CN103018383A - Oil chromatogram on-line monitoring noise data correction method - Google Patents

Abstract

The invention relates to a correction method for oil chromatogram on-line device noise data. The method comprises the steps of: 1) collecting oil chromatogram off-line test and on-line monitoring data; 2) acquiring an optimal combination of important parameters in a support vector machine regression model by a firefly algorithm; 3) training the support vector machine by the minority accurate oil chromatogram off-line test data obtained in step1) so as to the support vector machine regression model; 4) initializing the allowable deviation radius h of the on-line monitoring data, calculating a piecewise function between off-line tests, and judging whether the data of oil chromatogram on-line monitoring are in an error range allowed by the model; 5) correcting the on-line data; and 6) adjusting the parameters in the model according to a feedback result of on-site data correction. The method for oil chromatogram on-line data correction provided in the invention has the characteristics of stable and accurate effect, short time, and good timeliness, thus being very suitable for correction of on-site oil chromatogram on-line data.

Description

The bearing calibration of a kind of oil chromatography on-line monitoring noise data

Technical field

The invention belongs to converting equipment on-line monitoring technique field, be applied in the noise data trimming process of transformer online monitoring equipment, specifically a kind of oil chromatography on-line monitoring noise data bearing calibration.

Background technology

The transformer oil chromatographic on-line monitoring can in time be grasped the operation conditions of transformer, finds and follow the tracks of the latency fault, for the reliability service of transformer provides safeguard.But because the oil chromatography on-line monitoring is subject to the impact of the factors such as environment temperature, humidity and monitoring equipment self error, may there be distortion in online data, needs to carry out the data school before state evaluation and fault diagnosis.At present, Chinese scholars has been done a large amount of research work and has been proposed some algorithms the data Correction Problems.Principle component regression can be removed noise data effectively, but the error of match is large, and calibration accuracy is low; The neural network algorithm fitting effect is good, but data volume is when large, and the training time is long, has the problem of " crossing study ".

For the poor situation of present oil chromatography online data calibration result, the present invention proposes the method based on the Data correction of firefly support vector machine.At first by the firefly algorithm important parameter that affects the support vector machine performance is optimized.Then calculate the piecewise function between the oil chromatography off-line data, when online data exceeds the scope of piecewise function error permission, think that online data is unusual.Utilize minority accurately the oil chromatography off-line data support vector machine regression model is trained, when by the support vector machine regression model unusual online data is proofreaied and correct when unusual appears in online data.The algorithm that the present invention proposes can be applied among the oil chromatography on-Line Monitor Device, Condition Monitoring Data to transformer oil chromatographic carries out on-line calibration, so that the data of oil chromatography on-Line Monitor Device output true and accurate more, effectively reject the bad point data, under the prerequisite that guarantees data validity, improve availability of data; Also can be applied in the master system of power transmission and transformation equipment state monitoring the data of oil chromatography on-line monitoring be carried out verification, reasonably tell active data, propose the bad point data, the oil chromatography data be played the effect of filtration.The data that this algorithm process of process is crossed can directly apply to fault diagnosis and the state evaluation of the power transmission and transforming equipments such as transformer, thereby directly instruct production run, repair based on condition of component and fault diagnosis work, improve greatly production cost and management level, the industry developments such as advanced state monitoring, repair based on condition of component and intelligent O﹠M produce huge economic benefit and social benefit.

Summary of the invention

In order to solve the above-mentioned problems in the prior art, the invention provides a kind of oil chromatography on-line monitoring noise data bearing calibration that is suitable for transformer.

