A method for monitoring the fluid supply capacity of oil wells based on dynamometer diagram

By applying the deep learning network model of Gaussian heat map, feature pyramid and attention mechanism on the demonstration diagram, the Ver opening and closing points are identified, and the accuracy and cost of the oil production engine's liquid supply capacity measurement is solved, and the oil production engine supply is realized Accurate monitoring of liquid capacity and energy conservation and increase production.

CN114581697BActive Publication Date: 2024-11-26NANJING FUDAO OIL & GAS INTELLIGENT CONTROL TECH CO LTD

Patent Information

Application Number
CN202210082358.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-01-24
Publication Date
2024-11-26
Estimated Expiration
2042-01-24

AI Technical Summary

Technical Problem

In low-permeability oil fields, the oil pumping capacity of the oil harvester is difficult to accurately match the fluid supply capacity of the bottom layer of the oil well, resulting in low pumping efficiency, large power consumption and high oil pumping cost. The existing technology is positioned at the Ver opening and closing point in the power diagram. There are problems of poor real-time and high labor costs.

Method used

A deep learning network model integrating Gaussian heat map, feature pyramid and attention mechanism is used to identify the four Vere opening and closing points of the power map, and calculate the fluid supply capacity of the oil harvester in real time, improving measurement accuracy and reducing costs.

Benefits of technology

It realizes accurate monitoring of the fluid supply capacity of the oil harvester, provides reliable data support, provides energy-saving and production-enhancing technical support for the oil harvester's flush adjustment, and reduces measurement costs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114581697B_ABST
    Figure CN114581697B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for monitoring the fluid supply capacity of an oil well based on an indicator diagram. The method is based on the indicator diagram generated during the oil production process, and a convolutional neural network regression model is built to identify the four valve opening and closing points of the indicator diagram. By integrating Gaussian heat maps, feature pyramids, and attention mechanisms in the network model to meet the requirements of fine-grained and high-precision in the valve opening and closing point identification task, the four valve opening and closing points can be accurately and efficiently identified, and the fluid supply capacity of the oil well can be quantified in real time based on this, providing reliable technical support for the subsequent stroke adjustment and intermittent pumping control of the oil well. Compared with traditional methods, this method significantly improves the real-time recognition, meets the needs of enterprise engineering applications, and has good application value in reducing the energy consumption of the pumping system and increasing oil field production.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the field of oil production, in particular to a method for monitoring the fluid supply capacity of an oil production well by locating valve opening and closing points of the oil production machine through a ground indicator diagram of the oil production machine. Background Art

[0002] Most of my country's oil fields are low-permeability oil fields, with low abundance, low permeability, and low single-well production capacity, which makes it difficult to accurately match the pumping capacity of the current oil production machine with the bottom fluid supply capacity of the oil well. The oil production machine has low pump efficiency, high power consumption, and high pumping costs. Enterprises need to analyze the fluid supply capacity of the oil production machine well in real time and formulate a stroke adjustment strategy to achieve an accurate match between the oil production machine's pumping capacity and the oil well's fluid supply capacity, so as to achieve the goal of energy saving and increased production.

[0003] In the oil production process, companies traditionally use the dynamic liquid level to measure the fluid supply capacity of the oil well, but the measurement cost of the dynamic liquid level is high and the measurement result is not accurate when the dynamic liquid level is deep or affected by gas. With the continuous advancement of dynamometer test technology, companies currently rely on manual experience to locate the valve opening and closing points from the dynamometer and calculate the fluid supply capacity of the oil well. Although this method makes up for the shortcomings of traditional methods, it still has problems such as poor real-time performance and high labor costs.

[0004] In recent years, the development of artificial intelligence technology has provided technical support for the quantitative analysis of the fluid supply capacity of oil wells. The convolutional neural network regression model can be used for the regression task of key points, which makes it possible to locate the valve opening and closing points of the oil production machine from the dynamometer diagram. However, previous studies have shown that if the convolutional neural network is simply applied to the fine-grained, high-precision task of identifying the valve opening and closing points of the dynamometer diagram, it is very easy to cause the network model to have deficiencies in accuracy and spatial generalization ability. For example, our company's prior application (CN202110927362.5 A method for calculating the effective stroke of an oil pump) provides a model for obtaining the effective stroke of an oil pump, but because the deep learning model in this method does not fully take into account the particularity of the valve opening and closing point regression task, it has deficiencies in spatial generalization ability and calculation accuracy.

