CN114718861A - Intelligent diagnosis method for working condition of screw pump well based on deep learning - Google Patents
Intelligent diagnosis method for working condition of screw pump well based on deep learning Download PDFInfo
- Publication number
- CN114718861A CN114718861A CN202110010119.7A CN202110010119A CN114718861A CN 114718861 A CN114718861 A CN 114718861A CN 202110010119 A CN202110010119 A CN 202110010119A CN 114718861 A CN114718861 A CN 114718861A
- Authority
- CN
- China
- Prior art keywords
- screw pump
- working condition
- deep learning
- pump well
- training
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 85
- 238000003745 diagnosis Methods 0.000 title claims abstract description 61
- 238000013135 deep learning Methods 0.000 title claims abstract description 47
- 238000012549 training Methods 0.000 claims abstract description 104
- 230000008569 process Effects 0.000 claims abstract description 40
- 238000011161 development Methods 0.000 claims abstract description 10
- 230000008859 change Effects 0.000 claims abstract description 7
- 238000013527 convolutional neural network Methods 0.000 claims description 77
- 230000006870 function Effects 0.000 claims description 45
- 210000002569 neuron Anatomy 0.000 claims description 12
- 230000004913 activation Effects 0.000 claims description 10
- 238000011176 pooling Methods 0.000 claims description 8
- 238000012545 processing Methods 0.000 claims description 7
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 6
- 238000004458 analytical method Methods 0.000 claims description 4
- 238000007405 data analysis Methods 0.000 claims description 4
- 238000013507 mapping Methods 0.000 claims description 3
- 238000012216 screening Methods 0.000 claims description 3
- 238000004519 manufacturing process Methods 0.000 abstract description 20
- 239000003129 oil well Substances 0.000 abstract description 7
- 239000003921 oil Substances 0.000 description 23
- 238000013461 design Methods 0.000 description 15
- 238000010586 diagram Methods 0.000 description 11
- 238000000605 extraction Methods 0.000 description 8
- 238000005457 optimization Methods 0.000 description 6
- 230000002159 abnormal effect Effects 0.000 description 5
- 230000000694 effects Effects 0.000 description 5
- 238000007781 pre-processing Methods 0.000 description 5
- 239000004576 sand Substances 0.000 description 4
- 238000012360 testing method Methods 0.000 description 4
- 238000013528 artificial neural network Methods 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 3
- 230000007423 decrease Effects 0.000 description 3
- 238000013136 deep learning model Methods 0.000 description 3
- 238000002405 diagnostic procedure Methods 0.000 description 3
- 238000001556 precipitation Methods 0.000 description 3
- 230000008961 swelling Effects 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 2
- 238000005086 pumping Methods 0.000 description 2
- PXFBZOLANLWPMH-UHFFFAOYSA-N 16-Epiaffinine Natural products C1C(C2=CC=CC=C2N2)=C2C(=O)CC2C(=CC)CN(C)C1C2CO PXFBZOLANLWPMH-UHFFFAOYSA-N 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000010835 comparative analysis Methods 0.000 description 1
- 239000010779 crude oil Substances 0.000 description 1
- 238000013480 data collection Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000012938 design process Methods 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 239000007788 liquid Substances 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
Images
Classifications
-
- F—MECHANICAL ENGINEERING; LIGHTING; HEATING; WEAPONS; BLASTING
- F04—POSITIVE - DISPLACEMENT MACHINES FOR LIQUIDS; PUMPS FOR LIQUIDS OR ELASTIC FLUIDS
- F04C—ROTARY-PISTON, OR OSCILLATING-PISTON, POSITIVE-DISPLACEMENT MACHINES FOR LIQUIDS; ROTARY-PISTON, OR OSCILLATING-PISTON, POSITIVE-DISPLACEMENT PUMPS
- F04C14/00—Control of, monitoring of, or safety arrangements for, machines, pumps or pumping installations
- F04C14/28—Safety arrangements; Monitoring
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/047—Probabilistic or stochastic networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Probability & Statistics with Applications (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Mechanical Engineering (AREA)
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
Abstract
The invention provides a screw pump well working condition intelligent diagnosis method based on deep learning, which comprises the following steps: step 1, obtaining operation parameters of a screw pump from an oil field development database; step 2, according to the step 1, determining that the current is a characteristic parameter of the working condition of the screw pump well, and summarizing common working conditions of the screw pump and current change characteristics under each working condition; step 3, establishing a sample library for working condition diagnosis of the screw pump well; step 4, building a deep learning network model, and training the network model; and 5, inputting a new current card sample into the trained deep learning network model, and intelligently outputting the working condition number corresponding to the current card sample, thereby realizing intelligent diagnosis of the working condition. The intelligent diagnosis method for the working condition of the screw pump well based on the deep learning realizes intelligent, accurate and quick recognition of the working condition of the screw pump well, can help workers to quickly learn the operating state of the screw pump and timely process faults, improves the production efficiency, and ensures the production operation of an oil well.
Description
Technical Field
The invention relates to the technical field of oilfield development, in particular to a screw pump well working condition intelligent diagnosis method based on deep learning.
