CN108804740A - Long distance pipeline pressure monitoring method based on integrated improvement ICA-KRR algorithms - Google Patents

Abstract

The invention discloses a long-distance pipeline pressure monitoring method based on an integrated improved ICA-KRR algorithm, comprising the following steps: 1) constructing a long-distance pipeline pressure monitoring data matrix; 2) calculating the variable P∈R ^m×m ; 3) extracting Component matrix T=P ^T X; 4) Whiten the extracted component matrix T to obtain the whitened result; 5) Calculate the matrix S=C ^T Z; 6) Calculate the matrix C _n ; 7) Calculate the separation matrix W∈R ^{d × m} and mixing matrix A ∈ R ^{m × d} ; 8) get the source signal of the independent component, the independent relationship between the source signal of the independent component is reflected by non-Gaussianity, and the non-Gaussianity is quantified by the negative entropy function, Negative entropy function can choose three kinds of non-quadratic functions; 9) construct three kinds of component importance evaluation criteria; 10) form a double-layer comprehensive learning strategy by mixing; 11) form 9 component selection models; 12) get weight coefficient w; 13) Obtain the regressed fault signal data y; 14) Calculate the leakage location d, this method can realize the real-time monitoring and accurate positioning of the leakage location on the long-distance pipeline.

Description

A pressure monitoring method for long-distance pipelines based on integrated improved ICA-KRR algorithm

technical field

The invention belongs to the technical field of safety detection of long-distance oil and gas pipelines, and relates to a pressure monitoring method of long-distance pipelines based on an integrated improved ICA-KRR algorithm.

Background technique

In recent years, with the increasing service life of oil and gas pipelines, the frequency of pipeline leakage accidents caused by various non-human factors has been increasing, which has brought serious economic losses to enterprises. Therefore, real-time monitoring of pipeline pressure, precise location of leakage faults and timely early warning have important research significance.

Oil and gas pipeline leakage detection technology is an important means to ensure the safe production of pipelines. With the rapid development of information technology and modern control theory, leak detection methods have been widely used due to their high efficiency and flexible application methods, and have become the main means of continuous monitoring of pipeline leakage. Due to the suddenness of pipeline leakage, the pressure monitoring data is complex and redundant. Therefore, in order to monitor pipeline pressure and give early warning of leakage failures, it is necessary to identify effective information and characteristic rules in pipeline pressure monitoring data, and then establish appropriate methods and models to analyze pressure data in order to realize fault detection.

In this regard, the pressure monitoring and risk warning work of long-distance oil and gas pipelines is moving towards a quantitative and proactive strategy. At present, scholars at home and abroad have done a lot of research on pipeline leakage faults, and have put forward unique research methods for pipeline pressure monitoring and fault leakage location. Wang Mingda et al. used the method of independent component analysis combined with support vector machine to denoise and separate the pipeline pressure signal to realize leak detection. Zhang Yu et al. used dynamic pressure transmitter to measure the dynamic change of pipeline pressure and empirical mode decomposition to analyze the pressure signal respectively. , but the data analysis of fault pressure signals cannot achieve accurate positioning. Lin Weiguo et al. used wavelet denoising to identify and extract abnormal signals, and realized the discrimination between interference signals and leakage signals. However, the effectiveness of multi-point pressure monitoring schemes for pipelines is not high. , Zhong Lusheng et al. used independent component analysis and principal component analysis to separate signals and combined with Lasso regression method to realize the location and selection of the main abnormal variables when a fault occurs, and detect and locate the abnormal variables that cause faults step by step. However, Lasso regression is not suitable for medium-sized The processing speed effect of the data set is not ideal. Based on the independent component analysis model, ChudongTong expanded the double-layer Bayesian reference and strengthened the selection of independent components. To sum up, the existing research methods have different degrees of limitations in pipeline pressure monitoring and leakage fault location, and the accuracy of location does not match the actual situation ideally.

Therefore, based on the above analysis, the author proposes a comprehensive improved independent component analysis algorithm (Ensemble Modified Independent Component Analysis, EMICA) combined with Kernel Ridge Regression algorithm (Kernel Ridge Regression, KRR) integrated model of pipeline pressure monitoring and real-time leak location. Use normal offline data to train the EMICA model, construct statistics T ² and Q constraints, and improve model efficiency. The KRR algorithm performs regression analysis on the data before and after the fault, and obtains the change amplitude of the regression coefficient before and after the leakage, based on which the leakage location and diagnosis are realized. Finally, the numerical simulation experiment of TE (TennesseeEastman) process is carried out, and compared with the existing leakage diagnosis method, the performance of the proposed method is verified.

The above methods have achieved certain results through different pressure monitoring methods, but there are also problems in these applications that the selection of the original sequence is not screened, and the analysis of its applicability is not enough, resulting in weakened discrimination ability and reduced accuracy.

Contents of the invention

The purpose of the present invention is to overcome the above-mentioned shortcomings of the prior art, and provide a long-distance pipeline pressure monitoring method based on the integrated improved ICA-KRR algorithm, which can realize real-time monitoring and accurate positioning of the leakage position on the long-distance pipeline.

In order to achieve the above object, the long-distance pipeline pressure monitoring method based on the integrated improved ICA-KRR algorithm of the present invention comprises the following steps:

1) Collect the pressure monitoring data of several long-distance pipelines, and then construct the long-distance pipeline pressure monitoring data matrix X∈R ^m×n through the acquired pressure monitoring data of long-distance pipelines, where n is the obtained long-distance pipeline pressure monitoring data The number of , m is the number of variables;

2) Calculate the covariance matrix XX ^T /(n-1)= ^PΛPT of the long-distance pipeline pressure monitoring data, where Λ is the eigenvector of the long-distance pipeline pressure monitoring data matrix X, Λ=diag{λ ₁ ,λ ₂ ,...,λ _m }, and then calculate the variable P∈R ^m×m according to the covariance matrix XX ^T /(n-1)=PΛP ^T and the eigenvector Λ of the long-distance pipeline pressure monitoring data;

3) Extract the component matrix T=P ^T X from the long-distance pipeline pressure monitoring data matrix X according to the variable P;

4) Calculate the variable Q=Λ- ^1/2 P ^T , and then whiten the extracted component matrix T to obtain the whitened result _x =Λ- ^1/2 T=Λ- ^1/2 P ^T X=QX;

5) Calculate the matrix C∈R ^m×n according to the C ^T C = D condition, where D = diag{λ ₁ ,λ ₂ ,…,λ _d } are the eigenvectors of known d independent signals, and then according to the matrix C Calculation matrix S=C ^T Z;

6) Calculate matrix C _n according to C _n ^T =D ^-1/2 C ^T and C _n ^T C _n =I;

7) Calculate separation matrix W∈R ^d×m and mixing matrix A∈R ^m×d , where,

8) Separate the monitoring data matrix X∈R ^m×n through the separation matrix W to obtain the source signals of the independent components, and the independent relationship between the source signals of the independent components is reflected by the non-Gaussian property, and the non-Gaussian property is represented by the negative Entropy function J(W ^T X)=[E{G(W ^T X)}-E{G( _v )}] ² for quantization, where v is a Gaussian variable with zero mean and unit variance, and the negative entropy function can be selected Three kinds of non-quadratic functions, the three kinds of non-quadratic functions are respectively G ₂ (u)=exp(-a ₂ u ² /2) and G ₃ =u ⁴ , where 1≤a ₁ ≤2, a ₂ ≈1, cosh() is a hyperbolic cosine function;

9) construct three kinds of component importance evaluation standards, wherein, the three kinds of component importance evaluation standards are respectively EMICA-CPV, and EMICA-nG;

10) According to three non-quadratic functions and three component importance evaluation standards, a two-layer comprehensive learning strategy is formed, namely Among them, i∈{1,2,3}, i∈{1,2,3} respectively represent three kinds of non-quadratic functions, E∈R ^m×n is the initial residual matrix, is the element of row i and column j in mixing matrix A, the i and j mixing matrices, W _i ^j is the element of row i and column j in separation matrix W;

11) Various non-quadratic functions select the above three component importance evaluation criteria respectively to form 9 component selection models Among them, i∈{1,2,3}, i∈{1,2,3}. Then use Bayesian inference to combine multiple statistics in a probabilistic manner to form a unique index;

12) In the i-th component selection model, calculate the detection probability of the monitoring statistic T2 ^;

13) In the i-th component selection model, calculate the detection probability of the statistic Q;

14) Separate and obtain the abnormal component s _i ∈ S, and record the separated abnormal component _si as y _i , and then calculate the minimum loss function Get the weight coefficient w, where _xi is the true value of the abnormal component, λ is the regularization parameter determined, and ||.|| _F represents the Frobenius specification;

15) Use the kernel transformation method to expand the explicit regression function, and then use the Gaussian kernel function and polynomial function Obtain the regressed fault signal data y, where σ is the kernel parameter of the Gaussian kernel function set, u and d are the kernel parameters of the polynomial kernel function set respectively;

16) Calculate the transmission velocity v of the pressure wave through the regressed fault signal data y, and then use the transmission velocity v of the pressure wave to calculate the leakage position d.

Detection probability of monitoring statistic ^T2 for:

Let N and F denote normal operating conditions and abnormal operating conditions respectively, and are the confidence degrees α and 1-α for calculating control limits respectively, the probabilities of normal operating conditions N and abnormal operating conditions F are expressed as and where,

The selection model of each component is integrated into:

leak location L is the distance between the two pressure monitoring points, Δt is the time difference of data received by the sensors at the first and last ends, and u is the fluid velocity in the pipe.

The present invention has the following beneficial effects:

The long-distance pipeline pressure monitoring method based on the integrated improved ICA-KRR algorithm described in the present invention constructs a long-distance pipeline pressure monitoring data matrix through the collected pressure monitoring data during specific operations, and then from the long-distance pipeline pressure monitoring data matrix Extract the component matrix T, then whiten the component matrix, and then separate the abnormal components, and then combine the three non-quadratic functions with the three component importance evaluation criteria, and based on Bayesian reasoning and explicit regression function Regress and separate the pressure fault components to determine the fault signal data, and then calculate the leak position according to the fault signal data, so as to realize real-time monitoring and accurate positioning of the leak position on the long-distance pipeline, so as to remind the staff to repair the pipeline in time to avoid unnecessary Loss.

Description of drawings

Fig. 1 is a flow chart of the present invention;

Figure 2 is a schematic diagram of source variables in the numerical simulation process;

Fig. 3 is the detailed schematic diagram of independent component analysis monitoring based on comprehensive improvement;

Fig. 4 is a broken line schematic diagram of the average fault detection rate of different numbers of dominant ICs.

Detailed ways

The present invention is described in further detail below in conjunction with accompanying drawing:

The long-distance pipeline pressure monitoring method based on the integrated improved ICA-KRR algorithm of the present invention comprises the following steps:

1) Collect the pressure monitoring data of several long-distance pipelines, and then construct the long-distance pipeline pressure monitoring data matrix X∈R ^m×n through the acquired pressure monitoring data of long-distance pipelines, where n is the obtained long-distance pipeline pressure monitoring data The number of , m is the number of variables;

2) Calculate the covariance matrix XX ^T /(n-1)= ^PΛPT of the long-distance pipeline pressure monitoring data, where Λ is the eigenvector of the long-distance pipeline pressure monitoring data matrix X, Λ=diag{λ ₁ ,λ ₂ ,...,λ _m }, and then calculate the variable P∈R ^m×m according to the covariance matrix XX ^T /(n-1)=PΛP ^T and the eigenvector Λ of the long-distance pipeline pressure monitoring data;

3) Extract the component matrix T=P ^T X from the long-distance pipeline pressure monitoring data matrix X according to the variable P;

4) Calculate the variable Q=Λ- ^1/2 P ^T , and then whiten the extracted component matrix T to obtain the whitened result Z=Λ- ^1/2 T=Λ- ^1/2 P ^T X=QX;

5) Calculate the matrix C∈R ^m×n according to the C ^T C = D condition, where D = diag{λ ₁ ,λ ₂ ,…,λ _d } are the eigenvectors of known d independent signals, and then according to the matrix C Calculation matrix S=C ^T _Z ;

6) Calculate matrix C _n according to C _n ^T _=D ^-1/2 C ^T and C _n ^T C _n =I;

7) Calculate separation matrix W∈R ^d×m and mixing matrix A∈R ^m×d , where,

8) Separate the monitoring data matrix X∈R ^m×n through the separation matrix W to obtain the source signals of the independent components, and the independent relationship between the source signals of the independent components is reflected by the non-Gaussian property, and the non-Gaussian property is represented by the negative Entropy function J(W ^T X)=[E{G(W ^T X)}-E{G(v)}] ² for quantization, where v is a Gaussian variable with zero mean and unit variance, and the negative entropy function can be selected Three kinds of non-quadratic functions, the three kinds of non-quadratic functions are respectively G ₂ (u)=exp(-a ₂ u ² /2) and G ₃ =u ⁴ , where 1≤a ₁ ≤2, a ₂ ≈1, cosh() is a hyperbolic cosine function;

9) Construct three kinds of component importance evaluation criteria, wherein, the three kinds of component importance evaluation criteria are EMICA-CPV, EMICA-L2 and _EMICA -nG respectively;

10) According to three non-quadratic functions and three component importance evaluation standards, a two-layer comprehensive learning strategy is formed, namely Among them, i∈{1,2,3}, i∈{1,2,3} respectively represent three kinds of non-quadratic functions, E∈R ^m×n is the initial residual matrix, is the element of row i and column j in mixing matrix A, the i and j mixing matrices, W _i ^j is the element of row i and column j in separation matrix W;

11) Various non-quadratic functions select the above three component importance evaluation criteria respectively to form 9 component selection models Among them, i∈{1,2,3}, i∈{1,2,3}. Then use Bayesian inference to combine multiple statistics in a probabilistic manner to form a unique index;

12) In the i-th component selection model, calculate the detection probability of the monitoring statistic T2 ^;

13) In the i-th component selection model, calculate the detection probability of the statistic Q;

14) Separate and obtain the abnormal component s _i ∈ S, and record the separated abnormal component _si as _yi , and then calculate the minimum loss function Get the weight coefficient w, where _xi is the true value of the abnormal component, λ is the regularization parameter determined, and ||.|| _F represents the Frobenius specification;

15) Use the kernel transformation method to expand the explicit regression function, and then use the Gaussian kernel function and polynomial function and k(y,y _i )=(uyy _i +1) ^d , get the regressed fault signal data y, where σ is the kernel parameter of the Gaussian kernel function set, u and d are the polynomial kernel function set respectively The kernel parameters;

16) Calculate the transmission velocity v of the pressure wave through the regressed fault signal data y, and then use the transmission velocity v of the pressure wave to calculate the leakage position d.

Detection probability of monitoring statistic ^T2 for:

Let N and F denote normal operating conditions and abnormal operating conditions respectively, and are the confidence degrees a and 1 _- α for calculating control limits respectively, and the probabilities of normal operating conditions N and abnormal operating conditions F are expressed as and where,

The selection model of each component is integrated into:

leak location L is the distance between the two pressure monitoring points, Δt is the time difference of data received by the sensors at the first and last ends, and u is the fluid velocity in the pipe.

Simulation

This simulation experiment is carried out on the TE platform for numerical simulation experiment analysis. Firstly, the MICA model is used to realize the fault detection of the TE process; then, the constructed KRR regression model is used to further select and separate the fault variables of the fault data, and finally realize fault location and diagnosis.

The physical model of TE process consists of reactor, condenser, gas-liquid separator, cycle compressor and stripper. The TE process consists of 41 measured variables (including 22 continuous variables and 19 component-valued variables) and 12 manipulated variables. This simulation experiment selects 30 variables, including 10 operating variables and 20 measurement variables, of which 20 measurement variables are shown in Table 1. There are 21 programmable abnormal conditions in the TE process; compared with methods using different non-quadratic functions, the present invention has a higher degree of recognition of dominant ICs.

Table 1

In the actual operation of long-distance pipelines, most faults will cause two or more variable abnormalities. In view of this, more targeted inspections can be carried out according to the main abnormal variables; The average fault detection rate shows that the present invention has concise and intuitive variable selection ability.

The content that is not described in detail in the description of the present invention belongs to the prior art known to those skilled in the art.

The above embodiments are only used to illustrate the present invention, but not to limit the present invention. Although the relevant embodiments and accompanying drawings of the present invention are disclosed for the purpose of illustration, those skilled in the art can understand that; without departing from the spirit and scope of the present invention and the appended claims, various replacements, changes and modifications are possible. It is possible. Therefore, all equivalent technical solutions also belong to the category of the present invention, and the scope of patent protection of the present invention should be defined by the claims, and should not be limited to the content disclosed by the preferred embodiment and the accompanying drawings.

Claims (4)

1. A long-distance pipeline pressure monitoring method based on an integrated improved ICA-KRR algorithm is characterized by comprising the following steps:

1) collecting pressure monitoring data of a plurality of long-distance pipelines, and establishing a long-distance pipeline pressure monitoring data matrix X epsilon R through the obtained pressure monitoring data of the long-distance pipelines^m×nN is the number of the obtained pressure monitoring data of the long-distance pipeline, and m is a variable number;

2) covariance matrix XX for calculating pressure monitoring data of long-distance pipeline^T/(n-1)＝PΛP^TWherein Λ isCharacteristic vector of long-distance pipeline pressure monitoring data matrix X, Λ ═ diag { lambda₁,λ₂,…,λ_mAnd fourthly, according to a covariance matrix XX of pressure monitoring data of the long-distance pipeline^T/(n-1)＝PΛP^TAnd the eigenvector Lambda calculates the variable P epsilon R^m×m；

3) Extracting a component matrix T (P) from a long-distance pipeline pressure monitoring data matrix X according to the variable P^TX；

4) Calculating variable Q ═ Λ^-1/2P^TWhitening the extracted component matrix T to obtain a whitened result Z ═ Lambda^-1/2T＝Λ^-1/2P^TX＝QX；

5) According to C^TC ═ D conditional calculation matrix C ∈ R^m×nWherein D ═ diag { λ₁,λ₂,…,λ_dD are known eigenvectors of the d independent signals, and then a matrix S is calculated according to the matrix C^TZ；

6) According to C_n ^T＝D^-1/2C^TAnd C_n ^TC_nCalculating the matrix C as I_n；

7) Calculating a separation matrix W ∈ R^d×mAnd the mixing matrix A ∈ R^m×dWherein

8) monitoring data matrix X epsilon R through separation matrix W^m×nSignal separation is performed to obtain source signals of independent components, and the independent relation between the source signals of the independent components is reflected by non-Gaussian property which is formed by a negative entropy function J (W)^TX)＝[E{G(W^TX)}-E{G(v)}]²And quantizing, wherein v is a Gaussian variable with zero mean and unit variance, and three non-quadratic functions can be selected as the negative entropy function, and are respectivelyG₂(u)＝exp(-a₂u²/2) and G₃＝u⁴Wherein 1 is not more than a₁≤2，a₂≈1，cosh()Is a hyperbolic cosine function;

9) constructing three component importance evaluation standards, wherein the three component importance evaluation standards are EMICA-CPV and EMICA-L respectively₂And EMICA-nG;

10) forming a double-layer comprehensive learning strategy according to three non-quadratic functions and three component importance evaluation criteria in a mixed mode, namelyWherein, i belongs to {1,2,3} respectively represents three non-quadratic functions, E belongs to R^m×nFor the initially set residual matrix, the residual matrix is,is the element of ith row and jth column in the mixing matrix A, i, j mixing matrix, W_i ^jIs the element of the ith row and the jth column in the separation matrix W;

11) selecting the above three component importance evaluation criteria by various non-quadratic functions to form 9 component selection modelsWherein i belongs to {1,2,3}, and i belongs to {1,2,3 }. Then, Bayesian reasoning is utilized to form a unique index by combining the multiple statistics in a probability manner;

12) in the ith component selection model, a monitoring statistic T is calculated²The detection probability of (2);

13) calculating the detection probability of the statistic Q in the ith component selection model;

14) separating to obtain abnormal component s_iE.g. S, and separating the abnormal component S_iNotation y_iRecalculating the minimization of loss functionObtaining a weight coefficient w, wherein x_iIs the true value of the abnormal component, lambda is the determined regularization parameter, | | - | luminous flux_FRepresents the Frobenius norm;

15) using a kernel transform approachExpanding the dominant regression function, reusing the Gaussian kernel function and the polynomial functionAnd k (y, y)_i)＝(uyy_i+1)^dObtaining regression fault signal data y, wherein sigma is a kernel parameter of a set Gaussian kernel function, and u and d are kernel parameters of a set polynomial kernel function respectively;

16) and calculating the transmission speed v of the pressure wave through the regressed fault signal data y, and then calculating the leakage position d by using the transmission speed v of the pressure wave.

2. The long-distance pipeline pressure monitoring method based on the integrated improved ICA-KRR algorithm as claimed in claim 1, wherein the monitoring statistic T²Detection probability ofComprises the following steps:

setting N and F as normal operation condition and abnormal operation condition,andconfidence a and 1-a of the calculated control limit, respectively, and probabilities of normal operating condition N and abnormal operating condition F are respectively expressed asAndwherein,

3. the long-distance pipeline pressure monitoring method based on the integrated improved ICA-KRR algorithm as claimed in claim 1, wherein the component selection models are synthesized as follows:

4. the long distance pipeline pressure monitoring method based on the integrated improved ICA-KRR algorithm as claimed in claim 1, wherein the leakage positionL two-section distance of the pressure monitoring point, delta t is the time difference of data receiving of the sensors at the head end and the tail end, and u is the velocity of fluid in the pipe.