediction network; when the prediction result is gas influence, the multi-condition water cut prediction network is the gas influence water cut prediction network; and when the prediction result is loss influence, the multi-condition water cut prediction network is the loss influence water cut prediction network.

[0018] The structure of the liquid supply sufficient water cut prediction network, the liquid supply insufficient water cut prediction network, the heavy oil thermal recovery water cut prediction network, the gas influence water cut prediction network or the leakage influence water cut prediction network is the same, each multi-working condition water cut prediction network comprises a feature coding layer, a multi-scale pooling layer, a dynamic division layer, a feature fusion layer and a parameter prediction layer;

[0019] The feature coding layer comprises three convolution layers with different convolution kernel sizes and a nonlinear activation function ReLU, the inputs of the three convolution layers in the feature coding layer are the same, each convolution layer in the feature coding layer is used for inputting the normalized microwave difference frequency signal of a sample, and the outputs of the three convolution layers in the feature coding layer are features X3, X4 and X5;

[0020] The input of the nonlinear activation function ReLU is the features X3, X4 and X5, and the nonlinear activation function ReLU obtains outputs X3, X4 and X5;

[0021] The multi-scale pooling layer uses a pyramid pooling method to aggregate features of the outputs X3, X4 and X5, and obtains a feature X6;

[0022] The dynamic division layer divides the feature X6 into a plurality of patches and maps each patch to a d model dimensional space, and obtains a plurality of patch features;

[0023] The feature fusion layer adopts a multilayer perception MLP to extract and fuse high-order features of the plurality of patch features, and obtains a feature X7;

[0024] The parameter prediction layer adopts a Sigmoid activation function to perform nonlinear transformation on the feature X7, and outputs a water cut prediction value.

[0025] In the above technical solution, the convolution layer module is three convolution layers connected in sequence, and the convolution kernel sizes of the three convolution layers connected in sequence from front to back are 50x50, 30x30 and 10x10.

[0026] In the above technical solution, the fully connected layer adopts a Softmax activation function to flatten the feature X2.

[0027] In the above technical solution, the convolution kernel sizes of the three convolution layers in the feature coding layer are 50x1, 30x1 and 10x1 in sequence.

[0028] In the above technical solution, the pooling scales of the multi-scale pooling layer are 25x1, 50x1, 75x1 and 100x1 in sequence.

[0029] In the above technical solution, the length of each patch is 50, and d model is 256.

[0030] In the technical solution, the standardized microwave difference frequency signal has a dimension of Lx1, and L is the length of the standardized microwave difference frequency signal.

[0031] In the technical solution, the microwave difference frequency signal is collected by a high-frequency microwave sensor installed at the wellhead of the oil well.

[0032] In the technical solution, the pump indicator diagram image is collected by an indicator diagram instrument installed at the wellhead of the oil well.

[0033] In the technical solution, the microwave difference frequency signal is standardized by using the standard deviation.

[0034] An intelligent real-time monitoring method for the production state of an oil well comprises the following steps:

[0035] S1, obtaining a to-be-predicted sample, the to-be-predicted sample comprising a pump indicator diagram image collected at a time T' and a standardized microwave difference frequency signal in a time period of T'-t to T';

[0036] S2, using the trained OilWellNet network to predict the to-be-predicted sample to obtain a prediction result and a water cut prediction value of the to-be-predicted sample.

[0037] In the technical solution, the method for obtaining the trained OilWellNet network comprises the following steps:

[0038] Step 1, preparing a plurality of samples for inputting into the OilWellNet network as a training set; each sample in the training set corresponds to a label value and a working condition label, wherein the label value is the true value of the water cut of the oil wellhead production fluid, and the working condition label is an actual working condition obtained according to the working state of the oil well, and the working condition label is sufficient supply, insufficient supply, heavy oil thermal recovery, gas influence or leakage influence.

