Research on the quantitative analysis of subsurface defects for non
Transcription
Research on the quantitative analysis of subsurface defects for non
NDT&E International 45 (2012) 104–110 Contents lists available at SciVerse ScienceDirect NDT&E International journal homepage: www.elsevier.com/locate/ndteint Research on the quantitative analysis of subsurface defects for non-destructive testing by lock-in thermography Liu Junyan n, Tang Qingju, Liu Xun, Wang Yang School of Mechatronics Engineering, Harbin Institute of Technology, Harbin 150001, PR China a r t i c l e i n f o a b s t r a c t Article history: Received 12 April 2011 Received in revised form 1 September 2011 Accepted 2 September 2011 Available online 17 September 2011 This pa p paper ape per p er describes descri des cri ri b ri be bes e ess the th he quantitative he qua qua anti nti nt tittat at atiiv ive ve ve analysis ana ana aly llys ysis of y of the the boundary, th boundary, location and b a depth of subsurface defects by y llock-in ock oc ock c -in n thermography. th tthe h he errmo rm mo m ogra aph phy h . The Th T e phase ph has ase difference ase diif dif d iffe iffer fe fer errenc en enc e ncce nc e between betw be ween defective areas a and non-defective areas iillustrates ill il ll llust u rat ate at ess the th he e qualitative qual ua alit al iita tta ativ tive analysis ti an a alysis sis i of of the th he e boundary boun unda un dar dar ary and and the th the location llocation of the th subsurface defect. In order to accurately ac acc ura atel tte elly estimate e essti ttim imate t the size, location and and depth depth th of th of the the e defects, de efe f cts, the phase p is normalised, the heat tra ra an nsf ns sffe err partial par art a r ial rt ia differential equation equ qu uati atti a ation tion o (PDE) (PDE) model mod del de e iss used ussed se ed e d to filter filt lt r the noise lte no transfer of normalised phase image and nd d the e differential differenttial all normalised a no no nor orrmal m ise ed phase ph phase assse e profile prr file iss employed pro p empl m loye oye oy yed to determine yed det d de etermine the boundary and location of the defect. de ect. The profile de def prrofil ofi ofi fille of of the tth he differential h dif d ifffe fer e ent er en ia al normalised al normalise sed d phase phase ph a distribution dis istri ttri tr ribut ri b ion has maximum, minimum and zer erro points points that help elp lp p to to quantitatively qu ua uan a tit titati atti a ati tivel vel ve ely determine d determine the e boundary bound nd dary and an location an loc zero of the subsurface defect. An n artificial artificial neural network net etttw work (ANN) N)) is is proposed pr o pr op opo p sed to determine dete e rmi miin m ne e the the he depth de of the subsurface defect. Exp Ex E x erimental results for fo or a steel plate, plate te e, a carbon carb carb rb bon n fibre-reinforced fibre-rein in nfforcce n ed d polymer poly oly lymer (CFRP) sheet–foam sandwich, Experimental a an nd honeycomb structure structtu ure ur rre e composites ccom om mpos p ite po es with wiitth w h artificial artificial subsurface subssurf u ace ce defects ce de de def effe ect ec c s show s and good agreement with the a act ual values. actual 201 20 011 Elsevier 0 E & 2011 Ltd. All rights reserved. Keywords: Lock-in thermography PDE ANN Differential normalised phase Quantification 1. Introduction te n ht l: tp +8 :/ 6 /w 41 w 18 w 3 .il 72 gw 6 el 985 ls .c om Infrared thermography (IRT) T) has been be b een n successfully ssu ucc c es essf sful u ly used d ass a non-destructive testing and evaluation evaluatio io on (NDT&E) (ND NDT T& &E)) technique tech te c nique in many applications. Infrared thermography thermograph phy provides ph pro rovi ovi vide ide d s the th he e colour c lour co mens in which h local llo oca al changes chang hang ha nges in in the the th temperature image of specimens defect cttss.. Since S nc Si nce e the the he beginbegi be begi ginn surface temperature indicate subsurface defects. een successfully used use ed as as a NDT&E NDT ND T& & &E E ning of the 1960s, IRT has been technique for materials [1–3]. cilitates better subsurface defect Lock-in thermography facilitates detection than ordinary infrared thermography because the thermal wave is very sensitive to interfaces between materials and it is less sensitive to non-uniform emission and surrounding conditions. With the development of advanced inspection