Anda di halaman 1dari 12

965

Adaptive neuro-fuzzy controllers for an open-loop morphing wing system


T L Grigorie, R M Botez , and A V Popov cole de Technologie Suprieure, Laboratory of Research in Active Controls, Avionics and AeroServoElasticity LARCASE, Montral, Quebec, Canada The manuscript was received on 7 December 2008 and was accepted after revision for publication on 10 June 2009. DOI: 10.1243/09544100JAERO487

Abstract: A new method for the realization of two neuro-fuzzy controllers for a morphing wing design application is presented here. The controllers main function is to correlate each set of pressure differences, calculated between the optimized and the reference airfoil, with each of the airfoil deformations produced by the actuators system. The pressures are calculated at different chord positions and will also be measured during wind tunnel tests. During a rst identication phase, the two fuzzy inference systems (FISs) from the controllers structure are generated for 16 ight conditions characterized by Mach numbers and angles of attack. Next, the FIS are optimized with the Matlab function adaptive neuro-fuzzy inference system (ANFIS) by training over different epochs. Finally, the controllers are validated for the other 33 ight conditions of the open-loop morphing wing system. This is the rst time that such a method of relating the pressure differences to airfoil displacements has been conceived and used in an open-loop morphing wing controller system. Keywords: morphing, neuro-fuzzy controller, simulation, modelling, testing

INTRODUCTION

Recently, morphing wing system studies have branched out into new research directions. Extremely complex and catalogued as inter- and multidisciplinary studies, morphing wing studies continue to push the science, up to the extreme boundaries of mathematics and physics. These multidisciplinary studies therefore require knowledge in the following disciplines: aerodynamics and computational uid dynamics, aeroelasticity, automatic control, intelligent materials, signal detection using the latest miniaturized sensors, high computer-time calculations, wind tunnel and ight testing, instruments, and signal acquisition these signals have such speed that they are raising serious problems for the existing calculus technology [1]. Consequently, real-time system functioning is conditioned (in addition to other factors) by obtaining

Corresponding

author: Dpartement de Gnie de la Production,

Quebec University, 1100, rue Notre-Dame Ouest, Montral, Quebec H3C 1K3, Canada. email: ruxandra.botez@etsmtl.ca
JAERO487

the best data processing algorithms, easy to implement software within the command and control unit. Fuzzy logic theories, which offer remarkable facilities, may therefore be used in these algortihms. They facilitate signal processing by allowing empirical models to be designed based on experimental data; thus, the complex mathematical calculus currently in use can be avoided. In addition, fuzzy logic can be used to model highly non-linear, multidimensional systems, including those with parameter variations, or where the sensors signals are not accurate enough for other models [2, 3]. In order to conceive such a model, a fuzzy set must be designed, which may be given by the original fuzzy logic theory conceived by Lot A. Zadeh. The most serious problem arises from the determination of a complete set of rules and the membership functions corresponding to each input. The many attempts to reduce errors and to optimize the model are time-consuming and, very often, the results are far from what was expected. A modern design method allows fuzzy model design to be completed in a relatively short time interval. The adaptive neuro-fuzzy inference system (ANFIS) design technique allows the generation and optimization of the set of rules and the
Proc. IMechE Vol. 223 Part G: J. Aerospace Engineering

966

T L Grigorie, R M Botez, and A V Popov

membership functions parameters by use of neural networks [4]. Moreover, the ANFIS design technique already implemented in Matlabs neuro-fuzzy software tools should be relatively easy to use. In the following paragraphs are mentioned some of the most important applications of neural network and fuzzy logic theories in the aerospace eld. Neural network methods have been used in the aeronautical industry for the following applications: the detection and identication of structural damage [5], fault diagnostics (detection and isolation) leading to compensation of control surface failures [6], the modelling of aerodynamic characteristics from ight data [7, 8], generalized reference models for six degrees of freedom motion simulation using global aerodynamic models, including unsteady aerodynamics and dynamic stall [911], the detection of unanticipated effects such as icing [12], and autopilot controllers and advanced control laws for applications such as carefree manoeuvring [13, 14]. Neural network methods have also been used for model identication purposes, based on ight utter tests. A good example is the method proposed in reference [15] for the prediction of damping ratios during ight utter tests using a neural network trained on model data. That method was compared to a simple statistical extrapolation approach and was focussed on studying noise effects. To date, fuzzy logic theories have been used less extensively than neural network theories for the design of active control systems and for the identication and validation of an aircraft model. A new method for the conversion of aerodynamic forces from the frequency to the Laplace domain was validated on an F/A-18 aircraft for aeroservoelasticity studies [16]. The F/A-18 aircraft non-linear model was identied from ight utter tests by use of fuzzy logic [17] and a combination of fuzzy logic and neural network methods [18]. A neuro-fuzzy controller design, simulation, and validation for a morphing wing application, using

