Cooling Fan Bearing Fault Identification Uing Vibration Meaurement Qiang Miao * School of Mechanical, Electronic and Indutrial Engineering Univerity of Electronic Science and Technology of China Chengdu, Sichuan 673, China mqiang@uetc.edu.cn Michael Azarian, Michael Pecht Center for Advanced Life Cycle Engineering (CALCE) Univerity of Maryland College Park, MD 74, USA mazarian@calce.umd.edu pecht@calce.umd.edu Abtract A a commonly ued aembly in computer cooling ytem, the normal operation of a cooling fan i critical for guaranteeing ytem tability and reducing damage to electronic component. Reliability analye have hown that fan bearing failure i a maor failure mode. Therefore, it i neceary to conduct reearch on fault detection of cooling fan bearing. In thi paper we propoe vibration-baed fan bearing fault detection through the wavelet tranform and the Hilbert tranform. An experiment on fan bearing wa conducted to collect vibration data for the validation of our propoed method. The analyi reult how that the propoed method can identify different bearing fault. Keyword cooling fan bearing; fault identification; dicrete wavelet tranform; Hilbert tranform I. INTRODUCTION Nowaday, computer are ued in many different area, uch a telecommunication, education, manufacturing, marketing, health care, etc. Computer ytem failure may bring inconvenience in our daily live, or even caue evere economic loe and catatrophic accident under certain condition. The requirement of high operational reliability ha driven the reearch on diagnoi and failure analyi of computer ytem. However, it ha been a challenging tak due to the complicated interaction of ytem performance parameter and application environment (e.g., temperature, moiture, and vibration) and their effect on ytem degradation and failure []. A a commonly ued aembly in a computer cooling ytem, the mechanical part of a cooling fan include bearing, haft, fan blade, and fan houing. A fan i ued to move heated air away from the component in the cae. According to [], fan failure i a maor problem for many electronic ytem. Bearing failure i the top contributor to fan failure. The normal operation of a cooling fan impact a computer ytem by preventing intability, malfunction, and damage to electronic component caued by overheating. Therefore, it i neceary to conduct reearch on cooling fan bearing fault detection o a to guarantee the normal operation of a fan. The maor type of bearing ued in cooling fan i ball bearing, mainly becaue it ha a longer lifepan at higher temperature (63, hour at o C) compared to leeve bearing (4, hour at o C) [3]. Ball bearing are the fundamental rotating part in mechanical ytem, and much reearch ha been conducted in bearing fault diagnoi [4-8]. However, the literature on computer cooling fan bearing reliability i very limited [, 9, ]. It i a challenge to identify ball bearing fault ignature baed on vibration ignal becaue the bearing component, including inner race, outer race, cage, and roller, complicate the bearing vibration ignal. When a local fault exit in a ball bearing, the urface i locally affected and the vibration ignal exhibit modulation []. Therefore, it i neceary to implement filtering and demodulation o a to obtain faultenitive feature from the raw ignal. At preent, the Hilbert tranform ha been widely ued a a demodulation method in vibration-baed fault diagnoi [, 3]. It ha a quick algorithm and can extract the envelope of the vibration ignal. In addition, the wavelet analyi i able to decompoe a ignal into different cale correponding to different frequency bandwidth [, 4, ], which can be treated a band-pa filter. The purpoe of thi tudy wa to invetigate fan bearing fault identification method uing vibration meaurement that can be ued for cooling fan degradation aement and prognotic. A tet rig with a cooling fan wa etablihed, with no lubricant in the ball bearing o a to accelerate the experiment. To identify the type of bearing failure from the vibration meaurement at the end of experiment, a new method wa propoed baed on the wavelet and the Hilbert tranform. Thi paper i organized a follow. In Section, a brief introduction to the wavelet and Hilbert tranform i given. Section 3 preent our propoed method for fan bearing fault identification uing the vibration ignal. A cae tudy on fan bearing i preented in Section 4, including a decription of the experiment and validation of the propoed method. Our concluion are ummarized in Section. II. THEORETICAL BACKGROUND A. Wavelet Tranform The wavelet tranform i the time-frequency decompoition of a ignal into a et of wavelet bai function. It poee flexibility in both the time and frequency domain, and it ha * Correponding author. Email: mqiang@uetc.edu.cn. Phone: +86-8- 683-669. 978--444-986-//$6. IEEE