[0039] Sufficient supply refers to a working condition in which the supply of liquid in the formation is sufficient to meet the production demand; insufficient supply refers to a working condition in which the supply of liquid in the formation is insufficient, and in this working condition, the wellbore liquid level may be too low, the oil pumping unit may be empty pumping or the production efficiency may be reduced; heavy oil thermal recovery refers to a working condition in which, in the process of heavy oil recovery, a thermal recovery technology (such as steam stimulation, steam flooding or fire flooding) is used to improve the flowability of heavy oil, reduce the viscosity and improve the recovery efficiency; gas influence refers to the interference to the fluid flow caused by the presence of gas in the production process, which usually includes the influence of dissolved gas, free gas or externally injected gas; leakage influence refers to the escape of liquid or gas due to formation leakage or wellbore leakage, and the leakage influence working condition may significantly affect the wellbore pressure, production and recovery efficiency, and needs to be diagnosed and repaired in time.

[0040] Step 2, the training set is substituted into the OilWellNet network for training, and the OilWellNet network outputs a prediction result and a water cut prediction value during the training, a loss function of the multi-working condition water cut prediction network is calculated according to the water cut prediction value and the label value, a loss function value of the indicator diagram working condition judgment network is calculated according to the prediction result and the working condition label, and the total loss function of the OilWellNet network is the sum of the loss function of the multi-working condition water cut prediction network and the loss function of the indicator diagram working condition judgment network, and the parameters of the OilWellNet network are updated through a gradient back propagation method.

[0041] In the technical solution, the loss function of the indicator diagram working condition judgment network adopts a cross-entropy loss function.

[0042] In the technical solution, the loss function of the multi-working condition water cut prediction network adopts a mean square error function.

[0043] Compared with the prior art, the present application has the following beneficial effects:

[0044] 1. The OilWellNet network combines the indicator diagram working condition judgment network and the multi-working condition water cut prediction network, the indicator diagram working condition judgment network analyzes the pump indicator diagram image through a deep convolutional neural network (CNN) to accurately judge the working state of the oil well, and provides accurate working condition information for subsequent water cut prediction, and the multi-working condition water cut prediction network accurately predicts the water cut of the oil well based on the normalized microwave difference frequency signal under different working conditions.

[0045] 2. The oil well liquid production state intelligent real-time monitoring method uses the OilWellNet network, selects and activates the multi-working condition water cut prediction network corresponding to the working condition according to the real-time monitored working state of the oil well, the selection mechanism effectively improves the prediction accuracy, and can make real-time response to changes under different working conditions, thereby optimizing the management and decision-making in the oil well production process. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 The OilWellNet network is a schematic diagram. DETAILED DESCRIPTION

[0047] The oil well liquid production state intelligent real-time monitoring method of the present application will be described in detail below in combination with the drawings and examples.

[0048] Example 1

[0049] An OilWellNet network is as follows: Figure 1As shown, the OilWellNet network comprises a dynamometer card working condition judgment network and a multi-working condition water cut prediction network, each sample input to the OilWellNet network comprises a pump dynamometer card image collected at T time and a normalized microwave difference frequency signal in a T-t~T time period, the normalized microwave difference frequency signal is obtained by normalizing the microwave difference frequency signal; T is the collection time;

[0050] The dynamometer card working condition judgment network comprises a convolution layer module, a pooling layer and a full connection layer connected in sequence, the convolution layer module is used for inputting the pump dynamometer card image of the sample and extracting different scale local features of the pump dynamometer card image, and the convolution layer module outputs a feature X1, wherein the convolution layer module is three convolution layers connected in sequence, and the convolution kernel sizes of the three convolution layers connected in sequence from front to back are 50*50, 30*30 and 10*10 respectively;

[0051] The pooling layer is used for inputting the feature X1 and outputting a feature X2, and the pooling layer adopts a maximum pooling operation to reduce the spatial size of the feature X1;

[0052] The full connection layer inputs the feature X2 and outputs a prediction result, the full connection layer adopts a Softmax activation function to flatten the feature X2, and the prediction result is used for predicting the working condition of the oil well, and the working condition is liquid supply sufficient, liquid supply insufficient, heavy oil thermal recovery, gas influence or loss influence.

