Research ArticleModeling and Calculation of Dent Based onPipeline Bending Strain

Qingshan Feng12 Rui Li23 and Hong Zhang1

1China University of Petroleum Beijing 102249 China2Petrochina Pipeline Company Langfang 065000 China3School of Automation Science and Electrical Engineering Beihang University Beijing 100191 China

Correspondence should be addressed to Qingshan Feng qsfengpetrochinacomcn

Received 27 January 2016 Revised 10 March 2016 Accepted 3 April 2016

Academic Editor Guiyun Tian

Copyright copy 2016 Qingshan Feng et alThis is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

The bending strain of long-distance oil and gas pipelines can be calculated by the in-line inspection tool which used inertialmeasurement unit (IMU) The bending strain is used to evaluate the strain and displacement of the pipeline During the bendingstrain inspection the dent existing in the pipeline can affect the bending strain data as well This paper presents a novel methodto model and calculate the pipeline dent based on the bending strain The technique takes inertial mapping data from in-lineinspection and calculates depth of dent in the pipeline using Bayesian statistical theory and neural network To verify accuracy ofthe proposedmethod an in-line inspection tool is used to inspect pipeline to gather dataThe calculation of dent shows themethodis accurate for the dent and the mean relative error is 244The newmethod provides not only strain of the pipeline dent but alsothe depth of dent It is more benefit for integrity management of pipeline for the safety of the pipeline

1 Introduction

With the development of the oil and gas production the long-distance buried pipeline is used to the transport the produc-tion of oil and gas [1 2] Due to the reasons of time accu-mulated or construction defects such as corrosion gougedent and displacement seriously threat the safety operationof the pipeline In order to reduce the risk of these defectsthe pipeline company usually selects in-line inspection toolto inspect the defects for body of pipeline The magnetic fluxleakage (MFL) and ultrasonic tool are used to inspect metalloss of pipeline such as corrosion crack and gouge

Dent is another common defect in long-distance buriedoil and gas pipeline A lot of dents have been found inthe pipeline due to the construction of mechanical damageDents can deform the pipeline and affect the integrity ofbody for the pipeline [3ndash6] Specifically the dent locatedin the weld of pipeline or the dent that is very sharp canlead to the serious leak consequence The information aboutthe severity of the damage to a pipe due to dents is veryimportant for the pipeline industry The defect of dent is the

radial deformations of the pipe wall Generally these dentscan be measured by in-line inspection tools named geometrytool which equipped with instrumentation and mechanicalfingers is able to provide data on the pipe wall geometrydeformationHowever this tool only inspects the size (lengthdepth and width) of the dent In order to assess the hazardfor the dent the technical requires a complex modeling andcomputes the strain or other information Another methodfor dent testing can be using the profile gauge to measure thedepth width and length of the gouge or scratch damageThismethod also can be used to determine the extent of damageand determine if any sharp edges are present Recently 3Dlaser scanner is used to map and measure the dent Thistechnology can more accurately measure the dimensions ofthe dent by the laser scan However the method for denttesting of profile gauge and 3D laser scan only use dentswhichhave been dug They are the verification for dents which areinspected by geometry tool

Nowadays another Pipeline In-Line Inspection toolwhich installed inertial measurement unit (IMU) is usedto inspect the pipeline [7ndash10] As shown in Figure 1 this

Hindawi Publishing CorporationJournal of SensorsVolume 2016 Article ID 8126214 8 pageshttpdxdoiorg10115520168126214

2 Journal of Sensors

Figure 1 Measurement system of in-line inspection of pipelinecenterline

tool which consists of a Data Acquisition System (DAS)an IMU some weld detectors and odometers is driven tomove through the pipeline and collect data regarding dentsbends and navigation Petrochina Pipeline Company hasdeveloped an IMU system based on a navigation systemwhich is suitable for a geometry PIG The IMU can be usedto record the attitude data for the tool during the inspectionA method of calculation for pipeline centerline is proposedby fusion for multisensors Czyz et al [11 12] developed apure geometry-based algorithm for the bending strain of apipeline Another pipeline strain testing method is using thestrain gauge to be installed on the surface of pipeline Twotypes of strain gauge resistance and optical fiber are used tomeasure the strain of pipelineThe strain gauge of optical fiberis more precise for measuring the strain These two methodscan be used to monitor the strain changes of the pipeline Butthe limit for these two methods is only used to measure thepart sections of pipelineThey cannot inspect thewhole strainfor long-distance buried pipeline

The bending strain can be reported for the whole pipelinewhen the tool runs for once Generally the technical just usesthe bending strain to detect the position of pipeline enduringthe extra strain larger than the operation requirement andassessing the displacement for local pipeline However itis found that the bending strain can be affected by thedent [13ndash16] In this paper a method for modeling andcomputation of dent which is based on bending strain isproposedThe technique takes Bayesian statistical theory andneural network to improve a new method to compute thedepth of the dent It is more useful to assess the dent for thepipeline integrity

2 The Calculation Method of the PipelineBending Strain

In this section the calculation method of the pipelinebending strain is described Several factors such as internalpressure temperature differential external loads and bound-ary conditions are affecting the strain The strain can beconsisted of two primary components longitudinal and hoopstrain [17ndash19] The bending strain is one of the componentsfor longitudinal strain It can be described by the followingformula

120576 =

119863

2

119896

120576V =119863

2

119896V

120576ℎ =

119863

2

119896ℎ

(1)

z (zenith)

y(north)

x (east)

Pitch

Azimuth

ty

t

tx

tz

Figure 2 Pitch (119875) and azimuth (119860) for the pipeline centerline

where 120576 is the total bending strain 119896 is the total curvature 119896Vis vertical curvature and 119896ℎ is horizontal curvature

The total curvature of the centerline of a pipe is describedat each point along the pipeline by the curvature vector Inorder to calculate the pipeline curvature the centerline of apipe is considered as a 3D parametric curve described in aCartesian system by a vector V(119904) which is a function of adistance (119904) along the curve [20 21]

V (119904) = [119909 (119904) 119910 (119904) 119911 (119904)] (2)

Assume that the vector 119905 is tangent of V(119904) separating thevertical and horizontal curvature components as shown inFigure 2 The calculation of the pipeline bending strain canbe given as

119905119909 = cos119875 sin119860

119905119910 = cos119875 cos119860

119905119911 = sin119875

(3)

where the pitch (119875) and azimuth (119860) of the pipeline centerlinecan be measured by the PIG

Assume that the vector 119896 is the curvature vector of a 3Dcurve at a given point and 119896 consists of the vertical curvature119896V and the horizontal curvature 119896ℎ which can be given as

119896 (119904) =

119889119905

119889119904

119896 = radic1198962V + 1198962

ℎ

(4)

The above equation can be written separately for eachcomponent of the curvature vector in the Cartesian system

119896119909 =

119889119905119909

119889119904

119896119910 =

119889119905119910

119889119904

119896119911 =

119889119905119911

119889119904

(5)

Journal of Sensors 3

Based on (1)ndash(4) the components of the defined curva-ture vector can be calculated as follows

119896119909 = minus sin119875(119889119875

119889119904

) sin119860 + cos119875 cos119860(119889119860119889119904

)

119896119910 = minus sin119875(119889119875

119889119904

) cos119860 minus cos119875 sin119860(119889119860119889119904

)

119896119911 = cos119875(119889119875119889119904

)

(6)

The vertical curvature 119896V and the horizontal curvature 119896ℎcan be given as

119896V = minus119889119875

119889119904

119896ℎ = minus(

119889119860

119889119904

) cos119875(7)

From (7) it can be seen that the pipeline bending straincan be calculated with the attitude data which can beacquired with a PIG

3 Modeling and Computation of Dent

In Section 2 a computation of the pipeline bending strain isdiscussed The total of the features such as displacement anddent based on bending strain can be computed and identifiedHowever it is difficult to obtain the depth of dent by thebending strain measurement Because of the lots of factorsit is hard to find a simple relationship between bending strainand size of dent Neural network is a good solution for thispractical project It is based on complex connection of largenumber of neurons and deal with difficult language of themodeling information by self-learning self-organized andnonlinear dynamics

31 BP Neural Network The BP neural network consists ofan input layer a hidden layer and a linear output layer Eachlayer is composed of several neurons [22ndash24]Themost basicthree-layer BP neural network structure is shown in Figure 3It is assumed that each layer consisted of ldquo119873rdquo processingelements the training set included ldquo119872rdquo samples mode for(119883119896119884119896) For the ldquo119901rdquo training sample the sum of total inputis119873119901119895 for 119895 element If the output is 119874119901119895 it follows

119873119875119895 =

119873

sum

119894minus0

119908119895119894119874119875119895 119874119901119895 = 119891 (119873119875119895) (8)

where 119908119895119894 is the weight for neurons between 119894 and 119895 119891 is thefunction as

119891 (119909) =

1

1 + 119890minus119909 (9)

The error of neural network is as follows

119864 = sum

119901

119864119875 119864119875 =

1

2

sum

119894

(119889119875119895 minus 119874119901119895)

2

(10)

S1 S3

S2

X1 X3

X2

Figure 3 A typical signal of dent feature from bending strain

where 119889119901119895 is expect output of the 119895 output for input of 119901 If thelearning rule used a method for gradient descent the weightscan be changed by error function The modified function isas follows

119908119894119895 (119905 + 1) = 119908119894119895 (119905) + 120578120575119901119895119874119901119895

120575119901119895 = 1198911015840(119873119901119895) (119889119901119895 minus 119874119901119895) 120575

1015840

119901119895= 1198911015840(119873119901119895)sum

119896

120575119901119896119908119896119895

(11)

where 119905 is the times of learning 120578 is the learning factor and119901 is the value of modified error 120575119901119895 is the modify error foroutput and 1205751015840

119901119895is the modified error for hidden layer

32 Improving the Neural Network Based on BayesianTheoryThe traditional neural network has difficulty in controlling itsmodel complexity leading to network overfitting long timetraining and network model low stability [25 26] Howeverthe Bayesian neural network based onBayesian reasoning cansolve these problems effectively through correcting networktraining performance function

In the actual complex system the input quantity includesgiven input quantity and unknown disturbance quantity suchas (noise signals) and normally the unknown disturbancequantity has an effect on the system output [27 28] Let119909 = [1199091 1199092 119909119899] as an observable given input then therelationship between system output quantity 119910 and inputquantity 119909 can be described by the following formula

119910 = 119891 (119909) + 120576 (12)

where 120576 represents the effect of input on the output andthe random quantity submitting to some distribution Thetraining performance function of neural network adoptsmean square error function Assume error function is asfollows

1198640 =

1

2

119870

sum

119896minus1

119873

sum

119899minus1

(1199101015840

119899119896minus 119910119899119896)

2

(13)

where 119873 is the sample numbers 119870 is the neural networkoutput numbers 119910119899119896 is the expectation output and 1199101015840

119899119896is the

network actual output Generally the not unique 1198640 solutionmay lead the neural network training to fall into partial min-imum value To solve this problem a constraint item can beintroduced to make function 119891(119909) have interpolation ability

4 Journal of Sensors

and then the solution for 1198640 is stable The condition that119891(119909) has interpolation ability is that 119891(119909) is smooth Whilethe square sum of network weight and threshold is smallits output is smoother Therefore constraint item shouldrepresent smooth constraint and the objective function isexpressed as follows

119864ALL = 1205720 1

2

119882

sum

119894=1

1199082

119894+ 1205730 1

2

119870

sum

119896=1

119873

sum

119899=1

(1199101015840

119899119896minus 119910119899119896)

2

= 120572119864net (119908) + 120573119864net (119910)

(14)

119882 is the neural network weight and threshold numbers119908 is network weight and threshold 119864net(119908) is the square sumof weight and threshold 119864net(119910) is the square sum of networkactual output and objective expectation value residual and 120572and 120573 are hyperparameter which decide the value size of theneural network objective performance function and controlthe distribution form of weight and thresholdThrough usingnew objective performance function the network weightand threshold are as little as possible on condition that thenetwork training error is as small as possible According toBayesian posterior distribution probability of 120572 and 120573meetsthe following formula

119875 (120572 120573 | 119863119873ℎ) =

119875 (119863 | 120572 120573119873ℎ) 119875 (120572 120573 | 119873ℎ)

119875 (119863 | 119873ℎ)

(15)

119863 = 119909(119873) 119910(119873) is a sample data set which consistsof 119873 samples 119873ℎ is neural network hidden layer number119875(120572 120573 | 119873ℎ) is hyperparameter prior probability 119875(119863 | 119873ℎ)

is normalization factor and 119875(119863 | 120572 120573119873ℎ) is likelihoodfunction After partial derivatives to 120572 120573 the most salienthyperparameter can be carried out

120572119872119875 =

120574

2119864net (119908119872119875)

120573119872119875 =

119873 minus 120574

2119864net (119908119872119875)

(16)

120574 is the number that can reduce parameters numbers (0ndashW) in performance index function parameters in network119908119872119875 is the weight and threshold when 119864ALL is minimum119872119875 is the subscript for 119908 119908119872119875 is the weight and thresholdwhen 119864ALL is minimum After working out the most salienthyperparameter 120572 120573 determine whether 119864ALL is convergenceagain Through iterative judgment the best and most salientneural network model can be obtained

Relative to traditional neural network the Bayesian neu-ral network is focused on the probability distribution of thewhole parameter space and the predicted results are basedon the statistical average of parameters posterior distributionA single model is corresponding to one point in the spaceall models are corresponding to the whole parameter spaceTherefore the Bayesian neural network guarantees networkstronger generalization in theory

33 Modeling and Computation for Dent by Bending Strain Atypical signal of dent feature from bending strain is shown in

Input the training data

Initialize the weight and threshold fornetwork

Training the neural network with L-M algorithm tocalculate the minimum target function

Whether convergence

Calculated the posterior distribution probability to checkthe modeling significant

Other modeling

Training ended and output thenetwork modeling

Yes

Yes

No

No

To calculate the parameters for 120574120572 and 120573

Figure 4 Dent modeling and computation based on Bayesiantheory neural network

Figure 3The horizontal axis represents the dent length alongthe pipeline and the vertical axis is the bending strain for dentrespectively The bending strain can be represented for theseverity for the dent To the identity of the dent feature theinput can be related to the bending strain information for thedent The output selected the actual depth of the dent fromthe site measurement

According to the relationship for the bending strain anddent as shown in Figure 3 pattern in a plot of bending strainagainst distance shows the characteristic of environmentallyinduced deformation for dent The pattern is caused by aprimary action in the center with a reaction (bending strainin the opposite direction) on both sides In the case of a dentthe central deformation is an underbend with overbends ateither end The main influencing parameters are includingthe beginning bending strain 1198781 and ending bending strain1198783 which is related to the reaction area of dent along thepipeline respectively The largest bending strain 1198782 is relatedto the deepest of the dent 1198831 is the distance for reactionof the beginning bending strain for dent 1198832 is the distancefor primary action for dent 1198833 is the distance for anotherreaction of the ending bending strain for dent The processof modeling and computation for the dent based on Bayesiantheory neural network is shown in Figure 4

4 Field Test and Data Analysis

41 Equipment and Performance To test the proposedmethod experimentally an in-line inspection tool (shown in

Journal of Sensors 5

Table 1 Characteristics of sensors

Sensor Characteristics Magnitude

Gyroscope Bias lt001∘hRandom walk 0002

∘radich

Accelerometer Bias stability lt50 120583gScaling factor lt50 ppm

Odometer Scaling factor lt03White noise lt01ms

Landmark White noise ltplusmn1m

Figure 5 IMU Pipeline Inspection Tool

Figure 5) with the proposed method is used to inspect oilpipeline which is about 150 kilometers About 150 dents havebeen dug tomeasure and verify depth gauge in past five yearsThese actual data can be selected to be used for sample for theproposed algorithm

The IMUpipeline ILI toolmainly consists of the followingsensors to be used to calculate the bending strain and positionof the pipeline

(1) Inertialmeasurement unit is themain device to gatherthe ILI data Three gyroscopes and three accelerom-eters orthogonally mounted on the IMU to measurethe attitude and positon for ILI tool These gathereddata can be used to calculate the pipeline bendingstrain with other sensors mounted on the ILI tool

