Method for extracting fault characteristic parameters of reciprocating plunger pump under informationization condition

Abstract

This invention relates to a method for extracting fault characteristic parameters of a reciprocating piston pump under information-based conditions, comprising the following steps: 1. Acquiring the cylinder vibration signal Y of the reciprocating piston pump; 2. Obtaining the reconstructed signal X based on the LMD algorithm and the comprehensive evaluation index CEI; 3. Constructing a time-frequency matrix S based on time-frequency analysis. M×N 4. Determine S M×N 5. Obtain the optimal decomposition dimension k in nonnegative matrix decomposition; 6. Obtain the basis matrix W that can characterize different features of the reciprocating piston pump. M×k And coefficient matrix H k×N 6. Construct the fault basis matrix V based on the optimal decomposition dimension k and cosine similarity theory. M×k 7. Fault characteristic parameters of reciprocating piston pumps are obtained based on time-frequency analysis theory. This invention can overcome the influence of piston impact characteristics and accurately extract fault characteristics of reciprocating piston pumps under operation, providing reliable technical support for subsequent fault diagnosis of reciprocating piston pumps.

Description

Technical Field

[0001] This invention relates to reciprocating plunger pumps in the field of oil extraction, and more particularly to a method for extracting fault characteristic parameters of reciprocating plunger pumps under information-based conditions, belonging to the field of intelligent fault diagnosis technology for oil fields. Background Technology

[0002] With the widespread application and industrialization of hydraulic transmission, both national and corporate governments are increasingly emphasizing its development. Against this backdrop, reciprocating piston pumps have become a research hotspot in the mechanical energy to hydraulic energy conversion industry. They are not only energy-efficient and highly effective, but also capable of automating hydraulic transmission by incorporating modern electronic control technology. To ensure the proper operation and maintenance of reciprocating piston pumps, it is essential to promptly detect their faults to prevent major operational safety accidents. Therefore, there is an urgent need to research efficient fault diagnosis techniques for reciprocating piston pumps. One of the key aspects of fault diagnosis technology is the extraction of fault features from reciprocating piston pumps. Current methods for extracting fault features from reciprocating piston pumps are based on feature processing of the original signal. Due to interference from the piston impact frequency, it is difficult to extract weak fault features from the signal, such as:

[0003] The paper "BP Fault Diagnosis of Pulsating Pressure Signals in Piston Pumps Based on Wavelet Packet Transform," published in the March 2023 issue of the journal *Hydraulics, Pneumatics & Seals*, discloses a method that uses wavelet packet transform to eliminate noise in the original signal and then extracts a frequency band composed of fault information to complete the analysis of the signal's time-domain characteristics. When the frequency band is between 0 and 4 kHz, the pulsating pressure signal of the piston pump contains pump shaft frequency, fundamental frequency, and harmonic frequencies, including various information parameters of the pump's operation. In the high-frequency range, the line spectrum amplitude continuously decreases. Fault diagnosis results using a BP network established with MATLAB software show that the pulsating pressure index can accurately identify the internal operating conditions of the piston pump, meeting the testing requirements for fault diagnosis. The method is effective in diagnosis and has good practical guiding significance in the field of fault diagnosis of pulsating pressure signals in piston pumps. However, this method is susceptible to interference from piston impact frequencies, which can mask weak fault features, making them difficult to extract.

[0004] The paper "Fault Diagnosis of Oilfield Water Injection Plunger Pumps Based on Spectrum Analysis," published in the March 2023 issue of the journal *Fluid Machinery*, addresses the problems of traditional water injection plunger pump fault diagnosis, such as high reliance on experience, insufficient real-time processing capabilities, and low accuracy. It proposes an online frequency domain signal diagnosis method for plunger pumps. Vibration signals from the plunger pump under normal and 16 fault states were collected and analyzed using Fourier transform. A fault model library was established, and vibration signals were acquired online for fault feature extraction. Fault diagnosis was performed by comparing the root mean square error (RMSE) of multiple consecutive time intervals with the model library. Simultaneously, the contribution rate of sensors in the fault was ranked to reduce sensor dimensionality and computational load. The results show that using this method to diagnose axial plunger pump faults improves the accuracy from 66.7% to over 80%, while reducing the computational load to 26.7% of the original. However, this method struggles to distinguish between piston impact characteristics and fault characteristics, which affects the accuracy of the final diagnostic results. Summary of the Invention

[0005] The purpose of this invention is to overcome the problems existing in the prior art and provide a method for extracting fault characteristic parameters of reciprocating piston pumps under information conditions. This method can overcome the influence of piston impact characteristics, accurately extract fault characteristics of reciprocating piston pumps under operation, and provide reliable technical support for subsequent fault diagnosis of reciprocating piston pumps.