The present invention has done brand-new design at the algorithm of Data correction, and concrete technical scheme is as follows:

The bearing calibration of a kind of oil chromatography on-line monitoring noise data, its method may further comprise the steps:

Step 1), collection oil chromatography off-line testing and online monitoring data;

Step 2), pass through the optimum combination of important parameter in the supported vector machine regression model of firefly algorithm;

Step 3), utilize minority that step 1) obtains accurately oil chromatography off-line testing data support vector machine is trained supported vector machine regression model;

Step 4), initialization online monitoring data allow to depart from radius h, calculate the piecewise function between off-line testing, judge that the data of oil chromatography on-line monitoring are whether within the error range of model permission; If, think that then online data is normal; Otherwise, think that online data is unusual;

Step 5), online data is proofreaied and correct: judge one by one whether online data normal, if data exception, then by the support vector machine Regression Model Simulator should be constantly data, replace abnormal data with match value; Otherwise, think that the oil chromatography online monitoring data is normal, do not need to proofread and correct;

Step 6), according to the result of field data correction feedback, the parameter in the model is adjusted.

In the above scheme, described step 2) concrete steps in are as follows:

Step 2.1), initialization fluorescein volatility coefficient ρ, fluorescein enhancer γ, sensing range r _s, neighborhood rate of change β, neighbours' threshold value n _t, the position of moving step length s and every firefly;

Step 2.2), need in the support vector machine to determine the span of Optimal Parameters: determine wrong penalty factor, the span of insensitive parameter ε and nuclear parameter σ optimal value;

Step 2.3), in every group of parameter value scope, choose at random a class value as the position of firefly individuality; As fitness, error is less with the error of support vector machine match value and actual value, and performance that then should the group parameter is better, and fitness is larger;

Step 2.4), calculate fluorescein concentration, decision domain scope and the neighbours' of each firefly fluorescein concentration, determine the moving direction of firefly and move forward by neighbours' fluorescein concentration;

Step 2.5), judge whether the firefly algorithm reaches end condition; If reach, then choose the parameter that optimum solution is support vector machine; Otherwise, turn step 2.4).

In the above scheme, described step 2.1) in, because the important parameter that needs to optimize in the support vector machine is wrong penalty factor, insensitive parameter ε and nuclear parameter σ optimal value, colony is expressed as X=((C in the firefly ₁, ε ₁, σ ₁), (C ₂, ε ₂, σ ₂) ..., (C _N, ε _N, σ _N))

In the above scheme, in the described step 4), suppose constantly t _iAnd t _j(t _i≠ t _j) off-line data be respectively y _iAnd y _j, then these two constantly the piecewise function between off-line data be $f_{\mathrm{ij}}\left(t\right)=\frac{y_j-y_i}{t_j-t_i}t+\frac{y_i{t}_j-y_j{t}_i}{t_j-t_i};$

If the radius that piecewise function allows to depart between twice off-line testing is h, t constantly then _iAnd t _jThe upper limit function of online monitoring data is between twice off-line testing

The lower limit function of online monitoring data is

If online data has exceeded the upper limit or lower limit that the off-line data piecewise function allows, think that then online monitoring data is unusual, need to proofread and correct.

In the above scheme, in the described step 6), the error sum of squares of models fitting off-line data is less, and then model is more accurate.

In the above scheme, in the described step 6), the quadratic sum of two data difference that abnormal data corrected value and on-line monitoring are adjacent is less, and the data after then proofreading and correct are got over stable and continuous, and calibration result is better.

In the above scheme, in the described step 6), the training time of model is shorter, and then model is more suitable carries out real time correction to online data.

By the engineering application attestation, the present invention carries out oil chromatography online data calibration result exactly, and the time, short real-time was good, was fit to very much the oil chromatography online data at scene is proofreaied and correct.

Description of drawings

Fig. 1 is the schematic diagram of embodiment of the invention oil chromatography online data means for correcting;

Fig. 2 is the structural drawing that embodiment of the invention oil chromatography online data is proofreaied and correct;

Fig. 3 is the process flow diagram of embodiment of the invention firefly algorithm optimization support vector machine important parameter;

Fig. 4 is the process flow diagram that embodiment of the invention support vector machine regression model is proofreaied and correct online noise data.

Embodiment

The invention will be further described below in conjunction with the drawings and specific embodiments.