(2) Two odometers mounted on the ILI to measure thedistance and instantaneous and average speed whichcan be used to modify the inertial system errors

(3) Two weld detector sensors are used to measure andrecord time of passing each girth weld for tool Theycan use the alignment for every spool for repeatinspection

As is known most domestic oil pipelines are heatingtransportation and pass through mountains hills rivers andother complex environmental areas [29] the technical andsafety requirements of electrical equipment are extremelystrict To safely inspect the actual pipeline the IMU shouldmeet not only the technical performance but also the actualsituation of the pipe and the external environment shouldbe considered Technical performances of inertial deviceswhich are used in the field test are shown in Table 1

42 Data Analysis According to Section 2 a typical bendingstrain for dent is shown in Figure 6 whose depth is 5279mm

201

201

2

201

4

201

6

201

8

202

202

2

202

4

202

6

202

8

203

minus05minus04minus03minus02minus01

0010203

Distance (m)

Bend

ing

strai

n (

)

times104

Figure 6 Typical bending strain of a dent

1 2 3 4 5 6 702468

10121416

n

Dep

th (m

m)

Calculation for BPActual valueError

Figure 7 Calculation and error for BP neural network

for 65 of outer diameter measured with gauge in field Thebending strain for this dent is minus04125 which is located inbottom of the pipeline

For the method of Section 3 120 data sets for bendingstrain of dent are selected for the input of the new neuralnetwork Part of dents of training data are shown in Table 2The input of neural network is consisted of 1198781 1198782 1198783 11988311198832 and 1198833 introduced in Section 33 The output of neuralnetwork is the depth of the dent

The measured data input into the BP neural network tocalculate the result of the dent Parts of sample data inputinto the network for training and the other sample data toverify the accuracy The calculation is shown in Table 3 andFigure 7Themean relative error for BP calculation and actualdepth is 1226 Although the BP neural network can be usedto calculate the depth from the bending strain the accuracydegree is low to assess the dent

To verify the accuracy of the proposed method all ofdata input into the modified Bayesian neural network Theneural network can be trained by the sample data Meansquare error (MSE) is selected to assess the quality of theneural network To adjust the parameter of neural networkthe result of training is shown in Figure 8The best validationperformance ofMSE for themethod is 00040885 at the epoch95

6 Journal of Sensors

Table 2 Part of dents of training data

Depth ratio () Depth (mm) 1198781() 119878

2() 119878

3() 119883

1(m) 119883

2(m) 119883

3(m)

202 1645 0067 0067 0079 5 18 9176 1429 0111 0111 0023 9 16 2139 1128 0016 0016 0046 1 23 6149 1213 0128 0128 0149 12 17 13199 1615 0068 0068 0101 6 15 9188 153 0055 0055 0080 3 14 5

Table 3 Calculation and error for BP neural network

Actual depth (mm) Calculationof BP (mm)

Error for calculation of BPand actual depth (mm)

Relative error()

147153 1338198 133332 906158535 1273971 311379 1964151218 1285353 226827 1500151218 1310556 201624 1333113007 1332507 202437 179196747 1019502 052032 538152844 1443888 084552 553

Mea

n sq

uare

d er

ror (

MSE

)

Best validation performance is 00040885 at epoch 95

TrainValidation

TestBest

0 10 20 30 40 50 60 70 80 90 100101 epochs

100

10minus1

10minus2

10minus3

Figure 8 Training for modified neural network based on Bayesian

The calculation of modified neural network based onBayesian is shown in Figure 9 and Table 4 It is clear to seethe method can be used to calculate the depth of dent frombending strain From Table 4 the mean relative error for cal-culation and actual depth is 244The accuracy of modifiedneural network is raised more than BP neural network Thecalculation of modified Bayesian neural network can be usedto calculate the depth of dent from dent With this methodthe pipeline company runs the IMU tool for once not only

1 2 3 4 5 6 7minus05

0

05

1

15

2

n

Dep

th (m

m)

Actual valueCalculation for Bayesian neural networkError

Figure 9 Calculation and error for modified Bayesian neuralnetwork

to obtain the pipeline bending strain but also to calculate allof the depths of dent The proposed method offers a usefulmethod for pipeline integrity evaluation

5 Conclusions

The in-line inspection tool which loads the IMU used toinspect the centerline of the pipeline The attitude infor-mation can be used to compute the bending strain of thepipeline However there is no paper or report to researchthe relationship between the bending strain and dent In thispaper based on the analysis of the calculation method of

Journal of Sensors 7

Table 4 Calculation and error for modified Bayesian neural network

Actual depth (mm) Calculation ofBayesian (mm)

Error for calculation of Bayesianand actual depth (mm) Relative error ()

147153 1498349 0268193 182158535 157455 0108 068151218 1417717 094463 625151218 1496672 015508 103113007 1107863 022207 19796747 1010614 0431435 446152844 1541688 0132484 087

pipeline bending strain we propose amethod based onmod-ified Bayesian neural network to calculate dent depths frombending strain To test the proposed method experimentallya PIG with the proposed method is used to inspect a 150 kmpipeline It can be obtained that

(1) A new method is proposed to verify the relationshipbetween pipeline bending strain and dent based onthe calculation of bending strain

(2) According to the characteristic of signal betweendent and pipeline bending strain a new model ispresented in detail to be used for proposed calculationof algorithm

(3) The calculation of modified Bayesian neural networkis proposed to compute the depth dent with pipelinebending strain to compare with actual data Andtraditional BP neural network is used to calculatethe depth of dent According to the result of calcula-tion the mean accuracy of calculation for modifiedBayesian neural network is better than BP neuralnetwork for 982

(4) According to calculation the proposed method ismore accurate and suitable for the calculation ofdepth based on pipeline bending strain The meanrelative error for calculation of dent depth based themodified Bayesian neural network is 244

This paper provides a novel method for calculating dentdepths from the pipeline bending strain The dent can beevaluated by the bending strain not used to dig or use anothertool to reinspect The bending strain and calculation of dentdepth is also useful to the evaluation of pipeline integrity

Competing Interests

The authors declare that they have no competing interests

Acknowledgments

Thisworkwas supported by the project of Petrochina PipelineCompany ldquothe research of safety service of buried pipeline inpermafrost regionrdquo

References

[1] Q Aihua ldquoChinese oil and gas pipeline transport develop-ment status and analysis of associated problemsrdquo InternationalPetroleum Economics vol 17 no 12 pp 17ndash12 2009

[2] L-J Yang R-Q Li S-W Gao and Y Yang ldquoNavigating andpositioning technique for inner detection of pipelinerdquo Journalof Shenyang University of Technology vol 34 no 4 pp 427ndash4322012

[3] S Sha and Q Feng ldquoExcavated verification technology of pitdefects of buried pipelinerdquoOil and Gas Storage and Transporta-tion vol 8 pp 834ndash838 2014

[4] HGhaednia SDas RWang andRKania ldquoSafe burst strengthof a pipeline with dent-crack defect effect of crack depth andoperating pressurerdquo Engineering Failure Analysis vol 55 pp288ndash299 2015

[5] B Pinheiro I Pasqualino and S Cunha ldquoFatigue life assess-ment of damaged pipelines under cyclic internal pressurepipelines with longitudinal and transverse plain dentsrdquo Inter-national Journal of Fatigue vol 68 pp 38ndash47 2014

[6] M Allouti C Schmitt and G Pluvinage ldquoAssessment of agouge and dent defect in a pipeline by a combined criterionrdquoEngineering Failure Analysis vol 36 pp 1ndash13 2014

[7] S Paeper B Brown and T Beuker ldquoInline inspection of dentsand corrosion using lsquohigh qualityrsquo multi-purpose smart-piginspection datardquo in Proceedings of the 6th International PipelineConference (IPC rsquo06) pp 243ndash248 September 2006

[8] X Wang and H Song ldquoThe inertial technology based 3-dimensional information measurement system for under-ground pipelinerdquoMeasurement vol 45 no 3 pp 604ndash614 2012

[9] LMengjie ldquoThe study of accurate in-line inspection technologyto offshore pipeline routerdquo China Offshore Platform vol 19 no6 pp 46ndash49 2004

[10] J Yu J G Lee C G Park andH S Han ldquoAn off-line navigationof a geometry PIG using a modified nonlinear fixed-intervalsmoothing filterrdquoControl Engineering Practice vol 13 no 11 pp1403ndash1411 2005

[11] J A Czyz C Fraccaroli andA P Sergeant ldquoMeasuring pipelinemovement in geotechnically unstable areas using an inertialgeometry pipeline inspection pigrdquo in Proceedings of the ASME1st International Pipeline Conference Calgary Canada June1996

[12] J A Czyz and J Falk ldquoUse of geopig for prevention of pipelinefailures in environmentally sensitive areasrdquo in Proceedings ofthe Pipeline Pigging Integrity Assessment and Repair ConferenceHouston Tex USA February 2000

8 Journal of Sensors

[13] I B Iflefel D G Moffat and J Mistry ldquoThe interaction ofpressure and bending on a dented piperdquo International Journalof Pressure Vessels and Piping vol 82 no 10 pp 761ndash769 2005

[14] J Błachut and I B Iflefel ldquoCollapse of pipeswith plain or gougeddents by bending momentrdquo International Journal of PressureVessels and Piping vol 84 no 9 pp 560ndash571 2007

[15] J-H Baek Y-P Kim W-S Kim J-M Koo and C-S SeokldquoLoad bearing capacity of API X65 pipe with dent defect underinternal pressure and in-plane bendingrdquo Materials Science andEngineering A vol 540 pp 70ndash82 2012

[16] A Limam L-H Lee and S Kyriakides ldquoOn the collapse ofdented tubes under combined bending and internal pressurerdquoInternational Journal of Mechanical Sciences vol 55 no 1 pp1ndash12 2012

[17] D K Kim S H Cho S S Park H R Yoo and Y W RholdquoDesign and implementation of 3010158401015840 geometry PIGrdquo KSMEInternational Journal vol 17 no 5 pp 629ndash636 2003

[18] J A Czyz et al ldquoMulti-pipeline geographical informationsystem based on high accuracy inertial surveysrdquo in Proceedingsof the ASME 3rd International Pipeline Conference CalgaryCanada 2000

[19] J D Hart N Zulfiqar D H Moore and G R Swank ldquolsquoDigitalPiggingrsquo as a basis for improved pipeline structural integrityevaluationsrdquo in Proceedings of the International Pipeline Con-ference vol 2 of Integrity Management Poster Session StudentPaper Competition Calgary Canada September 2006

[20] J A Czyz and J RAdams ldquoComputations of pipelinemdashbendingstrains based on geopig measurementsrdquo in Proceedings of theipeline Pigging and Integrity Monitoring Conference pp 14ndash17Houston Tex USA February 1994

[21] J D Hart and G H Powel Geometry Monitoring of the Trans-Alaska Pipeline Trans-Alaska Pipeline System (TAPS) 2005

[22] D Yu L ZhangW Liang Y Ye and ZWang ldquoNoise reductionof signal and condition recognition of long-distance pipelinerdquoActa Petrolei Sinica vol 30 no 6 pp 937ndash941 2009

[23] J Zhou BGuiM Li andWLin ldquoAn application of the artificialneural net dominated by lithology to permeability predictionrdquoActa Petrolei Sinica vol 31 no 6 pp 985ndash988 2010

[24] D Li D Lu X Kong and Y Du ldquoProcessing of well log databased on backpropagation neural network implicit approxima-tionrdquo Acta Petrolei Sinica vol 28 no 3 pp 105ndash108 2007

[25] K-X Peng and S-Z Yu ldquoPrediction and control of strip thick-ness based on Bayesian neural networksrdquo Journal of Universityof Science and Technology Beijing vol 32 no 2 pp 256ndash2622010

[26] H Yang and H-Z Fu ldquoStock index forecast based on bayesianregularization BP neural networkrdquo Science Techno Logy andEngineering vol 6 pp 3306ndash3310 2009

[27] H Wang L Shi H Zhao and Y Yue ldquoTemperature predictionmodel for sunlight greenhouse based on bayesian regularizationBP neural networkrdquo Hubei Agricultural Sciences vol 9 pp4300ndash4303 2015

[28] Y Shenghua and D Juan ldquoGDP prediction based on principalcomponent analysis and bayesian regularization BP neuralnetworkrdquo Journal of Hunan University (Social Sciences) vol 11pp 42ndash45 2011

[29] R Li Q Feng M Cai et al ldquoMeasurement of long-distanceburied pipeline centerline based on multi-sensor data fusionrdquoActa Petrolei Sinica vol 35 no 5 pp 987ndash992 2014

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

2 Journal of Sensors

Figure 1 Measurement system of in-line inspection of pipelinecenterline

tool which consists of a Data Acquisition System (DAS)an IMU some weld detectors and odometers is driven tomove through the pipeline and collect data regarding dentsbends and navigation Petrochina Pipeline Company hasdeveloped an IMU system based on a navigation systemwhich is suitable for a geometry PIG The IMU can be usedto record the attitude data for the tool during the inspectionA method of calculation for pipeline centerline is proposedby fusion for multisensors Czyz et al [11 12] developed apure geometry-based algorithm for the bending strain of apipeline Another pipeline strain testing method is using thestrain gauge to be installed on the surface of pipeline Twotypes of strain gauge resistance and optical fiber are used tomeasure the strain of pipelineThe strain gauge of optical fiberis more precise for measuring the strain These two methodscan be used to monitor the strain changes of the pipeline Butthe limit for these two methods is only used to measure thepart sections of pipelineThey cannot inspect thewhole strainfor long-distance buried pipeline

The bending strain can be reported for the whole pipelinewhen the tool runs for once Generally the technical just usesthe bending strain to detect the position of pipeline enduringthe extra strain larger than the operation requirement andassessing the displacement for local pipeline However itis found that the bending strain can be affected by thedent [13ndash16] In this paper a method for modeling andcomputation of dent which is based on bending strain isproposedThe technique takes Bayesian statistical theory andneural network to improve a new method to compute thedepth of the dent It is more useful to assess the dent for thepipeline integrity

2 The Calculation Method of the PipelineBending Strain

In this section the calculation method of the pipelinebending strain is described Several factors such as internalpressure temperature differential external loads and bound-ary conditions are affecting the strain The strain can beconsisted of two primary components longitudinal and hoopstrain [17ndash19] The bending strain is one of the componentsfor longitudinal strain It can be described by the followingformula

120576 =

119863

2

119896

120576V =119863

2

119896V

120576ℎ =

119863

2

119896ℎ

(1)

z (zenith)

y(north)

x (east)

Pitch

Azimuth

ty

t

tx

tz

Figure 2 Pitch (119875) and azimuth (119860) for the pipeline centerline

where 120576 is the total bending strain 119896 is the total curvature 119896Vis vertical curvature and 119896ℎ is horizontal curvature

The total curvature of the centerline of a pipe is describedat each point along the pipeline by the curvature vector Inorder to calculate the pipeline curvature the centerline of apipe is considered as a 3D parametric curve described in aCartesian system by a vector V(119904) which is a function of adistance (119904) along the curve [20 21]

V (119904) = [119909 (119904) 119910 (119904) 119911 (119904)] (2)

Assume that the vector 119905 is tangent of V(119904) separating thevertical and horizontal curvature components as shown inFigure 2 The calculation of the pipeline bending strain canbe given as

119905119909 = cos119875 sin119860

119905119910 = cos119875 cos119860

119905119911 = sin119875

(3)

where the pitch (119875) and azimuth (119860) of the pipeline centerlinecan be measured by the PIG

Assume that the vector 119896 is the curvature vector of a 3Dcurve at a given point and 119896 consists of the vertical curvature119896V and the horizontal curvature 119896ℎ which can be given as

119896 (119904) =

119889119905

119889119904

119896 = radic1198962V + 1198962

ℎ

(4)

The above equation can be written separately for eachcomponent of the curvature vector in the Cartesian system

119896119909 =

119889119905119909

119889119904

119896119910 =

119889119905119910

119889119904

119896119911 =

119889119905119911

119889119904

(5)

Journal of Sensors 3

Based on (1)ndash(4) the components of the defined curva-ture vector can be calculated as follows

119896119909 = minus sin119875(119889119875

119889119904

) sin119860 + cos119875 cos119860(119889119860119889119904

)