[0005] Therefore, a convolutional neural network regression model customized for the Vertex opening and closing point recognition task is urgently needed to solve the above problems. Summary of the invention

[0006] In view of the above problems, the present invention discloses a method for monitoring the fluid supply capacity of an oil production well based on a dynamometer diagram. The method is based on the dynamometer diagram generated during the oil production process, and adopts a deep learning network model that integrates Gaussian heat maps, feature pyramids, and attention mechanisms to identify the four valve opening and closing points of the dynamometer diagram and calculate the fluid supply capacity of the oil production well in real time.

[0007] This method has the following steps:

[0008] 1) Obtain the dynamometer data set S of the historical operating condition records of the oil production machine:

[0009] S={S i |i=1,2,...,N}

[0010] Where S i represents the i-th dynamometer image sample, a total of N images;

[0011] 2) Calibrate the vane opening and closing points for the dynamometer data set, and use image processing to generate the pixel coordinate value set P of the vane opening and closing points:

[0012] P={P i |i=1,2,...,N}

[0013] P i ={(r ij ,c ij )|j=1,2,...,4}

[0014] Where P i Represents the pixel coordinates of the four opening and closing points of the i-th indicator diagram image in the indicator diagram data set, (r ij ,c ij ) corresponds to the pixel coordinate value of a single Vertex opening and closing point;

[0015] 3) The pixel coordinate value set P of the Vertex opening and closing points is normalized as follows and used as the label set L of the deep learning network model:

[0016]

[0017]

[0018] Where (w,h) is the pixel resolution of the dynamometer diagram;

[0019] 4) Use the holdout method to divide the calibrated dynamometer data set S into a training set S train , validation set S val and the test set S test ;

[0020] 5) Build a deep learning network model for predicting Gaussian heatmaps to meet the requirements of spatial generalization and translation equivalence for the Vertex opening and closing point recognition task;

[0021] 6) Introducing the feature pyramid network into the Gaussian heat map prediction model to solve the theoretical error lower bound problem caused by the resolution of the Gaussian heat map;

[0022] 7) An attention mechanism module is embedded in the feature extraction network of the Gaussian heat map prediction model to meet the requirements of fine-grained and high-precision Vertex opening and closing point recognition tasks;

[0023] 8) The built deep learning network model is trained on the training set S train and validation set S val Learning network parameters, based on the test set S test The results evaluate its generalization ability and obtain the final network model;

[0024] 9) Input the indicator diagram of the real-time working condition record of the oil production machine into the network model to obtain the corresponding valve opening and closing point label (x rj ,y rj );

[0025] 10) Calculate the pixel coordinates of the Vernal opening and closing points according to the Vernal opening and closing point labels (r rj ,c rj ):

[0026]

[0027]

[0028] 11) Calculate the effective stroke S based on the pixel coordinates of the valve opening and closing points pe :

[0029]

[0030] Where S max It is the maximum stroke of the oil production equipment;

[0031] 12) Determine the fluid supply capacity index β based on the ratio of the effective stroke to the stroke S during oil extraction machine operation:

[0032]

[0033] In this method, the pixel coordinates of the valve opening and closing points of the dynamometer diagram data set include: the coordinates of the fixed valve opening point A (r1, c1), the coordinates of the fixed valve closing point B (r2, c2), the coordinates of the mobile valve opening point C (r3, c3) and the coordinates of the mobile valve closing point D (r4, c4).

[0034] In this method, the pixel coordinate values ​​of the van der Waals opening and closing points are normalized to [-1, 1] based on the work diagram image resolution (w, h), thereby eliminating the dimension effect and making the deep learning network model easier to converge.

[0035] In this method, the allocation ratio of the holdout method is 6:2:2.