[0053] The multi-working condition water cut prediction network is a liquid supply sufficient water cut prediction network, a liquid supply insufficient water cut prediction network, a heavy oil thermal recovery water cut prediction network, a gas influence water cut prediction network or a loss influence water cut prediction network, the multi-working condition water cut prediction network is selected according to the prediction result of the dynamometer card working condition judgment network, and the water cut of the oil well produced liquid under the working condition is predicted. When the prediction result is liquid supply sufficient, the multi-working condition water cut prediction network is the liquid supply sufficient water cut prediction network; when the prediction result is liquid supply insufficient, the multi-working condition water cut prediction network is the liquid supply insufficient water cut prediction network; when the prediction result is heavy oil thermal recovery, the multi-working condition water cut prediction network is the heavy oil thermal recovery water cut prediction network; when the prediction result is gas influence, the multi-working condition water cut prediction network is the gas influence water cut prediction network; and when the prediction result is loss influence, the multi-working condition water cut prediction network is the loss influence water cut prediction network;

[0054] The liquid supply sufficient water cut prediction network, the liquid supply insufficient water cut prediction network, the heavy oil thermal recovery water cut prediction network, the gas influence water cut prediction network or the loss influence water cut prediction network have the same structure, and each multi-working condition water cut prediction network comprises a feature coding layer, a multi-scale pooling layer, a dynamic division layer, a feature fusion layer and a parameter prediction layer;

[0055] The feature encoding layer comprises three convolution layers with different convolution kernel sizes and a nonlinear activation function ReLU, the inputs of the three convolution layers in the feature encoding layer are the same, each convolution layer in the feature encoding layer is used for the standardized microwave difference frequency signal of the input sample, the outputs of the three convolution layers in the feature encoding layer are features X3, X4 and X5, wherein the convolution kernel sizes of the three convolution layers in the feature encoding layer are 50x1, 30x1 and 10x1 in turn, the feature encoding layer extracts the local time sequence features of the input signal through the three convolution layers, and each convolution layer captures different local features of the input signal.

[0056] The input of the nonlinear activation function ReLU is the features X3, X4 and X5, the nonlinear activation function ReLU processes the features X3, X4 and X5, and the nonlinear activation function ReLU obtains outputs X3, X4 and X5.

[0057] The multi-scale pooling layer uses a pyramid pooling method to aggregate features of the outputs X3, X4 and X5 to obtain a feature X6, and the pooling scales of the multi-scale pooling layer are 25x1, 50x1, 75x1 and 100x1 in turn, and the pyramid pooling method enhances the multi-scale perception ability of the multi-working-condition water cut prediction network.

[0058] The dynamic division layer divides the feature X6 into a plurality of patches and maps each patch to a d model dimensional space to obtain a plurality of patch features, the dynamic division layer captures local patterns of the feature X6 and retains global relationships, the length of each patch is 50, and d model is 256.

[0059] The feature fusion layer adopts a multi-layer perception MLP to extract and fuse high-order features of the plurality of patch features to obtain a feature X7.

[0060] The parameter prediction layer adopts a Sigmoid activation function to perform nonlinear transformation on the feature X7 to output a water cut prediction value.

[0061] The standardized microwave difference frequency signal has a dimension of Lx1; the feature X6 has a dimension of Lx1; the plurality of patch features have a dimension of (L / 50)x50x256; and the feature X7 has a dimension of 256x1, wherein L is the length of the standardized microwave difference frequency signal.

[0062] Embodiment 2

[0063] The microwave difference frequency signal is obtained by collecting a high-frequency microwave sensor installed at the wellhead of the oil well, and the high-frequency microwave sensor is not limited herein, for example, a double-helix high-frequency microwave sensor in CN202311628698.7 can be used, and the pump indicator diagram image is obtained by collecting an indicator diagram instrument installed at the wellhead of the oil well.

[0064] Preparing a multi-modal data set, a method for obtaining a multi-modal data set, comprising the following steps:

[0065] S1, the high-frequency microwave sensor collects 100 times, and collects the microwave difference frequency signal every 1 hour, 5 minutes each time;

[0066] In the 5 minutes of collecting the high-frequency microwave sensor, the indicator diagram of the pump is collected every 1 minute (the size of each pump indicator diagram is 224x224), a total of 500 times of collection (500 pump indicator diagrams); each time the indicator diagram of the pump is collected, the oil wellhead production fluid is sampled (a total of 500 times of manual sampling), and the oil wellhead production fluid of all manual sampling is artificially tested to detect the water content of the oil wellhead production fluid, and is used as a test value, 500 test values are obtained, and one test value is used as a label value.