Fuzzy Logic Toolbox and Matlab/Simulink, is presented in this study. A hybrid learning algorithm is used to tune the fuzzy inference systems (FISs) from the controllers structure; the least-squares and the back propagation gradient descent methods are combined for the FIS membership functions (mf ) parameters training to emulate a given training dataset. Another dataset is used to validate and test the robustness of the neuro-fuzzy controllers. 2 OPEN-LOOP ARCHITECTURE OF THE MORPHING WING SYSTEM

The functional diagram of the morphing wing system using an implemented controller is presented in Fig. 1. The reference airfoil is a modied WTEA type. The two shape memory alloy (SMA) actuators that deect the airfoils (from the reference to the optimized) are shown in the lower left side of Fig. 1, and the neurofuzzy controllers algorithm conceived here are shown in the upper right corner of Fig. 1. It is important to mention that the present study is entirely numerical and is done with the aim to integrate it into the real-time controller, to be used and then validated in a wind tunnel. Thirty pressure sensors are installed on the wing at different chord positions, as shown in Figs 1 and 2. These sensors are used to detect the airfoil transition points position from their pressure data [1]. Their installation is shown non-dimensionally on the wing airfoil between its leading and trailing edges in Fig. 2. The equipment presented in Figs 1 and 2 will be installed soon in a wind tunnel; experimental data will be compared to numerical results obtained with the present authors new algorithm for 49 ight cases dependent on seven attack angles varying between 1 and 2 , and on seven Mach numbers between 0.2 and 0.35. An optimized airfoil was determined for each of the 49 ight cases. The airfoil deformation is realized by using two SMA actuators installed at 25.3 and 47.6 per cent from its

Fig. 1

Open-loop architecture of the morphing wing system with the neuro-fuzzy controllers
JAERO487

Proc. IMechE Vol. 223 Part G: J. Aerospace Engineering

Adaptive neuro-fuzzy controllers for an open-loop morphing wing system

967

1000 800 600 dp [N/m2] 400 200 0 -200


alpha1 alpha2 alpha3 alpha4 alpha5 alpha6 alpha7

Fig. 2

Pressure sensors arrangement on the morphing airfoil

-400 -600 0 5 10 15 20 Sensor number 25 30

leading edge. For each ight case, the actuation system with two SMA actuators must ensure the two reference airfoil vertical displacements dY1 and dY2 until the optimized airfoil is obtained. For each ight case, the pressures detected by the 30 sensors on the reference airfoil and on the optimized airfoil are determined, and then one set of 30 pressure differences between the optimized and the reference airfoil (dp1 dp30 ) is calculated. In addition, for each ight case, the two vertical displacements of the reference airfoil with respect to the optimized airfoil (dY1 and dY2 ) correspond to each set of 30 pressure differences (dp1 dp30 ), and are generated by the actuation system. The role of the neuro-fuzzy controllers is to nd a correspondence between each set of pressure differences detected by the 30 sensors and each of the two displacements given by the two SMA actuators. Therefore, two controllers are required: dY1 = f1 (dp1 , dp2 , . . . , dp30 ) and dY2 = f2 (dp1 , dp2 , . . . , dp30 ). The two controllers identication is performed for 16 ight cases, while their validation is performed for 33 ight cases. Thus, neuro-fuzzy controllers are built for all 49 ight cases. 3 FUZZY LOGIC CONTROLLERS DESIGN

Fig. 3

dp1 dp30 for M = 0.2 and various

1500
alpha1 alpha2 alpha3 alpha4 alpha5 alpha6 alpha7

1000

dp [N/m2]

