An integrated working condition diagnosis method based on electric parameter and indicator diagram information fusion

Abstract

The application discloses a comprehensive working condition diagnosis method based on electric parameter and dynamometer card information fusion, and based on dictionary coding of current state of an oil pumping well motor as "0: stable, 1: rising, 2: falling", the working condition is classified as follows: A class which can use electric parameter data to quickly diagnose specific working condition; B class and C class which can use electric parameter data to quickly diagnose specific working condition category but cannot determine specific working condition, B class such as down pump (B1) and insufficient liquid supply (B2) working condition with current coding all as "0-02"; C class such as fixed valve leakage (C1) and pipe leakage (C2) working condition with current coding all as "2-0"; and D class which cannot use current change to represent working condition. The algorithm fully gives play to the advantages of electric parameter identification and dynamometer card identification method, has higher accuracy than traditional electric parameter identification method, and has better identification accuracy and speed than traditional dynamometer card identification method.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of oil well production, and particularly relates to a comprehensive working condition diagnosis method based on information fusion of electric parameters and dynamometer card. BACKGROUND

[0002] The dynamometer card of the pumping unit and the electric parameters of the pumping unit are the main means to understand the working conditions of the pipe, rod and pump in the pumping unit well. However, the current of the pumping unit cannot fully reflect the downhole information, and there are cases where the current characteristics are the same but the working conditions are different. Therefore, using the current technology, the accuracy of the algorithm for determining the working condition of the pumping well according to the electric parameters of the pumping unit is low. On the other hand, all abnormal phenomena in the operation of the deep well pump can be directly reflected on the dynamometer card, and analyzing and interpreting the dynamometer card is a main means to directly understand the working condition of the deep well pump. The pumping unit is the core equipment in the oil exploitation link, and the working condition of the pumping unit and the operation efficiency of the pumping unit are low, so it cannot meet the needs of the development of the industry. Under the background of the current artificial intelligence trend, especially using deep learning technology to mine the information expressed by the electric parameters and the dynamometer card, it is of great significance to improve the accuracy and efficiency of the working condition or fault recognition of the pumping unit and ensure the stable and reliable operation of the pumping unit. The current traditional pumping well recognition methods include mechanical model analysis and rod pump well fault diagnosis expert system, but due to the differences in geology in different places and the aging of machines, the recognition results will have large differences if the traditional dynamometer card recognition method is used. At present, the recognition of the pumping working condition mainly includes the electric parameter judgment method, but the overall accuracy is low. There is also an artificial recognition method of the dynamometer card, which has the problems of slow recognition speed, great influence of individual subjective factors and poor consistency. Machine recognition based on the dynamometer card also gradually appears, but it has the problems of slow recognition speed and low accuracy. SUMMARY

[0003] The present application discloses a pumping working condition recognition method based on information fusion of electric parameters and dynamometer card information. From the engineering applicability, the method combines data dictionary technology, machine deep learning and information fusion reasoning means. The specific steps of the method are as follows:

[0004] An integrated working condition diagnosis algorithm based on the fusion of electrical parameters and dynamometer card information, characterized in that: based on the dictionary coding of the current state of the pumping well motor as "0: stable, 1: rising, 2: falling", the working conditions are classified as follows: A class that can use electrical parameter data to quickly diagnose specific working conditions, such as "2-1" for overloading of the balance block and "212-121" for vibration; B and C classes that can use electrical parameter data to quickly diagnose a large class of specific working conditions but cannot determine the specific working condition, B class such as "0-02" for the current code of the down-pumping pump (B1) and insufficient liquid supply (B2) working condition; C class such as "2-0" for the current code of the fixed valve leakage (C1) and pipe leakage (C2) working condition; and D class that cannot be characterized by current changes; and the specific steps include the following:

[0005] (1) Use the electrical parameter algorithm to quickly match based on the current parameters of the pumping well motor;

[0006] (2) According to the matching result of step (1), determine whether the working condition result is an A class working condition,

[0007] (2-1) If the result is an A class working condition, output the result directly;

[0008] (2-2) If the result is not an A class working condition, go to the next step;

[0009] (3) According to the matching result of step (1), determine whether the working condition result is a B or C class working condition,

[0010] (3-1) If the result is a B or C class working condition, use the image classification algorithm to determine the working condition based on the dynamometer card using a lightweight model, and perform information fusion and comprehensive diagnosis on the results of the two judgment algorithms, and go to the next step;

[0011] (3-2) If the result is not a B or C class working condition, use the image classification algorithm to determine the working condition based on the dynamometer card using a lightweight model, and go to the next step;

[0012] (4) Output the judgment result.

