A Novel Acoustic Emission Source Location Method in the Crane Based on EEMD-FastICA

Transcription

1 Journal of Mechancs Engneerng and Automaton 7 (2017) do: / / D DAVID PUBLISHING A Novel Acoustc Emsson Source Locaton Method n the Crane Based on EEMD-FastICA Yu Jang 1 * and Feyun Xu 2 1. College of Intellgent Scence and control Engneerng, Jnlng Insttute of Technology, Nanng , Chna 2. School of Mechancal Engneerng, Southeast Unversty, Nanng , Chna Abstract: The study presents a novel source locaton method based on EEMD (ensemble emprcal mode decomposton) and optmzed FastICA (ndependent component analyss) for determnng the poston of the AE (acoustc emsson) sources n the damage structure of crane. Frstly, The AE sgnals are self-adaptve decomposed nto a number of IMFs (ntrnsc mode functons) by usng EEMD algorthm. Then, the man feature IMFs sgnals are extracted as the effectve AE source sgnal by optmzed Fast-ICA method, the PSD (power spectral densty) of each IMF and the real IMFs are obtaned. Accordng to the prncple of spectrum smlarty, tme delay s computed at the dfferent channels n combnaton wth the Cross Correlaton tme delay estmaton prncple. Thrdly, a novel source locaton algorthm based on EEMD-FastICA s proposed and the results of AE source locaton are obtaned. Fnally, the three-pont bendng test for the crane s mplemented n order to valdate the effcency of the proposed method. The expermental results ndcate that the proposed method has the ablty to determnate the poston of the AE sources and reduce the nterference noses. Moreover, compared wth the tradtonal locaton algorthm, a consderable mprovement s obtaned. Key words: Source locaton, EEMD, FastICA, cranet, three-pont bendng test. 1. Introducton AE source locaton technology can determne the damaged poston of the defect,especally of early damage of materals for dynamc montorng [1, 2]. However, Due to the nfluence of AE source features, AE sgnal propagaton paths, envronmental noses and measurement system, the AE sgnal s very complex, whch s a typcal nonlnear and non-statonary random sgnal and needs to use needle of the sgnal processng technque to extract the effectve nformaton from the AE sgnal. The EMD (emprcal mode decomposton) ntroduced by Huang [3] s a good tme-frequency tool. However, EMD decomposes AE sgnal nto a number of basc consttuent sgnals called IMFs (ntrnsc mode functons), whch obtaned IMFs s not a strct sngle component sgnal, but accompaned by the IMF mode Correspondng author: Yu Jang, teacher, Dr., research felds: acoustc emsson structural damage detecton, sgnal processng. alasng and false sgnals mxed, serously mpactng on the acoustc emsson data analyss and acoustc emsson source locaton accuracy. In 2009, Wu and Huang [4] proposed an mproved verson of EMD, by addng fnte whte nose to the nvestgated sgnal, ths mproved method called EEMD can elmnate the mode mxng problem of EMD automatcally. Based on these merts, the EEMD has lately attracted sgnfcant attenton and t has been proven to outperform EMD n decomposng random sgnals. Apart from usng EEMD as a non-statonary sgnal decomposng tool, Guo and Tse [5, 6] studed the EEMD, whch eases the problem of mode mxng n real AE sgnals generated from defectve area. They performed a seres of nvestgatons to reveal the relatonshp between the ampltude of the added whte nose and the number of ensemble members for the mnmzaton of the mode mxng problem, t concluded that a hgher number of ensemble members leads to smaller RMS error. Zvokel [7] et al. proposed the EEMD-MSPCA (EEMD-based multscale prncpal

