1

ARTICLE IN PRESS
Mechanical Systems and Signal Processing

Mechanical Systems and Signal Processing 21 (2007) 22122236 www.elsevier.com/locate/jnlabr/ymssp
Genetic fuzzy system for online structural health monitoring of composite helicopter rotor blades
Prashant M. Pawar, Ranjan Ganguli
Department of Aerospace Engineering, Indian Institute of Science, Bangalore 560012, India Received 13 January 2006; received in revised form 22 September 2006; accepted 25 September 2006 Available online 13 November 2006
Abstract A structural health monitoring (SHM) methodology is developed for composite rotor blades. An aeroelastic analysis of composite rotor blades based on the nite element method in space and time and with implanted matrix cracking and debonding/delamination damage is used to obtain measurable system parameters such as blade response, loads and strains. A rotor blade with a two-cell airfoil section and 0= 45=90s family of laminates is used for numerical simulations. The model based measurements are contaminated with noise to simulate real data. Genetic fuzzy systems (GFS) are developed for global online damage detection using displacement and force-based measurement deviations between damaged and undamaged conditions and for local online damage detection using strains. It is observed that the success rate of the GFS depends on number of measurements, type of measurements and training and testing noise level. The GFS work quite well with noisy data and is recommended for online SHM of composite helicopter rotor blades. r 2006 Elsevier Ltd. All rights reserved.
Keywords: Genetic fuzzy system; Online damage detection; Composite rotor blade; Matrix cracking; Debonding/delamination
1. Introduction Structural health monitoring (SHM) of helicopters has become an important area of research due to the need for reduction in operational accidents and maintenance costs. SHM also promises to help in reducing the maintenance cost which is about 24% of the direct operating cost of helicopters [13]. Due to evolution in onboard information processing technology and avionics, the comprehensive SHM systems are feasible and ight data recorder systems mitigate the cost of installing the SHM systems. Finally, SHM plays an important role in the case of aging helicopters (whose life span is increased due to budget restrictions) by promising to increase the safety and decrease in the maintenance cost. In general, SHM systems detect and interpret the adverse changes in a structure in order to improve reliability and to reduce the life cycle cost. The current study is focused on the development of a SHM methodology for composite rotor blades, which are routinely used in the modern helicopters [46]. For developing SHM for composite rotor blades, it is necessary to study the effect of various modes of damages on
Corresponding author. Fax: +91 080 334 5134.
E-mail address: ganguli@aero.iisc.ernet.in (R. Ganguli). 0888-3270/$ - see front matter r 2006 Elsevier Ltd. All rights reserved. doi:10.1016/j.ymssp.2006.09.006
ARTICLE IN PRESS
P.M. Pawar, R. Ganguli / Mechanical Systems and Signal Processing 21 (2007) 22122236 2213
Nomenclature A, B, D extension, extensionbending and bending stiffness matrix for virgin laminate Ac , Bc , Dc extension, extensionbending and bending stiffness matrix for laminate having matrix cracks c blade chord EA cross-sectional axial stiffness E L , E T Youngs modulus of plies in longitudinal and transverse directions EI y , EI z cross-sectional ap and lag bending stiffness GJ cross-sectional torsion stiffness G LT shear modulus of plies in principle plane K ij elements of the cross-sectional stiffness matrix 9 9 K 0ij elements of the cross-sectional stiffness matrix after static condensation 4 4 l length of blade M x , M y cross-sectional ap and lag bending moments Nk matrix of normal vectors for crack surface of ply k Qk plane stress stiffness matrix of ply k SR success rate sk average crack spacing of ply k tk thickness of ply k Ts cross-sectional twisting moment W cross-sectional longitudinal displacement DA, DB, DD residual extension, extensionbending and bending stiffness matrix due to the presence of matrix cracks DV , DW , Df peak-to-peak change in the tip lag, ap and torsion responses DF x , DF y , DF z peak-to-peak change in blade root longitudinal, lateral and vertical forces DM x , DM y , DM z peak-to-peak change blade root rolling, pitching and yawing moments dU , dT , dW virtual variations in strain energy, kinetic energy and virtual work mLT , mTL Poissions ratios of plies rk dimensionless crack density of ply k c azimuth fx , fy bending slopes of blade along lag and ap direction fz blade twist 0 denotes undamaged case
the composite rotor system properties. However, very limited work has been done on the modelling of damage in composite blades. Several researchers have investigated the modelling of a damaged helicopter rotor assuming isotropic structural properties [79]. Selected studies [1012] have investigated the inverse problem of nding the size, location and type of damage from measured system data which is contaminated with noise. Some studies have used modal data to detect damage in rotor blades [1114]. Along with analytical research on SHM mentioned above, selected studies have been reported on the experimental validation of rotor health monitoring methods [1517]. Most researchers have modelled damage in helicopter blades assuming isotropic materials. A rst step addressing damage modelling in composite rotor blades was taken by Lakshmanan and Pines [18] who modelled damage in the form of a transverse crack, which extends across the entire width of the exbeam of a bearingless rotor. However, there is a need for detailed modelling of the damage modes in a composite rotor blade. A beginning in this direction was made by Pawar and Ganguli [19] who modelled matrix cracking in the helicopter rotor blade with cross-ply laminates and studied its behaviour in forward ight. Pawar and Ganguli [20] further modelled the matrix cracking in a generalised layup box beams and an airfoil beam section and studied their behaviour at the static level. They also studied [21] the effect of more severe damage modes such
ARTICLE IN PRESS
2214 P.M. Pawar, R. Ganguli / Mechanical Systems and Signal Processing 21 (2007) 22122236
as debonding/delamination and bre breakage due to matrix cracking for static condition [21] and for forward ight [22]. Typically, the rst failure mode in the composite materials, which occurs in both monotonic and fatigue loading, is matrix cracking. Effects of matrix cracking must be studied carefully because matrix cracking may lead to severe damage modes like debonding/delamination. When the matrix cracking effect starts saturating, a local delamination can initiate from the matrix crack tips. Even though the effects of delamination are not very severe, it may lead to extreme failure damage like bre breakage. As a result of matrix cracking and debonding/delamination damage modes in composite materials, the matrix part of composite materials becomes weak and the proper transfer of shear loads at the broken bres may not take place. Therefore, the composite material acts as dry bundles of bres and leads to the extreme failure mechanism in composite materials, bre breakage. A detailed literature survey of all these key damage modes in composites is given in the authors previous paper [20]. SHM systems for composite structures are typically developed by solving an inverse problem using the changes in some measurable properties of the structure to detect damage. The inverse problem becomes complicated because of incomplete information and uncertainty in the modelling, measurements and signal processing. Therefore, the inverse problem for developing a SHM system can be solved using soft computing methods like neural network, genetic algorithm (GA) and fuzzy logic (FL) to extract precise conclusions from slightly imprecise data. Neural network is a powerful tool which forms a nonlinear mapping between input and output data sets using training data. Several researchers have recently used neural network methods for damage detection in composite structures. Garg et al. [23] used neural networks for monitoring various damage modes in composite structures using Fourier spectral data. Chakraborty [24] used neural networks for predicting delamination in composite plates using the natural frequencies. Yuan et al. [25] applied neural networks for online damage detection in thin-walled composite structures using wide-band Lamb wave measurements. Su and Ye [26] developed an online health monitoring system for in-service composite structures using PZT transducer network and neural networks. Another algorithm which has recently been used in damage detection is the GA. Genetic algorithms are search algorithms developed using the principles inspired by natural population genetics to evolve solutions to optimisation problems. Ramanujam et al. [27] used GA to minimise an error represented by the objective function relating the delamination parameters to the measurement parameters. Iwasaki and Todoroki [28] used GA to minimise an estimation error in the function generated by relating the damage modes and the measurement values. Rao et al. [29] used GA to minimise the global error derived from dynamic residual vectors, which are obtained by introducing simulated experimental data into the eigenproblem. Compared to the wide use of neural networks and GAs, very few works have used FL for damage detection. A fuzzy system (FS) is any FL-based system, which either uses FL as the basis for representing the different forms of knowledge or to model the interactions and relationships among the system variables. Fuzzy systems address uncertainty directly by using linguistic reasoning, which is more robust to uncertainty than pure numerical reasoning. Ramu and Johnson [30] used FL to handle uncertainties within the articial neural network strategy for making assessments of damage in composites. Ganguli [31] used FL to develop a groundbased health monitoring of a helicopter rotor blade. Much less use has been made of FL for structural damage detection, relative to the large amount of work on damage detection using neural networks possibly due to the unfamiliarity of structural engineers with the subject. Each of the methods discussed above have their advantages and shortcomings. Neural networks can be easily trained from inputoutput data but the training time can be quite large. Furthermore, neural networks have a black-box nature. Genetic algorithms can be too time consuming for use in an online setting. Fuzzy logic systems are difcult to develop as the rules need to be obtained from human experts or from inputoutput data. However, FL systems and neural networks are computationally efcient once the rule development and training phase has been completed. Hybridising the best features of two or more soft computing methods, a novel method can be formulated which is more accurate than the original ones. The genetic fuzzy system (GFS) is an advanced soft computing method developed by combining the approximate reasoning capabilities of FSs with the learning capabilities of GAs [32,33]. The rst GFS for structural damage detection was developed by Pawar and Ganguli [11].
ARTICLE IN PRESS
A GA was used for automating the process of rule generation for a FS for application to damage detection in beam-type structures and a helicopter rotor blade. The GFS was further demonstrated for the health monitoring of a thin walled composite beam [34]. However, these studies were limited to frequencies only. In the present study, the GFS is used for developing an online SHM system using blade response, loads and strains for composite helicopter rotor blades. The current study is focused on developing a GFS for online matrix cracking and debonding/delamination prediction in composite rotor blade which are the rst and second progressive damage mechanisms to occur in composites. Two possibilities of damage occurrence in the composite rotor blade are addressed using two GFSs. The rst GFS is developed for global damage detection using displacement, force and moment-based measurement deltas (deviations between the undamaged and damaged blade). The second GFS is developed for local damage detection using strain-based measurement deltas. The measurement deltas are obtained by a comprehensive aeroelastic analysis of a composite helicopter rotor blade. A progressive damage model, which includes the key damage modes in composite materials such as matrix cracking and debonding/delamination is used. The SHM system of a composite helicopter rotor blade is demonstrated using a two-cell airfoil sectionbeam with stiffness properties representing a stiff-inplane helicopter rotor blade with matrix cracking and debonding/delamination damages. 2. Analysis The analysis of the damaged composite rotor blade is explained in three parts. The rst part summarises the mathematical model of a helicopter rotor system. The second part discusses the composite rotor blade crosssection analysis. The composite cross-section properties are included in the mathematical model of the helicopter rotor system. The third part discusses the modelling of the key damage modes in composite materials. These key damage modes are included in the mathematical model of the helicopter rotor system through the composite cross-section properties. A detailed formulation of the forward ight simulation of damaged composite rotor blade is available in Pawar and Ganguli [22]. 2.1. Mathematical model of helicopter rotor system The helicopter is represented by a non-linear model of rotating elastic rotor blades dynamically coupled to a six-degree-of-freedom rigid fuselage. Each blade undergoes ap (out-of-plane) bending, lag (in-plane) bending, elastic twist and axial displacement. Governing equations are derived using a generalised Hamiltons principle applicable to non-conservative systems: Z c2 dU dT dW dc 0. (1)
c1
Here, dU and dT include energy contributions from components that are attached to the blade, e.g., pitch link, lag damper, etc. The effects of the damaged and undamaged composite materials are included to the virtual strain energy through the elastic stiffness matrix (discussed in the next section). External aerodynamic forces on the rotor blade contribute to the virtual work variational, dW . The aerodynamic forces and moments are calculated using an inow distribution from the BagaiLeishman free wake model and unsteady effects are accounted for using the LeishmanBeddoes model [35]. A nite element method is used to discretise the governing equation of motion and allows for the accurate representation of complex hub kinematics and nonuniform blade properties. After nite element discretisation, Hamiltons principle is written as Z cf X N dU i dT i dW i dc 0. (2)
ci i1
Each of the beam nite elements has 15 degrees of freedom (Fig. 1). These degrees of freedom correspond to cubic variations in axial elastic and (ap and lag) bending deections, and quadratic variation in elastic
ARTICLE IN PRESS
xi u1 v1 v1 w1 w1 1 li
Fig. 1. Beam nite element used for rotor blade structural modelling.
u33
u4
u2 v2 v2 w2 w2 2
torsion. Between the elements, there is continuity of slope and displacement for ap and lag bending deections and continuity of displacements for elastic twist and axial deections. An aeroelastic analysis is carried out at the specied trim condition. The nite element equations representing each rotor blade are transformed to normal mode space for an efcient solution of the blade response. The nonlinear, periodic, normal mode equations are then solved for steady response using a nite element in time method. Steady and vibratory components of the rotating frame blade loads (i.e. shear forces and bending/torsion moments) are calculated using the force summation method. In this approach, blade aerodynamic and inertia forces are integrated directly over the length of the blade. Fixed frame hub loads are calculated by summing the contributions of individual blades. A coupled trim procedure is carried out to solve for the blade response, pilot input trim controls, and vehicle orientation, simultaneously. The coupled trim procedure is essential for elastically coupled blades since elastic deections play an important role in the steady net forces and moments generated by the rotor. Further details of the aeroelastic analysis are available in [35,36]. 2.2. Composite rotor blade The composite helicopter blade is modelled as an one-dimensional thin walled beam undergoing extension, torsion, ap (out-of-plane), and lag (in-plane) bending using the Chandra and Chopra [37] model. Chandra and Chopras [37] thin walled composite beam model includes terms due to constrained warping torsion, and terms due to transverse shear. The effects of transverse shear are included by static condensation. The restrained warping effect is negligible for a closed section and is therefore ignored for this work. The stiffness matrix of order 9 9 is thus reduced to a stiffness matrix of order 4 4 [38] 3 2 03 2 3 2 K0 W K 012 K 013 K 015 N 11 6 7 6 7 0 0 0 6 Mx 7 6 K 22 K 23 K 25 7 6 f0y 7 6 7 6 7 6 7 (3) 6 7 6 0 7. K 033 K 035 7 4 M y 5 6 4 sym: 5 4 fx 5 K 055 Ts f0z The coefcients K 0ij of stiffness matrix are obtained by static condensation of the 9 9 stiffness matrix K. The terms in the K matrix depend on the beam cross-section and geometry and are expressed in terms of A, B and D matrices. Here EA K 011 , EI y K 022 , EI z K 033 , GJ K 055 are the axial, ap, lag and torsion stiffness, respectively. The effects of composite material are included in the forward ight simulation through the strain energy expression using the cross-sectional stiffness matrix derived in this section. The effects of the key damages in composite material are included in the forward ight simulation through the cross-sectional stiffness matrices using the progressive damage accumulation model given in the next section.
ARTICLE IN PRESS
2.3. Progressive damage accumulation Typically, matrix cracking is the rst failure mode in composites and is followed by debonding/ delamination. These damage modes are modelled at the lamina and laminate level and the procedure is summarised below. Matrix cracking: The effects of matrix cracking in composite materials are included through the extension (A), extensionbending (B) and bending (D) stiffness matrices. The stiffness matrices for the presence of matrix cracks Ac , Bc and Dc are obtained by subtracting the damage matrices DA, DB and DD from the stiffness matrices A, B and D of the virgin laminate: Ac A DA, B
c c
4 5 6
B DB, D DD.
These stiffness matrices reduce with increasing crack density. The dimensionless crack density rk for the ply k is dened by tk . (7) sk Here sk is the average crack spacing and tk is the thickness of ply k. The changes in stiffness matrices are obtained using the Adolfsson and Gudmundson [39] matrix crack model, which relates the strain increment produced by an array of cracks to the total crack displacement. Debonding/delamination: After a certain crack density, the stiffness get saturated at a crack density known as the saturation crack density (f0 ). However, as damage increases, matrix cracks may induce more severe damage at the crack tip such as debonding/delamination. The ply stiffness due to the presence of debonding/ delamination can be expressed as rk
d QM xx f rE xx f,
8 9 10 11 12
QM yy f QM yx f QM xy f QM ss f where
rE d yy f, d rmxy fE d xx f, d r md yx fE xx f, Gd xy f,
r 1 mxy fmyx f1 .
(13)
Stiffness reduces with increase in the effective strain ratio. At the end of the debonding/delamination phase, the bre breakage phase begins. This phase nally leads to catastrophic failure of the composite structure. A detailed discussion of the damage models is given in [22]. In present paper, we do not consider the bre breakage phase and consider the point of transition of debonding/ delamination to bre breakage to be the end of the debonding/delamination phase. 3. Numerical simulation of measurement deltas For a SHM system, blade response, root loads and strains are candidate measurements. However, these measurements are functions of the location of sensors, azimuth angle around the rotor disk 0pcp2p and damage level. A large amount of data is generated due to the many sensor locations, azimuth locations along the rotor disk and damage levels. For a specied sensor location, the measurement depends on an azimuth c for a specied damage level. By taking the peak-to-peak value of the total signal obtained for c values between 0 and 360 , the blade system response can be expressed as a scalar variable which is a function of a damage parameter only. A considerable amount of data reduction is thus obtained by considering peak-to-peak values
ARTICLE IN PRESS
of the system response. Comparing the peak-to-peak values between the undamaged and damaged rotor blade results in a measurement delta which is an indicator of damage. In this study, a four bladed stiff-inplane hingeless composite rotor with progressive damage accumulation in composite material is considered. The progressive damage model includes matrix cracking and debonding/ delamination. Geometric properties and ply orientation of the two-cell airfoil section are shown in Fig. 2. Ply elastic stiffness properties are E L 206 GPa (30 msi), E T 20:7 GPa (3 msi), GLT 8:3 GPa (1.2 msi) and mLT 0:3. The length (l) of the beam is 5 m (200 in). Since actual rotor blades do not exploit composite couplings, the sections are selected to have zero or minimum couplings. Therefore, in the present study, all the laminates are selected from the family of 0= 45=90s composite to minimise the coupling terms. Ten modes (ve ap, two lag, three torsion) are used for the trim analysis. The rst lag, rst ap and rst torsion rotating frequencies of the undamaged blade are 1:84O, 1:11O and 3:6O where O is the rotor rotational speed. The blade steady response is calculated by dividing the azimuth into six time elements with a fth order polynomial distribution within each time element. Results are obtained in forward ight at an advance ratio m V =OR of m 0:3, a moderate thrust condition C T =s 0:07, lock number g 6:34, radius of gyration 2 2 2 mk2 m1 =m0 R 0:000174, mkm2 =m0 R 0:00061 and m=m0 1. The transition from matrix cracking to debonding/delamination is assumed to occur at the matrix crack saturation density f0 3 and the transition of the debonding/delamination to the bre breakage damage mode is assumed to occur at an effective strain ratio of 0.94. Only the signicant stiffness reduction zones i.e. the debonding/delamination for effective strain ratios between 0.8 and 0.94 are considered. Fig. 3 shows the distribution of the bending and torsional stiffnesses for matrix cracking and debonding/delamination. Here, 0 0 EI 0 y , EI z and GJ are the cross-sectional stiffnesses of the beams with virgin laminate. From the rst column of Fig. 3, it is observed that stiffness decreases rapidly up to a crack density of 1.5 and thereafter the rate of decrease of stiffness reduces and gets saturated at crack density of 3. Matrix cracking is rarely the cause of blade failure; however, monitoring matrix cracking gives an idea about damage initiation in the composite rotor blade. When matrix crack density becomes very high, matrix cracks may induce debonding/ delamination. The second column of Fig. 3 shows the variation of stiffness due to an increase in effective strain
y (0)3 (45)4 (90) (90) (45)4 (0)3
t = 36 x 0.005 in
(0)6 (45)4 (90)4 (90)4 (45)4 (0)6 z 0.35c D-Spar
NACA0012 with chord (c)=12 in.
skin
Fig. 2. Details of two-cell airfoil section beam.
Fuzzifier
Fuzzy Inference Engine
Defuzzifier
Damage detection , location and size
Fuzzy Rule Base Fuzzy Logic System (FLS)

