You are on page 1of 4

2010 International Conference on Digital Manufacturing & Automation

Real-time Data Mining in Magnetic Flux Leakage Detecting in Boiler Pipeline

KeMinYi[1],LiaoPan[1]
Hubei University of Technology School of Computer Science and Technology Wuhan, China e-mail: kmy0095@sina.com e-mail: liaopan_117@yahoo.com.
AbstractFor boiler in magnetic flux leakage testing data characteristics on the basis of full analysis, Combining the application of industrial control integrated automation needs, proposed the pipeline magnetic flux leakage testing data mining system framework. Through analysis of magnetic flux leakage pipeline inspection data and mines the key data. It could be better to achieve detection and prediction of the pipe flaw. Keywords-Magnetic flux leakage ;Real-time; Time series; Data mining

SongXiaoChun[2]
Hubei University of Technology School of Mechanical Engineering Wuhan, China e-mail: songxc@mail.tsinghua.edu.cn

I.

INTRODUCTION

With the pipeline toward the rapid detection, automation and intelligent direction, magnetic flux leakage signal processing method of increasing the reliability requirements. On neural networks, pattern recognition, data mining and other new methods of research shows that they can make up the traditional method of magnetic flux leakage signal processing a number of deficiencies, which can improve the detection accuracy and speed. Now, these new theories and new methods applied to magnetic flux leakage test has become a research hotspot. A. Pipeline defect detection and shortcomings The defect identification process is as follows: First determine the detection signal detected signal is the horizontal or vertical detection signal, and then filtering the pre-processed detection signal, compare the absolute peak detection signal and cutting level, if the absolute peak detection signal level is less than cutting, you can determine the defect-free. If the detection signal is greater than the absolute peak cut level, then continues to compare the absolute peak level of size and defect, if defect is less than the absolute peak level, but also determine the defect. However, if the defect is greater than the absolute peak level, then determine the existence of a defect, then we must determine within the outer defect or defects. If the signal wave detection limit but less than generous in the pulse width determined the internal and external defects, defects are detected defects judged to be outside. If the signal wave detecting large defects in the internal and external pulse width determine the width and less than the ceiling, then the defect is detected within the defect is judged. Through the above two-dimensional magnetic flux leakage testing system identification technique of defect can be seen, the current magnetic flux leakage testing equipment for the presence or absence of defects, deficiencies or defects within
978-0-7695-4286-7/10 $26.00 2010 IEEE DOI 10.1109/ICDMA.2010.243 130 131

the defects outside the judge already has a comprehensive theoretical and technology. As for the type of defect identification and defect pipe size performance study is currently mainly used in a large number of test data, That large quantities of standard defect signal characteristic value, and establish standards for defect signal characteristics of the database, by comparing the defect detection signal and the signal characteristics of the standard values of the defect identification, the defect type and the size determine the parameters. As the relatively cumbersome process of artificial defects, and different materials need different pipe defect detection signal template, the workload too much. In order to warranty defects evaluation of reliability and accuracy of template matching method requires a lot of testing to collect test data and to establish a system of signal characteristics of effective defect detection algorithm. Therefore, it is the type of defect identification and the defect size parameter quantitative identification of the intelligence research has become a hot field of magnetic flux leakage testing. B. Pipeline defect detection real-time data mining In recent years, data mining technology has been in business, finance, management, industry and other widely used in many areas and have made significant progress, but mostly limited to business, business information. In the industrial control, application of the production process is still relatively small. Currently, In the industrial control process, As various kinds of advanced equipment and engineering technology widely used ,generation and accumulation of various types of historical data., and the current production of real-time dynamic data. In these massive production data contained a large amount of information .To manage and control .In order to hide in the mass of data useful, in-depth knowledge and information dug up, And extract its overall characteristics, correlation and prediction of development trends. Based on the characteristics of industrial control production data mining technology to solve industrial control of the vast amounts of information, one of the key technologies data processed. This article embarks based on the pipeline leakage examination gathering data's characteristic and the analysis demand, the establishment of a general boiler pipe, according to the number of magnetic flux leakage testing data mining architecture. And analysis of magnetic flux leakage pipeline inspection data mining boiler of the key issues, Discusses the data mining method in pipeline applications.

