Method for monitoring working condition of rod-pumped well based on load data symbol representation

Abstract

The invention provides a rod-pumped well working condition monitoring method based on load data symbol representation, and relates to the technical field of petroleum engineering. According to the method, the load time sequence data are fully utilized, the time dimension information is restored, the influence of redundant displacement information in the indicator diagram synthesis process is reduced, and the data utilization rate and the monitoring efficiency are improved; a symbol sequence is constructed, the dependence on expert experience is reduced, and the influence of a formalization method based on domain knowledge on a monitoring result when the shape features of the ground indicator diagram are quantified is weakened; feature word vectors are constructed, computer analysis and processing are facilitated, and retrieval and matching efficiency is improved; a working condition knowledge base is constructed, the efficiency of managing working condition load data resources is improved, the redundancy of knowledge in load data is reduced, and the reusability of the knowledge is improved; similarity measurement is introduced, the retrieval efficiency of the knowledge base is improved, calculation is easy and convenient, applicability is high, a weight matrix is introduced, contribution of part of feature words to similarity calculation is enhanced, and accuracy and flexibility of measurement are improved.

Description

Technical Field

[0001] The invention belongs to the technical field of petroleum engineering, and in particular relates to a method for monitoring the working condition of a rod pumping unit well based on load data symbol representation. Background Art

[0002] The rod pumping well is the most widely used oil production well, playing an important role in the domestic and foreign petroleum industry. The main equipment of the rod pumping well works thousands of meters underground in a harsh working environment. There are often complex working conditions such as gas impact, insufficient fluid supply, fixed valve or floating valve leakage, plunger disengagement, etc. Once these working conditions are not discovered and resolved in time, it will not only cause crude oil production losses, but may even lead to severe safety and environmental accidents. Therefore, the working condition monitoring of the rod pumping well is of great significance for improving the recovery rate and ensuring the safe production of oil fields. The closed curve formed by the suspension point displacement and the suspension point load, also known as the ground indicator diagram, is the main data source for the working condition monitoring of the rod pumping well. The difference in the shape of the ground indicator diagram reflects the changes in the working condition of the rod pumping well. The existing monitoring methods mainly rely on expert experience to extract the shape features of the indicator diagram, so as to build a classifier to realize the working condition monitoring of the rod pumping well. However, existing monitoring methods are overly dependent on subjective factors such as expert experience. The level of experts and the formalization of expert experience will have a great impact on the monitoring results. For example, for ground dynamometers with complex shape features, judging the corresponding working conditions requires not only rich domain knowledge and experience, but also accurate formalization methods to quantify the shape features. In addition, the load data of rod pumping wells with different working conditions vary greatly, while the displacement data vary less. The ground dynamometers not only increase redundant displacement information but also lose the time dimension information contained in the load data, which limits the improvement of the existing monitoring level.

[0003] The patent application document with application number CN202210739084.5 proposes "a method for downhole fault diagnosis of rod pumping units based on machine learning". The implementation scheme of the patent is as follows: first, the preprocessed indicator diagram is divided into multiple curve segments by using expert experience, and the geometric moment and central moment of each curve segment are calculated and integrated into vectors according to specific rules, so as to construct a spatial sample set of indicator diagrams of different working conditions; then, the spatial sample set is preliminarily reduced in dimension by the principal component analysis method to construct the cross-sectional space, and the cross-sectional space samples are projected to the new weak nonlinear cross-sectional space under the condition of minimizing the mean and variance at the same time; finally, the local linear embedding algorithm is used to extract the features of the samples in the weak nonlinear cross-sectional space and the support vector machine is used to judge the downhole working conditions of the rod pumping unit well.

[0004] The patent application document with application number CN202310977728.9 proposes a "deep learning-based fault diagnosis method for pumping unit indicator diagram". The implementation scheme of the patent is as follows: first, the data acquisition equipment is used to collect the suspension point load and suspension point displacement data of the rod pumping unit and draw them into images; then, the abnormal images are eliminated and the data is enhanced by flipping and rotating operations; finally, the deep neural network with attention mechanism is used to classify the images, so as to realize the working condition monitoring of the rod pumping unit well.

[0005] Chinese patent CN202210739084.5 "A method for downhole fault diagnosis of rod pumping units based on machine learning". Although this method can realize downhole working condition monitoring of rod pumping units, the ground dynamometer data used in this method contains redundant displacement information and ignores the time dimension information of load data, which makes it difficult to obtain the changing trend of working conditions during the continuous operation of rod pumping units. In addition, this method relies on expert experience to segment the ground dynamometer diagram, and requires complex calculations to quantify the shape characteristics of each curve segment.

[0006] Chinese patent CN202310977728.9 "Fault diagnosis method of pumping unit indicator diagram based on deep learning". Although this method inputs the ground indicator diagram as an image into the deep neural network for image classification, it reduces the reliance on expert experience. However, operations such as flipping and rotation in this method will change the working condition information contained in the ground indicator diagram. For example, the ground indicator diagram corresponding to the floating valve leakage is similar to the ground indicator diagram corresponding to the fixed valve leakage after rotation. Since the mechanism information contained in the ground indicator diagram is ignored, it is easy to cause misjudgment of the working condition. Summary of the invention

[0007] In view of the shortcomings of the prior art, the purpose of the present invention is to propose a method for realizing automatic detection of the working conditions of a rod pumping well by using load time series data, so as to solve the problem that the existing monitoring method is too dependent on subjective factors such as expert experience, which leads to the influence of the level of experts and the formal method of expert experience on the monitoring results. And solve the problem that the existing monitoring method uses ground indicator diagram data, which leads to the increase of redundant displacement information and the loss of time dimension information contained in the load data.

