Neural Network Based System for Nondestructive Testing of Composite Materials Using Low-Frequency Acoustic Methods

Transcription

1 Universl Journl of Engineering Science 1(3): , 2013 DOI: /ujes Neurl Network Bsed System for Nondestructive Testing of Composite Mterils Using Low-Frequency Acoustic Methods V.S. Eremenko, A.V. Pereidenko *, E.F. Suslov Deprtment of Informtion-mesuring Systems, Ntionl Avition University, 03680, Kiev, Ukrine *Corresponding Author: Copyright 2013 Horizon Reserch Pulishing All rights reserved. Astrct The im of the reserch ws: the scientific justifiction nd development of non-destructive testing system of products mde of composite mterils using low-frequency coustic methods. During the work the following prolems hve een pointed nd solved: 1. Anlysis of the sttisticl chrcteristics of informtion signls nd forming set of dignostic prmeters. 2. Justifiction of necessity to use rtificil neurl networks for the technicl stte clssifiction of products from composite mterils. Comprtive nlysis of clssifiction nd decision mking using of the sttisticl methods (sed on chi-squre sttistics, metric distnces, etc.), seprting hyperplnes nd neurl networks. The type of neurl network ws defined, s se for the neurl network sed clssifier of composite mterils defects. 3. Hrdwre nd softwre development of informtion-dignostic system for non-destructive testing of products from composite mterils. Developed softwre includes three min prts: mthemticl support, dtwre nd I/O module softwre. 4. Experimentl investigtion of developed informtion-dignostic system in generl. rudders nd elevtor (Figure 1), ilerons etc. (Imges tken from Keywords Nondestructive Testing System, Composite Mterils, Honeycom Pnel, Mechnicl Impednce Anlysis, Low-Velocity Impct, Neurl Network 1. Introduction It is hrd to imgine modern ircrft industry without composite mterils. Nowdys the use of composites growths rpidly in ircrft design especilly in mnufcturing the highly loded criticl elements of the ircrft. The sndwich pnels (sed on the styrofom) nd honeycom structures (which re mde of luminum fillers nd cldding from composite mterils tke specil plce in the ircrft structure. Composites re widely used in such prts of the ircrft like Composites re widely used for externl ircrft cldding, flps, htches, sides of the power components nd floors, Figure 1. Honeycom pnels in rudder () nd elevtor () structure There is lrge vriety of possile defects of products from composite mterils which used in modern ircrft [1]. The most widespred types of defects of products from composite mterils (Figure 2, imge tken fromhttp:// re the following: Type A delmintion etween plies of outer skin, prllel to surfce;

2 96 Neurl Network Bsed System for Nondestructive Testing of Composite Mterils Using Low-Frequency Acoustic Methods Type B disnding etween the outer skin nd the honeycom core; Type C crcked honeycom core prllel to the inspection surfce; Type D crushed honeycom core in prllel to the re; Type E disonding etween inner skin nd honeycom core; Type F fluid ingress in honeycom core. Figure 2. Typicl defects of composite mterils These defects re difficult to dignose, especilly on the erly stges of their development. The dignostics of composite mterils is chrcterized y the strong influence of rndom fctors which cused y chnges in the composites properties (cused due to the complexity of mnufcturing processes), imperfect inspection techniques nd flw detection equipment, high mount of types of possile defects nd other fctors [2]. Non-destructive testing of composites vi common methods such s x-ry nd ultrsonic methods is very complex ecuse of the high heterogeneity of the composite structure. It is difficult to mke enough stndrd smples for dignostic system trining tht is nother ig prolem of the non-destructive testing of products from composite mterils. Tht s why, the ccurcy of the dignosis of composite mterils is determined not only y physicl methods for otining experimentl dt ut lso with mthemticl models nd dt processing methods which re used for creting the set of informtive prmeters nd dignostic decisions mking. 2. Prolem Definition The nlysis of the set of signls informtive prmeters, i.e. performing multi-prmeter control, is the one of the possile wys to increse effectiveness nd reliility of the non-destructive testing of composites [3]-[5]. The methods of spectrum nlysis nd pttern recognition methods (in prticulr seprting hyperplnes) re often used in multi-prmeter control for dt processing nd decision rules mking. However, the ppliction of these methods hs numer of drwcks [6-8]. The choice of sis for spectrl trnsform is chllenge during spectrum nlysis. A lrge numer of spectrl components, which re not lwys sensitive to defects in the specific mteril, should e processed for signl nlysis. The components of higher frequencies tht hve low energy nd re strongly distorted y noise should e nlyzed too. There re dditionl difficulties with the development nd implementtion of pproprite mthemticl softwre for non-destructive testing system.. It is necessry to use sophisticted methods of informtion signls processing (sed on the chi-squre test or other sttisticl tests), which leds to the complex decision rules [9]. If multidimensionl spces of dignostic prmeters re linerly inseprle then construction of seprting hyperplnes leds to the solution of complex nonliner equtions with lrge numer of components (the equtions numer is determined y the numer of informtive prmeters which used for nlysis). This is quite difficult or even impossile in some cses from the computtionl point of view nd reduces the efficiency of the non-destructive testing system. The existing non-destructive testing systems do not llow to solve the cluster nlysis tsk in the utomtic mode. It is impossile to mke chnges quickly in the system memory out existing or new clss of the defect, s well s enlrge system knowledge without reclcultion of the ll system prmeters. There re mny methods nd systems tht perform dignosis nd technicl stte clssifiction of products from composite mterils [10]-[13]. However, their usge does not llow perform high qulity of nondestructive testing such products ecuse composite mteril hve complicted structure. Nondestructive testing of composites should e performed with methods tht would llow to collect the most comprehensive informtion out new defects, expnd existed se of defects nd increse dignostics system precision without restrting (ut in runtime). Since the numer of dimensions of the fetures spce (tht used for decision rules mking) is high it is difficult to use sttisticl methods for these purposes, ecuse the nlysis of high dimensionl proility distriution functions should e performed. It results in significnt usge of hrdwre nd time resources. Nowdys the developing of nondestructive testing systems for controlling the technicl stte of composites with complicted structure is ctul prolem. There re mny reserches devoted to prolems of dignostic systems construction nd such issues s improvement of efficiency, ccurcy, noise immunity, reliility, performnce, nd softwre polymorphism of these systems [14]-[16]. The solution requires systems tht could use severl control methods tht re fst nd efficient for dignostics of composites nd oriented for control wide set of possile products without lrge chnges in the min hrdwre nd softwre structure [14]. This work depicts the reserch of n lterntive wy of dt processing nd decision-mking rule mking during the multi-prmeter control of products from composite mterils, nmely, the ppliction of rtificil neurl networks. There re mny scientific ppers which descrie nd prove the effectiveness of neurl networks ppliction for the dt signls preprocessing in non-destructive testing. However, this is not the only re of their possile ppliction in non-destructive testing. A trined neurl network is le to recognize (clssify) signls received from