Fig. 3 is the process flow diagram of embodiment of the invention firefly algorithm optimization support vector machine important parameter; Fig. 4 is the process flow diagram that embodiment of the invention support vector machine regression model is proofreaied and correct online noise data.Such as Fig. 3, Fig. 4, the bearing calibration of a kind of oil chromatography on-line monitoring noise data, its method may further comprise the steps:

Step 1), collection oil chromatography off-line testing and online monitoring data;

Step 2), pass through the optimum combination of important parameter in the supported vector machine regression model of firefly algorithm;

Step 3), utilize minority that step 1) obtains accurately oil chromatography off-line testing data support vector machine is trained supported vector machine regression model;

Step 4), initialization online monitoring data allow to depart from radius h, calculate the piecewise function between off-line testing, judge that the data of oil chromatography on-line monitoring are whether within the error range of model permission; If, think that then online data is normal; Otherwise, think that online data is unusual;

Step 5), online data is proofreaied and correct: judge one by one whether online data normal, if data exception, then by the support vector machine Regression Model Simulator should be constantly data, replace abnormal data with match value; Otherwise, think that the oil chromatography online monitoring data is normal, do not need to proofread and correct;

Step 6), according to the result of field data correction feedback, the parameter in the model is adjusted.

The one-piece construction of oil chromatography online data correcting algorithm as shown in Figure 2, this algorithm adopts the off-line data verification, proofreaies and correct the thinking of online monitoring data, make up off-line oil chromatography Sample Storehouse by the data message under the various situations of collecting the oil chromatography off-line testing, optimize the parameter of on-line testing algorithm by offline database.The data communication device that on-Line Monitor Device is collected is crossed the online data checking algorithm to data analysis, algorithm is not done special processing to normal data, algorithm is proofreaied and correct irrational noise data of on-line equipment, and the result that then will proofread and correct sends into follow-up senior application relevant data analysis and diagnosis.

The principle of oil chromatography online data means for correcting is the module of online data correction as shown in Figure 1 in the frame, it obtains data by data acquisition module, and data results is turned back to display interface, thereby the guide field personnel carry out work.Affect comparatively accurately the important parameter of performance in the off-line data Support Vector Machines Optimized, then to the model training behind the Optimal Parameters, when judging that by piecewise function online data is unusual, be somebody's turn to do data constantly by the support vector machine Regression Model Simulator, replace abnormal data with match value, normally then do not proofread and correct if judge online data.

The algorithm that the oil chromatography online data is proofreaied and correct is described in detail, and algorithm mainly is divided into two parts.First is by the optimum combination of important parameter in the supported vector machine regression model of firefly algorithm, and as shown in Figure 3, the concrete steps of this some algorithm are as follows:

Step 1: the parameter of initialization firefly algorithm: initialization fluorescein volatility coefficient ρ, fluorescein enhancer γ, sensing range r _s, neighborhood rate of change β, neighbours' threshold value n _t, the position of moving step length s and every firefly;

Step 2: span that need in the support vector machine to determine Optimal Parameters; Determine wrong penalty factor, the span of insensitive parameter ε and nuclear parameter σ optimal value;

Step 3: in every group of parameter value scope, choose at random a class value as the position of firefly individuality.As fitness, error is less with the error of support vector machine match value and actual value, and performance that then should the group parameter is better, and fitness is larger;

Step 4: calculate fluorescein concentration, decision domain scope and the neighbours' of each firefly fluorescein concentration, determine the moving direction of firefly and move forward by neighbours' fluorescein concentration;

Step 5: judge whether the firefly algorithm reaches end condition; If reach, then choose the parameter that optimum solution is support vector machine; Otherwise, turn step 4).

Because the important parameter that needs to optimize in the support vector machine is wrong penalty factor, insensitive parameter ε and nuclear parameter σ optimal value, so colony is expressed as follows in the firefly:

X＝((C ₁,ε ₁,σ ₁),(C ₂,ε ₂,σ ₂),...,(C _N,ε _N,σ _N))……(1)。

The fluorescein concentration of i firefly is as follows:

l _i(t)＝(1-ρ)l _i(t-1)+γJ(x _i(t))……(2)，

Decision domain scope more new formula is as follows:

$r_d^i(t+1)=\min \{r_s,\max \{0,r_d^i\left(t\right)+\mathrm{\β}(n_t-|N_i\left(t\right)\left|\right)\left\}\right\}......\left(3\right),$