119896119910 = minus sin119875(119889119875

119889119904

) cos119860 minus cos119875 sin119860(119889119860119889119904

)

119896119911 = cos119875(119889119875119889119904

)

(6)

The vertical curvature 119896V and the horizontal curvature 119896ℎcan be given as

119896V = minus119889119875

119889119904

119896ℎ = minus(

119889119860

119889119904

) cos119875(7)

From (7) it can be seen that the pipeline bending straincan be calculated with the attitude data which can beacquired with a PIG

3 Modeling and Computation of Dent

In Section 2 a computation of the pipeline bending strain isdiscussed The total of the features such as displacement anddent based on bending strain can be computed and identifiedHowever it is difficult to obtain the depth of dent by thebending strain measurement Because of the lots of factorsit is hard to find a simple relationship between bending strainand size of dent Neural network is a good solution for thispractical project It is based on complex connection of largenumber of neurons and deal with difficult language of themodeling information by self-learning self-organized andnonlinear dynamics

31 BP Neural Network The BP neural network consists ofan input layer a hidden layer and a linear output layer Eachlayer is composed of several neurons [22ndash24]Themost basicthree-layer BP neural network structure is shown in Figure 3It is assumed that each layer consisted of ldquo119873rdquo processingelements the training set included ldquo119872rdquo samples mode for(119883119896119884119896) For the ldquo119901rdquo training sample the sum of total inputis119873119901119895 for 119895 element If the output is 119874119901119895 it follows

119873119875119895 =

119873

sum

119894minus0

119908119895119894119874119875119895 119874119901119895 = 119891 (119873119875119895) (8)

where 119908119895119894 is the weight for neurons between 119894 and 119895 119891 is thefunction as

119891 (119909) =

1

1 + 119890minus119909 (9)

The error of neural network is as follows

119864 = sum

119901

119864119875 119864119875 =

1

2

sum

119894

(119889119875119895 minus 119874119901119895)

2

(10)

S1 S3

S2

X1 X3

X2

Figure 3 A typical signal of dent feature from bending strain

where 119889119901119895 is expect output of the 119895 output for input of 119901 If thelearning rule used a method for gradient descent the weightscan be changed by error function The modified function isas follows

119908119894119895 (119905 + 1) = 119908119894119895 (119905) + 120578120575119901119895119874119901119895

120575119901119895 = 1198911015840(119873119901119895) (119889119901119895 minus 119874119901119895) 120575

1015840

119901119895= 1198911015840(119873119901119895)sum

119896

120575119901119896119908119896119895

(11)

where 119905 is the times of learning 120578 is the learning factor and119901 is the value of modified error 120575119901119895 is the modify error foroutput and 1205751015840

119901119895is the modified error for hidden layer

32 Improving the Neural Network Based on BayesianTheoryThe traditional neural network has difficulty in controlling itsmodel complexity leading to network overfitting long timetraining and network model low stability [25 26] Howeverthe Bayesian neural network based onBayesian reasoning cansolve these problems effectively through correcting networktraining performance function

In the actual complex system the input quantity includesgiven input quantity and unknown disturbance quantity suchas (noise signals) and normally the unknown disturbancequantity has an effect on the system output [27 28] Let119909 = [1199091 1199092 119909119899] as an observable given input then therelationship between system output quantity 119910 and inputquantity 119909 can be described by the following formula

119910 = 119891 (119909) + 120576 (12)

where 120576 represents the effect of input on the output andthe random quantity submitting to some distribution Thetraining performance function of neural network adoptsmean square error function Assume error function is asfollows

1198640 =

1

2

119870

sum

119896minus1

119873

sum

119899minus1

(1199101015840

119899119896minus 119910119899119896)

2

(13)

where 119873 is the sample numbers 119870 is the neural networkoutput numbers 119910119899119896 is the expectation output and 1199101015840

119899119896is the

network actual output Generally the not unique 1198640 solutionmay lead the neural network training to fall into partial min-imum value To solve this problem a constraint item can beintroduced to make function 119891(119909) have interpolation ability

4 Journal of Sensors

and then the solution for 1198640 is stable The condition that119891(119909) has interpolation ability is that 119891(119909) is smooth Whilethe square sum of network weight and threshold is smallits output is smoother Therefore constraint item shouldrepresent smooth constraint and the objective function isexpressed as follows

119864ALL = 1205720 1

2

119882

sum

119894=1

1199082

119894+ 1205730 1

2

119870

sum

119896=1

119873

sum

119899=1

(1199101015840

119899119896minus 119910119899119896)

2

= 120572119864net (119908) + 120573119864net (119910)

(14)

119882 is the neural network weight and threshold numbers119908 is network weight and threshold 119864net(119908) is the square sumof weight and threshold 119864net(119910) is the square sum of networkactual output and objective expectation value residual and 120572and 120573 are hyperparameter which decide the value size of theneural network objective performance function and controlthe distribution form of weight and thresholdThrough usingnew objective performance function the network weightand threshold are as little as possible on condition that thenetwork training error is as small as possible According toBayesian posterior distribution probability of 120572 and 120573meetsthe following formula

119875 (120572 120573 | 119863119873ℎ) =

119875 (119863 | 120572 120573119873ℎ) 119875 (120572 120573 | 119873ℎ)

119875 (119863 | 119873ℎ)

(15)

119863 = 119909(119873) 119910(119873) is a sample data set which consistsof 119873 samples 119873ℎ is neural network hidden layer number119875(120572 120573 | 119873ℎ) is hyperparameter prior probability 119875(119863 | 119873ℎ)

is normalization factor and 119875(119863 | 120572 120573119873ℎ) is likelihoodfunction After partial derivatives to 120572 120573 the most salienthyperparameter can be carried out

120572119872119875 =

120574

2119864net (119908119872119875)

120573119872119875 =

119873 minus 120574

2119864net (119908119872119875)

(16)

120574 is the number that can reduce parameters numbers (0ndashW) in performance index function parameters in network119908119872119875 is the weight and threshold when 119864ALL is minimum119872119875 is the subscript for 119908 119908119872119875 is the weight and thresholdwhen 119864ALL is minimum After working out the most salienthyperparameter 120572 120573 determine whether 119864ALL is convergenceagain Through iterative judgment the best and most salientneural network model can be obtained

Relative to traditional neural network the Bayesian neu-ral network is focused on the probability distribution of thewhole parameter space and the predicted results are basedon the statistical average of parameters posterior distributionA single model is corresponding to one point in the spaceall models are corresponding to the whole parameter spaceTherefore the Bayesian neural network guarantees networkstronger generalization in theory

33 Modeling and Computation for Dent by Bending Strain Atypical signal of dent feature from bending strain is shown in

Input the training data

Initialize the weight and threshold fornetwork

Training the neural network with L-M algorithm tocalculate the minimum target function

Whether convergence

Calculated the posterior distribution probability to checkthe modeling significant

Other modeling

Training ended and output thenetwork modeling

Yes

Yes

No

No

To calculate the parameters for 120574120572 and 120573

Figure 4 Dent modeling and computation based on Bayesiantheory neural network

Figure 3The horizontal axis represents the dent length alongthe pipeline and the vertical axis is the bending strain for dentrespectively The bending strain can be represented for theseverity for the dent To the identity of the dent feature theinput can be related to the bending strain information for thedent The output selected the actual depth of the dent fromthe site measurement

According to the relationship for the bending strain anddent as shown in Figure 3 pattern in a plot of bending strainagainst distance shows the characteristic of environmentallyinduced deformation for dent The pattern is caused by aprimary action in the center with a reaction (bending strainin the opposite direction) on both sides In the case of a dentthe central deformation is an underbend with overbends ateither end The main influencing parameters are includingthe beginning bending strain 1198781 and ending bending strain1198783 which is related to the reaction area of dent along thepipeline respectively The largest bending strain 1198782 is relatedto the deepest of the dent 1198831 is the distance for reactionof the beginning bending strain for dent 1198832 is the distancefor primary action for dent 1198833 is the distance for anotherreaction of the ending bending strain for dent The processof modeling and computation for the dent based on Bayesiantheory neural network is shown in Figure 4

4 Field Test and Data Analysis

41 Equipment and Performance To test the proposedmethod experimentally an in-line inspection tool (shown in

Journal of Sensors 5

Table 1 Characteristics of sensors

Sensor Characteristics Magnitude

Gyroscope Bias lt001∘hRandom walk 0002

∘radich

Accelerometer Bias stability lt50 120583gScaling factor lt50 ppm

Odometer Scaling factor lt03White noise lt01ms

Landmark White noise ltplusmn1m

Figure 5 IMU Pipeline Inspection Tool

Figure 5) with the proposed method is used to inspect oilpipeline which is about 150 kilometers About 150 dents havebeen dug tomeasure and verify depth gauge in past five yearsThese actual data can be selected to be used for sample for theproposed algorithm

The IMUpipeline ILI toolmainly consists of the followingsensors to be used to calculate the bending strain and positionof the pipeline

(1) Inertialmeasurement unit is themain device to gatherthe ILI data Three gyroscopes and three accelerom-eters orthogonally mounted on the IMU to measurethe attitude and positon for ILI tool These gathereddata can be used to calculate the pipeline bendingstrain with other sensors mounted on the ILI tool

(2) Two odometers mounted on the ILI to measure thedistance and instantaneous and average speed whichcan be used to modify the inertial system errors

(3) Two weld detector sensors are used to measure andrecord time of passing each girth weld for tool Theycan use the alignment for every spool for repeatinspection

As is known most domestic oil pipelines are heatingtransportation and pass through mountains hills rivers andother complex environmental areas [29] the technical andsafety requirements of electrical equipment are extremelystrict To safely inspect the actual pipeline the IMU shouldmeet not only the technical performance but also the actualsituation of the pipe and the external environment shouldbe considered Technical performances of inertial deviceswhich are used in the field test are shown in Table 1

42 Data Analysis According to Section 2 a typical bendingstrain for dent is shown in Figure 6 whose depth is 5279mm

201

201

2

201

4

201

6

201

8

202

202

2

202

4

202

6

202

8

203

minus05minus04minus03minus02minus01

0010203

Distance (m)

Bend

ing

strai

n (

)

times104

Figure 6 Typical bending strain of a dent

1 2 3 4 5 6 702468

10121416

n

Dep

th (m

m)

Calculation for BPActual valueError

Figure 7 Calculation and error for BP neural network

for 65 of outer diameter measured with gauge in field Thebending strain for this dent is minus04125 which is located inbottom of the pipeline

For the method of Section 3 120 data sets for bendingstrain of dent are selected for the input of the new neuralnetwork Part of dents of training data are shown in Table 2The input of neural network is consisted of 1198781 1198782 1198783 11988311198832 and 1198833 introduced in Section 33 The output of neuralnetwork is the depth of the dent

The measured data input into the BP neural network tocalculate the result of the dent Parts of sample data inputinto the network for training and the other sample data toverify the accuracy The calculation is shown in Table 3 andFigure 7Themean relative error for BP calculation and actualdepth is 1226 Although the BP neural network can be usedto calculate the depth from the bending strain the accuracydegree is low to assess the dent

To verify the accuracy of the proposed method all ofdata input into the modified Bayesian neural network Theneural network can be trained by the sample data Meansquare error (MSE) is selected to assess the quality of theneural network To adjust the parameter of neural networkthe result of training is shown in Figure 8The best validationperformance ofMSE for themethod is 00040885 at the epoch95

6 Journal of Sensors

Table 2 Part of dents of training data

Depth ratio () Depth (mm) 1198781() 119878

2() 119878

3() 119883

1(m) 119883

2(m) 119883

3(m)

202 1645 0067 0067 0079 5 18 9176 1429 0111 0111 0023 9 16 2139 1128 0016 0016 0046 1 23 6149 1213 0128 0128 0149 12 17 13199 1615 0068 0068 0101 6 15 9188 153 0055 0055 0080 3 14 5

Table 3 Calculation and error for BP neural network

Actual depth (mm) Calculationof BP (mm)

Error for calculation of BPand actual depth (mm)

Relative error()

147153 1338198 133332 906158535 1273971 311379 1964151218 1285353 226827 1500151218 1310556 201624 1333113007 1332507 202437 179196747 1019502 052032 538152844 1443888 084552 553

Mea

n sq

uare

d er

ror (

MSE

)

Best validation performance is 00040885 at epoch 95

TrainValidation

TestBest

0 10 20 30 40 50 60 70 80 90 100101 epochs

100

10minus1

10minus2

10minus3

Figure 8 Training for modified neural network based on Bayesian

The calculation of modified neural network based onBayesian is shown in Figure 9 and Table 4 It is clear to seethe method can be used to calculate the depth of dent frombending strain From Table 4 the mean relative error for cal-culation and actual depth is 244The accuracy of modifiedneural network is raised more than BP neural network Thecalculation of modified Bayesian neural network can be usedto calculate the depth of dent from dent With this methodthe pipeline company runs the IMU tool for once not only

1 2 3 4 5 6 7minus05

0

05

1

15

2

n

Dep

th (m

m)

Actual valueCalculation for Bayesian neural networkError

Figure 9 Calculation and error for modified Bayesian neuralnetwork

to obtain the pipeline bending strain but also to calculate allof the depths of dent The proposed method offers a usefulmethod for pipeline integrity evaluation

5 Conclusions

The in-line inspection tool which loads the IMU used toinspect the centerline of the pipeline The attitude infor-mation can be used to compute the bending strain of thepipeline However there is no paper or report to researchthe relationship between the bending strain and dent In thispaper based on the analysis of the calculation method of

Journal of Sensors 7

Table 4 Calculation and error for modified Bayesian neural network

Actual depth (mm) Calculation ofBayesian (mm)

Error for calculation of Bayesianand actual depth (mm) Relative error ()

147153 1498349 0268193 182158535 157455 0108 068151218 1417717 094463 625151218 1496672 015508 103113007 1107863 022207 19796747 1010614 0431435 446152844 1541688 0132484 087

pipeline bending strain we propose amethod based onmod-ified Bayesian neural network to calculate dent depths frombending strain To test the proposed method experimentallya PIG with the proposed method is used to inspect a 150 kmpipeline It can be obtained that

(1) A new method is proposed to verify the relationshipbetween pipeline bending strain and dent based onthe calculation of bending strain

(2) According to the characteristic of signal betweendent and pipeline bending strain a new model ispresented in detail to be used for proposed calculationof algorithm

(3) The calculation of modified Bayesian neural networkis proposed to compute the depth dent with pipelinebending strain to compare with actual data Andtraditional BP neural network is used to calculatethe depth of dent According to the result of calcula-tion the mean accuracy of calculation for modifiedBayesian neural network is better than BP neuralnetwork for 982

(4) According to calculation the proposed method ismore accurate and suitable for the calculation ofdepth based on pipeline bending strain The meanrelative error for calculation of dent depth based themodified Bayesian neural network is 244

This paper provides a novel method for calculating dentdepths from the pipeline bending strain The dent can beevaluated by the bending strain not used to dig or use anothertool to reinspect The bending strain and calculation of dentdepth is also useful to the evaluation of pipeline integrity

Competing Interests

The authors declare that they have no competing interests

Acknowledgments

Thisworkwas supported by the project of Petrochina PipelineCompany ldquothe research of safety service of buried pipeline inpermafrost regionrdquo

References

[1] Q Aihua ldquoChinese oil and gas pipeline transport develop-ment status and analysis of associated problemsrdquo InternationalPetroleum Economics vol 17 no 12 pp 17ndash12 2009

[2] L-J Yang R-Q Li S-W Gao and Y Yang ldquoNavigating andpositioning technique for inner detection of pipelinerdquo Journalof Shenyang University of Technology vol 34 no 4 pp 427ndash4322012

[3] S Sha and Q Feng ldquoExcavated verification technology of pitdefects of buried pipelinerdquoOil and Gas Storage and Transporta-tion vol 8 pp 834ndash838 2014

[4] HGhaednia SDas RWang andRKania ldquoSafe burst strengthof a pipeline with dent-crack defect effect of crack depth andoperating pressurerdquo Engineering Failure Analysis vol 55 pp288ndash299 2015