[0013] Preferably, the working condition is determined based on the dynamometer card using a lightweight model, and the specific steps are as follows:

[0014] (3-A) Obtain the dynamometer card of the pumping well to be detected;

[0015] (3-B) Input the dynamometer card into the trained SE-ShuffleNet lightweight network model for recognition;

[0016] (3-C) Output the recognition result.

[0017] Preferably, the specific steps of training the SE-ShuffleNet lightweight network model are:

[0018] (3-B-1) Obtain the load and displacement data collected by the oil pumping well dynamometer, and draw a dynamometer card based on the above data;

[0019] (3-B-2) Business experts classify and label the dynamometer card;

[0020] (3-B-3) Image enhancement is performed on the labeled dynamometer card to form a sample set;

[0021] (3-B-4) Put the sample set into the SE-ShuffleNet lightweight network model based on the improved SE-Net for training;

[0022] (3-B-5) Determine whether the training accuracy is greater than 98%:

[0023] If the training accuracy is not greater than 98%, return to step (3-B-4);

[0024] If the training accuracy is greater than 98%, go to the next step;

[0025] (3-B-6) After training, save the trained SE-ShuffleNet weight.

[0026] Specifically, three SE-Net are inserted at the end of Stage2, Stage3 and Stage4 of ShuffleNet v2 network, and the improved network structure diagram is shown in the following table:

[0027]

[0028] In the table, Image represents the input image, Conv1MaxPool represents a 3*3 convolution layer and a pooling layer, three Stages represent three repeated basic units of ShuffleNet, Conv5 is a 1*1 convolution layer unique to ShuffleNet, GlobalPool is a pooling layer with a feature map size of 1*1, FC is a full connection layer for outputting class prediction values; √ represents that SE-Net is used, Output Size is the output matrix size, KSize is the size of the convolution kernel, Stride is the unit that distinguishes the different strides of ShuffleNetv2, and Repeat is the number of times of repeating the used unit.

[0029] Preferably, the specific steps of the rapid working condition matching based on the oil pumping well motor current data are:

[0030] (1-1) Obtain the motor current data collected by the oil pumping well sensor;

[0031] (1-2) distinguish upstroke and downstroke: one method is to distinguish in a ratio close to 1:1, and another method is to divide upstroke and downstroke according to the opening and closing points of the valve;

[0032] (1-3) filter the current data and calculate the slope;

[0033] (1-4) determine the change of motor current in upstroke and downstroke according to the slope and generate current change code;

[0034] (1-5) determine the working condition according to the current code and refer to the fault diagnosis table;

[0035] (1-6) output the determination result.

[0036] Preferably, the current change code specifically comprises:

[0037]

[0038] Preferably, the specific steps of the comprehensive working condition diagnosis algorithm process based on information fusion are:

[0039] (3-1-A) record the (B1+B2) or (C1+C2) type results matched with the acquired electric parameters, and process the image classification results into the form of A, B1, B2, C1, C2 and D;

[0040] (3-1-B) according to the formula, perform DS evidence reasoning on the results of the two algorithms;

[0041] (3-1-C) perform comprehensive diagnosis according to the results after evidence reasoning;

[0042] (3-1-C) output the diagnosis result.

