CN106022352A - Submersible piston pump fault diagnosis method based on support vector machine - Google Patents

Abstract

The invention provides a submersible piston pump fault diagnosis method based on a support vector machine, and relates to the submersible piston pump fault diagnosis method. According to the invention, the problem of high error rate of the existing submersible piston pump fault diagnosis method is solved. The method comprises the steps that feature quantity extraction is carried out on the running data of each cycle under n kinds of operating conditions; the feature quantity is normalized to acquire the input vector of a fault diagnosis model; the relative distance between a class i sample and a class j sample is calculated; according to the relative distance, a partial binary tree is constructed through samples of different classes; an LS-SVM method is used to construct dichotomy SVM corresponding to each level; and finally, according to an SVM classification learning machine based on the partial binary tree, the running data of a submersible linear motor are classified to realize submersible piston pump fault diagnosis. The method is applicable to submersible piston pump fault diagnosis.

Description

Latent oil piston failure of pump diagnostic method based on support vector machine

Technical field

The present invention relates to latent oil piston failure of pump diagnostic method.

Background technology

Owing to beam pumping unit is most widely used oil pumper in domestic and international each elephant, use historical time relatively Long, domestic and international research worker is the most deep to the research of its method for diagnosing faults, therefore the method for diagnosing faults of beam pumping unit is Through the most perfect, most common of which is exactly surface dynamometer card method.But latent oil piston pump is that one passes through oil immersion line motor The new pumping unit directly driven, different from the pump body structure of traditional beam pumping unit.Therefore, it is impossible to use ground to show merit The running status of latent oil piston pump is diagnosed by the mode of figure.

Because the latent oil piston pump use time in actual production is the longest, so grinding its method for diagnosing faults both at home and abroad Study carefully and go back imperfection at present.Neural Network Fault Diagnosis Method on the basis of existing frequently-used is based on traditional statistics, wherein Main method has BP Neural Network Diagnosis Method and learning vector quantization (LVQ) Neural Network Diagnosis Method.Neutral net is fitted Sample for research tend to infinite many time operating mode, but the quantity of sample data gathered in reality is the most limited, therefore The algorithm of neutral net is difficult to have in sample data obtain effect the most reliably in limited time.And the fault diagnosis of oil piston pump of diving is One obvious small sample, the non-linear and problem of fuzzy relation, it is therefore desirable to the on-line fault diagnosis side that a kind of generalization is strong Method.And the approximation capability of BP and LVQ Neural Network Diagnosis Method and generalization ability are the most relatively weak, under condition of small sample, base The most higher in the False Rate of the method for diagnosing faults of neutral net.

Summary of the invention

The present invention is the problem higher in order to solve the method for diagnosing faults False Rate of existing latent oil piston pump.

Latent oil piston failure of pump diagnostic method based on support vector machine, comprises the steps:

Step 1, obtain oil immersion line motor history data under various operating modes, each under n kind operating mode The service data in cycle carries out Characteristic Extraction, and characteristic quantity includes: T_uu,T_ud,T_du,T_dd,I_uv,I_dv,I_um,I_dm,P_um,P_dm,P_uv, P_dv；T_uuFor upstroke load time, T_udFor upstroke discharge time, T_duFor down stroke load time, T_udUnload for down stroke Time, I_uvFor linear electric motors upstroke electric current variance, I_dvFor linear electric motors down stroke electric current variance, I_umFor linear electric motors upper punch The electric current average of journey, I_dmFor the electric current average of linear electric motors down stroke, P_umFor upstroke load average, P_dmFor down stroke load Average, P_uvFor upstroke load variance, P_dvFor down stroke load variance；

Step 2, structure input space vector:

Owing to the order of magnitude between the characteristic quantity of extraction differs greatly, so must be to the spy extracted before being trained The amount of levying is normalized, and makes all of data point reuse in the range of the same order of magnitude；

Normalize to all characteristic quantities of the linear electric motors of extraction, between [-1,1], to have

${\stackrel{\‾}{x}}_i=2\frac{x_i-x_{min}}{x_{\max }-x_{\min }}-1---(2-1)$

In formula:Represent normalization data；x_iThe raw value of representative feature；

x_maxRepresentative is the maximum in each characteristic quantity sequence；

x_minRepresentative is the minima in each characteristic quantity sequence；

Respectively the characteristic quantity T extracted_uu,T_ud,T_du,T_dd,I_uv,I_dv,I_um,I_dm,P_um,P_dm,P_uv,P_dvSubstitution formula (2-1) In, obtain characteristic quantity normalization data

The input vector G of fault diagnosis model can be obtained；

$G={\[{\stackrel{\‾}{T}}_{uu},{\stackrel{\‾}{T}}_{ud},{\stackrel{\‾}{T}}_{du},{\stackrel{\‾}{T}}_{dd},{\stackrel{\‾}{I}}_{uv},{\stackrel{\‾}{I}}_{dv},{\stackrel{\‾}{I}}_{um},{\stackrel{\‾}{I}}_{dm},{\stackrel{\‾}{P}}_{um},{\stackrel{\‾}{P}}_{dm},{\stackrel{\‾}{P}}_{uv},{\stackrel{\‾}{P}}_{dv}\]}^T---(2-2)$

Step 3, foundation svm classifier statistical learning machine based on inclined binary tree:

Due to its distinctive features of latent oil piston pump, it is therefore desirable to the fault diagnosis mould that a kind of generalization ability is stronger Type, if being the most also suitable for the Nonlinear Dynamic on-line fault diagnosis method of small sample；