[0008] A method for monitoring the working condition of a rod pumping well based on load data symbol representation comprises the following steps:

[0009] Step 1: Obtain the load time series data of the rod pumping well and preprocess the load time series data of the rod pumping well to obtain the preprocessed load time series data of the rod pumping well; the preprocessing is standardization; the load time series data of the rod pumping well are the load values ​​at the suspension point of the rod pumping well at several different times under different working conditions; the load value at the suspension point includes the sucker rod column load, the liquid column load acting on the plunger, the inertia load and the friction load;

[0010] Step 1.1: Collect the load values ​​at the suspension point of the rod pumping well at equal time intervals, and then obtain the load time series data X = {x1, x2, ..., x j ,…,x n}, where x j is the load value at the hanging point of the rod pumping well at the jth moment, n is the number of samples of the load value at the hanging point of the rod pumping well, 1≤j≤n;

[0011] Step 1.2: Calculate the mean and standard deviation of the load values ​​at the hanging point of the rod pumping well in the load time series data X, and standardize the load values ​​at the hanging point of the rod pumping well in the load time series data X to obtain standardized load time series data Z;

[0012] For the load time series data X within one stroke of the rod pumping well, the standardized load time series data Z = {z1, z2, ..., z j ,…,z n}:

[0013]

[0014] In the formula, z j is the load value at the suspension point of the rod pumping well at the jth moment in the standardized load time series data Z, is the mean of the load values ​​at the suspension point of the rod pumping well in the load time series data X, and S is the standard deviation of the load values ​​at the suspension point of the rod pumping well in the load time series data X;

[0015] Step 2: Use the segmented aggregation approximation algorithm to segment the standardized load time series data to obtain the segmented mean sequence;

[0016] Specifically, determine the number of segments ω, 1<ω<n, and convert the standardized load time series data Z={z1,z2,…,z j , …, z n} is divided into ω segments, and each segment is represented by the mean of the load values ​​at the suspension point of the rod pumping well falling within the segment. The mean values ​​of each segment form a segment mean sequence P = {p1, p2, ..., p h ,…,p ω};

[0017] According to the number of segments ω, the mean value of each segment of the standardized load time series data is calculated using formula (2):

[0018]

[0019] Among them, p h is the mean value of the hth segment in the standardized load time series data Z, where h is the segment number;

[0020] Step 3: Adopt the adaptive kernel density estimation method to calculate the probability density function and distribution function of the load value at the suspension point in the standardized load time series data;

[0021] Step 3.1: For the standardized load time series data Z = {z1, z2, ..., z j , …, z n}, the probability density function f(z) of the load value at the suspension point in the standardized load time series data is calculated using the adaptive kernel density estimation method, and the calculation formula is as follows:

[0022]

[0023] Where z is a point in the probability density function, δ is the bandwidth, δ is adaptively adjusted by the local density, and K(·) is the kernel function;

[0024] Step 3.2: Calculate the distribution function F(z) based on the probability density function f(z). The calculation formula is as follows:

[0025]

[0026] Among them, τ is the independent variable of the probability density function;

[0027] Step 4: Map the mean of each segment in the segment mean sequence into a symbol to form a symbol sequence of the segment mean sequence;

[0028] Step 4.1: By weighing the loss of information on the structure of the payload timing data and the effect of symbolic representation, determine the number of symbols k and the number of symbols {e1, e2, …, e k},e k is the kth symbol;

[0029] Step 4.2: According to the distribution function F(z) and the number of symbols k, determine the equally probable symbols representing the segment points t1, t2, ..., t k-1 , t k-1 The k-1th equally probable symbol representation segmentation point is a segmentation point of an equally probable interval determined according to a distribution function, and these segmentation points make the probability of the symbol falling in each interval equal;

[0030] Step 4.3: According to the determined symbols and equally probable symbols representing the segmentation points, the segment mean sequence P in step 2 is converted into {p1, p2, ..., p h ,…,p ω} is mapped to a symbol sequence S = {s1, s2, ..., s h ,…,s ω};

[0031]

[0032] Among them, g(p h ) is the mapping function, and the segmented mean sequence P is mapped to obtain the symbol sequence S = {s1, s2, ..., s h ,…,s ω},s h ∈{e1, e2, …, e k}, 1≤h≤ω;

[0033] Step 5: construct a feature word dictionary and use a sliding window to extract feature words in the symbol sequence, then use a vectorization method to obtain the feature word coordinates and use them to construct a feature word vector of the symbol sequence; the feature word is a word composed of several adjacent symbols in the symbol sequence, and the feature word dictionary is a collection of all feature words;

[0034] Step 5.1: Set the number of symbols in the feature word, obtain all feature words consisting of the set number of symbols, and then construct a feature word dictionary; the dimension of the feature dictionary is the number of feature words, and each dimension of the feature dictionary corresponds to a feature word;

[0035] Step 5.2: For the symbol sequence S = {s1, s2, ..., s h ,…,s ω}, use the sliding window to perform word segmentation, extract the feature words appearing in the feature word dictionary from the symbol sequence S and construct a set of extracted feature words;