Fig. 3. Change in stiffness due to progressive damage accumulation.
t = 24 x 0.005 in
ARTICLE IN PRESS
ratio. Reduction in bending stiffness in this damage mode is very small. However, the rate of torsional stiffness reduction is very high. Figs. 4 and 5 show the behaviour of displacements due to matrix cracking and debonding/delamination, respectively. The change in tip ap response and elastic twist increases steadily in both the damage modes, while the change in lag response is very small. However, the ap and torsion response decreases in the region where transitions between two damage modes occur. Figs. 6 and 7 show the behaviour of root forces and moments due to matrix cracking and debonding/delamination damage mode, respectively. The change in root forces increases steadily after a slight reduction at the initial stage of matrix cracking damage. In debonding/ delamination, the change in root lateral and vertical forces remains constant. The change in root longitudinal force remains constant up to an effective strain ratio of 0.88 and thereafter increases rapidly. The effect of matrix cracking on root rolling and pitching moments is negligible as compared to root yawing moment. In debonding/delamination, all the moments remain almost constant except for the pitching moment during transition of the damage mode.
Matrix cracking 1 1 Debonding / delamination
0 EIY / EIY
0.5
0.5
0 0 1 1 2 3
0 0.8 1 0.85 0.9 0.94
EIZ / EI0 Z
0.5
0.5
0 0 1 1 2 3
0 0.8 1 0.85 0.9 0.94
GJ / G0 J
0.5
0.5
0 0 1 2 3 crack density
0 0.8 0.85 0.9 0.94 Effective strain ratio
Fig. 4. Change in peak to peak tip lag (m), ap (m) and torsion (degrees) for increasing matrix crack density.
ARTICLE IN PRESS
0.1 0.08 0.06 V and W 0 0.5 1 1.5 crack density 2 2.5 3 0.04 0.02 0 -0.02 V W
10 9 8 7 6 5 4 3 2 1 0 0 0.5 1 1.5 crack density 2 2.5 3
Fig. 5. Change in peak to peak tip lag (m), ap (m) and torsion (degrees) for increasing effective strain ratio due to debonding/ delamination.
0.2 V 0.15 V and W W
12 11 10
0.1 0.05 0 -0.05 0.8 0.82 0.84 0.86 0.88 0.9 0.92 0.94 Effective strain ratio
9 8 7 6 5 4 0.8 0.82 0.84 0.86 0.88 0.9 0.92 0.94 Effective strain ratio
Fig. 6. Change in peak to peak blade root forces (N) and moments (N m) for increasing matrix crack density.
Strains are useful local indicators of damage. To study the effects of the local damage, the peak-to-peak changes in shear strains due to matrix cracking and debonding/delamination damage are calculated at ve locations on the blade ranging from the root to the tip. All the strains are calculated on the top side of the beam and along the line passing through the point 0.35c on two-cell airfoil section. These peak-to-peak changes in shear strains are shown in Figs. 8 and 9 for matrix cracking and debonding/delamination, respectively. The maximum change in shear strains occurs where the damage occurs along the blade length in cases of local damage. Therefore, the shear strain is a useful indicator to predict damage location. The current study is based on numerically simulated measurements such as blade tip response and blade root loads. Selected researchers have worked on developing methods to obtain these measurements experimentally. Optical blade tracking devices with day and night capability that are part of health and usage monitoring systems (HUMS) can measure relative height and leadlag deection differences between the blades [40]. Blade trackers can be mounted on a helicopter nose or fuselage, oriented upwards toward the under side of a rotating hub. The optical tracker views the blade through a window of discrete time and
ARTICLE IN PRESS
2500 2000 Fx , Fy and Fz 1500 1000 500 0 -500 -1000 -1500 0
F x Fy Mx, My and Mz Fz
2000 1000 0 -1000 -2000 -3000 -4000 -5000 Mx My Mz
0.5
1.5 crack density
2.5
0.5
1.5 crack density
2.5
Fig. 7. Change in peak to peak blade root forces (N) and moments (N m) for increasing effective strain ratio due to debonding/ delamination.
9000 8000 7000 Fx , Fy and Fz 6000 5000 4000 3000 2000 1000 0 -1000 0.8 0.82
Fx Fy Fz Mx, My and Mz
4500 4000 3500 3000 2500 2000 1500 1000 500 0 -500
Mx My Mz
0.84
0.86
0.88
0.9
0.92
0.94
0.8
0.82
0.84
0.86
0.88
0.9
0.92
0.94
Effective strain ratio
Effective strain ratio
Fig. 8. Change in peak to peak shear strain measured at various locations for various crack densities and cracks at various locations (tip 0 m and root 5 m) (crack locations: T tip, O outboard, C centre, I inboard, R root).
operates on the principle that a higher ying blade will remain in the eld of view longer than a lower ying blade. Tip torsional response measurement can be possible using a laser Doppler vibrometer (LDV) [41,42]. Unlike displacement measurement deltas, force and moment measurement deltas cannot be measured directly on rotorcraft. However, the force and moment measurement deltas can be estimated using vibration level measurements. Recently, some studies have been done on predicting the oscillatory loads in a rotor system from the xed system load information using a neural network [43]. For example, the pitch-link load is estimated by a neural network from roll, pitch and yaw rates; airspeed; and other xed system information measured by a ight control computer on the helicopter. Actual ight loads data from an AH-64A helicopter were used to demonstrate the process. The predicted loads agreed well with the measured data. Azzam [44] developed a learning network to predict damage in a helicopter rotor from the ight parameters. The network predicted the fatigue damage of two rotating components indirectly from the ight parameters with accuracy better than a strain gauge system and with a measurement error of 5%. Thus, measurement of blade response and loads are possible in an online setting and can be used for online health monitoring of rotor blades. Typically, response measurements can be measured directly whereas vibratory loads may require the use of the other processing and therefore can be thought of as virtual sensors.
ARTICLE IN PRESS
x 103 14 12 peak to peak shear strain 10 8 6 4 2 0 -2 -4 -6 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 gage location (crack density=0.4) 0.05 0.04 peak to peak shear strain 0.03 0.02 0.01 0 -0.01 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 gage location (crack density=1.6) T O C I R 0.06 0.05 peak to peak shear strain 0.04 0.03 0.02 0.01 0 -0.01 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 gage location (crack density=2.4) T O C I R T O C I R 0.045 0.04 peak to peak shear strain 0.035 0.03 0.025 0.02 0.015 0.01 0.005 0 -0.005 0 1 2 3 4 5 gage location (crack density=1.2) T O C I R
Fig. 9. Change in peak to peak shear strain measured at various locations for various effective strain ratios and debonding/delamination at various locations (tip 0 m and root 5 m) (crack locations: T tip, O outboard, C centre, I inboard, R root).
4. Fuzzy logic system Before developing the GFS, a brief description of a FL system is given in this section. Fuzzy logic is a unique soft computing method which simultaneously handles numerical data and linguistic knowledge. A FL system is a nonlinear mapping of an input feature vector into a scalar output [45]. Fuzzy set theory and FL provide the framework for the nonlinear mapping. Fuzzy logic systems have been widely used in engineering applications, because of the exibility they offer designers and their ability to handle uncertainty. A FL system can be expressed as a linear combination of fuzzy basis functions and is a universal function approximater. The schematic diagram of FL is shown in Fig. 10 for Do as crisp inputs and damage and damage location as crisp output. Here, Do is a measurement delta which can be change in frequency between the undamaged and damaged structure, for example. It could also be changes in strains, blade tip responses and loads in a helicopter rotor blade. A typical multi-input single-output (MISO) FL system performs a mapping from V 2 Rm to W 2 R using four basic components: rules, fuzzier, inference engine, and defuzzier. Here F : V 2 Rm ! W 2 R
ARTICLE IN PRESS
0.2 0.16 0.12 0.08 0.04 0 0 1 2 3 4 5 T O C O R
0.35 0.3 peak peak shear strain T O C O R
peak peak shear strain
0.25 0.2 0.15 0.1 0.05 0 -0.05 0
gage location (ESR = 0.8 ) 0.35 0.3 peak peak shear strain 0.25 0.2 0.15 0.1 0.05 0 -0.05 0 1 2 3 4 T O C I R
gage location ( ESR = 0.88 )
gage location ( ESR = 0.92 )