[0036] In this method, when building a deep learning network model for predicting Gaussian heat maps, in order to ensure that the theoretical error lower bound is as small as possible while reducing the number of model parameters, the resolution of the predicted Gaussian heat map is controlled to 1 / 4 of the resolution of the indicator diagram, that is,

[0037] In this method, the attention mechanism modules embedded in the deep learning network model are divided into spatial attention mechanism and channel attention mechanism, so as to increase the representation ability of the backbone network, so that the network model can locate the image area that is valuable for identifying the opening and closing points of the vanes, and pay attention to the features in the area that are more valuable for identifying the opening and closing points of the vanes.

[0038] Beneficial effects:

[0039] The method provided by the present invention, based on the dynamometer diagram generated during the oil production process, builds a deep learning network model to identify the four valve opening and closing points of the dynamometer diagram and calculates the fluid supply capacity of the oil production well in real time, thereby reducing the cost of measuring the fluid supply capacity of the oil production well.

[0040] Furthermore, the construction of the deep learning network model integrates Gaussian heat map, feature pyramid and attention mechanism, and improves the measurement accuracy with a customized convolutional neural network regression model. It provides reliable data support for the subsequent oil production machine stroke adjustment and has important application value for energy saving and production increase in oil fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 is a schematic diagram of calibration of a dynamometer data set in an embodiment;

[0042] Figure 2 is a structural diagram of a deep learning network model for identifying valve opening and closing points in an embodiment;

[0043] Figure 3 is a loss function diagram during deep learning network model training in an embodiment;

[0044] Figure 4 is a test result diagram of the deep learning network model after training on the test set in the embodiment;

[0045] Figure 5 It is a sample indicator diagram under real-time working conditions in the embodiment.

[0046] Specific implementation process

[0047] The following is a detailed calculation process and specific operation flow in conjunction with the accompanying drawings and specific examples to further illustrate the present invention. This implementation case is implemented based on the technical solution of the present invention, but the protection scope of the present invention is not limited to the following implementation case.

[0048] This case takes an oil production company as an example to illustrate the effectiveness and implementation process of this method. The training set, validation set, and test set are constructed using data from July 2020 to December 2020 to analyze the effectiveness of the method in identifying the fluid supply capacity of the pumping unit.

[0049] The specific implementation steps are as follows:

[0050] (1) A total of 8672 dynamometer diagram data sets with historical operating records of the oil production units were obtained.

[0051] (2) Calibrate the valve opening and closing points for the dynamometer data set as follows Figure 1 As shown, image processing is used to generate pixel coordinate values ​​of the Vall opening and closing points as label values ​​of the dynamometer diagram, as shown in Table 1. In this embodiment, w=h=256 is taken.

[0052] Table 1 shows the pixel coordinate values ​​of the valve opening and closing points on the power diagram

[0053] Vernal opening and closing point Pixel coordinate value Fixed the opening point A of the Val (30.23,41.15) Fixed valve closing point B (246.77,57.69) Swimming Vale Opening Point C (123.69,182.31) Moving valve closing point D (6.38,159.23)

[0054] (3) The pixel coordinate values ​​of the Van der Waals opening and closing points are normalized according to equations (1) and (2) as the labels of the deep learning network model, as shown in Table 2.

[0055] Table 2 Valve opening and closing point labels of power diagram

[0056] Vernal opening and closing point Tag value Fixed the opening point A of the Val (-0.764,-0.679) Fixed valve closing point B (0.928,-0.548) Swimming Vale Opening Point C (-0.034,0.424) Moving valve closing point D (-0.950,0.244)

[0057] (4) The hold-out method is used to divide the calibrated dynamometer data set into 5204 training sets, 1734 validation sets, and 1734 test sets according to the ratio of 6:2:2.

[0058] (5) Aiming at the requirements of spatial generalization ability and translation equivalence for the Vertex opening and closing point recognition task, a deep learning network model for predicting Gaussian heat maps is built, and the feature pyramid network and attention mechanism modules are integrated into it. The overall structure of the model is as follows: Figure 2 As shown. The whole model uses ResNet50 embedded with attention mechanism as the feature extraction network, where C1 is the initial convolution module, C2, C3, C4 and C5 are Bottleneck composed of residual blocks, and CBAM is the attention mechanism module. After extracting the features of the dynamometer diagram, the model uses the feature pyramid network FPN to fuse features from top to bottom in one direction, where M2, M3, M4 and M5 are the fused feature maps, and P2, P3, P4 and P5 correspond to the feature maps of M2, M3, M4 and M5 after 3x3 convolution to eliminate the aliasing effect. In this embodiment, P2 is selected as the input of the final Gaussian heat map prediction module, in which HeatMap represents the Gaussian heat map spectrum, and DSNT is a conversion module that converts HeatMap into coordinate point output.