technology, the quantitative non-destructive evaluation (QNDE) of defects plays an important role in NDT. In QNDE, defects are generally evaluated by the characteristics of position, size, depth, shape and type. Based on these parameters, the reliability and service life of the structure can be estimated [4]. To accomplish this, quantitative determination has been thoroughly studied both in theory and in practical applications [5]. Although lock-in thermography provides powerful amplitude and phase images, QNDE gives a clear bridge relation between the true size of de effe ectss and a am mp pliitu tude de and an a nd phase p defects their amplitude images. Saintey and Alm mond have h ve proposed ha prop op pos osed e a technique techn hnique for defect size determinaAlmond n in in which wh hic i h the he e defect’s de d efe f ct ct’ss size size e can be estimated as the full width tion ha alff maximum ma maxi ax xiimu mum (FWHM) (FW WH HM M) contrast contrast [6]. However, this method at half tend ds to to estimate est stim tiim matte the the he defect’s defe de fe ect’s size as larger when the contrast is tends sm ma alll. l. In In some som me e cases, case case ca ses, the th he contrast contrast profile profi is non-symmetric, and it small. diffi di fficcu ffi ult lt to to evaluate ev e val a ua ate te the quantitative quantitativ defect size. iss difficult In this tth hiiss case, ca ase e, we we suggest a simple, simple practical method for the In quan nti tita tta ati tive determination d termination of the boundary, de boun quantitative location and depth of g subsurface defects. To achieve this goal of quantitatively determining the defect size, the original phase data from the lock-in thermography are processed by means of a normalised method, and the noise of normalised phase image is filtered based on the heat transfer partial differential equation model. The normalised phase distribution profile is directly used to calculate its differential value by the numerical difference method, or it is fitted by a polynomial method and then used to calculate the differential value. The differential profile of the normalised phase has maximum, minimum and zero points; these points correspond to the boundary and location of subsurface defects. The artificial neural network is proposed as a tool to determine the depth of the defects based on the normalised phase for testing and training. In our experiments, the reference specimens are made from a steel plate, a carbon fibre-reinforced polymer (CFRP) sheet–foam sandwich and honeycomb structure composites with known sizes, locations and depths of subsurface defects. The results obtained by the proposed technique are compared with actual values. Corresponding author. Tel.: þ86 451 8640 3380; fax: þ86 451 8640 2755. E-mail address: [email protected] (L. Junyan). 0963-8695/$ - see front matter & 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.ndteint.2011.09.002 L. Junyan et al. / NDT&E International 45 (2012) 104–110 2. Theory 2.1. Lock-in thermography Lock-in thermography was first proposed by Carlomagno and Berardi. The lock-in principle is the technique of choice if a signal has to be extracted from statistical noise [7]. Basically, the heat source induced from outside has to be modulated and calibrated to make the surface temperature variation from the heat source truly sinusoidal. The infrared camera captures a series of thermal images and extracts a sinusoidal thermal wave pattern at each pixel of the thermal image at a frequency modulated by different thermal wave signal analysis methods [8]. The surface temperature T(x,y) (z ¼0) in one dimension is given by Tðx,yÞ ¼ Aðx,yÞej½otjðx,yÞ ð1Þ where A(x,y) is the surface temperature amplitude and j(x,y) is the surface temperature phase at each image pixel. The thermal wave can be e reconstructed using four equidistant x,y)– ) surface temperature (thermal mal wave) average data points poin nttss S1((x,y)– perio od d.. More More Mo re thermal th herma errm l S4(x,y), with a phase step off a quarter of a period. aged to obtain four ur data ur data points, da po p oints iin nttss, which wh hiicch h wave data points were averaged e noise effectt (Fig. (F Fiig. 1). ) ). were helpful in reducing the ase shift and and nd the th he amplitude am mp plliitu itu ud de e of of the th he Calculations of the phase thermal wave are given as follows: S ðx,yÞS1 ðx,yÞ x,yÞ ð2Þ ð2Þ jðx,yÞ ¼ arctan 3 x,yÞ S4 ðx,yÞS2 ðx,yÞ Aðx,yÞ ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ffiffiffiffiffiffiffiffiffiffi ffiffiffiffi ffiffiffiffiffiffi ffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ½S3 ðx,yÞS1 ðx,yÞ2 þ ½S4 ððx,yÞS x,y yÞS ÞS2 ðx Þ ðx,yÞ x, yÞ2 phase image. The heat transfer partial differential equation is described as follows: ( @uðx,y,tÞ ¼ a Duðx,y,tÞ @t ð5Þ uðx,y,0Þ ¼ u0 ðx,yÞ where u(x,y,t) is the two-dimensional temperature field, D is the Laplacian operator, a is the thermal diffusivity and u0(x,y) is the initial temperature. The temperature field can be obtained by the difference form of Eq. (5), which is given by uðx,y,tÞu0 ðx,yÞ ¼ aDuðx,y,0Þ t ð6Þ uðx,y,tÞ ¼ u0 ðx,yÞ þ t aDuðx,y,0Þ ð7Þ The normalised phase can be filtered by Eq. (7), and the filtered phase is given by jFNormal ðx,yÞ ¼ jNormal ðx,yÞ þ bDjNormal ðx,yÞ ð8Þ (x,y) , ) is the normalis normalised phase filtered at the pixel where w wh he errre jFNormal(x,y i the filtered coefficien coefficient b ¼ t a. (x,y) and (x x,,y y) a nd b is nd If b 4 If 40, 0, Eq. Eq. ((8) Eq 8 is the so-called regular 8) r heat transfer PDE filter and at low frequency band, a an nd tthe he h e normalised no n orm rmal alis ise is ed d phase p ase image is filtered ph fi o0, 0,, E 0 Eq Eq. q q.. (8) (8) is is the the so-called iinverse heat transfer PDE filter th a an nd iiff b o and and tthe he normalised n rm no ma alliisse ed d phase phase ph s image is filtered at high frequency ba b and n . Eq. (8) (8 8) is is used use ed to to filter filt lter the noise n band. of the normalised phase da d atta a from fro fr lock ck k-i - n thermography the hermog ography and to improve the contrast og data lock-in be etw twee en the defective defe fe ect c ive area and and n the healthy area on the normalbetween iis ssed ed e d ph phase image. ised ð3Þ e helpful hel elpfful el elpf ul for for estimating fo estimating the defect defe ectt size ssiiz ze e The phase image is more om mm mo on thermal the herm rmal image because becau use e the in comparison with the common han nges att defect defect spots, and itt is is less phase has more contrast changes rm surface sur urrffa a ace ce emission, ce emission, ambient em ambie ie en ntt distursensitive to the non-uniform bance and so on. 2.2. Phase noise filtering e variable varria ia abl ble bl le range, ran ange g , the he phase ph p hase ase is as is In order to reduce the phase normalised as follows: jmax jðx,yÞ jmax jmin ðð4Þ 4Þ 2.3. 3.. Di D Diff iff ffeerreen ntial calculation calcul ullatio io on of of the the he normalised no Differential phase The normalised phase pha ha ase se distribution dist di stri ribu ri ib bu utio along the x direction of the sp pe eccimen surface is analysed ana naly na l sed by means me specimen of the differential calculati ion on method. The differential differ erential er tia ti all normalised no orma tion phase profile in Eq. (9) is ca alc l ulated by subtraction subtra ra acttio on o the adjacent a calculated off the normalised phase. In a ad dd diiti tion, the normal allis a ised d phase ph p has ase distribution dist addition, normalised along the x direction iiss fit fi tte ted b y a poly ly ynomi mia all in i E q. (1 fitted by polynomial Eq. (10), and then the differential no n orma alised phase al phas asse is is calculated cal cal alccu ulate te ed by Eq. E (11): normalised jFFNor x,y yÞx ¼ jFNor FNormal Nor No orrmall ððx,yÞ FNormal No Nor N orrm o mal ma a ðx þ Dx,yÞjFNormal ðx,yÞ xÞ ¼ f jFN FNormal F FNor N mal al ððxÞ n X malised phase, jmaxx is i the th he maximum max axim imu um m where jNormal is the normalised e of of the the e phase. ph hasse. e value of the phase and jminn is the minimum value In order to enhance the contrast and reduce the background back kgro kg roun und sed phase data are processed based on noise influence, the normalised the heat transfer partial differential equation, which can be called the filter of the heat transfer partial differential equation (PDE) and it is always used to process image in signal processing fields. In this paper, the heat transfer PDE is utilised to filter normalised ð10Þ iUai xi1 ¼ a1 þ2a2 x þ3a3 x2 þ þ nan xn1 ð11Þ i¼0 f j0FFN FNormal FNor FNo N mal ðxÞ ¼ n X i¼1 where jFnormal (x,y)x is the differential normalising phase, Dx is the step in the x direction, fjFnormal (x) is the fitting of the normalised phase along the x direction, fj0 Fnormal (x) is the differential normalised phase of polynomial fitting and ai is the polynomial coefficient. The least square method (LEM) is applied to fit the polynomial normalised phase along the x direction of the spatial profile. The highest polynomial exponent should be selected to reduce the fitting deviation and avoid more oscillations. The allowance deviation of polynomial fitting is given as te e0 ¼ c 9maxðjNormal ðx,yÞ9y ÞminðjNormal ðx,yÞ9y Þ9 Fig. 1. Correlation principle of lock-in thermography. ð9Þ ai xi ¼ a0 þ a1 x þ a2 x2 þ þ an xn ht l: tp +8 :/ 6 /w 41 w 18 w 3 .il 72 gw 6 el 985 ls .c om jNormal ðx,yÞ ¼ 105 ð12Þ where e0 is the allowance deviation of polynomial fitting, c is the scale factor, usually, it should be selected that c o0.1, and jNormal(x,y)9y is the normalised phase distribution along the x direction of the spatial profile. The deviation of the polynomial fitting is defined as the maximum absolute difference value between the polynomial fitting values and actual values of the normalised phase along 106 L. Junyan et al. / NDT&E International 45 (2012) 104–110 the x direction of the spatial profile: 2.4. Artificial neural network classifier for depth of defect Dj ¼ maxð9f jNormal ðxÞjNormal ðx,yÞ9y 9Þ ð13Þ where Dj is the deviation of the polynomial fitting. This polynomial fitting process is repeated so that the criteria condition is satisfied that Dj o e0 (in Fig. 2). The polynomial fitting process is finished using the Matlab programme. In order to enhance the polynomial fitting accuracy and reduce oscillations at higher order, the normalised phase profile can be divided into many small segments, and then the above polynomial fitting is allowed to be applied on each segment of the normalised phase profile. This method is available for showing good fitting results. The differential normalised phase profile distribution provides maximum, minimum and zero values, the distance between characteristic points (maximum value and minimum value) is used for defect size determination and the zero value point is used for local defect centre determination. An artificial neural network is a massively parallel distributed processor that has a natural propensity for storing experiential knowledge and making it available for use. Artificial neural networks as classifiers for defect detection in non-destructive testing when using infrared thermography have been proposed in the past few years, and multilayer perceptrons were employed to detect and characterise defects using the normalised phase data from lock-in thermography. In this work, a multilayer perceptron classifier is characterised by the presence of an input layer of source nodes, a hidden layer and an output layer (in Fig. 3). The multilayer perceptron network is called a feedback network that can be trained using the back propagation algorithm, which has been widely the most used learning algorithm in recent years. The input vector is Pi that includes the modulated frequency fLockin and the normalised phase at different pixel (x,y) is given as follows: Pi ¼ ffff Lock Lockin ck kin in , jN Normal Norm No Nor orrmal ðx,yÞg or This Th T h his is feedback is fe eed edba dba b ck k network netw ne twork architecture can be used to identify the de defe fect ctt and an nd inspect in nssp pe ectt the the defect th defect depth, dept defect so the output layer of feed dbacck k network n tw ne twor ork has has a two ttw wo defect defect detectors’ dete feedback neuron cells and the ou utp t ut layerr value val a ue e are arre given a giv ve en n by output ϕ ϕ = ϕ −ϕ ð15Þ liinearððW O2 lo llogsigðW gssiig gðW W H P i þ bH ÞÞ þ bO2 Þ De epde Dep def eff ¼ linearððW e ð16Þ te ht l: tp +8 :/ 6 /w 41 w 18 w 3 .il 72 gw 6 el 985 ls .c om Δϕ > ε Fig. 2. Diagram off the polynomial fitting. logsigðW ogs gssig igððW H Pi þ bH ÞÞ þ bO logsigððW A ¼ llo og gssig igðð ð W O1 lo O1 Þ one detector network outputs for where e A is is on o ne of tthe he defect de d efe fectt d ete iidentifying id denti tifyi ti fyin fy ing defect; ing defect; its desired de esirred ed output outp ou outp tput p value is 0 for healthy pixel a an nd 1 for defective pixe xel. De D epde and pixel. Dep def ef is one of the defect detector netw work outputs for measuring mea ea asu suring riing defect def de efect depth. network Th he hidden hidden layer enab ble es tthe he e neural ne eura network to extract higherThe enables or rde er sstatistics, tatistics, especial allly ly when when wh en the th he size of input layer is large order especially en nou ugh h. T he supervised d lle earni rniing g me etho is concerned with