Fig. 10. Schematic representation of fuzzy logic system.
where V V 1 V 2 V n 2 Rm is the input space and W 2 R is the output space. A typical FL system maps crisp inputs to crisp outputs using four basic components: rules, fuzzier, inference engine, and defuzzier. Rules can come from experts or can be obtained from numerical data. In either case, engineering rules are expressed as a collection of IFTHEN statements such as IF u1 is HIGH, and u2 is LOW, THEN v is LOW. To formulate such a rule we need an understanding of 1. Linguistic variables versus numerical values of a variable (e.g., HIGH versus 3.5%). 2. Quantifying linguistic variables (e.g., u1 may have a nite number of linguistic terms associated with it, ranging from NEGLIGIBLE to VERY HIGH), which is done using fuzzy membership functions 3. Logical connections between linguistic variables (e.g., AND, OR etc.), and 4. Implications such as IF A THEN B. We also need to understand how to combine more than one rule. The fuzzier maps crisp input numbers into fuzzy sets. It is needed to activate rules that are expressed in terms of linguistic variables. An inference engine of the FL system maps fuzzy sets to fuzzy sets and determines the way in which the fuzzy sets are combined. In several applications, crisp numbers are needed as an output of the FL system. In those cases, a defuzzier is used to calculate crisp values from fuzzy values.
ARTICLE IN PRESS
4.1. Fuzzy sets A fuzzy set F is dened on a universe of discourse U and is characterised by a degree of membership mx, which can take on values between 0 and 1. A fuzzy set generalised the concept of an ordinary set whose membership function only takes two values, zero and unity. 4.2. Linguistic variables A linguistic variable u is used to represent the numerical value x, where x is an element of U. A linguistic variable is usually decomposed into a set of terms T u, which cover its universe of discourse. 4.3. Membership functions The most commonly used shapes for membership functions mx are triangular, trapezoidal, piecewise linear, or Gaussian. The designer selects the type of membership function used. There is no theoretical requirement that membership functions overlap. However, one of the major strengths of FL is that membership functions can overlap. Fuzzy logic systems are robust because decisions are distributed over more than one input class. For convenience, membership functions are normalised to one so they take values between 0 and 1, and thus dene the fuzzy set. 4.4. Inference engine Rules for the FS can be expressed as Ri : IF x1 is F 1 AND x2 is F 2 AND AND xm is F m THEN y C i i 1; 2; 3; . . . ; M where m and M are the number of input variables and rules, xi and y are the input and output variables, and F i 2 V i and C i 2 W are fuzzy sets characterised by membership functions mF i x and mC i x, respectively. Each rule can be viewed as a fuzzy implication F 12m F 1 F 2 F m ! C i which is a fuzzy set in V W V 1 V 2 . . . V m W with membership function given by mRi x; y mF i x1 mF 2 x2 mF m xm mC i y where the asterisk can be the min or product operator with x x1 ; x2 ; :::; xm 2 V and y 2 W . This sort of rule covers many applications. The algebraic product is one of the most widely used operators in applications and leads to product implication. In pattern recognition problems, the outputs are often crisp sets, and mC i y 1 is often used for the product implication formula [45]. 4.5. Defuzzication Popular defuzzication methods include maximum matching and centroid defuzzication. While centroid defuzzication is widely used for fuzzy control problems where a crisp output is needed, maximum matching is often used for pattern matching problems where we need to know the output class. Suppose there are K fuzzy rules and among them, K j rules (j 1; 2; . . . ; L and L is the number of classes) produce class C j . Let Dip be the measurements of how the pth pattern matched the antecedent conditions (IF part) of the ith rule, which is given by the product of membership grades of the pattern in the regions which the ith rule occupies Dip
m Y i 1
mli ,
(14)
where m is the number of inputs and mli is the degree of membership of measurement l in the fuzzy regions that the ith rule occupies. Let Dmax p C j be the maximum matching degree of the rules (rules j l , l 1; 2; . . . ; K j )
ARTICLE IN PRESS
generating class C j
jl Dmax p C j max Dp , l 1 Kj
(15)
then the system will output class C j provided that

max Dmax p C j max Dp C j . j
(16)
If there are two or more classes which achieve the maximum matching degree, we will select the class which has the largest number of red fuzzy rules (a red rule has a matching degree of greater than zero). 4.6. Difculties in fuzzy system design Two aspects in the design of the FS are difcult: (1) generating the best rule set and (2) tuning the membership functions. The rules and the membership functions must accurately capture the relationship between the independent and dependent variable. Unfortunately, the task of tuning of membership function and generating rules are not independent. The task of selecting membership functions and rule values is difcult since the information has to be obtained from numerical data of the system to be modelled. Another problem is selecting an appropriate number of fuzzy sets. Most studies use experience to come up with this number. The problems in the design of FS can be addressed by using a learning algorithm such as the GA. This leads to the GFS whose design for the rotor SHM problem is discussed in the next section. 5. Development of genetic fuzzy system There are two possibilities of damage in composite rotor blades. The rst possibility is that the damage will be approximately uniform along the whole blade which may occur due to the vibrating environment of the helicopter. The second possibility is that the damage will be localised which may occur due to a sudden impact by a foreign object or due to uneven loading conditions. In this section, we formulate two GFSs. The rst or global GFS is for matrix cracking and debonding/delamination detection along the whole blade. The second or local GFS is for matrix cracking and debonding/delamination detection in various parts of the blade. Both the global and local GFS systems are developed using the measurement deltas obtained from the aeroelastic analysis discussed earlier in the paper. The schematic diagram of the GFS system is shown in Fig. 11. Development of various components of the GFS for global and local damage detection system is explained below. 5.1. Global damage detection The global damage detection system is developed to predict the damage along the whole blade. 5.1.1. Input and output Inputs to the global GFS are measurement deltas based on displacement, force and moment. The outputs of the global GFS are various damage levels in the matrix cracking and debonding/delamination zone. The objective is to nd the mapping between measurement deltas and various damage levels. 5.1.2. Fuzzication In this FS, matrix crack densities for matrix cracking and the effective strain ratios (ESR) for debonding/ delamination are crisp numbers. To get a degree of resolution of the extent of matrix cracking and debonding/ delamination, these two damage modes are allowed several levels of damage and split into linguistic variables, as shown in Table 1. These classications are based on the numerical results obtained for the damaged blade. We rst consider the displacement measurement deltas for dening the rules, which show approximately monotonic behaviour with respect to damage (Figs. 4 and 5). Next, the force and moment measurement deltas are considered for damage prediction (Figs. 6 and 7). The rst step in FL is to transform typical measurements
ARTICLE IN PRESS
DESIGN PROCESS Genetic Algorithm Based Learning Process
Knowledge Base
Input Interface
Fuzzy System
Output Interface
Environment
Computation with fuzzy systems