[0043] Preferably, the specific formula for information fusion based on the results of the two algorithms of electric parameter rapid working condition matching and indicator diagram working condition diagnosis is:

[0044]

[0045] wherein m1 () is the result of diagnosis using the electric parameter algorithm, m2 () is the result of diagnosis using the image classification algorithm, and m 12 () is the result of comprehensive diagnosis using the two algorithms; wherein the normalization constant ; Y j represents A, B1, B2, C1, C2 and D working conditions; represents one of A, (B1+B2), (C1+C2) and D in the electric parameter algorithm, is the result of the electric parameter algorithm for the case of value; represent one of the six cases A, B1, B2, C1, C2, D in the image classification algorithm, for the image classification algorithm to the case value.

[0046] beneficial effects

[0047] 1. The algorithm fully utilizes the advantages of electrical parameter identification and indicator diagram identification methods, has higher accuracy than traditional electrical parameter identification methods, has better identification accuracy and speed than traditional indicator diagram identification methods, and has better stability and consistency brought by information fusion reasoning.

[0048] 2. The entire system uses data dictionary technology, machine deep learning and information fusion reasoning means, breaks through the limitations of traditional identification methods, and from the actual engineering, this method can reduce the performance requirements of embedded devices, ensure the accuracy and work efficiency of system working condition diagnosis, and has more practical application value.

[0049] 3. The improved SE-ShuffleNet network is used. ShuffleNetv2 uses channel splitting and channel splicing, reduces the number of parameters and calculations, and strengthens the exchange of information between different groups. SENet performs global average pooling on the input, followed by two full connections, then uses an activation function to fix the value in the [0, 1] interval, and finally multiplies the weight of each channel of the input layer obtained by the input layer, strengthens the ability of the network to extract useful information, and the model size growth is also negligible.

[0050] 4. An information fusion diagnosis method is proposed. A composite model can be constructed based on two algorithms and information fusion. The composite model has no obvious difference in model size compared with the single indicator diagram identification model, but can realize the real-time identification of the system, and can improve the speed and accuracy of the working condition judgment.

[0051] 5. The entire system has a dynamic learning function, which will send the judgment result to the upper position after each identification, and will put the graphs of various working conditions into the matching graph library, and upgrade and iterate the matching graph library. Gradually improve the identification accuracy when judging the working condition in the future. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 is the overall flowchart of the method of the present application

[0053] Figure 2 is the flowchart of the acquisition, labeling and classification of the indicator diagram and the training of the network model of the present application

[0054] Figure 3Flow chart for quick matching of oil well motor current data of the present application

[0055] Figure 4 Flow chart for condition diagnosis based on lightweight model using dynamometer card of the present application

[0056] Figure 5 Flow chart for comprehensive condition diagnosis algorithm based on information fusion of the present application

[0057] Figure 6 Results of diagnosis based on image classification algorithm in Example 2 DETAILED DESCRIPTION

[0058] A comprehensive condition diagnosis algorithm based on information fusion of electrical parameters and dynamometer card, based on the dictionary coding of the current state of the oil well motor as "0: stable, 1: rising, 2: falling", the conditions are classified as follows: A: conditions that can be quickly diagnosed using electrical parameter data, such as over-heavy balance block "2-1", heavy oil "1-2", damaged plunger and pump interval "0-0", waxing "1-2", double-valve leakage "2-2", vibration "212-121", etc.; B: lower pump (B1) and insufficient liquid supply (B2) conditions with current coding "0-02"; C: fixed valve leakage (C1) and pipe leakage (C2) conditions with current coding "2-0"; D: other conditions. In combination with Figure 1 , the above algorithm includes the following flow:

[0059] (1) Quick matching based on oil well motor current parameters;

[0060] (1) Determine whether the condition result is an A-class condition,

[0061] (2-1) If the result is an A-class condition, output the result directly;

[0062] (2-2) If the result is not an A-class condition, proceed to the next step;

[0063] (3) Determine whether the condition result is a B or C class condition,

[0064] (3-1) If the result is a B or C class condition, use a lightweight model based on the dynamometer card to determine the condition, and perform information fusion and comprehensive diagnosis of the results of the two algorithms, and proceed to the next step;

[0065] (3-2) If the result is not a B or C class condition, use a lightweight model based on the dynamometer card to determine the condition, and perform information fusion and comprehensive diagnosis of the results of the two algorithms, and proceed to the next step;

[0066] (4) Output the determination result.