II.

ARMA MODEL OVERVIEW

Box and Jenkins propose the ARMA model, and used for stationary time series modeling. In many fields has been extensively applied, often used to predict. ARMA model is the AR model and MA model of integrated, Describe their state of the system memory of the past and the system noise into the system over the last time the memory. General description of the following forms of ARMA( p, q ) ( B ) = 1 1 B p B p (1)
(B) = 11B q Bq (2)
Y = { yt | t = 0,1.... p} Is stationary time series, Y at t time step not only with the various values ( yt 1, yt 2 ....., yt p ) of

end of the scanning databases to save, But in the offline analysis, the system is realized data replays showed, defect recognition and quantitative analysis and generate test report etc. Pipeline magnetic flux leakage detection based on historical data. Use of data mining techniques and database technology to combine, through the historical data and real-time data analysis and mining to identify hidden in the pattern of relationships within these data. The use of multi-sensor data fusion method, the voltage signal data sampling, the amount of extract their characteristics so as to better achieve the magnetic flux leakage inspection of pipelines in real-time monitoring, predictive control.

the past, But also with each step past q interference ( t 1, t 2 , , t q ), Where { t } is the white noise sequence immediately and not related to the past time sequence y k (k < t ) , ~ N (0, 2 ) is the introduction of the delay operator, t
ARMA( p, q )

about
p B
p

to

be ( B) yt = ( B) t .

In

Formula,

(B) = 1 1B

for the p-order autoregressive linear


Figure 1 System Framework Chart

coefficient, ( B ) = 1 1B q B q for the q -order moving average coefficient of polynomial, When p is 0, ARMA transfer p -order AR model, When q is 0, ARMA model transfer MA
AR( p) = ARMA( p,0)

IV.

REAL-TIME DATA MINING PIPELINE MAGNETIC FLUX LEAKAGE

MA(q) = ARMA(0, q) 5

and MA(q ) is actually a special case of ARMA , statistical properties of AR and MA is the combination of the statistical characteristics, ARMA both the nature of AR and MA . In the application of ARMA models, stationary time series under the autocorrelation function, partial autocorrelation function of the properties to determine if the truncated autocorrelation function is used MA(q ) , if the partial autocorrelation function of truncated use the AR ( p ) .If the autocorrelation function and partial autocorrelation function of the trailing tail are used ARMA( p, q) . Selected parameter estimation maximum likelihood estimation of a law, the determination of the order is to determine the number of parameters, often used the AIC criteria, the final model is used for testing the models are appropriate.
AR( p)
ARMA

III.

FLUX LEAKAGE DETECTING SYSTEM ARCHIITECUTRE

Boiler Pipeline Magnetic Flux Leakage inspection system mainly by the magnetic flux leakage detection devices (including magnetic devices and integrated Hall elements), Signal pre-processing unit, Signal Acquisition Unit, Motion control unit and data acquisition and analysis and processing software, etc. Among them, magnetic flux leakage testing unit includes three semicircular, each detector probe by 8 SS495A1 type of integration, Side by side, such arrangement between 24 channel sensor array, TMS320 based on DSP chip development VC5509A data acquisition unit under the wheel in mileage trigger and interval samples, And through the USB port will be collected data real-time transmission to the PC, In order to realize dynamic data display and analysis, In the testing process, the main software system to read data from the USB port, real-time display sampling data and the data at the