The t moment neighborhood of i firefly is as follows:

$N_i\left(t\right)=\{j:||x_j\left(t\right)-x_i\left(t\right)||<r_d^i\left(t\right);1_i\left(t\right)<1_j\left(t\right)\}......\left(4\right).$

Firefly is decided its moving direction according to the fluorescein concentration of each firefly in its neighborhood in motion process, the probability that moves of i firefly j firefly in its neighborhood is as follows constantly for t:

$P_{\mathrm{ij}}\left(t\right)=\frac{1_j\left(t\right)-1_i\left(t\right)}{{\mathrm{\&Sigma;}}_{k\&Element;N_i\left(t\right)}{1}_k\left(t\right)-1_i\left(t\right)}......\left(5\right).$

I firefly is as follows in t+1 position constantly:

$x_i(t+1)=x_i\left(t\right)+s\frac{x_j\left(t\right)-x_i\left(t\right)}{\left|\right|x_j\left(t\right)-x_i\left(t\right)\left|\right|}......\left(6\right).$ Algorithm withdraws from circulation when firefly algorithm end condition being set being iterations and surpassing 1000 times.

Second portion is proofreaied and correct irrational noise data by the support vector machine regression model that trains, and as shown in Figure 4, the concrete steps of this some algorithm are as follows:

Step 1: collect oil chromatography off-line testing and online monitoring data;

Step 2: the parameter of Support Vector Machines Optimized regression model.Obtain the optimum combination of important parameter in the regression model by the firefly algorithm;

Step 3: utilize minority accurately oil chromatography off-line testing data support vector machine is trained supported vector machine regression model;

Step 4: the initialization online monitoring data allow to depart from radius h, calculate the piecewise function between off-line testing, judge that the data of oil chromatography on-line monitoring are whether within the error range of model permission.If, think that then online data is normal; Otherwise, think that online data is unusual;

Step 5: online data is proofreaied and correct.Judge one by one whether online data is normal, if data exception then is somebody's turn to do data constantly by the support vector machine Regression Model Simulator, replace abnormal data with match value; Otherwise, think that the oil chromatography online monitoring data is normal, do not need to proofread and correct;

Step 6: according to the result of field data correction feedback, the parameter in the model is adjusted, so that the method among the present invention is to the better effects if of the online noise compensation of oil chromatography.

It is as follows that support vector machine returns the function of estimating:

F (x)=ω φ (x)+b ... (7), the support vector machine employing minimizes structure and determines ω and b, namely

${\min R}_{\mathrm{str}}=\frac{1}{2}{\left|\right|\mathrm{\&omega;}\left|\right|}^2+{\mathrm{CR}}_{\mathrm{emp}}......\left(8\right),$

Wherein $R_{\mathrm{emp}}=\frac{1}{l}\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}{L}_{\mathrm{\&epsiv;}}(x_i,y_i-f\left(x_i\right))......\left(9\right)$ And L _ε(x, y-f (x))=max{0, | y-f (x)-ε | ... (10), it is as follows to find the solution the function that formula (8) finally can supported vector machine regression model:

$f\left(x\right)=\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}({\mathrm{\&alpha;}}_i^{*}-{\mathrm{\&alpha;}}_i)K(x_i,x)+b......\left(11\right),$

Select the most general radial basis RBF kernel function of usable range as follows in the regression model:

$K(x_i,x)=\exp [-\frac{{\left|\right|x_i-x\left|\right|}^2}{2{\mathrm{\&sigma;}}^2}]......\left(12\right).$

Judge by the piecewise function of off-line data whether online data is unusual, suppose constantly t _iAnd t _j(t _i≠ t _j) off-line data be respectively y _iAnd y _j, then these two constantly the piecewise function between off-line data be $f_{\mathrm{ij}}\left(t\right)=\frac{y_j-y_i}{t_j-t_i}t+\frac{y_i{t}_j-y_j{t}_i}{t_j-t_i}......\left(13\right).$ Suppose that the radius that piecewise function allows to depart between twice off-line testing is h, then t constantly _iAnd t _jThe upper limit function of online monitoring data is between twice off-line testing

The lower limit function of online monitoring data is

If online data has exceeded the upper limit or lower limit that the off-line data piecewise function allows, think that then online monitoring data is unusual, need to proofread and correct.

