阅读说明：本技术 一种小样本条件下管道微弱泄漏检测方法 (Pipeline weak leakage detection method under a kind of condition of small sample ) 是由 刘金海 臧东 付明芮 马艳娟 马大中 冯健 朱和贵 于 2019-08-22 设计创作，主要内容包括：本发明提供一种小样本条件下管道微弱泄漏检测方法,涉及管道泄漏检测技术领域。本发明步骤如下：步骤1：获取真实样本集,并根据真实样本集生成虚拟样本集；步骤2：对真实样本集和虚拟样本集进行组合特征提取,所述组合特征提取包括7种统计特征和1组符号化变换特征；步骤3：根据7种统计特征和1组符号化变换特征采用朴素贝叶斯方法和最小二乘支持向量机方法分别建立朴素贝叶斯网络管道小泄漏辨识模型和最小二乘支持向量机管道小泄漏辨识模型,并利用模型对管道进行小泄漏检测。本方法从增加微弱泄漏样本数量和深度挖掘微弱泄漏样本特征两方面来构造微弱泄漏辨识模型,大大提高了管道微弱泄漏的检测准确率,保障输油管道的安全运行。(The present invention provides pipeline weak leakage detection method under a kind of condition of small sample, is related to Discussion on Pipe Leakage Detection Technology field.Steps are as follows by the present invention: step 1: obtaining authentic specimen collection, and generate virtual sample collection according to authentic specimen collection；Step 2: feature extraction being combined to authentic specimen collection and virtual sample collection, it includes 7 kinds of statistical natures and 1 group code transform characteristics that the assemblage characteristic, which extracts,；Step 3: according to 7 kinds of statistical natures and 1 group code transform characteristics establish naive Bayesian network pipeline Small leak identification model respectively using Nae Bayesianmethod and least square method supporting vector machine method and least square method supporting vector machine pipeline Small leak recognizes model, and carry out Small leak detection to pipeline using model.This method constructs faint leakage identification model in terms of increasing faint leakage sample size and the faint leakage sample characteristics two of depth excavation, substantially increases the Detection accuracy of pipeline weak leakage, ensures the safe operation of oil pipeline.)

1. pipeline weak leakage detection method under a kind of condition of small sample, characterized by the following steps:

Step 1: obtaining authentic specimen collection, and virtual sample collection is generated according to authentic specimen collection, the authentic specimen collection includes true Real normal sample collection X^NWith true faint leakage sample set X^S, the virtual sample collection includes virtual normal sample collection G primary^NWith Virtual faint leakage sample set G primary^S；

Step 2: feature extraction being combined to authentic specimen collection and virtual sample collection, it includes 7 kinds of systems that the assemblage characteristic, which extracts, Count feature and 1 group code transform characteristics；

Step 3: 7 kinds of statistical natures and 1 group code transform characteristics according to obtained in step 2 using Nae Bayesianmethod and Least square method supporting vector machine method establishes naive Bayesian network pipeline Small leak identification model respectively and least square is supported Vector machine pipeline Small leak recognizes model, carries out leak detection to pipeline according to identification model.

2. pipeline weak leakage detection method under a kind of condition of small sample according to claim 1, it is characterised in that: described Specific step is as follows for step 1:

Step 1.1: obtaining authentic specimen collection X^N, and generated according to the true normal sample in the sample set coarse virtual normal Sample generates coarse virtual normal sample collection

Pass through fittings for statistical analysis, data to the sample in true normal sample collection, the analysis of the evaluation of similitude, hair The data rule for having known true normal sample, generates coarse virtual normal sample

Wherein t indicates time, i.e. the detection length of sample, t={ 1,2 ..., L }；k₁Indicate the slope of true normal sample；ε is Data fluctuations caused by various noises in transmission process；Y_O1For the pressure initial value of setting, which normally presses according to oil pipeline The setting of power range；

Step 1.2: the data in the coarse virtual normal sample collection generated in step 1.1 being carried out according to the constraint of normal sample The sample for not meeting constraint condition is rejected, generates virtual normal sample collection primary by adjustment

It is wherein constrained according to normal sample, virtual normal sampleMeet:

Wherein,Indicate coarse virtual normal sample G^NThe minimum value and maximum value of interior sample, t (G^N) represent slightly Rough virtual normal sample X^N′Detection length；t_min、t_maxIndicate coarse virtual normal sampleMinimum detection length and Maximum detection length；Indicate any normal sample X^NThe minimal ripple of middle element,Indicate the equal of any normal sample Value,Constant is represented, Indicate any normal sample X^NThe maximum fluctuation of middle element,Constant is represented,σ₁Indicate leakage alarm threshold value, △ Y_minIndicate the minimum value of high reject signal；Indicate normal sample fluctuation Maximum value；

Step 1.3: according to true faint leakage sample constraint to virtual normal sample collection primaryIt is interior Data be adjusted, the sample that will do not meet true faint leakage sample constraint condition is rejected, and primary virtual faint let out is generated Leak sample set

It is constrained according to true faint leakage sample, virtual faint leakage sampleMeet:

Wherein,Indicate virtual faint leakage sample set G^SThe minimum value and maximum value of interior sample； Indicate virtual faint leakage sample set G^SThe minimum detection length of interior sample and maximum detection length；Indicate virtual faint Leak sampleThe minimum value of middle element,Indicate the mean value of any Small leak sample,Constant is represented, Indicate virtual faint leakage sampleThe maximum value of middle element,Constant is represented,σ₂Indicate normal sample The maximum value of data fluctuations；△G^SIndicate the data fluctuations of virtual faint leakage；△ Z indicates that Operating condition adjustment causes data fluctuations Minimum value；

Step 1.4: according to Pearson correlation coefficients ρ and true normal sample collection X^NTo virtual normal sample collection G primary^NIn it is every One virtual normal sample is screened, if screening is qualified, is retained virtual normal sample, if unqualified, is rejected；According to skin Ademilson correlation coefficient ρ and true faint leakage sample set X^STo virtual faint leakage sample set G primary^SEach of it is virtual micro- Weak leakage sample is screened, if screening is qualified, is retained virtual normal sample, if unqualified, is rejected；It obtains virtual normal Sample setWith virtual faint leakage sample set

3. pipeline weak leakage detection method under a kind of condition of small sample according to claim 2, it is characterised in that: described Screening technique is to calculate virtual normal sample primary according to Pearson correlation coefficients formula in step 1.4With true normal sample This collection X^NThe Pearson correlation coefficients ρ of the true normal sample of each interior, if ρ ∈ [0.6,0.8], then it is assumed thatQualification, By virtual normal sample collection G primary^NThe virtual normal sample of each interior calculates one by one according to screening technique；According to Pearson's phase It closes coefficient formula and calculates virtual faint leakage sample primaryWith true faint leakage sample set X^SEach interior is true normal The Pearson correlation coefficients ρ of sample, if ρ ∈ [0.6,0.8], then it is assumed thatQualification, will virtual faint leakage sample set G^SInterior Each virtual faint leakage sample calculates one by one according to screening technique；

The Pearson correlation coefficients formula is as follows:

Wherein, x^UIndicate arbitrary authentic specimen, wherein U=(N, S)；E () is that expectation calculates；Indicate sample x^UIt is equal Value；g^UIt indicates arbitrary and generates sample,Indicate sample g^UMean value；Indicate sample x^UA-th interior of element；It indicates Sample g^UFirst interior of element.

4. pipeline weak leakage detection method under a kind of condition of small sample according to claim 2, it is characterised in that: described Specific step is as follows for step 2:

Step 2.1: respectively to true normal sample collection X^N, true faint leakage sample set X^S, virtual normal sample collection G^N′, it is virtual Faint leakage sample set G^S′7 kinds of statistical natures are extracted, the statistical nature set H={ V of true normal sample is obtained₁、V₂、…、 V_ξ, statistical nature set H '={ V ' of true faint leakage sample₁、V′₂、…、V′_ζ}；The statistical nature of virtual normal sample Set J={ D₁、D₂、…、D_n}；Statistical nature set J '={ D ' of virtual faint leakage sample₁、D′₂、…、D′_q}；Wherein, V_ξ Represent the statistical nature set of ξ true normal samples, V '_ξRepresent the statistical nature collection of ζ true faint leakage samples It closes, D_nRepresent the statistical nature set of n-th of virtual normal sample, D '_qThe statistics for representing q-th of virtual faint leakage sample is special Collection is closed；The statistical nature includes that the maximum pressure extracted in the l period rises information f_MPR, extract the l period in maximum pressure Decline information f_MPD, extract sample peak valley value information f_PV, extract sample coefficient of variation f_CV, extract signal amplitude root letter Cease f_SRA, extract root mean square information f_RMS, extract the set f of the maximum of fitting coefficient, minimum value_FC；

Expression formula is as follows；

Extract l₁Interior sample in time intervalMaximum pressure rise information

Wherein,R-th of sample in representative sample collection R；Wherein R={ X^N、X^S、G^N′、G^S′, r=1,2 ..., θ^R, it is described θ^RTotal sample number in representative sample collection R；Indicate sampleMean value, b_uIndicate sampleU-th of element, Representative sampleInterior element sum；

Extract l₁Interior sample in time intervalMaximum pressure decline information

Extract samplePeak valley value information

Extract sampleThe coefficient of variation

WhereinIndicate sampleVariance；

Extract sampleThe root information of signal amplitude

Extract sampleRoot mean square information

Extract sampleL₂The linear fit coefficient maximum value in period and the set of minimum value

Wherein fit (a_l2) indicate sampleMiddle l₂The linear fit coefficient of a continuous element, l₂∈L；

Step 2.2: respectively to true normal sample collection X^N, true faint leakage sample set X^S, virtual normal sample collection G^N′, it is virtual Faint leakage sample set G^S′Interior each sample extraction symbolism transform characteristics f_ST, obtain the symbolism change of true normal sample Change characteristic set F={ λ₁、λ₂、…、λ_ξ, symbolism transform characteristics set F '={ λ ' of true faint leakage sample₁、λ′₂、…、 λ′_ζ, the symbolism transform characteristics set T={ β of virtual normal sample₁、β₂、…、β_n, the symbolism of virtual faint leakage sample Transform characteristics set T '={ β '₁、β′₂、…、β′_q}；