Fig. 11. Genetic fuzzy rule-based system.
Environment
Table 1 Linguistic classication of damage for the genetic fuzzy system for global damage detection No. 1 2 3 4 5 6 7 8 9 10 11 12 Rule Undamaged Very small crack density Small crack density Considerable crack density High crack density Very high crack density Saturation crack density Transition of MC to DD Slight D/D Moderate D/D Severe D/D Extreme D/D Damage level Nil crack density Crack density 00.8 Crack density 0.81.2 Crack density 1.21.6 Crack density 1.62.0 Crack density 2.02.4 Crack density 2.43.0 Crack density 3.0a to ESR 0.8b ESR 0.80.88 ESR 0.880.9 ESR 0.90.92 ESR 0.920.94 Prog. action OK OK OK OK OK OK OK WATCH WATCH WATCH WATCH REMOVE
Prognostic actions:- (1) OK: blade is ok, no action is required. (2) WATCH: put blade under watch. (3) REMOVE: remove blade. Take for thorough inspection. a Saturation crack density. b The ESR from where the effects of D/D become considerable.
into fuzzy linguistic measures. This is done using the relation between peak-to-peak displacements and matrix crack density shown in Fig. 4 and from the relation between peak-to-peak displacements and debonding/ delamination shown in Fig. 5. Table 1 shows the linguistic and numerical measures based on the displacement measurement deltas. These linguistic classications allow the damage to be grouped into small levels which are more robust to the presence of uncertainty and also provide a fuzzy rule base which is similar to human reasoning. The different levels of damage can be used to create different alarm levels to be shown to the user. The measurement deltas DW , Df, DF x , DF y , DF z , DM x , DM y , and DM z are treated as fuzzy variables. Fuzzy sets with Gaussian membership functions are used to dene these input variables. The Gaussian membership function can be written as mx e0:5xm=s .
2
(17)
ARTICLE IN PRESS
Here m is the midpoint and s is the standard deviation. Gaussian fuzzy sets have the advantage of providing smooth transition between the different sets [45]. Furthermore, they always have nonzero values on the real number line and therefore every rule in the FS res to some degree. Changes in the measurement deltas are calculated using the aeroelastic analysis for different damage levels along the whole blade. The midpoints of the Gaussian function are calculated by normalising the changes in measurement deltas with their maximum values. The GFS is tested using the normalised noisy measurement delta (x) which is derived from the noise model given as x m ua. (18)
The addition of noise to the simulations is needed to make it realistic and develop a robust model-based diagnostic system. A noise level parameter a denes the maximum variance between the computed value of m (normalised) and simulated measured value x (normalised) which is a simulation of the practical measurement. 5.1.3. Rule generation Rules for the FS are obtained by fuzzication of the numerical values obtained from an aeroelastic analysis of the composite helicopter blade in forward ight. The fuzzy sets corresponding to DW , Df, DF x , DF y , DF z , DM x , DM y , and DM z are generated by taking the change in measurements obtained from the aeroelastic analysis solution as midpoints of membership functions corresponding to a damage level. For each measurement delta corresponding to a given damage level, the degree of membership in the fuzzy set is calculated. Each measurement delta is assigned to the fuzzy set with the maximum degree of membership. One rule is obtained for each damage level by relating the measurement deltas. The standard deviation is obtained by maximising the success rate for each set. The fuzzy rules provide a knowledge base and represent how a human engineer would interpret data to isolate a damage level using measurement deltas. The fuzzy rules represent a fuzzied model of the measurements obtained by an aeroelastic analysis for each damage level. Since, Gaussian fuzzy sets asymptotically approaching zero far from the midpoint are used, all of the rules re at some level. For any given input set of measurement deltas, the fuzzy rules are applied using product implication. Once the fuzzy rules are applied for a given measurement, we have degrees of membership for each of the damage levels. For damage level isolation, we are interested in the most likely damage level. The damage level with the highest degree of membership is selected as the most likely damage level. 5.1.4. Tuning of the rules The realistic measurement deltas will be noisy due to vibrating environment of the helicopter. By generating noisy measurement deltas and testing the FS for a known damage, we can dene a success rate. For example, if N is the total number of classications and N c is the number of correct classications, the success rate can be dened as SR Nc 100. N (19)
The success rate of the GFS is calculated by using N number of noisy training samples. Since the midpoints of the fuzzy sets are tuned using the aeroelastic analysis, the success rate is a function of the standard deviations of the Gaussian functions for the FS i.e. SR S R sij . (20)
To get a better performance in this environment, the FS is tuned with noisy measurement deltas. The uncertainty associated with variables, i.e. standard deviation of the Gaussian membership functions is calculated using GA for optimisation of the success rate. Maximise S R sij for smin psij psmax ; i 1; 2; . . . ; M and j 1; 2; . . . ; P, (21)
where M is the number of rules and P is the number of measurements. Therefore, the success rate of the GFS is the objective function and the standard deviations s corresponding to each rule are the design variables for
ARTICLE IN PRESS
the genetic algorithm. The use of formal optimisation to design the FS leads to an optimal diagnostic system which provides the best results for the given structure, measurement set and noise level in the data. 5.2. Local damage detection The local damage detection system is developed to predict the damage at various locations along the blade. 5.2.1. Input and output Inputs to the GFS are strain-based measurement deltas at ve locations and outputs are damage levels and location. The damage considered here is matrix cracking and debonding/delamination. The objective is to nd a mapping between the measurement deltas and the ve damage locations. 5.2.2. Fuzzcation In this FS, the matrix crack densities and effective strain ratios at a given location are crisp numbers. The different locations considered are as follows. Tip ranges from 0% to 20% of the blade from the free end, Outboard from 20% to 40%, Center from 40% to 60%, Inboard from 60% to 80% and Root from 80% to 100%. Each damage location is allowed several levels of damage and split into linguistic variables, as shown in Table 2. It is not computationally efcient to calculate the shear strains at all crack densities and effective strain ratios. Therefore, based on the experience of the results of the previous section, FL rules are dened based on the shear strains obtained for a few key matrix crack densities and effective strain ratios. The linguistic relations of the key matrix crack densities for matrix cracking and effective strain ratios for debonding/delamination are shown in Table 2. Strain-based measurement deltas DTip , DOutboard , DCenter , DInboard and DRoot are treated as fuzzy variables. Fuzzy sets with Gaussian membership functions (Eq. (18)) are used to dene these input variables. Change in strains (measurement deltas) is calculated using the aeroelastic analysis for combination of ve different locations and seven different levels of damages and are shown in Figs. 8 and 9. 5.2.3. Rule generation Formulation of the GFS system and calculation of the success rate is done using the algorithm discussed in the previous section. In this way the Undamaged level of damage is represented by one rule and the seven damage levels at ve locations are represented by ve rules each. Therefore, the complete matrix crack and debonding/delamination detection system can be represented using 36 fuzzy rules. For example Less Crack Density damage at 020% part of beam can be written as Less Crack Density at Tip.
Table 2 Linguistic classication of damage for local damage detection Damage name Undamaged Small crack density Moderate crack density High crack density Very high crack density Slight D/D Moderate D/D Severe D/D Damage level Matrix crack Matrix crack Matrix crack Matrix crack Matrix crack ESR 0.8 ESR 0.88 ESR 0.92 density density density density density zero 0.4 1.2 1.6 2.4 Prog. action OK OK OK OK OK WATCH WATCH REMOVE
Prognostic actions:- (1) OK: blade is ok, no action is required. (2) WATCH: put blade under watch. (3) REMOVE: remove blade. Take for thorough inspection.
ARTICLE IN PRESS