2 18 A Novel Acoustc Emsson Source Locaton Method n the Crane based on EEMD-FastICA component analyss) technques to overcome the non-adaptve nature n conventonal MSPCA. They appled ther n vbraton as well as acoustc emsson recordngs from large-sze slow-speed bearngs, t not only hghlghted the mprovement of sgnal-to-nose-rato, but also the enhanced dagnostc capablty. The only potental drawback s that the proposed technques are computatonally ntensve. However, as a matter of fact, among the approaches reported n the aforementoned lterature reles on nformaton comng both from the normal as well as the faulty class. In most real lfe applcatons, data from the possble faulty modes are not readly avalable, makng the bnary (or mult-class) classfcaton approach very dffcult and mpractcal. The man nnovaton of ths paper stems from the proposal of a combned and ntegrated anomaly detecton approach to sgnal processng. A novel CC (cross-correlaton) sgnal processng method based on EEMD (ensemble emprcal mode decomposton) and Fast-ICA (optmzed ndependent component analyss) s presented for auxlary determnng the poston of the AE sources. 2. Cross-Correlaton Source locaton 2.1 Cross-Correlaton Algorthm Tme Delay Estmaton s a typcal passve method for target parameter estmaton. CC s to use smlar sgnal n tme doman for tme delay estmaton [8], whch s the most basc method of tme delay estmaton. Eq. (1) s AE sgnal model. where, A ( ( s( A( e (1) s the envelope ampltude of AE sgnal, t s the frequency and phase of functon wth the attenuaton and oscllaton sgnal. Generally, ths model belongs to random dstrbuton form, whch reflects related features ncludng the acoustc emsson source, propagaton medum, and acoustc emsson nstrument response. Consderng the sgnal receved from the sensor wll be affected by the nose nterference, the receved sgnal model of the th sensor can be assumed as the followng [14, 15]. x ( n). s( ) v ( n) (2) where, s the attenuaton coeffcent of the acoustc wave propagaton [0,1], s the delay tme of acoustc source arrved the th sensor and v (n) s the added nose. Accordng to Eq. (2), Eq. (3) s as followng: (3) the receved sgnal model of the th sensor can be shown as Eq. (4). x ( n). s( ) v ( n) (4). s( ) v ( n). s( ) v ( n) (5) If s (n), v (n), and v (n) are ndependent of each other, then the cross-correlaton functon of sgnal receved from the two sensors wll be R, namely, R ( ) E[ x ( n) x ( )] E[ s( n) s( )]. Rss ( ) (6) where, s the delay tme and R ss s the self-correlaton functon of source sgnal s (n), based on the self-correlaton theory, when Rss ( ) Rss (0), the two receved sgnals wll have the max value. The peak pont of the correcton functon wll be the value of tme delay estmaton. On ths bass, accordng to the theory of lnear locaton, the tme dfference of the two sensors from the same acoustc source wll be, ( L L ) c (7) / where, L, L s the dstance between the acoustc and the two sensors, c s acoustc velocty and D s the

3 A Novel Acoustc Emsson Source Locaton Method n the Crane based on EEMD-FastICA 19 poston of two sensors. Then, L D c ) / 2 (8) ( L D (9) L L s the dstance from acoustc source to sensor. 2.2 EEMD The EMD, orgnally proposed by Huang [3] n 1998 s a non-lnear mult-resoluton self-adaptve decomposton technque, whch can self-adaptve decompose a complcated sgnal nto a set of complete and almost orthogonal components, IMFs (ntrnsc mode functons), wthout a prelmnary knowledge of the nature and the number of IMF components embedded n the data. However, EMD stll suffers from the mode mxng problem. To allevate the drawback of the mode mxng [9], Wu and Huang proposed [3, 4] an effectve nose-asssted method named EEMD whch sgnfcantly reduces the chance of undue mode mxng and preserves the dyadc property of the decomposton for any data. EEMD s an mproved verson of the orgnal EMD. The prncple of the EEMD s smple: The added whte nose populates the whole tme-frequency space unformly, facltatng a natural separaton of the frequency scales, whch reduces the occurrence of mode mxng. 2.3 Blnd Source Separaton and Fast-ICA Algorthm Blnd source separaton ams at the false IMFs component separaton resultng from the EEMD decomposton, whch s a new method of sgnal processng [10]. The core of ths algorthm s to estmate the separaton matrx W, and use y as the estmaton of the source sgnal x (, y shows the output of matrx W, namely, as shown n Fg. 1. ^ s( Wx( WAs( (10) The FastICA algorthm s an effcent and popular procedure for ndependence analyss. The algorthm s Fg. 1 Blnd source separaton chart. maxmzng non-gaussanty usng a fxed-pont teraton [11-13]. It can also calculate wth Newton teraton method. As the frst preprocess of the FastICA algorthm, the nput vector X should be centered: X = X-E{X} (11) Then, n the second preprocess, the nput vector data X s whtened so that the new components can be uncorrelated and have varance one. Let c = E{XX T } be the covarance matrx and d the dagonal matrx of egenvalues of c. Therefore, the whtened data are: X (12) d E T where, E s the matrx of egenvalues. The FastICA algorthm fnds a proecton drecton of Eq. (15) or maxmzes the non-gaussanty of the proecton W T τ for data X. In ths study, due to the extracton of the ndependent components, the functon f(ν) s utlzed as a non-quadrc functon wth the equaton below: f ( v) logcosh( v) (13) the dervatves g(ν) and second dervatves g (ν) are: g' v g( v) tanh (14) v 2 v 1 tanh (15) By teratvely renewng the unmxng matrx W, the fnal ndependent components can be calculated as: T T } E g ' W W E{ g W W (16) W W / W (17) After every fxed-pont teraton, the normalzaton of W has been added to mprove the stablty. When the old and new values of the W ponts are n the same

