F16 Icing Identification Based on Neural Networks

IFAC DECOM-TT 2004 Automatic Systems for Building the Infrastructure in Developing Countries Copyright © IFAC October 3 - 5, 2004 Bansko, Bulgaria AUTOMATIC SYSTEMS FOR BUILDING THE INFRASTRUCTURE IN DEVELOPING COUNTRIES Bansko, Bulgaria, 2004     

F16 ICING IDENTIFICATION BASED ON NEURAL NETWORKS 

2 Rahmi AYKAN1, Fikret CALISKAN , and Chingiz HAJIYEV3 

 

1Turkish

Airlines Technic, Engineering Department, Gate B, Yesilkoy, 34850, Istanbul, Republic of Turkey, Tel: 90-212-6636300-ex.9565; E-mail: [email protected]

2Istanbul

Technical University Electrical Engineering, Ayazaga, 34469, Istanbul, Republic of Turkey, Tel: 90-212-2853683; E-mail: [email protected]

3Istanbul

Technical University Aeronautical Engineering, Ayazaga 34469, Istanbul, Republic of Turkey, Tel: 90-212-2853105; E-mail: [email protected] 



Abstract: This study aims at the identification of in-flight wing icing of an F16 aircraft by using neural networks trained with flight data and by observing the changes of parameters affected by icing. In the light of the previous research on in-flight icing, five parameters are assumed to be affected and so identified. In order to obtain training data set for neural network model, F16 aircraft analytical model is simulated in the time-varying manner. With several simulations, the best neural network model of the F16 aircraft is obtained. The applied tests show that neural network model satisfactorily represents iced F16 aircraft. In this research, icing identification based on neural networks is applied for the first time to F16 aircraft. Copyright © 2004 IFAC 

Keywords: Parameter Identification, Neural Networks, Fault Detection, Aircraft Control     

1. INTRODUCTION 

The recent improvements and research on aviation have focused on the subject of aircraft safe flight even in the severe weather conditions. As one type of such weather conditions, aircraft icing has been found considerably negative effect on the aircraft flight performance. Furthermore, this phenomenon has resulted in several fatal accidents. The risks of the aircraft icing encountered during flight at freezing temperatures and humid air have been known early 1900s. Although it has been seen that icing could be hazardous at every flight phase, take-off and landing have been affected most importantly. In addition, ice may occur on wings, control surfaces, horizontal and vertical stabilizers, fuselage nose, landing gear doors, engine intakes, fuselage air data ports and sensors and drain system outputs. This study examines the only wing icing occurrences. Icing during grounding of aircraft is out of this study.

NASA has performed several flight test for in-flight icing of the aircraft DHC-6 Twin Otter since 1986. Ratvasky and Ranauda (1993) obtained very useful data regarding the effects of aircraft icing to aircraft stability and control. As soon as aircraft icing was announced as a prior issue in 1997, NASA established a team called Icing Research Group. Bragg, Perkins, Sarter, Baúar, Voulgaris, Gurbacki, Melody and McCray (1999) from Illinois University have investigated aircraft icing from several different viewpoints and proposed a Smart Icing System. Miller and Ribbens (1999) tried to detect tail icing by evaluating the decrease of elevator effectiveness via Failure Detection Filter. In another application, these researchers used a state estimator as a type of Luenberger Observer. These studies showed that icing detection via statistical error analysis of states was more effective than online parameter estimation (1999). With NASA support, Ratvasky and Zante (1999) examined experimentally and analytically the effects of tail icing. Bragg et al. (2000) proposed a method for flight envelope protection by identifying icing characterization. Melody, Baúar, Perkins, and