been widely ued in machinery fault diagnoi. The continuou wavelet tranform (CWT) of a finite energy time domain ignal x ( with the wavelet ψ ( i defined a [4]: + t b W ( a, = ψ ( ) dt () a a where denote the complex conugation, a R ; b R ; and a and b are the cale and tranlation parameter, repectively. A een in (), the wavelet coefficient W ( a, i defined on the a b plane. A mall cale parameter a correpond to a high-frequency component, and the tranlation parameter b repreent the location of the wavelet bai in the time domain. W ( a, i a meaure of imilarity between the ignal x ( and the wavelet ψ ( at different frequencie determined by the cale parameter a and different time location determined by the tranlation parameter b. Although the CWT offer the poibility of a detailed analyi of tranient with an arbitrary fine frequency cale, it i not computationally efficient and it reult in high redundancy. The dicrete wavelet tranform (DWT) i derived from the CWT through dicretization of the wavelet, and the mot common dicretization of the wavelet i baed on a power of (i.e., dyadic cale and poition [6]). That i, a =, b = k,, k Z () Therefore, the dicrete wavelet function and caling function can be defined a follow:, k k ψ ( = ψ ( t ) (3), k k φ ( = φ( t ) (4) Mallat [6] propoed a fat wavelet decompoition and recontruction method, which i a claical ignal proceing cheme, known a a two-channel ub-band coding. In the decompoition proce, the ignal i convolved with a low-pa filter and a high-pa filter, repectively. It produce two piece of decompoed ignal, namely, the approximation ignal and the detail ignal. For example, let x ( t ) = A ( t ) ; the decompoition proce can be repeated a follow: A ( = A ( + D ( () where A ( and D ( are the approximation ignal and detail ignal at the th decompoition level, repectively. In recontruction, a pair of low-pa and high-pa recontruction filter are convolved with A ( and (, repectively. The decompoition of a ignal uing dyadic orthogonal wavelet i a quadratic ub-band filtering. Suppoe a ignal i collected at a ampling frequency F. The information obtained by DWT on each cale correpond to a frequency bandwidth F +. The frequency band of the approximation A i expreed a f [, A F And the frequency band of the detail D + ] D i + (6) f D + [ F, F ] (7) Thi mean that the approximation and detail are the narrowbanded ub-ignal of the original ignal. B. Hilbert Tranform From a ignal proceing perpective, the Hilbert tranform can be interpreted a a filtering operation in which the amplitude of the frequency component i unchanged, while the phae i hifted by 9 o. It i a time-domain convolution that map one real-valued time-hitory into another. Given a time domain ignal x (, it Hilbert tranform H [ i defined a: + τ ) H[ = dτ (8) π t τ where t and τ are the time and tranlation parameter, repectively. In machinery fault detection, modulation caued by local fault i inevitable in collected ignal. In order to identify fault-related ignature, demodulation i a neceary tep, and it can be accomplihed by forming a complex-valued timedomain analytic ignal A [ with x ( and H [. That i, iϕ( A[ = + ih[ = a( e (9) H[ where a ( = x ( + H [, ϕ ( t ) = arctan, and i = ; a ( i the envelope of A [, which repreent an etimate of the modulation in the ignal. III. THE PROPOSED METHOD FOR FAN BEARING FAULT IDENTIFICATION A. Fan Bearing Failure Behavior The cooling fan bearing tudied in thi paper are typical of thoe ued in many computer. Therefore, it i neceary to undertand it failure behavior in order to conduct reearch on fan bearing fault detection. Typical failure of ball bearing include local fault on the inner race, outer race, cage, and roller. If there i a local fault on a certain part of a ball bearing, the correponding fault-related characteritic frequency and it harmonic can be identified through pectral analyi in the frequency domain [7]. The formulae for the variou characteritic frequencie are a follow: Ball pin frequency (BSF): D d BSF = = [ ( co β ) ] () d D Ball pa frequency, inner race (BPFI): n d BPFI = I = ( + co β ) () D Ball pa frequency, outer race (BPFO): n d BPFO = O = ( co β ) () D Fundamental train frequency (FTF): d FTF = C = ( co β ) (3) D