The inclined binary-tree support vector machine grader that the present invention uses is by constructing the realization of multiple SVM two-value graders 's；The main thought of the method: for n class classification problem, needs to construct n-1 two-value grader；Construct the 1st SVM two-value During grader, the sample belonging to the 1st class being designated as positive class, other all of samples are classified as negative class, and the rest may be inferred, until all of Till node the most only comprises a single classification, train SVM；Grader concrete structure is as shown in Figure 3；

Step 3.1, grader structure design in, use relative distance weigh the difference degree between two classes, The relative distance calculated between class i sample and class j sample is D_ij:

N kind operating mode is designated as n classification, and the i-th class operating mode is designated as class i；If X is the sample set comprising n classification, X_iFor class The sample set of i, the namely set of the input vector that multiple cycles are corresponding under the i-th class operating mode；Then:

The center of a sample of class i is:

$c_i=\frac{1}{n_i}\underset{{\stackrel{\→}{x}}_i\∈X_i}{\Σ}{\stackrel{\→}{x}}_i,i=1,2,...,n---(3-1)$

Wherein, c_iIt is the center of a sample of class i, n_iIt it is the sample size of class i；For the sample of class i, represent each cycle pair The input vector G answered；

Then the Euclidean distance at the center of class i sample and class j sample is:

d_ij=| | c_i-c_j||；I, j=1,2 ..., n (3-2)

The minimal hyper-sphere radius of class i sample is:

R_i=max{ | | c_i-x_i||} (3-3)

Relative distance between class i sample and class j sample is:

$D_{ij}=\frac{d_{ij}}{R_i+R_j},D_{ij}\∈\[0,+\mathrm{\∞})---(3-4)$

Step 3.2, according to relative distance D_ij, by the different classes of inclined binary tree of sample architecture；Priority in classification On, it is set to higher priority by high class can be indexed, identifies at first；Remaining class sorts successively；

Step 3.3, according to sample set and partially about every grade subclass of binary tree, utilize LS-SVM method construct every grade corresponding Two classification SVM；

Step 4, basis svm classifier statistical learning machine based on inclined binary tree, enter oil immersion line motor service data Row classification, it is achieved latent oil piston failure of pump diagnosis.

The present invention has the effect that

1, relative to conventional method for diagnosing faults such as Neural Network Diagnosis Method and Neural Network Diagnosis Method, the present invention Being more applicable for the small sample of latent oil piston pump condition and nonlinear feature, generalization is higher, effectively reduces latent oil reciprocal The False Rate of each fault of oil pumper, comprehensive False Rate can reach less than 5%, it is ensured that the accuracy of fault diagnosis, can be very The good requirement reached to produce during control latent oil piston failure of pump information.

2, the present invention just can analyze according only to the running status of oil immersion line motor and obtain under various operating mode that it loads and unloads Situation about carrying, the method can reflect the feature of the most various most common failure more accurately.By defining upper (lower) stroke In add (unloading) carry threshold value, and from its operational factor, extract characteristic quantity, characteristic quantity just can be as this fault diagnosis machine Input.

3, this method is based on based on binary tree sort rule and diagnostic method, can preferably reach latent oil piston pump Fault diagnosis requirement, and number of training when making False Rate minimum in simulated failure diagnostic test can be obtained.Therefore, can be latent The control system of oil piston pump provides and compares work information accurately and reliably, it is achieved its production efficiency is high, continuous and steady operation Target.

Accompanying drawing explanation

Fig. 1 (a) is the voltage curve of linear electric motors, the current curve of Fig. 1 (b) linear electric motors；

Fig. 2 is the starting current pretreatment schematic diagram of linear electric motors；

Fig. 3 is inclined binary-tree support vector machine many classification schematic diagram；

Fig. 4 is analog platform structure diagram；

Fig. 5 (a) is fault F₁、F₂、F₇、F₈False Rate curve, Fig. 5 (b) is fault F₃、F₄、F₅、F₆False Rate curve；

Fig. 6 is the False Rate curve of SVM and LVQ diagnostic method.

Detailed description of the invention

Detailed description of the invention one:

Latent oil piston failure of pump diagnostic method based on support vector machine, comprises the steps:

Step 1, obtain oil immersion line motor history data under various operating modes, each under n kind operating mode The service data in cycle carries out Characteristic Extraction, and characteristic quantity includes: T_uu,T_ud,T_du,T_dd,I_uv,I_dv,I_um,I_dm,P_um,P_dm,P_uv, P_dv；T_uuFor upstroke load time, T_udFor upstroke discharge time, T_duFor down stroke load time, T_udUnload for down stroke Time, I_uvFor linear electric motors upstroke electric current variance, I_dvFor linear electric motors down stroke electric current variance, I_umFor linear electric motors upper punch The electric current average of journey, I_dmFor the electric current average of linear electric motors down stroke, P_umFor upstroke load average, P_dmFor down stroke load Average, P_uvFor upstroke load variance, P_dvFor down stroke load variance；

Step 2, structure input space vector:

Owing to the order of magnitude between the characteristic quantity of extraction differs greatly, so must be to the spy extracted before being trained The amount of levying is normalized, and makes all of data point reuse in the range of the same order of magnitude；

Normalize to all characteristic quantities of the linear electric motors of extraction, between [-1,1], to have