[0059] (6) The built deep learning network model is trained and verified through learning network parameters. The loss function value during the training process is as follows: Figure 3 As shown in the figure, train_loss represents the trend curve of the loss function value of the training set, and val_loss represents the trend curve of the loss function value of the validation set.

[0060] (7) Input the test set into the trained deep learning network model and obtain the recognition result of the network model on the test set as follows: Figure 4 shown.

[0061] (8) Select the indicator diagram of the real-time operating condition record of an oil well of an oil production enterprise in February 2021, input it into the trained deep learning network model, and obtain the corresponding valve opening and closing point prediction label value. Figure 5 Taking the dynamometer diagram shown in the figure as an example, after inputting it into the network model, the output is {(-0.88,-0.68),(0.90,-0.52),(-0.67,0.01),(-0.90,-0.06)}, which correspond to the label value of the fixed valve opening point A (-0.88,-0.68), the label value of the fixed valve closing point B (0.90,-0.52), the label value of the floating valve opening point C (-0.67,0.01) and the label value of the floating valve closing point D (-0.90,-0.06). After calculating these label values ​​according to formula (3) and (4), the pixel coordinates of the four points are: the pixel coordinates of the fixed valve opening point A (15.36, 40.96), the pixel coordinates of the fixed valve closing point B (243.20, 61.44), the pixel coordinates of the floating valve opening point C (42.24, 129.28) and the pixel coordinates of the floating valve closing point D (12.8, -120.32). According to the pixel coordinates of these four valve opening and closing points and formula (5), the effective stroke S is calculated. pe for:

[0062]

[0063] Then, according to formula (6), the fluid supply capacity index β of the oil well is calculated as:

[0064]

[0065] Check that the liquid production measured by the wellhead measuring device is 4.4m 3 The theoretical liquid production of the well is 36.7m 3 , the calculated pump efficiency is 11.98%, which is very close to the value of the fluid supply capacity index β calculated above. It can be seen that the calculation results of this method can fully and accurately reflect the actual fluid supply capacity of the oil well.

[0066] From the above analysis, it can be seen that the method for monitoring the fluid supply capacity of the oil production machine of the present invention can monitor the fluid supply capacity of the oil production machine in real time and accurately, and provides reliable technical support for dynamically adjusting the stroke frequency of the oil production machine.