6. Testing of genetic fuzzy system The GFS for global and local damage detection are tested at various noise levels. All the measurements are normalised with their maximum value. 6.1. Global damage detection Two key issues are discussed in the next two subsections using numerical simulations. The rst is the use of different sets of measurement deltas and the second is the effect of training noise on the GFS. 6.1.1. Effect of different measurement sets The midpoints and standard deviations of the Gaussian membership functions which form the rule base of the global GFS are given in Table 3. This rule base is obtained using a training noise level of 0.15. Success rates of the GFS are calculated for various test noise levels of 0.03, 0.05, 0.10, 0.15 and 0.20. The global GFS is tested using four sets of measurement deltas: (1) tip displacement only, (2) blade root forces only, (3) blade root moments only and (4) all three. The performance of the global GFS for the four sets of measurement deltas gives an idea about the reliability of the SHM based on different measurement instruments. It may be possible to implement the displacement-based SHM in an old helicopter to achieve economy and simplicity. The root load-based SHM can be implemented when the root loads can be obtained from the existing instrumental arrangements. The SHM based on the all measurement deltas can be implemented when more accurate prediction is needed. Table 4 shows the success rates for all the rules for four different sets of measurement deltas at various noise levels and these results are summarised in Fig. 12. From Table 4 and Fig. 12, it is observed that the GFS with displacement-based measurement deltas gives good results at the noise level of 0.05 with an average success rate of 95.55% and a minimum success rate of 80.90% (transition of MC to DD) and starts deteriorating rapidly with further increase in the noise level. The critical performance of the SHM system can be judged based on the success rate of the key fuzzy rules. The key fuzzy rules for the current SHM are rule number 7 and rule number 12. Rule number 7 indicates the beginning of a severe damage mode, debonding/delamination and changing to a new prognostic action Put blade under watch. Rule number 12 indicates about the beginning of a catastrophic damage, bre breakage, and suggests an action of removing the blade for detailed NDT inspection and possible repair or replacement. The SHM performs satisfactorily for the two key rules by giving success rates of 98% and 100% for rule number 7 and 12, respectively, at a noise level of 0.05. The GFS with force-based measurement deltas gives a success rate of 100% at the noise level of 0.05. For a noise level of 0.1, the force-based GFS gives an average success rate of 99% and a minimum success rate of 91.3% (severe D/D). At higher noise levels of 0.15 the average success rates fall to 96% and 89.8%. At a noise
Table 3 Midpoints and standard deviations of rules for the genetic fuzzy system for global damage detection No. 1 2 3 4 5 6 7 8 9 10 11 12 DW 0.00 0.07 0.03 0.15 0.27 0.42 0.57 0.51 0.42 0.50 0.65 0.98 (0.10) (0.11) (0.17) (0.11) (0.17) (0.10) (0.14) (0.16) (0.17) (0.15) (0.14) (0.19) Dy 0.00 0.09 0.32 0.49 0.64 0.76 0.84 0.67 0.52 0.61 0.77 1.00 (0.12) (0.13) (0.11) (0.11) (0.17) (0.15) (0.17) (0.17) (0.15) (0.18) (0.11) (0.12) DF x 0.00 0.11 0.12 0.11 0.38 0.75 1.00 0.49 0.21 0.42 0.45 0.51 (0.19) (0.13) (0.15) (0.12) (0.16) (0.13) (0.19) (0.13) (0.14) (0.16) (0.16) (0.10) DF y 0.00 0.52 0.96 0.70 0.30 0.29 0.83 1.00 0.76 0.50 0.37 0.07 (0.16) (0.17) (0.18) (0.16) (0.15) (0.15) (0.19) (0.16) (0.11) (0.18) (0.18) (0.18) DF z 0.00 0.05 0.06 0.02 0.08 0.20 0.33 0.40 0.44 0.49 0.64 1.00 (0.14) (0.14) (0.17) (0.13) (0.13) (0.18) (0.14) (0.13) (0.16) (0.16) (0.16) (0.11) DM x 0.00 0.07 0.04 0.18 0.43 0.80 1.00 0.77 0.50 0.49 0.60 0.71 (0.18) (0.16) (0.17) (0.11) (0.13) (0.19) (0.18) (0.19) (0.13) (0.13) (0.16) (0.20) DM y 0.00 0.18 0.34 0.25 0.14 0.07 0.29 0.18 0.07 0.20 0.41 1.00 (0.12) (0.19) (0.19) (0.15) (0.20) (0.15) (0.12) (0.13) (0.11) (0.18) (0.20) (0.12) DM z 0.00 0.49 1.00 0.95 0.75 0.33 0.06 0.22 0.16 0.06 0.10 0.03 (0.19) (0.13) (0.17) (0.13) (0.12) (0.15) (0.14) (0.15) (0.18) (0.19) (0.18) (0.17)
ARTICLE IN PRESS
2230 P.M. Pawar, R. Ganguli / Mechanical Systems and Signal Processing 21 (2007) 22122236 Table 4 Success rate for various testing noise levels and training noise level of 0:15 for global damage detection using displacement, force, moments and all measurements No. Displacements SR0:05 1 2 3 4 5 6 7 8 9 10 11 12 Avg. Min. SR0:1 SR0:15 Forces SR0:20 SR0:05 60.20 50.50 65.10 29.10 49.90 24.40 41.70 21.10 48.20 23.70 34.00 96.60 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 SR0:1 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 97.20 91.30 100.00 99.04 91.30 SR0:15 SR0:20 Moments SR0:05 SR0:1 SR0:15 SR0:20 All SR0:05 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 SR0:1 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 SR0:15 100.00 100.00 100.00 98.30 100.00 100.00 100.00 100.00 95.30 98.90 98.00 100.00 99.21 95.30 SR0:20 100.00 99.90 99.30 90.20 100.00 100.00 100.00 99.50 86.30 94.40 94.60 100.00 97.02 86.30
96.90 79.70 69.90 97.80 81.60 68.70 100.00 97.50 77.60 100.00 81.20 46.10 100.00 92.40 69.40 98.10 61.40 35.50 97.70 73.60 55.10 80.90 48.80 27.80 98.40 73.90 60.40 82.30 46.60 31.60 94.50 66.00 46.90 100.00 100.00 100.00 95.55 80.90 75.23 46.60
99.60 95.50 100.00 100.00 100.00 97.50 97.00 90.30 100.00 100.00 100.00 96.40 99.70 95.90 100.00 99.90 94.70 90.60 89.30 66.30 100.00 94.90 74.30 45.80 100.00 100.00 100.00 100.00 99.90 97.80 100.00 99.80 100.00 100.00 99.50 89.60 100.00 100.00 100.00 100.00 95.50 89.00 100.00 100.00 100.00 99.80 87.00 68.00 96.40 91.30 100.00 91.40 75.40 60.80 88.30 78.00 100.00 93.20 76.70 65.20 76.30 63.90 100.00 97.40 79.70 67.70 100.00 96.60 100.00 100.00 100.00 100.00 95.55 76.30 89.80 100.00 63.90 100.00 98.05 91.40 90.23 74.30
57.42 45.38 100.00 27.80 21.10 100.00
80.70 100.00 100.00 45.80 100.00 100.00
levels of 0.20 the minimum success rates are 76% and 63.90%, respectively. The force-based GFS gives a 100% success rate for both the key rules even at a higher noise level of 0.15. The GFS with moment-based measurement deltas gives a success rate of 100% at the noise level of 0.05. For a noise level of 0.1, the moment-based GFS gives an average success rate of 98.05% and a minimum success rate of 91.4% (slight D/D). At noise level of 0.15 and 0.20 the average success rate is 96% and 89.8%, respectively. Furthermore, the minimum success rates at noise levels of 0.15 and 0.20 are 71.30% and 45.80%, respectively. The moment-based GFS also gives good performance at a noise level of 0.15 for the key rules by giving a success rate of about 96% for rule 7 and a success rate of about 100% for rule 12. When all measurement deltas are considered for damage detection, the GFS gives a success rate of 100% up to a noise level of 0.10. The GFS also gives good results with higher noise levels of 0.15 and 0.20 by giving average success rates of 99.21% and 97.02%, respectively, and minimum success rates of 95.30% and 86.30%, respectively. Thus, the use of both displacement and load-based measurements results in a health monitoring system with high levels of accuracy even with noisy data. The GFS using all the available measurement deltas gives a success rate of about 100% for rule 12 which indicates the beginning of catastrophic damage. Therefore, the GFS with all measurements can be used for giving a warning about the beginning of catastrophic damage in a highly noisy environment. It is clear from the results that the noise tolerance capacity of the GFS is dependent on the type of measurement delta. However, the accuracy of the measurement instrument will also inuence the accuracy of damage prediction. For example, displacements are measured directly using optical blade tracking and LDV and are typically quite accurate. In contrast, the force and moment measurements are estimated using vibration level measurements and may be less accurate. The individual accuracy of the available sensors should be considered when using the GFS. 6.1.2. Effect of training noise In this section, the effect of different levels of noise in the training data is considered. The GFS is trained using noisy data with training noise levels of 0.05, 0.10, and 0.15. These GFSs are then tested using noisy data at noise levels of 0.05, 0.10 and 0.15. The effects of various training noise levels are studied by comparing the minimum and average success rates of the GFS with training noise levels of 0.05 and 0.10 with the success rates with the GFS with a training noise level of 0.15 for various testing noise levels. DS R0:05 indicates the difference between the success rates with a training noise level of 0.05 compared with the success rates with a training noise level of 0.15. A similar denition is used for DSR0:10 .
ARTICLE IN PRESS
displacement force moment all 100 90 Minimum SR 80 70 60 50 40 30 20 0.03 0.05 0.1 displacement force moment all 100 90 Average SR 80 70 60 50 40 0.03 0.05 0.10
Fig. 12. Effect of various measurements on minimum and average success rate of genetic fuzzy logic system for global damage detection.
0.15
0.20
0.15
0.20
The effects of training noise on the minimum success rates for different measurement data sets at various (testing) noise levels are shown in Table 5. The minimum success rate gives an indication of the reliability of the GFS. From Table 5, it is observed that the highest values of DSR0:05 and DS R0:10 for minimum success rate with displacement-based measurement deltas are about 16.1 and 11.4, respectively. This result indicates that the training noise level of 0.15 is a good option for the GFS with the displacement-based measurement deltas. The highest value of DS R0:05 for minimum success rate with the force-based measurement deltas is about 3:20 and the highest value of DS R0:10 is 14.2. This shows that for the force-based measurement deltas the training noise level of 0.05 is a good option. The highest value of DSR0:05 for minimum success rate with moment-based measurement deltas is about 4.90 and the highest value of DS R0:10 is 4.80. Thus for the moment-based measurement deltas, the training noise level of 0.15 is a good option. The highest value of DS R0:05 for
ARTICLE IN PRESS
2232 P.M. Pawar, R. Ganguli / Mechanical Systems and Signal Processing 21 (2007) 22122236 Table 5 Effect of training noise level on minimum success rate at various test noise levels for global damage detection using displacement, force, moments and all measurements a Displacements DSR0:05 0.05 0.10 0.15 15.60 16.10 13.30 DSR0:1 8.10 11.40 6.90 Forces DSR0:05 0 2.20 3.20 DSR0:1 0.02 4.80 14.20 Moments DSR0:05 0 4.90 1.90 DSR0:1 0 1.70 4.80 All DSR0:05 0 0 2.70 DSR0:1 0 0.02 0
DSR0:05 Difference between the minimum SR for training at a 0:15 and 0.05, DSR0:10 Difference between the minimum SR for training at a 0:15 and 0.10, negative value indicates better SR than the SR for training at a 0:15.