3 Universl Journl of Engineering Science 1(3): , the sensors s well s to sve importnt informtion out the reltionship etween the signl s wveform nd the technicl stte of the oject under control [17,18]. This llows the neurl network to clssify new signls nd possile defects tht were not previously known nd never met during its studies, i.e. dynmiclly enlrge system knowledge. Appliction of neurl networks in non-destructive testing of products from composite mterils tsks is descried very smll nd therefore it hs gret scientific interest. And the tsk of developing neurl network clssifier forinformtive-dignostic systems of composites is ctul nd importnt. through the oject under control. The sensor s output signl is sequence of dmping impulses. An exmple of the impednce split-comined sensor output signl is shown on Figure 4. Chnge of impednce in the re under control ffects the mplitude, frequency, phse, nd wveshpe of the sensor's output signl. Tht is why these prmeters re used s informtive for the clssifiction nd dignostic decision mking in cse of the impednce non-destructive testing method. 3. Solution Description Low frequency coustic methods sed on dry point contct re widely used for non-destructive testing of composites [19,20]. The most populr mong them is the pulse vrint of the mechnicl impednce nlysis (MIA) [21-22] nd the low-velocity impct method [23,24]. These methods of dignostic llows to find the iggest numer of the most dngerous defects of composite mterils [3,4,24,25]. The growth of composites use in ircrft requires dditionl reserches nd improvements to this clss of methods. The impednce method is sed on the using of ending virtion. The method ws first proposed in the 1960 yer nd ws improved since tht time. Impednce method is the one of the min nondestructive testing methods, which re used for the control of honeycom sndwich pnels nd multilyer structures tht mde from polymeric composite mterils, metls nd their comintions. This is nondestructive testing method tht sed on the differences of mechnicl impednce in defective nd defect free res of oject under control. The mechnicl impednce Z is complex rtio of the force F which ffects on surfce of the mechnicl system to the men oscilltory speed of the surfce in the direction of the force: F ZКZН Z = (1) v Z + Z ZК 1 jωk К elstic resistnce; K contct flexiility; К Z Н mechnicl impednce in the contct re etween the sensor nd mteril; ω = 2πf circulr frequency. This vlue depends on the contct resistnce nd elstic flexiility in the re under control nd sed on chnge of the mteril stiffness in the re under control. The impulse impednce method involves the usge of two types of sensors: comined nd split-comined sensors. When comined sensor is used the estimtion of mechnicl impednce performed in the sme re s the excittion of elstic wves. In this sensor different piezoelectric elements re used for virtion receiving nd excittion. The split-comined sensor (Figure 3) hs two cousticlly nd electriclly seprted virtors which re connected only К Н Figure 3. Mechnicl impednce split-comined sensor structure: 1 sensor cse; 2 emitting piezoelectric oscilltor; 3 receiving piezoelectric oscilltor; 4 stel pltes; 5 piezo; 6 contctor; 7 testing oject; 8 defect. Figure 4. Signls receiving from impednce split-comined sensor The low-velocity impct method is sed on mesuring prmeters of shock impct on the oject under control. The specilly creted force sensor (shown on Figure 5) is used s trnsducer. The min elements of the sensor re the firing pin nd the piezocermic plte tht converts the energy of deformtion into n electricl signl.