[0006] To solve the above technical problems, the method for extracting fault characteristic parameters of a reciprocating plunger pump under information-based conditions of the present invention includes the following steps in sequence:

[0007] Step 1: Collect the cylinder vibration signal Y of the reciprocating piston pump;

[0008] Step 2: Obtain the reconstructed signal X based on the LMD algorithm and the comprehensive evaluation index CEI;

[0009] Step 3: Construct the time-frequency matrix S based on time-frequency analysis method M×N ;

[0010] Step 4: Determine S M×N The optimal decomposition dimension k value in nonnegative matrix factorization;

[0011] Step 5: Obtain the basis matrix W that can characterize different features of the reciprocating piston pump. M×k And coefficient matrix H k×N ;

[0012] Step 6: Construct the fault basis matrix V based on the optimal decomposition dimension k and cosine similarity theory. M×k ;

[0013] Step 7: Obtain the fault characteristic parameters of the reciprocating piston pump based on time-frequency analysis theory.

[0014] Further, step 2 is as follows: the cylinder vibration signal Y is decomposed based on the LMD algorithm to obtain multiple components P, the comprehensive evaluation index CEI of each component P is calculated, and the comprehensive evaluation index vector CEI' is constructed; Q fusion indices with values ​​greater than the mean of the elements of CEI' are selected to obtain the weight vector ω, and the components P are reconstructed by weighting based on the weight vector ω to obtain the reconstructed signal X of the piston pump cylinder vibration.

[0015] Furthermore, step 3 is as follows: Perform time-frequency analysis on the reconstructed signal X, calculate the amplitude of the reconstructed signal X at each time and at each frequency, and construct the time-frequency matrix S of the reconstructed signal X using all amplitudes as elements. M×N .

[0016] Furthermore, step 4 is as follows: Based on matrix decomposition theory, the time-frequency matrix S from step 3 is... M×N Singular value decomposition is performed, and the number of characteristic frequencies of the signal is predicted by analyzing the singular values ​​of the matrix, thus determining S. M×N The optimal decomposition dimension k value in nonnegative matrix decomposition.

[0017] Furthermore, step 5 is as follows: Based on the nonnegative matrix factorization algorithm and the loss model of Euclidean distance, the time-frequency matrix S... M×N Decomposition is performed to obtain the basis matrix W, which can characterize different features of the reciprocating piston pump. M×k And coefficient matrix H k×N .

[0018] Furthermore, step 6 is as follows: Based on the optimal decomposition dimension k and cosine similarity theory, the basis matrix W is... M×k The column vectors α1, α2, ..., α3 in the array are... k Classify; if k=2, it indicates that the basis matrix W M×k If the matrix contains only characteristic information of piston impact and residual component information from nonnegative matrix decomposition, it is considered fault-free; if k > 2, it indicates that the basis matrix W M×k The system contains fault information. Keeping the column vectors corresponding to the fault information unchanged, the column vectors corresponding to the characteristic information of piston impact and the residual components in the non-negative matrix decomposition are replaced with zero vectors to construct the fault basis matrix V. M×k .

[0019] Furthermore, step 7 is as follows: The fault basis matrix V... M×k With coefficient matrix H k×N Perform matrix multiplication to obtain the fault characteristic matrix Z of the reciprocating piston pump. M×N .

[0020] Furthermore, based on time-frequency analysis theory, the fault characteristic matrix Z... M×NAn inverse transform is performed to obtain a time-domain signal Y' that can characterize the fault features of the reciprocating piston pump. A suitable mother wavelet is selected to perform three-level wavelet packet decomposition on the time-domain signal Y', and the energy values ​​of each sub-band are used as fault feature parameters of the reciprocating piston pump for input to the subsequent fault diagnosis model.

[0021] Furthermore, step 2 specifically includes the following sub-steps:

[0022] Step 2.1: For all local extreme points n of Y i Perform calibration and calculate n for two adjacent extreme points. i and n i+1 Local mean m i and envelope estimate a i ;

[0023] Step 2.2: Connect adjacent m i and a i Connect the segments with line segments, and smooth the extended polyline using the moving average method. Let the resulting local mean function be m. 11 (t), and let the envelope estimation function be a. 11 (t);

[0024] Step 2.3: Use the smoothed local mean function m extracted from the signal Y. 11 (t), and estimate the function a through the envelope. 11 (t) The demodulation result yields the frequency modulated signal s. 11 (t);

[0025] Step 2.4: Determine the frequency modulation signal s 11 The signal morphology of (t); when s 11 When (t) exhibits a non-pure frequency modulation signal morphology, its envelope estimation function a 12 (t)≠1; if the condition for a non-pure frequency modulated signal is met, s 11 (t) will be used as the original data and the above process will be repeated p times, eventually s 1p (t) represents a pure frequency modulated signal, at which point a 1(p+1) (t) = 1;

[0026] Step 2.5: Based on three evaluation indicators—Pearson correlation coefficient, impulse factor, and information entropy—this invention sets a comprehensive evaluation index (CEI) for component P.