Background
The screw pump has the advantages of simple structure, light weight, convenient maintenance and management, compact ground equipment, small occupied area, easy adjustment of discharge capacity and high working efficiency, and has better effect than other lifting modes when in exploitation of crude oil with high viscosity, high gas content and high sand content. The screw pump is gradually popularized and used in oil fields, and the share of the screw pump in each oil field is increased year by year. However, compared with the pumping unit which can quickly and accurately diagnose different working conditions according to the shape of the indicator diagram, the working condition diagnosis method of the screw pump is not perfect. The current diagnostic methods mainly have the following problems: (1) the diagnostic parameters are difficult to obtain, such as a torque test diagnostic method, related instruments are required to be installed after the well is stopped, the operation is complex, and the production of an oil well is influenced; (2) the diagnostic method has damage to the oil well, such as a pressure holding method, and is easy to interfere with an oil well pressure system to further damage an oil layer; (3) the method depends too much on manual experience, and the diagnosis error is large, and specific fault reasons cannot be directly judged if a liquid level position change diagnosis method is adopted.
In recent years, an artificial intelligence algorithm is rapidly developed, an artificial neural network is used for diagnosing the oil pumping working condition of the screw pump, the manual intervention is eliminated, more useful information is lost during fault feature extraction, the diagnostic result has errors, the algorithm is low in training speed, and the precision and the efficiency cannot meet the requirements.
The existing screw pump working condition diagnosis method is not rapid and accurate enough, so that problems are found and solved in the oil extraction process, the yield and the operation of an oil well are influenced, and the popularization and the use of the screw pump are seriously restricted; the method for efficiently and accurately diagnosing the working condition of the screw pump is established, and has important significance for guaranteeing the production operation of an oil well and further popularizing the oil production mode of the screw pump.
In the application No.: chinese patent application CN201810781901.7 relates to a design method of screw pump production system based on effective power, which comprises the following steps: collecting related parameter data related to a screw pump production system which normally works in the working process; the collected data are sorted and cleaned, and a parameter database is established; training a Deep Belief Network (DBN) by using the classified working conditions, enabling a neural network to learn a region range of a reasonable working condition, and finding out an optimal solution based on the maximum effective power; when a screw pump production system needs to be designed in the development of a new well, the trained DBN is used for guiding the pump type of the screw pump to be optimized and the design of corresponding matching equipment such as a motor, a sucker rod and the like. The method mainly aims to establish the corresponding relation between the screw pump type and the oil extraction parameters by utilizing a deep belief network so as to guide the matching problem of a rod, a pump and a machine in production, and does not provide a method for intelligently diagnosing the working condition of the screw pump.
In the application No.: CN201811472378.6 relates to a CNN-LSTM deep learning method and a fault diagnosis method based on multi-attribute time sequence data. Which comprises the following steps: s1, collecting historical operation data of the system and carrying out data preprocessing, and then establishing a fault diagnosis model based on CNN and LSTM; and S2, acquiring real-time operation data of the system, performing data preprocessing, sending the data into the fault diagnosis model established in S1 for processing, and outputting a diagnosis result. The innovation point of the patent lies in that attribute information and time delay information during system operation are preferably integrated, so that the accuracy and the noise resistance of fault diagnosis are improved. From the application field, the patent mainly aims at a large industrial system in a broad sense, and is not in the field of screw pump diagnosis of oil extraction engineering; from the specific implementation point of view, the patent does not specify any index as a characteristic parameter for diagnosing the working condition of the screw pump, and has no practical guiding significance for the patent.
In the application No.: chinese patent application CN201810949099.8 relates to a method for diagnosing mechanical failure of unsupervised deep learning network, which comprises the following steps: (1) installing corresponding sensors near parts such as a bearing of mechanical equipment to acquire vibration signals to obtain mechanical vibration signals; (2) converting the acquired vibration signals into a mixed domain fault characteristic data set, and dividing the mixed domain fault characteristic data set into a test and training sample characteristic subset; (3) inputting a training sample feature subset into a constructed Unsupervised Deep Learning Network (UDLN) model for learning and training, wherein the UDLN model consists of two layers of improved sparse filtering (L12SF) unsupervised feature extraction layers and one layer of weighted Euclidean distance similar affine (WE-AP) clustering layer; (4) and inputting the test sample into the trained diagnostic model to realize the whole-process unsupervised feature learning and fault clustering. (5) And calculating the recognition rate according to the membership degree of the test sample clustering division, thereby realizing fault recognition and diagnosis. The patent aims at machine faults, the established characteristic data set is mainly vibration signals, and the corresponding relation with the machine faults is established through the characteristics of the vibration signals, so that the method is different from intelligent diagnosis of the working conditions of the screw pump.
Therefore, a novel intelligent diagnosis method for the working condition of the screw pump well based on deep learning is invented, and the technical problems are solved.
Disclosure of Invention
The invention provides a screw pump well working condition intelligent diagnosis method based on deep learning, which can intelligently and accurately identify the working condition of a screw pump in real time.
The object of the invention can be achieved by the following technical measures: the intelligent diagnosis method for the working condition of the screw pump well based on deep learning comprises the following steps:
step 1, obtaining operation parameters of a screw pump from an oil field development database;
step 3, establishing a sample library for working condition diagnosis of the screw pump well;
step 4, building a deep learning network model, and training the network model;
and 5, inputting a new current card sample into the trained deep learning network model, and intelligently outputting the working condition number corresponding to the current card sample, thereby realizing intelligent diagnosis of the working condition.
The object of the invention can also be achieved by the following technical measures:
in step 1, the operating parameters of the screw pump are obtained from the oil field development database, and parameters which can be easily acquired and the variation characteristics of which can reflect the working conditions of different screw pumps are searched for as characteristic parameters for expressing the working conditions of the screw pump well through mass data analysis and screening.
In step 1, the obtained operating parameters of the screw pump include current, voltage, power, wellhead oil pressure/temperature, rotating speed, instantaneous/accumulated flow and daily power consumption.