4 20 A Novel Acoustc Emsson Source Locaton Method n the Crane based on EEMD-FastICA drecton, the fxed-pont teraton s converged. 3. EEMD-Fastca Cross Correlaton Source Locaton Algorthm The EEMD-FastICA-CC Source locaton s an nnovaton combnaton algorthm, ths method contans three steps: frstly, the sgnal s decomposed by usng the EEMD algorthm; secondly, amng at the false sgnal resultng from the EEMD, the Fast ICA algorthm s ntroduced to realze the decomposton completely and remove the false component decomposton. The hgher frequency IMF component that reflects the physcal characters of sgnal s extracted; fnally, the combnaton source locaton theory and the CC technque, tme delay of the IMF components at dfferent channels s obtaned and the postons of AE source are determned. The detaled flowchart s shown n Fg Experment Descrpton 4.1 Experment Layout In order to demonstrate the valdty of the proposed source locaton algorthm based on EEMD-FastICA, three pont bendng experment s carred out. SJ9-3 of a Q235B (n Chna Code) mld steel sheet was selected as the experment sample accordng to the engneerng proects of crane. The specmen area was a 500 mm 60 mm and the thckness was 8 mm. A weld was n the mddle of the sample, two sensors resonant at 150 khz (DP15I) coupled drectly on the face of the sample usng vaselne as acoustc couplant (Fgs. 3 and 4). Sensor senstvty was evaluated usng the classcal PLB (pencl lead break) technque.e. by means Hsu-Nelsen artfcal sources. Germany Zwck/Roell Z100MST, loadng velocty was 6mm/mn. AE actvty was recorded usng a 48-channel SAMOS PAC equpment. Thresh value s 40 db, sample rate s 4, Result Analyss Seen from Fg. 5a, both energy and mpact parameters have the same trend n the two dfferent channels. Fg. 2 EEMD-FastICA-CC algorthm flow chart. Fg. 3 Three pont bendng test. So takng the frst channel for example, AE dynamc performance s descrbed by the data analyss. Seen from Fg. 5b, the experment has four dfferent stages from the loadng curve durng the whole loadng course: elastc stage, plastcty and yeld stage, crack propagaton stage and nstablty stage.