[0027] Step 2.6: Calculate each component P j CEI jThe comprehensive evaluation index of each component signal is normalized to construct an index vector. The mean of the index is calculated, and Q fusion indexes with values ​​greater than the mean are selected to obtain the weight vector ω. At the same time, the corresponding component signals can be found. The residual components reflect the slowly changing quantities in the signal, so they are not considered. The reconstructed signal X is represented as follows:

[0028]

[0029] Furthermore, in step 3, the time-frequency matrix S is constructed using the short-time Fourier transform. M×N .

[0030] Furthermore, in step 4, based on matrix factorization theory, the time-frequency matrix S is... M×N Singular value decomposition is performed, and the formula for singular value decomposition is as follows:

[0031] S M×N =U∑V

[0032] Where U is an M-dimensional left singular matrix, V is an N-dimensional left singular matrix, and ∑ is an M×N diagonal matrix with non-zero elements σ1, σ2, ..., σ on the diagonal. k It is S M×N The singular values ​​of S, where K is the number of singular values; based on the number of singular values ​​K, for S M×N Predicting the optimal decomposition dimension k in nonnegative matrix factorization; establishing the objective function:

[0033]

[0034] h(k) is the ratio of the standard deviations of singular values ​​within a window transformed by the value of k, reflecting the smoothness of the signal; the smaller h(k) is, the smoother the signal. Therefore, the choice of the value of k satisfies the following formula:

[0035] k=arg min h(K),(k=2,...,K-2)

[0036] Furthermore, in step 5, after obtaining the optimal decomposition dimension k, the time-frequency matrix S of the cylinder vibration signal X is... M×N Perform nonnegative matrix decomposition; S M×N The nonnegative matrix factorization expression is as follows:

[0037] S M×N ≈W M×k H k×N

[0038] Among them, the left matrix W M×k Defined as a basis matrix, with the right matrix H k×N Defined as a coefficient matrix; in the time-frequency matrix S M×N In this context, N M-dimensional column vectors represent the potential features of the signal at the corresponding frequencies.

[0039] Furthermore, step 5 specifically includes the following sub-steps:

[0040] Step 5.1: Convert the basis matrix W M×k And coefficient matrix H k×N The elements in the array are initialized under the condition of nonnegation;

[0041] Step 5.2, Calculate S M×N With W M×k H k×N Euclidean distance D ED ;

[0042] Step 5.3: According to the gradient descent algorithm, for W... M×k and H k×N Update;

[0043] Step 5.4, when the Euclidean distance D ED When the basis matrix W reaches a stable value or tends to converge, M×k And coefficient matrix H k×N Updates stopped;

[0044] Step 5.5: Obtain the final basis matrix W M×k And coefficient matrix H k×N .

[0045] Compared with the prior art, the present invention has achieved the following beneficial effects: 1. The present invention defines a comprehensive evaluation index CEI, which can effectively characterize the time-varying nature of the fault characteristics of the component signal of the reciprocating plunger pump. The reconstruction weight constructed in this way gives the fault information in the reconstructed signal a larger weight, effectively highlighting the local features of the signal.

[0046] 2. Based on the singular value standard deviation ratio of the time-frequency matrix, this invention can predict the optimal decomposition dimension, solve the problem of decomposition dimension selection, effectively estimate the amount of various feature information contained in the time-frequency matrix, and improve the accuracy of non-negative matrix decomposition.

[0047] 3. This invention measures the similarity of each column vector in the time-frequency matrix of a reciprocating piston pump by using a similarity coefficient. This can effectively identify the time-frequency information representing piston impact and faults in the time-frequency matrix and obtain column vectors that can characterize the fault features of the reciprocating piston pump.

[0048] 4. Taking the cylinder vibration signal of a reciprocating piston pump as the research and analysis object, a time-frequency feature extraction method for reciprocating piston pumps based on the local mean decomposition (LMD) algorithm and the nonnegative matrix factorization (NMF) algorithm is proposed. This method can overcome the influence of piston impact characteristics and accurately extract the fault characteristics of the reciprocating piston pump under operation. Attached Figure Description

[0049] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The drawings are provided for reference and illustration only and are not intended to limit the present invention.