In step 2, through data comparison and analysis, the current is very sensitive to different working conditions and has strong regular characteristics, so that the current is determined as a characteristic parameter for expressing the working conditions of the screw pump well.
In step 3, preprocessing the current data in the database, drawing current card samples, manually classifying the samples, and forming a sample library for diagnosing the working conditions of the screw pump well, wherein each class corresponds to one working condition of the screw pump.
In the step 4, selecting a convolutional neural network CNN as a deep learning network model, building the convolutional neural network, taking a sample library as a training set, and extracting the mapping relation between the current characteristic of each sample and the working condition of the corresponding screw pump through a deep learning algorithm; by adjusting various parameters of the convolutional neural network, the training loss is continuously reduced, and when the training loss is controlled at a lower level, the training is completed.
In step 4, the number of input layers, convolutional layers, pooling layers, full-link layers and Softmax classifiers of the used CNN architecture is determined according to the size of the current card and the principle of less model training loss and less training time.
In step 4, in the process of building the CNN model, a cross entropy loss function with a more stable and smooth descending process and less fluctuation is selected as a loss function of the CNN model.
In step 4, in the process of building the CNN model, at least 3 different convolution kernel combination schemes are designed according to the determined number of the convolution layers, and the combination mode of the convolution kernels is determined according to the principle of minimum training loss.
In step 4, in the process of building the CNN model, at least 4 schemes are designed for training respectively in order to optimize the number of convolution kernels of each layer, and the number of convolution kernels of each layer is determined according to the principle of low training loss and convergence rate.
In step 4, in the process of building the CNN model, a Sigmoid function, a hyperbolic tangent function and a ReLU function are respectively adopted to train the CNN model, and an activation function is determined according to the principle of low training loss and convergence rate.
In step 4, in the process of building the CNN model, at least 4 schemes are designed for optimizing the number of the neurons of the full connection layer, and the number of the neurons of the full connection layer is determined on the basis of small training loss and short training time.
In step 4, in the process of building the CNN model, at least 4 different iteration times are set for optimizing the iteration times, and the iteration times are determined on the basis of the minimum training loss.
In step 4, in the process of building the CNN model, an Adam optimizer, an SGD optimizer and an Adadelta optimizer are respectively used for training, and the optimizer is determined on the basis of small training loss.
In step 4, in the process of building the CNN model, at least 4 schemes are designed for training in order to optimize the batch sample number, and the batch processing number is determined on the basis of small training loss.
In step 4, in the process of building the CNN model, Dropout is set to at least 3 different values respectively, and the value of Dropout is determined on the basis of small training loss.
The intelligent diagnosis method for the working condition of the screw pump well based on the deep learning overcomes the defects that the conventional screw pump well diagnosis excessively depends on artificial experience, the diagnosis process is slow, the efficiency is low and the accuracy is poor, determines the current as a characteristic parameter, realizes the intelligent, accurate and quick recognition of the working condition of the screw pump well by building the deep learning network, can help workers to quickly learn the operating state of the screw pump and timely process faults, improves the production efficiency and ensures the production and operation of an oil well.
Drawings
FIG. 1 is a flow chart of an embodiment of the intelligent diagnosis method for the working condition of the screw pump well based on deep learning of the present invention;
FIG. 2 is a schematic diagram of a current signature card corresponding to different operating conditions in an embodiment of the present disclosure;
FIG. 3 is a schematic diagram of current card samples at different fault times under the same operating condition in an embodiment of the present invention;
FIG. 4 is a diagram of a convolutional neural network architecture used in an embodiment of the present invention;
FIG. 5 is a graph illustrating loss values of different loss functions in an embodiment of the present invention;
FIG. 6 is a diagram illustrating training loss for different combinations of numbers of convolution kernels in accordance with an embodiment of the present invention;
FIG. 7 is a diagram illustrating training loss for different activation functions in an embodiment of the present invention;
FIG. 8 is a diagram illustrating training loss at different iterations in an embodiment of the present invention;
FIG. 9 is a diagram illustrating training loss for different optimizers in an embodiment of the present invention;
FIG. 10 is a graph illustrating training loss for different sample numbers in different batches according to an embodiment of the present invention;
FIG. 11 is a diagram illustrating training loss for different Dropout values according to an embodiment of the present invention;
FIG. 12 is a diagram of an operation mode of an intelligent diagnostic system for working conditions of a screw pump well according to an embodiment of the present invention;
fig. 13 is a Web client interface display diagram of the intelligent diagnosis and warning system for screw pump well working conditions in an embodiment of the present invention.
Detailed Description
It is to be understood that the following detailed description is exemplary and is intended to provide further explanation of the invention as claimed. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.
It is noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of exemplary embodiments according to the invention. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, and it should be understood that when the terms "comprises" and/or "comprising" are used in this specification, they specify the presence of the stated features, steps, operations, and/or combinations thereof, unless the context clearly indicates otherwise.
The invention discloses a screw pump well working condition intelligent diagnosis method based on deep learning, which comprises the following steps:
step 1, acquiring operating parameters of a screw pump from an oil field development database, and searching parameters which can be easily acquired and have variation characteristics capable of reflecting working conditions of different screw pumps through mass data analysis and screening to be used as characteristic parameters for expressing the working conditions of a screw pump well;
the obtained operating parameters of the screw pump mainly comprise current, voltage, power, wellhead oil pressure/temperature, rotating speed, instantaneous/accumulated flow and daily power consumption;
the screw pump well working condition covers nine common working conditions of the screw pump well oil production mode, and the working conditions are respectively as follows: rod breakage, pump leakage, pump jamming, wax precipitation/stator swelling, eccentric wear/sand production, rubber damage, rotor breakage, abnormal working conditions and normal working conditions. Through data comparison and analysis, the current is very sensitive to different working conditions, and almost all screw pumps can easily acquire the data, so that the current is determined as a characteristic parameter for expressing the working conditions of the screw pump wells.