[0067] The indicated power diagram acquisition, labeling classification and training steps include acquisition of the indicated power diagram, labeling classification of the existing indicated power diagram data set, training of the SE-ShuffleNet light network model using the labeled indicated power diagram data set, and saving the trained weight value; combined with Figure 2 The specific steps are:

[0068] (3-B-1) Obtain the load and displacement data collected by the oil pumping well dynamometer, and draw the indicated power diagram based on the above data;

[0069] (3-B-2) The business expert classifies and labels the indicated power diagram;

[0070] (3-B-3) The labeled indicated power diagram is image enhanced to form a sample set;

[0071] (3-B-4) Put the sample set into the SE-ShuffleNet light network model based on the improved SE-Net for training;

[0072] For example:

[0073] Example 1: The three SE-Net are respectively inserted at the end of the ShuffleNet v2 network Stage2, Stage3 and Stage4, and the improved network structure diagram is shown in the following table. In the table, Image represents the input image, Conv1MaxPool represents a 3*3 size convolution layer and a pooling layer, three Stages represent three repeated basic units of ShuffleNet, Conv5 is a 1*1 convolution layer unique to ShuffleNet, GlobalPool is a 1*1 pooling layer, FC is a full connection layer for outputting class prediction value; represents that SE-Net is used, Output Size is the size of the output matrix, KSize is the size of the convolution kernel, Stride is the unit that distinguishes the different strides of ShuffleNetv2, and Repeat is the number of times of repeating the unit. The improved network structure diagram is shown in the following table:

[0074]

[0075] (3-B-5) Determine whether the training accuracy is greater than 98%,

[0076] If the training accuracy is not greater than 98%, return to step (2-4);

[0077] If the training accuracy is greater than 98%, go to the next step;

[0078] (3-B-6) After training, save the trained SE-ShuffleNet weight value;

[0079] (3-B-7) identifying the indicator diagram of the sample set;

[0080] (3-B-8) outputting the training result.

[0081] In combination Figure 3 The specific steps of the quick working condition matching based on the electric motor current data of the pumping well are as follows:

[0082] (1-1) acquiring the electric motor current data collected by the pumping well sensor;

[0083] (1-2) distinguishing the upstroke and downstroke;

[0084] (1-3) filtering the current data and calculating the slope;

[0085] (1-4) judging the change of the upstroke and downstroke motor current according to the slope and generating the current change code;

[0086] (1-5) judging the working condition according to the current code and the fault diagnosis table;

[0087] (1-6) outputting the judgment result.

[0088] The current change code specifically includes:

[0089]

[0090] The working condition diagnosis based on the indicator diagram using the lightweight model, in combination Figure 4 The specific steps are as follows:

[0091] (3-A) acquiring the indicator diagram to be detected of the pumping well;

[0092] (3-B) inputting the indicator diagram into the trained SE-ShuffleNet lightweight network model for identification;

[0093] (3-C) outputting the identification result.

[0094] In combination Figure 5 The specific steps of the comprehensive working condition diagnosis algorithm process based on information fusion are as follows:

[0095] (3-1-A) recording the (B1+B2) or (C1+C2) type results matched by the acquired electric parameters, and processing the image classification results into the form of A, B1, B2, C1, C2 and D;

[0096] (3-1-B) performing DS evidence reasoning on the two algorithm results according to the formula;

[0097] (3-1-C) performing comprehensive diagnosis according to the result after the evidence reasoning;

[0098] (3-1-D) output diagnosis results.

[0099] The specific formula of information fusion based on the results of the two algorithms of quick working condition matching based on electrical parameters and indicator diagram working condition diagnosis is:

[0100]

[0101] In the formula, m1 () is the result of diagnosis using the electrical parameter algorithm, m2 () is the result of diagnosis using the image classification algorithm, and m 12 () is the result of comprehensive diagnosis using the two algorithms; wherein the normalization constant ; Y j represents A, B1, B2, C1, C2, and D working conditions. represents one of A, (B1+B2), (C1+C2), and D four cases in the electrical parameter algorithm, is the value of the electrical parameter algorithm to case; is the value of the electrical parameter algorithm to case. represents one of A, B1, B2, C1, C2, and D six cases in the image classification algorithm, is the value of the image classification algorithm to case.