${\stackrel{\‾}{x}}_i=2\frac{x_i-x_{min}}{x_{\max }-x_{\min }}-1---(2-1)$

In formula:Represent normalization data；x_iThe raw value of representative feature；

x_maxRepresentative is the maximum in each characteristic quantity sequence；

x_minRepresentative is the minima in each characteristic quantity sequence；

Respectively the characteristic quantity T extracted_uu,T_ud,T_du,T_dd,I_uv,I_dv,I_um,I_dm,P_um,P_dm,P_uv,P_dvSubstitution formula (2-1) In, obtain characteristic quantity normalization data

The input vector G of fault diagnosis model can be obtained；

$G={\[{\stackrel{\‾}{T}}_{uu},{\stackrel{\‾}{T}}_{ud},{\stackrel{\‾}{T}}_{du},{\stackrel{\‾}{T}}_{dd},{\stackrel{\‾}{I}}_{uv},{\stackrel{\‾}{I}}_{dv},{\stackrel{\‾}{I}}_{um},{\stackrel{\‾}{I}}_{dm},{\stackrel{\‾}{P}}_{um},{\stackrel{\‾}{P}}_{dm},{\stackrel{\‾}{P}}_{uv},{\stackrel{\‾}{P}}_{dv}\]}^T---(2-2)$

Step 3, foundation svm classifier statistical learning machine based on inclined binary tree:

Due to its distinctive features of latent oil piston pump, it is therefore desirable to the fault diagnosis mould that a kind of generalization ability is stronger Type, if being the most also suitable for the Nonlinear Dynamic on-line fault diagnosis method of small sample；

The inclined binary-tree support vector machine grader that the present invention uses is by constructing the realization of multiple SVM two-value graders 's；The main thought of the method: for n class classification problem, needs to construct n-1 two-value grader；Construct the 1st SVM two-value During grader, the sample belonging to the 1st class being designated as positive class, other all of samples are classified as negative class, and the rest may be inferred, until all of Till node the most only comprises a single classification, train SVM；Grader concrete structure is as shown in Figure 3；

Step 3.1, grader structure design in, use relative distance weigh the difference degree between two classes, The relative distance calculated between class i sample and class j sample is D_ij:

N kind operating mode is designated as n classification, and the i-th class operating mode is designated as class i；If X is the sample set comprising n classification, X_iFor class The sample set of i, the namely set of the input vector that multiple cycles are corresponding under the i-th class operating mode；Then:

The center of a sample of class i is:

$c_i=\frac{1}{n_i}\underset{{\stackrel{\→}{x}}_i\∈X_i}{\Σ}{\stackrel{\→}{x}}_i,i=1,2,...,n---(3-1)$

Wherein, c_iIt is the center of a sample of class i, n_iIt it is the sample size of class i；For the sample of class i, represent each cycle pair The input vector G answered；

Then the Euclidean distance at the center of class i sample and class j sample is:

d_ij=| | c_i-c_j||；I, j=1,2 ..., n (3-2)

The minimal hyper-sphere radius of class i sample is:

R_i=max{ | | c_i-x_i||} (3-3)

Relative distance between class i sample and class j sample is:

$D_{ij}=\frac{d_{ij}}{R_i+R_j},D_{ij}\∈\[0,+\mathrm{\∞})---(3-4)$

Step 3.2, according to relative distance D_ij, by the different classes of inclined binary tree of sample architecture；Priority in classification On, it is set to higher priority by high class can be indexed, identifies at first；Remaining class sorts successively；

Step 3.3, according to sample set and partially about every grade subclass of binary tree, utilize LS-SVM method construct every grade corresponding Two classification SVM；

Step 4, basis svm classifier statistical learning machine based on inclined binary tree, enter oil immersion line motor service data Row classification, it is achieved latent oil piston failure of pump diagnosis.

Detailed description of the invention two:

The detailed process of the Characteristic Extraction described in present embodiment step 1 is as follows:

Step 1.1, linear electric motors service data pretreatment:

Latent oil piston pump work is in batch (-type) pattern, and its cycle of operation includes " in upstroke under interval down stroke Intermittently " four states；Due to the impact of linear electric motors starting current, all can in submersible electric machine with oil upstroke just section and down stroke just section Occur that the capital of a peak current, i.e. upstroke on submersible electric machine with oil current curve just section and down stroke just section occurs one Significantly spike, as shown in Fig. 1 (a) and (b)；

In order to avoid its impact on fault diagnosis, the peak value by peak current (starting current) is needed to filter；Process side Method is as in figure 2 it is shown, for upstroke just section and down stroke just section, be respectively processed: by the peak current peak on current curve First electric current minimum after value (starting current peak value) is as reference value and identical with reference value before finding current peak Sampled point (is actually intended to look for the point of identical value, but owing to sampling process is discrete data, is difficult to find the most equal Point, so finding the sampled point close with reference value in practical operation), by (t between this sampled point and electric current minimum₁、t₂It Between) current value all with reference value replace (i.e. straight line portion in Fig. 2)；

Step 1.2, from the operational factor of linear electric motors extract characteristic quantity:

In order to preferably study the operating mode of linear electric motors, for the upstroke process of linear electric motors, now it is defined as follows:

$I_{um}=\frac{1}{t_u}\underset{t=1}{\overset{t_u}{\Σ}}{i}_u\left(t\right)---(1-1)$

In formula: i_uT () is the real-time current of linear electric motors upstroke；

t_uTotal sample points for linear electric motors upstroke；

I_umElectric current average for linear electric motors upstroke；

I_uu=K_uuI_um, wherein K_uu∈(0,1) (1-2)