500

-500

-1000

10

15 20 Sensor number

25

30

Fig. 4

dp1 dp30 for M = 0.225 and various

1500
alpha1 alpha2 alpha3 alpha4 alpha5 alpha6 alpha7

1000

For all 49 ight cases, the pressure differences (dp1 dp30 ) between the optimized and the reference airfoils are traced for each sensor, as shown in Figs 3 to 9, for seven incidence angles: = 1 , 0.5 , 0 , 0.5 , 1 , 1.5 , and 2 and for seven Mach numbers: M = 0.2 (Fig. 3), M = 0.225 (Fig. 4), M = 0.25 (Fig. 5), M = 0.275 (Fig. 6), M = 0.3 (Fig. 7), M = 0.325 (Fig. 8), and M = 0.35 (Fig. 9). The vertical deection differences (dY1 and dY2 ) are depicted in Fig. 10 as functions of Mach number for each of the seven incidence angles. By analyzing the pressures differences (dp1 dp30 ) and vertical position differences (dY1 and dY2 ) for all 49 ight cases, 16 identication cases (I1 I16 ) and 33 validation cases (V1 V33 ) are chosen for the controllers identication and validation; these cases are shown in Table 1.
JAERO487

dp [N/m2]

500

-500

-1000

10

15 20 Sensor number

25

30

Fig. 5

dp1 dp30 for M = 0.25 and various

The 16 identication cases are selected in order to obtain the most robust controllers, which would contain the generated FISs covering the maximum domain for the 30 pressure differences (dp1 dp30 )
Proc. IMechE Vol. 223 Part G: J. Aerospace Engineering

968

T L Grigorie, R M Botez, and A V Popov

2000 1500 1000 dp [N/m2] 500 0 -500


alpha1 alpha2 alpha3 alpha4 alpha5 alpha6 alpha7

3000 2500 2000 1500 dp [N/m2] 1000 500 0 -500


alpha1 alpha2 alpha3 alpha4 alpha5 alpha6 alpha7

-1000 -1500

-1000
0 5 10 15 20 Sensor number 25 30

-1500

10

15 20 Sensor number

25

30

Fig. 6
2500 2000 1500 1000 dp [N/m2] 500 0 -500 -1000 -1500 0

dp1 dp30 for M = 0.275 and various

Fig. 8
3500

dp1 dp30 for M = 0.325 and various

dp [N/m2]

alpha1 alpha2 alpha3 alpha4 alpha5 alpha6 alpha7

3000 2500 2000 1500 1000 500 0 -500 -1000 -1500

alpha1 alpha2 alpha3 alpha4 alpha5 alpha6 alpha7

10

15 20 Sensor number

25

30

10

15 20 Sensor number

25

30

Fig. 7

dp1 dp30 for M = 0.3 and various

Fig. 9

dp1 dp30 for M = 0.35 and various

and for the vertical displacements between airfoils (dY1 and dY2 ). As shown in Table 2, each identication case is chosen because it maximizes or minimizes the values of dp1 dp30 , dY1 and dY2 . The minimum and maximum values of the indexed parameter are denoted with indices of min and max, respectively.
8 7 6 dY1 [mm] 5 4 3 2 0.2

The maximum and minimum values of dp1 dp30 , dY1 and dY2 are covered by 15 identication cases (with no case I9 ), which can be observed in Table 2. From Table 1, it is obvious that these 15 cases cover all of the values of Mach number and incidence angles except for M = 0.225 and = 1 . For this reason,
alpha1 alpha2 alpha3 alpha4 alpha5 alpha6 alpha7

8 7 6 dY2 [mm] 5 4 3 2 0.2

0.22

0.24

0.26 0.28 0.3 Mach number

0.32

0.34

0.36

0.22

0.24

0.26 0.28 0.3 Mach number

0.32

0.34

0.36

Fig. 10

dY1 and dY2 as functions of M for various


JAERO487

Proc. IMechE Vol. 223 Part G: J. Aerospace Engineering

Adaptive neuro-fuzzy controllers for an open-loop morphing wing system

969

Table 1

Identication (I1 I16 ) (V1 V33 ) ight cases


1 0.2 2 0.225 V1 V5 V11 V16 I9 V28 V32 3 0.25 V2 V6 V12 V17 V22 I10 V33 4 0.275 I2 V7 V13 V18 V23 I11 I13

and

validation

Mach () 1 2 3 4 5 6 7 1 0.5 0 0.5 1 1.5 2

5 0.3 I3 I5 V14 V19 V24 V29 I14

6 0.325 I4 V8 V15 I8 V25 V30 I15

7 0.35 V3 V9 I6 V20 V26 I12 I16

The Sugeno fuzzy model was proposed by Takagi, Sugeno, and Kang to generate the fuzzy rules from a given inputoutput dataset [19]. For the system presented in this paper, a rst-order model is considered for both FISs, and for N rules is given by [19, 20]
1 1 1 Rule 1: If x1 is A1 and x2 is A2 and . . . and x30 is A30 , 1 1 1 then y 1 (x1 , x2 , . . . , x30 ) = b0 + a1 x1 + a2 x2 + 1 + a30 x30 ,