Table 6 Effect of training noise level on average success rate at various test noise levels for global damage detection using displacement, force, moments and all measurements a Displacements DSR0:05 0.05 0.10 0.15 0.78 0.35 0.25 DSR0:1 0.96 0.59 2.18 Forces DSR0:05 0 0.16 0.22 DSR0:1 0.02 0.34 1.17 Moments DSR0:05 0 0.20 0.19 DSR0:1 0 0.57 0.28 All DSR0:05 0 0 0.37 DSR0:1 0 0.02 0.10
DSR0:05 Difference between the average SR for training at a 0:15 and 0.05, DSR0:10 Difference between the average SR for training at a 0:15 and 0.10, negative value indicates better SR than the SR training at a 0:15.
minimum success rate with all measurement deltas together is about 2:70 and the highest value of DS R0:10 is 0.02. Thus when the all measurement deltas are used, the training noise level of 0.05 is a good option. The effects of training noise on the average success rates for different measurement sets at various (testing) noise levels is shown in Table 6. The average success rate gives an indication of the overall performance of the system. From Table 6, it is observed that the GFS with training noise level of 0.15 gives a good success rate except for some cases of the training noise level of 0.05. The highest negative value DSR0:05 for average success rates is 0:37 at a testing noise level of 0.15. This result indicates that at a testing noise level of 0.15, the success rate of the GFS with training noise level of 0.05 is more than the success rate of the GFS with training noise level of 0.15 by an amount of 0.37. Therefore, from the results of minimum success rates and average success rates, the training noise level of 0.15 appears to be a judicious selection because this system also performs well with single measurements like displacement and moments-based measurements deltas. 6.2. Local damage detection The midpoints and standard deviations of the Gaussian membership functions which form the rule base of the local GFS are shown in Table 7. These results are obtained at a training noise level of 0.05, which is found to be good for the strain-based measurement deltas used for local damage detection. The midpoints in this table are normalised with the maximum value of the corresponding measurement deltas and the standard deviations are given in the parentheses. The success rates for local damage detection are calculated at noise levels of 0.03, 0.05, and 0.10. Table 8 shows the success rate for all the rules with various noise levels. From Table 8, it is observed that the GFS for local matrix crack detection gives a success rate of 100% up to a noise level of 0.03. For local debonding/ delamination detection, the GFS gives a success rate of 100% for a noise level of 0.03 except for two rules, moderate D/D at root and severe D/D at root. For a noise level of 0.05, the GFS gives an average success rate of 99.42% and a minimum success rate of 91.8%. The GFS gives a minimum success rate of 97.70% for matrix crack detection and a minimum success rate of 91.8% for debonding/delamination detection for a noise
ARTICLE IN PRESS
P.M. Pawar, R. Ganguli / Mechanical Systems and Signal Processing 21 (2007) 22122236 Table 7 Midpoints and standard deviations of rules for local damage detection using strain-based measurement deltas No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 Rule Undamaged Small CD at tip Small CD at outboard Small CD at centre Small CD at inboard Small CD at root Moderate CD at tip Moderate CD at outboard Moderate CD at centre Moderate CD at inboard Moderate CD at root High CD at tip High CD at outboard High CD at centre High CD at inboard High CD at root Very high CD at tip Very high CD at outboard Very high CD at centre Very high CD at inboard Very high CD at root Slight D/D at tip Slight D/D at outboard Slight D/D at centre Slight D/D at inboard Slight D/D at root Moderate D/D at tip Moderate D/D at outboard Moderate D/D at centre Moderate D/D at inboard Moderate D/D at root Severe D/D at tip Severe D/D at outboard Severe D/D at centre Severe D/D at inboard Severe D/D at root DTip 0 (0.18) 0.08 (0.13) 0.00 (0.19) 0.00 (0.16) 0.01 (0.13) 0.01 (0.14) 0.17 (0.12) 0.01 (0.11) 0.02 (0.10) 0.04 (0.13) 0.07 (0.19) 0.19 (0.20) 0.01 (0.17) 0.02 (0.11) 0.05 (0.17) 0.09 (0.13) 0.21 (0.17) 0.02 (0.14) 0.03 (0.14) 0.07 (0.15) 0.11 (0.17) 0.41 (0.16) 0.04 (0.16) 0.10 (0.12) 0.28 (0.17) 0.48 (0.14) 0.59 (0.19) 0.07 (0.20) 0.23 (0.15) 0.42 (0.20) 0.83 (0.15) 0.76 (0.13) 0.64 (0.17) 0.48 (0.11) 1.00 (0.18) 0.88 (0.14) DOutboard 0 (0.14) 0.00 (0.16) 0.02 (0.11) 0.01 (0.10) 0.01 (0.14) 0.01 (0.12) 0.00 (0.17) 0.05 (0.13) 0.01 (0.18) 0.02 (0.18) 0.03 (0.13) 0.00 (0.11) 0.05 (0.19) 0.02 (0.11) 0.02 (0.12) 0.03 (0.18) 0.00 (0.18) 0.06 (0.14) 0.02 (0.19) 0.03 (0.19) 0.04 (0.14) 0.01 (0.15) 0.12 (0.15) 0.04 (0.13) 0.09 (0.16) 0.15 (0.19) 0.01 (0.20) 0.18 (0.11) 0.08 (0.12) 0.13 (0.20) 0.24 (0.18) 0.01 (0.11) 1.00 (0.19) 0.15 (0.17) 0.28 (0.15) 0.26 (0.19) DCenter 0 (0.17) 0.00 (0.10) 0.01 (0.14) 0.05 (0.15) 0.01 (0.19) 0.02 (0.13) 0.00 (0.18) 0.02 (0.20) 0.11 (0.12) 0.04 (0.14) 0.06 (0.18) 0.00 (0.14) 0.02 (0.15) 0.13 (0.15) 0.04 (0.20) 0.07 (0.12) 0.00 (0.17) 0.02 (0.14) 0.15 (0.17) 0.05 (0.17) 0.08 (0.14) 0.00 (0.16) 0.03 (0.10) 0.33 (0.16) 0.18 (0.19) 0.31 (0.12) 0.00 (0.12) 0.05 (0.17) 0.59 (0.15) 0.27 (0.19) 0.48 (0.15) 0.00 (0.11) 0.27 (0.13) 1.00 (0.15) 0.54 (0.14) 0.51 (0.13) DInboard 0 (0.13) 0.00 (0.13) 0.00 (0.16) 0.01 (0.16) 0.04 (0.13) 0.00 (0.13) 0.00 (0.16) 0.01 (0.19) 0.02 (0.12) 0.09 (0.12) 0.02 (0.14) 0.01 (0.17) 0.01 (0.17) 0.02 (0.19) 0.11 (0.17) 0.02 (0.11) 0.01 (0.19) 0.01 (0.17) 0.02 (0.18) 0.13 (0.14) 0.03 (0.19) 0.01 (0.14) 0.02 (0.15) 0.05 (0.14) 0.36 (0.18) 0.18 (0.15) 0.01 (0.17) 0.03 (0.10) 0.09 (0.10) 0.51 (0.12) 0.31 (0.18) 0.02 (0.16) 0.19 (0.15) 0.18 (0.15) 1.00 (0.20) 0.33 (0.14) DRoot 0 (0.15) 0.00 (0.13) 0.00 (0.16) 0.00 (0.16) 0.01 (0.12) 0.05 (0.18) 0.00 (0.15) 0.00 (0.19) 0.00 (0.19) 0.01 (0.18) 0.13 (0.16) 0.00 (0.17) 0.00 (0.15) 0.00 (0.20) 0.01 (0.18) 0.16 (0.11) 0.00 (0.12) 0.00 (0.12) 0.00 (0.13) 0.01(0.14) 0.19 (0.17) 0.00 (0.11) 0.00 (0.10) 0.01 (0.11) 0.06 (0.16) 0.62 (0.16) 0.00 (0.18) 0.01 (0.18) 0.04 (0.13) 0.12 (0.19) 0.94 (0.12) 0.01 (0.11) 0.15 (0.12) 0.12 (0.12) 0.40 (0.10) 1.00 (0.17) 2233
level of 0.05. It can be noted that the GFS gives a success rate of 100% for debonding/delamination detection up to a noise level of 0.05, except for two rules, moderate D/D at root and severe D/D at root. However, for a higher noise level of 0.1, the system success rate starts falling rapidly for the rules, which dene matrix cracking towards the blade tip zone, and for the rules, which dene moderate and severe D/D at root. During the analysis of misclassication of the rules [34], the reason for the sudden fall in success rates at higher noise levels is that the GFS is unable to isolate the damage levels due to higher noise in the data. However, it is observed that even at higher noise levels, the GFS can isolate damage locations accurately. This is the rst work towards development of an online SHM system for the composite helicopter rotor blades. For implementation of the SHM system on helicopters, for global GFS, the sensors to measure tip deection and root forces, and for local GFS, the sensors to measure the strains at ve locations are required. The data obtained from these sensors can be compared with the data base of measurements from the undamaged blade at a given trim condition. Further, using the data reduction algorithm, the measurement deltas can be calculated and input to the GFS to predict the damage level of the blade. As this SHM system directly gives linguistic output, the maintenance norms can be developed for giving direct instructions to maintenance engineers.
ARTICLE IN PRESS
2234 P.M. Pawar, R. Ganguli / Mechanical Systems and Signal Processing 21 (2007) 22122236 Table 8 Success rate of various rules for local damage detection using strain-based measurement deltas No. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 Rule Undamaged Small CD at tip Small CD at outboard Small CD at centre Small CD at inboard Small CD at root Moderate CD at tip Moderate CD at outboard Moderate CD at centre Moderate CD at inboard Moderate CD at root High CD at tip High CD at outboard High CD at centre High CD at inboard High CD at root Very high CD at tip Very high CD at outboard Very high CD at centre Very high CD at inboard Very high CD at root Slight D/D at tip Slight D/D at outboard Slight D/D at centre Slight D/D at inboard Slight D/D at root Moderate D/D at tip Moderate D/D at outboard Moderate D/D at centre Moderate D/D at inboard Moderate D/D at root Severe D/D at tip Severe D/D at outboard Severe D/D at centre Severe D/D at inboard Severe D/D at root Avg. Min. S R0:03 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 99.90 100.00 100.00 100.00 100.00 99.90 99.99 99.90 SR0:05 100.00 100.00 100.00 100.00 100.00 100.00 98.60 99.60 99.90 100.00 100.00 99.90 100.00 99.80 100.00 100.00 99.30 97.70 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 92.90 100.00 100.00 100.00 100.00 91.80 99.43 91.80 SR0:10 100.00 100.00 100.00 100.00 100.00 100.00 81.10 81.50 86.00 95.50 99.90 47.20 84.20 68.40 96.10 83.80 84.10 69.20 89.90 93.50 99.40 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 74.90 86.00 100.00 100.00 100.00 73.90 91.52 47.20
7. Conclusions An automated online SHM approach for damage detection in composite helicopter rotors is presented in this paper. An aeroelastic analysis is used to obtained simulated blade response, loads and strain measurements for a damaged rotor. These simulated rotor system parameters are then contaminated with noise to simulate real data and the deviations (deltas) between the damaged and undamaged parameters are used to develop a GFS. Global and local damage detection is investigated. The GFS for global damage detection is developed using blade response and loads-based measurements. The GFS for local damage detection is developed using the strain-based measurement deltas. The following conclusions are drawn from the study. 1. The success rate of the GFS for global damage detection depends on number of measurements, type of measurements, training noise level and testing noise level.