applying enough. The learning method ma any y training trra ning examples trai exampl plless to modify m diiffy mo y the connecting weights of the many ne eu urron n cells. ce ells. Each example ex xa am mpl ple e includes inc nclu l d de es a se neuron series of input vectors and the corr respo p nd ndin i g des essiirrred ed e d re esp spon onse se e. The artificial neural network is corresponding desired response. pres sen ente ted with with h plenty plle ent nty of of examples, exam mples, and the connecting weights are mp presented adju usstte ed d to m mi ini ini nimi m sse e the the difference th difffe ference between bet adjusted minimise the desired output and th he a ctua al rresponse espons es po on nsse of of the the e network. network. The network training is repeated the actual u un ntil ttiil a steady sttea eady dy sst ta atte iiss rreached. eached. At that time, ea t until state the connecting weights a ar re no no longer lon ong ge er v arrie ied. In this work, this network was trained using are varied. 731 iinput–output 73 731 n ut np u –o ou uttput pair points extract from fro the metal samples, which ϕ Δϕ = ð14Þ Fig. 3. An artificial neural network with one hidden layer and an output. L. Junyan et al. / NDT&E International 45 (2012) 104–110 contained air (bottom-flat hole) delamination defects in different depths and configurations, 603 input vectors were the modulated frequency and normalised phase pair points over healthy area and 32 were over at each defect depth area (4 different defect depth areas). The network was trained using 1240 input–output pair points extract from the CFRP sheet sandwich samples, which contained Teflon insert defects in different configurations, 1000 input vectors were the modulated frequency and normalised phase pair point over healthy area and 120 were over at defective area (2 defective area). After different simulation and experimental tests, the ANN is settled for a 30 2 defect detector network for the metal sample and settled for a 60 2 defect detector network for the CFRP sheet sandwich sample, so the reported experiments were executed with 60 nodes for the metal material and 120 nodes for the CFRP sheet sandwich material. The artificial neural network is applied to recognise the depth of the subsurface defects. 107 12.8 10 6 m2/s, and the thermal diffusivity of the CFRP material is 0.42 10 6 m2/s. The experimental system consists of the infrared camera, the power amplifier controller, the function generator, the system controller and the heat source (2 halogen lamps, the power of each lamp being 1 KW), as shown in Fig. 5. The heat source was driven by a power amplifier controller, and the heat source and the infrared camera were synchronised using the sinusoidal function of the function generator and the lock-in modulus. The phase angle data were calculated by the lock-in modulus at the optimum modulated frequency that produces the maximum phase difference between the defective area and the healthy area. 4. Results and discussion 4.1. Specimen 1: flat bottom holes of ANSI 1045 steel mental procedure 3. Specimens and experimental ht l: tp +8 :/ 6 /w 41 w 18 w 3 .il 72 gw 6 el 985 ls .c om ns, with known defects deffec ects ts of of different diff di ffffer ere en nt The reference specimens, re made as shown show wn in in Fig. Fig. Fi g 4. 4. Specimen Sp pe eci cm me en 1 en depths, shapes and sizes, were 5 steel with four fo our ur rows wss of w of flat-bottom fla fl atatt--bo bottom was made from ANSI 1045 pecimen 2 was was made wa ma ad de from frro om CFRP CF FR RP sheet she h et holes at different depths. Specimen wit ith two two w rectangular rre ectta an ngular Teflon face and foam sandwich material with hs with h the tth he same same ame sizes. am siz izes. Specimen Spe ecciime im me en 3 inserts at the same depths eet face ce and aluminium ce alu l m miinium honeycomb ho on ne eyc yco om mb was made from CFRP sheet rts at at the th same s me sa e depths depths with the de th he same sa am me me with four round Teflon inserts ty off the the steel steel material st material (ANSI 1045) 10 045 45) is sizes. The thermal diffusivity Fig. 6 shows the defects of specimen sp 1 detected using the no n orm rmal rmal alis ised e phase image at the optimum opti normalised modulated frequency of 0. .12 2 Hz. Hz 0.12 All that form the subsurface A ll of of tthe he defects defe de fect c s at at different depths de are e cclearly le ear arlly y observed obsse obs ob e erv