Engineering, 2010, 2, 60-64 doi:10.4236/eng.2010.21008 Published Online January 2010 (http://www.scirp.org/journal/eng/). Research on Early Fault Sel-Recovery Monitoring o Aero-Engine Rotor Syste Abstract Zhongsheng WANG, Shiwe MA School o Aeronautics Northwestern Polytechnical University Xi an, China Eail: sazs_wang@nwpu.edu.cn Received August 19, 2009; revised Septeber 14, 2009; accepted Septeber 20, 2009 In order to increase robustness o the AERS (Aero-engine Rotor Syste) and to solve the proble o lacking ault saples in ault diagnosis and the diiculty in identiying early weak ault, we proposed a new ethod that it not only can identiy the early ault o AERS but also it can do sel-recovery onitoring o ault. Our ethod is based on the analysis o the early ault eatures on AERS, and it cobined the SVM (Support Vector Machine) with the stochastic resonance theory and the wavelet packet decoposition and ault sel-recovery. First, we zoo the early ault eature signals by using the stochastic resonance theory. Second, we extract the eature vectors o early ault using the ulti-resolution analysis o the wavelet packet. Third, we input the eature vectors to a ault classiier, which can be used to identiy the early ault o AERS and carry out sel-recovery onitoring o ault. In this paper, eatures o early ault on AERS, the zoo o early ault characteristics, the extraction ethod o early ault characteristics, the construction o ulti-ault classiier and way o ault sel-recovery onitoring are studied. Results show that our ethod can eectively identiy the early ault o AERS, especially or identiying o ault with sall saples, and it can carry on sel-recovery onitoring o ault. Keywords: AERS, Early Fault, Support Vector Machine, Classiication Identiication o Fault, Sel-Recovery Monitoring o Fault 1. Introduction With the developent o the odern aviation industry, the saety o aircrat and reliability are ore and ore attracted. The engine is heart o aircrat and AERS is central part o engine. I AERS has ault in light, the aircrat would be severely threatened in saety. Because the structure o AERS is coplex, load is bigger and the change o operation conditions is requency, it will be the ore diicult in the ault identiication. Especially in the early ailures occurred, i we can catch tiely the ault inoration, and can eectively identiy it and carry on sel-recovery onitoring o ault, it will have iportant signiicance to eradicate or eliinate the potential ault caused o an accident. Fault identiication on AERS is widely studied and any results are obtained [1 3]. At present now, eective identiication o early ault has ore diiculty. In particular, obtaining o the early ault inoration is ore diiculty, and eectiveness o ault *Project (50675178, 60472116) Supported by National Natural Science Foundation o China identiication is not satisactory. Because operation o AERS is in strong noise density, the early ault inoration is very weak and signals are easily looded in noises, satisactory results are very diicult by general ethods. Our ethod can not only ast identiy the early ault o the AERS but also carry on sel-recovery onitory o ault. 2. Features o Early Fault on AERS In the light, the wear, deoration, corrosion and racture o structure coponents and eects o work stress, external environent and huan actors will lead to aults o AERS. The early ault o AERS oten shows a or in icro-cracks, icro-creeping, icro-corrosion and icro-wear. These aults, with the exception o little sudden aults, the ajority have a developent process ro nothing to ault, inor to general, evolution to ast. In this process, the structure o syste, properties and internal energy will change. We can onitor the early aults through catching these ault inoration in tie,
Z. S. WANG ET AL. 61 and it can be identiied and be sel-recovery. These early aults o AERS have the ollowing eatures: 1) Fault signal is very weak. When ault is at early stage or just sprouted, the changes o the ault signal is very weak in aplitude, phase and tie-requency characteristics, and the little ault characterization is oten diicult to detect. 2) Fault signal will be drowned by noise signal. During the light, the noise signal is usually aong the ault signal. When the ault signal is very weak and the noise signal is very strong, the early ault signal will be drowned by noise signal. In order to detect the early ault, we ust reduce noise or extracted early ault inoration ro the noise signal. 3) Fault signal is oten a transient one. The daage structure coponents o aircrat is generated by ipact loading, and this ault is aniested by transient signal, such as expansion o early cracks, is a process ro gradual to utation. 4) Fault occurred in the area o stress concentration. When area o stress concentration is acted by strong regional load, the ault o structural part is easy occurrence in the creeping. For the erroagnetic etal part, the agnetic eory ethod can be used or the early ault detection and localization [4]. 