ARTICLE IN PRESS
2. As displacement-based measurement deltas show approximately monotonic behaviour with respect to the matrix cracking and debonding/delamination, the fuzzy rules for global damage detection are dened based on the displacement trend. 3. The GFS developed at a training noise level of 0.15 for global damage detection using displacement-based measurement deltas gives good success rates up to a noise level of 0.05. The use of load-based measurement deltas in the GFS leads to good performance up to a noise level of 0.1. When both loads and response are used as measurements, the GFS shows good performance up to a noise level of 0.2. 4. Even though the GFS for global damage detection gives a slightly better success rate for the training noise level of 0.05 using all measurement deltas, the training noise level of 0.15 is a judicious selection to avoid the sudden fall in success rate due to high errors in a few measurement deltas. 5. The GFS developed for local damage detection gives good results up to a noise level of 0.05. For higher noise level, success rates of the GFS developed using strain-based measurement delta starts falling drastically because this system is unable to isolate the damage levels. However, it is observed that the system can isolate the matrix crack location accurately even at higher noise level.
References
[1] J.D. Cronkhite, Practical application of health and usage monitoring (HUMS) to helicopter rotor, engine and drive system, Proceedings of the American Helicopter Society 49th Annual Forum, May 1921, 1993, St. Louis, MO, USA, pp. 14451455. [2] J. Land, C. Weitzman, How HUMS have the potential of signicantly reducing the direct operating costs for modern helicopters through monitoring, Proceedings of the American Helicopter Society 51st Annual Forum, May 911 1995, Fort Worth, TX, USA, pp. 744757. [3] P.D. Samuel, D.J. Pines, A review of vibration-based techniques for helicopter transmission diagnostics, Journal of Sound and Vibration 282 (12) (2005) 475508. [4] J. Alkahe, O. Rand, Analytic extraction of the elastic coupling mechanisms in composite blades, Composite Structures 49 (4) (2000) 399413. [5] G.L. Ghiringhelli, P. Masarati, P. Mantegazza, Analysis of an actively twisted rotor by multi-body global modelling, Composite Structures 52 (1) (2001) 113122. [6] E.V. Morozov, S.A. Sylantiev, E.G. Evseev, Impact damage tolerance of laminated composite helicopter blades, Composite Structures 62 (34) (2003) 367371. [7] H. Azzam, M.J. Andrew, The use of math-dynamic model to aid the development of integrated health and usage monitoring, Proceedings of the Institution of Mechanical Engineers, Part G: Journal of Aerospace Engineering 206 (G1) (1992) 7176. [8] R. Ganguli, I. Chopra, D.J. Haas, Simulation of helicopter rotor-system structural damage, blade mistracking, friction and freeplay, Journal of Aircraft 35 (4) (1998) 591597. [9] M. Yang, I. Chopra, D.J. Haas, Sensitivity of rotor-fault-induced vibrations to operational and design parameters, Journal of the American Helicopter Society 49 (3) (2004) 328339. [10] R. Ganguli, Health monitoring of helicopter rotor in forward ight using fuzzy logic, AIAA Journal 40 (12) (2002) 27732781. [11] P.M. Pawar, R. Ganguli, Genetic fuzzy system for damage detection in beams and helicopter rotor blades, Computer Methods in Applied Mechanics and Engineering 192 (1618) (2003) 20312057. [12] J. Alkahe, O. Rand, Y. Oshman, Helicopter health monitoring using an adaptive estimator, Journal of the American Helicopter Society 48 (3) (2003) 199210. [13] H.Y. Hwang, C. Kim, Damage detection in structures using a few frequency response measurements, Journal of Sound and Vibration 270 (12) (2004) 114. [14] N. Roy, R. Ganguli, Helicopter, rotor blade frequency evolution with damage growth and signal processing, Journal of Sound and Vibration 283 (35) (2005) 821851. [15] J. Cattarius, D.J. Inman, Experimental verication of intelligent fault detection in rotor blades, International Journal of Systems Science 31 (11) (2003) 13751379. [16] A. Ghoshal, J. Harrison, M.J. Sundaresan, D. Hughes, M.J. Schulz, Damage detection testing on a helicopter exbeam, Journal of Intelligent Material Systems and Structures 12 (5) (2001) 315330. [17] J. Kiddy, D. Pines, Experimental validation of a damage detection technique for helicopter main rotor blades, Proceedings of the Institution of Mechanical Engineers, Part I Journal of Systems and Control in Engineering 215 (I3) (2001) 209220. [18] K.A. Lakshmanan, D.J. Pines, Damage identication of chordwise crack size and location in uncoupled composite rotorcraft exbeams, Journal of Intelligent Material Systems and Structures 9 (2) (1998) 146155. [19] P.M. Pawar, R. Ganguli, On the effect of matrix cracks in composite helicopter rotor blade, Composites Science and Technology 65 (34) (2005) 581594. [20] P.M. Pawar, R. Ganguli, Modeling multi-layer matrix cracking in thin-walled composite rotor blade, Journal of the American Helicopter Society 50 (3) (2005) 354366.
ARTICLE IN PRESS
2236 P.M. Pawar, R. Ganguli / Mechanical Systems and Signal Processing 21 (2007) 22122236 [21] P.M. Pawar, R. Ganguli, Modeling progressive damage accumulation in thin walled composite beams for rotor blade applications, Composites Science and Technology 66 (13) (2006) 23372349. [22] P.M. Pawar, R. Ganguli, On the effect of progressive damage on composite helicopter rotor system behaviour, Composite Structures, in press, doi:10.1016/j.compstruct.2005.11.043. [23] A.K. Garg, D.R. Mahapatra, S. Suresh, S. Gopalakrishnan, S.N. Omkar, Estimation of composite damage model parameters using spectral nite element and neural network, Composites Science and Technology 64 (16) (2004) 24772493. [24] D. Chakraborty, Articial neural network based delamination prediction in laminated composites, Materials and Design 26 (1) (2005) 17. [25] S.F. Yuan, L. Wang, G. Peng, Neural network method based on a new damage signature for structural health monitoring, ThinWalled Structures 43 (4) (2005) 553563. [26] Z.Q. Su, L. Ye, Quantitative damage prediction for composite laminates based on wave propagation and articial neural networks, Structural Health Monitoring 4 (1) (2005) 5766. [27] N. Ramanujam, T. Nakamura, M. Urago, Identication of embedded interlaminar aw using inverse analysis, International Journal of Fracture 132 (2) (2005) 153173. [28] A. Iwasaki, A. Todoroki, Delamination identication of CFRP structure by discriminant analysis using Mahalanobis Distance, Advances in Nondestructive Evaluation, PT 1-3 Key Engineering Materials 270273 (13) (2004) 18591865. [29] M.A. Rao, J. Srinivas, B.S.N. Murthy, Damage detection in vibrating bodies using genetic algorithms, Computers and Structures 82 (1112) (2004) 963968. [30] S.A. Ramu, V.T. Johnson, Damage assessment of composite structuresa fuzzylogic integrated neural network approach, Computers and Structures 57 (3) (1995) 491502. [31] R. Ganguli, A fuzzy logic system for ground based structural health monitoring of a helicopter rotor using modal data, Journal of Intelligent Material Systems and Structures 12 (6) (2001) 397407. n, F. Gomide, F. Herrera, F. Hoffmann, L. Magdalena, Genetic Fuzzy Systems, Evolutionary Tuning and Learning of [32] O. Cordo Fuzzy Knowledge Bases, World Scientic, Singapore, 2001. n, F. Gomide, F. Herrera, F. Hoffmann, L. Magdalena, Ten years of genetic fuzzy systems: current framework and new [33] O. Cordo trends, Fuzzy Sets and Systems 141 (1) (2004) 531. [34] P.M. Pawar, R. Ganguli, Matrix cracking detection in thin-walled composite beam using genetic fuzzy system, Journal of Intelligent Material Systems and Structures 16 (5) (2005) 381468. [35] G. Bir, et al., University of Maryland Advanced Rotorcraft Code (UMARC) Theory Manual, UM-AERO Report 92-02 1992. [36] R. Ganguli, Optimum design of a helicopter rotor for low vibration using aeroelastic analysis and response surface methods, Journal of Sound and Vibration 258 (2002) 327344. [37] R. Chandra, I. Chopra, Structural response of composite beams and blades with elastic couplings, Composite Engineering 2 (56) (1992) 347374. [38] R. Ganguli, I. Chopra, Aeroelastic optimisation of a helicopter rotor with two-cell composite blades, AIAA Journal 34 (4) (1996) 835841. [39] E. Adolfsson, P. Gudmundson, Thermoelastic properties in combined bending and extension of thin composite laminates with transverse matrix cracks, International Journal of Solids and Structures 34 (16) (1997) 20352060. [40] G.P. Cleveland, C. Trammel, Integrated health and usage monitoring system for the SH-60B helicopter, Proceedings of American Helicopter Society 52nd Annual Forum, American Helicopter Society, Alexandria, VA, USA, 1996, pp. 17671787. [41] T.J. Miles, M. Lucas, N.A. Halliwell, S.J. Rothberg, Torsional and bending vibration measurements on rotors using laser technology, Journal of Sound and Vibration 266 (3) (1999) 441467. [42] P. Castellini, M. Martarelli, E.P. Tomasini, Laser Doppler vibrometry: development of advanced solutions answering to technologys needs, Mechanical Systems and Signal Processing 20 (6) (2006) 12651285. [43] R.H. Cabell, C.R. Fuller, W.F. OBrien, Neural network modelling of oscillatory loads and fatigue damage estimation of helicopter components, Journal of Sounds and Vibration 209 (2) (1998) 329342. [44] H. Azzam, A practical approach for the indirect prediction of structural fatigue from measured ight parameters, Proceedings of the Institution of Mechanical Engineers, Part G: Journal of Aerospace Engineering 211 (G1) (1997) 2938. [45] B. Kosko, Fuzzy Engineering, Prentice-Hall, NJ, 1997.