The error sum of squares of models fitting off-line data is less, and then model is more accurate.The quadratic sum of two data difference that abnormal data corrected value and on-line monitoring are adjacent is less, and the data after then proofreading and correct are got over stable and continuous, and calibration result is better.The training time of model is shorter, and then model is more suitable carries out real time correction to online data.By the engineering application attestation, the present invention carries out oil chromatography online data calibration result exactly, and the time, short real-time was good, was fit to very much the oil chromatography online data at scene is proofreaied and correct.

Claims (4)

1. oil chromatography on-line monitoring noise data bearing calibration, it is characterized in that: its method may further comprise the steps:

Step 1), collection oil chromatography off-line testing and online monitoring data;

Step 2), pass through the optimum combination of important parameter in the supported vector machine regression model of firefly algorithm;

Step 3), utilize minority that step 1) obtains accurately oil chromatography off-line testing data support vector machine is trained supported vector machine regression model;

Step 4), initialization online monitoring data allow to depart from radius h, calculate the piecewise function between off-line testing, judge that the data of oil chromatography on-line monitoring are whether within the error range of model permission; If, think that then online data is normal; Otherwise, think that online data is unusual;

Step 5), online data is proofreaied and correct: judge one by one whether online data normal, if data exception, then by the support vector machine Regression Model Simulator should be constantly data, replace abnormal data with match value; Otherwise, think that the oil chromatography online monitoring data is normal, do not need to proofread and correct;

Step 6), according to the result of field data correction feedback, the parameter in the model is adjusted.

2. oil chromatography on-line monitoring noise data as claimed in claim 1 bearing calibration, it is characterized in that: the concrete steps described step 2) are as follows:

Step 2.1), initialization fluorescein volatility coefficient ρ, fluorescein enhancer γ, sensing range r _s, neighborhood rate of change β, neighbours' threshold value n _t, the position of moving step length s and every firefly;

Step 2.2), need in the support vector machine to determine the span of Optimal Parameters: determine wrong penalty factor, the span of insensitive parameter ε and nuclear parameter σ optimal value;

Step 2.3), in every group of parameter value scope, choose at random a class value as the position of firefly individuality; With the error of support vector machine match value and actual value as fitness;

Step 2.4), calculate fluorescein concentration, decision domain scope and the neighbours' of each firefly fluorescein concentration, determine the moving direction of firefly and move forward by neighbours' fluorescein concentration;

Step 2.5), judge whether the firefly algorithm reaches end condition; If reach, then choose the parameter that optimum solution is support vector machine; Otherwise, turn step 2.4).

3. oil chromatography on-line monitoring noise data as claimed in claim 2 bearing calibration, it is characterized in that: described step 2.1), because the important parameter that needs to optimize in the support vector machine is wrong penalty factor, insensitive parameter ε and nuclear parameter σ optimal value, colony is expressed as X=((C in the firefly ₁, ε ₁, σ ₁), (C ₂, ε ₂, σ ₂) ..., (C _N, ε _N, σ _N)).

4. oil chromatography on-line monitoring noise data as claimed in claim 1 bearing calibration is characterized in that: in the described step 4), suppose constantly t _iAnd t _j(t _i≠ t _j) off-line data be respectively y _iAnd y _j, then these two constantly the piecewise function between off-line data be $f_{\mathrm{ij}}\left(t\right)=\frac{y_{j-}{y}_i}{t_j-t_i}t+\frac{y_i{t}_j-y_j{t}_i}{t_j-t_i};$

If the radius that piecewise function allows to depart between twice off-line testing is h, t constantly then _iAnd t _jThe upper limit function of online monitoring data is between twice off-line testing

The lower limit function of online monitoring data is If online data has exceeded the upper limit or lower limit that the off-line data piecewise function allows, think that then online monitoring data is unusual, need to proofread and correct.

Images