ved ed in in the normalised normalis phase image with different T normalised normali phase image is obviously fil lte tere red coefficients re coe effi ffici cie en nts t b. The filtered ¼0.4 (Fig. 6(c)), and the contrast ssmoothed sm oothed d at at filtered fillte terre ed coefficient coe effi fficient b ¼ main inly improved im mprrov o ed d at at higher h gh hi her values of the filtered coefficient. The iss ma mainly detta de ails ls of of the normalised no orm rmal alised phase phas ph a e image im details are clearly recognised for 0.4 0.4 (Fig. (F 6(d)). Fig. 6(e) shows the tth he filtered fillttered coefficient coeffficcie i nt b ¼ the no orrm m mal a issed phase difference al dif iffere re enc nce induced in nd duce by 0.12 Hz at the central line normalised te Fig. 4. Specimen structure and geometry. Fig. 5. Experimental system. 108 L. Junyan et al. / NDT&E International 45 (2012) 104–110 Fig. 6. Normalised phase filtered images, ages, normalised phase pha ha h ase e distribution dist istrib rib rib ri ibu uti t io ti on n and and nd differential diiff ifffffere renti re ntti tiall normalised tial no orm or rrm mal ali a lliised ised sed se ed phase pha ph p h se e filtered filt ltte lte ered r images of specimen speci 1: (a) original phase image; tered coefficient b ¼ 0.4 0 0.4; .4; (d (d) d) filtered filltte fi ter ered e d coefficient coefficient b ¼ 0.4; 0.4 .4 4; (e) (e (e) n no normalised orm rrma ma m alli alis lis iissed s d phase pha has ase distribution in Line I and (f) b ¼ =0 Differential (b) filtered coefficient b ¼0.3; (c) filtered normalised phase filtered image. ht l: tp +8 :/ 6 /w 41 w 18 w 3 .il 72 gw 6 el 985 ls .c om Fig. 7. Defect location and boundary y identifications identificcati a ons nss of of specimen spec eccim ime me m en 1: (a) defect defect def ect location and nd d boundary boun ndar da y identification id de en ent nttiific fica ati tion ti on in in Line Liine II II and (b) defect location and boundary identification in Line III. Table T Tab Ta a ab blle e1 Comparison Co C Com om mpar ariso ison iso n of measurement mea me m ea e asurement results and actual actua values. Rows Measurement Results (mm) Ac Actual size (mm) Errors (%) MCL MD ACL AD CLE DE 1 23.53 50.84 77.95 8.82 10.08 9.24 24.30 50.50 76.80 9.20 10.30 9.85 3.17 0.67 1.50 4.13 0.48 6.20 2 23.53 50.42 77.74 8.40 9.24 7.56 24.50 50.00 75.80 8.20 8.60 8.10 3.96 0.84 2.56 2.44 7.44 6.67 3 25.11 50.54 75.32 7.15 6.85 6.72 24.60 50.20 75.60 6.80 6.50 6.50 2.07 0.68 0.37 5.15 5.38 3.38 MCL—measurement centre location, MD—measurement diameter, ACL—actual centre location, AD—actual diameter, CLE—centre location error and DE—diameter error. te of the third column (Line I) for defects that are are re 1.0 1 0 mm 1. mm deep. de ee ep p.. Although the defect locationss can be estimated from m the th maxmax ximum normalised phase difference ence value, it is difficult to evaluate evalu luat lu ate ze. The subsurface defects can be the defect’s boundary and size. identified in the differential normalised phase image shown in Fig. 6(f). Fig. 7 shows the differential normalised phase profile for the central line of the first row (Line II) and the third column (Line III) defects. The defect locations and sizes are quantitatively determined from the differential normalised phase image in Fig. 6(f) and the differential normalised phase diagram in Fig. 7. The distance between the maximum and minimum positions of the differential normalised phase is calculated as the size of the subsurface defect, and the zero value position of the differential normalised phase defines its location. The sizes and central location of the three rows of defects are listed in Table 1. The measurement results are compared with the actual values, and the measurement errors are calculated. The maximum errors in the defect locations are less than 4%, and their size errors are less than 7.5%. These values also show good agreement with the actual locations and sizes. The depths of the defects were estimated by the artificial neural network. Each pixel of the normalised phase image and the frequency is an input variable, and the depth of the defect is an output parameter. The defect depths that were estimated by the ANN are listed in Table 2. It is noted that the maximum error of the ANN calculations is less L. Junyan et al. / NDT&E International 45 (2012) 104–110 than 5%. These values also show good agreement with the actual depth. 