4 98 Neurl Network Bsed System for Nondestructive Testing of Composite Mterils Using Low-Frequency Acoustic Methods During the dignostic process the sensor consistently pplies to the entire dignostic surfce surgicl strike with some energy (which does not entil the destruction of the oject under control). At the moment when firing-pin with kinetic energy A ffects on the oject under control n impulse of shock interction force rises. This impulse is chrcterized y n mplitude, durtion, wveshpe nd rings rich informtion out technicl stte of oject under control. Oscillogrms of detected impulses using low-velocity impct method re shown on Figure. As depicted on Figure 6 the dmge level vrition of the smple cuses chnging of impulse mplitude, durtion nd wveshpe. Therefore, these chrcteristics were used s dignostic prmeters during the control of products from composite mterils using low-velocity impct method. The min prt of informtion out technicl stte of oject under control is concentrted in signl wveshpe. The method of low-velocity impct is effective for dignostics of ftigue dmge, undle defects, link reks, defects which egin to orn nd lso crck-like defects with millimeter size etc. The lock-digrm of the developed NDT system is depicted on Figure 7. The system s hrdwre prt consists the next locks: 1) Set of sensors force sensor nd impednce sensor, trnsformer initil physicl prmeters into the electric signl (Figure 8); 2) Dt cquisition unit dt cquisition pltform sed on the ADC nd specilized modules or specilized defectoscope tht is used for experimentl dt cquisition; 3) I/O Module intercts with defectoscopes or with specilized dt retrieving system nd trnsmits dt for susequent dt processing. Figure 5. Mechnicl impednce split-comined sensor structure: 1 ody; 2 solenoid ; 3 rod, 4 nchor, 5 clutch, 6 peen, 7 the force trnsducer shock interction, 8 nut, 9 return spring, 10 wsher 11 lod mss, 12 cover, 13 tip; Figure 6. Impulses of impct interction force

5 Universl Journl of Engineering Science 1(3): , Figure 7. Block-digrm of the NDT system Figure 8. Impulse sensors: force sensor () nd impednce sensor () The dt cquisition unit in the system is sed on n nlog-digitl converter (ADC) Acute DS which is shown on Figure 9 (The imge is tken from the wesite of the compny Tem Solutions Inc., Aville t: The Acute DS-1202 is PC sed dul chnnel ADC with dynmic rnge of ± 10V, ndwidth of 100 MHz/chnnel nd smpling rte of 200 MS/s. It is directly connected to PC vi USB 2.0 connection. Figure 9. ADC Acute DS-1202 As shown on the lock digrm (Figure 7) the sensor contcts with the oject under control nd forms informtive signl. This informtive signl crries informtion out the technicl stte of the oject. Then vi dt cquisition unit nd specil I/O module detected signl is sent to the PC. The next stge is the signl's processing nd dignostic decision mking tht performed vi specilly developed softwre. A key element of system softwre is clssifiction module sed on neurl networks. Appliction of neurl networks for the experimentl dt processing gives the system some dvntges tht re sent when using sttisticl methods nd seprting hyperplnes for clssifiction of defects [26-31]. These dvntges re: 1) Solve oth clustering nd pttern recognition tsks; 2) Allow to work with lrge dimension vectors of dignostic prmeters which incresing the reliility of control. 3) Flexiility nd dptility of the network rchitecture for vrious tsks; 4) High performnce ecuse of prllel dt processing; 5) It is possile quickly crete functioning systems without extensive construction of comprehensive

6 100 Neurl Network Bsed System for Nondestructive Testing of Composite Mterils Using Low-Frequency Acoustic Methods mthemticl models nd their reserch; 6) No needs of lrge numer of stndrds for network trining; 7) Aility to retrin nd chnge network chrcteristics while modifiction is needed. In the Tle 1 is given the comprison of the chrcteristics of the min known NDT systems of products from composite mterils with the chrcteristics of the developed system. Presented tle shows tht the developed system hs severl dvntges which include: the ility to perform cluster nlysis, non-liner clssifiction, prior trining on stndrd smples re not required nd others. This pproch lso hs some disdvntges tht should e considered when developing neurl network sed systems for dt processing: The sence of forml theory for the pproprite neurl network models nd rchitectures selection. There re no scientific sed rules for setting specific network prmeters, which mkes necessry to rely on empiricl estimtes of the nlysis of such systems. For choosing neurl network rchitecture tht cn e used for solving stndrdless dignostic tsks it should stisfy the following conditions: 1) Perform unsupervised lerning 2) Solve oth clustering nd pttern recognition tsks. Tle 2 shows tht mong the most populr neurl network rchitectures the sic requirements re stisfied y set of neurl networks of dptive resonnce theory (ART) [32-36]. So these types of network rchitecture cn e used for solving tsks of nondestructive testing of composites. Also, for solving these tsks cn e used comined (hyrid) neurl network [11,17] sed on Kohonen network (cluster nlysis) [37] nd multilyer perceptron [38] or RBF networks (detection nd clssifiction of defects type). Tle 1. Comprison developed system with existing systems Existing informtion-mesuring systems Developed informtion-mesuring system Dignostic prmeters Amplitude-time chrcteristics, mplitude nd phse spectrum, wvelet spectrum, sttisticl chrcteristics Amplitude-time chrcteristics, shpe of informtion signls Methods of clssifiction using rtificil neurl Comprison with treshold, the difference of mplitude Dignostic decision- networks spectrum, sttisticl hypothesis testing, metric mking rules (metric distnce, seprting hyperplnes, pttern distnce, discriminnt functions recognition) Clssifiction type Liner, using mthemticl nd sttisticl methods Nonliner, using neurl networks Aility to perform cluster nlysis Not Aville Aville Previous study with stndrd smples Required Not Required Automtic clssifiction of defects Not Aville Aville Tle 2. Neurl Networks Clssifiction Prdigm Lerning rule Architecture Lerning Algorithm Possile Tsks Supervised lerning Unsupervised lerning Mixed Error correction Single-lyer or Multi-lyer perceptron Perceptron s lerning lgorithms, Bck propgtion Pttern recognition, Approximtion function, Prognosis Boltzmnn Recurrent neurl network Boltzmnn lerning lgorithm Pttern recognition He's rule Feedforwrd neurl network Liner discriminnt nlysis Dt mining, Pttern recognition He's rule Hopfield network Associtive memory lerning Associtive memory Computtionl, Clustering, LVQ network Vector quntiztion Dt compression Clustering, Computtionl SOM (Kohonen mp) SOM (Kohonen mp) Dt mining ART-1, ART-2, Clustering, ART networks Fuzzy-ART Pttern recognition Pttern recognition, Error correction nd RBF network RBF lerning lgorithm Approximtion function, Computtionl Prognosis