[0050] Figure 1 This is a flowchart of the method for extracting fault characteristic parameters of a reciprocating plunger pump under information-based conditions according to the present invention;

[0051] Figure 2 This is a flowchart illustrating the process of obtaining the basis matrix and coefficient matrix based on the nonnegative matrix decomposition algorithm in this invention. Detailed Implementation

[0052] In the following description of the present invention, the terms "upper", "lower", "front", "rear", "left", "right", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not mean that the device must have a specific orientation.

[0053] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below with reference to specific illustrations.

[0054] Unless otherwise defined, 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 pertains. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of the invention.

[0055] like Figure 1 As shown, the method for extracting fault characteristic parameters of a reciprocating piston pump under information-based conditions according to the present invention includes the following steps:

[0056] Step 1: When a multi-cylinder reciprocating plunger pump is working, it relies on the reciprocating motion of the piston or plunger inside the cylinder to change the volume of the working chamber, thereby transporting liquid. During the reciprocating motion of the piston, the cylinder of the plunger pump will experience periodic impacts. When the multi-cylinder reciprocating plunger pump is working, a vibration sensor is placed in the axial direction on the pump valve end face of one of the plunger pumps to collect the cylinder vibration signal of the multi-cylinder reciprocating plunger pump during the time period from 0 to T, which is recorded as Y.

[0057] Step 2: Decompose the cylinder vibration signal Y using the Local Means Decomposition (LMD) algorithm to obtain multiple components P. Calculate the Comprehensive Evaluation Index (CEI) for each component and construct the comprehensive evaluation index vector CEI'. Select Q fusion indices with values ​​greater than the mean of the elements in CEI' to obtain a weight vector ω. Based on the weight vector ω, reconstruct the components P using weighted averages to obtain the reconstructed signal X of the piston pump cylinder vibration. The specific steps are as follows:

[0058] Step 2.1: For all local extreme points n of Y i Perform calibration and calculate n for two adjacent extreme points. i and n i+1 Local mean m i and envelope estimate a i ,have to:

[0059]

[0060]

[0061] Step 2.2: Connect adjacent m i and a i Connect the segments with line segments, and smooth the extended polyline using the moving average method. Let the resulting local mean function be m. 11 (t), and let the envelope estimation function be a. 11 (t). The formulas for separating the smoothed local mean function and envelope estimation function from the signal Y are as follows:

[0062] h 11 (t)=Y(t)-m 11 (t) (3)

[0063]

[0064] Step 2.3: Use equation (3) to separate the smoothed local mean function m from the signal Y. 11 (t), and estimate the function a through the envelope. 11 (t) The demodulation result yields the frequency modulated signal s. 11 (t).

[0065] Step 2.4: Determine the frequency modulation signal s 11 The signal shape of (t). When s 11 When (t) exhibits a non-pure frequency modulation signal morphology, its envelope estimation function a 12(t)≠1; if the condition for a non-pure frequency modulated signal is met, s 11 (t) will be used as the original data to repeat the above process, specifically to calculate s. 11 The local mean function m of (t) 12 (t), m 12 (t) from s 11 Separate from (t), using a 12 (t) Demodulation h 12 (t); Repeated p times, finally s 1p (t) represents a pure frequency modulated signal, at which point a 1(p+1) (t) = 1.

[0066] Throughout the iteration process, the envelope signal a1(t) is obtained by multiplying the local envelope functions.

[0067]

[0068] Through the envelope signal a1(t) and the pure frequency modulation signal s 1p Multiplying (t) yields the first component P1(t) of Y. Then, further extracting component P1(t) from Y, denoted as u1(t), and using u1(t) as the new original signal, this process is repeated g times until u... g (t) is a monotonic function:

[0069]

[0070] After multiple iterations, Y is finally decomposed into g P components and one residual component u. g The sum of (t):

[0071]

[0072] The selection of components containing the most latent feature information from the g components obtained by decomposing the original signal often becomes a direct factor affecting the diagnostic effect. The decomposed components contain important or even multiple fault component information; selecting components based on a single indicator may lead to inaccurate results, causing diagnostic errors or missed diagnoses. Therefore, realizing the latent feature mining of component signals plays a positive role in fault diagnosis.

[0073] Step 2.5: Based on three evaluation indicators—Pearson correlation coefficient, impulse factor, and information entropy—this invention sets a comprehensive evaluation index (CEI) for component P, and constructs a reconstructed signal based on this index. The specific steps are as follows:

[0074] The Pearson correlation coefficient is an effective measure of the similarity between a component signal and the original signal, and is widely used in signal analysis. However, it is easily affected by external noise, which can influence the selection results. The specific formula is as follows:

[0075]

[0076] Where P(n) and Y(n) are the component signal and the original signal, respectively. The mean of the component signals. is the mean of the original signal, and N is the truncation length of the signal.