[0102] For example:

[0103] Example 2: According to Figure 6 , the working condition result of image classification is normal, and the probability value is 0.452, but the electrical parameter diagnosis result is fixed valve leakage or pipe leakage, and the actual working condition result is pipe leakage. Therefore, information fusion is introduced at this time.

[0104] As shown in the following table. In the table, m1 () is the result of diagnosis using the electrical parameter algorithm, m2 () is the result of diagnosis using the image classification algorithm, and m 12 () is the result of comprehensive diagnosis using the two algorithms.

[0105]

[0106] The value of m12 () in the table is calculated according to the formula , wherein the normalization constant ; Y j represents A, B1, B2, C1, C2, and D working conditions. represents one of A, (B1+B2), (C1+C2), and D four cases in the electrical parameter algorithm, is the value of the electrical parameter algorithm to case; is the value of the electrical parameter algorithm to representing one of the six cases A, B1, B2, C1, C2, D in the image classification algorithm, for the image classification algorithm the case of value.

[0107] According to the table, the result of the electrical parameter diagnosis algorithm is C. However, since the current codes of class C are the same, it is not possible to determine whether it is C1 (fixed valve leakage) or C2 (pipe leakage) condition, so the system performs image classification. In the result of image classification, the probability value of pipe leakage is only second, which is 0.345, but after information fusion, it becomes 0.99995. According to the result of comprehensive diagnosis, it is diagnosed as pipe leakage, which avoids the diagnosis error caused by directly accepting the result of image classification.

Claims

1. A comprehensive operating condition diagnosis method based on the fusion of electrical parameters and dynamometer card information, characterized in that, Based on the dictionary coding of the current status of the pumping well motor as "0: stable, 1: rising, 2: falling", the operating conditions are classified as follows: Category A, which uses electrical parameter data to quickly diagnose specific operating conditions; Categories B and C, which use electrical parameter data to quickly diagnose the general category of specific operating conditions but cannot determine the specific operating conditions. Category B includes the down-pump pump B1 and insufficient fluid supply B2 operating conditions with current change coding of "0-02"; Category C includes the fixed valve leakage C1 and pipe leakage C2 operating conditions with current change coding of "2-0"; and Category D, which includes other operating conditions that cannot be characterized by current changes. The specific steps include: (1) Use electrical parameter algorithm to quickly match based on the current parameters of the pumping well motor; (2) Based on the matching results of step (1), determine whether the working condition result is a Class A working condition; (2-1) If the judgment result is Class A working condition, output the judgment result directly; (2-2) If the judgment result is not Class A working condition, proceed to step (3); (3) Based on the matching results of step (1), determine whether the working condition result is a B or C type working condition; (3-1) If the judgment result is a B or C type of working condition, the working condition is judged using an image classification algorithm, specifically including: judging the working condition using a lightweight model based on the dynamometer diagram; and after the results of the two judgment algorithms are fused and comprehensively diagnosed, proceed to step (4); wherein, the specific steps for fusion and comprehensive diagnosis of the results of the two judgment algorithms are as follows: (3-1-A) Let B or C be the result of the electrical parameter matching obtained, and process the image classification result into the form of A, B1, B2, C1, C2, D; (3-1-B) Based on the formula for information fusion using the results of the two judgment algorithms, perform DS evidence reasoning on the results of the two judgment algorithms; the specific formula is as follows: In the formula, m1() represents the result diagnosed using the electrical parameter algorithm, and m2() represents the result diagnosed using the image classification algorithm. 12 () represents the combined diagnostic results using two algorithms; where the normalization constant is... ;Y j This indicates the operating conditions of A, B1, B2, C1, C2, and D. This represents one of the four cases A, B, C, and D in the electrical parameter algorithm. For electrical parameter algorithm situation value; This represents one of the six cases A, B1, B2, C1, C2, and D in an image classification algorithm. For image classification algorithms situation value; (3-1-C) Make a comprehensive diagnosis based on the results of reasoning from the evidence; (3-1-D) Output diagnostic results; (3-2) If the judgment result is not a B or C type working condition, use an image classification algorithm to judge the working condition, specifically including: using a lightweight model to judge the working condition based on the dynamometer diagram; proceed to step (4); (4) Output the judgment result.