[5] B Pinheiro I Pasqualino and S Cunha ldquoFatigue life assess-ment of damaged pipelines under cyclic internal pressurepipelines with longitudinal and transverse plain dentsrdquo Inter-national Journal of Fatigue vol 68 pp 38ndash47 2014

[6] M Allouti C Schmitt and G Pluvinage ldquoAssessment of agouge and dent defect in a pipeline by a combined criterionrdquoEngineering Failure Analysis vol 36 pp 1ndash13 2014

[7] S Paeper B Brown and T Beuker ldquoInline inspection of dentsand corrosion using lsquohigh qualityrsquo multi-purpose smart-piginspection datardquo in Proceedings of the 6th International PipelineConference (IPC rsquo06) pp 243ndash248 September 2006

[8] X Wang and H Song ldquoThe inertial technology based 3-dimensional information measurement system for under-ground pipelinerdquoMeasurement vol 45 no 3 pp 604ndash614 2012

[9] LMengjie ldquoThe study of accurate in-line inspection technologyto offshore pipeline routerdquo China Offshore Platform vol 19 no6 pp 46ndash49 2004

[10] J Yu J G Lee C G Park andH S Han ldquoAn off-line navigationof a geometry PIG using a modified nonlinear fixed-intervalsmoothing filterrdquoControl Engineering Practice vol 13 no 11 pp1403ndash1411 2005

[11] J A Czyz C Fraccaroli andA P Sergeant ldquoMeasuring pipelinemovement in geotechnically unstable areas using an inertialgeometry pipeline inspection pigrdquo in Proceedings of the ASME1st International Pipeline Conference Calgary Canada June1996

[12] J A Czyz and J Falk ldquoUse of geopig for prevention of pipelinefailures in environmentally sensitive areasrdquo in Proceedings ofthe Pipeline Pigging Integrity Assessment and Repair ConferenceHouston Tex USA February 2000

8 Journal of Sensors

[13] I B Iflefel D G Moffat and J Mistry ldquoThe interaction ofpressure and bending on a dented piperdquo International Journalof Pressure Vessels and Piping vol 82 no 10 pp 761ndash769 2005

[14] J Błachut and I B Iflefel ldquoCollapse of pipeswith plain or gougeddents by bending momentrdquo International Journal of PressureVessels and Piping vol 84 no 9 pp 560ndash571 2007

[15] J-H Baek Y-P Kim W-S Kim J-M Koo and C-S SeokldquoLoad bearing capacity of API X65 pipe with dent defect underinternal pressure and in-plane bendingrdquo Materials Science andEngineering A vol 540 pp 70ndash82 2012

[16] A Limam L-H Lee and S Kyriakides ldquoOn the collapse ofdented tubes under combined bending and internal pressurerdquoInternational Journal of Mechanical Sciences vol 55 no 1 pp1ndash12 2012

[17] D K Kim S H Cho S S Park H R Yoo and Y W RholdquoDesign and implementation of 3010158401015840 geometry PIGrdquo KSMEInternational Journal vol 17 no 5 pp 629ndash636 2003

[18] J A Czyz et al ldquoMulti-pipeline geographical informationsystem based on high accuracy inertial surveysrdquo in Proceedingsof the ASME 3rd International Pipeline Conference CalgaryCanada 2000

[19] J D Hart N Zulfiqar D H Moore and G R Swank ldquolsquoDigitalPiggingrsquo as a basis for improved pipeline structural integrityevaluationsrdquo in Proceedings of the International Pipeline Con-ference vol 2 of Integrity Management Poster Session StudentPaper Competition Calgary Canada September 2006

[20] J A Czyz and J RAdams ldquoComputations of pipelinemdashbendingstrains based on geopig measurementsrdquo in Proceedings of theipeline Pigging and Integrity Monitoring Conference pp 14ndash17Houston Tex USA February 1994

[21] J D Hart and G H Powel Geometry Monitoring of the Trans-Alaska Pipeline Trans-Alaska Pipeline System (TAPS) 2005

[22] D Yu L ZhangW Liang Y Ye and ZWang ldquoNoise reductionof signal and condition recognition of long-distance pipelinerdquoActa Petrolei Sinica vol 30 no 6 pp 937ndash941 2009

[23] J Zhou BGuiM Li andWLin ldquoAn application of the artificialneural net dominated by lithology to permeability predictionrdquoActa Petrolei Sinica vol 31 no 6 pp 985ndash988 2010

[24] D Li D Lu X Kong and Y Du ldquoProcessing of well log databased on backpropagation neural network implicit approxima-tionrdquo Acta Petrolei Sinica vol 28 no 3 pp 105ndash108 2007

[25] K-X Peng and S-Z Yu ldquoPrediction and control of strip thick-ness based on Bayesian neural networksrdquo Journal of Universityof Science and Technology Beijing vol 32 no 2 pp 256ndash2622010

[26] H Yang and H-Z Fu ldquoStock index forecast based on bayesianregularization BP neural networkrdquo Science Techno Logy andEngineering vol 6 pp 3306ndash3310 2009

[27] H Wang L Shi H Zhao and Y Yue ldquoTemperature predictionmodel for sunlight greenhouse based on bayesian regularizationBP neural networkrdquo Hubei Agricultural Sciences vol 9 pp4300ndash4303 2015

[28] Y Shenghua and D Juan ldquoGDP prediction based on principalcomponent analysis and bayesian regularization BP neuralnetworkrdquo Journal of Hunan University (Social Sciences) vol 11pp 42ndash45 2011

[29] R Li Q Feng M Cai et al ldquoMeasurement of long-distanceburied pipeline centerline based on multi-sensor data fusionrdquoActa Petrolei Sinica vol 35 no 5 pp 987ndash992 2014

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Journal of Sensors 3

Based on (1)ndash(4) the components of the defined curva-ture vector can be calculated as follows

119896119909 = minus sin119875(119889119875

119889119904

) sin119860 + cos119875 cos119860(119889119860119889119904

)

119896119910 = minus sin119875(119889119875

119889119904

) cos119860 minus cos119875 sin119860(119889119860119889119904

)

119896119911 = cos119875(119889119875119889119904

)

(6)

The vertical curvature 119896V and the horizontal curvature 119896ℎcan be given as

119896V = minus119889119875

119889119904

119896ℎ = minus(

119889119860

119889119904

) cos119875(7)

From (7) it can be seen that the pipeline bending straincan be calculated with the attitude data which can beacquired with a PIG

3 Modeling and Computation of Dent

In Section 2 a computation of the pipeline bending strain isdiscussed The total of the features such as displacement anddent based on bending strain can be computed and identifiedHowever it is difficult to obtain the depth of dent by thebending strain measurement Because of the lots of factorsit is hard to find a simple relationship between bending strainand size of dent Neural network is a good solution for thispractical project It is based on complex connection of largenumber of neurons and deal with difficult language of themodeling information by self-learning self-organized andnonlinear dynamics

31 BP Neural Network The BP neural network consists ofan input layer a hidden layer and a linear output layer Eachlayer is composed of several neurons [22ndash24]Themost basicthree-layer BP neural network structure is shown in Figure 3It is assumed that each layer consisted of ldquo119873rdquo processingelements the training set included ldquo119872rdquo samples mode for(119883119896119884119896) For the ldquo119901rdquo training sample the sum of total inputis119873119901119895 for 119895 element If the output is 119874119901119895 it follows

119873119875119895 =

119873

sum

119894minus0

119908119895119894119874119875119895 119874119901119895 = 119891 (119873119875119895) (8)

where 119908119895119894 is the weight for neurons between 119894 and 119895 119891 is thefunction as

119891 (119909) =

1

1 + 119890minus119909 (9)

The error of neural network is as follows

119864 = sum

119901

119864119875 119864119875 =

1

2

sum

119894

(119889119875119895 minus 119874119901119895)

2

(10)

S1 S3

S2

X1 X3

X2

Figure 3 A typical signal of dent feature from bending strain

where 119889119901119895 is expect output of the 119895 output for input of 119901 If thelearning rule used a method for gradient descent the weightscan be changed by error function The modified function isas follows

119908119894119895 (119905 + 1) = 119908119894119895 (119905) + 120578120575119901119895119874119901119895

120575119901119895 = 1198911015840(119873119901119895) (119889119901119895 minus 119874119901119895) 120575

1015840

119901119895= 1198911015840(119873119901119895)sum

119896

120575119901119896119908119896119895

(11)

where 119905 is the times of learning 120578 is the learning factor and119901 is the value of modified error 120575119901119895 is the modify error foroutput and 1205751015840

119901119895is the modified error for hidden layer

32 Improving the Neural Network Based on BayesianTheoryThe traditional neural network has difficulty in controlling itsmodel complexity leading to network overfitting long timetraining and network model low stability [25 26] Howeverthe Bayesian neural network based onBayesian reasoning cansolve these problems effectively through correcting networktraining performance function

In the actual complex system the input quantity includesgiven input quantity and unknown disturbance quantity suchas (noise signals) and normally the unknown disturbancequantity has an effect on the system output [27 28] Let119909 = [1199091 1199092 119909119899] as an observable given input then therelationship between system output quantity 119910 and inputquantity 119909 can be described by the following formula

119910 = 119891 (119909) + 120576 (12)

where 120576 represents the effect of input on the output andthe random quantity submitting to some distribution Thetraining performance function of neural network adoptsmean square error function Assume error function is asfollows

1198640 =

1

2

119870

sum

119896minus1

119873

sum

119899minus1

(1199101015840

119899119896minus 119910119899119896)

2

(13)

where 119873 is the sample numbers 119870 is the neural networkoutput numbers 119910119899119896 is the expectation output and 1199101015840

119899119896is the

network actual output Generally the not unique 1198640 solutionmay lead the neural network training to fall into partial min-imum value To solve this problem a constraint item can beintroduced to make function 119891(119909) have interpolation ability

4 Journal of Sensors

and then the solution for 1198640 is stable The condition that119891(119909) has interpolation ability is that 119891(119909) is smooth Whilethe square sum of network weight and threshold is smallits output is smoother Therefore constraint item shouldrepresent smooth constraint and the objective function isexpressed as follows

119864ALL = 1205720 1

2

119882

sum

119894=1

1199082

119894+ 1205730 1

2

119870

sum

119896=1

119873

sum

119899=1

(1199101015840

119899119896minus 119910119899119896)

2

= 120572119864net (119908) + 120573119864net (119910)

(14)

119882 is the neural network weight and threshold numbers119908 is network weight and threshold 119864net(119908) is the square sumof weight and threshold 119864net(119910) is the square sum of networkactual output and objective expectation value residual and 120572and 120573 are hyperparameter which decide the value size of theneural network objective performance function and controlthe distribution form of weight and thresholdThrough usingnew objective performance function the network weightand threshold are as little as possible on condition that thenetwork training error is as small as possible According toBayesian posterior distribution probability of 120572 and 120573meetsthe following formula

119875 (120572 120573 | 119863119873ℎ) =

119875 (119863 | 120572 120573119873ℎ) 119875 (120572 120573 | 119873ℎ)

119875 (119863 | 119873ℎ)

(15)

119863 = 119909(119873) 119910(119873) is a sample data set which consistsof 119873 samples 119873ℎ is neural network hidden layer number119875(120572 120573 | 119873ℎ) is hyperparameter prior probability 119875(119863 | 119873ℎ)

is normalization factor and 119875(119863 | 120572 120573119873ℎ) is likelihoodfunction After partial derivatives to 120572 120573 the most salienthyperparameter can be carried out

120572119872119875 =

120574

2119864net (119908119872119875)

120573119872119875 =

119873 minus 120574

2119864net (119908119872119875)

(16)

120574 is the number that can reduce parameters numbers (0ndashW) in performance index function parameters in network119908119872119875 is the weight and threshold when 119864ALL is minimum119872119875 is the subscript for 119908 119908119872119875 is the weight and thresholdwhen 119864ALL is minimum After working out the most salienthyperparameter 120572 120573 determine whether 119864ALL is convergenceagain Through iterative judgment the best and most salientneural network model can be obtained

Relative to traditional neural network the Bayesian neu-ral network is focused on the probability distribution of thewhole parameter space and the predicted results are basedon the statistical average of parameters posterior distributionA single model is corresponding to one point in the spaceall models are corresponding to the whole parameter spaceTherefore the Bayesian neural network guarantees networkstronger generalization in theory

33 Modeling and Computation for Dent by Bending Strain Atypical signal of dent feature from bending strain is shown in

Input the training data

Initialize the weight and threshold fornetwork

Training the neural network with L-M algorithm tocalculate the minimum target function

Whether convergence

Calculated the posterior distribution probability to checkthe modeling significant

Other modeling

Training ended and output thenetwork modeling

Yes

Yes

No

No

To calculate the parameters for 120574120572 and 120573

Figure 4 Dent modeling and computation based on Bayesiantheory neural network

Figure 3The horizontal axis represents the dent length alongthe pipeline and the vertical axis is the bending strain for dentrespectively The bending strain can be represented for theseverity for the dent To the identity of the dent feature theinput can be related to the bending strain information for thedent The output selected the actual depth of the dent fromthe site measurement

According to the relationship for the bending strain anddent as shown in Figure 3 pattern in a plot of bending strainagainst distance shows the characteristic of environmentallyinduced deformation for dent The pattern is caused by aprimary action in the center with a reaction (bending strainin the opposite direction) on both sides In the case of a dentthe central deformation is an underbend with overbends ateither end The main influencing parameters are includingthe beginning bending strain 1198781 and ending bending strain1198783 which is related to the reaction area of dent along thepipeline respectively The largest bending strain 1198782 is relatedto the deepest of the dent 1198831 is the distance for reactionof the beginning bending strain for dent 1198832 is the distancefor primary action for dent 1198833 is the distance for anotherreaction of the ending bending strain for dent The processof modeling and computation for the dent based on Bayesiantheory neural network is shown in Figure 4

4 Field Test and Data Analysis

41 Equipment and Performance To test the proposedmethod experimentally an in-line inspection tool (shown in

Journal of Sensors 5

Table 1 Characteristics of sensors

Sensor Characteristics Magnitude

Gyroscope Bias lt001∘hRandom walk 0002

∘radich

Accelerometer Bias stability lt50 120583gScaling factor lt50 ppm

Odometer Scaling factor lt03White noise lt01ms

Landmark White noise ltplusmn1m

Figure 5 IMU Pipeline Inspection Tool

Figure 5) with the proposed method is used to inspect oilpipeline which is about 150 kilometers About 150 dents havebeen dug tomeasure and verify depth gauge in past five yearsThese actual data can be selected to be used for sample for theproposed algorithm

The IMUpipeline ILI toolmainly consists of the followingsensors to be used to calculate the bending strain and positionof the pipeline

(1) Inertialmeasurement unit is themain device to gatherthe ILI data Three gyroscopes and three accelerom-eters orthogonally mounted on the IMU to measurethe attitude and positon for ILI tool These gathereddata can be used to calculate the pipeline bendingstrain with other sensors mounted on the ILI tool

(2) Two odometers mounted on the ILI to measure thedistance and instantaneous and average speed whichcan be used to modify the inertial system errors

(3) Two weld detector sensors are used to measure andrecord time of passing each girth weld for tool Theycan use the alignment for every spool for repeatinspection

As is known most domestic oil pipelines are heatingtransportation and pass through mountains hills rivers andother complex environmental areas [29] the technical andsafety requirements of electrical equipment are extremelystrict To safely inspect the actual pipeline the IMU shouldmeet not only the technical performance but also the actualsituation of the pipe and the external environment shouldbe considered Technical performances of inertial deviceswhich are used in the field test are shown in Table 1

42 Data Analysis According to Section 2 a typical bendingstrain for dent is shown in Figure 6 whose depth is 5279mm

201

201

2

201

4

201

6

201

8

202

202

2

202

4

202

6

202

8

203

minus05minus04minus03minus02minus01

0010203

Distance (m)

Bend

ing

strai

n (

)

times104

Figure 6 Typical bending strain of a dent

1 2 3 4 5 6 702468

10121416

n

Dep

th (m

m)

Calculation for BPActual valueError

Figure 7 Calculation and error for BP neural network

for 65 of outer diameter measured with gauge in field Thebending strain for this dent is minus04125 which is located inbottom of the pipeline