I_ud=K_udI_um, wherein K_ud∈(0,1) (1-3)

In formula: K_uuFor upstroke current load coefficient；K_udFor upstroke electric current off-loading coefficient；

I_uuFor upstroke loading current gate value；I_udElectric current gate value is unloaded for upstroke；

In formula: T_uuFor the upstroke load time；L is the upstroke a certain sampled point correspondence moment；

I_lCurrent value for l moment correspondence sampled point；

In formula: T_udFor upstroke discharge time；H is the upstroke a certain sampled point correspondence moment；

I_hCurrent value for h moment correspondence sampled point；

Corresponding with the upstroke process of linear electric motors, equally the down stroke of linear electric motors is defined:

$I_{dm}=\frac{1}{t_d}\underset{t=1}{\overset{t_d}{\Σ}}{i}_d\left(t\right)---(1-6)$

In formula: i_dT () is the real-time current of linear electric motors down stroke；

t_dTotal sample points for linear electric motors down stroke；

I_dmElectric current average for linear electric motors down stroke；

I_du=K_duI_dm, wherein K_du∈(0,1) (1-7)

I_dd=K_ddI_dm, wherein K_dd∈(0,1) (1-8)

In formula: K_duFor down stroke current load coefficient；K_ddFor down stroke electric current off-loading coefficient；

I_duFor down stroke loading current gate value；I_ddElectric current gate value is unloaded for down stroke；

In formula: T_duFor the down stroke load time；M is the down stroke a certain sampled point correspondence moment；

I_mCurrent value for m moment correspondence sampled point；

In formula: T_udFor down stroke discharge time；N is the down stroke a certain sampled point correspondence moment；

I_nCurrent value for n moment correspondence sampled point；

$I_{uv}=\frac{1}{t_u}\left(\underset{t=1}{\overset{t_u}{\Σ}}{\left(i_u\left(t\right)-I_{um}\right)}^2\right)---(1-11)$

In formula: I_uvFor linear electric motors upstroke electric current variance；

$I_{dv}=\frac{1}{t_d}\left(\underset{t=1}{\overset{t_d}{\Σ}}{\left(i_d\right(t)-I_{dm})}^2\right)---(1-12)$

In formula: I_dvFor linear electric motors down stroke electric current variance；

In addition it is also necessary to load average and load variance to linear electric motors upstroke and down stroke are defined:

$P_{um}=\frac{1}{t_u}\underset{t=1}{\overset{t_u}{\Σ}}{P}_u\left(t\right)---(1-13)$

$P_{uv}=\frac{1}{t_u}\left(\underset{t=1}{\overset{t_u}{\Σ}}{\left(P_u\left(t\right)-P_{um}\right)}^2\right)---(1-14)$

In formula: P_uT () is the real-time load of linear electric motors upstroke, P_umFor upstroke load average；

P_uvFor upstroke load variance；

$P_{dm}=\frac{1}{t_d}\underset{t=1}{\overset{t_d}{\Σ}}{P}_d\left(t\right)---(1-15)$

$P_{dv}=\frac{1}{t_u}\left(\underset{t=1}{\overset{t_u}{\Σ}}{\left(P_d\left(t\right)-P_{dm}\right)}^2\right)---(1-16)$

In formula: P_dT () is the real-time load of linear electric motors down stroke, P_dmFor down stroke load average；

P_dvFor down stroke load variance.

During all of Characteristic Extraction, I_um,I_dm,P_um,P_dm,P_uv,P_dvBe linear electric motors run time intrinsic Parameter, and T_uu,T_ud,T_du,T_dd,I_uv,I_dvThe value of characteristic quantity is by sample data and parameter K_uu,K_ud,K_du,K_ddTogether decide on, Therefore to the accuracy improving fault diagnosis needs to select suitably to load and the coefficient of unloading.

Other step and parameter are identical with detailed description of the invention one.

Detailed description of the invention three:

The detailed process constructing inclined binary tree described in present embodiment step 3.2 is as follows:

(1) for a n class problem, it is D according to class i and class j and relative distance_ijStructure relative distance matrix D, relatively The first of Distance matrix D is classified as the label of class i, and second is classified as the label of class j, and the 3rd is classified as class i, relative distance D of j_ij；Will be each The label of class is ascending to be stored in set Q；

(2) according to D, find in Q, with other all classes, there is class i of maximum relative distance, class i is stored in set Q₁In, And remaining class in Q is stored in set Q by the size of class label₂In；

(3) respectively by Q₁、Q₂As the left and right subtree of binary tree, so far, the left and right subclass of a sub-classifier is obtained；

(4) Q=Q is made₂, return (2), right subtree be divided further into 2 subtrees, until each class becomes y-bend Till the leaf node of tree；

(5) construct inclined binary tree to terminate.

Other step and parameter are identical with detailed description of the invention one or two.