I1 V4 V10 I7 V21 V27 V31

. . .
i i i Rule i: If x1 is A1 and x2 is A2 and . . . and x30 is A30 , i i i then y i (x1 , x2 , . . . , x30 ) = b0 + a1 x1 + a2 x2 + i + a30 x30 ,

Table 2

Correlation of the identication cases with the extreme values of dp1 dp30 , dY1 , and dY2
Maximized or minimized parameters (dp1 dp30 , dY1 , and dY2 ) dp2min dp5min , dp11max dp20max dp1min dp30min , dp8max dp10max dp29min dY1min , dY2min dp28min dp6min dp7max dp22max , dp23max dp21max dp9min , dp12min , dp24max dp28max dY1max , dp7min , dp8min , dp10min dY2max dp21min , dp23min dp26min , dp29max , dp30max dp11min , dp13min dp20min , dp22min , dp27min , dp1max dp6max

Identication cases I1 I2 I3 I4 I5 I6 I7 I8 I10 I11 I12 I13 I14 I15 I16

. . .
N N N Rule N : If x1 is A1 and x2 is A2 and . . . and x30 is A30 , N N N then y N (x1 , x2 , . . . , x30 ) = b0 + a1 x1 + a2 x2 + N + a30 x30

(1) where xq (q = 1, 30) are individual input variables, i Aq (i = 1, N ) are the associated individual antecedent fuzzy sets of each input variable, and y i (i = 1, N ) is the rst-order polynomial function in the consequent. i ak (k = 1, 30, i = 1, N ) are parameters of the linear i function and b0 (i = 1, N ) denotes a scalar offset. The i i parameters ak , b0 (k = 1, 30, i = 1, N ) are optimized by the least-squares method. For any input vector x = [x1 , x2 , . . . , x30 ]T , if the singleton fuzzier, the product fuzzy inference, and the centre average defuzzier are applied, the output of the fuzzy model y is inferred as follows (weighted average)
N i=1

the I9 (M = 0.225 and = 1 ) identication case is considered. For the 16 identication cases, two Sugeno-type FISs are generated using the Matlab gens2 function [4]. Therefore, the working eld is separated into regions where, for each region, the local parameters to be processed are represented by a low-order linear model. The non-linear process is locally linearized around a functioning point by using the least-squares method, and the obtained model is considered valid in the entire region around this point. The limitation of the operating regions implies the existence of overlapping among these different regions so their denition is given in a fuzzy manner. Thus, for each model input, several fuzzy sets are associated with their corresponding denitions of their mf. By combining these fuzzy inputs, the input space is divided into fuzzy regions. For each of these regions, a local linear model is used, while the global model is obtained by defuzzication with the gravity centre method (Sugeno), which is used to interpolate the local models outputs [2, 4].
JAERO487

y=

w i (x)y i (2) w i (x)

N i=1

where w i (x) is the degree of fulllment of the antecedent; that is, the level of ring of the ith rule
i i i w i (x) = A1 (x1 ) A2 (x2 ) A30 (x30 )

(3)

The Matlab gens2 function generates Gaussian-type membership functions, dened as [3, 18]
i Aq (x) = exp0.5 i x cq i q 2

(4)

i i where cq is the cluster center and q is the dispersion of the cluster.

Proc. IMechE Vol. 223 Part G: J. Aerospace Engineering

970

T L Grigorie, R M Botez, and A V Popov

Inputs

Input mf
in1cluster1 ... ... ...

Rules

Output mf

Output

in1 dp1

in1clusterk ... in1cluster30 ... ...

out1cluster1

in1cluster1 ... ... ... inj dpj in1clusterk ... in1cluster30 ... ... k out1clusterk
Agregated Output1

out1

in1cluster1 ... ... in30 dp30 in1clusterk

30 ...

out1cluster30

in1cluster30

Thus, for the system presented here, a set of 16 rules is obtained, which are of the following type: if (in1 is in1clusterk) and (in2 is in2clusterk) and . . . and (inj is inj cluster k) and . . . and (in30 is in30clusterk) then (out1 is out1clusterk). For each input, out of the 30 inputs, 16 Gaussian-type mf are generated; within the set of rules, they are denoted by inj cluster k; where j is the input number (1 30), and k is the number of the membership function (1 16). The two FISs generated for dY1 and dY2 displacements, denoted by dY1Fis and dY2Fis, have the structure shown in Fig. 11, while the corresponding controllers have the structure presented in Fig. 12. 4 DESIGN OF ADAPTIVE FUZZY LOGIC CONTROLLERS