For the method of Section 3 120 data sets for bendingstrain of dent are selected for the input of the new neuralnetwork Part of dents of training data are shown in Table 2The input of neural network is consisted of 1198781 1198782 1198783 11988311198832 and 1198833 introduced in Section 33 The output of neuralnetwork is the depth of the dent

The measured data input into the BP neural network tocalculate the result of the dent Parts of sample data inputinto the network for training and the other sample data toverify the accuracy The calculation is shown in Table 3 andFigure 7Themean relative error for BP calculation and actualdepth is 1226 Although the BP neural network can be usedto calculate the depth from the bending strain the accuracydegree is low to assess the dent

To verify the accuracy of the proposed method all ofdata input into the modified Bayesian neural network Theneural network can be trained by the sample data Meansquare error (MSE) is selected to assess the quality of theneural network To adjust the parameter of neural networkthe result of training is shown in Figure 8The best validationperformance ofMSE for themethod is 00040885 at the epoch95

6 Journal of Sensors

Table 2 Part of dents of training data

Depth ratio () Depth (mm) 1198781() 119878

2() 119878

3() 119883

1(m) 119883

2(m) 119883

3(m)

202 1645 0067 0067 0079 5 18 9176 1429 0111 0111 0023 9 16 2139 1128 0016 0016 0046 1 23 6149 1213 0128 0128 0149 12 17 13199 1615 0068 0068 0101 6 15 9188 153 0055 0055 0080 3 14 5

Table 3 Calculation and error for BP neural network

Actual depth (mm) Calculationof BP (mm)

Error for calculation of BPand actual depth (mm)

Relative error()

147153 1338198 133332 906158535 1273971 311379 1964151218 1285353 226827 1500151218 1310556 201624 1333113007 1332507 202437 179196747 1019502 052032 538152844 1443888 084552 553

Mea

n sq

uare

d er

ror (

MSE

)

Best validation performance is 00040885 at epoch 95

TrainValidation

TestBest

0 10 20 30 40 50 60 70 80 90 100101 epochs

100

10minus1

10minus2

10minus3

Figure 8 Training for modified neural network based on Bayesian

The calculation of modified neural network based onBayesian is shown in Figure 9 and Table 4 It is clear to seethe method can be used to calculate the depth of dent frombending strain From Table 4 the mean relative error for cal-culation and actual depth is 244The accuracy of modifiedneural network is raised more than BP neural network Thecalculation of modified Bayesian neural network can be usedto calculate the depth of dent from dent With this methodthe pipeline company runs the IMU tool for once not only

1 2 3 4 5 6 7minus05

0

05

1

15

2

n

Dep

th (m

m)

Actual valueCalculation for Bayesian neural networkError

Figure 9 Calculation and error for modified Bayesian neuralnetwork

to obtain the pipeline bending strain but also to calculate allof the depths of dent The proposed method offers a usefulmethod for pipeline integrity evaluation

5 Conclusions

The in-line inspection tool which loads the IMU used toinspect the centerline of the pipeline The attitude infor-mation can be used to compute the bending strain of thepipeline However there is no paper or report to researchthe relationship between the bending strain and dent In thispaper based on the analysis of the calculation method of

Journal of Sensors 7

Table 4 Calculation and error for modified Bayesian neural network

Actual depth (mm) Calculation ofBayesian (mm)

Error for calculation of Bayesianand actual depth (mm) Relative error ()

147153 1498349 0268193 182158535 157455 0108 068151218 1417717 094463 625151218 1496672 015508 103113007 1107863 022207 19796747 1010614 0431435 446152844 1541688 0132484 087

pipeline bending strain we propose amethod based onmod-ified Bayesian neural network to calculate dent depths frombending strain To test the proposed method experimentallya PIG with the proposed method is used to inspect a 150 kmpipeline It can be obtained that

(1) A new method is proposed to verify the relationshipbetween pipeline bending strain and dent based onthe calculation of bending strain

(2) According to the characteristic of signal betweendent and pipeline bending strain a new model ispresented in detail to be used for proposed calculationof algorithm

(3) The calculation of modified Bayesian neural networkis proposed to compute the depth dent with pipelinebending strain to compare with actual data Andtraditional BP neural network is used to calculatethe depth of dent According to the result of calcula-tion the mean accuracy of calculation for modifiedBayesian neural network is better than BP neuralnetwork for 982

(4) According to calculation the proposed method ismore accurate and suitable for the calculation ofdepth based on pipeline bending strain The meanrelative error for calculation of dent depth based themodified Bayesian neural network is 244

This paper provides a novel method for calculating dentdepths from the pipeline bending strain The dent can beevaluated by the bending strain not used to dig or use anothertool to reinspect The bending strain and calculation of dentdepth is also useful to the evaluation of pipeline integrity

Competing Interests

The authors declare that they have no competing interests

Acknowledgments

Thisworkwas supported by the project of Petrochina PipelineCompany ldquothe research of safety service of buried pipeline inpermafrost regionrdquo

References

[1] Q Aihua ldquoChinese oil and gas pipeline transport develop-ment status and analysis of associated problemsrdquo InternationalPetroleum Economics vol 17 no 12 pp 17ndash12 2009

[2] L-J Yang R-Q Li S-W Gao and Y Yang ldquoNavigating andpositioning technique for inner detection of pipelinerdquo Journalof Shenyang University of Technology vol 34 no 4 pp 427ndash4322012

[3] S Sha and Q Feng ldquoExcavated verification technology of pitdefects of buried pipelinerdquoOil and Gas Storage and Transporta-tion vol 8 pp 834ndash838 2014

[4] HGhaednia SDas RWang andRKania ldquoSafe burst strengthof a pipeline with dent-crack defect effect of crack depth andoperating pressurerdquo Engineering Failure Analysis vol 55 pp288ndash299 2015

[5] B Pinheiro I Pasqualino and S Cunha ldquoFatigue life assess-ment of damaged pipelines under cyclic internal pressurepipelines with longitudinal and transverse plain dentsrdquo Inter-national Journal of Fatigue vol 68 pp 38ndash47 2014

[6] M Allouti C Schmitt and G Pluvinage ldquoAssessment of agouge and dent defect in a pipeline by a combined criterionrdquoEngineering Failure Analysis vol 36 pp 1ndash13 2014

[7] S Paeper B Brown and T Beuker ldquoInline inspection of dentsand corrosion using lsquohigh qualityrsquo multi-purpose smart-piginspection datardquo in Proceedings of the 6th International PipelineConference (IPC rsquo06) pp 243ndash248 September 2006

[8] X Wang and H Song ldquoThe inertial technology based 3-dimensional information measurement system for under-ground pipelinerdquoMeasurement vol 45 no 3 pp 604ndash614 2012

[9] LMengjie ldquoThe study of accurate in-line inspection technologyto offshore pipeline routerdquo China Offshore Platform vol 19 no6 pp 46ndash49 2004

[10] J Yu J G Lee C G Park andH S Han ldquoAn off-line navigationof a geometry PIG using a modified nonlinear fixed-intervalsmoothing filterrdquoControl Engineering Practice vol 13 no 11 pp1403ndash1411 2005

[11] J A Czyz C Fraccaroli andA P Sergeant ldquoMeasuring pipelinemovement in geotechnically unstable areas using an inertialgeometry pipeline inspection pigrdquo in Proceedings of the ASME1st International Pipeline Conference Calgary Canada June1996

[12] J A Czyz and J Falk ldquoUse of geopig for prevention of pipelinefailures in environmentally sensitive areasrdquo in Proceedings ofthe Pipeline Pigging Integrity Assessment and Repair ConferenceHouston Tex USA February 2000

8 Journal of Sensors

[13] I B Iflefel D G Moffat and J Mistry ldquoThe interaction ofpressure and bending on a dented piperdquo International Journalof Pressure Vessels and Piping vol 82 no 10 pp 761ndash769 2005

[14] J Błachut and I B Iflefel ldquoCollapse of pipeswith plain or gougeddents by bending momentrdquo International Journal of PressureVessels and Piping vol 84 no 9 pp 560ndash571 2007

[15] J-H Baek Y-P Kim W-S Kim J-M Koo and C-S SeokldquoLoad bearing capacity of API X65 pipe with dent defect underinternal pressure and in-plane bendingrdquo Materials Science andEngineering A vol 540 pp 70ndash82 2012

[16] A Limam L-H Lee and S Kyriakides ldquoOn the collapse ofdented tubes under combined bending and internal pressurerdquoInternational Journal of Mechanical Sciences vol 55 no 1 pp1ndash12 2012

[17] D K Kim S H Cho S S Park H R Yoo and Y W RholdquoDesign and implementation of 3010158401015840 geometry PIGrdquo KSMEInternational Journal vol 17 no 5 pp 629ndash636 2003

[18] J A Czyz et al ldquoMulti-pipeline geographical informationsystem based on high accuracy inertial surveysrdquo in Proceedingsof the ASME 3rd International Pipeline Conference CalgaryCanada 2000

[19] J D Hart N Zulfiqar D H Moore and G R Swank ldquolsquoDigitalPiggingrsquo as a basis for improved pipeline structural integrityevaluationsrdquo in Proceedings of the International Pipeline Con-ference vol 2 of Integrity Management Poster Session StudentPaper Competition Calgary Canada September 2006

[20] J A Czyz and J RAdams ldquoComputations of pipelinemdashbendingstrains based on geopig measurementsrdquo in Proceedings of theipeline Pigging and Integrity Monitoring Conference pp 14ndash17Houston Tex USA February 1994

[21] J D Hart and G H Powel Geometry Monitoring of the Trans-Alaska Pipeline Trans-Alaska Pipeline System (TAPS) 2005

[22] D Yu L ZhangW Liang Y Ye and ZWang ldquoNoise reductionof signal and condition recognition of long-distance pipelinerdquoActa Petrolei Sinica vol 30 no 6 pp 937ndash941 2009

[23] J Zhou BGuiM Li andWLin ldquoAn application of the artificialneural net dominated by lithology to permeability predictionrdquoActa Petrolei Sinica vol 31 no 6 pp 985ndash988 2010

[24] D Li D Lu X Kong and Y Du ldquoProcessing of well log databased on backpropagation neural network implicit approxima-tionrdquo Acta Petrolei Sinica vol 28 no 3 pp 105ndash108 2007

[25] K-X Peng and S-Z Yu ldquoPrediction and control of strip thick-ness based on Bayesian neural networksrdquo Journal of Universityof Science and Technology Beijing vol 32 no 2 pp 256ndash2622010

[26] H Yang and H-Z Fu ldquoStock index forecast based on bayesianregularization BP neural networkrdquo Science Techno Logy andEngineering vol 6 pp 3306ndash3310 2009

[27] H Wang L Shi H Zhao and Y Yue ldquoTemperature predictionmodel for sunlight greenhouse based on bayesian regularizationBP neural networkrdquo Hubei Agricultural Sciences vol 9 pp4300ndash4303 2015

[28] Y Shenghua and D Juan ldquoGDP prediction based on principalcomponent analysis and bayesian regularization BP neuralnetworkrdquo Journal of Hunan University (Social Sciences) vol 11pp 42ndash45 2011

[29] R Li Q Feng M Cai et al ldquoMeasurement of long-distanceburied pipeline centerline based on multi-sensor data fusionrdquoActa Petrolei Sinica vol 35 no 5 pp 987ndash992 2014

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

4 Journal of Sensors

and then the solution for 1198640 is stable The condition that119891(119909) has interpolation ability is that 119891(119909) is smooth Whilethe square sum of network weight and threshold is smallits output is smoother Therefore constraint item shouldrepresent smooth constraint and the objective function isexpressed as follows

119864ALL = 1205720 1

2

119882

sum

119894=1

1199082

119894+ 1205730 1

2

119870

sum

119896=1

119873

sum

119899=1

(1199101015840

119899119896minus 119910119899119896)

2

= 120572119864net (119908) + 120573119864net (119910)

(14)

119882 is the neural network weight and threshold numbers119908 is network weight and threshold 119864net(119908) is the square sumof weight and threshold 119864net(119910) is the square sum of networkactual output and objective expectation value residual and 120572and 120573 are hyperparameter which decide the value size of theneural network objective performance function and controlthe distribution form of weight and thresholdThrough usingnew objective performance function the network weightand threshold are as little as possible on condition that thenetwork training error is as small as possible According toBayesian posterior distribution probability of 120572 and 120573meetsthe following formula

119875 (120572 120573 | 119863119873ℎ) =

119875 (119863 | 120572 120573119873ℎ) 119875 (120572 120573 | 119873ℎ)

119875 (119863 | 119873ℎ)

(15)

119863 = 119909(119873) 119910(119873) is a sample data set which consistsof 119873 samples 119873ℎ is neural network hidden layer number119875(120572 120573 | 119873ℎ) is hyperparameter prior probability 119875(119863 | 119873ℎ)

is normalization factor and 119875(119863 | 120572 120573119873ℎ) is likelihoodfunction After partial derivatives to 120572 120573 the most salienthyperparameter can be carried out

120572119872119875 =

120574

2119864net (119908119872119875)

120573119872119875 =

119873 minus 120574

2119864net (119908119872119875)

(16)

120574 is the number that can reduce parameters numbers (0ndashW) in performance index function parameters in network119908119872119875 is the weight and threshold when 119864ALL is minimum119872119875 is the subscript for 119908 119908119872119875 is the weight and thresholdwhen 119864ALL is minimum After working out the most salienthyperparameter 120572 120573 determine whether 119864ALL is convergenceagain Through iterative judgment the best and most salientneural network model can be obtained

Relative to traditional neural network the Bayesian neu-ral network is focused on the probability distribution of thewhole parameter space and the predicted results are basedon the statistical average of parameters posterior distributionA single model is corresponding to one point in the spaceall models are corresponding to the whole parameter spaceTherefore the Bayesian neural network guarantees networkstronger generalization in theory

33 Modeling and Computation for Dent by Bending Strain Atypical signal of dent feature from bending strain is shown in

Input the training data

Initialize the weight and threshold fornetwork

Training the neural network with L-M algorithm tocalculate the minimum target function

Whether convergence

Calculated the posterior distribution probability to checkthe modeling significant

Other modeling

Training ended and output thenetwork modeling

Yes

Yes

No

No

To calculate the parameters for 120574120572 and 120573

Figure 4 Dent modeling and computation based on Bayesiantheory neural network

Figure 3The horizontal axis represents the dent length alongthe pipeline and the vertical axis is the bending strain for dentrespectively The bending strain can be represented for theseverity for the dent To the identity of the dent feature theinput can be related to the bending strain information for thedent The output selected the actual depth of the dent fromthe site measurement

According to the relationship for the bending strain anddent as shown in Figure 3 pattern in a plot of bending strainagainst distance shows the characteristic of environmentallyinduced deformation for dent The pattern is caused by aprimary action in the center with a reaction (bending strainin the opposite direction) on both sides In the case of a dentthe central deformation is an underbend with overbends ateither end The main influencing parameters are includingthe beginning bending strain 1198781 and ending bending strain1198783 which is related to the reaction area of dent along thepipeline respectively The largest bending strain 1198782 is relatedto the deepest of the dent 1198831 is the distance for reactionof the beginning bending strain for dent 1198832 is the distancefor primary action for dent 1198833 is the distance for anotherreaction of the ending bending strain for dent The processof modeling and computation for the dent based on Bayesiantheory neural network is shown in Figure 4

4 Field Test and Data Analysis

41 Equipment and Performance To test the proposedmethod experimentally an in-line inspection tool (shown in

Journal of Sensors 5

Table 1 Characteristics of sensors

Sensor Characteristics Magnitude

Gyroscope Bias lt001∘hRandom walk 0002

∘radich

Accelerometer Bias stability lt50 120583gScaling factor lt50 ppm

Odometer Scaling factor lt03White noise lt01ms

Landmark White noise ltplusmn1m

Figure 5 IMU Pipeline Inspection Tool

Figure 5) with the proposed method is used to inspect oilpipeline which is about 150 kilometers About 150 dents havebeen dug tomeasure and verify depth gauge in past five yearsThese actual data can be selected to be used for sample for theproposed algorithm

The IMUpipeline ILI toolmainly consists of the followingsensors to be used to calculate the bending strain and positionof the pipeline