Detailed description of the invention four:

The detailed process of the two classification SVM that LS-SVM method construct every grade is corresponding is utilized described in present embodiment step 3.3 As follows:

Support vector machine is based on structural risk minimization, the nonlinear problem of lower dimensional space is transformed to higher-dimension empty Between linear problem, and in higher dimensional space, obtain its Generalized optimal hyperplane, thus build an optimal hyperlane statistics Practise machine；

LS-SVM method construct two is utilized to classify SVM, if there is nonlinear mapping φ: an x → φ (x), by sample It is mapped to a certain m from former n dimension lower dimensional space and ties up high-dimensional feature space；Then its Optimal Separating Hyperplane can be expressed as

F (x)=ω^T×φ(x)+b (3-5)

In formula: ω, φ (x) is m dimensional vector, and b is amount of bias, requires that 2/ | | ω | | is maximum, therefore on optimal hyperlane It is converted into and solves quadratic programming problem by Lagrangian method；

LS-SVM have selected error ξ_iQuadratic term, its optimization problem is

$\underset{\mathrm{\ω},\mathrm{\ξ}}{min}J=\frac{1}{2}\left|\right|\mathrm{\ω}|{|}^2+\frac{1}{2}C\underset{1}{\overset{n}{\Σ}}{{\mathrm{\ξ}}_i}^2---(3-6)$

s.t. y_i(ω^T×φ(x_i)+b)=1-ξ_i

In formula: ξ refers to total error, ξ_iRefer to the error of class i；y_iFor corresponding x_iDesired output；C is penalty coefficient；

Can seek optimal solution by the dual form of formula (3-6), dual form can be set up according to object function and constraints Lagrangian:

$L(\mathrm{\ω},b,\mathrm{\ξ},a)=\frac{1}{2}\left|\right|\mathrm{\ω}|{|}^2+\frac{1}{2}C\underset{i=1}{\overset{n}{\Σ}}{\mathrm{\ξ}}_i-\underset{i=1}{\overset{n}{\Σ}}{a}_i\left(y_i\right({\mathrm{\ω}}^T\×\mathrm{\φ}\left(x_i\right)+b)-1+{\mathrm{\ξ}}_i)---(3-7)$

In formula: a is Lagrange multiplier, according to Kuhn-Tucker condition

$\frac{\∂L}{\∂\mathrm{\ω}}=0,\frac{\∂L}{\∂b}=0,\frac{\∂L}{\∂\mathrm{\ξ}}=0,\frac{\∂L}{\∂a}=0,$

Can obtain

$\left[\begin{array}{cc}0& {\left({\mathrm{\η}}_i\right)}^T\\ {\mathrm{\η}}_i& \mathrm{\Ω}+C^{-1}I\end{array}\right]\left[\begin{array}{c}b\\ a\end{array}\right]=\left[\begin{array}{c}0\\ Y\end{array}\right]---(3-8)$

In formula: Ω=[φ (x_i)]^Tφ(x_i)；Y=[y₁,y₂,…,y_i]^T；

η_i=[1,1 ..., 1]^T；A=[a₁,a₂,…,a_i]

Available least square solution obtains a and b, thus obtains categorised decision function and is:

$y\left(x\right)=sign\left(\underset{S_V}{\Σ}{a}_i\right(\mathrm{\φ}{\left(x\right)}^T\mathrm{\φ}\left(x_i\right)+b\left)\right)---(3-9)$

In formula: S_VFor supporting vector set；

Can be seen that in formula (3-9) the most relevant with the inner product operation of sample, according to the correlation theory of functional, if a kind of core Function k (x_i,x_j) meeting Mercer condition, it is with regard to the inner product in corresponding a certain transformation space；Therefore if there is a certain kernel function k(x_i,x_j)=φ (x_i)φ(x_j) meet Mercer condition, in the case of concrete form the unknown of φ (x), can obtain former Discriminant function corresponding to the input space is:

$y\left(x\right)=sign\left(\underset{S_V}{\Σ}{a}_i\right(k\left(x,k_i\right)+b)---(3-10)$

Find out the support vector of corresponding level after every one-level SVM is trained, set up optimal classification surface；Owing to n-1 SVM is Arrange from high to low according to priority, when new model produces, only need to scan for from high to low according to binary tree, so that it may draw Conclusion.

Additionally, another group linear electric motors service data under each operating mode test specimens as this diagnostic method can be recorded This, be used for checking this method for diagnosing faults False Rate under given each SVM parameter and threshold value parameter, thus verify that this is dived The quality of oil piston failure of pump diagnostor.

Other structure and parameter are identical with one of detailed description of the invention one to three.

Detailed description of the invention five:

In present embodiment step 3.3 construct two classification SVM during discriminant function kernel function selection course such as Under:

This fault diagnosis model is selected radial direction base core as the kernel function of fault diagnosis, the form of RBF such as formula (4-1) shown in；

k(x,μ_i)=exp (-| | x-μ_i||/σ²) (4-1)

In formula: σ is the core width control system factor；μ_iIt it is the center of corresponding kernel function.

By kernel function characteristic quantity G changed the C further determining that model_i、σ_i, finally obtain under each operating mode is inclined Binary tree SVM classifier, as shown in Figure 3.

Other structure and parameter are identical with detailed description of the invention four.

Embodiment

The present invention is utilized to carry out emulation experiment:

The major part of experiment porch is a set of oil well simulation circulating system, and this experiment porch is mainly by indoor oil well simulation Blood circulation, volume control device, latent oil piston pump and control device composition thereof, its structure diagram is as shown in Figure 4.

Experiment detects the oil immersion line motor running status under the various operating modes of simulation in real time, and it is corresponding to record it The important parameter such as operating current, power.

