Vous êtes sur la page 1sur 7

Modified delta method for estimation of parameters

from flight data of stable and unstable aircraft

Sanjay Singh A K Ghosh


Amity Institute of Aerospace Engineering Department of Aerospace Engineering
Amity University Indian Institute of Technology Kanpur
Noida, India Kanpur, India
ssingh10@amity.edu akg@iitk.ac.in

Abstract-The aim of the study described herein was to modeling wherein no physical significance can be attached to
develop and verify an efficient neural network based either the network structure or the network weights [2]. The
method for extracting aircraft stability and control recent interest in, and fascinating with the evolving
derivatives from real flight data using feed-forward neural applications of ANNs to diverse fields such as signal
networks. The proposed method (Modified Delta method) processing, pattern recognition, system identification and
draws its inspiration from feed forward neural network control have led many researchers to explore their capabilities
based the Delta method for estimating stability and control for aircraft aerodynamic modeling and estimation of
derivatives. The neural network is trained using differential aerodynamic coefficients (stability and control derivatives).
Significant contributions have been made in this direction by
variation of aircraft motion/control variables and
Hess [1], Linse and Stengel [3], Youseff and Juang [4], and
coefficients, as the network inputs and outputs respectively.
Raol and Jategaonkar [6]. In all these papers, the emphasis has
For the purpose of parameter estimation, the trained been on aerodynamic modeling and estimation of aerodynamic
neural network is presented with a suitably modified input coefficients using FFNNs, without any attempt to estimate
file and the corresponding predicted output file of aircraft parameters as conventionally defined in terms of
aerodynamic coefficients is obtained. An appropriate stability and control derivatives in the equations of motion of
interpretation and manipulation of such input-output files an aircraft. Raol and Jategaonkar [6] have used the recurrent
yields the estimates of the parameter. The method is neural network (RNN) to model aircraft aerodynamics in a
validated first on the simulated flight data and then on real way that allows aircraft parameters to be estimated from
flights data obtained by digitizing analogue data from flight data. However, as the authors [6] pointed out, RNNs
published reports. A new technique is also proposed for have only a limited scope for aircraft identification
validating the estimated parameters using feed-forward applications and it is the FFNNs which may prove to be more
neural networks. flexible, and thereby to have a higher potential for future
applications for aircraft identification and parameter
Keywords— Modeling, Neural networks, Parameter estimation, estimation.
Validation
Only recently, Raisinghani, Ghosh and Kalra [7, 8]
proposed two new methods for explicitly estimating aircraft
I. INTRODUCTION parameters from the flight data using FFNNs. The results
A new thrust area is emerging in the area of aircraft obtained for simulated flight data and real flight data have
aerodynamic modeling and parameter estimation: shown the success and the potential of the proposed methods. It
development of techniques using artificial neural networks was further shown that the number of estimated values for each
(ANN) for flight vehicle identification. Recently artificial of the parameters was equal to the number of data points used
neural networks modeling has been attempted for aircraft for training the network. For real flight data, in addition to
dynamics where aircraft motion variables and control inputs training being less than perfect, the parameters may not be
are mapped to predict the total aerodynamic coefficients [1-4]. strictly constant, i.e., the parameters may vary slightly with
In the past, the most widely used parameter estimation other motion and control variables. Furthermore, all of the
methods have been equation error method, output error corrections and axes transformations done on the data would
method and filter error method. Applications of these methods introduce their own uncertainties. All these factors, contribute
require a priori postulations of an aircraft model. On the other toward different estimates at different time points. However, a
hand, a class of neural networks called the feed forward neural near normal distribution of the estimated values may be
networks (FFNNs) work as a general function approximators, observed for most of the parameters. For calculating standard
and are capable of approximating any continuous function to deviation of the estimates, the estimated parameters were first
any desired accuracy by an appropriate network structure [5]. arranged in descending/ascending order then 25 % of the
This ability of FFNNs has been utilized to model aircraft estimates were removed from both the ends of the order set
dynamics. However, the FFNNs lead to a black-box type of before calculating the mean and sample standard deviations [7,