Here, i the rotating peed of the bearing haft (Hz), n i the number oolling element, d i the mean diameter of the rolling element (mm), D i the pitch diameter of the bearing (mm), and β i the contact angle ( o ). B. The Propoed Method The function of fan bearing i to reduce friction and allow a fan to operate at high peed with lower noie. In cae where noie can be heard, eriou fault may have developed on different part of the bearing, complicating the vibration meaurement. In addition, fan bearing ued in computer cooling fan are mall ( 8 mm in pitch diameter), which make it almot impoible to place everal enor on different poition around the bearing for data collection. The DWT i a quadratic ub-band filtering technique that can decompoe the original ignal into different band, and the Hilbert tranform provide a mean of ignal demodulation. Therefore, thee two technique are ued in fan bearing vibration analyi for the identification of fault. A flow chart of the fan bearing fault identification method baed on the wavelet and the Hilbert tranform i hown in Fig.. Hilbert tranform. To identify the exitence of characteritic frequency component (uch a BSF, BPFI, BPFO, FTF) of the bearing, perform the pectrum analyi of a ( by + πfti ES ( f ) = a ( e dt () Here, ES ( f ) denote the abolute value of the Fourier tranform amplitude of the th envelope a (. IV. CASE STUDY A. Decription of Experiment The normal lifepan of a cooling fan can be everal year, and it i unrealitic to conduct an experiment for uch a long time. For a bearing to have it nominal lifepan at it nominal maximum load, lubrication ha a critical impact on the lifepan of the bearing. Therefore, it i reaonable to accelerate the fan bearing experiment through the reduction of the lubrication level. Auming that the nominal amount of lubricant i at the % level, then a certain lubrication level p% decribe the percentage of lubricant being added in the bearing. In thi tudy, an experiment wa conducted to validate the propoed method for fan bearing fault identification. To accelerate the experiment, the lubrication level wa %, which mean that no lubricant wa in the bearing. Figure how the cooling fan with bearing teted in thi experiment. Figure. Flow chart of the propoed method. Given a piece of ignal x (, a pre-proceing tep i neceary to reduce the trend and the DC component in the ignal. That i, x y( = (4) σ where y ( i the pre-proceed ignal, x i the mean value of x (, and σ i the tandard deviation of x (. In order to obtain ub-ignal correponding to different frequency band, the DWT with the Daubechie wavelet i utilized, and the decompoition level i J. Therefore, a erie of detail ignal D ( with =,,..., J can be obtained, and the frequency range of D ( i decribed in (7). When a local fault exit in the bearing, there i modulation in the ignal. To reduce the impact of modulation, the Hilbert tranform wa performed on all of the detail ignal uing (8) and (9), and the correponding analytical ignal and their envelope, a ( ( =,,..., J ), can be obtained. Here, a ( i the envelope ignal of the th detail ignal D ( after the Figure. The cooling fan with bearing being teted in thi experiment. The pecification of the fan bearing in our experiment are given in Table. The cooling fan ued in thi experiment wa in it brand new tate with ungreaed ball bearing before the experiment. At the beginning, the vibration ignal of the fan wa collected at a ampling frequency of.4 khz with - econd period under an ambient environment. The fan peed wa 4, rpm. After that, the cooling fan wa treed in a chamber at a high temperature (7 o C). After 7 hour of running at it maximum peed of 48 rpm, the vibration ignal of the fan wa collected at a ampling frequency of.4 khz with -econd period under an ambient environment. The fan peed wa 4, rpm. Therefore, the rotation peed of the bearing during data collection wa = 66.67 Hz. Uing () (3), the variou fault characteritic frequencie were determined, a lited in Table.
TABLE I. FAN BEARING SPECIFICATIONS Number oolling element n 6 Diameter oolling element d Pitch diameter of bearing D Contact angle β TABLE II..9 mm.mm.4 o CHARACTERISTIC FREQUENCIES OF FAN BEARING Ball pin frequency ( ).96 Hz Ball pa frequency, inner race ( I ) 6.87 Hz Ball pa frequency, outer race ( O ) 43.3 Hz Fundamental train frequency ( C ) 3.8 Hz B. Experimental Reult To validate the propoed method, the vibration ignal collected before and after the 7-hour treing experiment were ued to conduct the following analyi. The Daubechie wavelet db wa ued for ignal decompoition; the decompoition level wa J =6. The Hilbert tranform wa applied to demodulate the detail D ( with =,,..., 6. x - ()Spectral of D ignal x -6 ()Spectral of D ignal x -6 (3)Spectral of D3 ignal 3 8 Figure 3. Analyi reult of the fan bearing before the treing experiment. 