Time series can be established from the regression model, moving average models, autoregressive moving average model, and so. Autoregressive model AR for a unilateral test data, and such data are characterized by large fluctuations do not occur in the case of the data there is a certain monotony. Moving average model ( MA ) is mainly used for the treatment of white noise curve, this model and the multi-regression model used for self-complementary and fine-tuning. Autoregressive moving average model ( ARMA ) is a combination of the above two kinds of mathematical models which combines the autoregressive grasp the general direction of the main features of curves at the same time .With moving average as a basis for fine-tuning curve. This article mainly aims at different voltage in the pipeline in the normal voltage. So, from a theoretical point of view, should be chosen autoregressive moving average model as a magnetic flux leakage pipeline inspection data in the voltage time series data mining algorithms. Under normal circumstances, against a group of dynamic real-time data for the N models with the ARMA to identify and fit, the whole process is usually divided into the following steps:1)dynamic data pre-processing;2)Preliminary assessment of model parameters; 3)Precise assessment of model parameters; 4) Model Selection and Design Parameters; 5) Comparison of time series prediction. We will be used for Pipeline Magnetic Flux Leakage data collected examples of data mining. A. Real-time data pre-processing ARMA model describing the data to meet for smooth zero mean condition data fitting such model is generally required smoothly before and zero mean treatment, these work collectively referred to as preprocess. This collection of data stored in ACCESS .We format the data in the database, in which the peak voltage includes the following fields: (Channel-number(CH),effect-depth(DEP),voltage(VOLT),time

131 132

points),in this data, time is by ten seconds to update, because it is the voltage value of the excavation. Therefore, the need to retain the time, ensuring a certain collection points of the voltage peak ,adopt the SQL sentence to abstract out the time collecting point and the voltage peak value data, we make use of ARMA to build a model, The voltage change shown in Figure 2. Figure 2 can be found from this time sequence is not a smooth time-series, X-axis represents time, the collection point, Y axis represents voltage amplitude value (the figure Volt curve). As can be seen for some time before the sequence of voltage fluctuation is smaller than in the defect occurred at the voltage peak.

coefficient, they are the determination using the ARMA model time series whether to achieve the steady sequence the key parameter and the weight target, Based on the magnetic flux leakage of real-time data from the mathematical statistics modeling and analysis, finite curve of the third-order Autocorrelation (from) and transportation Correlation(Partial Correlation) good convergence in a first and second order differential curve, And less computational complexity, Based on the above characteristics. The choice of the form of the model, To make full use of data points of the map data of physical laws and intuitive information contained in the background, So the comprehensive consideration chooses third-order difference, namely parameter d = 2. Selecting parameters, we still need to model for judgment, we choose the AIC (red pool information standards), Be within certain order number range, The combination of the order, the numerical information AIC minimum order of the corresponding model of order, Under this exponent number estimated results in model senate for model parameter best estimated value, Its formula is:

Volti =1+0.0427Volti-1+0.029(Volti-2 +DEP*Volti-1+6.0135*10-2


Figure 2 Original signal graph

B. Preliminary assessment of model parameters Preliminary estimates of model parameters and functions with two meanings, first, that a rough estimate, but the form of deductions and apply the formula, it is estimated based on some kind of criteria is neither the result nor the guarantee limit theory. Second, this estimate is estimated as the initial starting point for further fine, which as a further iteration of the initial value. The initial estimate used moment method and inverse function method, but in practice, we often bypass the first stage is estimated to model the smooth reversible within a point as the initial value, directly into the refined estimate Diego generation, this model-based theory of guaranteed stability. C. Precise assessment of model parameters Precise estimates of model parameters (referred to as refined estimate) can generally be in accordance with the least squares or maximum likelihood criteria were. And has approximate maximum likelihood estimation is equivalent to least squares estimation, part of the calculation involves more specialized, thus omitted. D. Model Selection and Design Parameters According to the data features of the map we use the choose ARIMA( p, d , q ) model fit the data to predict. The procedure is as follows: stationary of time series processing. There are many ways of smoothing processing, the most commonly used are first-order differential, second-order differential and third-order differential, For the time series of the difference frequency, the parameter d is usually taken 0,1,2. The model depends on identifying key associated with the partial correlation analysis of, before carrying out analysis on the time series, answer the sampling data first taking a logarithm,purpose is to remove the different method difference there existing in possibility in the data, And then analyze their relevance, At last gets number reason statistics result amounts by differences way. In the mathematical statistic measurement, we usually defined the autocorrelation(AC) that is the autocorrelation coefficient, Partial Cor-relation is the partial correlation