7 Universl Journl of Engineering Science 1(3): , Figure 10 nd Figure 11 shows the lock digrms nd functionlity lgorithm, respectively, of the neurl networks of dptive resonnce theory, nmely ART-2 nd Fuzzy-ART [31, 34-36]. ART network contins two lyers of neurons: 1) Bottom-up lyer or comprison lyer which performs signl preprocessing 2) Top-down lyer or recognition lyer which performs clssifiction nd definition of mtching the input signl to stndrd in network memory. 3) performed gin ut without suppressed neuron. signl from the winner neuron in the recognition lyer comes ck together with weights of n pproprite neuron to the comprison lyer. In the comprison lyer input signl is checked for complince with the pttern in the network memory. If this pttern does not correspond to the input signl, the ctivity of the winner neuron in the recognition lyer must e suppressed nd the serch phse for the new winner neuron (pttern) in the network memory should e. Figure 11. The ART neurl network lgorithm Figure 10. Block digrm of ART-2 () nd Fuzzy-ART () neurl networks (originl works [34-36]) For comprison of the degree of difference etween the input nd stndrd signls, specil mtching unit is used. During the clssifiction process, the input signl is processing ccording to the weights of the neurons in comprison lyer of n ART-network. The signl is then fed to recognition lyer where the competitive serch of the one single winning neuron (the neuron with n ctive output signl) is performed. Winner neuron is neuron whose weight vector lies closest to the input vector (signl). This neuron, tht ecomes ctive under the influence of the input signl, corresponds to prticulr clss (pttern), which ws formed in the memory of neurl network nd could include n input signl. On the next stge feedck The lgorithm is repeted until correct pttern is found or ll neurons in the recognition lyer re suppressed. If ll neurons in the recognition lyer re suppressed, then it mens tht the input signl elongs to the new pttern. This new pttern ws not registered y neurl network erlier nd it didn t similr to ny of the previously registered ptterns, stored in the network memory. In this cse, the neurl network will generte new clss nd will provide new neuron in the recognition lyer. If during the serch phse correct pttern (pproprite neuron with the corresponding weights) is found in the recognition lyer, then the serch phse stops nd neurl network genertes s its output signl tht corresponds to one of the existing clsses. Weights of the winning neuron in the recognition lyer will e dpted ccording to pproprite network teching rules. So desired signl pttern in network memory will in more generl form descrie the group of signls which elong to the pproprite clss. Interction etween dt cquisition unit nd PC, informtion strem orgniztion, dt processing lgorithms reliztion nd displying the result ws implemented y uthors in softwre prt of the system (mthemticl support, dtwre nd I/O Module softwre). Figure 12 depicts the developed softwre structure of the descried nondestructive