6 8 6 x -6 ()Spectral of D ignal x - (4)Spectral of D4 ignal x -4 (6)Spectral of D6 ignal 3 8 The analyi reult of the fan bearing before the treing experiment are hown in Fig. 3. Since the ball bearing wa in a brand new tate at the beginning, no fault-related frequency component were found from the pectra of the detail ignal. Fig. 4 how the analyi reult of the fan bearing after the 7-hour treing experiment. From the pectra of detail ignal D (,..., D6 (, many fault-related characteritic frequencie and their harmonic were identified, which indicated that there were everal local fault at different part of the fan bearing. In fact, noie coming from the fan wa heard at the end of the experiment. C x -3 3 C. ()Spectral of D ignal x -4 ()Spectral of D ignal x -4 (3)Spectral of D3 ignal x -4 Figure 4. Analyi reult of the fan bearing after 7-hour treing experiment. V. CONCLUSIONS Cooling fan bearing failure i a maor failure mode in computer cooling ytem. Cooling fan bearing failure can reult in ytem intability, malfunction, and damage to the computer ytem. In thi paper, a fan bearing fault identification method baed on the dicrete wavelet tranform and the Hilbert tranform i propoed. In order to validate the propoed method, an experiment with ungreaed fan bearing wa conducted to obtain the vibration ignal before and after failure. The analyi reult howed that the propoed method O O (4)Spectral of D4 ignal x -4 ()Spectral of D ignal x -3 (6)Spectral of D6 ignal. 4 fro 3
can identify the fault-related characteritic frequencie of the fan bearing. The work decribed in thi paper provide a promiing way to etablih potential metric for the decription of bearing health degradation. Therefore, it i deirable to develop a condition monitoring ytem baed on the propoed method and realize on-line health evaluation of cooling fan. With uch a function, the critical failure of cooling ytem can be avoided, and the reliability and availability of electronic ytem can be guaranteed. ACKNOWLEDGMENT Thi reearch wa partially upported by National Natural Science Foundation of China (Grant No. 98), Fundamental Reearch Fund for the Central Univeritie (Grant No. ZYGX9X). We would like to thank Mr. Hyuneok Oh at CALCE PHM Group, Univerity of Maryland, for hi work on the experiment and dicuion on the improvement of thi reearch. REFERENCES [] M. Pecht, Prognotic and Health Management of Electronic. London: Wiley-Intercience, 8. [] X. Tian, Cooling fan reliability: failure criteria, accelerated life teting, modeling and qualification, Proceeding of 6 Reliability and Maintainability Sympoium, 6, pp. 38-384. [3] M. William, Ball v. leeve: a comparion in bearing performance, Technical Paper from NMB Technologie Corp. [4] N. Tandon, and A. Choudhury, A review of vibration and acoutic meaurement method for the detection of defect in rolling element bearing, Tribology International, Vol.3, No.8, 999, pp. 469-48. [] W. He, Z.N. Jiang, and K. Feng, Bearing fault detection baed on optimal wavelet filter and pare code hrinkage, Meaurement, Vol.4, No.7, 9,pp. 9-. [6] V.K. Rai, and A.R. Mohanty, Bearing fault diagnoi uing FFT of intrinic mode function in Hilbert-Huang tranform, Mechanical Sytem and Signal Proceing, Vol., No.6, 7, pp. 67-6. [7] H. Qiu, J. Lee, J. Lin, and G. Yu, Wavelet filter-baed weak ignature detection method and it application on rolling element bearing prognotic, Journal of Sound and Vibration, Vol.89, No.4-, 6, pp. 66-9. [8] R.B. Randall, and J. Antoni, Rolling element bearing diagnotic A tutorial, Mechanical Sytem and Signal Proceing, Vol., No.,, pp. 48-. [9] H. Oh, M. H. Azarian, M. Pecht, C. H. White, R. C. Sohaney, and E. Rhem, Phyic-of-failure approach for fan PHM in electronic application, Proceeding of the IEEE Prognotic and Sytem Health Management Conference,, pp. -6. [] H. Oh, T. Shibutani, and M. Pecht, Precuror monitoring approach for reliability aement of cooling fan, Journal of Intelligent Manufacturing, 9, DOI.7/84-9-34-. [] J. Antoni, and R.B. Randall, Differential diagnoi of gear and bearing fault, Journal of Vibration and Acoutic, Vol.4, No.,, pp. 6-7. [] D. Wang, Q. Miao, X. Fan, and H.Z. Huang, Rolling element bearing fault detection uing an improved combination of Hilbert and Wavelet tranform. Journal of Mechanical Science and Technology, Vol.3, No., 9, pp. 39-33. [3] Y. Qin, S. Qin, and Y. Mao, Reearch on iterated Hilbert tranform and it application in mechanical fault diagnoi, Mechanical Sytem and Signal Proceing, Vol., No.8, 8, pp. 967-98. [4] D. Wang, Q. Miao, and R. Kang, Robut health evaluation of gearbox ubect to tooth failure with wavelet decompoition, Journal of Sound and Vibration, Vol.34, No.3-, 9, pp. 4-7. [] G. Niu, A. Widodo, J.D. Son, B.S. Yang, D.H. Hwang, and D.S. Kang, Deciion-level fuion baed on wavelet decompoition for induction motor fault diagnoi uing tranient current ignal, Expert Sytem with Application, Vol.3, No.3, 8, pp. 98-98. [6] S.G. Mallat, A theory for multireolution ignal decompoition: the wavelet repreentation, IEEE Tranaction on Pattern Analyi and Machine Intelligence, Vol., No.7, 989, pp. 674-693. [7] Q. Miao, D. Wang, and H.Z. Huang, Identification of characteritic component in frequency domain from ignal ingularitie, Review of Scientific Intrument, Vol.8, No.3,, 33.