2. The method according to claim 1, characterized in that, The specific steps for determining operating conditions using a lightweight model based on dynamometer diagrams are as follows: (3-A) Obtain the dynamometer card of the oil well to be tested; (3-B) Input the dynamometer diagram into the trained SE-ShuffleNet lightweight network model for recognition; (3-C) Output the recognition results.

3. The method according to claim 2, characterized in that, The specific steps for training the SE-ShuffleNet lightweight network model are as follows: (3-B-1) Obtain the load and displacement data collected by the dynamometer of the oil well, and draw the dynamometer diagram based on the load and displacement data; (3-B-2) Business experts classify and label the indicator diagrams; (3-B-3) Perform image enhancement on the labeled dynamometer diagrams to form a sample set; (3-B-4) The sample set is fed into the SE-ShuffleNet lightweight network model, which is an improvement on SENet, for training; (3-B-5) Determine if the training accuracy is greater than 98%: If the training accuracy is not greater than 98%, return to step (3-B-4). If the training accuracy is greater than 98%, proceed to step (3-B-6). (3-B-6) After training is completed, save the trained SE-ShuffleNet weights.

4. The method according to claim 1, characterized in that, The specific steps for rapid matching based on the pumping well motor current parameters are as follows: (1-1) Obtain motor current data collected by the oil well sensor; (1-2) Differentiate between upper and lower strokes: differentiate them according to a ratio of approximately 1:1, or divide the upper and lower strokes according to the opening and closing points of the valve; (1-3) Filter the current data and calculate the slope; (1-4) Determine the changes in motor current during the up and down strokes based on the slope and generate current change codes; (1-5) Determine the operating condition based on the current change code and the fault diagnosis table; (1-6) Output the judgment result.

5. The method according to claim 4, characterized in that, The specific relationship between operating conditions and current changes in coding is as follows: When the balance block is too heavy, the current change trend is "down" - "up", the current change code is 2-1, and the code classification belongs to Class A; When the operating condition is a bottom-impact pump, the current change trend is "flat" - "flat bottom", the current change code is 0-02, and the code classification belongs to B1. When the operating condition is insufficient liquid supply, the current change trend is "flat" - "flat down", the current change code is 0-02, and the code classification belongs to B2. When the operating condition is heavy oil, the current change trend is "up" - "down", the current change code is 1-2, and the code classification belongs to category D; When the operating condition is leakage of the floating valve, the current change trend is "up" - "flat", the current change code is 1-0, and the code classification belongs to Class A; When the rod breaks, the current change trend is "down" - "down", the current change code is 2-2, and the code classification belongs to category D. When the operating condition is a fixed valve leakage, the current change trend is "down" - "flat", the current change code is 2-0, and the code classification belongs to C1. When the operating condition is pipe leakage, the current change trend is "down" - "flat", the current change code is 2-0, and the code classification belongs to C2. When the operating condition is waxing, the current change trend is "up" - "down", the current change code is 1-2, and the code classification belongs to category D; When the operating condition is double valve leakage, the current change trend is "down"-"down", the current change code is 2-2, and the code classification belongs to Class D; When the operating condition is influenced by gas, the current change trend is "flat down" - "down flat", the current change code is 02-20, and the code classification belongs to Class A; When the operating condition is vibration, the current change trend is "down-up-down" - "up-down-up", the current change code is 212-121, and the code classification belongs to Class A; When the operating condition is top-impact pump, the current change trend is "flat-top-flat" - "flat", the current change code is 010-0, and the code classification belongs to Class A.

Images