[0077] The impulse factor reflects the impact characteristics of a fault signal and is helpful in assessing the severity of the fault. However, its sensitivity for fault diagnosis decreases in the presence of intermittent interference noise. The specific formula for the impulse factor is as follows:

[0078]

[0079] Information entropy describes the uncertainty of a signal. The more fault information a component signal contains, the more ordered the signal appears, and the better its time-frequency distribution clustering. Therefore, the magnitude of information entropy can directly reflect the degree to which a component carries fault information. However, using information entropy as a single indicator makes it difficult to extract weak fault features. The specific formula for information entropy is as follows:

[0080]

[0081] P n It represents the probability of element P(n) in the component signal, and N is the truncation length of the signal.

[0082] A comprehensive evaluation index, CEI, is proposed by integrating Pearson correlation coefficient, impulse factor, and information entropy. j This is used to calculate the degree to which each component carries fault information. The j-th component P... j CEI j The value can be expressed as:

[0083]

[0084] Where j represents the j-th component P j (j=1,2,···,g), C is the Pearson correlation coefficient, I is the impulse factor, H is the information entropy, γ1, γ2, and γ3 are their respective influencing factors, representing the degree of participation of the corresponding indicators in the process of evaluating fault information. The influencing factors can be determined according to the actual situation of the research object. In this invention, the three indicators are equally important to the evaluation results, and γ1=γ2=γ3=1 is set.

[0085] Step 2.6: Based on this, calculate each component P.j CEI j Considering the relative importance of each component signal, the membership degree of the component signals is set based on the soft thresholding concept. The comprehensive evaluation index of each component signal is normalized to construct the index vector CHI'=(CEI′1,CEI′2,...,CEI′ g Calculate the mean and select Q fusion indicators that are greater than the mean to obtain the weight vector ω = (ω1, ω2, ..., ω...). Q Meanwhile, the corresponding component signals can be found. The residual components reflect slowly changing quantities in the signal, so they are not considered. The reconstructed signal X is then represented as follows:

[0086]

[0087] Step 3: Based on the working principle of a reciprocating piston pump, the cylinder vibration signal is an approximately periodic signal. Therefore, this invention requires the use of a periodic signal time-frequency analysis method, including Fourier series expansion, Fourier transform, and short-time Fourier transform. The short-time Fourier transform not only converts the signal from the time domain to the frequency domain but also allows for signal truncation through a sliding window function and successive transformation operations to obtain the spectrum. Therefore, this invention uses the short-time Fourier transform to construct the time-frequency matrix. The specific steps are as follows:

[0088] Time-frequency analysis of X is performed based on the short-time Fourier transform. The formula for the Fourier transform is as follows:

[0089] F(ω)=∫X(t)ω(t-τ)e -2jtft dt (13)

[0090] Where F(ω) is the spectrum of X at frequency ω, X(t) is the time-domain representation of the cylinder vibration signal, and ω(t-τ) represents a window function centered at τ. -2jπft It is a complex exponential function, where j is the imaginary unit. The time-frequency information of the reconstructed signal X can be obtained through the short-time Fourier transform. Based on this, the amplitude of X(t) at each time and each frequency is calculated, and all amplitudes are used as elements to construct the time-frequency matrix S of the reconstructed signal X. M×N Where M is S M×N The time dimension, N is the time dimension of S. M×N The frequency dimension.

[0091] Step 4: Based on matrix factorization theory, analyze the time-frequency matrix S. M×N Singular value decomposition is performed, and the formula for singular value decomposition is as follows:

[0092] S M×N =U∑V (14)

[0093] Where U is an M-dimensional left singular matrix, V is an N-dimensional left singular matrix, and ∑ is an M×N diagonal matrix with non-zero elements σ1, σ2, ..., σ on the diagonal. k It is S M×N The singular values ​​of S are given by K, where K is the number of singular values. Based on the number of singular values ​​K, for S... M×N Predicting the optimal dimension k in nonnegative matrix factorization. Establishing the objective function:

[0094]

[0095] h(k) is the ratio of the standard deviations of singular values ​​within a window transformed by the value of k, reflecting the smoothness of the signal. The smaller h(k) is, the smoother the signal. Therefore, the choice of the value of k satisfies the following formula:

[0096] k=arg min h(k),2,...,K-2) (16)

[0097] Step 5: After obtaining the optimal decomposition dimension k, the time-frequency matrix S of the cylinder vibration signal X is... M×N Perform nonnegative matrix decomposition. S M×N The nonnegative matrix factorization expression is as follows:

[0098] S M×N ≈W M×k H k×N (17)