5) Fault has volatile. That eans soeties we can not ind the ault trace in ault condition. But with the change o action or tie, the syste can recover autoatically. Volatile ault can be shown by eventuality ault, transient ault and interruption ault. 3. Extraction o Early Fault Characreristic Because the vibration o engine is larger in operation, ault characteristics o the AERS will be suberged in the strong background noise in the early. In order to extract characteristics o early ault ro being suberged signals by noise, we use the stochastic resonance theory to zoo characteristic signals o early ault [5]. In the ailure, the energy o AERS will change in all requency bands and dierent aults have dierent eects to the signal energy in each requency band. So, we can use wavelet packet to decopose the output signal in stochastic resonance and select signal energy in the characteristics requency band as a eature vector. I the data length o original signal x t is N, the data length o discrete signals x k, i is reduced to 2 N by decoposition o wavelet packet and its energy can be expressed as: E n 2 N k, k, x x i 2 i 1 1 2 (1) N 1 where, N = the length o original data, k = layer nuber by wavelet packet decoposition, = the serial nuber o decoposition requency band location = 0, 1, k 2,... 2 1. The energy within the each requency band can be calculated by Equation (1) and the eature vector can be constructed by the energy. 4. Classiication Identiication o Fault The SVM can do the classiication very well in the ew nubers o ault saples and it can solve the identiication probles on nonlinear and high-diensional pattern. The ault identiication o the AERS belongs to proble o ulti-classiication and it needs to construct ultiault classiier. In this paper, we adopt iproved classiication algorith o the one-to-any. To the classiication o K type, way o one-to-any is need to construct K two classiiers. In this way constructing every two classiiers, all n training saples o K type should be operated. In the testing and classiication, the scale o classiiers is larger and speed is slower. An iproved the one-to-any classiication algorith is to construct two classiiers o K, and the training saple o type in the K classiiers is y i y i 1 and other types arking is 1. Established output unction in M classiiers can be expressed as: i i i b (2) x sgn y K x x SVM The algorith overcoes deects that one-on-one approach needs to establish any nuber classiiers and it can control the training saples nuber in traditional one-to-any. Its operate speed is ast and eect o ault classiication is well. 5. Sel-Recovery Monitoring o Fault Fault onitor process o AERS is shown in Figure 1. Fault sel-recovery database consist o any intelligent odels. It can carry on sel-recovery onitoring based on dierent ault sources and ault characteristics. Sel-recovery onitoring o ault can be realized by sart structure [6], aixing agnetic ield [7], regenerating aterials [8], etc. When AERS has ault, the syste can not work norally. Sel-recovery odel based on the ault copensation can restore original unction o syste by eans o the ault sel-recovery copensator. I syste equations under noral condition are x ( t ) Ax ( t ) Bu ( t ) (3) y( Cx( (4) Copensator equations are
62 Z. S. WANG ET AL. AERS Fault Detection Fault Identiication Database Classiication Identiication Fault Sel-recovery Database Can Do Sel-recovery? y Sel-recovery Repair n Module Replace Parts Repair Figure 1. Fault sel-recovery onitor process o AERS. z ( Ey( (5) u( Fz( Hy( (6) The syste loop equations depicted in (3) and (4) are x ( ( A BHC) x( BFz( (7) z ( ECx( (8) When ault o coponent or sub-syste occurred during the light, the loop eedback equations are x ( ( A B HC ) x( B Fz( (9) z ( EC x( (10) To ake the perorance o ault syste as close as possible to the perorance o the original syste, we can design the appropriate sel-recovery copensator D, E, F, H to achieve the ault sel-recovery copensation. During light, when the aircrat cockpit, wings and other iportant coponents have severe vibration or chatter, distributed piezoelectric driver copensator can weaken or oset the ipact o vibration by vibration control and active vibration absorber. Sel-recovery onitoring based on intelligent structure is shown in Figure 2. Controller Sel-recovery AERS Detect odule Sel-recovery odel Figure 2. Sel-recovery onitoring based on sart structure. When the sensor odule detects the ault inoration, the signals will transit to sel-recovery odule and act to the controller. Fault o AERS will be restored by sart structure. It is based on dierent ault sources and ault eature, and the ault sel-recovery tactics is adopted. 6. Experiental Results and Its Analysis In order to veriy the useulness o the ethod, we choose the our conditions o the AERS. They are noral condition, early rotor isalignent, early rotor unbalance and early rotor crack. These signals are preprocessed and the ault eatures are extracted. Fault is identiied by the ulti-ault classiiers and ault is onitored by sel-recovery odule. In the experient, according to the character o AERS, we collected 10 group data by the acceleration sensor. They respectively correspond to the over our conditions in the 1800 rp. The sapling requency is 256 Hz and the rotation requency is 30 Hz. The characteristic requency o ault and its concoitant requency are shown in Table 1. Figure 3 shows the output waveor o the stochastic resonance syste and its spectru. The requency coponent in 30 Hz is obvious in Figure 2. Because the undaental requency is 30 Hz, the rotor isalignent ault is oten accopanied by 1x (30 Hz), 2 x (60 Hz) and 3x (90 Hz). Table 1. 2x (60Hz) crossed concoitant requency and unbalance. Type Misalignent Unbalance Crack Characteristic requency Concoitant requency 1x (30Hz), 2 x (60Hz) 1x (30Hz) 2x (60Hz) 3 x (90Hz) 4x (120Hz)