The jig frequency cycle of oil immersion line motor is set to 8 times/min by this experiment, measures every kind of simulated condition respectively Under the related data in 300 cycles, one input vector of data genaration in each cycle owing to measuring, therefore can be by front 230 input vectors are as the training sample of grader, and rear 70 input vectors are as test sample.By training sample to latent Oil piston failure of pump diagnostic classification device is trained, and is further tested this grader by test sample, wherein K_uu, K_ud,K_du,K_ddParameter value be all artificial selected, if the numerical value of these parameters is different, then training sample changes the most therewith, Thus parameter C and σ obtained also differs.The False Rate of each fault, as shown in Table 6-1 (F under different parameters₁: gas shadow Ring, F₂: liquid is not fully filled, F₃: travelling valve is missed, F₄: standing valve is missed, F₅: double valves are missed, F₆: oil pipe is missed, F₇: Shake out, F₈: wax deposition, F₉: properly functioning).

The False Rate of each fault of table 6-1SVM method

It can be seen that work as K_uu=0.85, K_ud=0.8, K_du=0.8, K_ddErroneous judgement when=0.75, under every kind of running status Rate is in more satisfactory scope.

From Fig. 5 (a) and Fig. 5 (b), along with number of training purpose is gradually increased, the False Rate of fault diagnosis machine is also Decrease.Curve from figure is it can be seen that the erroneous judgement of the fault such as " gases affect ", " liquid is not fully filled ", " sand card ", " wax deposition " The False Rate downward trend of the faults such as rate has obvious downward trend in first half section, oil pipe leakage is the most inconspicuous.But It is when training sample increases to some, is but difficult to make their False Rate reduce further, can produce certain on the contrary Fluctuation, the False Rate situation of the most comprehensive each fault just can obtain suitable number of training.

Error analysis and comparing:

Experiment also use LVQ Neural Network Diagnosis Method to be analyzed to group training and test sample.By LVQ god Being defined as 12-50-9 through the structure of network, the dimension of the input layer number amount of being characterized vector, output layer is object run shape State correspondence classification, competition layer neuron is 50.Training sample is inputted LVQ diagnostor, and by test sample, it is tested. Analyzing the False Rate of each running status diagnostic result under different number of training, its test result is as shown in table 6-2.And with wax deposition As a example by fault, provide the contrast of SVM and LVQ False Rate curve under 50～340 varying number training samples, such as Fig. 6 institute Show.

The False Rate of each fault of table 6-2LVQ method

As can be seen from Figure 6, LVQ diagnostic method is in the incipient stage along with the increase of training sample, and False Rate has under substantially Fall trend, but along with the further increase of training sample, the decrease speed of its False Rate starts to slow down, and its diagnostic accuracy can not Quickly reach in claimed range.Therefore, the method for diagnosing faults of this support vector machine is compared to LVQ Neural Network Diagnosis Method more Add the fault diagnosis on applicable small sample problem.

Claims (5)

1. latent oil piston failure of pump diagnostic method based on support vector machine, it is characterised in that comprise the steps:

Step 1, acquisition oil immersion line motor history data under various operating modes, for each cycle under n kind operating mode Service data carry out Characteristic Extraction, characteristic quantity includes: T_uu,T_ud,T_du,T_dd,I_uv,I_dv,I_um,I_dm,P_um,P_dm,P_uv,P_dv； T_uuFor upstroke load time, T_udFor upstroke discharge time, T_duFor down stroke load time, T_udFor down stroke discharge time, I_uvFor linear electric motors upstroke electric current variance, I_dvFor linear electric motors down stroke electric current variance, I_umElectricity for linear electric motors upstroke Stream average, I_dmFor the electric current average of linear electric motors down stroke, P_umFor upstroke load average, P_dmFor down stroke load average, P_uv For upstroke load variance, P_dvFor down stroke load variance；

Step 2, structure input space vector:

Normalize to all characteristic quantities of the linear electric motors of extraction, between [-1,1], to have

${\stackrel{\‾}{x}}_i=2\frac{x_i-x_{min}}{x_{\max }-x_{min}}-1---(2-1)$

In formula:Represent normalization data；x_iThe raw value of representative feature；

x_maxRepresentative is the maximum in each characteristic quantity sequence；

x_minRepresentative is the minima in each characteristic quantity sequence；

Respectively the characteristic quantity T extracted_uu,T_ud,T_du,T_dd,I_uv,I_dv,I_um,I_dm,P_um,P_dm,P_uv,P_dvIn substitution formula (2-1), To characteristic quantity normalization data

The input vector G of fault diagnosis model can be obtained；

$G={\[{\stackrel{\‾}{T}}_{uu},{\stackrel{\‾}{T}}_{ud},{\stackrel{\‾}{T}}_{du},{\stackrel{\‾}{T}}_{dd},{\stackrel{\‾}{I}}_{uv},{\stackrel{\‾}{I}}_{dv},{\stackrel{\‾}{I}}_{um},{\stackrel{\‾}{I}}_{dm},{\stackrel{\‾}{P}}_{um},{\stackrel{\‾}{P}}_{dm},{\stackrel{\‾}{P}}_{uv},{\stackrel{\‾}{P}}_{dv}\]}^T---(2-2)$

Step 3, foundation svm classifier statistical learning machine based on inclined binary tree:

Step 3.1, the relative distance calculated between class i sample and class j sample are D_ij:

N kind operating mode is designated as n classification, and the i-th class operating mode is designated as class i；If X is the sample set comprising n classification, X_iSample for class i This collection, the namely set of the input vector that multiple cycles are corresponding under the i-th class operating mode；Then:

The center of a sample of class i is:

$c_i=\frac{1}{n_i}\underset{{\stackrel{\→}{x}}_i\∈X_i}{\Σ}{\stackrel{\→}{x}}_i,i=1,2,...,n---(3-1)$