8 102 Neurl Network Bsed System for Nondestructive Testing of Composite Mterils Using Low-Frequency Acoustic Methods testing system. The softwre structure consists from the next min prts: mthemticl support, dtwre nd I/O Module softwre. There re different softwre pckges such s MthL [39], NeuroSolution [40], NeuroPro [41], STATISTICA [42] nd others tht provide the ility to work with mthemticl pprtus of neurl networks (NN). These pckges could e used for solving the wide verity of prolems. But in other hnd they restrict the ville NN rchitecture usge nd do not llow performing djustment on neuron level, creting dd-ons nd developing new lerning nd functionl lgorithms of NN. Some rchitectures of neurl networks (for exmple some custom hyrid rchitectures) re sent in such softwre pckges. Bsed on this fct softwre development of the nondestructive testing system ws performed y uthors in NI LVIEW 2011 development environment [43,44]. This environment provides grphicl progrmming lnguge G nd does not restrict the NN model implementtion. NI LVIEW llows developing new lerning nd functionl NN lgorithms, creting new types of NN or merging/joining existing nets in one functionl system. Also it is possile to use existing progrmming modules nd DLLs. LVIEW 2011 environment performs prllel execution of lgorithms which increses the performnce of creted nondestructive testing system softwre. The nondestructive testing system softwre is developed y uthors like set of independent modules ecuse of the need to solve the prolem of stndrdless dignostics for different oject under control types with usge of only one system with possile minor chnges for ech specific tsk (difference could e in chnging few existing functions or including new one for specific dignostics tsks). It llows connecting nd integrting creted erlier lirries nd modules (which were developed with high level lnguges) in the system core without its significnt chnges. This pproch permits dding new or excluding unnecessry functionlity, performing fst reconfigurtion, moderniztion, dpttion, etc. without ny complexity. There re next min modules of system softwre: 1) ADC control module. This module includes units for configuring ADC prmeters (smpling rte, dynmic rnge of input signls, numer of ADC chnnels, synchroniztion source, etc.). 2) System s regime control module. It is used for configuring such prmeters of the system regime like the input dt source (defectoscope, ADC or physicl/opticl dt storge), system regime, etc. 3) Dt storges interction control module. It llows reding/writing dt from/to dt storge. It lso forms rry or surry of necessry informtive prmeters for importing them to other dignostics systems or for creting reports. In cse when reding opertion is performed this module forms the rry from the stored dt nd represents it in required formt. 4) Reporting fcilities nd system output control modules. These modules perform dt trnsformtion to convenient user-friendly formt in cse of dt displying or to required dt storge formt in cse of dt sving. Also it is possile to disply or sve on storge prmeters of neurl networks which re used in the system. 5) Control module for system dtse. It llows red/write set of formed clsses during system work s well s their dditionl specific prmeters. 6) Control module for neurl networks. It llows user to choose the type of neurl network tht will e used for dt processing nd clssifiction nd setup its prmeters. This module contins units which perform clss forming of OUC, finding new clsses, nomlies, nd converting the neurl network output to user-friendly formt. 7) Dt uffer control module. This module writes retrieved dt from ADC or defectoscope to rry which further could e processed or sved for processing in future y current or nother dignostics system. 8) Dt processing module. It llows choosing the necessry dignostics chrcteristics which the most informtive for current cse nd use them insted the full set of chrcteristics. In some cses the dt vectors of smll dimension size could e used. For efficient dignostics system work the input dt should e preprocessed to in specific wy. Tht s why in this module the necessry lgorithms re executed. The user interfce (Figure 13) of developed system s softwre contins control elements for ADC setting up, prepring nd normlizing input dt, choosing neurl network type, choosing the NN lerning lgorithm nd other prmeters of NN. System's user interfce lso contins elements for sving detected dt signls, loding dt form physicl or opticl storges, output reports sving, etc.

9 Universl Journl of Engineering Science 1(3): , Figure 12. Nondestructive testing system softwre rchitecture Figure 13. User interfce of the NDT system sed on ART-2 nd Fuzzy-ART neurl networks The system cn e used in oth mnul nd utomtic modes. The system llows: 1) Input signl visuliztion; 2) Detecting input signls informtive prmeters; 3) Performing input signl preprocessing; 4) Performing input dt cluster nlysis; 5) Determining nd uilding seprting hyperplnes; 6) Clssifying the stte of OUC nd displying the chrcteristics of its clss; 7) Sving the vlues of selected informtive prmeters; 8) Creting decision rule; 9) Converting otined results to user-friendly formt. 4. Experimentl Prt For the developed system testing were used smples of honeycom pnels which used in ircrft nd presented on Figure 14. Experimentl smples were provided y the Stte Enterprise "Antonov" nd used in the "An" ircrft models mnufcture. Two smples (S1 nd S2) of honeycom pnels from wings of ircrft type An-70 (Figure 14, 14) of the mteril ELUR-P-0.1 nd honeycom filler PSP with 10 mm of thickness were used. As defect models the res with rtificilly formed defects were used. The following type of defects ws under control: peeling skin on the inside of the filler cldding. The smples hd six types of res: one defect-free nd five with different defectiveness level. Ares of defects under control were divided on the points with distnce etween ech point equl 2 mm. Control ws mde from the outside of the cldding (s shown on this

10 5 104 Neurl Network Bsed System for Nondestructive Testing of Composite Mterils Using Low-Frequency Acoustic Methods figure). Scnning ws performed with discrete step equl 2 mm in the direction indicted in Figure 15. The numers ner the orders of res with defects men numer of scnning points, were re with defect is egun nd ended. Figure 14. Smples S1 () nd S2 () of honeycom pnels (ottom view) Top view Smple 1 Scnning Defect 3 10 Scnning Defect 1 5 cm 4 2 Scnning Defect 2 5 cm 2 cm Figure 15. Loction of defects in smples S1 () nd S2 (). Scnning ws mde with step 2mm The developed softwre hs een tested for the detection nd dignosis of defects in descried smples of composite mterils. Figure 16, 17 depicts exmples of otined informtive signls during nondestructive testing. The dignosis of defects in descried smples of honeycom pnels with rtificil defects nd signls processing ws performed vi three groups of methods: 1) Anlyzing the mplitude nd durtion of the registered signls in individul nd complex (y grouping these two prmeters to the ggregte feture spce) wy, nd constructing etween different oject sttes (defect without defects) seprting hyperplnes vi itertive methods descried in [6,30]. 2) Sttisticl nlysis of the registered signls, which includes spectrl trnsformtion of informtion signls nd susequent nlysis of the spectrl densities using chi-squre distriution. 3) Anlyzing the chnge of the registered signls shpe vi neurl networks of dptive resonnce theory (ART-2 nd Fuzzy-ART) [33,34], ecuse chnges of signls shpe llow with high reliility detect defects presence in honeycom pnels nd clssify its type.