Voulgaris (1999) applied H-infinity algorithm to icing identification problem. They claimed that proposed method are better than least square estimation methods and Extended Kalman Filter methods. Schuchard, Melody, Baúar, Perkins and Voulgaris (2000) have worked on tail icing detection and classification by estimating icing affected parameters and sensor information via neural networks. Johnson and Rokhsan (2000) have proposed a method detecting icing via neural networks and Kohonen Self Organizing Maps (SOMs). By observing neural network connection weights’ changes, they have tried to find iced and clean aircraft model via SOMs. In that research, the effects of atmospheric turbulences and elevator input signal to icing identification were presented. With respect to identification of degradation in aerodynamic parameters and characteristics of flight dynamics due to aircraft icing, Dynamic Icing Detection System (DIDS) was proposed by Myers, Klyde and Magdeleno (2000). Bragg et al. (2001) used hinge moment sensors in order to detect icing on control surfaces. They improved a neural network model to estimate stability and control derivatives (Bragg, 2002). In this research, icing identification based on neural networks is applied to F16 aircraft. A proposed neural network is trained with nine measurements. 2. AIRCRAFT ICING In-flight icing decreases the aerodynamic quality of aircraft such that aircraft weight increases, drag increases, lift decreases, and hence the effectiveness of angle of attack and pitch angle change. The experimental studies have showed that, in the result of wing icing, drag may increase to values of 500%, and lift may decrease to values of 40% (Bragg et al. 1999; Whalen et al., 2002). The effect on moments may vary. Accordingly, the effectiveness of control surfaces may decrease. All these directly affect aircraft safe flight. As well as this subject is clearly important in the respect of aviation safety, by taking into account extra fuel consumption due to icing, it is important for economical reasons. By including icing effects, the control and stability of more correct aircraft model improve flight performance, passenger comfort and tight flight plans. It is clearly obvious that military aircraft have to fly at all places on air and weather conditions as possible. On the other hand, intensive air traffic has forced the aircraft could fly at all weather conditions. Civil aviation authorities and other organizations such as Federal Aviation Authority, FAA, and Joint Aviation Authority, JAA, which provide aircraft certificates and quality assurances, have restricted the flight of the aircraft which are not installed antiicing system and icing detection system. In order to make sure that whether aircraft is safe on icing weather conditions, or not, the flight tests are mandated by these organizations prior to first aircraft

approval. These flight tests are too time-consuming and expensive. Instead of these tests, flight simulations of exactly modelled iced aircraft by using modern technology products would be better for aircraft manufacturers. At least, these simulations could support to flight tests data. Some icing sensors in nose sections are used on some modern aircraft to detect in-flight icing. However, these sensors only show an indication or a possibility for icing when it comes to some levels. They do not measure the icing effects such as its shape, thickness and location. It is impossible to evaluate the degradation of aircraft performance due to in-flight wing and tail icing. Hence, the existing sensors do not provide enough information to pilot or autopilot. By an improved icing detection, monitoring and control system, pilot/aircraft system can safely continue its route regardless of weather icing conditions within acceptable safety margins. On the subject of aircraft in-flight icing detection and evaluation, there have been not enough academic research for the time being. Especially, some icing related accidents in Turkey and worldwide during last 20-25 years, and additional requirements of civil aviation authorities have forced researchers to work on this subject in deeper. The several works related aircraft is continuing with the support of NATO, NASA and some universities. The special working groups have been assigned by these research centers for icing.

3 .F16 AIRCRAFT DYNAMIC MODEL By an improved icing detection, monitoring and control system, pilot/aircraft system can safely continue its route regardless of weather icing conditions within acceptable safety margins. In-flight icing detection and identification is applied to an unstable multi-input multi-output model of an AFTI/F-16 fighter. The fighter is stabilized by means of a linear quadratic optimal controller. The control gain brings all eigenvalues that are outside the unit circle, inside the unit circle. It also keeps the mechanical limits on the deflections of control surfaces. The model of the fighter is as follows (Lyshevski, 1997): x (k + 1) = Ax(k ) + Bu(k ) + F( x (k )) + Gw(k )

(1)

where A is the system matrix with the dimensions 9x9, B is the control matrix with the dimensions 9x5, F is a vector with the dimensions 9x1 of nonlinear part of the system, and G is the process noise transition matrix of F16 fighter for the sampling period of 0.03 sec. The aircraft state variables are: x = [v Į q ș ȕ p r I < ]T

where, v is the forward velocity, D is the angle of attack, q is the pitch rate, T is the pitch angle, E is the side-slip angle, p is the roll rate, r is the yaw rate, I is the roll angle, and \ is the yaw angle. The fighter has six control surfaces and hence six control inputs are: u = [į HR

į HL

į FR

į FL

įC

accumulation generally increases the drag parameters, here C D D , it decreases moment and lift parameters, here C L D , C L q , C M D , and C M q (Melody et al. 2001; Bragg et al. 1999 and 2002).