Wherein, c_iIt is the center of a sample of class i, n_iIt it is the sample size of class i；For the sample of class i, represent that each cycle is corresponding Input vector G；

Then the Euclidean distance at the center of class i sample and class j sample is:

d_ij=| | c_i-c_j||；I, j=1,2 ..., n (3-2)

The minimal hyper-sphere radius of class i sample is:

R_i=max{ | | c_i-x_i||} (3-3)

Relative distance between class i sample and class j sample is:

$D_{ij}=\frac{d_{ij}}{R_i+R_j},D_{ij}\∈\[0,+\mathrm{\∞})---(3-4)$

Step 3.2, according to relative distance D_ij, by the different classes of inclined binary tree of sample architecture；In the priority of classification, will High class can be indexed and be set to higher priority, identify at first；Remaining class sorts successively；

Step 3.3, according to sample set and partially about every grade subclass of binary tree, utilize that LS-SVM method construct every grade is corresponding two points Class SVM；

Step 4, according to svm classifier statistical learning machine based on inclined binary tree, oil immersion line motor service data is carried out point Class, it is achieved latent oil piston failure of pump diagnosis.

Latent oil piston failure of pump diagnostic method based on support vector machine the most according to claim 1, it is characterised in that step The detailed process of the Characteristic Extraction described in rapid 1 is as follows:

Step 1.1, linear electric motors service data pretreatment:

Latent oil piston pump work is in batch (-type) pattern, and its cycle of operation includes " interval under interval down stroke in upstroke " Four states；A peak current is all there will be, i.e. at submersible electric machine with oil electricity in submersible electric machine with oil upstroke just section and down stroke just section There is an obvious spike in the capital of the just section of the upstroke on flow curve and down stroke just section；

For upstroke just section and down stroke just section, it is respectively processed: by the after the peak current peak value on current curve One electric current minimum is as reference value, and sampled point identical with reference value before finding peak current peak value, by this sampled point And the current value between electric current minimum all replaces with reference value；

Step 1.2, from the operational factor of linear electric motors extract characteristic quantity:

For the upstroke process of linear electric motors, define:

$I_{um}=\frac{1}{t_u}\underset{t=1}{\overset{t_u}{\Σ}}{i}_u\left(t\right)---(1-1)$

In formula: i_uT () is the real-time current of linear electric motors upstroke；

t_uTotal sample points for linear electric motors upstroke；

I_umElectric current average for linear electric motors upstroke；

I_uu=K_uuI_um, wherein K_uu∈(0,1) (1-2)

I_ud=K_udI_um, wherein K_ud∈(0,1) (1-3)

In formula: K_uuFor upstroke current load coefficient；K_udFor upstroke electric current off-loading coefficient；

I_uuFor upstroke loading current gate value；I_udElectric current gate value is unloaded for upstroke；

In formula: T_uuFor the upstroke load time；L is the upstroke a certain sampled point correspondence moment；

I_lCurrent value for l moment correspondence sampled point；

In formula: T_udFor upstroke discharge time；H is the upstroke a certain sampled point correspondence moment；

I_hCurrent value for h moment correspondence sampled point；

Corresponding with the upstroke process of linear electric motors, the down stroke of linear electric motors is defined:

$I_{dm}=\frac{1}{t_d}\underset{t=1}{\overset{t_d}{\Σ}}{i}_d\left(t\right)---(1-6)$

In formula: i_dT () is the real-time current of linear electric motors down stroke；

t_dTotal sample points for linear electric motors down stroke；

I_dmElectric current average for linear electric motors down stroke；

I_du=K_duI_dm, wherein K_du∈(0,1) (1-7)

I_dd=K_ddI_dm, wherein K_dd∈(0,1) (1-8)

In formula: K_duFor down stroke current load coefficient；K_ddFor down stroke electric current off-loading coefficient；

I_duFor down stroke loading current gate value；I_ddElectric current gate value is unloaded for down stroke；

In formula: T_duFor the down stroke load time；M is the down stroke a certain sampled point correspondence moment；

I_mCurrent value for m moment correspondence sampled point；

In formula: T_udFor down stroke discharge time；N is the down stroke a certain sampled point correspondence moment；

I_nCurrent value for n moment correspondence sampled point；

$I_{uv}=\frac{1}{t_u}\left(\underset{t=1}{\overset{t_u}{\Σ}}{\left(i_u\left(t\right)-I_{um}\right)}^2\right)---(1-11)$

In formula: I_uvFor linear electric motors upstroke electric current variance；

$I_{dv}=\frac{1}{t_d}\left(\underset{t=1}{\overset{t_d}{\Σ}}{\left(i_d\left(t\right)-I_{dm}\right)}^2\right)---(1-12)$

In formula: I_dvFor linear electric motors down stroke electric current variance；

Load average and load variance to linear electric motors upstroke and down stroke are defined:

$P_{um}=\frac{1}{t_u}\underset{t=1}{\overset{t_u}{\Σ}}{P}_u\left(t\right)---(1-13)$

$P_{uv}=\frac{1}{t_u}\left(\underset{t=1}{\overset{t_u}{\Σ}}{\left(P_u\left(t\right)-P_{um}\right)}^2\right)---(1-14)$

In formula: P_uT () is the real-time load of linear electric motors upstroke, P_umFor upstroke load average；

P_uvFor upstroke load variance；

$P_{dm}=\frac{1}{t_d}\underset{t=1}{\overset{t_d}{\Σ}}{P}_d\left(t\right)---(1-15)$

$P_{dv}=\frac{1}{t_u}\left(\underset{t=1}{\overset{t_u}{\Σ}}{\left(P_d\left(t\right)-P_{dm}\right)}^2\right)---(1-16)$

In formula: P_dT () is the real-time load of linear electric motors down stroke, P_dmFor down stroke load average；

P_dvFor down stroke load variance.