[0099] Among them, the left matrix W M×k Defined as a basis matrix, with the right matrix H k×N Defined as a coefficient matrix. In the time-frequency matrix S... M×N In the original time-frequency matrix, N M-dimensional column vectors represent the potential signal features at corresponding frequencies. However, the feature data is too numerous, making it difficult to distinguish between the impact frequency features and fault features in a reciprocating piston pump. After decomposition, the time-frequency information in the original time-frequency matrix is ​​sparsified to filter out redundant time-frequency information. In the basis matrix W... M×k The matrix contains k M-dimensional column vectors, which include information on the piston pump's impact frequency, fault characteristics, and residual components. Macroscopically, it exhibits sparsity, a property that makes fault characteristics more prominent. The coefficient matrix H... k×N It has weighting implications and is represented by the time-frequency matrix S. M×N In the basis matrix W M×k The projection of W in the spatial direction M×k With H k×N By performing a product operation, the time-frequency matrix S can be approximately reconstructed. M×N .

[0100] like Figure 2 As shown in step 5.1, in the initial stage of nonnegative matrix decomposition, the basis matrix W is...M×k And coefficient matrix H k×N The elements in the matrix are initialized under non-negativity conditions. To improve the accuracy and speed of non-negative matrix decomposition, it is necessary to control the original time-frequency matrix S. M×N With basis matrix W M×k coefficient matrix H k×N The error between them needs to be addressed by selecting an appropriate objective function. Commonly used objective functions for nonnegative matrix factorization include loss models based on Euclidean distance and loss models based on KL divergence.

[0101] Step 5.2: This invention uses a loss model based on Euclidean distance to analyze the time-frequency matrix S. M×N With basis matrix W M×k coefficient matrix H k×N To control the error between them, the Euclidean distance D ED The calculation formula is as follows:

[0102]

[0103] Step 5.3: According to the gradient descent algorithm, the basis matrix W... M×k And coefficient matrix H k×N The update rules are as follows:

[0104]

[0105] Where β is the balancing parameter, and to ensure the convergence of the algorithm, this invention takes β as 1 / (k×N). ε is a parameter set to prevent the denominator term from being zero in the update formula, relative to the basis matrix W. M×k And coefficient matrix H k×N The element ε is a very small positive number.

[0106] Step 5.4, when the Euclidean distance D ED When the basis matrix W reaches a stable value or tends to converge, M×k And coefficient matrix H k×N Updates stopped;

[0107] Step 5.5: Obtain the final basis matrix W M×k And coefficient matrix H k×N .

[0108] Step 6, Basis matrix W M×k The column vectors α1, α2, ..., α3 in the array are... k These represent different characteristic information during the operation of the plunger pump. Based on the working principle and signal characteristics of the plunger pump, the column vectors α1, α2, ..., α... k It contains one impact frequency feature and zero or several fault feature information. This invention is based on the optimal decomposition dimension k and cosine similarity theory for the basis matrix W.M×k The column vectors α1, α2, ..., α3 in the array are... k Classification is performed to obtain column vectors that characterize the fault features of reciprocating piston pumps. The specific steps are as follows:

[0109] Based on column vector α i The present invention proposes a column vector α based on the average cosine similarity between (i = 1, 2, ..., k) and the other k-1 column vectors. i similarity coefficient Cos i Cosine similarity is used to measure the similarity between each column vector and other column vectors. i The calculation formula is as follows:

[0110]

[0111] Where, α i ·α j It is the dot product of two vectors, ||α i ||×||α j || is the product of the magnitudes of two vectors. Based on this, we obtain the basis matrix W. M×k Each column vector α1, α2, ..., α k The corresponding similarity coefficients Cos1, Cos2, ..., Cos k .

[0112] Furthermore, based on the value of k and the similarity coefficient of each column vector, the column vector α is... i There are two possible classification methods:

[0113] 1. When k = 2, it represents the basis matrix W. M×k The information contains only two types of time-frequency information. Since piston impact is inevitable during the operation of a reciprocating piston pump, the time-frequency information of the two types of information is the characteristic information of piston impact and the residual component information in non-negative matrix decomposition. That is, when k=2, the reciprocating piston pump is operating normally and there is no fault.

[0114] 2. When k > 2, it represents the basis matrix W. M×k It contains time-frequency information with three or more features, so the basis matrix W M×k The column vectors in the matrix contain time-frequency information representing fault characteristics. Based on the magnitude of the similarity coefficient, the column vector corresponding to the largest similarity coefficient represents the characteristic information of piston impact; the column vector corresponding to the smallest similarity coefficient represents the residual component information in the non-negative matrix decomposition; the column vectors corresponding to similarity coefficients other than the largest and smallest similarity coefficients represent the time-frequency information of fault characteristics, and these column vectors are defined as fault feature vectors.