į R ]T

where GHR and GHL are the deflections of the right and left horizontal stabilizers, GFR and GFL are the deflections of the right and left flaps, GC and GR are the canard and rudder deflections. w(k) is the model noise vector, which is E[w(k)] = 0. We assume the following hard bounds (mechanical limits) on the deflections of control surfaces: ~GHR, GHL~d 0.44 rad, ~GFR, GFL~d 0.35 rad, ~GC~d 0.47 rad and ~GR~d 0.52 rad. The measurement equations can be written as: z(k) = H x(k) + v (k)

tail. For most conventional aircraft, it is generally true that the wing contribution 85-90 percent to the value of C L D (McLean, 1990). As well as ice

Stability and control derivatives are usually found from wind-tunnel tests at first. Unfortunately, unavoidable differences between test environment and flight conditions, the wind-tunnel test data are considered only as initial estimates. In this study, as being in the previous research on icing, the changes of other derivatives are assumed small and negligible. When aircraft linearized equations are examined, it is easily found that all these derivatives are in the matrix A shown in the section 3. Icing affected parameters of F16 aircraft dynamic model are A(1,2), A(2,2), A(2,3), A(3,2) and A(3,3), which could be written as follow:

(2) A(1,2) = k1 (C D D

where z(k) is a vector of nine measurements, H is the measurement matrix, which is 9x9 unit matrix, v (k) is the measurement disturbance, and its mean and correlation matrix respectively are: E[v(k)] = 0;

T

E[v(k) v (j)] = R(k)G(kj)

4. PARAMETERS AFFECTED BY THE ICING As explained in the second chapter of this paper, icing results in decreasing aircraft aerodynamic performance, which are affected by changes in lift, drag and pitch moment, and their effectiveness with regard to aircraft position angles and velocities. In common representatives of aircraft linearized dynamic equations, this effect may be reflected by stability and control derivatives. Especially, the researches in NASA Icing Research Group and Icing Institute of Illinois University (Melody et al. 2001) have showed that the most affected parameters from in-flight wing icing are followings: CDD

wC D , wD CM D

CLD wC M , wD

wC L , CLq wD wC M CM q wq

wC L , wq

where CD CL CM

: drag coefficient : lift coefficient : pitching moment coefficient

The change in lift coefficient with a change in angle of attack, C L D , often called the lift curve slope. The lift curve slope for the total airframe includes the components due to the wing, the fuselage, and the

CL )

A(2,2) = k 2 (C L D + C D )

(3) (4)

A(2,3) = k 3C L q

(5)

A(3,2) = k 4 C M D

(6)

A(3,3) = k 5C M q

(7)

where k i , i = 1,2,3,4,5, consist of all other flight parameters which are considered constant for a certain time. These constants may be calculated from certain flight conditions such as take off, climb, cruise, and landing. Total aircraft speed, wing reference area, wing chord length, and aircraft inertial moment per aircraft pitch axis must be provided to calculate these constants. In the simulations at section 6, A(1,2), A(2,2), A(2,3), A(3,2), and A(3,3) are expressed as a12, a22, a23, a32, and a33, respectively. In this study, duration of two minutes, four ice-affected parameters are assumed to decrease their halves, and one parameter to increase fifty percent more. 5. DESIGN OF NEURAL NETWORK MODEL TO ESTIMATE THE PARAMETERS Neural networks have increasingly been shown as viable tools for mapping nonlinear systems and for the purpose of parameter identification. It is very efficient method in the analysis of nonlinear and complex models if enough data are available for its training phase. Unfortunately, icing in flight occurs in many different ways, and there is no enough training data available regarding stability and control derivatives. There are little data only for a few research aircraft obtained from tunnel test or flight test. After enough data are picked up from other methods, neural networks may be used effectively for control. This study aims to find the stability and

The neural network has many interesting, complex, and attractive features such as, parallel processing, learning, self-organizing, nonlinear capabilities. Neural networks have inherent parallel properties that provide a robust and fault-tolerant structure. Networks are practical for aircraft applications because, following initial training, they process information very rapidly. Rapid computation can be achieved because the majority of mathematical operations involve addition, subtraction, or multiplication. A quick response in a certain time frame is especially critical for icing determination since ice accretion during flight at low altitude requires immediate action. Neural networks also have the capability to be trained on-line using real data or off-line with recorded or simulated data. In this study, since there are nine states measured and five parameters to be estimated, a neural network structure having nine inputs and five outputs is presented. Two hidden layers are proposed. These three layers have these activation functions, respectively: logarithmic, tangent and linear. This neural network is trained with the measurements obtained by simulating nonlinear unstable F16 model.

Fig. 1 and 2 show the history of aircraft model angle of attack and side slip angle history (actual-simulated values) together with their noisy measurements. Errors are at reasonable levels which modern sensors are capable to measure. Similar results are obtained for other seven states, which are forward speed, pitch rate, pitch angle, roll rate, yaw rate, roll angle, yaw angle. Fig. 3 shows the flap deflection. Other five control surface deflections are not shown due to the limit of available space. Fig. 4 shows the training details for five epochs, at which the proposed network has reached the desired error value, called goal. In fact the network almost generalizes the inputs after second epoch. This is very quick convergence rate because of the training algorithm. Fig. 5, 6, 7, and 8 show the training outputs of some parameters and all test outputs of five parameters. Similar training results are obtained for other parameters.

0.3 alpha and alpham

control derivatives of clear and iced configuration. By monitoring the flight data, changes in these derivatives are found, and a fault signal can be built up according to change level..

For training method the Levenberg-Marquardt Algorithm (LMA) is used to maintain second-order training speed. The aim in LMA is to shift towards Newton’s method as quickly as possible.

0.2 0.1 0 -0.1 -0.2

0

50

100

150

200

250

150

200

250

error

0 -0.1

iteration

Fig. 1. Angle of attack history, simulated and measured, in radian

beta and betam

0.3 0.2 0.1 0 -0.1

0

50

100

0

50

100

150

200

250

150

200

250

0.1

error

0.05 0 -0.05 -0.1

In the all of the graphics, subscripts “e”, “12”, “22”, “23”, “32”, and “33” are not written in their exact place for convenience. For instance, “ve” instead of “ve”, “a12” instead of “a12” and so on.

100

0.1

-0.2

For the batch size of 6 sec, 201 data sets are obtained by simulating F16 aircraft model. Data sets comprise five parameters and corresponding nine states. 101 data sets are used for training, 50 data are used for validation, and 50 data are used for testing.

50

0.2

6. SIMULATIONS The proposed method is applied to iced model configuration such that ice is occurring during the simulated time frame. It is assumed that five parameters affected by icing are linearly changing due to wing leading edge icing. Whilst one of them, related to drag and angle of attack, is increasing, the other four are increasing at the same gradient. Training, validation and testing are performed for only iced F16 model. Batch size for training is chosen as 6 seconds such that icing should be detected within a certain time frame before it is hazardous.

0

iteration

Fig. 2. Side-slip angle history, simulated and measured, in radian

0.05 0.1 init a22 error

0.04

right flap

0 -0.1

0.03 0.02 0.01

-0.2 -0.3

0 0

50

100

150

200

250 2 final a22 error

left flap

0.3 0.2 0.1

50

100

150

200

40

60

80

100

120

20

40

60 iteration

80

100

120

x 10

1 0 -1 -2

0

20 -5

0.4

0

0

0

250

iteration

Fig. 6. Training output error history of parameter a22

Fig. 3. Flap deflection angle history, in radian

initial a22

0.96

Performance is 8.78532e-007, Goal is 1e-006

1

0.94

0.92

10

test

0.9

0

10

0

5

10

15

20

25

30

35

40

45

50

0

5

10

15

20

25 iteration

30

35

40

45

50

0.94

-1

10

validation trained a22

Network error

-2

10

-3

training

10

0.92

0.9

-4

10

0.88 goal

-5

10

Fig. 7. Test output history of parameter a22

-6

10

-7

10

0

0.5

1

1.5

2

2.5 5 Epochs

3

3.5

4

4.5

5

Fig. 4. Neural network training, validation and test performance

0.032

initial a23

0.0315 0.031 0.0305 0.03 0.0295

0

5

10

15

20

25

30

35

40

45

50

0

5

10

15

20

25 iteration

30

35

40

45

50

trained a23

0.032

0.94

initial a22

0.92

0.031 0.0305 0.03

0.9 0.88 0.86

0.0315

Fig. 8. Test output history of parameter a23 0

20

40

60

80

100

120

trained a22

0.94

7. CONCLUSION AND COMMENTS

0.93 0.92 0.91 0.9

0

20

40

60 iteration

80

100

Fig. 5. Training output history of parameter a22

120

In-flight icing affects several aircraft dynamic parameters. In order to evaluate icing effect on aircraft performance these parameters are to be calculated. Five parameters primarily affected by icing are taken in accordance with the previous study results. In order to identify in-flight icing, aircraft should have sufficient dynamical properties or certain level model noises. Determination of stability derivatives affected by icing is very difficult. As a nonlinear F-16 aircraft model is simulated in a time