Extracting method is to sampleCalculate its transition probability matrixSample is converted to obtain by maximum entropy dividing methodAccording toIt obtainsOne-dimensional vectorRepeat the symbolism transform characteristics that this extracting method extracts all samples； Calculate transition probability matrix

Wherein p_ηyElement changes to the probability of state y from state η in expression sample；

It carries outWithIt deforms, wherein p_η+Indicate the probability that element up changes from state η in sample, p_η-Indicate the probability that element changes down from state η in sample；Deforming postscript is

It willIt is transformed into one-dimensional vector

To sample set { X^N、X^S、G^N′、G^S′In each sample execute the feature extraction operation of step 2.2, until completing all samples This feature extraction；

Step 2.3: will be in the symbolism transform characteristics set of the statistical nature set of true normal sample and true normal sample Feature merged, obtain the fusion feature set of true normal sampleIt will be true micro- The statistical nature set of weak leakage sample is melted with the feature in the symbolism transform characteristics set of true faint leakage sample It closes, obtains the fusion feature set of true faint leakage sampleBy virtual normal sample Statistical nature set is merged with the feature in the symbolism transform characteristics set of virtual normal sample, obtains virtual normal sample This fusion feature setBy the statistical nature set and void of virtual faint leakage sample Intend the symbolism transform characteristics set of faint leakage sample

The Fusion Features of n-th of virtual normal sample are as follows:

Step 2.4: obtaining the authentic specimen set of unknown operating statusStep 2.1 is repeated to step Rapid 2.3, obtain the fusion feature set of the authentic specimen of unknown operating status

5. pipeline weak leakage detection method under a kind of condition of small sample according to claim 4, it is characterised in that: described Specific step is as follows for step 3:

Step 3.1: the F obtained according to step 2^XN、F^XS、F^GN、F^GS、And the L ' (X with sample label^N,X^S,G^N′,G^S′) As outputting and inputting for naive Bayesian network pipeline Small leak identification model, the identification of naive Bayesian network pipeline Small leak Model is as follows:

WhereinRepresent the model of maximum accuracy rate；Represent the authentic specimen set X that output has label^*；It is defeated " 0 " represents the sample for having Small leak sample label out, and output " 1 " represents the sample for having normal sample label；

Step 3.2: the F obtained according to step 2^XN、F^XS、F^GN、F^GS、And the L ' (X with sample label^N,X^S,G^N′,G^S′) As outputting and inputting for least square method supporting vector machine pipeline Small leak identification model, least square method supporting vector machine pipeline is small Leakage identification model is as follows:

Technical field

The present invention relates to pipeline weaks under Discussion on Pipe Leakage Detection Technology field more particularly to a kind of a small amount of sample conditions to leak Detection method.

Background technique

Pipeline transportation is the prevailing traffic mode of petroleum resources, and safe and stable operation is of great significance, and is passed to oil gas The accurate detection of pipeline Small leak during defeated is the means for ensureing that one kind of pipe safety stable operation is effective.In recent years Come, machine learning method is used widely in terms of data-driven modeling, but there is also some drawbacks；

In pipeline weak leak detection, it there are problems that following two:

(1) in actual production process, the sample of faint leakage is less, is unable to satisfy machine learning for high-precision modeling Demand, and faint leakage will also tend to lead to more serious consequence, have to accurately detect and handle in time.

(2) feature of faint leakage signal is often more obscure, traditional feature extracting method for feature excavation not It is enough abundant, it is unable to satisfy the demand of the faint leakage identification model of high-precision.

Summary of the invention

The technical problem to be solved by the present invention is in view of the above shortcomings of the prior art, provide a kind of condition of small sample down tube The faint leakage detection method in road, this method excavate faint two side of leakage sample characteristics from the faint leakage sample size of increase and depth Face recognizes model to construct faint leakage, substantially increases the Detection accuracy of pipeline weak leakage, ensures the peace of oil pipeline Row for the national games.

In order to solve the above technical problems, the technical solution used in the present invention is:

The present invention provides pipeline weak leakage detection method under a kind of condition of small sample, includes the following steps:

Step 1: obtaining authentic specimen collection, and virtual sample collection, the authentic specimen Ji Bao are generated according to authentic specimen collection Include true normal sample collection X^NWith true faint leakage sample set X^S, the virtual sample collection includes virtual normal sample collection primary G^NWith virtual faint leakage sample set G primary^S；

Step 2: feature extraction being combined to authentic specimen collection and virtual sample collection, it includes 7 that the assemblage characteristic, which extracts, Kind statistical nature and 1 group code transform characteristics；

Step 3: 7 kinds of statistical natures and 1 group code transform characteristics according to obtained in step 2 use naive Bayesian side Method and least square method supporting vector machine method establish naive Bayesian network pipeline Small leak identification model and least square respectively Support vector machines pipeline Small leak recognizes model, carries out leak detection to pipeline according to identification model.