(1) Inertialmeasurement unit is themain device to gatherthe ILI data Three gyroscopes and three accelerom-eters orthogonally mounted on the IMU to measurethe attitude and positon for ILI tool These gathereddata can be used to calculate the pipeline bendingstrain with other sensors mounted on the ILI tool

(2) Two odometers mounted on the ILI to measure thedistance and instantaneous and average speed whichcan be used to modify the inertial system errors

(3) Two weld detector sensors are used to measure andrecord time of passing each girth weld for tool Theycan use the alignment for every spool for repeatinspection

As is known most domestic oil pipelines are heatingtransportation and pass through mountains hills rivers andother complex environmental areas [29] the technical andsafety requirements of electrical equipment are extremelystrict To safely inspect the actual pipeline the IMU shouldmeet not only the technical performance but also the actualsituation of the pipe and the external environment shouldbe considered Technical performances of inertial deviceswhich are used in the field test are shown in Table 1

42 Data Analysis According to Section 2 a typical bendingstrain for dent is shown in Figure 6 whose depth is 5279mm

201

201

2

201

4

201

6

201

8

202

202

2

202

4

202

6

202

8

203

minus05minus04minus03minus02minus01

0010203

Distance (m)

Bend

ing

strai

n (

)

times104

Figure 6 Typical bending strain of a dent

1 2 3 4 5 6 702468

10121416

n

Dep

th (m

m)

Calculation for BPActual valueError

Figure 7 Calculation and error for BP neural network

for 65 of outer diameter measured with gauge in field Thebending strain for this dent is minus04125 which is located inbottom of the pipeline

For the method of Section 3 120 data sets for bendingstrain of dent are selected for the input of the new neuralnetwork Part of dents of training data are shown in Table 2The input of neural network is consisted of 1198781 1198782 1198783 11988311198832 and 1198833 introduced in Section 33 The output of neuralnetwork is the depth of the dent

The measured data input into the BP neural network tocalculate the result of the dent Parts of sample data inputinto the network for training and the other sample data toverify the accuracy The calculation is shown in Table 3 andFigure 7Themean relative error for BP calculation and actualdepth is 1226 Although the BP neural network can be usedto calculate the depth from the bending strain the accuracydegree is low to assess the dent

To verify the accuracy of the proposed method all ofdata input into the modified Bayesian neural network Theneural network can be trained by the sample data Meansquare error (MSE) is selected to assess the quality of theneural network To adjust the parameter of neural networkthe result of training is shown in Figure 8The best validationperformance ofMSE for themethod is 00040885 at the epoch95

6 Journal of Sensors

Table 2 Part of dents of training data

Depth ratio () Depth (mm) 1198781() 119878

2() 119878

3() 119883

1(m) 119883

2(m) 119883

3(m)

202 1645 0067 0067 0079 5 18 9176 1429 0111 0111 0023 9 16 2139 1128 0016 0016 0046 1 23 6149 1213 0128 0128 0149 12 17 13199 1615 0068 0068 0101 6 15 9188 153 0055 0055 0080 3 14 5

Table 3 Calculation and error for BP neural network

Actual depth (mm) Calculationof BP (mm)

Error for calculation of BPand actual depth (mm)

Relative error()

147153 1338198 133332 906158535 1273971 311379 1964151218 1285353 226827 1500151218 1310556 201624 1333113007 1332507 202437 179196747 1019502 052032 538152844 1443888 084552 553

Mea

n sq

uare

d er

ror (

MSE

)

Best validation performance is 00040885 at epoch 95

TrainValidation

TestBest

0 10 20 30 40 50 60 70 80 90 100101 epochs

100

10minus1

10minus2

10minus3

Figure 8 Training for modified neural network based on Bayesian

The calculation of modified neural network based onBayesian is shown in Figure 9 and Table 4 It is clear to seethe method can be used to calculate the depth of dent frombending strain From Table 4 the mean relative error for cal-culation and actual depth is 244The accuracy of modifiedneural network is raised more than BP neural network Thecalculation of modified Bayesian neural network can be usedto calculate the depth of dent from dent With this methodthe pipeline company runs the IMU tool for once not only

1 2 3 4 5 6 7minus05

0

05

1

15

2

n

Dep

th (m

m)

Actual valueCalculation for Bayesian neural networkError

Figure 9 Calculation and error for modified Bayesian neuralnetwork

to obtain the pipeline bending strain but also to calculate allof the depths of dent The proposed method offers a usefulmethod for pipeline integrity evaluation

5 Conclusions

The in-line inspection tool which loads the IMU used toinspect the centerline of the pipeline The attitude infor-mation can be used to compute the bending strain of thepipeline However there is no paper or report to researchthe relationship between the bending strain and dent In thispaper based on the analysis of the calculation method of

Journal of Sensors 7

Table 4 Calculation and error for modified Bayesian neural network

Actual depth (mm) Calculation ofBayesian (mm)

Error for calculation of Bayesianand actual depth (mm) Relative error ()

147153 1498349 0268193 182158535 157455 0108 068151218 1417717 094463 625151218 1496672 015508 103113007 1107863 022207 19796747 1010614 0431435 446152844 1541688 0132484 087

pipeline bending strain we propose amethod based onmod-ified Bayesian neural network to calculate dent depths frombending strain To test the proposed method experimentallya PIG with the proposed method is used to inspect a 150 kmpipeline It can be obtained that

(1) A new method is proposed to verify the relationshipbetween pipeline bending strain and dent based onthe calculation of bending strain

(2) According to the characteristic of signal betweendent and pipeline bending strain a new model ispresented in detail to be used for proposed calculationof algorithm

(3) The calculation of modified Bayesian neural networkis proposed to compute the depth dent with pipelinebending strain to compare with actual data Andtraditional BP neural network is used to calculatethe depth of dent According to the result of calcula-tion the mean accuracy of calculation for modifiedBayesian neural network is better than BP neuralnetwork for 982

(4) According to calculation the proposed method ismore accurate and suitable for the calculation ofdepth based on pipeline bending strain The meanrelative error for calculation of dent depth based themodified Bayesian neural network is 244

This paper provides a novel method for calculating dentdepths from the pipeline bending strain The dent can beevaluated by the bending strain not used to dig or use anothertool to reinspect The bending strain and calculation of dentdepth is also useful to the evaluation of pipeline integrity

Competing Interests

The authors declare that they have no competing interests

Acknowledgments

Thisworkwas supported by the project of Petrochina PipelineCompany ldquothe research of safety service of buried pipeline inpermafrost regionrdquo

References

[1] Q Aihua ldquoChinese oil and gas pipeline transport develop-ment status and analysis of associated problemsrdquo InternationalPetroleum Economics vol 17 no 12 pp 17ndash12 2009

[2] L-J Yang R-Q Li S-W Gao and Y Yang ldquoNavigating andpositioning technique for inner detection of pipelinerdquo Journalof Shenyang University of Technology vol 34 no 4 pp 427ndash4322012

[3] S Sha and Q Feng ldquoExcavated verification technology of pitdefects of buried pipelinerdquoOil and Gas Storage and Transporta-tion vol 8 pp 834ndash838 2014

[4] HGhaednia SDas RWang andRKania ldquoSafe burst strengthof a pipeline with dent-crack defect effect of crack depth andoperating pressurerdquo Engineering Failure Analysis vol 55 pp288ndash299 2015

[5] B Pinheiro I Pasqualino and S Cunha ldquoFatigue life assess-ment of damaged pipelines under cyclic internal pressurepipelines with longitudinal and transverse plain dentsrdquo Inter-national Journal of Fatigue vol 68 pp 38ndash47 2014

[6] M Allouti C Schmitt and G Pluvinage ldquoAssessment of agouge and dent defect in a pipeline by a combined criterionrdquoEngineering Failure Analysis vol 36 pp 1ndash13 2014

[7] S Paeper B Brown and T Beuker ldquoInline inspection of dentsand corrosion using lsquohigh qualityrsquo multi-purpose smart-piginspection datardquo in Proceedings of the 6th International PipelineConference (IPC rsquo06) pp 243ndash248 September 2006

[8] X Wang and H Song ldquoThe inertial technology based 3-dimensional information measurement system for under-ground pipelinerdquoMeasurement vol 45 no 3 pp 604ndash614 2012

[9] LMengjie ldquoThe study of accurate in-line inspection technologyto offshore pipeline routerdquo China Offshore Platform vol 19 no6 pp 46ndash49 2004

[10] J Yu J G Lee C G Park andH S Han ldquoAn off-line navigationof a geometry PIG using a modified nonlinear fixed-intervalsmoothing filterrdquoControl Engineering Practice vol 13 no 11 pp1403ndash1411 2005

[11] J A Czyz C Fraccaroli andA P Sergeant ldquoMeasuring pipelinemovement in geotechnically unstable areas using an inertialgeometry pipeline inspection pigrdquo in Proceedings of the ASME1st International Pipeline Conference Calgary Canada June1996

[12] J A Czyz and J Falk ldquoUse of geopig for prevention of pipelinefailures in environmentally sensitive areasrdquo in Proceedings ofthe Pipeline Pigging Integrity Assessment and Repair ConferenceHouston Tex USA February 2000

8 Journal of Sensors

[13] I B Iflefel D G Moffat and J Mistry ldquoThe interaction ofpressure and bending on a dented piperdquo International Journalof Pressure Vessels and Piping vol 82 no 10 pp 761ndash769 2005

[14] J Błachut and I B Iflefel ldquoCollapse of pipeswith plain or gougeddents by bending momentrdquo International Journal of PressureVessels and Piping vol 84 no 9 pp 560ndash571 2007

[15] J-H Baek Y-P Kim W-S Kim J-M Koo and C-S SeokldquoLoad bearing capacity of API X65 pipe with dent defect underinternal pressure and in-plane bendingrdquo Materials Science andEngineering A vol 540 pp 70ndash82 2012

[16] A Limam L-H Lee and S Kyriakides ldquoOn the collapse ofdented tubes under combined bending and internal pressurerdquoInternational Journal of Mechanical Sciences vol 55 no 1 pp1ndash12 2012

[17] D K Kim S H Cho S S Park H R Yoo and Y W RholdquoDesign and implementation of 3010158401015840 geometry PIGrdquo KSMEInternational Journal vol 17 no 5 pp 629ndash636 2003

[18] J A Czyz et al ldquoMulti-pipeline geographical informationsystem based on high accuracy inertial surveysrdquo in Proceedingsof the ASME 3rd International Pipeline Conference CalgaryCanada 2000

[19] J D Hart N Zulfiqar D H Moore and G R Swank ldquolsquoDigitalPiggingrsquo as a basis for improved pipeline structural integrityevaluationsrdquo in Proceedings of the International Pipeline Con-ference vol 2 of Integrity Management Poster Session StudentPaper Competition Calgary Canada September 2006

[20] J A Czyz and J RAdams ldquoComputations of pipelinemdashbendingstrains based on geopig measurementsrdquo in Proceedings of theipeline Pigging and Integrity Monitoring Conference pp 14ndash17Houston Tex USA February 1994

[21] J D Hart and G H Powel Geometry Monitoring of the Trans-Alaska Pipeline Trans-Alaska Pipeline System (TAPS) 2005

[22] D Yu L ZhangW Liang Y Ye and ZWang ldquoNoise reductionof signal and condition recognition of long-distance pipelinerdquoActa Petrolei Sinica vol 30 no 6 pp 937ndash941 2009

[23] J Zhou BGuiM Li andWLin ldquoAn application of the artificialneural net dominated by lithology to permeability predictionrdquoActa Petrolei Sinica vol 31 no 6 pp 985ndash988 2010

[24] D Li D Lu X Kong and Y Du ldquoProcessing of well log databased on backpropagation neural network implicit approxima-tionrdquo Acta Petrolei Sinica vol 28 no 3 pp 105ndash108 2007

[25] K-X Peng and S-Z Yu ldquoPrediction and control of strip thick-ness based on Bayesian neural networksrdquo Journal of Universityof Science and Technology Beijing vol 32 no 2 pp 256ndash2622010

[26] H Yang and H-Z Fu ldquoStock index forecast based on bayesianregularization BP neural networkrdquo Science Techno Logy andEngineering vol 6 pp 3306ndash3310 2009

[27] H Wang L Shi H Zhao and Y Yue ldquoTemperature predictionmodel for sunlight greenhouse based on bayesian regularizationBP neural networkrdquo Hubei Agricultural Sciences vol 9 pp4300ndash4303 2015

[28] Y Shenghua and D Juan ldquoGDP prediction based on principalcomponent analysis and bayesian regularization BP neuralnetworkrdquo Journal of Hunan University (Social Sciences) vol 11pp 42ndash45 2011

[29] R Li Q Feng M Cai et al ldquoMeasurement of long-distanceburied pipeline centerline based on multi-sensor data fusionrdquoActa Petrolei Sinica vol 35 no 5 pp 987ndash992 2014

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Journal of Sensors 5

Table 1 Characteristics of sensors

Sensor Characteristics Magnitude

Gyroscope Bias lt001∘hRandom walk 0002

∘radich

Accelerometer Bias stability lt50 120583gScaling factor lt50 ppm

Odometer Scaling factor lt03White noise lt01ms

Landmark White noise ltplusmn1m

Figure 5 IMU Pipeline Inspection Tool

Figure 5) with the proposed method is used to inspect oilpipeline which is about 150 kilometers About 150 dents havebeen dug tomeasure and verify depth gauge in past five yearsThese actual data can be selected to be used for sample for theproposed algorithm

The IMUpipeline ILI toolmainly consists of the followingsensors to be used to calculate the bending strain and positionof the pipeline

(1) Inertialmeasurement unit is themain device to gatherthe ILI data Three gyroscopes and three accelerom-eters orthogonally mounted on the IMU to measurethe attitude and positon for ILI tool These gathereddata can be used to calculate the pipeline bendingstrain with other sensors mounted on the ILI tool

(2) Two odometers mounted on the ILI to measure thedistance and instantaneous and average speed whichcan be used to modify the inertial system errors

(3) Two weld detector sensors are used to measure andrecord time of passing each girth weld for tool Theycan use the alignment for every spool for repeatinspection

As is known most domestic oil pipelines are heatingtransportation and pass through mountains hills rivers andother complex environmental areas [29] the technical andsafety requirements of electrical equipment are extremelystrict To safely inspect the actual pipeline the IMU shouldmeet not only the technical performance but also the actualsituation of the pipe and the external environment shouldbe considered Technical performances of inertial deviceswhich are used in the field test are shown in Table 1

42 Data Analysis According to Section 2 a typical bendingstrain for dent is shown in Figure 6 whose depth is 5279mm

201

201

2

201

4

201

6

201

8

202

202

2

202

4

202

6

202

8

203

minus05minus04minus03minus02minus01

0010203

Distance (m)

Bend

ing

strai

n (

)

times104

Figure 6 Typical bending strain of a dent

1 2 3 4 5 6 702468

10121416

n

Dep

th (m

m)

Calculation for BPActual valueError

Figure 7 Calculation and error for BP neural network

for 65 of outer diameter measured with gauge in field Thebending strain for this dent is minus04125 which is located inbottom of the pipeline

For the method of Section 3 120 data sets for bendingstrain of dent are selected for the input of the new neuralnetwork Part of dents of training data are shown in Table 2The input of neural network is consisted of 1198781 1198782 1198783 11988311198832 and 1198833 introduced in Section 33 The output of neuralnetwork is the depth of the dent

The measured data input into the BP neural network tocalculate the result of the dent Parts of sample data inputinto the network for training and the other sample data toverify the accuracy The calculation is shown in Table 3 andFigure 7Themean relative error for BP calculation and actualdepth is 1226 Although the BP neural network can be usedto calculate the depth from the bending strain the accuracydegree is low to assess the dent

To verify the accuracy of the proposed method all ofdata input into the modified Bayesian neural network Theneural network can be trained by the sample data Meansquare error (MSE) is selected to assess the quality of theneural network To adjust the parameter of neural networkthe result of training is shown in Figure 8The best validationperformance ofMSE for themethod is 00040885 at the epoch95

6 Journal of Sensors

Table 2 Part of dents of training data

Depth ratio () Depth (mm) 1198781() 119878

2() 119878

3() 119883

1(m) 119883

2(m) 119883

3(m)