Z. S. WANG ET AL. 63 (a) Waveor beore onitoring Figure 3. The output waveor o stochastic resonance syste and its spectru. (b) Waveor ater onitoring Figure 5. Frequency doain waveor o rotor isalignent. Figure 4. Energy distribution o rotor isalignent. Table 2. Test results o stochastic resonance syste. Fault type Misjudgent saples Diagnosis saples Diagnosis rate Misalignent 1 40 97.5 % Unbalance 0 40 100 % Crack 2 40 95 % The energy eature extraction is done by the wavelet packet analysis. It is based on the signal waveor o stochastic resonance syste. Seven-layer wavelet packet is resolved in the db3 wavelet and we obtained 64 requency bands. In order to reduce the aount o coputation, we divide the requency bands as 6 segents: 0 ~ 0.4X, 0.4 ~ 0.8X, 0.8 ~ 1.2 X, 1.8 ~ 2.2X, 2.8 ~ 3.2X and greater than 3.6X. Thus 64 requency bands will be coposed o the 6 groups: 1 ~ 4 bands, 5 ~ 8 bands, 9 ~ 11 bands, 12 ~ 15 bands, 16 ~ 30 bands and 55 ~ 64 bands. Energy value o each group is added together and they are processed on the noralization. The energy distribution o rotor isalignent ault will be acquired. It is shown in Figure 4. Repeated the above process, the energy distribution o each condition can be obtained and it is taken as training saples o SVM. According to the training saples obtained ro our conditions, we take 40 groups energy distributions ro each condition as training saples and input to the iproved ault classiier o one-to-any. We choose Gaussian RBF kernel unction as a classiication unction and ake the paraeters 0. 01, punishent actor C = 100. The classiication results are shown in Table 2. We can see that classiication results by the stochastic resonance syste are signiicantly high than classiication result by direct wavelet packet eature extraction. The classiication tie o the each testing saples is saller (in 0.05s, 1.8 GHZ coputers). Its accuracy is higher and speed is quick in early ault identiication. In order to onitor isalignent, we adopt a principle o the electroagnetic eect [9]. The nuber o isalignent is detected by our acceleration sensors, and our electroagnetic sets are controlled by the output signal o our sensor. When isalignent occurred, the isalignent orce F by rotor produced can be adjusted by alignent orce F o electroagnet produced. The alignent orce F is equal to the isalignent orce F in nuber and they are contrary in direction. Figure 5 shows the result o sel-recovery onitoring on rotor isalignent.
64 Z. S. WANG ET AL. We can ind out that vibration o rotor is obviously reduced in Figure 5(a) and Figure 5(b). Because o the isalignent ault o rotor is counteracted by the electroagnet orce, the rotor isalignent is inhibited and the noral operation condition is restored well. 7. Conclusions 1) Our ethod can eectively extract the early ault eature o the AERS by cobination the stochastic resonance with the wavelet packet resolving. Energy eigenvectors o constructed by this ethods can accurately relect the condition changes o AERS. 2) Multi-ault classiier based on the SVM has characteristics that its algorith is sipler, the classiication eect is well and identiication eiciency is higher. It particularly suits to the classiication identiication o sall saple and sel-recovery onitoring o early ault on AERS. 3) The ault diagnosis is aied at inding ailure in tie and to ensure sae operation o plant. The ethod o ault sel-recovery onitoring provides an eective way. 8. Reerences [1] D. Sion, A coparison o iltering approaches or aircrat engine health estiation, Aerospace Science and Technology, Vol. 12, No. 4, pp. 276 284, 2008. [2] S. Borguet and O. Léonard, Coupling principal coponent analysis and Kalan iltering algoriths or on-line aircrat engine diagnostics, Control Engineering Practice, Vol. 17, No. 4, pp. 494 502, 2009. [3] T. Raesh Babu and A. S. Sekhar, Detection o two cracks in a rotor-bearing syste using aplitude deviation curve, Journal o Sound and Vibration, Vol. 314, No. 3 5, pp. 457 464, 2008. [4] A. Dubov and S. Kolokoinikov, Review o welding probles and allied processes and their solution using the etal agnetic eory eect, Welding in the World, Vol. 49, No.9, pp. 306 313, 2005. [5] K. Tanaka and M. Kawakatsu, Stochastic resonance in auditory steady-state responses in a agnetoencephalogra, Clinical Neurophysiology, Vol. 119, No. 9, pp. 2104 2110, 2008. [6] S. Hurlebaus and L. Gaul, Sart structure dynaics, Mechanical Systes and Signal Processing, Vol. 20, No. 2, pp. 255 281, 2006. [7] A. Ignatios and B. Alexey, Anoaly induced eects in a agnetic ield, Nuclear Physics B, Vol. 793, No. 1 2, pp. 246 259, 2008. [8] Mueller and S. N. Sokolova, Characteristics o lightweight aggregate ro priary and recycled raw aterials, Construction and Building Materials, Vol. 22, No. 4, pp. 703 712, 2008. [9] Y. Asher, P. Yose, and L. Yuri, Spectral and variational principles o electroagnetic ield excitation in wave guides, Physics Letters A, Vol. 344, No. 1, pp. 18 28, 2005.