4.2. Specimen 2: rectangular Teflon insert of CFRP sheet face foam sandwich Fig. 8 shows the defects of specimen 2 detected using the normalised phase image at the optimum modulated frequency of 0.042 Hz. All of the rectangular defects that form the subsurface at a depth of 2.0 mm are clearly distinguished from the normalised phase image with different filtered coefficients b. The background noise of the normalised phase image is critically reduced at higher filtered coefficients b Z0.3, as seen in Fig. 8(c), and the contrast is mainly enhanced at higher filtered coefficients. The subsurface defects can be identified in the differential normalised phase image shown in Fig. 8(e). In Fig. 8(f), the differential normalised Table 2 Results of the ANN and errors. Actual depth h (mm) Depth of ANN AN NN (mm) ((m mm) Error Erro or (%) ((% %) 8 8 8 10 10 1.0 1.4 1.8 1.0 1.4 1.0035 1.003 35 1.4617 1.461 4617 461 4 1 1. .77 772 7 7 6 1.7726 1 . 29 .01 9 1.0129 1.4651 0 0. 35 4.40 4. 4 .40 1.52 1 .29 1.29 4 .65 65 65 4.65 phase profile presents the location and boundary of the defect. However, in the first-order differential normalised phase profile, two maximum normalised phase points exist in the position range from 140 to 180 mm, and the defect boundary is difficult to distinguish from the first-order differential normalised phase profile. The second-order differential normalised phase profile is used to detect the defect boundary, and the zero value point corresponds to the boundary of the defect in the second-order differential normalised phase profile. The defect classifier and the depth of the defect obtained from the ANN are presented in Fig. 9. Two defects can be identified by the defect classifier of the ANN, as shown in Fig. 9(a). In addition, there are other defects visible in Fig. 9(a) that may be the CFRP defect or manufacturing defects. The depth estimation of the ANN is less than that of the actual depth. The estimation error reaches approximately 25% for the second defect, and the size of the defect is smaller in comparison with the differential normalised phase profile in Fig. 8(f).The estimation error is less than 5% for the first defect, and the size of the defect is similar to the diffffer di eren eren enttiial process in Fig. 8(f). differential 4 3. 4. 3. Specimen Sp pec ecim cim men en 3: 3 round ro oun und Teflon insert of CFRP sheet face honeycomb 4.3. 0 shows ssh how ws the the e defects de d efe fects of specimen sp Fig. 10 3 detected using the no orm rmal a ised phase ph has ase e image im ma age e at at a modulated mo normalised frequency of 0.1 Hz. Th Th he e entire ent ntire round nd d defects defe de fects forming for orming the subsurface at a depth of The ht l: tp +8 :/ 6 /w 41 w 18 w 3 .il 72 gw 6 el 985 ls .c om Diameter (mm) 109 te Fig. 8. Normalised phase filtered images and differential normalised phase filtered images of specimen 2: (a) filtered coefficient b ¼ 0.1; (b) filtered coefficient b ¼0.2; (c) filtered coefficient b ¼ 0.3; (d) filtered coefficient b ¼ 0.4; (e) differential normalised phase filtered image; and (f) defect location and boundary identification in Line IV. Fig. 9. Defect classifier of specimen 2 by the ANN: (a) defect classifier and (b) depth distribution profile in Line V by the ANN. 110 L. Junyan et al. / NDT&E International 45 (2012) 104–110 Fig. 10. Phase and differential normalised phase filtered images of specimen 3: (a) normalised phase image; (b) differential normalised phase filtered image; and (c) defect location and boundary identification in Line VI. 1.0 mm are detectable from the normalised phase image shown in Fig. 10(a) and the differential normalised phase image shown in dif ifffe ere renenn Fig. 10(b) at the modulated frequency of 0.1 Hz. The differenn is is plotted pllot p otte t d in in tial normalised phase at line VI of the specimen ions and boundaries bounda da arriie ess (sizes) (siz zess) can can an be be Fig.10(c), and the defect locations differenti tial ti al normalised norm no rm ma alliisse ed d phase p ase ph obtained and calculated from the differential profile. 