Step 3, preprocessing the current data in the database, drawing current card samples, manually classifying the samples, and forming a sample library for diagnosing the working conditions of the screw pump well, wherein each type corresponds to one working condition of the screw pump;
each current characteristic has a large number of samples, and the samples not only cover the fault degree of the corresponding working condition, but also cover the occurrence time of the corresponding working condition.
Step 4, building a deep learning network model, taking a sample library as a training set, and automatically extracting the mapping relation between the current characteristic of each sample and the working condition of the corresponding screw pump through a deep learning algorithm; continuously reducing training loss by adjusting various parameters of the convolutional neural network, and finishing training when the training loss is controlled at a lower level;
the current card sample library is a two-dimensional image, belongs to the field of image recognition, and selects a Convolutional Neural Network (CNN) which is widely applied in the field of image recognition and mature in technology as a deep learning model; further, according to the size of the current card designed by the invention, the number of input layers, convolution layers, pooling layers, full-link layers and Softmax classifiers of the CNN framework is determined according to the principle that the model training loss is small and the training time is short.
In an embodiment, according to the size of a current card, the accuracy and the speed of model operation are combined, namely the number of input layers, convolutional layers, pooling layers, full-link layers and Softmax classifiers are set as several different parameters, the model is operated, training loss and convergence speed are counted, and the number of the input layers, the convolutional layers, the pooling layers, the full-link layers and the Softmax classifiers of the used CNN architecture is determined by taking the model training loss and the convergence speed as parameter preference standards.
Specifically, in the process of building the CNN model, a cross entropy loss function with a more stable and smooth descending process and less fluctuation is selected as a loss function of the CNN model.
Specifically, in the process of building a CNN model, at least 3 different convolution kernel combination schemes are designed according to the determined number of the convolution layer, and the combination mode of convolution kernels is determined according to the principle of minimum training loss.
Specifically, in the process of building the CNN model, at least 4 schemes are designed for training respectively in order to optimize the number of convolution kernels of each layer, and the number of the convolution kernels of each layer is determined according to the principle that the training loss is small and the convergence rate is considered.
Specifically, in the process of building the CNN model, a Sigmoid function, a hyperbolic tangent function and a ReLU function are respectively adopted to train the CNN model, and an activation function is determined according to the principle that the training loss is small and the convergence rate is considered.
Specifically, in the process of building the CNN model, at least 4 schemes are designed for optimizing the number of neurons in the full connection layer, and the number of the neurons in the full connection layer is determined on the basis of small training loss and short training time.
Specifically, in the process of building the CNN model, at least 4 different iteration times are set for optimizing the iteration times, and the iteration times are determined on the basis of the minimum training loss.
Specifically, in the process of building the CNN model, an Adam optimizer, an SGD optimizer and an Adadelta optimizer are used for training respectively, and the optimizers are determined on the basis of small training loss.
Specifically, in the process of building the CNN model, at least 4 schemes are designed for training in order to optimize the batch sample number, and the batch processing number is determined on the basis of small training loss.
Specifically, in the process of building the CNN model, Dropout is set to at least 3 different values respectively, and the value of Dropout is determined on the basis of small training loss.
and 6, developing a Web client of the intelligent screw pump well working condition diagnosis and early warning system according to the deep learning algorithm framework, connecting the Web client with an oil field development database, realizing real-time data acquisition and analysis of the screw pump, quickly, accurately and intelligently diagnosing the real-time working condition of the screw pump well, and pushing the real-time working condition to relevant workers for fault treatment.
The Web client architecture of the intelligent diagnosis and early warning system for the working condition of the screw pump well mainly comprises six parts, namely a screw pump production large database, a local screw pump well working condition diagnosis workstation, a screw pump well working condition diagnosis CNN model, a local database, a Web client and oil field management and control personnel; the method mainly comprises three functions of working condition statistics, working condition query, alarm early warning and the like.
In an embodiment 1 to which the present invention is applied, as shown in fig. 1, a method for intelligently diagnosing the working condition of a screw pump well based on deep learning mainly includes the following components: acquiring operating parameters of the screw pump from an oil field development database; classifying and numbering common working conditions of the screw pump well; finding out the operating parameters which are most sensitive to the working conditions and have strong regular characteristics as the characteristic parameters for indicating the working conditions of the screw pump; on the basis, reading the massive examples of the characteristic parameters in the database, preprocessing the massive examples, and drawing a card sample in an image form; classifying the card samples into folders according to characteristics and corresponding to working conditions to form a training set; building a Convolutional Neural Network (CNN) and initializing various parameters, and training the CNN by taking a training set as input and a working condition number as a target value; adjusting various parameters of the CNN network according to the training result until the training loss value is controlled at a lower level, and finishing training to obtain a better CNN network model; and inputting the new characteristic parameters or the characteristic parameters read in real time from the database into the model in the form of card samples, and outputting the working condition number corresponding to the specific characteristic parameters, thereby realizing the intelligent diagnosis of the working condition.