978-1-4673-4529-3/12/$31.00 2012
c IEEE 775
8]. One of the reasons responsible for getting different II. MODIFIED DELTA METHOD
estimates at different time point through the Delta method was The proposed Modified Delta method is based on
attributed to training being less than perfect. Further a careful interpreting the stability and control derivatives as follows: If
look into the Delta method proposed in [7] reveals that it does we could obtain variation in the value of an aerodynamic
not suggest any procedure for validating the estimates by coefficient due to variation in only one of the motion/control
comparing estimated response (with the help of estimated variables while the variation in other motion/control variables
parameters) with the flight generated response (real flight data) are identically zero, then the ratio of the variation of the
for a known control input, other than used for generating real aerodynamic coefficient to variation of the non-zero
flight data for parameter estimation purpose. The scheme motion/control variable will yield the corresponding
proposed in [7] to calculate confidence level in the estimates stability/control derivative. Let us say that the FFNN is trained
does not work always, specially, if the distribution of the to map differential variations in input variables, ( Δα , Δq , and
numerical values estimated parameters are skewed. The
motivation to persue this work lies in improving the Delta Δδ e ) to the network output variable (variation), ΔCL . Now
method so that the estimated parameters have larger confidence one input (say Δα ) at a time is chosen to be at its original
bound (lesser scatter) and propose a methodology to validate value while the rest of the network inputs ( Δq , and Δδ e ) are
the extracted model by comparing the estimated response with set to zero. The predicted value of the aerodynamic coefficient
the flight response generated by a control input not used for
ΔCL corresponding to such a modified file is divided by the
estimation purpose. It is in this context that the present work
explores the suitability of the newly proposed Modified Delta non-zero variation in motion variable ( Δα ) to yield the
method [9] by applying it on simulated flight data as well as on corresponding stability/control derivatives, CLα .
flight data obtained via discretization of analog plots of real-
flight data of a stable [10] and an unstable aircraft [11]. Further Similarly, all the parameters can be estimated by suitably
the scheme to validate the estimated model using time histories modifying the input file. Fig. 1 schematically represents the
of measured motion/control variables with the help FFNN training strategy for application of the Modified Delta method
modeling has also been demonstrated. However, due to non using FFNN. A detail study was carried out for a few sets of
availability of complete set of flight data of a stable aircraft and simulated data to understand the influence of various network
an unstable aircraft, such FFNN based scheme (by passing the parameters on training and prediction capability. A matrix of
requirement of solving equations of motion) to validate tuning parameters such as the number of hidden layers, the
extracted model has been demonstrated using simulated flight number of nodes in each of the hidden layer(s), the learning
data only. The proposed scheme to validate the estimated rate, the momentum rate, the logistic gain factor of the
model using FFNN will increase the acceptability of the Delta sigmoidal function, the initial network weights, and the scaling
and the Modified Delta method by practicing engineers for of input-output data was generated, wherein each parameters
field application. The maximum likelihood (ML) estimator in were varied with in a prescribed range [1-3], and the network
its several forms has been the most widely and successfully was trained to arrive at the best possible set that led to
used for estimating aircraft stability and control parameters of minimum MSE for the given flight data. After the training, the
stable aircraft. However, for an unstable aircraft, difficulties are same input data are passed to check the prediction capability of
encountered because of 1) integration of equations of motion of the network. The predicted aerodynamic coefficients are
the open loop model generally resulting in diverging solutions, deemed acceptable only if the MSE is less than the specified
2) the potential for correlation between input and output value.
variables, 3) controllers suppressing the transients and thereby
reducing information content in measured signals [10, 11]. The III. SCHEME TO VALIDATE ESTIMATED MODEL USING FFNN
FFNNs based Delta method is shown to be a good alternative BASED METHOD
to the existing methods for estimating parameters of an
unstable aircraft [11]. One of the procedures to validate aircraft parameter
estimation method is to compare the estimated response
The work carried out in this paper is presented in seven (generated with estimated parameters) with the flight
sections. In section I, the subject of parameter estimation is measured response generated with control input other than
reviewed. The present status of application of FFNNs for used for estimation purpose. Only way to validate by
aerospace engineering problem is highlighted. Finally, the comparing motion variables, generated using new control
motivation and objective of the present work are outlined. input, would require solving of equations of motion. In this
Section II and III present in detail the method used for paper a scheme using FFNN to validate the estimated model
estimation and validation of estimates using neural network. by comparing flight measured variables generated using
Section IV describes briefly simulated flight data used for different control inputs (not used for estimation) is presented.
parameter estimation via the proposed method along with the
Delta method. The implementation of the proposed method for
parameter estimation and validation using neural network has
been discussed in chapter V and VI. The paper is concluded in
Section VII with a summary of work, strengths and limitations
of the work, and possibilities of future applications and
expansion of the work.