[0036] Furthermore, the window size is set to b, the sliding step length is set to l, and starting from symbol s1, the feature words contained in the window after each sliding are extracted and stored until symbol s ω , extracted feature words, and construct a set of extracted feature words;

[0037] Step 5.3: Calculate the feature word weight according to the extracted feature word set, obtain the feature word coordinates and construct the feature word vector based on them;

[0038] Furthermore, the frequencies of all the feature words in the extracted feature word set are counted, and the frequencies of the feature words are used as weights. For feature words that do not appear, the frequencies are recorded as 0. The frequencies of the feature words are used as the feature word coordinates under the corresponding dimensions of the feature word dictionary to construct the feature word vector W = {w 1 , w 2 ,…,w p}, p is the dimension of the feature word dictionary;

[0039] Step 6: According to steps 1 to 5, construct the feature word vectors corresponding to the load time series data under several working conditions as templates and use them to build a working condition knowledge base;

[0040] Further, according to the method for obtaining load time series data described in step 1, load time series data in multiple strokes during the historical production process of the rod pumping well are collected, and m load time series data containing c working conditions are uniformly selected, c < m, and recorded as X1, X2, ..., X m ; For each load time series data X i ={x i1 , x i2 , …, x ij , …, x in}, 1≤f≤m, follow the method described in steps 1 to 5 to construct the feature word vector W i As a template and stored, traverse m load time series data to establish the working condition knowledge base KB, as shown in the following formula:

[0041]

[0042] in, 1≤q≤p, 1≤i≤m represents the feature word coordinates of the qth dimension in the i-th feature word vector, label i represents the working condition category label corresponding to the i-th feature word vector; i is the number of the load time series data, x ij is the load value at the suspension point of the rod pumping well at the jth moment in the i-th load time series data, W i is the feature word vector corresponding to the i-th load time series data;

[0043] Step 7: Introduce the similarity measurement of working conditions, collect new load time series data and construct the corresponding feature word vector, calculate its similarity with each template in the working condition knowledge base, and determine the working condition category through template matching;

[0044] Step 7.1: Collect new load time series data and construct a feature word vector corresponding to the newly collected load time series data according to the method described in steps 1 to 5;

[0045] Further, according to the method of step 1 to step 5, construct the new collected load time series data Xα ={x α1 , x α2 , …, x αn}The corresponding feature word vector

[0046] Step 7.2: Introduce the similarity measurement of working conditions and establish the calculation formula of similarity measurement;

[0047] Specifically: construct the new collection load time series data feature word vector W α Any template W in the working condition knowledge base i , introduce the similarity measurement calculation formula (9) of the working conditions:

[0048]

[0049] Among them, Sim(W α , W i ) represents the feature word vector W α Any template W in the working condition knowledge base i Similarity measure between them, D = dag(d1, d2, ..., d p ) is the weight matrix, d q is the qth diagonal element of matrix D, 1≤q≤p, and d is calculated by the normalized formula (10): q value:

[0050]

[0051] in, is the variance of the feature word coordinates of the qth dimension in all feature word vectors in the working condition knowledge base, is the set of variances of feature word coordinates of all feature word vectors from 1 to p dimensions, min(S 2 ) is the minimum value of the variance of the feature word coordinates of all p dimensions, max(S 2 ) is the maximum variance of the feature word coordinates of all p dimensions, Calculated by formula (11):

[0052]

[0053] in, is the mean of the feature word coordinates of the qth dimension of all feature word vectors;

[0054] Step 7.3: Calculate the similarity measure between the feature word vector of the newly collected load time series data and each template in the working condition knowledge base using the calculation formula of similarity measure, and determine the working condition type through template matching;

[0055] Specifically: traverse the working condition knowledge base KB and calculate W in sequence α The similarity measure with each template in KB is arranged from high to low according to the similarity measure size, and a majority vote is adopted. According to the voting results, the working condition labels corresponding to the K templates with the largest similarity measure are the working condition categories corresponding to the newly collected load time series data.

[0056] Compared with the prior art, the beneficial effects of the present invention are:

[0057] 1. The present invention uses the load time series data at the suspension point as the data source for analyzing the working conditions of the rod pumping unit well, makes full use of the load time series data that reflects the changes in working conditions, restores the time dimension information, and reduces the influence of redundant displacement information in the process of dynamometer diagram synthesis, thereby improving data utilization and monitoring efficiency.

[0058] 2. The present invention proposes a symbolic representation method for constructing a symbolic sequence corresponding to load time series data, which reduces the dependence on expert experience and weakens the impact of formal methods based on domain knowledge on monitoring results when quantifying the shape characteristics of the ground dynamometer diagram.

[0059] 3. The present invention proposes a symbol vectorization method for constructing a feature word vector of a symbol sequence. The data type is converted by the vectorization method, which is convenient for computer analysis and processing. The constructed feature word vector is convenient for calculating similarity, thereby improving the efficiency of retrieval and matching.

[0060] 4. The present invention establishes a working condition knowledge base using the characteristic word vector of load time series data as a template to manage the load data resources of the working condition, thereby reducing the redundancy of knowledge in the load data and increasing the reusability of knowledge.