The operating parameters of the screw pump mainly comprise current, voltage, power, wellhead oil pressure/temperature, rotating speed, instantaneous/accumulated flow and daily power consumption; the working conditions of the screw pump well comprise nine common working conditions which are respectively as follows: rod breakage, pump leakage, pump jamming, wax precipitation/stator swelling, eccentric wear/sand production, rubber damage, rotor breakage, abnormal working conditions and normal working conditions. The abnormal working condition refers to an abnormal condition that current data acquisition is discontinuous and incomplete and diagnosis cannot be given. Data analysis shows that parameters such as wellhead oil pressure/temperature and the like do not change greatly under different working conditions, namely the parameters are not sensitive enough to the working conditions, so that the parameters cannot be selected as characteristic parameters for expressing the working conditions of the screw pump well; the parameters such as the rotating speed, the instantaneous/accumulated flow and the like can not be collected on all screw pumps, namely the data collection integrity of the parameters is not high, and the parameters can not be used as characteristic parameters for expressing the working conditions of the screw pump wells. Through comparative analysis, the current is very sensitive to different working conditions, and almost all screw pumps can easily acquire the data, so that the current is determined as a characteristic parameter for expressing the working conditions of the screw pump wells (as shown in figure 2).
Wherein, in the card sample drawing process, to same kind of operating mode, because the fault degree is different, the time that the trouble took place is different, and the current curve is also not completely the same, consequently, for the accuracy of guaranteeing the discernment, every kind of electric current characteristic all can have magnanimity sample to guarantee not only to cover the fault degree of corresponding operating mode, cover the emergence time of corresponding operating mode moreover (as shown in fig. 3). In the specific implementation process, the number of samples is as follows: 5000 sets of pump leakage, 2000 sets of wax precipitation/stator swelling, 2300 sets of eccentric wear/sand production, 2700 sets of rubber skin, 3500 sets of rod fracture, 4500 sets of pump clamps, 2500 sets of rotor fracture, 6000 sets of normal production and 4000 sets of abnormal production.
The deep learning network model is built, and a Convolutional Neural Network (CNN) is selected because a current card sample library related in the invention is a two-dimensional image, belongs to the field of image recognition, and the CNN has the characteristic of invariance to the position, shape and scale of image features, and is widely applied in the field of image recognition and mature in technology.
In the design process of the convolutional neural network architecture, two structures of a classical LeNet-5 model and an AlexNet model are applied in sequence. The input of the LeNet-5 model is a 32 multiplied by 32 gray level image which mainly comprises 2 convolution layers, 2 pooling layers and 3 full-connection layers; the input of the AlexNet model is a gray scale image of 224 × 224, and mainly comprises 5 convolution layers and 3 full-connection layers. The LeNet-5 model has fast convergence but poor precision, while the AlexNet model has high precision but long operation time; in combination with the size of the current card designed by the present invention, the accuracy and the convergence are both considered, and a CNN architecture used in this example is designed, which mainly includes 1 input layer, 3 convolutional layers, 3 pooling layers, 1 full-link layer, and 1 Softmax classifier (as shown in fig. 4).
The CNN network design requires a determination of a loss function. For this purpose, the average absolute error function, the mean square error loss function and the cross entropy loss function are applied in a traversal mode, and the relationship between the loss value and the iteration number of three different loss functions is counted (as shown in fig. 5). The result shows that when the average absolute error function is used, the model has poor convergence performance, and the loss decline curve fluctuates up and down all the time; the mean square error loss function and the cross entropy loss function both show good convergence effect, wherein the descending process of the cross entropy loss function is more stable and smooth, and the fluctuation is less. The cross entropy loss function is determined as the loss function of the CNN model.
The CNN network design requires optimization of the convolution kernel size. The large convolution kernel can provide a larger receptive field, can capture more information, and obtains more features, but increases parameters and calculation amount involved in training, is not beneficial to increase of model depth, and reduces calculation performance. The small convolution kernel can capture more detailed features, the large convolution kernel is replaced by a stack of a plurality of small convolution kernels, the calculation efficiency can be improved, and the feature information cannot be effectively extracted from sparse data due to the fact that the convolution kernel is too small. The present invention designs different convolution kernel combination schemes for 3 convolution layers (as shown in table 1):
TABLE 1 three-layer convolution kernel combination design scheme
The pooling windows of different schemes are all designed by adopting a common 2 x 2, the convolutional neural network is trained, and the combination mode of three layers of convolutional kernels is determined to be 3 x 3&3 x 3 (as shown in table 2) according to the principle of minimum training loss.
TABLE 2 training loss under different convolution kernel size combinations
The CNN network design requires optimization of the number of convolution kernels. Too many convolution kernels increase the model training time; too few in number may result in insufficient feature extraction, resulting in difficulty in model convergence. The number of convolution kernels is generally in an exponential form of 2, and the number of convolution kernels is increased along with the deepening of the number of layers; 4 schemes (as shown in table 3) were designed and trained separately, and the loss value of each scheme was counted (as shown in fig. 6).
TABLE 3 convolution kernel number design
According to the training results, although the convergence is rapid, the loss values of the scheme 1 and the scheme 2 are as high as 0.25; scheme 4 has large fluctuation of loss values; therefore, option 3 is chosen, i.e., the number of convolution kernels for each layer is 32, 64, 128, respectively.