[0067] Among them, the sampling frequency of the high-frequency microwave sensor is 1000Hz, and the sampling frequency of the indicator diagram of the pump is 1Hz.

[0068] S2, using a synchronous alignment technology (Wang Guojing, Wang Weidong. Data synchronous acquisition method between multi-lead sleep monitoring equipment and intestinal sound equipment [J]. China medical equipment, 2024, 39 (05): 9-14.) to time align the collected microwave difference frequency signal and the collected pump indicator diagram;

[0069] S3, denoising, screening out abnormal and converting the image format to RGB for the time-aligned pump indicator diagram, and labeling the RGB format pump indicator diagram according to the working state of the oil well to obtain the working condition label of each pump indicator diagram, the working condition label includes: "sufficient liquid supply", "insufficient liquid supply", "thick oil hot mining", "gas influence" and "leakage influence";

[0070] The time-aligned microwave difference frequency signal is standardized by using a dispersion standardization (dispersion standardization, see: Xu Yifang, Chen Jin, Li Lin, etc. Credit prediction based on Isomap fusion naive Bayes classifier [J]. Computer knowledge and technology, 2021, 17 (35): 125-126+139.), that is, scaled to [0, 1] range, and the standardized microwave difference frequency signal is segmented according to the collection time sequence by using a sliding window method (Zhao Yuxuan, Gong Jianning, Liu Han, etc. Front-end processing of mobile robot based on multi-source sensing information fusion [J]. Manufacturing automation, 2023, 45 (04): 217-220.), and the sliding window length is set to L=6x10 4 and non-overlapping segmentation, so that the microwave difference frequency signal collected in 5 minutes obtains 5 standardized microwave difference frequency signals;

[0071] This method yields 500 samples. Each sample consists of a pump dynamometer image acquired at time T and a normalized microwave difference frequency signal from Tt to T (t = 1 minute). Each sample corresponds to a working condition label and a label value. These 500 samples constitute a multimodal dataset.

[0072] Example 3

[0073] Based on Example 2, a method for training and evaluating an OilWellNet network includes the following steps: training and validating the OilWellNet network using a ten-fold cross-validation method on a multimodal dataset to obtain a trained OilWellNet network. Nine samples of the multimodal dataset are used as training sets, and one sample is used as a validation set, so that all samples in the multimodal dataset are used as the validation set.

[0074] Each fold is trained for 100 rounds. During each fold training, the OilWellNet network predicts the working conditions and water cut of the oil wells on the training set, and outputs the prediction results and water cut prediction values. The loss function of the multi-condition water cut prediction network is calculated based on the water cut prediction values ​​and label values. The loss function value of the dynamometer diagram working condition judgment network is calculated based on the prediction results and working condition labels. The loss function values ​​of the multi-condition water cut prediction network and the loss function values ​​of the dynamometer diagram working condition judgment network are added together to obtain the total loss function value. The parameters of the OilWellNet network are updated using the gradient backpropagation method.

[0075] Among them, the loss function of the dynamometer working condition judgment network adopts the cross entropy loss function (Ren Jinjun, Wang Ning. Research on loss function in artificial neural network[J]. Journal of Gansu Normal University, 2018, 23(02):61-63.);

[0076] The loss function of the multi-condition moisture content prediction network adopts the mean square error function (MSE);

[0077] The total loss function of the OilWellNet network is the sum of the cross entropy loss function and the mean square error (MSE) function.

[0078] Example 4

[0079] An intelligent real-time monitoring method for oil well fluid production status, comprising:

[0080] S1, obtaining 50 samples to be predicted, each sample to be predicted includes: a pump dynamometer image acquired at time T' and a normalized microwave difference frequency signal within the time period T'-t to T', t = 1 minute;

[0081] S2, using the OilWellNet network trained in example 3, the prediction of the to-be-predicted sample is carried out, and the prediction result and the water cut prediction value of the to-be-predicted sample are obtained.