776 2013 3rd IEEE International Advance Computing Conference (IACC)


Input
™ f Cme = Cme + Cme α 2 + Cme ( q2 c / 2u ) + Cme δ e2 (4)
Δα 2 0 α q δ e

Δq Where q is the pitch rate, rad/sec, Į is theangle-of-attack,


ΔC L
rad, įe is the elevator deflection, rad. The already FFNN
™ f or
trained neural model is now used to predict estimated motion
ΔC m
Output Layer
variables (α e )t +1 , and ( qe )t +1 corresponding to new control
Δδ e input containing time histories of (α 2 )t , ( q2 )t , (δ e2 )t ,
™ f
Inputs Layer Hidden Layer
(CLe )t , (Cme )t as shown in Fig. 2b. Estimated responses of
2 2
Fig.1. Schematic of feed forward neural network for proposed aerodynamic (α e )t +1 , ( qe )t +1 are then compared with measured responses
modeling.
of (α 2 )t +1 , ( q2 )t +1 generated using control inputs δ e2 .
This proposed scheme does not require solving of
equations of motion for validation. As the first step, the
longitudinal parameters CLe , C Le , CLe , CLe , Cm
e e
, Cm ,
0 α q δe 0 α
e
Cm , and Cme are estimated by applying Modified Delta
q δe
method on flight data generated by a known control inputs
(say δ e1 ). For training, a neural mapping between input vector
containing time histories (at time, t , s) of (α1 )t , ( q1 )t ,
(δ e1 )t , (C L1 )t , (Cm1 )t , and the output vector containing time
histories (at time, t + 1 , s) of (α1 )t +1 , ( q1 )t +1 , is established Fig.2. Scheme for validation of proposed method using FFNN; a) training, b)
using back propagation algorithm (BPA) as shown in Fig. 2a. validation.
For the case of real flight data, the (C L1 )t , and (Cm1 )t would
be computed using the measured value of acceleration a z and IV. GENERATION OF SI,ULATED FLIGHT DATA
q through the following equations: For generating longitudinal simulated data of a stable
aircraft, the example aircraft chosen is the stable aircraft used
CL1 = − 2 m a z ρ u 2 S (1) in [12]. The stability and control parameters of the chosen
example aircraft are listed in Table I. For simulated
Cm1 = 2 q I y ρ u 2 S c (2) longitudinal flight data of an unstable aircraft, the chosen
stable aircraft was modified by assigning a positive value of
Where C L and C m are the lift and moment coefficients, 0.15 to Cmα . Rest of the parameters, inertia characteristics
az is measured acceleration along z body axis, meter per were kept same as the chosen example stable aircraft. An
second square, c is the mean aerodynamic chord, meter, Iy is unstable aircraft would generally have feedback control
the moments of inertia about y axis, kilogram-meter square, m system. We assume that feedback, if any, is turned off and
is mass of aircraft, kilogram, q is the pitch acceleration, radian only open-loop response is considered. For both the stable and
unstable aircraft, simulated flight data were generated for the
per second square, S is the reference wing area, meter square,
short period mode of an aircraft given by [13]:
u is the aircraft speed, meter per second square, ȡ is the air
density, kilogram per meter cube. α = q − ( ρ u S 2 m ) C L (5)
Once the training has been established by satisfying the
MSE criteria, a new set of real flight data generated through ( )
q = ρ u 2 S c 2 Iy Cm (6)
flight using a different control input, say δ e2 (other than used
for estimation purposes) is used for validation. The time CL = CL0 + CLα α + CLq ( qc / 2u ) + CLδ δ e (7)
e
histories of motion variables α 2 , q2 corresponding to this
control input designated δ e2 are recorded. However, the Cm = Cm0 + Cmα α + Cmq ( qc / 2u ) + Cmδ e δ e (8)