11 Universl Journl of Engineering Science 1(3): , Figure 16. Informtive signls from S1 () nd S2 () otined using the impednce dignostic method Figure 17. Informtive signls from S1 () nd S2 () otined using the low-velocity impct dignostic method

12 106 Neurl Network Bsed System for Nondestructive Testing of Composite Mterils Using Low-Frequency Acoustic Methods For the nondestructive testing of descried smples for ech smple nd ech re (without defect nd with different defects) 250 reliztions of the informtive signls were otined. During the control process registered signls were given one y one in csul order to the neurl network entrnce. After n pproprite signl ws given to system, it ws removed from input smples collection. This procedure ws repeted until the input smples collection ecme empty. After the clssifier hs formed its dt se the verifiction of the nondestructive testing reliility vi descried softwre hs een performed. For these purposes for ech re 100 new signls were otined nd given to the neurl network for clssifiction. Figure 18 depicts the results of investigtions of the control reliility using the developed non-destructive testing system. As the physicl dignostic method the impednce method ws used. For the comprison, clssifiction of defects ws performed using the chi-squre sttistic, seprting hyperplnes nd descried the ART-2 nd the Fuzzy-ART neurl networks. Amplitude, frequency, phse, nd wveshpe of the sensor's output signl were used s informtive prmeters for the clssifiction nd dignostic decision mking in cse of the impednce non-destructive testing method. Otined results shows tht the decision rule sed on sttisticl nlysis using chi-squre distriution s well s the decision rule sed on seprting hyperplnes constructed in the coordinte system mplitude-durtion filed to detect defects with plottge less then 2 cm 2 (defect 3 on the smple 1 nd defect 5 on the smple 2). At the sme time using the mplitude nd durtion of the dt signls s informtive prmeters during control defects with plottge greter then 5 cm 2 (defect 1 on the smple 1 nd defect 4 on smple 2) it ws otined reliility of the control out 80%. Using sttisticl nlysis during control the sme defects it ws otined reliility of the control out 83-85%. Usge of ART-2 nd Fuzzy-ART neurl networks for the registered informtion signls processing during the control of honeycom pnels llows to determine ounds of defective res 1 nd 2 with plottge 5 nd 2 cm 2 respectively on the smple 1 nd defective re 1 with plottge 5 cm 2 on the smple 2. Also, unlike the groups of methods 1 nd 2, clssifier sed on ART-2 nd Fuzzy-ART neurl networks hs detected ounds of the defective res with plottge 2 cm 2 (defect 3 on the smple 1 nd defect 5 on the smple 2). So, it could e extended tht usge of clssifier sed on ART-2 nd Fuzzy-ART neurl networks in nondestructive testing of honeycom pnels llows to detect defects with plottge greter then 2 cm 2 which is difficult or impossile with most currently used methods of dignosis. The reliility of control vi descried clssifier using impednce dignostic method is over then 90%. Figure 18. The reliility of non-destructive testing of S1 () nd S2 () using the impednce dignostic method Figure 19 depicts the results of investigtions of the control reliility using the developed non-destructive testing system. As the physicl dignostic method the low-velocity impct method ws used. Clssifiction of defects ws performed using the chi-squre sttistic, seprting hyperplnes nd descried the ART-2 nd the Fuzzy-ART neurl networks. Amplitude, durtion nd wveshpe of the sensor's output signl were used s informtive prmeters for the clssifiction nd dignostic decision mking in cse of the low-velocity impct method. Otined results lso shows tht the decision rule sed on sttisticl nlysis using chi-squre distriution nd sed on seprting hyperplnes constructed in the coordinte system mplitude-durtion filed to detect defects with plottge less then 2 cm 2 (defect 3 on the smple 1 nd defect 5 on the smple 2). During control defects using low-velocity impct method with plottge greter then 5 cm 2 (defect 1 on the smple 1 nd defect 4 on smple 2) it ws otined reliility of the control out 90% which on 10% greter then with using impednce method. Using sttisticl nlysis during control the sme defects it ws otined reliility of the control out 90-93% which is greter on 7-8% greter then with impednce method.