[0082] Example 5

[0083] An intelligent real-time monitoring method of oil well liquid production state is basically consistent with example 4, and the only difference is that only the water cut prediction network in example 1 is reserved. In this example, each to-be-predicted sample is a normalized microwave difference frequency signal in the T'-t-T' time period.

[0084] Example 6

[0085] An intelligent real-time monitoring method of oil well liquid production state is basically consistent with example 4, and the only difference is that the convolution layer module of the dynamometer card working condition judgment network in example 1 is replaced by a convolution layer with a convolution kernel size of 50*50, and the three convolution layers of the feature encoding layer of the water cut prediction network in example 1 are replaced by a convolution layer with a convolution kernel size of 50*1.

[0086] Example 7

[0087] An intelligent real-time monitoring method of oil well liquid production state is basically consistent with example 4, and the only difference is that the multi-scale pooling layer of the water cut prediction network in example 1 is replaced by a pooling layer, and the pooling scale of the pooling layer is 50*1.

[0088] The mean square error (MSE), mean absolute error (MAE) and mean absolute percentage error (MAPE) are used to evaluate example 4, example 5, example 6 and example 7, and the results are as follows:

[0089] Table 1

[0090]

[0091]

[0092] As shown in table 1, the OilWellNet network of the present application combines the dynamometer card working condition judgment network with the multi-condition water cut prediction network, effectively judges the working condition of the oil well, provides accurate working condition information for subsequent water cut prediction, and through the multi-scale analysis technology, fuses and extracts multi-scale features, significantly improves the accuracy of water cut prediction.

[0093] The above is an exemplary description of the present application, it should be noted that without departing from the core of the present application, any simple modification, modification or other equivalent replacement which can not cost creative labor of those skilled in the art falls within the protection scope of the present application.

Claims

1. An intelligent real-time monitoring method for oil well production status, characterized in that: include: S1, obtaining a sample to be predicted, the sample to be predicted including: a pump dynamometer image acquired at time T′ and a normalized microwave difference frequency signal within a time period from T′-t to T′; S2, using the trained OilWellNet network to predict the sample to be predicted, and obtain the prediction result and water content prediction value of the sample to be predicted; The OilWellNet network includes: a dynamometer diagram operating condition judgment network and a multi-operating condition water content prediction network. Each sample input to the OilWellNet network includes: a pump dynamometer diagram image acquired at time T and a standardized microwave difference frequency signal within the time period Tt to Tt. The standardized microwave difference frequency signal is obtained by standardizing the microwave difference frequency signal. T is the acquisition time. The dynamometer working condition judgment network includes: a convolutional layer module, a pooling layer, and a fully connected layer connected in sequence. The convolutional layer module is used to input the sample pump dynamometer image and extract local features of different scales of the pump dynamometer image. The convolutional layer module outputs feature X1; The pooling layer is used to input feature X1 and output feature X2; The fully connected layer inputs feature X2 and outputs prediction results, which are used to predict the operating conditions of the oil well, including sufficient fluid supply, insufficient fluid supply, heavy oil thermal recovery, gas impact, or leakage impact. The multi-condition water cut prediction network includes the water cut prediction network for sufficient fluid supply, the water cut prediction network for insufficient fluid supply, the water cut prediction network for heavy oil thermal recovery, the water cut prediction network affected by gas, or the water cut prediction network affected by leakage; The multi-condition water cut prediction network is selected according to the prediction result of the indicator diagram working condition judgment network: when the prediction result is sufficient fluid supply, the multi-condition water cut prediction network is a sufficient fluid supply water cut prediction network; when the prediction result is insufficient fluid supply, the multi-condition water cut prediction network is a insufficient fluid supply water cut prediction network; when the prediction result is heavy oil thermal recovery, the multi-condition water cut prediction network is a heavy oil thermal recovery water cut prediction network; when the prediction result is gas influence, the multi-condition water cut prediction network is a gas influence water cut prediction network; when the prediction result is leakage influence, the multi-condition water cut prediction network is a leakage influence water cut prediction network; The structures of the water cut prediction network for sufficient fluid supply, insufficient fluid supply, heavy oil thermal recovery, gas-affected water cut, or leakage-affected water cut prediction networks are the same. Each multi-condition water cut prediction network includes: a feature encoding layer, a multi-scale pooling layer, a dynamic partitioning layer, a feature fusion layer, and a parameter prediction layer. The feature coding layer includes: three convolutional layers with different convolution kernel sizes and a nonlinear activation function ReLU. The inputs of the three convolutional layers in the feature coding layer are the same. Each convolutional layer in the feature coding layer is used to input the standardized microwave difference frequency signal of the sample. The outputs of the three convolutional layers in the feature coding layer are feature X3, feature X4, and feature X5. The input of the nonlinear activation function ReLU is feature X3, feature X4 and feature X5, and the nonlinear activation function ReLU obtains output X3, output X4 and output X5; The multi-scale pooling layer uses the pyramid pooling method to aggregate the features of output X3, output X4 and output X5 to obtain feature X6; The dynamic partitioning layer divides the feature X6 into multiple patches and maps each patch to d model Dimension, get multiple patch features; The feature fusion layer uses a multi-layer perceptron (MLP) to extract and fuse high-order features from multiple patch features to obtain feature X7. The parameter prediction layer uses the Sigmoid activation function to perform nonlinear transformation on feature X7 and output the moisture content prediction value.