1

Diunggah oleh

Informasi Dokumen

Hak Cipta

Format Tersedia

Bagikan dokumen Ini

Bagikan atau Tanam Dokumen

Opsi Berbagi

Apakah menurut Anda dokumen ini bermanfaat?

Apakah konten ini tidak pantas?

Hak Cipta:

Format Tersedia

1

Diunggah oleh

Hak Cipta:

Format Tersedia

ARTICLE IN PRESS

Mechanical Systems and Signal Processing

y (0)3 (45)4 (90) (90) (45)4 (0)3

(0)6 (45)4 (90)4 (90)4 (45)4 (0)6 z 0.35c D-Spar

NACA0012 with chord (c)=12 in.

Fig. 2. Details of two-cell airfoil section beam.

Fuzzy Inference Engine

Damage detection , location and size

Fuzzy Rule Base Fuzzy Logic System (FLS)

0 0.8 1 0.85 0.9 0.94

0 0.8 1 0.85 0.9 0.94

0 0.8 0.85 0.9 0.94 Effective strain ratio

10 9 8 7 6 5 4 3 2 1 0 0 0.5 1 1.5 crack density 2 2.5 3

0.2 V 0.15 V and W W

2500 2000 Fx , Fy and Fz 1500 1000 500 0 -500 -1000 -1500 0

2000 1000 0 -1000 -2000 -3000 -4000 -5000 Mx My Mz

1.5 crack density

1.5 crack density

Effective strain ratio

Effective strain ratio

0.2 0.16 0.12 0.08 0.04 0 0 1 2 3 4 5 T O C O R

0.35 0.3 peak peak shear strain T O C O R

peak peak shear strain

0.25 0.2 0.15 0.1 0.05 0 -0.05 0

gage location ( ESR = 0.88 )

gage location ( ESR = 0.92 )

then the system will output class C j provided that

DESIGN PROCESS Genetic Algorithm Based Learning Process

Computation with fuzzy systems

57.42 45.38 100.00 27.80 21.10 100.00

80.70 100.00 100.00 45.80 100.00 100.00

Anda mungkin juga menyukai