Latent oil piston failure of pump diagnostic method based on support vector machine the most according to claim 1 and 2, it is characterised in that The detailed process constructing inclined binary tree described in step 3.2 is as follows:

(1) it is D according to class i and class j and relative distance_ijStructure relative distance matrix D, the first of relative distance matrix D is classified as The label of class i, second is classified as the label of class j, and the 3rd is classified as class i, relative distance D of j_ij；Deposit ascending for all kinds of label Enter to gather in Q；

(2) according to D, find in Q, with other all classes, there is class i of maximum relative distance, class i is stored in set Q₁In, and by Q In remaining class by the size of class label be stored in set Q₂In；

(3) respectively by Q₁、Q₂As the left and right subtree of binary tree, so far, the left and right subclass of a sub-classifier is obtained；

(4) Q=Q is made₂, return (2), right subtree be divided further into 2 subtrees, until each class becomes binary tree Till leaf node；

(5) construct inclined binary tree to terminate.

Latent oil piston failure of pump diagnostic method based on support vector machine the most according to claim 3, it is characterised in that step The detailed process utilizing the two classification SVM that LS-SVM method construct every grade is corresponding described in rapid 3.3 is as follows:

LS-SVM method construct two is utilized to classify SVM, if there is nonlinear mapping φ: an x → φ (x), by sample from former n Dimension lower dimensional space is mapped to a certain m and ties up high-dimensional feature space；Then its Optimal Separating Hyperplane can be expressed as

F (x)=ω^T×φ(x)+b (3-5)

In formula: ω, φ (x) is m dimensional vector, and b is amount of bias, requires that 2/ | | ω | | is maximum, therefore convert on optimal hyperlane For solving quadratic programming problem by Lagrangian method；

LS-SVM have selected error ξ_iQuadratic term, its optimization problem is

$\underset{\mathrm{\ω},\mathrm{\ξ}}{min}J=\frac{1}{2}\left|\right|\mathrm{\ω}|{|}^2+\frac{1}{2}C\underset{1}{\overset{n}{\Σ}}{\mathrm{\ξ}}_i^2---(3-6)$

s.t.y_i(ω^T×φ(x_i)+b)=1-ξ_i

In formula: ξ refers to total error, ξ_iRefer to the error of class i；y_iFor corresponding x_iDesired output；C is penalty coefficient；

Seeking optimal solution by the dual form of formula (3-6), dual form sets up Lagrange according to object function and constraints Function:

$L(\mathrm{\ω},b,\mathrm{\ξ},a)=\frac{1}{2}\left|\right|\mathrm{\ω}|{|}^2+\frac{1}{2}C\underset{i=1}{\overset{n}{\Σ}}{\mathrm{\ξ}}_i-\underset{i=1}{\overset{n}{\Σ}}{a}_i\left(y_i\right({\mathrm{\ω}}^T\×\mathrm{\φ}\left(x_i\right)+b)-1+{\mathrm{\ξ}}_i)---(3-7)$

In formula: a is Lagrange multiplier, according to Kuhn-Tucker condition

$\frac{\∂L}{\∂\mathrm{\ω}}=0,\frac{\∂L}{\∂b}=0,\frac{\∂L}{\∂\mathrm{\ξ}}=0,\frac{\∂L}{\∂a}=0,$

?

$\left[\begin{array}{cc}0& {\left({\mathrm{\η}}_i\right)}^T\\ {\mathrm{\η}}_i& \mathrm{\Ω}+C^{-1}I\end{array}\right]\left[\begin{array}{c}b\\ a\end{array}\right]=\left[\begin{array}{c}0\\ Y\end{array}\right]---(3-8)$

In formula: Ω=[φ (x_i)]^Tφ(x_i)；Y=[y₁,y₂,…,y_i]^T；

η_i=[1,1 ..., 1]^T；A=[a₁,a₂,…,a_i]

Obtaining a and b by least square solution, obtaining categorised decision function is:

$y\left(x\right)=sign\left(\underset{S_V}{\mathrm{\Σ}}{a}_i\right(\mathrm{\φ}{\left(x\right)}^T\mathrm{\φ}\left(x_i\right)+b\left)\right)---(3-9)$

In formula: S_VFor supporting vector set；

Obtaining discriminant function corresponding to the former input space is:

$y\left(x\right)=sign\left(\underset{S_V}{\Σ}{a}_i\right(k\left(x,x_i\right)+b)---(3-10)$

Find out the support vector of corresponding level after every one-level SVM is trained, set up optimal classification surface.

Latent oil piston failure of pump diagnostic method based on support vector machine the most according to claim 4, it is characterised in that step In rapid 3.3, during structure two classification SVM, the selection course of the kernel function of discriminant function is as follows:

Selecting radial direction base core as the kernel function of fault diagnosis, the form of RBF such as formula (4-1) is shown；

k(x,μ_i)=exp (-| | x-μ_i||/σ²) (4-1)

In formula: σ is the core width control system factor；μ_iIt it is the center of corresponding kernel function.