The CNN network design requires a preferred activation function. The convolution operation is followed by an activation function before the values can be input into the next layer of the network. Commonly used activation functions include Sigmoid functions, hyperbolic tangent functions, and ReLU functions. The Sigmoid function is:
the hyperbolic tangent function is:
the ReLU function is:
and (3) respectively adopting the three activation functions to train the convolutional neural network model, and counting the loss value of each scheme (as shown in fig. 7). The training result shows that when the ReLU activation function is used, the loss of the model is reduced most quickly, and the convergence effect is the best, so the ReLU is selected as the activation function.
The CNN network design requires optimization of the number of full-link layer neurons. Too small a number of neurons in the fully connected layer leads to undesirable convergence of the model, and too large a number increases the training burden of the convolutional neural network model. 4 protocols were designed and statistical training losses and training times were calculated (as shown in Table 4).
TABLE 4 training loss and training time Table for different neuron numbers in the Total connexion layer
According to the training result, when the number of the neurons exceeds 512, the amplitude of the continuous decline of the training loss is small, but the training time is obviously increased; the number of full connectivity layer neurons was thus determined to be 512.
The CNN network design requires an optimized number of iterations. The iteration number is also the number of training rounds, insufficient number of rounds can result in model under-fitting, and excessive number of rounds can result in model over-fitting. And setting different iteration times, training the model, and counting the training loss (as shown in figure 8). The training result shows that the loss value does not obviously decrease after 100 times of training, and even an overfitting phenomenon occurs when the number of training is close to 150 times, so that the number of iterations is determined to be 100 times.
The CNN network design requires a preference optimizer. The optimizer is the training method of CNN. Commonly used neural network optimizers include Adam optimizers, SGD optimizers, adapelta optimizers, and the like. Training was performed using the three optimizers described above, respectively, and training loss was counted (as shown in fig. 9). The result shows that the model convergence is poor when the optimizer SGD is used, the fluctuation in the early stage of training is large when the Adadelta optimizer is used, the model loss using the Adam optimizer is reduced smoothly, and the minimum loss value of the model can reach 0.0016, so that the Adam optimizer is determined to be used.
The CNN network design requires optimization of batch sample numbers. The batch sample size should not be too small, for example, batch size is 1, which means that each sample is corrected in the gradient direction, which results in large differences between samples, unrepresentative statistics, and increased noise, thus making the model difficult to converge. If the number of the selected samples in a large batch is too large, the gradient direction is relatively stable, so that the model is easy to fall into the local optimal condition, and the precision of the model is reduced. Training was performed with the number of samples in the batch set to 1, 50 and 100, respectively, and the loss values were counted (as shown in fig. 10). The results show that when the batch sample amount is 1, the convergence effect of the model is poor; when the batch data volume is 50, the loss change is smooth and the loss value is small; when the batch processing amount is 100 and the iteration times are close to 100, the loss value is increased; the batch size is set to 50.
The CNN network design requires optimization Dropout. Dropout is that in the actual training process of the network, the connection number of neurons between layers of the convolutional neural network is reduced according to a certain probability, so that when the weight of the network is updated in the back propagation process, the weight connected with the node is not updated any more, and the overfitting phenomenon caused by insufficient training samples is effectively reduced. The values of Dropout are set to 0.1, 0.35, 0.5, respectively, and training is performed to count the loss values (as shown in fig. 11). The result shows that when Dropout is 0.1, the model convergence is best, and the loss value is also lowest, so that the value of Dropout is determined to be 0.1.
In an embodiment 2 to which the present invention is applied, the optimization results of parameters of the CNN network are shown in table 5.
TABLE 5 deep learning model parameter Table
In a specific embodiment 3 to which the present invention is applied, in order to facilitate application of the design result of the present invention, a Web client of the intelligent diagnosis and early warning system for working conditions of the screw pump well is programmed in this example, so as to implement real-time connection of a production database of the screw pump well and intelligent diagnosis and push of the working conditions. The intelligent diagnosis system is operated in the mode shown in fig. 12, and the interface is shown in fig. 13; the system comprises functions of working condition statistics, working condition query, alarm early warning and the like, oil field management and control personnel can perform corresponding processing immediately by monitoring the diagnosis result displayed on the Web client side if finding that the fault of the screw pump well occurs, and the processing content and the result are stored in a local database.
In summary, the invention defines the current as the characteristic parameter reflecting the working condition of the screw pump, establishes the intelligent diagnosis method of the working condition of the screw pump well based on deep learning, can quickly and accurately diagnose the working condition of the screw pump well, helps oil field management and control personnel to know and process faults at the first time, and has great significance for the popularization of the screw pump oil extraction technology and the production and operation guarantee of the oil field extraction well.
Finally, it should be noted that: although the present invention has been described in detail with reference to the foregoing embodiments, it will be apparent to those skilled in the art that changes may be made in the embodiments and/or equivalents thereof without departing from the spirit and scope of the invention. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims (16)
1. The intelligent diagnosis method for the working condition of the screw pump well based on deep learning is characterized by comprising the following steps of:
step 1, obtaining operation parameters of a screw pump from an oil field development database;
step 2, according to the step 1, determining that the current is a characteristic parameter of the working condition of the screw pump well, and summarizing common working conditions of the screw pump and current change characteristics under each working condition;
step 3, establishing a sample library for working condition diagnosis of the screw pump well;
step 4, building a deep learning network model, and training the network model;
and 5, inputting a new current card sample into the trained deep learning network model, and intelligently outputting the working condition number corresponding to the current card sample, thereby realizing intelligent diagnosis of the working condition.