Specific step is as follows for the step 1:

Step 1.1: obtaining authentic specimen collection X^N, and generated according to the true normal sample in the sample set coarse virtual Normal sample generates coarse virtual normal sample collection

Pass through point of fittings for statistical analysis, data to the sample in true normal sample collection, the evaluation of similitude Analysis finds the data rule of known true normal sample, generates coarse virtual normal sample

Wherein t indicates time, i.e. the detection length of sample, t={ 1,2 ..., L }；k₁Indicate the oblique of true normal sample Rate；ε is data fluctuations caused by various noises in transmission process；Y_O1For the pressure initial value of setting, the value according to oil pipeline just Normal pressure limit setting；

Step 1.2: according to the constraint of normal sample to the data in the coarse virtual normal sample collection generated in step 1.1 It is adjusted, the sample for not meeting constraint condition is rejected, virtual normal sample collection primary is generated

It is wherein constrained according to normal sample, virtual normal sampleMeet:

Wherein,Indicate coarse virtual normal sample G^NThe minimum value and maximum value of interior sample, t (G^N) generation The coarse virtual normal sample X of table^N′Detection length；t_min、t_maxIndicate coarse virtual normal sampleMinimum detection it is long Degree and maximum detection length；Indicate any normal sample X^NThe minimal ripple of middle element,Indicate any normal sample Mean value,Constant is represented, Indicate any normal sample X^NThe maximum fluctuation of middle element,It represents normal Number,σ₁Indicate leakage alarm threshold value, △ Y_minIndicate the minimum value of high reject signal；Indicate normal sample fluctuation Maximum value；

Step 1.3: according to true faint leakage sample constraint to virtual normal sample collection primaryInterior data are adjusted, and the sample for not meeting true faint leakage sample constraint condition is picked It removes, generates virtual faint leakage sample set primary

It is constrained according to true faint leakage sample, virtual faint leakage sampleMeet:

Wherein,Indicate virtual faint leakage sample set G^SThe minimum value and maximum value of interior sample；Indicate virtual faint leakage sample set G^SThe minimum detection length of interior sample and maximum detection length；It indicates Virtual faint leakage sampleThe minimum value of middle element,Indicate the mean value of any Small leak sample,Constant is represented, Indicate virtual faint leakage sampleThe maximum value of middle element,Constant is represented,σ₂Table Show the maximum value of normal sample data fluctuations；△G^SIndicate the data fluctuations of virtual faint leakage；△ Z indicates that Operating condition adjustment causes The minimum value of data fluctuations；

Step 1.4: according to Pearson correlation coefficients ρ and true normal sample collection X^NTo virtual normal sample collection G primary^NIn Each virtual normal sample screened, if screening is qualified, retains virtual normal sample, if unqualified, reject；Root According to Pearson correlation coefficients ρ and true faint leakage sample set X^STo virtual faint leakage sample set G primary^SEach of void Intend faint leakage sample to be screened, if screening is qualified, retains virtual normal sample, if unqualified, reject；It obtains virtual Normal sample collectionWith virtual faint leakage sample set

Screening technique is to calculate virtual normal sample primary according to Pearson correlation coefficients formula in the step 1.4 With true normal sample collection X^NThe Pearson correlation coefficients ρ of the true normal sample of each interior, if ρ ∈ [0.6,0.8], then recognize ForQualification, by virtual normal sample collection G primary^NThe virtual normal sample of each interior calculates one by one according to screening technique；Root Virtual faint leakage sample primary is calculated according to Pearson correlation coefficients formulaWith true faint leakage sample set X^SInterior is each The Pearson correlation coefficients ρ of a true normal sample, if ρ ∈ [0.6,0.8], then it is assumed thatQualification, will virtual faint leakage sample This collection G^SThe virtual faint leakage sample of each interior calculates one by one according to screening technique；

The Pearson correlation coefficients formula is as follows:

Wherein, x^UIndicate arbitrary authentic specimen, wherein U=(N, S)；E () is that expectation calculates；Indicate sample x^U's Mean value；g^UIt indicates arbitrary and generates sample,Indicate sample g^UMean value；Indicate sample x^UA-th interior of element；Table This g of sample^UFirst interior of element.

Specific step is as follows for the step 2:

Step 2.1: respectively to true normal sample collection X^N, true faint leakage sample set X^S, virtual normal sample collection G^N′、 Virtual faint leakage sample set G^S′7 kinds of statistical natures are extracted, the statistical nature set H={ V of true normal sample is obtained₁、 V₂、…、V_ξ, statistical nature set H '={ V ' of true faint leakage sample₁、V′₂、…、V′_ζ}；The statistics of virtual normal sample Characteristic set J={ D₁、D₂、…、D_n}；Statistical nature set J '={ D ' of virtual faint leakage sample₁、D′₂、…、D′_q}；Its In, V_ξRepresent the statistical nature set of ξ true normal samples, V '_ξThe statistics for representing ζ true faint leakage samples is special Collection is closed, D_nRepresent the statistical nature set of n-th of virtual normal sample, D '_qRepresent the system of q-th of virtual faint leakage sample Count characteristic set；The statistical nature includes that the maximum pressure extracted in the l period rises information f_MPR, extract the l period in maximum Pressure declines information f_MPD, extract sample peak valley value information f_PV, extract sample coefficient of variation f_CV, extract signal amplitude side Root information f_SRA, extract root mean square information f_RMS, extract the set f of the maximum of fitting coefficient, minimum value_FC；