5. Conclusions Acknowledgement This Th is wo w ork r was supported by the Chinese National Natural This work S ie Sc en ncce Foundation Fou und ndatio attio ion under under Contract no. no 60776802, no. 51074208 Science an a nd no. no. 51173034, 51 5 117 117 1730 3034 34 4, the the he Fundamental Re and Research Funds for the Central Univ Un iver ersi s ttiies es u un nder der C de ontr on trac a t no. HIT.NSR Universities under Contract HIT.NSRIF.2009025, the 111 Project (B0701 018) 01 8)) and 8 an a nd the th he e Heilongjiang He H eil i on ng gjjia jia iang Provence Provenc Project under Contract no. (B07018) GB B06 0 A512. GB06A512. Re R effe erenc er ces e References [1] Xavier Xav vie ier err P Maldague. e M Maldague. Theory Th heor eory and and practice an prrac p act a c ice of infrared technology for nondessttr tru r cti ctive ve e testing. tes esting. New w York: Y Yo ork: k:: Wiley; Wiley Wi Wil y; 2001. y; 2001 destructive [ Manyong [2] Man any an yo yon on o ng Choi, Ch C hoi, o Kisoo Kang, Ka an a ng, Jeonghak JJe eo on ong ng nghak ha h ak ak Park, Pa Wontae Kim, Koungsuk Kim. Qua Q Qu u ntitative determination n off a subsurface ssu ubs bsu b s rface defect of reference specimen by Quantitative lo ock-in thermography. NDT&E oc NDT DT T&E Int Int 2008;41:119–24. In 200 2 08; 8;;41: lock-in [3] Busse Bu usse G, Wu D, Karpen W. W. Thermal Th al w a wave imaging with phase sensitive modulated mo odulated thermography. J Appl od Appl Phys 1992;71(8):3962–5. 1992 [[4 4] Achenbach Acch A Ach chenbach JD, Rajapakse Y. Y. Solid Soli lid mechanics li m cch me han [4] research for quantitative nondes d estr tr ctive evaluation (QNDE). tru (QN QN QND NDE E). London: Londo Lo n n: Springer; Spr destructive 1988. Gularannella Spagnolo P. Defect detection in aircraft [5]] D’ D’ Orazio Oraz r io T, Gularannell ra la C C, Le Leo e M M,, S p n pag composites com mpo pos o ites by using a neural ne eural approach appro ap oach in i the analysis of thermographic images es. NDT&E es NDT&E Int 2005;38:665–73. 200 00 0 05;3 5;; 8:6 5 :665– :6 66 65– 65 5–73. 5 3.. images. [6]] Saintey Sa ntey Sai y MB, MB, Almond Almon on nd DP. D . Defect DP De D Defec efe t sizing sizing by b transient thermography II: a num n nu um u meriicca cal al treatment. a treatm ment en ntt. J Phys n Ph hy hys ys D Appl ys Appl Phys App Phys 1995;28:2546. 19 numerical [7] Carlomagno Car a lo ar lo om mag gn g no GM, GM M, Berardi Be B erarrdi di PG. PG. G Unsteady Uns nst sstteady thermotopography ther in non-destructive tes esstin ng g.. In: In:: Warren Wa War arr ar r en n C, C, editor. editor. e r.. Proceedings Proceeding of the III infrared information testing. ex xccch x hange; 1976 hange; 197 19 1 9 6 pp. pp. 33–40. pp 33 3 3–40. –40 –40 exchange; [8 [[8] 8 8]] Liu Liiiu L u Junyan, Jun JJu un u nyan ya , Wang Wa an ng g yang, ya an ang n , Dai Da ai Jingmin. Jingmin. Research Resear on thermal wave processing of lloc oc ock-i k in n thermography th therm erm rmog ogr o grra g aph phy based p b sed on analyzing image ba i lock-in sequences for NDT. Infrared Phy hys Te hy T ech chn hnoll 2010;53:348–57. hn 2010 10;53 10 ;53 ;;5 53 5 3:348–57. Phys Technol te ht l: tp +8 :/ 6 /w 41 w 18 w 3 .il 72 gw 6 el 985 ls .c om antita tative ta tat e analysis anallysis of the boundary, bound nd dar ary y,, This paper proposes the quantitative face e defects defe de fect cts by by lock-in lock-in thermograthermogrraalocation and depth of subsurface o reduce re ed du uce ce noise noisse disturbance, no disturbance, and the th he phy. The phase is normalised to al equation eq qu ua attiio on (PDE) (PD P E) model is used d to to heat transfer partial differential mage ge g e noise. nois nois no isse e. The e. Th T he differential normalnorm malfilter the normalised phase image em mploye y d to to determine determine the e defect def efect ised phase profile method is employed e maximum, ma ax xiim mu um, minimum and and nd zeros location, boundary and size. The fectt boundary bou und ndarry and and location loca ati tion in the points correspond to the defect profi file. An fil An artificial arrttiificial neural neurrra al network n net e work differential normalised phase profile. epth of of the th he defect. deffec ect. t The location loc oca attiio on and nd nd is proposed to estimate the depth ment with with ith the it the he actual accttu ual a value valu ue for for o all of of size of the defect are in agreement thes esse values va allu ue es are are re less l ss than le n 10%. 10 0%. the specimens, and the errors for these mated by y the t e ANN th ANN NN is is close clo clos cl osse to to the The depth of the defect estimated eel. The proposed propo posed po o d method met met etho h d is is helpful hel elpf p ul actual value for ANSI 1045 steel. def e eccts ts using usiin us ng lock-in loc ock ck-iin n in the quantitative analysis of subsurface defects thermography.