estimated numerical value of CLe , and Cme


are computed where ȡ = 1.076 kilogram per meter cube, u = 204.0 meter
2 2
per second, S = 180.78 meter, c = 4.663 meter, m = 130642.3
using estimated values of estimated parameters through the
kilogram, Iy = 8677233.1 kilogram meter square. The stability
following equations:
derivatives CL0 , CLα … Cmδ were assigned the true values
CLe = CLe + CLe α 2 + CLe
2 0 α q
( q2 c / 2u ) + CLeδ δ e2 (3) as given in Table I to II. The values of CL0 , Cm0 were
e
assigned zero for unstable aircraft. The fourth-order Runga-

2013 3rd IEEE International Advance Computing Conference (IACC) 777


Kutta method was employed to integrate Eqs. (5) and (6), and
the time histories of α and q were obtained for a multi-step 3-
2-1-1 elevator input signal of maximum amplitude 5.73o for
stable aircraft and 1.43o for unstable aircraft. If real flight data
of an unstable aircraft were to be generated by separate
surface excitation with feedback off, then control input
employed for data generation have to be both limited duration
and small amplitude [11]. The responses in α and q were
obtained for duration of 7 s at an interval of 0.01 and the
corresponding values of aerodynamic coefficients CL and
Cm were obtained for the true values of parameters as shown
in Table I to II. The true values of parameters are required and
used for the sole purpose of generating simulated flight data.
The FFNN based methods (both the Delta and Modified Delta)
do not require even an order of magnitude information about Fig.3. Simulated flight data used for; parameter estimation ( − ) and
validation ( ⋅ ⋅ ⋅ )
the parameters. This is in contrast to the conventional methods
like the ML estimator, where such information is required in
the form on initial values. It may be clarified here that in the TABLE I. ESTIMATED PARAMETERS FROM SIMULATED FLIGHT DATA OF
case of real flight data, since the true values of the parameters A STABLE AIRCRAFT
are not known, CL and Cm is to be calculated from the
Noise = 0 % Noise = 5%
measured values of a z and q using the Eqs. (1) and (2). Once Parameters True DM MDM DM MDM
the flight data is simulated, the input vector to FFNN was CLα 2.922 2.922 2.911 2.900
2.922
(0.01)* (0.008) (0.024) (0.004)
constructed to have Δα , Δq , and Δδ e as its elements, and 14.699 14.700 14.657 14.690
CLq 14.700
ΔCL or ΔCm constituted the output vector for FFNN (0.042) (0.017) (0.709) (0.133)
training. Once the training was achieved to an acceptable CLδ 0.431 0.434 0.417 0.445
0.435
(0.016) (0.005) (0.015) (0.004)
limit, Modified Delta method was applied for parameter 1.660 1.660 1.610 1.680
estimation. The simulated flight data containing information - Cm 1.660
α (0.001) (0.0003) (0.031) (0.002)
about angle of attack, pitch rate, linear acceleration, and pitch - Cm 34.771 34.750 34.854 34.704
34.750
acceleration for a given elevator input, used for parameter q (0.026) (0.0063) (0.771) (0.077)
estimation and validation is presented in Fig. 3. - Cm 2.578
2.581 2.577 2.663 2.504
δ (0.005) (0.001) (0.105) (0.098)