[0115] Based on this, the basis matrix W M×kOnly the fault feature vector is retained. The column vectors that originally represented the normal operation information of the plunger pump and the column vectors corresponding to the residual component information in the non-negative matrix decomposition are replaced with zero column vectors to obtain the fault basis matrix V. M×k .

[0116] Step 7: Convert the fault basis matrix V M×k With coefficient matrix H k×N Perform matrix multiplication to obtain the fault characteristic matrix Z of the reciprocating piston pump. M×N Based on time-frequency analysis theory, the fault characteristic matrix Z is analyzed. M×N An inverse transform is performed to obtain a time-domain signal Y' that characterizes the fault features of the reciprocating piston pump. Finally, a suitable mother wavelet is selected to perform three-level wavelet packet decomposition on the time-domain signal Y', and the energy values ​​of each sub-band are used as fault feature parameters of the reciprocating piston pump for input to the subsequent fault diagnosis model.

[0117] The technology of this invention was tested in the field at the Hua 24 transfer station of Jiangsu Oilfield and achieved the expected results.

[0118] The above description is merely a preferred embodiment of the present invention, showing and describing the basic principles, main features, and advantages of the present invention. It is not intended to limit the scope of patent protection of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. In addition to the above embodiments, the present invention may have other implementations without departing from the spirit and scope of the invention. Various changes and modifications to the present invention are possible, and all technical solutions formed by equivalent substitutions or equivalent transformations fall within the scope of protection claimed by the present invention. The scope of protection of the present invention is defined by the appended claims and their equivalents. Technical features not described in the present invention can be implemented by or using existing technology, and will not be elaborated here.

Claims

1. A method for extracting fault characteristic parameters of a reciprocating plunger pump under information-based conditions, characterized in that, The steps are as follows: Step 1: Collect the cylinder vibration signal Y of the reciprocating piston pump; Step 2: Decompose the cylinder vibration signal Y based on the LMD algorithm to obtain multiple components P, calculate the comprehensive evaluation index CEI of each component P, and perform weighted reconstruction to obtain the reconstructed signal X. Step 3: Perform time-frequency analysis on the reconstructed signal X, calculate the amplitude of the reconstructed signal X at each time and at each frequency, and construct the time-frequency matrix S of the reconstructed signal X using all amplitudes as elements. M×N ; Step 4: Based on matrix factorization theory, analyze the time-frequency matrix S. M×N Singular value decomposition is performed, and the number of characteristic frequencies of the signal is predicted by analyzing the singular values ​​of the matrix, thus determining S. M×N The optimal decomposition dimension k value in nonnegative matrix factorization; Step 5: Apply the loss model based on the nonnegative matrix factorization algorithm and Euclidean distance to the time-frequency matrix S. M×N Decomposition is performed to obtain the basis matrix W, which can characterize different features of the reciprocating piston pump. M×k And coefficient matrix H k×N ; Step 6: Based on the optimal decomposition dimension k and cosine similarity theory, adjust the basis matrix W. M×k The column vectors α1, α2, ..., α3 in the array are... k Classify; if k > 2, it indicates that the basis matrix W M×k Fault information exists; construct the fault basis matrix V. M×k ; Step 7: Based on time-frequency analysis theory, the fault basis matrix V M×k With coefficient matrix H k×N Perform matrix multiplication to obtain the fault characteristic matrix Z of the reciprocating piston pump. M×N .

2. The method for extracting fault characteristic parameters of a reciprocating plunger pump under information-based conditions according to claim 1, characterized in that, In step 2: After calculating the comprehensive evaluation index CEI for each component P, a comprehensive evaluation index vector CEI' is constructed; Q fusion indices with values ​​greater than the mean of the elements of CEI' are selected to obtain a weight vector ω, and the component P is reconstructed by weighting based on the weight vector ω to obtain the reconstructed signal X of the piston pump cylinder vibration.

3. The method for extracting fault characteristic parameters of a reciprocating plunger pump under information-based conditions according to claim 1, characterized in that, In step 6: If k = 2, it indicates that the basis matrix W M×k If the matrix contains only the characteristic information of piston impact and the residual component information of nonnegative matrix decomposition, it is considered to be fault-free; if k > 2, the column vectors corresponding to the fault information are kept unchanged, and the column vectors corresponding to the characteristic information of piston impact and the residual component information of nonnegative matrix decomposition are all replaced with zero vectors to construct the fault basis matrix V. M×k .