[0061] 5. The present invention introduces a similarity measure of working conditions for knowledge base retrieval, which is simple to calculate and has strong applicability. By allocating weights, the contribution of some feature words to similarity calculation is enhanced, thereby improving the accuracy and flexibility of the measure. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 It is a flow chart of a method for monitoring the working condition of a rod pumping unit well based on the symbolic representation of load data in an embodiment of the present invention;

[0063] Figure 2 It is a schematic diagram of visualization of time series data of loads of six working conditions that have been standardized in an embodiment of the present invention;

[0064] Figure 3 It is a visualization diagram of the segmented mean sequence of 6 working conditions in the embodiment of the present invention;

[0065] Figure 4Schematic diagram of probability density functions and equal-area regions of six operating conditions in an embodiment of the present invention;

[0066] Figure 5 A schematic diagram of extracting symbol sequence feature words and constructing feature word vectors in an embodiment of the present invention. DETAILED DESCRIPTION

[0067] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0068] The present invention provides a method for realizing automatic monitoring of the working condition of a rod pumping well by using load time series data. The load time series data at the suspension point is used as the data source for analyzing the working condition of the rod pumping well. First, a symbol sequence is constructed by using segmented aggregation approximation and symbol mapping; secondly, a feature word dictionary is constructed and feature words in the symbol sequence are extracted, and the feature word coordinates are obtained by using a vectorization method and a feature word vector is constructed based on this; then, according to historical production data, several working condition feature word vectors are constructed as templates to form a working condition knowledge base; finally, a similarity measure of working conditions is introduced, and the similarity between the feature word vector of the newly collected load time series data and each template in the working condition knowledge base is calculated, and the working condition monitoring result is given by template matching.

[0069] A method for monitoring the working condition of a rod pumping well based on the symbolic representation of load data, such as Figure 1 As shown, the following steps are included:

[0070] Step 1: Obtain the load time series data of the rod pumping well and preprocess the load time series data of the rod pumping well to obtain the preprocessed load time series data of the rod pumping well; the preprocessing is standardization; the load time series data of the rod pumping well are the load values ​​at the suspension point of the rod pumping well at several different times under different working conditions; the load value at the suspension point includes the sucker rod column load, the liquid column load acting on the plunger, the inertia load and the friction load;

[0071] Step 1.1: Collect the load value at the suspension point of the rod pumping well at equal time intervals, and then obtain the load time series data X = {x1, x2, ..., x j , …, x n}, where x j is the load value at the hanging point of the rod pumping well at the jth moment, n is the number of samples of the load value at the hanging point of the rod pumping well, 1≤j≤n;

[0072] In this embodiment, the load sensor between the clamping plates of the rope hanger is used to collect the load time series data, and the collected load time series data X={45.06, 45.85, ..., 42.31}.

[0073] Step 1.2: Calculate the mean and standard deviation of the load values ​​at the hanging point of the rod pumping well in the load time series data X, and standardize the load values ​​at the hanging point of the rod pumping well in the load time series data X to obtain standardized load time series data Z;

[0074] For the load time series data X within one stroke of the rod pumping well, the standardized load time series data Z = {z1, z2, ..., z j , …, z n}:

[0075]

[0076] In the formula, z j is the load value at the suspension point of the rod pumping well at the jth moment in the standardized load time series data Z, is the mean of the load values ​​at the suspension point of the rod pumping well in the load time series data X, and S is the standard deviation of the load values ​​at the suspension point of the rod pumping well in the load time series data X;

[0077] In this embodiment, the load time series data is standardized according to formula (1), and Z={-0.7466, -0.6872, ..., -0.9540} is obtained. The schematic diagram of the visualization of the standardized load time series data is as follows: Figure 2 shown.

[0078] Step 2: Use the segmented aggregation approximation algorithm to segment the standardized load time series data to obtain the segmented mean sequence;

[0079] Specifically, determine the number of segments ω (1<ω<n), and convert the standardized load time series data Z = {z1, z2, ..., z j , …, z n} is divided into ω segments, and each segment is represented by the mean of the load values ​​at the suspension point of the rod pumping well falling within the segment. The mean values ​​of each segment form a segment mean sequence P = {p1, p2, ..., p h ,…,p ω};

[0080] According to the number of segments ω, the mean value of each segment of the standardized load time series data is calculated using formula (2):

[0081]

[0082] Among them, p h is the mean value of the hth segment in the standardized load time series data Z, where h is the segment number;

[0083] In this implementation, it is recommended that the number of segments ω=16, and the standardized load time series data Z={-0.7466, -0.6872, ..., -0.9540} is equally divided into 16 segments, each segment is represented by the mean of the load values ​​at the suspension point of the rod pumping well falling within the segment, and the segmented mean sequence P={-0.4297, 0.6573, ..., -1.0069} composed of the mean values ​​of each segment after segmentation is calculated according to the segmented aggregation approximate mean calculation formula (2). Figure 3 The visualization of the segmented mean sequences of the six working conditions in the embodiment is given.

[0084] Step 3: Adopt the adaptive kernel density estimation method to calculate the probability density function and distribution function of the load value at the suspension point in the standardized load time series data;

[0085] Step 3.1: For the standardized load time series data Z = {z1, z2, ..., z j , …, z n}, the probability density function f(z) of the load value at the suspension point in the standardized load time series data is calculated using the adaptive kernel density estimation method, and the calculation formula is as follows:

[0086]

[0087] Where z is a point in the probability density function, δ is the bandwidth, δ is adaptively adjusted by the local density, and K(·) is the kernel function;

[0088] A Gaussian kernel function with wide applicability and good smoothness properties is selected as formula (4):

[0089]

[0090] in, is the independent variable of kernel function;

[0091] In this implementation, the distribution of the standardized load time series data is relatively complex, and the probability density function has single-peak, multi-peak, symmetric, asymmetric and other distribution forms. Selecting the Gaussian kernel function can improve the stability of the probability density estimation.

[0092] In this embodiment, for the standardized load time series data Z={-0.7466, -0.6872, ..., -0.9540}, the probability density function f(z) is calculated according to formula (3) and formula (4): Figure 4 Schematic diagram of probability density functions and equal-area regions of six operating conditions in an embodiment of the present invention.

[0093] Step 3.2: Calculate the distribution function F(z) based on the probability density function f(z). The calculation formula is as follows:

[0094]

[0095] In order to distinguish it from the independent variable of F(z), let τ be the independent variable of the probability density function;

[0096] In this embodiment, the distribution function F(z) is calculated according to formula (5).

[0097] Step 4: Map the mean of each segment in the segment mean sequence into a symbol to form a symbol sequence of the segment mean sequence;

[0098] Step 4.1: By weighing the loss of information on the structure of the payload timing data and the effect of symbolic representation, determine the number of symbols k and the number of symbols {e1, e2, …, e k},e k is the kth symbol; the structural information loss refers to the degree of distortion or loss of information contained in the original data when the data type is converted;

[0099] In this implementation, it is suggested through experiments that the number of symbols k is 3, and it is suggested that e1=L, e2=M, and e3=H are used to represent the segmented mean sequence. The L symbol contains the semantics of "low load", the M symbol contains the semantics of "medium load", and the H symbol contains the semantics of "high load". The number of symbols obtained is 3, which can not only meet the requirements of the representation effect, but also reduce the loss of structural information and reduce the calculation in the subsequent steps;

[0100] Step 4.2: According to the distribution function F(z) and the number of symbols k, determine the equally probable symbols representing the segment points t1, t2, ..., t k-1 , t k-1 The k-1th equally probable symbol representation segmentation point is a segmentation point of an equally probable interval determined according to a distribution function, and these segmentation points make the probability of the symbol falling in each interval equal;

[0101] In this implementation, based on the distribution function and the number of symbols, the F(z) equal probability interval segmentation points t1=-0.9456 and t2=0.9567 are determined.

[0102] Step 4.3: According to the determined symbols and equally probable symbols representing the segmentation points, the segment mean sequence P in step 2 is converted into {p1, p2, ..., p h ,…,p ω} is mapped to a symbol sequence S = {s1, s2, ..., s h ,…,s ω};

[0103]

[0104] Among them, g(p h ) is the mapping function, g(p h)=e1 represents the hth mean value p in the segment mean sequence P h When it is less than or equal to t1, the symbol e1 is used to represent the mean value p h The corresponding load value at the suspension point in this load time series data is mapped through the segmented mean sequence P to obtain the symbol sequence S = {s1, s2, ..., s h ,…,s ω},s h ∈{e1, e2, …, e k}, 1≤h≤ω;

[0105] In this embodiment, for the segmented mean sequence P = {-0.4297, 0.6573, ..., -1.0069}, according to the determined symbols and equal probability interval segmentation points t1 = -0.9456 and t2 = 0.9567, the segmented mean sequence P is traversed from the 1st mean to the 16th mean, and each segmented mean is mapped to a symbol according to formula (7) to form a symbol sequence S = {M, M, ..., L} of the segmented mean sequence.

[0106]

[0107] Step 5: construct a feature word dictionary and use a sliding window to extract feature words in the symbol sequence, then use a vectorization method to obtain the feature word coordinates and use them to construct a feature word vector of the symbol sequence; the feature word is a word composed of several adjacent symbols in the symbol sequence, and the feature word dictionary is a collection of all feature words;

[0108] Step 5.1: Set the number of symbols in the feature word, obtain all feature words consisting of the set number of symbols, and then construct a feature word dictionary; the dimension of the feature word dictionary is the number of feature words, and each dimension of the feature word dictionary corresponds to a feature word;

[0109] In this embodiment, it is recommended that the feature word is a word composed of three adjacent symbols in the symbol sequence, and the feature word is a description of the change of the load value at the suspension point in the load time series data during this period of time. There are 27 feature words with a length of 3 composed of symbols L, M, and H, which are introduced into the feature word dictionary shown in Table 1.

[0110] Table 1 Feature word dictionary

[0111]

[0112]

[0113] In the feature word dictionary, each feature word corresponds to a dimension of the feature word dictionary, and all feature words constitute a feature word dictionary of 27 dimensions.

[0114] Step 5.2: For the symbol sequence S = {s1, s2, ..., sh , …, s ω}, perform word segmentation using a sliding window, extract the feature words that appear in the feature word dictionary from the symbol sequence S, and construct a set of the extracted feature words;

[0115] Further, set the window size to b and the sliding step size to l. Starting from symbol s1, extract the feature words contained in the window after each slide and store them until symbol sω, and a total of feature words are extracted to construct a set of the extracted feature words. Figure 5 This is a schematic diagram for extracting the feature words of the symbol sequence and constructing the feature word vector in the embodiment of the present invention.

[0116] In this embodiment, it is recommended that b = 3 and l = 1. Count the extracted feature words and construct a set of the extracted feature words.

[0117] Step 5.3: Calculate the feature word weights according to the set of the extracted feature words, obtain the feature word coordinates, and construct the feature word vector based on this;

[0118] Further, count the frequencies of all the feature words that appear in the set of the extracted feature words, use the frequencies of the feature words as weights, and for the feature words that do not appear, record the frequency as 0. Use the frequencies of the feature words as the feature word coordinates under the corresponding dimensions of the feature word dictionary, and construct the feature word vector W = {w 1 , w 2 , …, w p}, where p is the dimension of the feature word dictionary;

[0119] In this embodiment, the feature word vector W = {w 1 , w 2 , …, w 27} = {0.2143, 0, …, 0.2857}

[0120] Step 6: Construct the feature word vectors corresponding to the load time series data under several working conditions as templates according to Steps 1 to 5, and construct a working condition knowledge base based on this;

[0121] Further, according to the method for obtaining the load time series data described in Step 1, collect the load time series data within multiple strokes during the historical production process of the rod pumping well, uniformly select m pieces of load time series data (c < m) containing c working conditions, denoted as X1, X2, …, X m . For each piece of load time series data X i = {x i1 , x i2 , …, x ij , …, x in} (1 ≤ i ≤ m), construct the feature word vector W according to the method described in Steps 1 to 5.i As a template and store it, traverse m load time series data to establish the working condition knowledge base KB, where i is the number of the load time series data, x ij is the load value at the suspension point of the rod pumping well at the jth moment in the i-th load time series data, W i is the feature word vector corresponding to the i-th load time series data;

[0122]

[0123] in, Represents the feature word coordinate of the qth dimension in the i-th feature word vector, label i Indicates the working condition category label corresponding to the i-th feature word vector;

[0124] In this implementation, production data of a certain oil field over a period of time is selected, including 6 types of working conditions, each type of working condition includes 6 sets of load time series data, each set of load time series data includes 144 equally spaced sampling points, a total of 36 sets of load time series data, recorded as X1, X2, ..., X 36 ,The label , name , cause of the working condition and the presentation form of the ,load time series data are shown in Table 2.

[0125] Table 2 Labels, names, causes and presentation of load time series data for working conditions

[0126]

[0127] According to the method described in steps 1 to 5, construct the feature word vectors W1, W2, ..., W corresponding to the working condition 36 It is used as a template and stored to build a working condition knowledge base KB of a certain oil field.

[0128] Step 7: Introduce the similarity measurement of working conditions, collect new load time series data and construct the corresponding feature word vector, calculate its similarity with each template in the working condition knowledge base, and determine the working condition category through template matching;

[0129] Step 7.1: Collect new load time series data and construct the feature word vector corresponding to the newly collected load time series data according to the methods of steps 1 to 5;

[0130] Further, according to the method described in step 1 to step 5, a new collection load time series data X is constructed. α ={x α1 , x α2 , …, x αn}The corresponding feature word vector

[0131] Step 7.2: Introduce the similarity measurement of working conditions and establish the calculation formula of similarity measurement;

[0132] Specifically: construct the new collection load time series data feature word vector W α Any template W in the working condition knowledge base i , the similarity measurement calculation formula (9) of the working conditions is introduced. In this similarity measurement, the closer the value is to 1, the higher the similarity is, and the closer the value is to 0, the lower the similarity is;

[0133]

[0134] Among them, Sim(W α , W i ) represents the feature word vector W α Any template W in the working condition knowledge base i Similarity measure between them, D = dag(d1, d2, ..., d p ) is the weight matrix, d q (1≤q≤p) is the qth diagonal element of the matrix D, and d is calculated by the normalization formula (10) q value:

[0135]

[0136] in, is the variance of the feature word coordinates of the qth dimension in all feature word vectors in the working condition knowledge base, is the set of variances of feature word coordinates of all feature word vectors from 1 to p dimensions, min(S 2 ) is the minimum value of the variance of the feature word coordinates of all p dimensions, max(S 2 ) is the maximum variance of the feature word coordinates of all p dimensions, Calculated by formula (11):

[0137]

[0138] in, is the mean of the feature word coordinates of the qth dimension of all feature word vectors;

[0139] Step 7.3: Calculate the similarity measure between the feature word vector of the newly collected load time series data and each template in the working condition knowledge base using the calculation formula of similarity measure, and determine the working condition type through template matching;

[0140] Specifically: traverse the working condition knowledge base KB and calculate W in sequence α The similarity measure with each template in KB is arranged from high to low according to the similarity measure size, and a majority vote is adopted. According to the voting results, the working condition labels corresponding to the K templates with the largest similarity measure are the working condition categories corresponding to the newly collected load time series data.

[0141] In this implementation, K=3 is recommended. According to the feature word dictionary introduced in step 5, the weight matrix D=diag(0.4878283, 0.2427467, ..., 1). After experiments, the accuracy of the monitoring system reached 0.882, and the average monitoring time was 0.668 seconds.

[0142] The present invention is verified on the load time series data collected in a certain oil field. For the 6 working conditions of normal, insufficient liquid supply, gas influence, fixed valve leakage, floating valve leakage, and plunger disengagement, a total of 36 load time series data are generated by the rod pumping well. According to the method described in the present invention, a feature word vector is constructed for each load time series data according to steps 1 to 5, and the 36 load time series data are divided into two parts: one part is used to construct the working condition knowledge base according to step 6, and the remaining part is tested according to step 7. Through repeated experiments, the accuracy of the working condition monitoring method of the rod pumping well proposed by the present invention reaches 0.882. The technical method proposed by the present invention has achieved good results in the working condition monitoring of the rod pumping well.

[0143] In addition, the symbolic representation method proposed in the present invention is verified on other machine learning methods. For 36 load time series data, the classification accuracy of the decision tree is 0.735, the classification accuracy of the naive Bayes is 0.765, and the classification accuracy of the support vector machine is 0.844. The test results show that the symbolic representation method proposed in the present invention can be combined with a variety of different machine learning methods to realize the monitoring of working conditions, and has wide applicability.

[0144] Compared with some existing technologies for monitoring the working conditions of rod pumping wells, the present invention directly uses the load time series data at the suspension point as the data source for analyzing the working conditions of rod pumping wells, thereby improving data utilization and monitoring efficiency. Compared with the use of Fourier descriptors, directed chain codes, curve moments and other methods to extract the shape features of the ground dynamometer diagram, the method proposed in the present invention reduces the reliance on expert experience and formal methods based on domain knowledge, avoids complex calculations when extracting shape features, and realizes automatic monitoring of the working conditions of rod pumping wells. In addition, the symbolic representation of the working conditions of rod pumping wells can intuitively reflect the significant characteristics of different working conditions, and is an effective method for judging the working conditions of rod pumping wells.

Claims

1. A method for monitoring the working condition of a rod pumping well based on load data symbol representation, characterized in that: The steps include: Step 1: Obtain the load time series data of the rod pumping well and preprocess the load time series data of the rod pumping well to obtain the preprocessed load time series data of the rod pumping well; Step 2: Use the segmented aggregation approximation algorithm to segment the standardized load time series data to obtain the segmented mean sequence; Step 3: Adopt the adaptive kernel density estimation method to calculate the probability density function and distribution function of the load value at the suspension point in the standardized load time series data; Step 4: Map the mean of each segment in the segment mean sequence into a symbol to form a symbol sequence of the segment mean sequence; Step 5: construct a feature word dictionary and use a sliding window to extract feature words in the symbol sequence, then use a vectorization method to obtain the feature word coordinates and use them to construct a feature word vector of the symbol sequence; the feature word is a word composed of several adjacent symbols in the symbol sequence, and the feature word dictionary is a collection of all feature words; Step 6: According to steps 1 to 5, construct the feature word vectors corresponding to the load time series data under several working conditions as templates and use them to build a working condition knowledge base; Step 7: Introduce the similarity measurement of working conditions, collect new load time series data and construct the corresponding feature word vector, calculate its similarity with each template in the working condition knowledge base, and determine the working condition category through template matching.

2. A method for monitoring the working condition of a pumping rod well based on load data symbol representation according to claim 1, characterized in that: The preprocessing in step 1 is standardization; the load time series data of the pumping rod well are the load values ​​at the suspension point of the pumping rod well at several different times under different working conditions; the load values ​​at the suspension point include the sucker rod column load, the liquid column load acting on the plunger, the inertia load and the friction load.

3. A method for monitoring the working condition of a pumping rod well based on load data symbol representation according to claim 2, characterized in that: The step 1 specifically includes: Step 1.1: Collect the load values ​​at the suspension point of the rod pumping well at equal time intervals, and then obtain the load time series data X = {x1, x2, ..., x j ,…,x n }, where x j is the load value at the hanging point of the rod pumping well at the jth moment, n is the number of samples of the load value at the hanging point of the rod pumping well, 1≤j≤n; Step 1.2: Calculate the mean and standard deviation of the load values ​​at the hanging point of the rod pumping well in the load time series data X, and standardize the load values ​​at the hanging point of the rod pumping well in the load time series data X to obtain standardized load time series data Z; For the load time series data X within one stroke of the rod pumping well, the standardized load time series data Z = {z1, z2, ..., z j ,…,z n }: In the formula, z j is the load value at the suspension point of the rod pumping well at the jth moment in the standardized load time series data Z, is the mean of the load values ​​at the suspension point of the rod pumping well in the load time series data X, and S is the standard deviation of the load values ​​at the suspension point of the rod pumping well in the load time series data X.

4. A method for monitoring the working condition of a pumping rod well based on load data symbol representation according to claim 3, characterized in that: The specific content of step 2 is as follows: determine the number of segments ω, where 1 < ω < n, and equally divide the standardized load time series data Z = {z1, z2, …, z j , …, z n} into ω segments. Each segment is represented by the mean value of the load values at the polished rod suspension point of the rod pumping well that fall within that segment. The segment mean value sequence P = {p1, p2, …, p h , …, p ω} is composed of the mean values of each segment; According to the number of segments ω, the mean value of each segment of the standardized load time series data is calculated using formula (2): Among them, p h is the mean value of the hth segment in the standardized load time series data Z, where h is the segment number.

5. A method for monitoring the working condition of a pumping rod well based on load data symbol representation according to claim 4, characterized in that: The step 3 specifically includes: Step 3.1: For the standardized load time series data Z = {z1, z2, …, z j ,…,z n }, the probability density function f(z) of the load value at the suspension point in the standardized load time series data is calculated using the adaptive kernel density estimation method, and the calculation formula is as follows: Where z is a point in the probability density function, δ is the bandwidth, δ is adaptively adjusted by the local density, and K(·) is the kernel function; Step 3.2: Calculate the distribution function F(z) based on the probability density function f(z). The calculation formula is as follows: Among them, τ is the independent variable of the probability density function.

6. A method for monitoring the working condition of a rod pumping unit based on load data symbol representation according to claim 5, characterized in that: The step 4 specifically includes: Step 4.1: By weighing the loss of information on the structure of the payload timing data and the effect of symbolic representation, determine the number of symbols k and the number of symbols {e1, e2, …, e k },e k is the kth symbol; Step 4.2: According to the distribution function F(z) and the number of symbols k, determine the equally probable symbols representing the segment points t1, t2, …, t k-1 , t k-1 The k-1th equally probable symbol representation segmentation point is a segmentation point of an equally probable interval determined according to a distribution function, and these segmentation points make the probability of the symbol falling in each interval equal; Step 4.3: According to the determined symbols and equally probable symbols representing the segmentation points, the segment mean sequence P in step 2 is converted into {p1, p2, …, p h ,…,p ω } is mapped to a symbol sequence S = {s1, s2, ..., s h ,…,s ω }; Among them, g(p h ) is the mapping function, which traverses the segmented mean sequence P and maps it to obtain the symbol sequence S = {s1, s2, ..., s h ,…,s ω },s h ∈{e1,e2,…,e k },1≤h≤ω.

7. A method for monitoring the working condition of a pumping rod well based on load data symbol representation according to claim 6, characterized in that: The step 5 specifically includes: Step 5.1: Set the number of symbols in the feature word, obtain all feature words consisting of the set number of symbols, and then construct a feature word dictionary; the dimension of the feature word dictionary is the number of feature words, and each dimension of the feature word dictionary corresponds to a feature word; Step 5.2: For the symbol sequence S = {s1, s2, ..., s h ,…,s ω }, use the sliding window to perform word segmentation, extract the feature words appearing in the feature word dictionary from the symbol sequence S and construct a set of extracted feature words; Set the window size to b, the sliding step length to l, start from symbol s1, extract and store the feature words contained in the window after each sliding, until symbol s ω , extracted feature words, and construct a set of extracted feature words; Step 5.3: Calculate the feature word weight according to the extracted feature word set, obtain the feature word coordinates and construct the feature word vector based on them; The frequency of all the feature words in the extracted feature word set is counted, and the frequency of the feature words is used as the weight. For the feature words that do not appear, the frequency is recorded as 0. The frequency of the feature words is used as the feature word coordinates under the corresponding dimension of the feature word dictionary to construct the feature word vector W = {w 1 ,w 2 ,…,w p }, p is the dimension of the feature word dictionary.

8. A method for monitoring the working condition of a rod pumping unit based on load data symbol representation according to claim 7, characterized in that: The specific steps of step 6 are as follows: According to the method for obtaining load time series data described in step 1, collect the load time series data within multiple strokes during the historical production process of the rod pumping well, and uniformly select m pieces of load time series data containing c working conditions, where c < m, denoted as X1, X2, …, X m ; for each piece of load time series data X i = {x i1 , x i2 , …, x ij , …, x in}, 1 ≤ i ≤ m, construct the feature word vector W i as a template and store it, and traverse the m pieces of load time series data to establish the working condition knowledge base KB; in, 1≤q≤p,1≤i≤m represents the feature word coordinates of the qth dimension in the i-th feature word vector, label i represents the working condition category label corresponding to the i-th feature word vector; i is the number of the load time series data, x ij is the load value at the suspension point of the rod pumping well at the jth moment in the i-th load time series data, W i is the feature word vector corresponding to the i-th load time series data.

9. A method for monitoring the working condition of a pumping rod well based on load data symbol representation according to claim 8, characterized in that: The step 7 is specifically as follows: Step 7.1: Collect new load time series data and construct a feature word vector corresponding to the newly collected load time series data according to the method described in steps 1 to 5; According to the methods from step 1 to step 5, construct the newly collected load time series data X α ={x α1 ,x α2 ,…,x αn }The corresponding feature word vector Step 7.2: Introduce the similarity measurement of working conditions and establish the calculation formula of similarity measurement; Specifically: construct the new collected load time series data feature word vector W α Any template W in the working condition knowledge base i , introduce the similarity measurement calculation formula (9) of the working conditions: Among them, Sim(W α ,W i ) represents the feature word vector W α Any template W in the working condition knowledge base i The similarity measure between them is D = diag(d1, d2, …, d p ) is the weight matrix, d q is the qth diagonal element of matrix D, 1≤q≤p, and d is calculated by normalization formula (8): q value: in, is the variance of the feature word coordinates of the qth dimension in all feature word vectors in the working condition knowledge base, is the set of variances of feature word coordinates of all feature word vectors from 1 to p dimensions, min(S 2 ) is the minimum value of the variance of the feature word coordinates of all p dimensions, max(S 2 ) is the maximum variance of the feature word coordinates of all p dimensions, Calculated by formula (9): in, is the mean of the feature word coordinates of the qth dimension of all feature word vectors; Step 7.3: Use the similarity metric calculation formula to calculate the similarity metric between the feature word vector of the newly collected load time series data and each template in the working condition knowledge base, and determine the working condition type through template matching.

10. A method for monitoring the working condition of a rod pumping unit based on load data symbol representation according to claim 9, characterized in that: The step 7.3 is specifically as follows: traverse the working condition knowledge base KB, calculate W in sequence α The similarity measure with each template in KB is arranged from high to low according to the similarity measure size, and a majority vote is adopted. According to the voting results, the working condition labels corresponding to the K templates with the largest similarity measure are the working condition categories corresponding to the newly collected load time series data.

Images