ANFIS error

0.075 0.07 0.065 0.06 0.055 0.05 0.045 1 2 3 4 5 6 7 Training epochs number 8 9 10

ANFIS error

ANFIS error

0.08 0.06 0.04 0.02 0 0 100 200 300 400 500 600 700 800 900 1000 Training epochs number

ANFIS error

Proc. IMechE Vol. 223 Part G: J. Aerospace Engineering

... ...
...

... ...

Fig. 11

...

Structure of the dY1Fis and the dY2Fis

...

in1

dp1

inj

dpj

Sugeno FIS

out1 dY1 or dY2

in30 dp 30

Fig. 12
0.095 0.09 0.085 0.08

Structure of the controllers

Adaptive fuzzy logic controllers can be obtained by optimizing the previously generated FISs (dY1Fis and dY2Fis). The neural networks are used to optimize the FIS mf parameters. Therefore, two neuro-fuzzy models (trained FISs) are obtained to control the dY1 and dY2 deections based on the dp1 dp30 pressure differences, by use of Matlabs ANFIS function [4]. Matlabs ANFIS function uses an adaptive learning algorithm to identify and modify the membership function parameters for the two Sugeno-type FISs
0.11 0.1 0.09 0.08 0.07 0.06 0.05 0.04 0.03 0.02 0 10 20 30 40 50 60 70 80 90 100

... ...
0.14 0.12 0.1

Training epochs number 0.14 0.12 0.1 0.08 0.06 0.04 0.02 0 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000 Training epochs number

Fig. 13 The ANFIS training errors for dY1Fis over 10, 102 , 103 , and 104 training epochs
JAERO487

Adaptive neuro-fuzzy controllers for an open-loop morphing wing system

971

0.1 0.09 0.08 ANFIS error ANFIS error 0.07 0.06 0.05 0.04 0.03 0.02 1 2 3 4 5 6 7 8 9 10

0.14 0.12 0.1 0.08 0.06 0.04 0.02 0

10

20

30

40

50

60

70

80

90

100

Training epochs number 0.14 0.12 0.1 ANFIS error ANFIS error 0.08 0.06 0.04 0.02 0 0.14 0.12 0.1 0.08 0.06 0.04 0.02 0

Training epochs number

100 200 300 400 500 600 700 800 900 1000 Training epochs number

1000 2000 3000 4000 5000 6000 7000 8000 9000 10000 Training epochs number

Fig. 14 The ANFIS training errors for dY2Fis over 10, 102 , 103 , and 104 training epochs
8
Data FIS model

8
Data FIS model

7 6 dY1 [mm] dY1 [mm] 0 2 4 6 8 10 12 14 16 5 4 3 2

7 6 5 4 3 2 0 2 4 6 8 10 12 14 16 All 16 airfoil identification cases for 10 training epochs All 16 airfoil identification cases for 100 training epochs 8
Data FIS model Data FIS model

8 7 6 dY1 [mm] 5 4 3 2 0 2 4 6 8 10 12 14 16 All 16 airfoil identification cases for 1000 training epochs dY1 [mm]

7 6 5 4 3 2 0 2 4 6 8 10 12 14 16 All 16 airfoil identification cases for 10000 training epochs

Fig. 15 The inputoutput data (dY1 ) and the outputs of trained dY1Fis over different training epochs
JAERO487 Proc. IMechE Vol. 223 Part G: J. Aerospace Engineering

972

T L Grigorie, R M Botez, and A V Popov

with 30 inputs and one output [4]. This is a hybrid algorithm that uses a combination of the gradient method along with the least-squares method. An inputoutput dataset and the FIS models generated with gens2 are considered as a starting point; then, the membership functions parameters are optimized with the ANFIS algorithm for a number of training epochs (set by the user). This optimization is realized by means of a quality parameter in the training algorithm [4], which is there in order to improve the process approximation performed by the neuro-fuzzy model. The data from the 16 identication cases are used as the inputoutput dataset. The training errors, dened by the quality parameter in the training algorithm, are shown in Figs 13 and 14 for the dY1Fis and the dY2Fis, respectively. These two FISs are trained over 10, 102 , 103 , and 104 training epochs. The post-training FIS evaluations are shown in Figs 15 and 16; the star shapes indicate the deections dY1 and dY2 from the inputoutput data, while the dY1 and dY2 deections obtained from the fuzzy models (the outputs of dY1Fis and dY2Fis trained FIS, respectively) are shown as circles.