dependent manner by entering changed stability derivatives for a certain time which ice is detectable, the necessary training samples can be collected. A suitable neural network can estimate uncertain stability derivatives. This method is one of off-line estimations. This may be performed as on-line in future studies. The more model noise exists, the less validation noise becomes, but training performance gets worse. Too many numbers of neurons in the network decrease the generalization of the network. Some validation samples outside the border of the training data values would have much more errors and result in rough estimation of ice-affected parameters. In this research, icing identification based on neural networks is applied for the first time to F16 aircraft model. The proposed method may be applied for more than five parameters of aircraft model. With this method, aircraft simulation model can be obtained from the flight data downloaded from aircraft computer, called Flight Data Recorder or Black Box. REFERENCES Bragg, M.B., Perkins, W.R., Sarter, N.B., Baúar, T., Voulgaris, P.G., Gurbacki, H.M., Melody, J.W., and McCray, S.A. (1998). An interdisciplinary approach to in-flight aircraft icing safety, in Proc. 36th AIAA Aerospace Sciences Meeting and Exhibit, Reno, NV, AIAA-98-0095. Bragg, M.B., Hutchison, T., Oltman, R., Pokhariyal, D. and Merritt, J. (2000). Effect of ice accretion on aircraft flight dynamics, in Proc. 38th AIAA Aerospace Sciences Meeting and Exhibit, Reno, NV, AIAA-2000-0360. Bragg, M. B., Perkins, W.R., Basar, T., Sarter, N. B., Voulgaris, P. G., Selig, M., and Melody, J. (2002). Smart Icing Systems for Aircraft Icing Safety, Reno NV, AIAA-2002-0813. Broeren, E. P., Eddy, H. E, and Bragg, M. B. (2002). Effect of Intercycle Ice Accretions on Airfoil Performance, AIAA-2002-0240. Caliskan, F., Hajiyev, C. (2003) Actuator Failure Detection And Reconfigurable Control For F-16 Aircraft Model, IFAC Automatic Systems for Building the Infrastructure in Developing Countries, Istanbul. Campa, G., Napolitano, M.R., Seanor, B., Fravolini, M.L., Song, Y. (2002). Application of an Improved LWR Method to Real-Time Aircraft Parameter Identification Problems, American Control Conference, Anchorage, AK, USA. Campa, G., Fravolini, M.L., Napolitano, M.R. (2002) A Library of Adaptive Neural Networks for Control Purposes, IEEE International Symposium on Computer Aided Control System Design, Glasgow, Scotland, UK. Gurbacki, M. H. and Bragg, M. B. (1999). Sensing Aircraft Icing Effects by Flap Hinge Moment Measurement, Norfolk VA, AIAA-99-3149. Gurbacki, H.M. and Bragg, M.B. (2001). Sensing Aircraft Icing Effects by Unsteady Flap Hinge-

Moment Measurement, Journal of Aircraft, Vol. 39, No. 3, pp. 575-577. Jackson, D.G., Bragg, M.B. (1999). Aerodynamic Performance of an NLF Airfoil with Simulated Ice, AIAA 99-0373. Johnson, M.D., Rokhsaz, K. (2000). Using Artificial Neural Networks And Self Organizing Maps for Detection of Airframe Icing, The 2000 Atmospheric Flight Mechanics Conference, AIAA-2000-4099. Lyshevski, S.E. (1997). State-space identification of nonlinear flight dynamics, In proceedings of the Conference on Control Applications, Hartford, Connecticut, pp. 496-498. McLean, D. (1990). Automatic Flight Control Systems, Prentice Hall, Cambridge, UK. Melody, J.W., Pokhariyal, D., Merret, J., Baúar, T., Bragg, M.B. (2001). Sensor Integration for Inflight Icing Characterization Using Neural Networks, 39th Aerospace Science Meeting and Exhibit, Reno, Nevada, AIAA-2001-0542. Melody, J.W., Hillbrand, T., Baúar, T., Perkins, W.R. (2001). H-Infinity Parameter Identification for In-flight Detection of Aircraft Icing: The Time Varying Case, IFAC Control Engineering Practice, 1327-1335. Myers, T.T., Klyde, D.H., Magdaleno, R.E. (2000). The Dynamic Icing Detection System, 38th AIAA Aerospace Sciences Meeting and Exhibit, Reno, NV. Ribbens, W., Miller, R.H. (1999). Detection of Icing and Related Loss of Control Effectiveness in Regional and Corporate Aircraft, 37th Aerospace Sciences, AIAA-99-0637. Whalen, E., Lee, S., Ratvasky, T. (2002). Characterizing the Effect of Ice on Aircraft Performance and Control From Flight Data, Aerospace Science Meeting and Exhibit, Reno, NV, AIAA-2002-816.