2. The intelligent real-time monitoring method for oil well production status according to claim 1, characterized in that: The convolution layer module consists of three sequentially connected convolution layers, and the convolution kernel sizes of the three sequentially connected convolution layers from front to back are 50×50, 30×30 and 10×10 respectively.

3. The intelligent real-time monitoring method for oil well production status according to claim 1, characterized in that: The fully connected layer uses the Softmax activation function to flatten the feature X2.

4. The intelligent real-time monitoring method for oil well production status according to claim 1, characterized in that: The convolution kernel sizes of the three convolution layers in the feature coding layer are 50×1, 30×1 and 10×1 respectively.

5. The intelligent real-time monitoring method for oil well production status according to claim 1, characterized in that: The pooling scales of the multi-scale pooling layer are 25×1, 50×1, 75×1 and 100×1, respectively.

6. The intelligent real-time monitoring method for oil well production status according to claim 1, characterized in that: The length of each patch output by the dynamic partitioning layer is 50, d ,odel is 256.

7. The intelligent real-time monitoring method for oil well production status according to claim 1, characterized in that: The microwave difference frequency signal is acquired by a high-frequency microwave sensor installed at the wellhead of the oil well, and the pump dynamometer image is acquired by a dynamometer installed at the wellhead of the oil well. The microwave difference frequency signal is standardized using deviation standardization, and the dimension of the standardized microwave difference frequency signal is L×1, where L is the length of the standardized microwave difference frequency signal.

8. The intelligent real-time monitoring method for oil well production status according to claim 1, characterized in that: Methods for obtaining the trained OilWellNet network include: Step 1: Prepare multiple samples for input into the OilWellNet network as a training set. Each sample in the training set corresponds to a label value and an operating condition label. The label value is the true value of the water content of the oil wellhead production fluid, and the operating condition label is the actual operating condition obtained based on the oil well operating status. The operating condition labels include sufficient fluid supply, insufficient fluid supply, heavy oil thermal recovery, gas impact, or leakage impact. Step 2: Substitute the training set into the OilWellNet network for training. During training, the OilWellNet network outputs the prediction results and water content prediction values. The loss function of the multi-condition water content prediction network is calculated based on the water content prediction values ​​and label values. The loss function value of the dynamometer working condition judgment network is calculated based on the prediction results and working condition labels. The total loss function of the OilWellNet network is the sum of the loss function of the multi-condition water content prediction network and the loss function of the dynamometer working condition judgment network. The parameters of the OilWellNet network are updated by the gradient back propagation method.

9. The method for obtaining a trained OilWellNet network according to claim 8, wherein: The loss function of the dynamometer diagram working condition judgment network adopts the cross entropy loss function, and the loss function of the multi-working condition moisture content prediction network adopts the mean square error function.

Images