Figures 13 and 14 show decreasing oscillations in the training errors over the rst 3200 training epochs for dY1Fis, with similar decreases over the rst 2700 training epochs for dY2Fis. After these intervals, oscillations with approximately constant amplitudes for both FISs can be observed. Therefore, a training of over 104 epochs gives the best results. Figures 15 and 16 reiterate the same observations as those obtained for Figs 13 and 14. There is an overlapping of the neuro-fuzzy models outputs and the dY1 and dY2 values from the inputoutput data. This superposition is dependent on the number of training epochs, and improves as the number of epochs increases. Table 3 presents the numerical values of the absolute maximum relative errors between the inputoutput data and their neuro-fuzzy models, for displacements dY1 and dY2 . The errors are evaluated over 10, 102 , 103 , and 104 training epochs. These errors decrease with the increase in the training epochs number; the minimum values of these errors are obtained for 104 training epochs: 0.5305 per cent for dY1Fis and 0.6753 per cent for dY2Fis.

8
Data FIS model

8
Data FIS model

7 6 dY2 [mm] dY2 [mm] 0 2 4 6 8 10 12 14 16 All 16 airfoil identification cases for 10 training epochs 8
Data FIS model

7 6 5 4 3 2 0 2 4 6 8 10 12 14 16 All 16 airfoil identification cases for 100 training epochs

5 4 3 2

9 8 7 dY2 [mm] 6 5 4
Data FIS model

7 6 dY2 [mm] 5 4 3 2 0 2 4 6 8 10 12 14 16 All 16 airfoil identification cases for 1000 training epochs

3 2 0 2 4 6 8 10 12 14 16 All 16 airfoil identification cases for 10000 training epochs

Fig. 16 The inputoutput data (dY2 ) and the outputs of trained dY2Fis over different training epochs
Proc. IMechE Vol. 223 Part G: J. Aerospace Engineering JAERO487

Adaptive neuro-fuzzy controllers for an open-loop morphing wing system

973

Table 3 The absolute maximum relative errors between the inputoutput data and their neuro-fuzzy models
Training epochs dY1Fis dY2Fis 10 2.2842% 2.1716% 100 1.1196% 1.1872% 1000 0.8657% 0.8165% 10 000 0.5305% 0.6753%

with six-pointed star symbols in Figs 17 and 18). These relative errors are obtained by training the original FISs, generated from the identication data, over 104 training epochs.

5 VALIDATION OF ADAPTIVE FUZZY LOGIC CONTROLLERS The validation and testing of the two controllers robustness is performed by using the Simulink diagram shown in Fig. 19. This diagram also shows the controllers implementation. The data to be validated are those corresponding to the 33 validation ight cases presented in Table 1 (V1 V33 ). Figures 20 and 21 show the controllers evaluation for these data, which consists of the validation data and the outputs of the two neuro-fuzzy models (the trained FISs) for all 33 validation ight cases. There is very good overlapping of these two data categories for both controllers, which is clearly visible in Figs 20 and 21. The values of the absolute relative errors are shown in Fig. 22, where they are at their maximum at 2.5158 per cent for controller 1 and at 1.0887 per cent for controller 2. The maximal values are obtained for the 20th validation case, V20 , for controller 1 and for the 23rd validation case , V23 , for controller 2.

The variations of the relative errors between the inputoutput data and their neuro-fuzzy models for all 16 identication cases are shown in Figs 17 and 18 the different training epochs 10, 102 , 103 , and 104 are considered for the FIS training. Both diagrams show that the training of the two FISs diminished the most relative error for case 6 (I6 M = 0.35, = 0 ) and the least for case 12 (I12 M = 0.35, = 1.5 ). The maximum values of the absolute relative error were found with case 4 (I4 M = 0.325, = 1 ) for dY1 and for case 11 (I11 M = 0.275, = 1.5 ) for dY2 . Two fuzzy logic controllers are conceived; their models give relative errors smaller than 0.7 per cent for all 16 identication ight cases (see the black curves
0.01 0.005 0

Relative errors

-0.005 -0.01

dY1 Fuzzy Logic Controller 1 To Workspace

dp
-0.015 -0.02 -0.025
10 training epochs 100 training epochs 1000 training epochs 10000 training epochs