The current time is Volti1 peak voltage, Volti2 is the smallest unit of time coordinates a moment ago peak voltage 1 + 0.0427 *Volti1 is the autoregressive polynomial, 0.029(Volti 2 + DEP *Volti 1 ) + 6.0135 *102 is a part of the moving average MA , DEP is the depth of magnetic flux leakage pipeline defects, and finally the residual Namely, the impact of other factors to consider factors, This depth of the main flaws, the variable i for different time points of sampling voltage. E. Model testing and improvement After treatment of the above steps, we have the dynamic data (Figure 3) has achieved the model fitting, but the whole process of data processing is not at this point, through the model values and spectral density of the structural analysis model may allow further simplification can be justified, by fitting the measured data related to inspection and forecast accuracy of the calculation of, could perhaps be found to model the need for further improvements or other means must be used to simulate for example, the data do not fit a linear model, Should be implemented non-linear fitting. In short, through the testing and improvement of the model is expected to finally get a good model for, using the mathematical model to fit the experimental data on the sample, Voltage value to be fitted curve Figure 3.

Figure 3 Model selection signal graph

Through the research and analysis of the data mining technology, applied to pipe data from a new perspective of pipeline Materials processing, and using the algorithm for time series data digging pipeline, compared with the use of traditional statistical methods, mining speed can be improved, and has long been troubling us to solve the face the piles of data to ignore overwhelming, the situation, combined with
132 133

other data model and method for evaluating model and the continuous improvement, resulting in an ideal prediction model, also further proved association rules in data analysis is effective and practical. F. Comparison of time series prediction Based on the above mathematical model of the voltage curves of the peak to predict ,axis scale is now forecast to 50 units as a unit peak voltage data curves have the following Figure 4.

REFERENCE
[1] Song XiaChun,Zhang HaiBin. Key Technologies of the Boiler Tubes MFL Inspection System Based on LabVIEW, Hubei Industrial university journal,2007,(9).65-68.(in Chinese). Funkhouser H G,historical. Development of the graphical representation of statistical data, The American Mathematical Monthly, 1938, 45(8): 541542. Fayyad U, Piatetsky S G, Smyth P, Knowledge discovery and data mining towards a unifying framework, Proceedings of the Second International Conference on Knowledge Discovery and Data Mining, Oregon, Portland, 1996, 8288. Zaiane O R, Han J, Zhu H, Mining recurrent items in multimedia with progressive resolution refinement, Proceedings of International Conference on Data Engineering, California, U.S.A., 2000, 2,46147. Srivastava J, Cooley R, Deshpande M, Web usage mining: Discovery and application of usage paterns from web data, SIG KDD Explorations, 2000, 1(2): 1223. R.Lippmann, D. Fried, I. Graf, J. Haines, K. Kendall,D. McClung, D. Weber, S. Webster, D. Wyschogrod,R. Cunninghan, and M. Zissman. Evaluating intrusion detection systems: The 1998 darpa off-line intrusion detection evaluation. In Proceedings of the 2000 DARPA Information Survivability Conference and Exposition, January 2000.

[2] [3]

[4] [5] [6]

Figure 4 Model transformation forecast graph

Model selection with Figure 3 and Figure 4 shows the voltage prediction model transformation can be seen, although the data mining prior to the corresponding real-time data were related to pre-treatment, but the detection of waveform still contains a lot of noise. But through the model, parameter selection, a smaller scale, its amplitude will increase with the scale decreases, adjacent scale values of the local modulus maximum is almost in the same location, and has the same symbols. Objectively speaking, applying the mathematical model predicted the peak voltage of the future able to achieve certain effects. The analysis, forecasting curve and actual similarity of Pipeline Magnetic Flux Leakage curve has reached a basic engineering and safety requirements, can be used as magnetic flux leakage pipeline inspection reference for safe production. V. CONCLUSIONS All above has introduced time series data application in excavating in pipeline flux leakage, to use the classic time series models and algorithms, through the mass pipeline magnetic flux leakage data extraction and analysis of real-time, And thus the establishment of the relevant data model, based on model-based qualitative analysis to achieve better prediction and detection of pipeline defects. And by this establishment related data model, based on model foundation in qualitative analysis, it is better to achieve detection and prediction of the pipe flaw. VI. ACKNOWLEDGMENT This work was supported by the National Natural Science Foundation of China, under grant 50875077; and was supported by the Natural Science Foundation of Hubei Province, under grant 2008CDA022.

133 134

You might also like