13 Universl Journl of Engineering Science 1(3): , The developed neurl network clssifier sed on networks of dptive resonnce theory ws used for the clssifiction of defects of products from composite mterils. This increses the reliility of technicl stte clssifiction of products from composite mterils compred with sttisticl methods nd seprting hyperplnes y 24%. Developed softwre dvntges re in its rchitecture flexiility, high performnce nd reliility of dt signl processing, humn-engineered interfce. The results of the reserch integrted to eductionl nd reserch processes t the deprtment of informtion-mesuring systems of the Ntionl Avition University. REFERENCES [1] NDT Toolox for Honeycom Sndwich Structures comprehensive pproch for mintennce inspections : (ATA NDT Forum 2010 Aluquerque) / Wolfgng Bisle // E-ook Aville: Figure 19. The reliility of non-destructive testing of S1 () nd S2 () using the low-velocity impct dignostic method Usge of ART-2 nd Fuzzy-ART neurl networks for the registered informtion signls processing during the control of honeycom pnels llows to determine ounds of defective res with plottge 5 cm 2 s well s with plottge 2 cm 2 on the smple 1 nd on the smple 2. The reliility of control vi descried clssifier using low-velocity impct dignostic method is over then 95% which is greter on 5% then using impednce method. Bsed on results it could e mentioned the following. Using low-velocity method is preferly for dt cquisition during nondestructive testing of honeycom pnels with defects of type: peeling skin on the inside of the filler cldding. Processing of cquired dt using neurl networks ART-2 nd Fuzzy-ART llow increse reliility of control to 95%. 5. Conclusion The nondestructive testing system of products from composite mterils ws creted nd investigted. Artificil neurl networks re used for dt processing, defects clssifiction nd mking decision rules in the dignostic system. The developed dignostic system provides high reliility of the nondestructive testing: 90-95% tht is etter thn nlogous existing systems. Also there is no need for mking the stndrd smples for its trining. [2] Eremenko V.S. Informtion-mesuring system fordignostic s of products from composite mterils: PhD Thesis / V.S. Eremenko Кiev, (In Ukrinin) [3] Lnge Y., Nikolev S., Muzhitskiy V.F., Nefedov S.V. Spectr of low-frequency pulse signl trnsducers of coustic defectoscopes. - Nondestructive Testing, 1996, 5, P (In Russin) [4] Lnge Y., Nikolev S.I., Muzhitskiy V.F., Lpshin V.C., Sidorenko A.S., Igonin M.A., Pvlyuchenkov N.F. The use of spectrl nlysis in the low-frequency coustic defectoscopes. - Nondestructive Testing, 1995, 10, P (In Russin) [5] Lnge Y., Nefedov S.V. Correltion signl processing of impednce defectoscope. - Control. Dignosis, P (In Russin) [6] Eremenko V.S. Appliction of liner recognition methods in prolems of nondestructive testing of composite mterils / V.S. Eremenko, О.О. Gilev // «Electromgnetic nd coustic methods of nondestructive testing of mterils nd products LЕОТЕSТ-2009»: mterils of the 14th interntionl scientific conference, Ferury 16-21, 2009.: proceedings Slvskoe,2009. P (In Ukrinin) [7] Eremenko V.S. The study of defects in cellulr screens low-frequency coustic techniques / V.S. Eremenko, О.A. Gilev, V.Y.Derech, E.F.Suslov, E.O.Pikolenko // Nuchni izvesti P (In Russin) [8] Murshov V.V. Defects in prts of monolithic nd multilyer structures mde of polymer composite mterils nd methods for their detection / V.V. Murshov, A.F. Rumyncev // Control. Dignostics ,5. P (In Russin) [9] Mokijchuk V.M. Dignostics system of products from composite mterils / V.М. Mokijchuk, V.S. Eremenko // Bulletin of the Khmelnytsky Ntionl University V.2. P (In Ukrinin)

14 108 Neurl Network Bsed System for Nondestructive Testing of Composite Mterils Using Low-Frequency Acoustic Methods [10] Cwley P. Clyton D.L. R. A virtion technique for the mesurement of contct stiffness // Mechnicl systems nd signl processing P [11] Suslov E.F. The control system of products from composite mterils / E.F. Suslov, V.S. Eremenko, V.М. Mokijchuk // «Eductionl, scientific nd engineering pplictions in LVIEW nd Ntionl Instruments technology»: mterils of the 8th interntionl scientific conference, Novemer 20-21, 2009.: proceedings. Moscow, P (In Russin) [12] Lihopoy A.A., Sysoev A.M. Devices for control with low-frequency coustic methods. - Scientific Conference "High Technologies nd Intelligent Systems 2003." Collection of scientific works. - Moscow, 2003., P (In Russin) [13] Benlri S. nd Melo W. Polymorphism Mesure for Erly Risk Prediction. In Proceeding of 21st ICSE 1999, Los Angeles USA. [14] Zeeshn, A. Towrds Performnce Mesurement nd Metrics sed Anlysis of PLA Applictions. Int. Jr. SE. & App (3): [15] Zeeshn, A. Mesurement Anlysis nd Fult Proneness Indiction in Product Line Applictions (PLA). Frontiers in Artificil Intelligence nd Applictions 2007, IOS Press, 161: Scientific & Acdemic Pulishing [16] Victor Bsili, Jens Heidrich, Mikel Lindvll, Jürgen Münch, Myrn Regrdie, Dieter Romch, Crolyn Semn, Adm Trendowicz. Linking Softwre Development nd Business Strtegy Through Mesurement. IEEE Computer 2010, 43(4): [17] Pereidenko A.V. Demerit rting system sed on rtificil neurl networks / A.V. Pereidenko, Y.V.Kuts, V.S. Eremenko// Dys of nondestructive control 2010 : mtetils of XXV Ntionl Conference with Interntionl prticiption "Defektoskopiy'10", June 13-17, 2010.: proceedings Sophi,2010. p (In Russin) [18] Pereidenko A.V. The system of stndrdless dignosis of composite mterils sed on hyrid neurl network./ A.V. Pereidenko, V.S. Eremenko, E.F. Suslov, P.A. Shegedin// Eductionl, scientific nd engineering pplictions in LVIEW nd Ntionl Instruments technology : mterils of the 9th interntionl scientific conference, Decemer 3-4, 2010.: proceedings Moscow,2010. P (In Russin) [19] Adms R.D., Cwley P. A review of defect types nd nondestructive techniques for composites nd onded joints // NDT Interntionl, v.21, 4, P [20] Muzhitskiy V.F. Control of multilyer structures with low-frequency coustic defectoscopes / V.F. Mujitskiy, A.A. Lihopoy // New mterils, nondestructive testing nd high technology in mechnicl engineering: Proceedings of the III Interntionl Scientific nd Technicl Conference. - Tyumen, 2005, p (In Russin) [21] Cwley P. The sensitivity of the mechnicl impednce method of nondestructive testing, - NDT Interntionl Vol. 20, No. 4, August p [22] Lnge Y. Acoustic impednce method of nondestructive testing (Review) / Y. Lnge // Defectoscopy P (In Russin) [23] Adms R.D., Cwley P. A review of defect types nd nondestructive techniques for composites nd onded joints // NDT Interntionl, v.21, 4, P [24] Eremenko V.S. Detection of impct dmge honeycom pnels using low-velocity impct/ V.S. Eremenko, V.M. Mokijchuk, A.M. Ovsnkin// Technicl dignostics nd nondestructive testing P (In Russin) [25] Bker A., Sturt D., Kelly D, Composite Mterils for Aircrft Structures. Second Edidion / Aln Bker, Sturt Dutton, Donld Kelly. Reston, Virgini: Americn Institute of Aeronutics nd Astronutics, Inc., p. [26] Brhtov V.A. Detection of defects using rtificil neurl network of specil type / V.А. Brhtov // Defectoscopy P (In Russin) [27] Nidenko A.G. Determining the reliility of the dimond cutters composite superhrd mterils y recording nd nlyzing coustic emission: PhD Thesis / А.G. Nidenko. Кiev, (In Ukrinin) [28] Zjickiy O.V. Recognition of ircrft engines ldes stte vi neurl networks during virocoustic monitoring: PhD Thesis / О.V. Zjickiy. Кiev, (In Ukrinin) [29] Brhtov V.A. Detection of signls nd their clssifiction using pttern recognition / V.А. Brhtov // Defectoscopy P (In Russin) [30] Pereidenko A.V. Construction of decision rules in multiprmeter NDT / A.V. Pereidenko, V.S. Eremenko, О.A. Gilev, E.F.Suslov // «Modern methods of of nondestructive testing nd technicl dignosis»: mterils of the 18th interntionl scientific conference, Octoer 5-9, 2010.: proceedings. Ylt, P (In Ukrinin). [31] Pereidenko A.V. Softwre of Informtion-Mesurement System for Stndrdless Dignostic of Composite Mterils / A.V. Pereidenko, V.S. Eremenko // Softwre Engineering v.2, 3. p [32] Pereidenko A.V. Uses of the modified ART-2 neurl network rchitecture in the nondestructive testing system of products from composite mterils / A.V. Pereidenko, V.S. Eremenko, O.V. Monchenko // "Modern methods nd instruments of nondestructive testing nd technicl dignosis": mterils of 19th Interntionl Conference, Octoer 3-7, 2011.: proceedings Gurzuf, p (In Ukrinin) [33] Pereidenko A.V. System Of Stndrtless Dignostic Of Cell Pnels Bsed On Fuzzy-ART Neurl Network / A.V. Pereidenko, V.S. Eremenko, V.O. Rognkov // " MRRS-2011": Proceedings of the Third Microwves, Rdr nd Remote Sensing Symposium, August 25-27, 2011: proceedings. Kyiv, p [34] Crpenter G.A. ART 2: Stle self-orgniztion of pttern recognition codes for nlog input ptterns / G.A. Crpenter, S. Grosserg // Applied Optics Р [35] Crpenter G.A. ART 2-A: An dptive resonnce lgorithm for rpid ctegory lerning nd recognition / G.A. Crpenter, S. Grosserg, D.B. Rosen // Neurl Networks Р [36] Crpenter G.A. Fuzzy ART: Fst stle lerning nd

15 Universl Journl of Engineering Science 1(3): , ctegoriztion of nlog ptterns y n dptive resonnce system / G.A. Crpenter, S. Grosserg, D.B. Rosen // Neurl Networks Р [37] Pereidenko A.V. The system of cluster nlysis of nondestructive testing products from composite mterils/ A.V. Pereidenko, V.S. Eremenko, V.A Rognkov// Knowledge-intensive technologies P (In Ukrinin) [38] Pereidenko A.V. The system of defects clssifiction sed on rtificil neurl networks / A.V. Pereidenko, V.S. Eremenko, J.O. Pvlenko // Bulletin of the Ntionl Technicl University of Ukrine "Kyiv Polytechnic Institute". Series Instrument Engineering p (In Ukrinin) [39] Sumthi S., Surekh P. Computtionl Intelligence Prdigms: Theory & Applictions using MATLAB. CRC Press, Inc., p. [40] [41] [42] Borovikov V.P. STATISTICA Neurl Networks: The methodology nd technology of modern dt nlysis. Мoscow: Hotline Telecom, p. (In Russin) [43] Fudoseev V.P. Digitl Signl Processing in LVIEW. Мoscow: DMK Press, p. (In Russin) [44] Zgidulin R.Sh. LVIEW in reserch nd development. Мoscow: Hotline Telecom, p. (In Russin)