5 A Novel Acoustc Emsson Source Locaton Method n the Crane based on EEMD-FastICA 21 Fg. 4 Q235B weldng sample. Fg. 5 (a) ht vs tme AE Performance parameters for (a) and (b). (b) Energy vs tme The frst s the elastc stage (I), and ths course has a lttle AE sgnal and sparsely the energy dstrbuton, acoustc source s manly nose source. In the Plastcty and yeld stage (II), AE sgnal has the trend of stable growth and the curvature vares obvously from pont A to pont B. Ths result shows that the poston of weldng at the area of the stress concentrated begns to produce the plastc deformaton and quckly nto the yeld stage, the energy dstrbuton s stll sparse, even the ampltude of energy has an downwards tendency, pont A and pont B correspondng to the upper and lower yeld ponts respectvely. Acoustc source s manly contnuous AE source. In the crack propagaton stage (III), the dstrbuton of energy s both gradually ntensve and steady growth and the macro-crack occurs ntal deformaton. At ths moment, although the speed of the loadng s on the declne, the crack s stll rapdly and sharply expandng. After the stress of ths peak, the loadng velocty keeps ths constant value to make the sample reach a maxmum bendng sze that Unversal Test Machne allowed and stop loadng. At ths moment, the crack s stll growng. Acoustc source s mostly sudden AE source. In the nstablty stage (IV), the crack s n the unstable state tll the test ends. Acoustc source s nterrupted contnuous AE source, complex and unstable. Accordng to the above four dfferent stages of AE source character analyss, takng the crack propagaton stage for example, two channel (S1, S2) AE sgnals are extracted for the proposed new algorthm. Fg. 6 s the crack sgnals of S1 and S2. Seen from Fg. 7, eleven dfferent frequency IMF components and a resdual tem are obtaned by usng the EEMD, accordng to the IMF performance ndexes above mentoned lterature [10], the reasonable threshold value s The frst three IMF components of S1 and S2 n the crack propagaton stage are extracted to reset nput maxs for FastICA algorthm, respectvely. The results of algorthm are shown as n Fgs. 8 and 9, whch show the sgnals of the frst three IMF and

6 22 A Novel Acoustc Emsson Source Locaton Method n the Crane based on EEMD-FastICA 2.5 Channel 1 (S1) 1.5 Channel 2 (S2) AMP [db] AMP [db] Samples Samples Fg. 6 (a) crack sgnal S1 Crack propagaton stage AE sgnals for (a) and (b). (b) crack sgnal S2 Fg. 7 S1-EEMD AE sgnals EEMD for (a) and (b). S2-EEMD FastICA IMF. Fgs. 10 and 11 are the correspondng Power Spectrum Dstrbuton. In order to realze AE source locaton, hgh frequency IMF components of S1 and S2 n FastICA algorthm are extracted wth Energy-rato and spectral smlarty factors [14]. Seen from Tables 1 and 2, the sgnals of Fast-ICA-IMF3 and Fast-ICA-IMF22 are used to perform the cross-correlaton algorthm (Fg. 10). Fnally, accordng to tme delay estmaton of cross-correlaton and velocty measured v at the begnnng of experment, AE source locaton far from sensor S1 s mm n the crack propagaton stage by usng the lnear locaton calculaton, compared to the actual AE loadng pont locaton (150 mm), the error of the locaton s 0.36% (0.546 mm). It shows that the proposed algorthm s very effectve and feasble.

7 A Novel Acoustc Emsson Source Locaton Method n the Crane based on EEMD-FastICA 23 (a) FastICA IMF1-IMF3 (S1) Fg. 8 FastICA sgnals processng. (b) FastICA IMF21-IMF23 (S2) (a) PSD IMF1-IMF3 (S1) Fg. 9 PSD FastICA sgnals processng. (b) PSD IMF21-IMF23 (S2) Table 1 Hgh frequency IMF components extracted (S1). IMF Energy-rato Spectral smlarty FastICA-IMF FastICA-IMF FastICA-IMF Fg. 10 Tme delay of the EEMD-FastICA CC algorthm. Table 2 Hgh frequency IMF components extracted (S2). IMF Energy-rato Spectral smlarty FastICA-IMF FastICA-IMF Fast-ICA-IMF