Expression formula is as follows；

Extract l₁Interior sample in time intervalMaximum pressure rise information

Wherein,R-th of sample in representative sample collection R；Wherein R={ X^N、X^S、G^N′、G^S′, r=1,2 ..., θ^R, The θ^RTotal sample number in representative sample collection R；Indicate sampleMean value, b_uIndicate sampleU-th of element, Representative sampleInterior element sum；

Extract l₁Interior sample in time intervalMaximum pressure decline information

Extract samplePeak valley value information

Extract sampleThe coefficient of variation

WhereinIndicate sampleVariance；

Extract sampleThe root information of signal amplitude

Extract sampleRoot mean square information

Extract sampleL₂The linear fit coefficient maximum value in period and the set of minimum value

Wherein fit (a_l2) indicate sampleMiddle l₂The linear fit coefficient of a continuous element, l₂∈L；

Step 2.2: respectively to true normal sample collection X^N, true faint leakage sample set X^S, virtual normal sample collection G^N′、 Virtual faint leakage sample set G^S′Interior each sample extraction symbolism transform characteristics f_ST, obtain the symbol of true normal sample Change transform characteristics set F={ λ₁、λ₂、…、λ_ξ, symbolism transform characteristics set F '={ λ ' of true faint leakage sample₁、λ ′₂、…、λ′_ζ, the symbolism transform characteristics set T={ β of virtual normal sample₁、β₂、…、β_n, virtual faint leakage sample Symbolism transform characteristics set T '={ β '₁、β′₂、…、β′_q}；

Extracting method is to sampleCalculate its transition probability matrixSample is converted by maximum entropy dividing method It obtainsAccording toIt obtainsOne-dimensional vectorRepeat the symbolism transformation spy that this extracting method extracts all samples Sign；Calculate transition probability matrix

Wherein p_ηyElement changes to the probability of state y from state η in expression sample；

It carries outWithIt deforms, wherein p_η+Element up changes from state η in expression sample Probability, p_η-Indicate the probability that element changes down from state η in sample；Deforming postscript is

It willIt is transformed into one-dimensional vector

To sample set { X^N、X^S、G^N′、G^S′In each sample execute step 2-2 feature extraction operation, until complete institute There is the feature extraction of sample；

Step 2.3: by the symbolism transform characteristics collection of the statistical nature set of true normal sample and true normal sample Feature in conjunction is merged, and the fusion feature set of true normal sample is obtainedIt will be true Feature in the statistical nature set of real faint leakage sample and the symbolism transform characteristics set of true faint leakage sample into Row fusion obtains the fusion feature set of true faint leakage sampleBy virtual normal sample This statistical nature set is merged with the feature in the symbolism transform characteristics set of virtual normal sample, is obtained virtually just The fusion feature set of normal sampleBy the statistical nature set of virtual faint leakage sample With the symbolism transform characteristics set of virtual faint leakage sampleN-th of virtual normal sample This Fusion Features are as follows:

Step 2.4: obtaining the authentic specimen set of unknown operating statusRepeat step 2.1 To step 2.3, the fusion feature set of the authentic specimen of unknown operating status is obtained

Specific step is as follows for the step 3:

Step 3.1: the F obtained according to step 2^XN、F^XS、F^GN、F^GS、And the L ' (X with sample label^N,X^S,G^N′, G^S′) outputting and inputting as naive Bayesian network pipeline Small leak identification model, naive Bayesian network pipeline Small leak It is as follows to recognize model:

WhereinRepresent the model of maximum accuracy rate；Represent the authentic specimen set that output has label X^*；It exports " 0 " and represents the sample for having Small leak sample label, output " 1 " represents the sample for having normal sample label；

Step 3.2: the F obtained according to step 2^XN、F^XS、F^GN、F^GS、And the L ' (X with sample label^N,X^S,G^N′, G^S′) outputting and inputting as least square method supporting vector machine pipeline Small leak identification model, least square method supporting vector machine pipe It is as follows that road Small leak recognizes model:

The beneficial effects of adopting the technical scheme are that pipeline under a kind of condition of small sample provided by the invention Faint leakage detection method can effectively solve the problems, such as that the faint leakage faced in pipeline transmission process is difficult to detect；This method In assemblage characteristic extracting method can be comprehensive extraction Small leak information so that feature vector for Small leak characterization more Add sufficiently；This method is faint to construct in terms of increasing faint leakage sample size and the faint leakage sample characteristics two of depth excavation Leakage identification model, substantially increases the Detection accuracy of pipeline weak leakage, ensures the safe operation of oil pipeline.