202 1645 0067 0067 0079 5 18 9176 1429 0111 0111 0023 9 16 2139 1128 0016 0016 0046 1 23 6149 1213 0128 0128 0149 12 17 13199 1615 0068 0068 0101 6 15 9188 153 0055 0055 0080 3 14 5

Table 3 Calculation and error for BP neural network

Actual depth (mm) Calculationof BP (mm)

Error for calculation of BPand actual depth (mm)

Relative error()

147153 1338198 133332 906158535 1273971 311379 1964151218 1285353 226827 1500151218 1310556 201624 1333113007 1332507 202437 179196747 1019502 052032 538152844 1443888 084552 553

Mea

n sq

uare

d er

ror (

MSE

)

Best validation performance is 00040885 at epoch 95

TrainValidation

TestBest

0 10 20 30 40 50 60 70 80 90 100101 epochs

100

10minus1

10minus2

10minus3

Figure 8 Training for modified neural network based on Bayesian

The calculation of modified neural network based onBayesian is shown in Figure 9 and Table 4 It is clear to seethe method can be used to calculate the depth of dent frombending strain From Table 4 the mean relative error for cal-culation and actual depth is 244The accuracy of modifiedneural network is raised more than BP neural network Thecalculation of modified Bayesian neural network can be usedto calculate the depth of dent from dent With this methodthe pipeline company runs the IMU tool for once not only

1 2 3 4 5 6 7minus05

0

05

1

15

2

n

Dep

th (m

m)

Actual valueCalculation for Bayesian neural networkError

Figure 9 Calculation and error for modified Bayesian neuralnetwork

to obtain the pipeline bending strain but also to calculate allof the depths of dent The proposed method offers a usefulmethod for pipeline integrity evaluation

5 Conclusions

The in-line inspection tool which loads the IMU used toinspect the centerline of the pipeline The attitude infor-mation can be used to compute the bending strain of thepipeline However there is no paper or report to researchthe relationship between the bending strain and dent In thispaper based on the analysis of the calculation method of

Journal of Sensors 7

Table 4 Calculation and error for modified Bayesian neural network

Actual depth (mm) Calculation ofBayesian (mm)

Error for calculation of Bayesianand actual depth (mm) Relative error ()

147153 1498349 0268193 182158535 157455 0108 068151218 1417717 094463 625151218 1496672 015508 103113007 1107863 022207 19796747 1010614 0431435 446152844 1541688 0132484 087

pipeline bending strain we propose amethod based onmod-ified Bayesian neural network to calculate dent depths frombending strain To test the proposed method experimentallya PIG with the proposed method is used to inspect a 150 kmpipeline It can be obtained that

(1) A new method is proposed to verify the relationshipbetween pipeline bending strain and dent based onthe calculation of bending strain

(2) According to the characteristic of signal betweendent and pipeline bending strain a new model ispresented in detail to be used for proposed calculationof algorithm

(3) The calculation of modified Bayesian neural networkis proposed to compute the depth dent with pipelinebending strain to compare with actual data Andtraditional BP neural network is used to calculatethe depth of dent According to the result of calcula-tion the mean accuracy of calculation for modifiedBayesian neural network is better than BP neuralnetwork for 982

(4) According to calculation the proposed method ismore accurate and suitable for the calculation ofdepth based on pipeline bending strain The meanrelative error for calculation of dent depth based themodified Bayesian neural network is 244

This paper provides a novel method for calculating dentdepths from the pipeline bending strain The dent can beevaluated by the bending strain not used to dig or use anothertool to reinspect The bending strain and calculation of dentdepth is also useful to the evaluation of pipeline integrity

Competing Interests

The authors declare that they have no competing interests

Acknowledgments

Thisworkwas supported by the project of Petrochina PipelineCompany ldquothe research of safety service of buried pipeline inpermafrost regionrdquo

References

[1] Q Aihua ldquoChinese oil and gas pipeline transport develop-ment status and analysis of associated problemsrdquo InternationalPetroleum Economics vol 17 no 12 pp 17ndash12 2009

[2] L-J Yang R-Q Li S-W Gao and Y Yang ldquoNavigating andpositioning technique for inner detection of pipelinerdquo Journalof Shenyang University of Technology vol 34 no 4 pp 427ndash4322012

[3] S Sha and Q Feng ldquoExcavated verification technology of pitdefects of buried pipelinerdquoOil and Gas Storage and Transporta-tion vol 8 pp 834ndash838 2014

[4] HGhaednia SDas RWang andRKania ldquoSafe burst strengthof a pipeline with dent-crack defect effect of crack depth andoperating pressurerdquo Engineering Failure Analysis vol 55 pp288ndash299 2015

[5] B Pinheiro I Pasqualino and S Cunha ldquoFatigue life assess-ment of damaged pipelines under cyclic internal pressurepipelines with longitudinal and transverse plain dentsrdquo Inter-national Journal of Fatigue vol 68 pp 38ndash47 2014

[6] M Allouti C Schmitt and G Pluvinage ldquoAssessment of agouge and dent defect in a pipeline by a combined criterionrdquoEngineering Failure Analysis vol 36 pp 1ndash13 2014

[7] S Paeper B Brown and T Beuker ldquoInline inspection of dentsand corrosion using lsquohigh qualityrsquo multi-purpose smart-piginspection datardquo in Proceedings of the 6th International PipelineConference (IPC rsquo06) pp 243ndash248 September 2006

[8] X Wang and H Song ldquoThe inertial technology based 3-dimensional information measurement system for under-ground pipelinerdquoMeasurement vol 45 no 3 pp 604ndash614 2012

[9] LMengjie ldquoThe study of accurate in-line inspection technologyto offshore pipeline routerdquo China Offshore Platform vol 19 no6 pp 46ndash49 2004

[10] J Yu J G Lee C G Park andH S Han ldquoAn off-line navigationof a geometry PIG using a modified nonlinear fixed-intervalsmoothing filterrdquoControl Engineering Practice vol 13 no 11 pp1403ndash1411 2005

[11] J A Czyz C Fraccaroli andA P Sergeant ldquoMeasuring pipelinemovement in geotechnically unstable areas using an inertialgeometry pipeline inspection pigrdquo in Proceedings of the ASME1st International Pipeline Conference Calgary Canada June1996

[12] J A Czyz and J Falk ldquoUse of geopig for prevention of pipelinefailures in environmentally sensitive areasrdquo in Proceedings ofthe Pipeline Pigging Integrity Assessment and Repair ConferenceHouston Tex USA February 2000

8 Journal of Sensors

[13] I B Iflefel D G Moffat and J Mistry ldquoThe interaction ofpressure and bending on a dented piperdquo International Journalof Pressure Vessels and Piping vol 82 no 10 pp 761ndash769 2005

[14] J Błachut and I B Iflefel ldquoCollapse of pipeswith plain or gougeddents by bending momentrdquo International Journal of PressureVessels and Piping vol 84 no 9 pp 560ndash571 2007

[15] J-H Baek Y-P Kim W-S Kim J-M Koo and C-S SeokldquoLoad bearing capacity of API X65 pipe with dent defect underinternal pressure and in-plane bendingrdquo Materials Science andEngineering A vol 540 pp 70ndash82 2012

[16] A Limam L-H Lee and S Kyriakides ldquoOn the collapse ofdented tubes under combined bending and internal pressurerdquoInternational Journal of Mechanical Sciences vol 55 no 1 pp1ndash12 2012

[17] D K Kim S H Cho S S Park H R Yoo and Y W RholdquoDesign and implementation of 3010158401015840 geometry PIGrdquo KSMEInternational Journal vol 17 no 5 pp 629ndash636 2003

[18] J A Czyz et al ldquoMulti-pipeline geographical informationsystem based on high accuracy inertial surveysrdquo in Proceedingsof the ASME 3rd International Pipeline Conference CalgaryCanada 2000

[19] J D Hart N Zulfiqar D H Moore and G R Swank ldquolsquoDigitalPiggingrsquo as a basis for improved pipeline structural integrityevaluationsrdquo in Proceedings of the International Pipeline Con-ference vol 2 of Integrity Management Poster Session StudentPaper Competition Calgary Canada September 2006

[20] J A Czyz and J RAdams ldquoComputations of pipelinemdashbendingstrains based on geopig measurementsrdquo in Proceedings of theipeline Pigging and Integrity Monitoring Conference pp 14ndash17Houston Tex USA February 1994

[21] J D Hart and G H Powel Geometry Monitoring of the Trans-Alaska Pipeline Trans-Alaska Pipeline System (TAPS) 2005

[22] D Yu L ZhangW Liang Y Ye and ZWang ldquoNoise reductionof signal and condition recognition of long-distance pipelinerdquoActa Petrolei Sinica vol 30 no 6 pp 937ndash941 2009

[23] J Zhou BGuiM Li andWLin ldquoAn application of the artificialneural net dominated by lithology to permeability predictionrdquoActa Petrolei Sinica vol 31 no 6 pp 985ndash988 2010

[24] D Li D Lu X Kong and Y Du ldquoProcessing of well log databased on backpropagation neural network implicit approxima-tionrdquo Acta Petrolei Sinica vol 28 no 3 pp 105ndash108 2007

[25] K-X Peng and S-Z Yu ldquoPrediction and control of strip thick-ness based on Bayesian neural networksrdquo Journal of Universityof Science and Technology Beijing vol 32 no 2 pp 256ndash2622010

[26] H Yang and H-Z Fu ldquoStock index forecast based on bayesianregularization BP neural networkrdquo Science Techno Logy andEngineering vol 6 pp 3306ndash3310 2009

[27] H Wang L Shi H Zhao and Y Yue ldquoTemperature predictionmodel for sunlight greenhouse based on bayesian regularizationBP neural networkrdquo Hubei Agricultural Sciences vol 9 pp4300ndash4303 2015

[28] Y Shenghua and D Juan ldquoGDP prediction based on principalcomponent analysis and bayesian regularization BP neuralnetworkrdquo Journal of Hunan University (Social Sciences) vol 11pp 42ndash45 2011

[29] R Li Q Feng M Cai et al ldquoMeasurement of long-distanceburied pipeline centerline based on multi-sensor data fusionrdquoActa Petrolei Sinica vol 35 no 5 pp 987ndash992 2014

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

6 Journal of Sensors

Table 2 Part of dents of training data

Depth ratio () Depth (mm) 1198781() 119878

2() 119878

3() 119883

1(m) 119883

2(m) 119883

3(m)

202 1645 0067 0067 0079 5 18 9176 1429 0111 0111 0023 9 16 2139 1128 0016 0016 0046 1 23 6149 1213 0128 0128 0149 12 17 13199 1615 0068 0068 0101 6 15 9188 153 0055 0055 0080 3 14 5

Table 3 Calculation and error for BP neural network

Actual depth (mm) Calculationof BP (mm)

Error for calculation of BPand actual depth (mm)

Relative error()

147153 1338198 133332 906158535 1273971 311379 1964151218 1285353 226827 1500151218 1310556 201624 1333113007 1332507 202437 179196747 1019502 052032 538152844 1443888 084552 553

Mea

n sq

uare

d er

ror (

MSE

)

Best validation performance is 00040885 at epoch 95

TrainValidation

TestBest

0 10 20 30 40 50 60 70 80 90 100101 epochs

100

10minus1

10minus2

10minus3

Figure 8 Training for modified neural network based on Bayesian

The calculation of modified neural network based onBayesian is shown in Figure 9 and Table 4 It is clear to seethe method can be used to calculate the depth of dent frombending strain From Table 4 the mean relative error for cal-culation and actual depth is 244The accuracy of modifiedneural network is raised more than BP neural network Thecalculation of modified Bayesian neural network can be usedto calculate the depth of dent from dent With this methodthe pipeline company runs the IMU tool for once not only

1 2 3 4 5 6 7minus05

0

05

1

15

2

n

Dep

th (m

m)

Actual valueCalculation for Bayesian neural networkError

Figure 9 Calculation and error for modified Bayesian neuralnetwork

to obtain the pipeline bending strain but also to calculate allof the depths of dent The proposed method offers a usefulmethod for pipeline integrity evaluation

5 Conclusions

The in-line inspection tool which loads the IMU used toinspect the centerline of the pipeline The attitude infor-mation can be used to compute the bending strain of thepipeline However there is no paper or report to researchthe relationship between the bending strain and dent In thispaper based on the analysis of the calculation method of

Journal of Sensors 7

Table 4 Calculation and error for modified Bayesian neural network

Actual depth (mm) Calculation ofBayesian (mm)

Error for calculation of Bayesianand actual depth (mm) Relative error ()

147153 1498349 0268193 182158535 157455 0108 068151218 1417717 094463 625151218 1496672 015508 103113007 1107863 022207 19796747 1010614 0431435 446152844 1541688 0132484 087

pipeline bending strain we propose amethod based onmod-ified Bayesian neural network to calculate dent depths frombending strain To test the proposed method experimentallya PIG with the proposed method is used to inspect a 150 kmpipeline It can be obtained that

(1) A new method is proposed to verify the relationshipbetween pipeline bending strain and dent based onthe calculation of bending strain

(2) According to the characteristic of signal betweendent and pipeline bending strain a new model ispresented in detail to be used for proposed calculationof algorithm

(3) The calculation of modified Bayesian neural networkis proposed to compute the depth dent with pipelinebending strain to compare with actual data Andtraditional BP neural network is used to calculatethe depth of dent According to the result of calcula-tion the mean accuracy of calculation for modifiedBayesian neural network is better than BP neuralnetwork for 982

(4) According to calculation the proposed method ismore accurate and suitable for the calculation ofdepth based on pipeline bending strain The meanrelative error for calculation of dent depth based themodified Bayesian neural network is 244

This paper provides a novel method for calculating dentdepths from the pipeline bending strain The dent can beevaluated by the bending strain not used to dig or use anothertool to reinspect The bending strain and calculation of dentdepth is also useful to the evaluation of pipeline integrity

Competing Interests

The authors declare that they have no competing interests

Acknowledgments

Thisworkwas supported by the project of Petrochina PipelineCompany ldquothe research of safety service of buried pipeline inpermafrost regionrdquo

References

[1] Q Aihua ldquoChinese oil and gas pipeline transport develop-ment status and analysis of associated problemsrdquo InternationalPetroleum Economics vol 17 no 12 pp 17ndash12 2009

[2] L-J Yang R-Q Li S-W Gao and Y Yang ldquoNavigating andpositioning technique for inner detection of pipelinerdquo Journalof Shenyang University of Technology vol 34 no 4 pp 427ndash4322012

[3] S Sha and Q Feng ldquoExcavated verification technology of pitdefects of buried pipelinerdquoOil and Gas Storage and Transporta-tion vol 8 pp 834ndash838 2014

[4] HGhaednia SDas RWang andRKania ldquoSafe burst strengthof a pipeline with dent-crack defect effect of crack depth andoperating pressurerdquo Engineering Failure Analysis vol 55 pp288ndash299 2015

[5] B Pinheiro I Pasqualino and S Cunha ldquoFatigue life assess-ment of damaged pipelines under cyclic internal pressurepipelines with longitudinal and transverse plain dentsrdquo Inter-national Journal of Fatigue vol 68 pp 38ndash47 2014

[6] M Allouti C Schmitt and G Pluvinage ldquoAssessment of agouge and dent defect in a pipeline by a combined criterionrdquoEngineering Failure Analysis vol 36 pp 1ndash13 2014

[7] S Paeper B Brown and T Beuker ldquoInline inspection of dentsand corrosion using lsquohigh qualityrsquo multi-purpose smart-piginspection datardquo in Proceedings of the 6th International PipelineConference (IPC rsquo06) pp 243ndash248 September 2006

[8] X Wang and H Song ldquoThe inertial technology based 3-dimensional information measurement system for under-ground pipelinerdquoMeasurement vol 45 no 3 pp 604ndash614 2012

[9] LMengjie ldquoThe study of accurate in-line inspection technologyto offshore pipeline routerdquo China Offshore Platform vol 19 no6 pp 46ndash49 2004

[10] J Yu J G Lee C G Park andH S Han ldquoAn off-line navigationof a geometry PIG using a modified nonlinear fixed-intervalsmoothing filterrdquoControl Engineering Practice vol 13 no 11 pp1403ndash1411 2005

[11] J A Czyz C Fraccaroli andA P Sergeant ldquoMeasuring pipelinemovement in geotechnically unstable areas using an inertialgeometry pipeline inspection pigrdquo in Proceedings of the ASME1st International Pipeline Conference Calgary Canada June1996

[12] J A Czyz and J Falk ldquoUse of geopig for prevention of pipelinefailures in environmentally sensitive areasrdquo in Proceedings ofthe Pipeline Pigging Integrity Assessment and Repair ConferenceHouston Tex USA February 2000

8 Journal of Sensors

[13] I B Iflefel D G Moffat and J Mistry ldquoThe interaction ofpressure and bending on a dented piperdquo International Journalof Pressure Vessels and Piping vol 82 no 10 pp 761ndash769 2005

[14] J Błachut and I B Iflefel ldquoCollapse of pipeswith plain or gougeddents by bending momentrdquo International Journal of PressureVessels and Piping vol 84 no 9 pp 560ndash571 2007

[15] J-H Baek Y-P Kim W-S Kim J-M Koo and C-S SeokldquoLoad bearing capacity of API X65 pipe with dent defect underinternal pressure and in-plane bendingrdquo Materials Science andEngineering A vol 540 pp 70ndash82 2012

[16] A Limam L-H Lee and S Kyriakides ldquoOn the collapse ofdented tubes under combined bending and internal pressurerdquoInternational Journal of Mechanical Sciences vol 55 no 1 pp1ndash12 2012

[17] D K Kim S H Cho S S Park H R Yoo and Y W RholdquoDesign and implementation of 3010158401015840 geometry PIGrdquo KSMEInternational Journal vol 17 no 5 pp 629ndash636 2003

[18] J A Czyz et al ldquoMulti-pipeline geographical informationsystem based on high accuracy inertial surveysrdquo in Proceedingsof the ASME 3rd International Pipeline Conference CalgaryCanada 2000

[19] J D Hart N Zulfiqar D H Moore and G R Swank ldquolsquoDigitalPiggingrsquo as a basis for improved pipeline structural integrityevaluationsrdquo in Proceedings of the International Pipeline Con-ference vol 2 of Integrity Management Poster Session StudentPaper Competition Calgary Canada September 2006

[20] J A Czyz and J RAdams ldquoComputations of pipelinemdashbendingstrains based on geopig measurementsrdquo in Proceedings of theipeline Pigging and Integrity Monitoring Conference pp 14ndash17Houston Tex USA February 1994

[21] J D Hart and G H Powel Geometry Monitoring of the Trans-Alaska Pipeline Trans-Alaska Pipeline System (TAPS) 2005

[22] D Yu L ZhangW Liang Y Ye and ZWang ldquoNoise reductionof signal and condition recognition of long-distance pipelinerdquoActa Petrolei Sinica vol 30 no 6 pp 937ndash941 2009

[23] J Zhou BGuiM Li andWLin ldquoAn application of the artificialneural net dominated by lithology to permeability predictionrdquoActa Petrolei Sinica vol 31 no 6 pp 985ndash988 2010

[24] D Li D Lu X Kong and Y Du ldquoProcessing of well log databased on backpropagation neural network implicit approxima-tionrdquo Acta Petrolei Sinica vol 28 no 3 pp 105ndash108 2007

[25] K-X Peng and S-Z Yu ldquoPrediction and control of strip thick-ness based on Bayesian neural networksrdquo Journal of Universityof Science and Technology Beijing vol 32 no 2 pp 256ndash2622010

[26] H Yang and H-Z Fu ldquoStock index forecast based on bayesianregularization BP neural networkrdquo Science Techno Logy andEngineering vol 6 pp 3306ndash3310 2009

[27] H Wang L Shi H Zhao and Y Yue ldquoTemperature predictionmodel for sunlight greenhouse based on bayesian regularizationBP neural networkrdquo Hubei Agricultural Sciences vol 9 pp4300ndash4303 2015

[28] Y Shenghua and D Juan ldquoGDP prediction based on principalcomponent analysis and bayesian regularization BP neuralnetworkrdquo Journal of Hunan University (Social Sciences) vol 11pp 42ndash45 2011

[29] R Li Q Feng M Cai et al ldquoMeasurement of long-distanceburied pipeline centerline based on multi-sensor data fusionrdquoActa Petrolei Sinica vol 35 no 5 pp 987ndash992 2014

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Journal of Sensors 7

Table 4 Calculation and error for modified Bayesian neural network

Actual depth (mm) Calculation ofBayesian (mm)

Error for calculation of Bayesianand actual depth (mm) Relative error ()

147153 1498349 0268193 182158535 157455 0108 068151218 1417717 094463 625151218 1496672 015508 103113007 1107863 022207 19796747 1010614 0431435 446152844 1541688 0132484 087

pipeline bending strain we propose amethod based onmod-ified Bayesian neural network to calculate dent depths frombending strain To test the proposed method experimentallya PIG with the proposed method is used to inspect a 150 kmpipeline It can be obtained that

(1) A new method is proposed to verify the relationshipbetween pipeline bending strain and dent based onthe calculation of bending strain

(2) According to the characteristic of signal betweendent and pipeline bending strain a new model ispresented in detail to be used for proposed calculationof algorithm

(3) The calculation of modified Bayesian neural networkis proposed to compute the depth dent with pipelinebending strain to compare with actual data Andtraditional BP neural network is used to calculatethe depth of dent According to the result of calcula-tion the mean accuracy of calculation for modifiedBayesian neural network is better than BP neuralnetwork for 982

(4) According to calculation the proposed method ismore accurate and suitable for the calculation ofdepth based on pipeline bending strain The meanrelative error for calculation of dent depth based themodified Bayesian neural network is 244

This paper provides a novel method for calculating dentdepths from the pipeline bending strain The dent can beevaluated by the bending strain not used to dig or use anothertool to reinspect The bending strain and calculation of dentdepth is also useful to the evaluation of pipeline integrity

Competing Interests

The authors declare that they have no competing interests

Acknowledgments

Thisworkwas supported by the project of Petrochina PipelineCompany ldquothe research of safety service of buried pipeline inpermafrost regionrdquo

References

[1] Q Aihua ldquoChinese oil and gas pipeline transport develop-ment status and analysis of associated problemsrdquo InternationalPetroleum Economics vol 17 no 12 pp 17ndash12 2009

[2] L-J Yang R-Q Li S-W Gao and Y Yang ldquoNavigating andpositioning technique for inner detection of pipelinerdquo Journalof Shenyang University of Technology vol 34 no 4 pp 427ndash4322012

[3] S Sha and Q Feng ldquoExcavated verification technology of pitdefects of buried pipelinerdquoOil and Gas Storage and Transporta-tion vol 8 pp 834ndash838 2014

[4] HGhaednia SDas RWang andRKania ldquoSafe burst strengthof a pipeline with dent-crack defect effect of crack depth andoperating pressurerdquo Engineering Failure Analysis vol 55 pp288ndash299 2015

[5] B Pinheiro I Pasqualino and S Cunha ldquoFatigue life assess-ment of damaged pipelines under cyclic internal pressurepipelines with longitudinal and transverse plain dentsrdquo Inter-national Journal of Fatigue vol 68 pp 38ndash47 2014

[6] M Allouti C Schmitt and G Pluvinage ldquoAssessment of agouge and dent defect in a pipeline by a combined criterionrdquoEngineering Failure Analysis vol 36 pp 1ndash13 2014

[7] S Paeper B Brown and T Beuker ldquoInline inspection of dentsand corrosion using lsquohigh qualityrsquo multi-purpose smart-piginspection datardquo in Proceedings of the 6th International PipelineConference (IPC rsquo06) pp 243ndash248 September 2006

[8] X Wang and H Song ldquoThe inertial technology based 3-dimensional information measurement system for under-ground pipelinerdquoMeasurement vol 45 no 3 pp 604ndash614 2012

[9] LMengjie ldquoThe study of accurate in-line inspection technologyto offshore pipeline routerdquo China Offshore Platform vol 19 no6 pp 46ndash49 2004

[10] J Yu J G Lee C G Park andH S Han ldquoAn off-line navigationof a geometry PIG using a modified nonlinear fixed-intervalsmoothing filterrdquoControl Engineering Practice vol 13 no 11 pp1403ndash1411 2005

[11] J A Czyz C Fraccaroli andA P Sergeant ldquoMeasuring pipelinemovement in geotechnically unstable areas using an inertialgeometry pipeline inspection pigrdquo in Proceedings of the ASME1st International Pipeline Conference Calgary Canada June1996

[12] J A Czyz and J Falk ldquoUse of geopig for prevention of pipelinefailures in environmentally sensitive areasrdquo in Proceedings ofthe Pipeline Pigging Integrity Assessment and Repair ConferenceHouston Tex USA February 2000

8 Journal of Sensors

[13] I B Iflefel D G Moffat and J Mistry ldquoThe interaction ofpressure and bending on a dented piperdquo International Journalof Pressure Vessels and Piping vol 82 no 10 pp 761ndash769 2005

[14] J Błachut and I B Iflefel ldquoCollapse of pipeswith plain or gougeddents by bending momentrdquo International Journal of PressureVessels and Piping vol 84 no 9 pp 560ndash571 2007

[15] J-H Baek Y-P Kim W-S Kim J-M Koo and C-S SeokldquoLoad bearing capacity of API X65 pipe with dent defect underinternal pressure and in-plane bendingrdquo Materials Science andEngineering A vol 540 pp 70ndash82 2012

[16] A Limam L-H Lee and S Kyriakides ldquoOn the collapse ofdented tubes under combined bending and internal pressurerdquoInternational Journal of Mechanical Sciences vol 55 no 1 pp1ndash12 2012

[17] D K Kim S H Cho S S Park H R Yoo and Y W RholdquoDesign and implementation of 3010158401015840 geometry PIGrdquo KSMEInternational Journal vol 17 no 5 pp 629ndash636 2003

[18] J A Czyz et al ldquoMulti-pipeline geographical informationsystem based on high accuracy inertial surveysrdquo in Proceedingsof the ASME 3rd International Pipeline Conference CalgaryCanada 2000

[19] J D Hart N Zulfiqar D H Moore and G R Swank ldquolsquoDigitalPiggingrsquo as a basis for improved pipeline structural integrityevaluationsrdquo in Proceedings of the International Pipeline Con-ference vol 2 of Integrity Management Poster Session StudentPaper Competition Calgary Canada September 2006

[20] J A Czyz and J RAdams ldquoComputations of pipelinemdashbendingstrains based on geopig measurementsrdquo in Proceedings of theipeline Pigging and Integrity Monitoring Conference pp 14ndash17Houston Tex USA February 1994

[21] J D Hart and G H Powel Geometry Monitoring of the Trans-Alaska Pipeline Trans-Alaska Pipeline System (TAPS) 2005

[22] D Yu L ZhangW Liang Y Ye and ZWang ldquoNoise reductionof signal and condition recognition of long-distance pipelinerdquoActa Petrolei Sinica vol 30 no 6 pp 937ndash941 2009

[23] J Zhou BGuiM Li andWLin ldquoAn application of the artificialneural net dominated by lithology to permeability predictionrdquoActa Petrolei Sinica vol 31 no 6 pp 985ndash988 2010

[24] D Li D Lu X Kong and Y Du ldquoProcessing of well log databased on backpropagation neural network implicit approxima-tionrdquo Acta Petrolei Sinica vol 28 no 3 pp 105ndash108 2007

[25] K-X Peng and S-Z Yu ldquoPrediction and control of strip thick-ness based on Bayesian neural networksrdquo Journal of Universityof Science and Technology Beijing vol 32 no 2 pp 256ndash2622010

[26] H Yang and H-Z Fu ldquoStock index forecast based on bayesianregularization BP neural networkrdquo Science Techno Logy andEngineering vol 6 pp 3306ndash3310 2009

[27] H Wang L Shi H Zhao and Y Yue ldquoTemperature predictionmodel for sunlight greenhouse based on bayesian regularizationBP neural networkrdquo Hubei Agricultural Sciences vol 9 pp4300ndash4303 2015

[28] Y Shenghua and D Juan ldquoGDP prediction based on principalcomponent analysis and bayesian regularization BP neuralnetworkrdquo Journal of Hunan University (Social Sciences) vol 11pp 42ndash45 2011

[29] R Li Q Feng M Cai et al ldquoMeasurement of long-distanceburied pipeline centerline based on multi-sensor data fusionrdquoActa Petrolei Sinica vol 35 no 5 pp 987ndash992 2014

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

8 Journal of Sensors

[13] I B Iflefel D G Moffat and J Mistry ldquoThe interaction ofpressure and bending on a dented piperdquo International Journalof Pressure Vessels and Piping vol 82 no 10 pp 761ndash769 2005

[14] J Błachut and I B Iflefel ldquoCollapse of pipeswith plain or gougeddents by bending momentrdquo International Journal of PressureVessels and Piping vol 84 no 9 pp 560ndash571 2007

[15] J-H Baek Y-P Kim W-S Kim J-M Koo and C-S SeokldquoLoad bearing capacity of API X65 pipe with dent defect underinternal pressure and in-plane bendingrdquo Materials Science andEngineering A vol 540 pp 70ndash82 2012

[16] A Limam L-H Lee and S Kyriakides ldquoOn the collapse ofdented tubes under combined bending and internal pressurerdquoInternational Journal of Mechanical Sciences vol 55 no 1 pp1ndash12 2012

[17] D K Kim S H Cho S S Park H R Yoo and Y W RholdquoDesign and implementation of 3010158401015840 geometry PIGrdquo KSMEInternational Journal vol 17 no 5 pp 629ndash636 2003

[18] J A Czyz et al ldquoMulti-pipeline geographical informationsystem based on high accuracy inertial surveysrdquo in Proceedingsof the ASME 3rd International Pipeline Conference CalgaryCanada 2000

[19] J D Hart N Zulfiqar D H Moore and G R Swank ldquolsquoDigitalPiggingrsquo as a basis for improved pipeline structural integrityevaluationsrdquo in Proceedings of the International Pipeline Con-ference vol 2 of Integrity Management Poster Session StudentPaper Competition Calgary Canada September 2006

[20] J A Czyz and J RAdams ldquoComputations of pipelinemdashbendingstrains based on geopig measurementsrdquo in Proceedings of theipeline Pigging and Integrity Monitoring Conference pp 14ndash17Houston Tex USA February 1994

[21] J D Hart and G H Powel Geometry Monitoring of the Trans-Alaska Pipeline Trans-Alaska Pipeline System (TAPS) 2005

[22] D Yu L ZhangW Liang Y Ye and ZWang ldquoNoise reductionof signal and condition recognition of long-distance pipelinerdquoActa Petrolei Sinica vol 30 no 6 pp 937ndash941 2009

[23] J Zhou BGuiM Li andWLin ldquoAn application of the artificialneural net dominated by lithology to permeability predictionrdquoActa Petrolei Sinica vol 31 no 6 pp 985ndash988 2010

[24] D Li D Lu X Kong and Y Du ldquoProcessing of well log databased on backpropagation neural network implicit approxima-tionrdquo Acta Petrolei Sinica vol 28 no 3 pp 105ndash108 2007

[25] K-X Peng and S-Z Yu ldquoPrediction and control of strip thick-ness based on Bayesian neural networksrdquo Journal of Universityof Science and Technology Beijing vol 32 no 2 pp 256ndash2622010

[26] H Yang and H-Z Fu ldquoStock index forecast based on bayesianregularization BP neural networkrdquo Science Techno Logy andEngineering vol 6 pp 3306ndash3310 2009

[27] H Wang L Shi H Zhao and Y Yue ldquoTemperature predictionmodel for sunlight greenhouse based on bayesian regularizationBP neural networkrdquo Hubei Agricultural Sciences vol 9 pp4300ndash4303 2015

[28] Y Shenghua and D Juan ldquoGDP prediction based on principalcomponent analysis and bayesian regularization BP neuralnetworkrdquo Journal of Hunan University (Social Sciences) vol 11pp 42ndash45 2011

[29] R Li Q Feng M Cai et al ldquoMeasurement of long-distanceburied pipeline centerline based on multi-sensor data fusionrdquoActa Petrolei Sinica vol 35 no 5 pp 987ndash992 2014

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of