Claims

1. A method for monitoring the fluid supply capacity of oil wells based on a dynamometer diagram. Features The method is based on the indicator diagram generated during the oil production process. The deep learning network model is used to identify the four valve opening and closing points of the indicator diagram and calculate the oil production machine's fluid supply capacity in real time based on the four valve opening and closing points. The method has the following steps: 1) Obtain the dynamometer data set S of the historical operating condition records of the oil production machine: S={S i |i=1,2,...,N} Where S i represents the i-th dynamometer image sample, a total of N images; 2) Calibrate the vane opening and closing points for the dynamometer data set, and use image processing to generate the pixel coordinate value set P of the vane opening and closing points: P={P i |i=1,2,...,N} P i ={(r ij ,c ij )|j=1,2,...,4} Where P i Represents the pixel coordinates of the four opening and closing points of the i-th indicator diagram image in the indicator diagram data set, (r ij ,c ij ) corresponds to the pixel coordinate value of a single Vertex opening and closing point; 3) The pixel coordinate value set P of the Vertex opening and closing points is normalized as follows and used as the label set L of the deep learning network model: L={L i |i=1,2,...,N} L i ={(x ij ,y ij )|j=1,2,...,4} Where L i represents the network model label of the i-th dynamometer picture in the dynamometer dataset, (x ij ,y ij ) represents the network model label of a single valve switch point, w and h are the width and height pixel resolutions of the indicator diagram respectively; 4) Use the holdout method to divide the calibrated dynamometer data set S into a training set S train , validation set S val and the test set S test ; 5) Build a deep learning network model to predict Gaussian heat maps to meet the requirements of spatial generalization and translation equivalence for the Vertex opening and closing point recognition task; 6) Introducing the feature pyramid network into the network model to solve the theoretical error lower bound problem caused by the resolution in the traditional Gaussian heat map deep learning network; Specifically, after extracting the features of the indicator diagram, the model uses the feature pyramid network FPN to unidirectionally fuse features from top to bottom, where M 2 、M 3 、M 4 and M 5 is the fused feature map, P 2 , P 3 , P 4 and P 5 Corresponding to M 2 、M 3 、M 4 and M 5 Feature map after 3x3 convolution to eliminate aliasing effect; Select P 2 As the input of the final Gaussian heat map prediction module, HeatMap in this module represents the Gaussian heat map spectrum. DSNT is a conversion module that converts HeatMap into coordinate point output; 7) An attention mechanism module is embedded in the feature extraction network of the network model to meet the fine-grained and high-precision requirements of the Vertex opening and closing point recognition task; specifically, the network model as a whole uses ResNet50 embedded with the attention mechanism as the feature extraction network, and the ResNet50 embedded with the attention mechanism includes sequentially cascaded C 1 CBAM, C 2 , C 3 , C 4 and C 5 , CBAM module; among which C 1 is the initial convolution module, C 2 , C 3 , C 4 and C 5 Bottleneck is composed of residual blocks, and CBAM is the attention mechanism module; 8) The built deep learning network model is trained on the training set S train and validation set S val Learning network parameters, based on the test set S test The results evaluate its generalization ability and obtain the final network model; 9) Input the indicator diagram of the real-time working condition record of the oil production machine into the network model to obtain the corresponding valve opening and closing point label (x rj ,y rj ); 10) Calculate the pixel coordinates of the Vernal opening and closing points according to the Vernal opening and closing point labels (r rj ,c rj ): 11) Calculate the effective stroke S based on the pixel coordinates of the valve opening and closing points pe : Where S max It is the maximum stroke of the oil production equipment; 12) Determine the fluid supply capacity index β based on the ratio of the effective stroke to the stroke S during oil extraction machine operation:

2. A method for monitoring the fluid supply capacity of an oil well based on a dynamometer diagram according to claim 1, Features The valve opening and closing points of the indicator diagram data set include: fixed valve opening point A, fixed valve closing point B, floating valve opening point C and floating valve closing point D.

3. The method for monitoring the fluid supply capacity of an oil well based on a dynamometer diagram according to claim 1, Features The pixel coordinate values ​​of the valve opening and closing points are normalized to [-1,1] based on the image resolution (w,h) of the dynamometer diagram.

4. The method for monitoring the fluid supply capacity of an oil well based on a dynamometer diagram according to claim 1, Features The allocation ratio of the set-aside method is 6:2:

2.

5. The method for monitoring the fluid supply capacity of an oil well based on a dynamometer diagram according to claim 1, Features The resolution of the predicted Gaussian heat map is 1 / 4 of the input image size.

6. A method for monitoring the fluid supply capacity of an oil well based on a dynamometer diagram according to claim 1, Features The embedding attention mechanism module is divided into spatial attention mechanism and channel attention mechanism.

Citation Information

Patent Citations

  • Method for calculating effective stroke of oil pumping unit

    CN113486599A

  • Extraction method of valve opening and closing points of indicator diagram of pumping well based on physical significance

    CN101660402A

  • Indicator diagram recognition method based on regularized attention convolutional neural network

    CN110163302A

  • Oil metering method and system of oil pumping unit based on data driving

    CN110363337A

  • General target detection method of adaptive attention guidance mechanism

    CN111259930A

Cited By

  • Oil extraction machine variable speed control method giving consideration to underground liquid supply capacity and time-of-use electricity price

    CN117494970A

  • A variable speed control method of oil pumping machine considering downhole liquid supply capacity and time-of-use electricity price

    CN117494970B