2. The intelligent diagnosis method for the working conditions of the screw pump well based on the deep learning of claim 1 is characterized in that in the step 1, the operation parameters of the screw pump are obtained from an oil field development database, and the parameters which can be easily collected and the variation characteristics of which can reflect the working conditions of different screw pumps are searched through mass data analysis and screening and are used as the characteristic parameters for expressing the working conditions of the screw pump well.
3. The intelligent deep learning-based screw pump well working condition diagnosis method according to claim 2, wherein in the step 1, the obtained screw pump operation parameters comprise current, voltage, power, wellhead oil pressure/temperature, rotating speed, instantaneous/accumulated flow and daily power consumption.
4. The intelligent deep learning-based screw pump well working condition diagnosis method according to claim 1, wherein in the step 2, through data comparison and analysis, the current is very sensitive to different working conditions and has strong regular characteristics, so that the current is determined as a characteristic parameter for expressing the screw pump well working condition.
5. The intelligent diagnosis method for the working condition of the screw pump well based on the deep learning of claim 1, wherein in the step 3, the current data in the database are preprocessed, current card samples are drawn, the samples are manually classified, and each class corresponds to one screw pump working condition to form a sample library for the working condition diagnosis of the screw pump well.
6. The intelligent deep learning-based screw pump well working condition diagnosis method according to claim 1, characterized in that in step 4, a convolutional neural network CNN is selected as a deep learning network model, the convolutional neural network is built, a sample library is used as a training set, and a mapping relation between the current characteristic of each sample and the corresponding screw pump working condition is extracted through a deep learning algorithm; by adjusting various parameters of the convolutional neural network, the training loss is continuously reduced, and when the training loss is controlled at a lower level, the training is completed.
7. The intelligent deep learning-based screw pump well working condition diagnosis method according to claim 6, wherein in step 4, the number of input layers, convolution layers, pooling layers, full-link layers and Softmax classifiers of the CNN architecture to be used is determined according to the current card size and the principle of low model training loss and short training time.
8. The intelligent deep learning-based screw pump well working condition diagnosis method according to claim 6, wherein in the step 4, in the CNN model building process, a cross entropy loss function which is more stable and smooth in a descending process and less in fluctuation is selected as a loss function of the CNN model.
9. The intelligent deep learning-based screw pump well working condition diagnosis method according to claim 6, characterized in that in step 4, in the process of building the CNN model, at least 3 different convolution kernel combination schemes are designed for the determined number of convolution layers, and the combination mode of the convolution kernels is determined according to the principle of minimum training loss.
10. The intelligent deep learning-based screw pump well working condition diagnosis method according to claim 6, characterized in that in step 4, in the process of building the CNN model, at least 4 schemes are designed for training respectively to optimize the number of convolution kernels in each layer, and the number of convolution kernels in each layer is determined according to the principle that the training loss is small and the convergence rate is considered.
11. The intelligent deep learning-based screw pump well working condition diagnosis method according to claim 6, wherein in step 4, in the process of building the CNN model, a Sigmoid function, a hyperbolic tangent function and a ReLU function are respectively adopted to train the CNN model, and the activation function is determined according to the principle of low training loss and convergence rate.
12. The intelligent deep learning-based screw pump well working condition diagnosis method according to claim 6, wherein in step 4, in order to optimize the number of neurons in the full connection layer during the CNN model building process, at least 4 schemes are designed, and the number of neurons in the full connection layer is determined on the basis of small training loss and short training time.
13. The intelligent deep learning-based screw pump well working condition diagnosis method according to claim 6, wherein in step 4, in the CNN model building process, at least 4 different iteration times are set for optimizing the iteration times, and the iteration times are determined on the basis of the minimum training loss.
14. The intelligent deep learning-based screw pump well working condition diagnosis method according to claim 6, wherein in step 4, in the CNN model building process, an Adam optimizer, an SGD optimizer and an Adadelta optimizer are used for training respectively, and the optimizers are determined on the basis of low training loss.
15. The intelligent deep learning-based screw pump well working condition diagnosis method according to claim 6, wherein in step 4, in the process of building the CNN model, at least 4 schemes are designed for training in order to optimize batch sample quantity, and batch processing quantity is determined on the basis of low training loss.