Detailed description of the invention

Fig. 1 is the pipeline weak leak detection flow chart under condition of small sample provided in an embodiment of the present invention；

Fig. 2 is that the assemblage characteristic of sample provided in an embodiment of the present invention extracts flow chart；

Fig. 3 establishes two kinds of faint leakages identification model framework charts to be provided in an embodiment of the present invention；

Fig. 4 is the test result figure provided in an embodiment of the present invention for generating virtual sample；

Fig. 5 is the test result figure of true pipeline weak leakage provided in an embodiment of the present invention, wherein a is simple pattra leaves The test result figure of the true pipeline weak leakage of this network model, b is the true pipeline weak of least square method supporting vector machine model The test result figure of leakage.

Specific embodiment

With reference to the accompanying drawings and examples, specific embodiments of the present invention will be described in further detail.Implement below Example is not intended to limit the scope of the invention for illustrating the present invention.

As shown in Figure 1, the method for the present embodiment is as described below.

Using the pressure information of pipeline in pressure transmitter acquisition oil gas transmission process, and manually choose a small amount of normal sample Originally with faint leakage sample, because faint leakage sample probability of occurrence is low in actual production, quantity is few, and it is accurate to be unable to satisfy foundation The requirement of model.Therefore the present invention provides pipeline weak leakage detection method under a kind of condition of small sample, includes the following steps:

Step 1: obtaining authentic specimen collection, and virtual sample collection, the authentic specimen Ji Bao are generated according to authentic specimen collection Include true normal sample collection X^NWith true faint leakage sample set X^S, the virtual sample collection includes virtual normal sample collection primary G^NWith virtual faint leakage sample set G primary^S；Specific step is as follows:

Step 1.1: obtaining authentic specimen collection X^N, and generated according to the true normal sample in the sample set coarse virtual Normal sample generates coarse virtual normal sample collection

Pass through fittings for statistical analysis, data to the sample in true normal sample collection, the evaluation equal part of similitude Analysis finds the data rule of known true normal sample, generates coarse virtual normal sample

Wherein t indicates time, i.e. the detection length of sample, t={ 1,2 ..., L }；k₁Indicate the oblique of true normal sample Rate；ε is data fluctuations caused by various noises in transmission process；Y_O1For the pressure initial value of setting, the value according to oil pipeline just Normal pressure limit setting；

The analytical mathematics of the sample are as follows: passing through the analysis of normal sample for magnanimity and our domain knowledge (research in terms of long campaigns pipeline leakage testing, the understanding for having comparison deep pipeline pressure data) approximation obtains slightly The rough virtual normal sample similar with true normal sample rule generates model, is containing magnanimity just in our database Why normal sample has first done the generation of normal sample, and reason has at 3 points: the cost of artificial sample is reduced, it is time-consuming, laborious；It tests Demonstrate,prove the validity of generation method；Small leak is all progress derivation on the basis of normal sample.Our analyses for data Include: the statistical analysis of normal sample, the fitting of data, evaluation of similitude etc., finally realize three purposes: obtaining number pressure According to magnitude --- the zone of reasonableness of pipeline pressure i.e. in normal productive process；Pressure with the attenuation coefficient of time, i.e. k1's Value；Pressure oscillation ε caused by noise；

Step 1.2: according to the constraint of normal sample to the data in the coarse virtual normal sample collection generated in step 1.1 It is adjusted, the sample for not meeting constraint condition is rejected, virtual normal sample collection primary is generated

It is wherein constrained according to normal sample, virtual normal sampleMeet:

WhereinIndicate coarse virtual normal sample G^NThe minimum value and maximum value of interior sample, t (G^N) generation The coarse virtual normal sample X of table^N′Detection length；t_min、t_maxIndicate coarse virtual normal sampleMinimum detection it is long Degree and maximum detection length；Indicate any normal sample X^NThe minimal ripple of middle element,Indicate any normal sample Mean value,Constant is represented, Indicate any normal sample X^NThe maximum fluctuation of middle element,It represents normal Number,σ₁Indicate leakage alarm threshold value, △ Y_minIndicate the minimum value of high reject signal；Indicate normal sample fluctuation Maximum value；

Step 1.3: according to true faint leakage sample constraint to virtual normal sample collection primaryInterior data are adjusted, and the sample for not meeting true faint leakage sample constraint condition is picked It removes, generates virtual faint leakage sample set primary

It is constrained according to true faint leakage sample, virtual faint leakage sampleMeet:

Wherein,Indicate virtual faint leakage sample set G^SThe minimum value and maximum value of interior sample；Indicate virtual faint leakage sample set G^SThe minimum detection length of interior sample and maximum detection length；It indicates Virtual faint leakage sampleThe minimum value of middle element,Indicate the mean value of any Small leak sample,Constant is represented, Indicate virtual faint leakage sampleThe maximum value of middle element,Constant is represented,σ₂Table Show that the maximum value of normal sample data fluctuations, so-called normal sample data refer to that front and back that time occurs for Small leak sample Normal sample, fluctuation are less than data fluctuations when Small leak occurs；△G^SIndicate the data fluctuations of virtual faint leakage； △ Z indicates that Operating condition adjustment causes the minimum value of data fluctuations；

Step 1.4: according to Pearson correlation coefficients ρ and true normal sample collection X^NTo virtual normal sample collection G primary^NIn Each virtual normal sample screened, if screening is qualified, retains virtual normal sample, if unqualified, reject；Root According to Pearson correlation coefficients ρ and true faint leakage sample set X^STo virtual faint leakage sample set G primary^SEach of void Intend faint leakage sample to be screened, if screening is qualified, retains virtual normal sample, if unqualified, reject；It is obtained Virtual normal sample collectionWith virtual faint leakage sample setThis The upper limit that the lower limit of ρ is 0.6, ρ in invention is 0.8, and the sample for allowing for generating so had both existed with limited true sample Statistical similitude, and have certain otherness, efficiently solve the small sample problem in engineering.