Signal From Workspace Fuzzy Logic Controller 2

dY2 To Workspace1

4 6 8 10 12 All airfoil 16 identification cases

14

16

Fig. 17
0.015 0.01

Relative errors for dY1Fis

Fig. 19

Simulink implementation of the two controllers

7 6.5 6
Data FIS model

0.005

Relative errors

5.5 dY1 [mm]


10 training epochs 100 training epochs 1000 training epochs 10000 training epochs

0 -0.005 -0.01 -0.015 -0.02 -0.025

5 4.5 4 3.5 3 2.5

4 6 8 10 12 All airfoil 16 identification cases

14

16

10 15 20 25 All 33 airfoil validation cases

30

35

Fig. 18
JAERO487

Relative errors for dY2Fis

Fig. 20 The validation of controller 1


Proc. IMechE Vol. 223 Part G: J. Aerospace Engineering

974

T L Grigorie, R M Botez, and A V Popov

6.5 6 5.5 5 dY2 [mm] 4.5 4 3.5 3 2.5 0 5 10 15 20 25 All 33 airfoil validation cases 30 35
Data FIS model

Fig. 21 The validation of controller 2


0.03 0.025 0.02
controller1 controller2

0.5305 per cent for the rst controller (corresponding to the displacement dY1 ) for the fourth identication case, I4 (M = 0.325, = 1 ), and 0.6753 per cent for the second controller (corresponding to the displacement dY2 ) for the 11th identication case, I11 (M = 0.275, = 1.5 ) (see the black curves with six-pointed star symbols in Figs 17 and 18, respectively). The validation of the two controllers led to maximum absolute values of the relative errors that are less than 2.6 per cent, more precisely, 2.5158 per cent for controller 1 and 1.0887 per cent for controller 2. The maximum values were obtained for the 20th validation case, V20 , (M = 0.35, = 0.5 ) for controller 1 (black curve with circle symbols in Fig. 22) and for the 23rd validation case, V23 , (M = 0.275, = 1 for controller 2 (grey curve with square symbols in Fig. 22). Finally, it can be concluded that the results presented here are extremely useful for morphing wing applications.

ACKNOWLEDGEMENTS The authors would like to thank Dr Mahmoud Mamou and Dr Mahmood Khalid from the NSERC (National Sciences and Engineering Research Council) for modelling the WTEA-TE1 airfoils, and Dr Octavian Trifu for the 49 optimized airfoils shapes. The authors also want to thank Mr George-Henri Simon from Thales Avionics for initiating this project, and Mr Philippe Molaret from Thales Avionics and Mr Eric Laurendeau from Bombardier Aeronautics for their collaboration on this article. The authors want to acknowledge the nancial support of Thales Avionics and Bombardier Aerospace, as well as that of CRIAQ (Consortium of Research in the Aerospace Industry in Quebec), and NSERC (National Sciences and Engineering Research Council of Canada). Authors 2009

Relative errors

0.015 0.01 0.005 0 -0.005 -0.01 -0.015 0 5 10 15 20 25 All airfoil 33 validation cases 30 35

Fig. 22 The relative errors obtained when validating the controllers for the 33 validation ight cases

From Fig. 22, it can be seen that the absolute values of the relative errors exceed 1 per cent in three validation cases for controller 1 and in two validation cases for controller 2, while most of the absolute values of the relative errors remain smaller than 0.5 per cent; therefore, it can be said that the two controllers are robust and accurate. 6 CONCLUSIONS

REFERENCES
1 Popov, A. V., Botez, R., and Labib, M. Transition point detection from the surface pressure distribution for controller design. J. Aircr., 2008, 45(1), 2328. 2 Sivanandam, S. N., Sumathi, S., and Deepa, S. N. Introduction to Fuzzy Logic using MATLAB, 2007 (Springer, Berlin Heidelberg). 3 Kosko, B. Neural networks and fuzzy systems a dynamical systems approach to machine intelligence, 1992 (Prentice Hall, New Jersey). 4 Matlab Fuzzy Logic and Neural Network Toolboxes Help. Available from http://www.mathtools.net/ MATLAB/Books/Neural_Network_and_Fuzzy_Logic/. 5 Tsou, P. and Shen, M. H. H. Structural damage detection and identication using neural networks. AIAA J., 1994, 32(1), 176183.
JAERO487