8 24 A Novel Acoustc Emsson Source Locaton Method n the Crane based on EEMD-FastICA Fg. 13 Tme delay of four dfferent algorthms. Table 3 The results of tme delay estmaton and locaton calculaton n dfferent methods. Crack sgnal Tme delay (s) Source locaton (mm) CC 1e EMD-CC 1.8e EEMD-CC e EEMD-FastICA-CC 1.8e Compared to Other Algorthms To further verfy the feasblty of the proposed algorthm, the results compared to the other algorthms such as CC, EMD-CC, EEMD-CC are as follow. Takng the crack propagaton stage for example, seen from Fg. 13 and Table. 3, four dfferent sgnal processng algorthms are correspondng to the four dfferent source locaton results. And compared to the tradtonal algorthm, the proposed EEMD-FastICA CC algorthm spends less tme delay and obtans much hgher postonng accuracy. 5. Conclusons In ths study, focusng on the characterstcs of AE sgnals Non-Statonary and tme delay estmaton accuracy under the uncertan correlated noses nterference condton, the EEMD-Fast-ICA-CC acoustc emsson source locaton method s proposed n ths paper. At frst, Fast-ICA algorthm s adopted n blnd source separaton of AE source sgnal by usng the EEMD sgnal processng algorthm, and the man feature IMF sgnals are extracted as the effectve acoustc emsson source locaton sgnal. On ths bass, tme delay s obtaned at dfferent channels n combnaton wth the cross-correlaton tme delay estmaton prncple, and the EEMD-Fast-ICA-CC source locaton algorthm s proposed, reducng the nterference noses caused by AE testng and mprovng the postonng accuracy. Then, the effectveness of the proposed algorthm s verfed by conductng the

9 A Novel Acoustc Emsson Source Locaton Method n the Crane based on EEMD-FastICA 25 three-pont bendng loadng experment n Q235B steel plate. Fnally, compared to the tradtonal algorthms such as CC, EMD-CC, EEMD-CC, the locaton accuracy of the proposed EEMD-Fast-CA-CC algorthm obtans an obvous mprovement. Acknowledgments Ths work s supported by the Natonal Natural Scence Foundaton of Chna ( , ), the funds of Hgh-level talents of JIT (JIT , FHXM201608). References [1] Rndorf, H Acoustc Emsson Source Locaton n Theory and n Practce. Bruel and Kaer Techncal Revew 2: [2] Jang, Y., Xu, F., and Xu, B Acoustc Emsson Tomography Based on Smultaneous Algebrac Reconstructon Technque to Vsualze Damage Source Locaton n Q235B Steel Plate. Mechancal System and Sgnal Processng 65: [3] Norden, H. E., and Zhaohua, W Ensemble Emprcal Mode Decomposton: A Nose-asssted Data Analyss Method. World Scentfc 1: [4] Zhaohua, W., and Norden, H. E The Mult-dmensonal Ensemble Emprcal Mode Decomposton Method. Advances n Adaptve Data Analyss 1: [5] Jan, T., and Zhao, L. J Vbraton Analyss Based on Emprcal Mode Decomposton and Partal Least Square. Proceda Engneerng 16: [6] George, G., Loutas, T., and Stylos, C Bearng Fault Detecton Based on Hybrd Ensemble Detector and Emprcal Mode Decomposton. Mechancal Systems and Sgnal Processng 16: 1-9. [7] Zvokel, M Multvarate and Multscale Montorng of Large-Sze Low-Speed Bearngs Usng Ensemble Emprcal Mode Decomposton Method Combned wth Prncpal Component Analyss. Mechancal System and Sgnal Processng 24: [8] Jn, Z Acoustc Emsson Locatzaton Technque Based on Generalzed Cross Correlaton Tme Dfference Estmston Algorthm. Chnese Journal of Sensors and Actuators 26 (11): [9] Ryan, D., and James, F. K The Use of a Maskng Sgnal to Improve Emprcal Mode Decomposton. Internatonal Conference on Acoustcs, Speech and Sgnal Processng 1: [10] Tang, B. P., and Dong, S. J Method for Elmnatng Mode Mxng of Emprcal Mode Decomposton Based on the Revsed Blnd Source Separaton. Sgnal Processng 92: [11] Muñoz, C. G., and Márquez, F. G A New Fault Locaton Approach for Acoustc Emsson Technques n Wnd Turbnes. Energy & Fuels 9 (1): 40. [12] L, X., and Adal, T Noncrcular Complex ICA by Generalzed Householder Reflectons. IEEE Trans. Sgnal Process. 61 (24): [13] Dermoune, A., and We, T FastICA Algorthm: Fve Crtera for the Optmal Choce of the Nonlnearty Functon. Sgnal Processng IEEE Transactons on 61 (8): [14] Fu, X., Sdropoulos, N., Trante, J., and Ma, W A Factor Analyss Framework for Power Spectra Separaton and Multple Emtter Localzaton. Sgnal Processng IEEE Transactons on 63 (24):