The screening technique is to calculate virtual normal sample primary according to Pearson correlation coefficients formulaWith it is true normal Sample set X^NThe Pearson correlation coefficients ρ of the true normal sample of each interior, if ρ ∈ [0.6,0.8], then it is assumed thatQualification, By virtual normal sample collection G primary^NThe virtual normal sample of each interior calculates one by one according to screening technique；According to Pearson's phase It closes coefficient formula and calculates virtual faint leakage sample primaryWith true faint leakage sample set X^SEach interior is true normal The Pearson correlation coefficients ρ of sample, if ρ ∈ [0.6,0.8], then it is assumed thatQualification, by virtual faint leakage sample set G primary^S The virtual faint leakage sample of each interior calculates one by one according to screening technique；

The Pearson correlation coefficients formula is as follows:

Wherein, x^UIndicate arbitrary authentic specimen, wherein U=(N, S)；E () is that expectation calculates；Indicate sample x^U's Mean value；g^UIt indicates arbitrary and generates sample,Indicate sample g^UMean value；Indicate sample x^UA-th interior of element；Table This g of sample^UFirst interior of element.

Step 2: feature extraction being combined to authentic specimen collection and virtual sample collection, it includes 7 that the assemblage characteristic, which extracts, Kind statistical nature and 1 group code transform characteristics；As shown in Figure 2, the specific steps are as follows:

Step 2.1: respectively to true normal sample collection X^N, true faint leakage sample set X^S, virtual normal sample collection G^N′、 Virtual faint leakage sample set G^S′7 kinds of statistical natures are extracted, the statistical nature set H={ V of true normal sample is obtained₁、 V₂、…、V_ξ, statistical nature set H '={ V ' of true faint leakage sample₁、V′₂、…、V′_ζ}；The statistics of virtual normal sample Characteristic set J={ D₁、D₂、…、D_n}；Statistical nature set J '={ D ' of virtual faint leakage sample₁、D′₂、…、D′_q}；Its In, V_ξRepresent the statistical nature set of ξ true normal samples, V '_ξThe statistics for representing ζ true faint leakage samples is special Collection is closed, D_nRepresent the statistical nature set of n-th of virtual normal sample, D '_qRepresent the system of q-th of virtual faint leakage sample Count characteristic set；The statistical nature includes that the maximum pressure extracted in the l period rises information f_MPR, extract the l period in maximum Pressure declines information f_MPD, extract sample peak valley value information f_PV, extract sample coefficient of variation f_CV, extract signal amplitude side Root information f_SRA, extract root mean square information f_RMS, extract the set f of the maximum of fitting coefficient, minimum value_FC；

Sample is sufficiently excavated, the expression formula of various features is as follows；

Extract l₁Interior sample in time intervalMaximum pressure rise information

Wherein,R-th of sample in representative sample collection R；Wherein R={ X^N、X^S、G^N′、G^S′, r=1,2 ..., θ^R, The θ^RTotal sample number in representative sample collection R；Indicate sampleMean value, b_uIndicate sampleU-th of element, Representative sampleInterior element sum；

Extract l₁Interior sample in time intervalMaximum pressure decline information

Extract samplePeak valley value information

Extract sampleThe coefficient of variation

WhereinIndicate sampleVariance；

Extract sampleThe root information of signal amplitude

Extract sampleRoot mean square information

Extract sampleL₂The linear fit coefficient maximum value in period and the set of minimum value

Wherein fit (a_l2) indicate sampleMiddle l₂The linear fit coefficient of a continuous element, l₂∈L；

Step 2.2: respectively to true normal sample collection X^N, true faint leakage sample set X^S, virtual normal sample collection G^N′、 Virtual faint leakage sample set G^S′Interior each sample extraction symbolism transform characteristics f_ST, obtain the symbol of true normal sample Change transform characteristics set F={ λ₁、λ₂、…、λ_ξ, symbolism transform characteristics set F '={ λ ' of true faint leakage sample₁、λ ′₂、…、λ′_ζ, the symbolism transform characteristics set T={ β of virtual normal sample₁、β₂、…、β_n, virtual faint leakage sample Symbolism transform characteristics set T '={ β '₁、β′₂、…、β′_q}；

Extracting method is to sampleCalculate its transition probability matrixSample is converted by maximum entropy dividing method It obtainsAccording toIt obtainsOne-dimensional vectorRepeat the symbolism transformation spy that this extracting method extracts all samples Sign, i.e.,WithWhereinIndicate ξ true normal samples,I-th of expression generation is virtual just Normal sample, S_ξ、Indicate the transformation vector of the transformation virtual normal sample of vector sum of true normal sample；Calculate transition probability square Battle array

Wherein p_ηyElement changes to the probability of state y from state η in expression sample；

To solveDimension disaster, carry outWithIt deforms, wherein p_η+It indicates in sample The probability that element up changes from state η, p_η-Indicate the probability that element changes down from state η in sample；Deforming postscript is

Finally, willIt is transformed into one-dimensional vector

To sample set { X^N、X^S、G^N′、G^S′In each sample execute step 2-2 feature extraction operation, realize all samples This feature extraction work.