Two robust neuro-fuzzy controllers were obtained for a morphing wing application, which relates the airfoil displacements caused by the actuator system to the pressure differences detected by the sensors system on the reference airfoil and on the optimized airfoils. The two controllers were trained for 16 identication ight cases and validated for 33 other ight cases. The nal FISs training was performed on 104 training epochs. The maximum absolute values of the relative errors among the identication data and those obtained for the controllers outputs were
Proc. IMechE Vol. 223 Part G: J. Aerospace Engineering

Adaptive neuro-fuzzy controllers for an open-loop morphing wing system

975

6 Rauch, H. E., Kline-Schoder, R. J., Adams, J. C., and Youssef, H. M. Fault detection, isolation, and reconguration for aircraft using neural networks. AIAA paper 1993-3870, 1993. 7 Linse, D. J. and Stengel, R. F. Identication of aerodynamic coefcients using computational neural networks. J. Guid. Control and Dyn., 1993, 16(6), 10181025. 8 Amin, S. M., Gerhart,V., and Rodin, E. Y. System identication via articial neural networks: applications to online aircraft parameter estimation. AIAA paper 97-5612, 1997. 9 Faller, W. E. and Shreck, S. J. Neural networks: applications and opportunities in aeronautics. Prog. Aerosp. Sci., 1996, 32, 433456. 10 Rockhsak, K. and Steck, J. E. Use of neural networks in control of high-alpha manoeuvres. J. Guid. Control Dyn., 16, 934939. 11 Scharl, J. and Mavris, D. Building parametric and probabilistic dynamic vehicle models using neural networks. AIAA paper 2001-4373, 2001. 12 Johnson, M. D. and Rokhsak, K. Using articial neural network and self-organizing maps for detection of airframe icing. J. Aircr., 2001, 38(2), 224230. 13 Napolitano, M. R. and Kincheloe, M. On-line learning neural network controllers for autopilot systems. J. Guid. Control Dyn., 1995, 33(6), 10081015. 14 Yavrucuk, I., Prasad, J. V. R., and Calise, A. Adaptive limit detection and avoidance for carefree manoeuvring. AIAA paper 2001-4003, 2001. 15 Crowther, J. W. and Cooper, E. J. Flight test utter prediction using neural networks. Proc. IMechE, Part G: J. Aerospace Engineering, 2001, 215(1), 3747. DOI: 10.1243/0954410011531736. 16 Hiliuta, A., Botez, R. M., and Brenner, M. Approximation of unsteady aerodynamic forces Q(k, M ) by use of fuzzy techniques. AIAA J., 2005, 43(10), 20932099. 17 Kouba, G., Botez, R. M., and Boely, N. Identication of F/A-18 model from ight tests using the fuzzy logic method. In Proceedings of the 47th AIAA Aerospace Sciences Meeting, Orlando, Florida, USA, 58 January 2009. 18 Boely, N., Botez, R. M., and Kouba, G. Identication of an F/A-18 nonlinear model between control and structural deections. In Proceedings of the 47th AIAA Aerospace Sciences Meeting, Orlando, Florida, USA, 58 January 2009.

19 Mahfouf, M., Linkens, D. A., and Kandiah, S. Fuzzy Takagi-Sugeno Kang model predictive control for process engineering, 1999, p. 4 (IEE, London, UK). 20 Kung, C. C. and Su, J. Y. Afne Takagi-Sugeno fuzzy modelling algorithm by fuzzy c-regression models clustering with a novel cluster validity criterion. IET Control Theory Appl., 2007, 1(5), 12551265.

APPENDIX Notation
i ak i Aq i b0 i cq dpj

dY1

dY2

FIS Il M N Vm wi x xq y yi i q

parameters of the linear function (k = 1, 30, i = 1, N ) associated individual antecedent fuzzy sets of each input variable (i = 1, N ) scalar offset (i = 1, N ) cluster centre (q = 1, 30) jth sensor pressure differences between the optimized and the reference airfoil ( j = 1, 30) vertical displacement between optimized and reference airfoil produced by the rst SMA vertical displacement between optimized and reference airfoil produced by the second SMA fuzzy inference system lth identication case (l = 1, 16) Mach number number of rules mth validation case (m = 1, 33) degree of fulllment of the antecedent, i.e. the level of ring of the ith rule input vector individual input variables (q = 1, 30) output of the fuzzy model rst-order polynomial function in the consequent (i = 1, N ) airfoil incidence angle dispersion of the cluster

JAERO487

Proc. IMechE Vol. 223 Part G: J. Aerospace Engineering

Anda mungkin juga menyukai