4. The method for extracting fault characteristic parameters of a reciprocating plunger pump under information-based conditions according to claim 1, characterized in that, Based on time-frequency analysis theory, the fault feature matrix Z M×N An inverse transform is performed to obtain a time-domain signal Y' that can characterize the fault features of the reciprocating piston pump. A suitable mother wavelet is selected to perform three-level wavelet packet decomposition on the time-domain signal Y', and the energy values ​​of each sub-band are used as fault feature parameters of the reciprocating piston pump for input to the subsequent fault diagnosis model.

5. The method for extracting fault characteristic parameters of a reciprocating plunger pump under information-based conditions according to claim 2, characterized in that, Step 2 specifically includes the following sub-steps: Step 2.1: For all local extreme points n of Y i Perform calibration and calculate n for two adjacent extreme points. i and n i+1 Local mean m i and envelope estimate a i ; Step 2.2: Connect adjacent m i and a i Connect the segments with line segments, and smooth the extended polyline using the moving average method. Let the resulting local mean function be m. 11 (t), and let the envelope estimation function be a. 11 (t); Step 2.3: Use the smoothed local mean function m extracted from the signal Y. 11 (t), and estimate the function a through the envelope. 11 (t) The demodulation result yields the frequency modulated signal s. 11 (t); Step 2.4: Determine the frequency modulation signal s 11 The signal morphology of (t); when s 11 When (t) exhibits a non-pure frequency modulation signal morphology, its envelope estimation function a 12 (t)≠1; if the condition for a non-pure frequency modulated signal is met, s 11 (t) will be used as the original data and the above process will be repeated p times, eventually s 1p (t) represents a pure frequency modulated signal, at which point a 1(p+1) (t) = 1; Step 2.5: Based on three evaluation indicators—Pearson correlation coefficient, impulse factor, and information entropy—this invention sets a comprehensive evaluation index (CEI) for component P. Step 2.6: Calculate each component P j CEI j The comprehensive evaluation index of each component signal is normalized to construct an index vector. The mean of the index is calculated, and Q fusion indexes with values ​​greater than the mean are selected to obtain the weight vector ω. At the same time, the corresponding component signals can be found. The residual components reflect the slowly changing quantities in the signal, so they are not considered. The reconstructed signal X is represented as follows:

6. The method for extracting fault characteristic parameters of a reciprocating plunger pump under information-based conditions according to claim 2, characterized in that, In step 3, the time-frequency matrix S is constructed using the short-time Fourier transform. M×N .

7. The method for extracting fault characteristic parameters of a reciprocating plunger pump under information-based conditions according to claim 1, characterized in that, In step 4, based on matrix factorization theory, the time-frequency matrix S is... M×N Singular value decomposition is performed, and the formula for singular value decomposition is as follows: S M×N =U∑V Where U is an M-dimensional left singular matrix, V is an N-dimensional left singular matrix, and ∑ is an M×N diagonal matrix with non-zero elements σ1, σ2, ..., σ on the diagonal. k It is S M×N The singular values ​​of S, where K is the number of singular values; based on the number of singular values ​​K, for S M×N Predicting the optimal decomposition dimension k in nonnegative matrix factorization; establishing the objective function: h(k) is the ratio of the standard deviations of singular values ​​within a window transformed by the value of k, reflecting the smoothness of the signal; the smaller h(k) is, the smoother the signal. Therefore, the choice of the value of k satisfies the following formula: k=arg minh(k),(k=2,...,K-2).

8. The method for extracting fault characteristic parameters of a reciprocating plunger pump under information-based conditions according to claim 1, characterized in that, In step 5, after obtaining the optimal decomposition dimension k, the time-frequency matrix S of the cylinder vibration signal X is... M×N Perform nonnegative matrix decomposition; S M×N The nonnegative matrix factorization expression is as follows: S M×N ≈Wx ×k H k×N Among them, the left matrix W M×k Defined as a basis matrix, with the right matrix H k×N Defined as a coefficient matrix; in the time-frequency matrix S M×N In this context, N M-dimensional column vectors represent the potential features of the signal at the corresponding frequencies.

9. The method for extracting fault characteristic parameters of a reciprocating plunger pump under information-based conditions as described in claim 8, characterized in that, Step 5 specifically includes the following sub-steps: Step 5.1: Convert the basis matrix W M×k And coefficient matrix H k×N The elements in the array are initialized under the condition of nonnegation; Step 5.2, Calculate S M×N With W M×k H k×N Euclidean distance D ED ; Step 5.3: According to the gradient descent algorithm, for W... M×k and H k×N Update; Step 5.4, when the Euclidean distance D ED When the basis matrix W reaches a stable value or tends to converge, M×k And coefficient matrix H k×N Updates stopped; Step 5.5: Obtain the final basis matrix W M×k And coefficient matrix H k×N .

Images