16. The intelligent deep learning-based screw pump well working condition diagnosis method according to claim 6, wherein in step 4, in the process of building the CNN model, Dropout is set to at least 3 different values respectively, and the Dropout value is determined on the basis of low training loss.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110010119.7A CN114718861A (en) | 2021-01-04 | 2021-01-04 | Intelligent diagnosis method for working condition of screw pump well based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110010119.7A CN114718861A (en) | 2021-01-04 | 2021-01-04 | Intelligent diagnosis method for working condition of screw pump well based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114718861A true CN114718861A (en) | 2022-07-08 |
Family
ID=82234437
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110010119.7A Pending CN114718861A (en) | 2021-01-04 | 2021-01-04 | Intelligent diagnosis method for working condition of screw pump well based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114718861A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116256589A (en) * | 2023-05-15 | 2023-06-13 | 南京研控科技有限公司 | Intelligent diagnosis method and device for electric pump well, storage medium and server |
| CN116906025A (en) * | 2023-07-24 | 2023-10-20 | 西南石油大学 | Oil extraction engineering production abnormality diagnosis device |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105781527A (en) * | 2016-03-24 | 2016-07-20 | 中国海洋石油总公司 | Parameter diagnosis and analysis method for electric submersible pump well working condition instrument |
| WO2017161126A1 (en) * | 2016-03-16 | 2017-09-21 | University Of Houston System | System and method for detecting, diagnosing, and correcting trips or failures of electrical submersible pumps |
| CN107687332A (en) * | 2017-09-08 | 2018-02-13 | 张雅芝 | Screw pump and working conditions of oil extractor on-line analysis diagnosis and control method of adjustment and device are realized using electrical parameter |
| CN108266178A (en) * | 2018-01-14 | 2018-07-10 | 常州大学 | A kind of sucker rod pump monitoring of working condition method based on rich information indicator card and DBN |
| CN109360120A (en) * | 2018-10-22 | 2019-02-19 | 常州大学 | A kind of electric immersible pump well operating condition intelligent diagnosing method based on convolutional neural networks |
| CN110284875A (en) * | 2019-06-13 | 2019-09-27 | 中国石油化工股份有限公司 | A kind of Diagnosing The Faults of Esp ' method and apparatus based on current parameters |
| US20200209109A1 (en) * | 2018-12-28 | 2020-07-02 | Shanghai United Imaging Intelligence Co., Ltd. | Systems and methods for fault diagnosis |
-
2021
- 2021-01-04 CN CN202110010119.7A patent/CN114718861A/en active Pending
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2017161126A1 (en) * | 2016-03-16 | 2017-09-21 | University Of Houston System | System and method for detecting, diagnosing, and correcting trips or failures of electrical submersible pumps |
| CN105781527A (en) * | 2016-03-24 | 2016-07-20 | 中国海洋石油总公司 | Parameter diagnosis and analysis method for electric submersible pump well working condition instrument |
| CN107687332A (en) * | 2017-09-08 | 2018-02-13 | 张雅芝 | Screw pump and working conditions of oil extractor on-line analysis diagnosis and control method of adjustment and device are realized using electrical parameter |
| CN108266178A (en) * | 2018-01-14 | 2018-07-10 | 常州大学 | A kind of sucker rod pump monitoring of working condition method based on rich information indicator card and DBN |
| CN109360120A (en) * | 2018-10-22 | 2019-02-19 | 常州大学 | A kind of electric immersible pump well operating condition intelligent diagnosing method based on convolutional neural networks |
| US20200209109A1 (en) * | 2018-12-28 | 2020-07-02 | Shanghai United Imaging Intelligence Co., Ltd. | Systems and methods for fault diagnosis |
| CN110284875A (en) * | 2019-06-13 | 2019-09-27 | 中国石油化工股份有限公司 | A kind of Diagnosing The Faults of Esp ' method and apparatus based on current parameters |
Non-Patent Citations (1)
| Title |
|---|
| 王辉等: "电潜泵抽油井工况诊断分析方法综述", 石油工程技术, pages 62 - 64 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116256589A (en) * | 2023-05-15 | 2023-06-13 | 南京研控科技有限公司 | Intelligent diagnosis method and device for electric pump well, storage medium and server |
| CN116906025A (en) * | 2023-07-24 | 2023-10-20 | 西南石油大学 | Oil extraction engineering production abnormality diagnosis device |
| CN116906025B (en) * | 2023-07-24 | 2024-06-11 | 西南石油大学 | Oil extraction engineering production abnormality diagnosis device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111914883B (en) | Spindle bearing state evaluation method and device based on deep fusion network | |
| CN108985380B (en) | Point switch fault identification method based on cluster integration | |
| CN107305601B (en) | Method for analyzing efficiency factors of oil pumping well system | |
| CN114718861A (en) | Intelligent diagnosis method for working condition of screw pump well based on deep learning | |
| CN111784093B (en) | Enterprise reworking auxiliary judging method based on power big data analysis | |
| CN114358124B (en) | New fault diagnosis method for rotary machinery based on deep countermeasure convolutional neural network | |
| CN113137211B (en) | Oil well production parameter self-adaptive control method based on fuzzy comprehensive decision | |
| CN107025468A (en) | Highway congestion recognition methods based on PCA GA SVM algorithms | |
| CN112576492A (en) | Intelligent diagnosis method for electric submersible pump production well fault | |
| CN108509701A (en) | A kind of direct intelligent diagnosing method of rotating machinery fault based on vibration signal | |
| CN114596010B (en) | BiGRU network drilling condition identification method combined with attention mechanism | |
| CN106022352A (en) | Submersible piston pump fault diagnosis method based on support vector machine | |
| CN113095414A (en) | Indicator diagram identification method based on convolutional neural network and support vector machine | |
| CN110318731A (en) | A kind of oil well fault diagnostic method based on GAN | |
| CN112664185A (en) | Indicator diagram-based rod-pumped well working condition prediction method | |
| CN112664184A (en) | Pumping unit well pump detection period prediction method based on support vector regression algorithm | |
| CN114819315A (en) | Bearing degradation trend prediction method based on multi-parameter fusion health factor and time convolution neural network | |
| CN109894495A (en) | A kind of extruder method for detecting abnormality and system based on energy consumption data and Bayesian network | |
| CN114021620A (en) | Electrical submersible pump fault diagnosis method based on BP neural network feature extraction | |
| CN108678728A (en) | A kind of oilwell parameter analysis combinational algorithm based on k-means | |
| CN117475191A (en) | Bearing fault diagnosis method for feature alignment domain antagonistic neural network | |
| CN109000924B (en) | Method for monitoring state of ball screw pair based on K mean value | |
| CN115880472A (en) | Intelligent diagnosis and analysis system for electric power infrared image data | |
| CN112861957B (en) | Method and device for detecting running state of oil well | |
| CN114004151A (en) | Oil pumping well fault prediction method based on LSTM-VGGNet |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |