CN1904581A - Oil gas water multiphase flow type identification method based on main component analysis and supporting vector machine - Google Patents

Abstract

The invention relates to an oil/gas/water multi-phase flow type identification method based on main constituent analysis and supporting vector machine. The feature is that: taking collection to flow parameter signal of different flow type by using horizontal tubes, pressure sampling ring, pressure sampling tube, pressure difference transmitter, data collection card and computer; taking denoising processing to pressure difference fluctuation signal between the collected different pressure sampling distance; taking empiric model decomposition to the signal to gain the fixed modal function, and forming a matrix X from the modal functions, using the main constituent analysis to gain the eigenvector of the flow type; using the supporting vector machine to accomplish the mapping from feature space to flow type space to realize the final flow type identification. The eigenvector fusing the information from plural sensors and would fully reflect the information of flow type. It has the feature of rapid learning speed and strong classifying ability.

Description

Oil gas water multiphase flow type identification method based on principal component analysis (PCA) and support vector machine

Technical field

The present invention relates to the multiphase flow measurement technical field, relate in particular to a kind of oil gas water multiphase flow type identification method based on principal component analysis (PCA) and support vector machine.

Background technology

Traditional flow type identification method has two big classes: a class is to adopt experimental technique to make flow regime map; Another kind of is the transformation criterion relational expression that the analysis according to the convection transformation mechanism obtains, and utilizes on-the-spot flow parameter to determine flow pattern.Because flow parameter needs the problem that solves often in production reality, therefore traditional method can't be used widely.

The modern surveying method of flow pattern can be divided into direct measurement and indirect measurement method on principle of work.Direct measuring method commonly used has ocular estimate, high-speed photography method etc., and this method has certain subjectivity, may draw different conclusions for a plurality of observers.Indirect measurement method is the fluctuation signals such as pressure, pressure reduction and void fraction that utilize the instrument and equipment measurement of measuring, and by signal being analyzed, extracted feature, discerns flow pattern in conjunction with neural network isotype recognition technology.This method has two key issues: the one, and the extraction of flow pattern proper vector; The 2nd, the choosing of network model, wherein feature extraction is particularly crucial.Existing method mainly is to utilize single pressure differential pressure fluctuation signal to extract the single feature of signal, and the feature that causes extracting can't reflect the information of flow pattern comprehensively, causes the flow pattern discrimination not high.

Summary of the invention

The objective of the invention is to overcome the defective of above-mentioned prior art, propose a kind of flow pattern identification accuracy, reliability height, the oil gas water multiphase flow type identification method based on principal component analysis (PCA) and support vector machine of highly versatile.

The objective of the invention is to be realized by following technical scheme: a kind of oil gas water multiphase flow type identification method based on principal component analysis (PCA) and support vector machine is characterized in that it may further comprise the steps:

(1) the flow parameter signal of different flow patterns is gathered: utilize the pressure spacing to be respectively 5 times, 10 times and 15 times of caliber D; Sample frequency is 256Hz; Sampling number is 1024 points; The differential pressure transmitter of 1 pressure unit and 3 different pressure spacings is gathered in real time to the parameter signal that flows;

(2) remove the noise pre-service: utilize wavelet packet that pressure and pressure-difference fluctuation signal are removed noise processed.

(a) to the pressure-difference fluctuation signal S of 1 pressure surge and 3 different pressure spacings ₁, S ₂, S ₃And S ₄Carry out 4 layers of WAVELET PACKET DECOMPOSITION according to following recursion (1):

$\left\{\begin{array}{c}{u}_{2n}\left(t\right)=\sqrt{2}\underset{k}{\mathrm{\Σ}}h\left(k\right)u_n(2t-k)\\ u_{2n-1}\left(t\right)=\sqrt{2}\underset{k}{\mathrm{\Σ}}g\left(k\right)u_n(2t-k)\end{array}\right.---\left(1\right)$

H in the formula (k) is the Hi-pass filter group; G (k) is the low-pass filter group; u ₀(t)=and  (t), be scaling function; u ₁(t)=and ψ (t), be wavelet function; The yardstick of k for decomposing.H (k) and g (k) satisfy orthogonality relation:

g(k)＝(-1) ^kh(1-k) (2)

(b) each frequency band difference setting threshold to decomposing adopts following floating threshold form

$t_k=\sqrt{21n\left(n\right)}\mathrm{\σ}/\sqrt{n}---\left(3\right)$

Wherein σ is a noise intensity, with median MAD (υ _k)/0.6745 (υ _kBe the WAVELET PACKET DECOMPOSITION coefficient) estimate that n is the length of signal.

(c) with nonlinear function ζ=sgn (υ _k) (| υ _k|-t _k) act on the WAVELET PACKET DECOMPOSITION coefficient υ of signal _k, obtain coefficient υ _k, soon the above coefficient of floating threshold deducts after the threshold value and remains.

(d) by coefficient υ _kRebuild original signal, thereby reach the removal noise, the signal that obtains removing behind the noise is S ₁', S ₂', S ₃' and S ₄'.

(3) to 1 pressure and 3 pressure-difference fluctuation signal S after removing noise ₁', S ₂', S ₃', S ₄' carry out empirical mode decomposition, extract the proper vector of flow pattern.The EMD method is a signal to be carried out tranquilization handle in essence, consequently the fluctuation or the trend of different scale in the signal are decomposed out step by step, produce a series of data sequences with different characteristic yardstick, each sequence is called an intrinsic mode function IMF.The IMF component must satisfy following two conditions: the one, and its limit number and number at zero point identical (or differing at most), the 2nd, envelope is local symmetrical about time shaft up and down for it.

The EMD method step is as follows: suppose that any signal all is made up of different IMF, each IMF can be linear, also can be non-linear, any one signal just can be decomposed into limited IMF sum like this, and then " screening " (Sift) by the following method for IMF:

(a) determine all Local Extremum of signal x (t), all maximum points are coupled together with the cubic spline line form the coenvelope line, all minimum points are coupled together with the cubic spline line form the lower envelope line, these two envelope envelopes all signal datas.

(b) average with two envelopes is designated as μ ₁, obtain

y ₁(t)＝x(t)-μ ₁ (4)

(c) judge y ₁(t) whether be IMF, if y ₁(t) do not satisfy the IMF condition, then with y ₁(t) as raw data, at this moment, note y ₁(t)=c ₁(t), c then ₁(t) be signal x ₁(t) first IMF component, its representation signal x ₁(t) component of highest frequency in;

(d) with c ₁(t) from x ₁(t) separate in, promptly obtain a difference signal r who removes high fdrequency component ₁(t), promptly have

r ₁(t)＝x ₁(t)-c ₁(t) (5)

With r ₁(t) as raw data, repeating step (1), (2) and (3) obtain second IMF component c ₂(t), repeat n time, obtain n IMF component.So just have

$\left.\begin{array}{c}{r}_1-c_2=r_2\\ M\\ r_{n-1}-c_n=r_n\end{array}\right\}---\left(6\right)$

Work as c _n(t) or r _n(t) satisfy given end condition and (make r usually _n(t) become a monotonic quantity) time, loop ends can be obtained by (2) formula and (3) formula:

$x\left(t\right)=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{c}_i\left(t\right)+r_n\left(t\right)---\left(7\right)$

In the formula, r _n(t) be remaining function, the average tendency of representation signal.And each IMF component c ₁(t), c ₂(t), Λ, c _n(t) comprised the signal composition of different frequency section from high to low respectively, the frequency content that each frequency band comprised all is different, and changes with the variation of signal itself.

(e) extraction of proper vector.N IMF component of 1 Pressure Fluctuation Signal and 3 pressure-difference fluctuation signals is designated as c respectively ₁₁(t), c ₁₂(t), Λ, c _1n1(t), c ₂₁(t), c ₂₂(t), Λ, c _2n2(t), c ₃₁(t), c ₃₂(t), Λ, c _3n3(t), c ₄₁(t), c ₄₂(t), Λ, c _4n4(t).These IMF components are formed this matrix of spy X, promptly

$X_{p\×n}=={\left[\begin{array}{cccc}{x}_{11}& x_{12}& \mathrm{\Λ}& x_{1p}\\ x_{21}& x_{22}& \mathrm{\Λ}& x_{2p}\\ M& M& M& M\\ x_{n1}& x_{n1}& \mathrm{\Λ}& x_{\mathrm{np}}\end{array}\right]}^T={\left[\begin{array}{c}{c}_{11}\\ M\\ c_{1n_1}\\ c_{21}\\ M\\ c_{2n_2}\\ c_{31}\\ M\\ c_{3n_3}\\ c_{41}\\ M\\ c_{4n_4}\end{array}\right]}^T---\left(8\right)$

In the formula: X is the eigenmatrix of p * n, and p is a sampling number, n=n ₁+ n ₂+ n ₃+ n ₄Be the number of all IMF components of 4 original signals, x _i(i=1,2, Λ n) is the p dimensional vector.

(f) matrix X is transformed to correlation matrix with following formula, i.e. major component matrix R, and calculate m the eigenwert of correlation matrix R.

1. original variable is carried out standardization

Each index dimension is often different, and during analysis, the different dimensions and the order of magnitude can be drawn new problem, so want standardization, making standardized average of variable is 0, and variance is 1, promptly $x_{\mathrm{ij}}^{*}=\frac{x_{\mathrm{ij}}-{\stackrel{\‾}{x}}_i}{\sqrt{\mathrm{var}\left(x_{\mathrm{ij}}\right)}}$ i＝1，2，Λ，n；j＝1，2，Λ，p (9)

In the formula: x _Ij ^*Be the observed quantity after the standardization, x _iBe mean value,

Be standard deviation.

2. calculate the correlation matrix R of X

After the standardization, relevant battle array is identical with covariance matrix, and the subcomponent of R can be expressed as

$r_{\mathrm{ij}}=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_{\mathrm{ti}}^{*}{x}_{\mathrm{ij}}^{*}---\left(10\right)$

In the formula: x _Ti ^*, x _Ij ^*Be respectively capable i row of t and j column element in the eigenmatrix.

3. ask eigenwert and the proper vector of correlation matrix R

Calculate each eigenwert (λ ₁＞λ ₂＞Λ＞λ _n) and each proper vector (u ₁, u ₂, Λ, u _n).

4. select major component

Calculate the contribution rate and the contribution rate of accumulative total of each major component, the number of major component is decided on particular problem, and it is that 90% a pairing m variable is as major component that the present invention gets contribution rate of accumulative total.

5. count the score

Calculate each principal component scores according to the major component expression formula.In principal component analysis (PCA), ${\mathrm{\μ}}_j={\mathrm{\λ}}_j/\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\λ}}_i$ Be the contribution rate of major component, ratio The percentage of the primary data information (pdi) of k major component representative before the reflection.Order

T＝[μ ₁，μ ₂，Λ，μ _m] (11)

Then T is the proper vector of different flow patterns.

(4) support vector machine classifier.Adopt support vector machine to realize the identification of flow pattern.Oil gas water multiphase flow type has 4 kinds, i.e. stratified flow, intermittent flow, bubble flow and annular flow in the horizontal tube.According to above-mentioned classification situation, adopt the strategy of " one-to-many ", construct 4 two-value sorters, to support vector machine of each flow pattern structure, the output of 4 support vector machine is one 4 dimensional vectors, whether each representation in components sample is to should flow pattern, and the support vector machine that trains can realize the identification of convection.

The advantage that the present invention is based on the oil gas water multiphase flow type identification method of principal component analysis (PCA) and support vector machine is:

1. the signal to 1 pressure and 3 differential pressure transmitters merges, and the feature of extraction can reflect the information of flow pattern comprehensively, therefore, and identification accuracy and reliability height.

2. adopt support vector machine to realize the identification of flow pattern, computing velocity is fast, helps the ONLINE RECOGNITION of flow pattern.

Description of drawings

Fig. 1 is based on the oil gas water multiphase flow type identification method process flow diagram of principal component analysis (PCA) and support vector machine.

Fig. 2 is based on the structural representation of the oil gas water multiphase flow type signal pickup assembly of principal component analysis (PCA) and support vector machine.

Embodiment

Utilize drawings and Examples that the oil gas water multiphase flow type identification method that the present invention is based on principal component analysis (PCA) and support vector machine is described further below.

As shown in Figure 2, oil gas water multiphase flow type signal pickup assembly based on principal component analysis (PCA) and support vector machine, has the horizontal pipeline 16 of polyphasic flow, the pressure ring 1a that on the horizontal pipeline 16 of polyphasic flow, is nested with, 2b, 3c, 4d, 5e, 6f, 7g, pressure ring 1a is connected with pressure pipe 8a, pressure ring 2b, 3c is connected with pressure pipe 8b, pressure ring 4d, 5e is connected with pressure pipe 8c, pressure ring 6f, 7g is connected with pressure pipe 8d, pressure pipe 8a is connected with pressure unit 9a, pressure pipe 8b, 8c, 8d respectively with differential pressure transmitter 10b, 11c, 12d connects, pressure unit 9a, differential pressure transmitter 10b, 11c, 12d all is electrically connected with data collecting card 13, and data collecting card 13 is electrically connected with computing machine 14.The spacing of described pressure ring 2b, 3c is 5 times of caliber D, and the spacing of pressure ring 4d, 5e is 10 times of caliber D, and the spacing of pressure ring 6f, 7g is 15 times of caliber D.Wherein pressure ring 1a is apart from inlet 300mm, pressure ring 1a and the pressure ring 2b 200mm of being separated by, pressure ring 3c and the pressure ring 4d 200mm of being separated by, pressure ring 5e and the pressure ring 6f 200mm of being separated by.Pressure unit 9a, differential pressure transmitter 10b, 11c, 12d model are that outputting standard signal and the model of PD-23 is that the data collecting card 13 of IMP3595 links to each other, data collecting card 13 links to each other with computing machine 14, finish the data processing of pressure difference signal by computing machine 14, and then realize collection the oil gas water multiphase flow type signal.Computing machine 14 also can connect printer 15.According to Automatic Measurement Technique and microcomputer data processing establishment, be the technology that those skilled in the art were familiar with based on the software program of the oil gas water multiphase flow type signal pickup assembly of principal component analysis (PCA) and support vector machine.

As shown in Figure 1, the oil gas water multiphase flow type identification method that the present invention is based on principal component analysis (PCA) and support vector machine comprises the steps:

(1) the flow parameter signal of different flow patterns is gathered: in oil gas water multiphase flow type signal pickup assembly, utilize the pressure spacing to be respectively 5 times, 10 times and 15 times of caliber D based on principal component analysis (PCA) and support vector machine; Sample frequency is 256Hz; Sampling number is 1024 points, and the sampling time is 16s; Differential pressure transmitter 10b, 11c, 12d with 1 pressure unit 9a and 3 different pressure spacings gather in real time to the parameter signal that flows;

(2) use wavelet theory the pressure-difference fluctuation signal of gathering is removed the noise pre-service: (a) to the pressure-difference fluctuation signal S of 1 pressure surge and 3 different pressure spacings ₁, S ₂, S ₃And S ₄Carry out 4 layers of WAVELET PACKET DECOMPOSITION according to following recursion (1):

$\left\{\begin{array}{c}{u}_{2n}\left(t\right)=\sqrt{2}\underset{k}{\mathrm{\Σ}}h\left(k\right)u_n(2t-k)\\ u_{2n-1}\left(t\right)=\sqrt{2}\underset{k}{\mathrm{\Σ}}g\left(k\right)u_n(2t-k)\end{array}\right.---\left(1\right)$

H in the formula (k) is the Hi-pass filter group; G (k) is the low-pass filter group; u ₀(t)=and  (t), be scaling function; u ₁(t)=and ψ (t), be wavelet function; The yardstick of k for decomposing.H (k) and g (k) satisfy orthogonality relation:

g(k)＝(-1) ^kh(1-k) (2)

(b) each frequency band difference setting threshold to decomposing adopts following floating threshold form

$t_k=\sqrt{21n\left(n\right)}\mathrm{\σ}/\sqrt{n}---\left(3\right)$

Wherein σ is a noise intensity, with median MAD (υ _k)/0.6745 (υ _kBe the WAVELET PACKET DECOMPOSITION coefficient) estimate that n is the length of signal.

(c) with nonlinear function ζ=sgn (υ _k) (| υ _k|-t _k) act on the WAVELET PACKET DECOMPOSITION coefficient υ of signal _k, obtain coefficient υ _k, soon the above coefficient of floating threshold deducts after the threshold value and remains.

(d) by coefficient υ _kRebuild original signal, thereby reach the purpose of removing noise.

(3) feature of extraction flow pattern: utilization EMD decomposes as follows to the pressure-difference fluctuation signal x (t) that collects:

(a) determine all Local Extremum of signal x (t), all maximum points are coupled together with the cubic spline line form the coenvelope line, all minimum points are coupled together with the cubic spline line form the lower envelope line, these two envelope envelopes all signal datas.

(b) average with two envelopes is designated as μ ₁, obtain

y ₁(t)＝x(t)-μ ₁ (4)

(c) judge y ₁(t) whether be IMF, if y ₁(t) do not satisfy the IMF condition, then with y ₁(t) as raw data, at this moment, note y ₁(t)=c ₁(t), c then ₁(t) be signal x ₁(t) first IMF component, its representation signal x ₁(t) component of highest frequency in;

(d) with c ₁(t) from x ₁(t) separate in, promptly obtain a difference signal r who removes high fdrequency component ₁(t), promptly have

r ₁(t)＝x ₁(t)-c ₁(t) (5)

With r ₁(t) as raw data, repeating step (1), (2) and (3) obtain second IMF component c ₂(t), repeat n time, obtain n IMF component.So just have

$\left.\begin{array}{c}{r}_1-c_2=r_2\\ M\\ r_{n-1}-c_n=r_n\end{array}\right\}---\left(6\right)$

Work as c _n(t) or r _n(t) satisfy given end condition and (make r usually _n(t) become a monotonic quantity) time, loop ends can be obtained by (2) formula and (3) formula:

$x\left(t\right)=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{c}_i\left(t\right)+r_n\left(t\right)---\left(7\right)$

In the formula, r _n(t) be remaining function, the average tendency of representation signal.And each IMF component c ₁(t), c ₂(t), Λ, c _n(t) comprised the signal composition of different frequency section from high to low respectively, the frequency content that each frequency band comprised all is different, and changes with the variation of signal itself.

(e) extraction of proper vector.N IMF component of 1 Pressure Fluctuation Signal and 3 pressure-difference fluctuation signals is designated as c respectively ₁₁(t), c ₁₂(t), Λ, c _1n1(t), c ₂₁(t), c ₂₂(t), Λ, c _2n2(t), c ₃₁(t), c ₃₂(t), Λ, c _3n3(t), c ₄₁(t), c ₄₂(t), Λ, c _4n4(t).These IMF components are formed this matrix of spy X, promptly

$X_{p\×n}=={\left[\begin{array}{cccc}{x}_{11}& x_{12}& \mathrm{\Λ}& x_{1p}\\ x_{21}& x_{22}& \mathrm{\Λ}& x_{2p}\\ M& M& M& M\\ x_{n1}& x_{n1}& \mathrm{\Λ}& x_{\mathrm{np}}\end{array}\right]}^T={\left[\begin{array}{c}{c}_{11}\\ M\\ c_{1n_1}\\ c_{21}\\ M\\ c_{2n_2}\\ c_{31}\\ M\\ c_{3n_3}\\ c_{41}\\ M\\ c_{4n_4}\end{array}\right]}^T---\left(8\right)$

In the formula: X is the eigenmatrix of p * n, and p is a sampling number, n=n ₁+ n ₂+ n ₃+ n ₄It is the number of all IMF components of 4 original signals.

(f) matrix X is transformed to correlation matrix with following formula, i.e. major component matrix R, and calculate m the eigenwert of correlation matrix R.

1. original variable is carried out standardization

Each index dimension is often different, and during analysis, the different dimensions and the order of magnitude can be drawn new problem, so want standardization, making standardized average of variable is 0, and variance is 1, promptly

$x_{\mathrm{ij}}^{*}=\frac{x_{\mathrm{ij}}-{\stackrel{\‾}{x}}_i}{\sqrt{\mathrm{var}\left(x_{\mathrm{ij}}\right)}}$ i＝1，2，Λ，n；j＝1，2，Λ，p (9)

In the formula: x _Ij ^*Be the observed quantity after the standardization, x _iBe mean value,

Be standard deviation.

2. calculate the correlation matrix R of X

After the standardization, relevant battle array is identical with covariance matrix, and the subcomponent of R can be expressed as

$r_{\mathrm{ij}}=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_{\mathrm{ti}}^{*}{x}_{\mathrm{ij}}^{*}---\left(10\right)$

In the formula: x _Ti ^*, x _Ij ^*Be respectively capable i row of t and j column element in the eigenmatrix.

3. ask eigenwert and the proper vector of correlation matrix R

Calculate each eigenwert (λ ₁＞λ ₂＞Λ＞λ _n) and each proper vector (u ₁, u ₂, Λ, u _n).

4. select major component

Calculate the contribution rate and the contribution rate of accumulative total of each major component, the number of major component is decided on particular problem, and it is that 90% a pairing m variable is as major component that the present invention gets contribution rate of accumulative total.

5. count the score

In principal component analysis (PCA), ${\mathrm{\μ}}_j={\mathrm{\λ}}_j/\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\λ}}_i$ Be the contribution rate of major component, ratio

The percentage of the primary data information (pdi) of k major component representative before the reflection.Order

T＝[μ ₁，μ ₂，Λ，μ _m] (11)

Then T is the proper vector of different flow patterns.

(4) adopt support vector machine to realize the identification of flow pattern: oil gas water multiphase flow type has 4 kinds, i.e. stratified flow, intermittent flow, bubble flow and annular flow in the horizontal tube.According to above-mentioned classification situation, adopt the strategy of " one-to-many ", construct 4 two-value sorters, be expressed as SVM1, SVM2, SVM3, SVM4 respectively.As shown in Figure 2.To support vector machine of each flow pattern structure, the output y of each support vector machine _i∈ 1 ,+1}, (i=1,2,3,4), then the output of 4 support vector machine is one 4 dimensional vectors, and whether each representation in components sample is to should flow pattern, (1 ,-1 ,-1 ,-1) represents stratified flow, (1,1 ,-1 ,-1) represents intermittent flow, (1 ,-1,1 ,-1) represents bubble flow, (1 ,-1 ,-1,1) represents annular flow.With 120 (4 kinds of each 30 samples of flow pattern) support vector machine is trained,, determine that kernel function of the present invention is radially basic kernel function, promptly by relatively $K(x,y)=e^{-{\left|\right|x-y\left|\right|}^2/{2\mathrm{\σ}}^2},$ σ gets 0.05 in the formula.The support vector machine that trains can realize the identification of convection.Test result's (partial test sample) as shown in table 1 again with 200 samples.

Table 1

Test sample book							Output	Flow pattern
									0.47561	0.29038	0.11616	0.08258	0.02529	0.00434	0.00473	(-1，-1，-1，1)	Annular flow
0.48657	0.28142	0.10764	0.07964	0.02637	0.00538	0.00456	(-1，-1，-1，1)	Annular flow
									0.48841	0.27038	0.11524	0.08357	0.02735	0.00469	0.00504	(-1，-1，-1，1)	Annular flow
0.57743	0.31873	0.06160	0.01991	0.00848	0.00563	0.00469	(-1，-1，1，-1)	Bubble flow
									0.57628	0.31772	0.06036	0.01814	0.01584	0.00433	0.00297	(-1，-1，1，-1)	Bubble flow
0.57628	0.31772	0.06036	0.01814	0.01584	0.00433	0.00297	(-1，-1，1，-1)	Bubble flow
									0.69828	0.11918	0.09776	0.06454	0.01191	0.004421	0.00229	(-1，1，-1，-1)	Intermittent flow
0.69828	0.11918	0.09776	0.06454	0.01191	0.004421	0.00229	(-1，1，-1，-1)	Intermittent flow
									0.68550	0.23403	0.05095	0.02419	0.00263	0.001564	0.00115	(-1，1，-1，-1)	Intermittent flow
0.47509	0.28644	0.13284	0.07738	0.01721	0.005168	0.00364	(1，-1，-1，-1)	Stratified flow
									0.42308	0.21099	0.16566	0.10569	0.074812	0.01078	0.00551	(1，-1，-1，-1)	Stratified flow
0.41356	0.23654	0.15324	0.12341	0.06244	0.01025	0.00631	(1，-1，-1，-1)	Stratified flow

Claims (5)

1. oil gas water multiphase flow type identification method based on principal component analysis (PCA) and support vector machine, it is characterized in that, principal component analytical method is combined with the support vector machine recognition technology, automatically oil gas water multiphase pipe stream flow pattern is discerned, its method comprises the steps:

1) the flow parameter signal of different flow patterns is gathered;

2) pressure and pressure-difference fluctuation signal are carried out the pre-service of wavelet packet removal noise;

3) pressure and the pressure-difference fluctuation signal of removing behind the noise carried out empirical mode decomposition, the natural mode function of each signal that decomposition is obtained is formed an eigenmatrix, utilizes principal component analytical method to extract the proper vector of flow pattern again;

4) with the energy feature vector of the empirical mode decomposition input sample as support vector machine, input is finished by support vector machine with the relation of output, and constructs 4 two-value sorters, is finished the identification of convection by radially basic kernel function support vector machine.

2. the oil gas water multiphase flow type identification method based on principal component analysis (PCA) and support vector machine according to claim 1 is characterized in that: described step 1 is to utilize the pressure spacing to be respectively 5 times, 10 times and 15 times of calibers (D); Sample frequency is 256Hz; Sampling number is 1024 points; The differential pressure transmitter of 1 pressure unit and 3 different pressure spacings is gathered in real time to the parameter signal that flows.

3. the oil gas water multiphase flow type identification method based on principal component analysis (PCA) and support vector machine according to claim 1 is characterized in that: the wavelet packet of described step 2 is removed the noise pre-service,

(a) to the pressure-difference fluctuation signal S of 1 pressure surge and 3 different pressure spacings ₁, S ₂, S ₃And S ₄Carry out 4 layers of WAVELET PACKET DECOMPOSITION according to following recursion (1):

$\left\{\begin{array}{c}{u}_{2n}\left(t\right)=\sqrt{2}\underset{k}{\mathrm{\Σ}}h\left(k\right)u_n(2t-k)\\ u_{2n-1}\left(t\right)=\sqrt{2}\underset{k}{\mathrm{\Σ}}g\left(k\right)u_n(2t-k)\end{array}\right.---\left(1\right)$

H in the formula (k) is the Hi-pass filter group, and g (k) is the low-pass filter group, u ₀(t)= (t) is a scaling function, u ₁(t)=and ψ (t) is a wavelet function, the yardstick of k for decomposing, and h (k) and g (k) satisfy orthogonality relation:

g(k)＝(-1) ^kh(1-k) (2)

(b) each frequency band difference setting threshold to decomposing, adopt following floating threshold form:

$t_k=\sqrt{2\ln \left(n\right)}\mathrm{\σ}/\sqrt{n}---\left(3\right)$

Wherein σ is a noise intensity, with median MAD (υ _k)/0.6745 (υ _kBe the WAVELET PACKET DECOMPOSITION coefficient) estimate that n is the length of signal;

(c) with nonlinear function ξ=sgn (υ _k) (| υ _k|-t _k) act on the WAVELET PACKET DECOMPOSITION coefficient υ of signal _k, obtain coefficient υ _k, soon the above coefficient of floating threshold deducts after the threshold value and remains;

(d) by coefficient υ _kRebuild original signal, thereby reach the removal noise.

4. the oil gas water multiphase flow type identification method based on principal component analysis (PCA) and support vector machine according to claim 1 is characterized in that: the extraction of described step 3 flow pattern proper vector,

(1) utilizes the empirical mode decomposition technology that 1 Pressure Fluctuation Signal and 3 pressure-difference fluctuation signals are decomposed, obtain n IMF component and be designated as c respectively ₁₁(t), c ₁₂(t), Λ, c _1n1(t), c ₂₁(t), c ₂₂(t), A, c _2n2(t), c ₃₁(t), c ₃₂(t), Λ, c _3n3(t), c ₄₁(t), c ₄₂(t), Λ, c _4n4(t); These IMF components are formed this matrix of spy X, promptly

$X_{p\×n}={\left[\begin{array}{cccc}{x}_{11}& x_{12}& \mathrm{\Λ}& x_{1p}\\ x_{21}& x_{22}& \mathrm{\Λ}& x_{2p}\\ M& M& M& M\\ x_{n1}& x_{n1}& \mathrm{\Λ}& x_{\mathrm{np}}\end{array}\right]}^T={\left[\begin{array}{c}{c}_{11}\\ M\\ c_{{1n}_1}\\ c_{21}\\ M\\ c_{{2n}_2}\\ c_{31}\\ M\\ c_{3n_3}\\ c_{41}\\ M\\ c_{{4n}_4}\end{array}\right]}^T$

In the formula: X is the eigenmatrix of p * n, and p is a sampling number, n=n ₁+ n ₂+ n ₃+ n ₄Be the number of all IMF components of 4 original signals, x _i(i=1,2, Λ n) is the p dimensional vector;

(2) original variable is carried out standardization: the different dimensions of each variable can not directly be calculated, so want standardization, making standardized average of variable is 0, and variance is 1, that is:

$x_{\mathrm{ij}}^{*}=\frac{x_{\mathrm{ij}}-{\stackrel{\‾}{x}}_i}{\sqrt{\mathrm{var}\left(x_{\mathrm{ij}}\right)}}i=\mathrm{1,2},\mathrm{\Λ},n;j=\mathrm{1,2},\mathrm{\Λ},p---\left(7\right)$

In the formula: x _Ij ^*Be the observed quantity after the standardization, x _iBe mean value,

Be standard deviation;

(3) the correlation matrix R of calculating X: after the standardization, relevant battle array is identical with covariance matrix, and the subcomponent of R can be expressed as:

$r_{\mathrm{ij}}=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{x}_{\mathrm{ti}}^{*}{x}_{\mathrm{ij}}^{*}---\left(8\right)$

In the formula: x _Ti ^*, x _Ij ^*Be respectively capable i row of t and j column element in the eigenmatrix;

(4) ask eigenwert and the proper vector of correlation matrix R: calculate each eigenwert (λ ₁＞λ ₂＞Λ＞λ _n) and each proper vector (u ₁, u ₂, Λ, u _n);

(5) select major component: calculate the contribution rate and the contribution rate of accumulative total of each major component, the number of major component is decided on particular problem, and it is that 90% a pairing m variable is as major component that the present invention gets contribution rate of accumulative total;

(6) count the score: calculate each principal component scores according to the major component expression formula, in principal component analysis (PCA), ${\mathrm{\μ}}_j={\mathrm{\λ}}_j/\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\λ}}_i$ Be the contribution rate of major component, ratio

The percentage of the primary data information (pdi) of k major component representative before the reflection, order:

T＝[μ ₁，μ ₂，Λ，μ _m] (9)

Then T is the proper vector of different flow patterns.

5. the oil gas water multiphase flow type identification method based on principal component analysis (PCA) and support vector machine according to claim 1, it is characterized in that: described step 4 is to adopt support vector machine to realize the identification of flow pattern, oil gas water multiphase flow type has 4 kinds in the horizontal tube, it is stratified flow, intermittent flow, bubble flow and annular flow, adopt the strategy of " one-to-many ", construct 4 two-value sorters, to support vector machine of each flow pattern structure, the output of 4 support vector machine is one 4 dimensional vectors, whether each representation in components sample is to should flow pattern, and the support vector machine that trains can realize the identification of convection.

Images