Step 2.3: by the symbolism transform characteristics collection of the statistical nature set of true normal sample and true normal sample Feature in conjunction is merged, and the fusion feature set of true normal sample is obtainedIt will be true Feature in the statistical nature set of real faint leakage sample and the symbolism transform characteristics set of true faint leakage sample into Row fusion obtains the fusion feature set of true faint leakage sampleBy virtual normal sample This statistical nature set is merged with the feature in the symbolism transform characteristics set of virtual normal sample, is obtained virtually just The fusion feature set of normal sampleBy the statistical nature set of virtual faint leakage sample With the symbolism transform characteristics set of virtual faint leakage sampleIt completes from initial data domain To the transformation of property field.

The Fusion Features of n-th of virtual normal sample are as follows:

Step 2.4: obtaining the authentic specimen set of unknown operating statusRepeat step 2.1 To step 2.3, the fusion feature set of the authentic specimen of unknown operating status is obtained

Step 3: simple pattra leaves is respectively adopted in 7 kinds of statistical natures and 1 group code transform characteristics according to obtained in step 2 This method and least square method supporting vector machine method establish naive Bayesian network pipeline Small leak identification model and minimum respectively Two multiply support vector machines pipeline Small leak identification model, carry out leak detection to pipeline according to identification model, realize for pipeline The accurate detection of Small leak.Because the quantity of Small leak sample is usually less, more accurate Small leak identification mould can not be established Type, therefore traditional leakage detection method can have rate of failing to report high (in the too low situation of leakage alarm sensitivity) or rate of false alarm is high The problem of (in the too high situation of leakage alarm sensitivity), the first job in the present invention is exactly that solve Small leak sample size few The problem of；[2] feature of Small leak sample is unobvious, and traditional feature extracting method not enough fills the excavation of Small leak sample Point, the extraction Small leak information that the assemblage characteristic extracting method in the present invention can be comprehensive, so that feature vector is let out for small The characterization of leakage is more abundant.So establishing accurate Small leak identification model, the accurate recognition for Small leak, accurate inspection are realized It surveys；As shown in figure 3, specifically including:

Step 3.1: the F obtained according to step 2^XN、F^XS、F^GN、F^GS、And the L ' (X with sample label^N,X^S,G^N′, G^S′) outputting and inputting as naive Bayesian network pipeline Small leak identification model, naive Bayesian network pipeline Small leak It is as follows to recognize model (i.e. faint leakage recognizes model 1):

WhereinRepresent the model of maximum accuracy rate；Represent the authentic specimen set that output has label X^*；It exports " 0 " and represents the sample for having Small leak sample label, output " 1 " represents the sample for having normal sample label；

Step 3.2: the F obtained according to step 2^XN、F^XS、F^GN、F^GS、And the L ' (X with sample label^N,X^S,G^N′, G^S′) outputting and inputting as least square method supporting vector machine pipeline Small leak identification model, least square method supporting vector machine pipe It is as follows that road Small leak recognizes model (i.e. faint leakage recognizes model 2):

The reliability of the adjustment model is tested in the present embodiment.Embodiment of the present invention is respectively with the virtual sample of generation and true Real faint leakage sample tests the model in the present invention.

Since the leakage frequency of occurrences faint in actual production process is low, we acquire 40 from historical data and faint let out Sample is leaked, for the balance for keeping positive negative sample, normal sample also chooses 50, with the virtual sample generation side designed in the present invention Method generates 1000 virtual normal samples and 1000 virtual faint leakage samples.Parameter selection is as follows: training sample length L =120, the partitioning parameters N=5 of symbolism transformation when using Nae Bayesianmethod, when using least square method supporting vector machine When symbolism variation partitioning parameters N=4.

As shown in figure 4, test result of the two methods proposed by the present invention in virtual sample, can be found that from result The accuracy rate of model is very high, illustrates effectiveness of the invention.

As shown in figure 5, being when being tested using 80 authentic specimens results, it can be seen that using Nae Bayesianmethod When have 5 samples (No.10, No.22, No.23, No.36 and No.47) by mistaken diagnosis, test accuracy rate 93.75%.Using most Small two when multiplying support vector machines, and by mistaken diagnosis, test accuracy rate is up to only 4 samples (No.10, No.23, No.36 and No.47) 95%.This has absolutely proved method validity designed in the present invention.

Finally, it should be noted that the above embodiments are merely illustrative of the technical solutions of the present invention, rather than its limitations；Although Present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that: it still may be used To modify to technical solution documented by previous embodiment, or some or all of the technical features are equal Replacement；And these are modified or replaceed, model defined by the claims in the present invention that it does not separate the essence of the corresponding technical solution It encloses.