V. PARAMETER ESTIMATION FROM SIMULATED FLIGHT


C L0 0.340 0.340 0.340 0.333 0.344
DATA Cm0 0.252 0.251 0.252 0.274 0.232
The parameters to be estimated are the longitudinal *
Standard deviation
parameters CL0 , CLα , CLq , CLδ , Cm0 , Cmα , Cmq , and
e
TABLE II. ESTIMATED PARAMETERS FROM SIMULATED FLIGHT DATA OF
Cmδ . It was also of interest to see how the accuracy of AN UNSTABLE AIRCRAFT
e
estimates is affected by the presence of measurement noise. Noise = 0 % Noise = 5%
Further, it was also decided to compare the accuracy of Parameters True DM MDM DM MDM
estimates obtained using the Delta and Modified Delta CLα 2.925 2.920 2.909 2.931
2.922
(0.010)* (0.001) (0.022) (0.013)
method. To study the effect of measurement noise on
CLq 14.455 14.736 15.215 14.544
parameter estimates, simulated pseudo noise was added to the 14.700
(0.169) (0.011) (0.581) (0.025)
simulated flight data of stable and unstable aircraft. The noise 0.428 0.437 0.483 0.466
was simulated by generating successively uncorrelated pseudo CLδ 0.435
(0.006) (0.003) (0.090) (0.008)
random numbers having a normal distribution with a zero Cmα 0.150
0.149 0.149 0.148 0.148
mean and an assigned standard deviation, the standard (0.004) (0.003) (0.010) (0.006)
deviation corresponding approximately to a designated - Cm 34.658 34.696 34.667 34.691
q
34.750
(0.108) (0.079) (0.142) (0.022)
percentage (5%) of the maximum amplitude of the motion 2.601 2.582 2.647 2.599
variables ( α , q ) and the aerodynamic coefficient ( CL , Cm ). - Cm 2.578
δ (0.032) (0.018) (0.119) (0.007)
*
Table I to II list the numerical values of the estimated Standard deviation
parameters obtained from flight data of stable and unstable
aircraft respectively. From Table I to II, it is observed that for noise and no
noise case both the method yielded acceptable values of
aerodynamic parameters. However a careful look in to the
confidence level associated with the numerical values of the
estimates reveals that the estimates obtained with Modified

778 2013 3rd IEEE International Advance Computing Conference (IACC)


Delta method are consistently having larger values of Modified Delta method). It is interesting to observe from Fig.
confidence level (lesser value of standard deviation). The 6 that the estimated CL via the Modified Delta method show a
standard deviation of the estimated parameters obtained using slightly better fit with the true CL than does the estimated CL
Modified Delta method is almost one order of magnitude less
when compared with the estimates obtained using the Delta via the Delta and ML method.
method. This observation indicates the advantage of the
Modified Delta method over the Delta method for estimating
parameters from flight data of a stable aircraft in general and
unstable aircraft in particular. For validation, the estimated
longitudinal parameters as listed in Table I were used. As a
first step, a neural mapping between input vector containing
time histories (at time, t , sec) of (α )t , ( q )t , (δ e )t , (CL )t ,
(Cm )t , and the output vector containing time histories at
( t + 1 )th s of (α )t +1 , ( q )t +1 , is established using scheme as
shown in Fig. 2a. The lift coefficient (CL )t and the moment
coefficient (Cm )t corresponding to tth sec for known values of
measured normal and pitch accelerations are computed using
Eqs. (1) and (2) respectively. Once the training is completed, a
new set of flight data (generated earlier using different control Fig.4. Validation of estimates using simulated flight data of stable aircraft
input) is used for prediction of estimated response. It may be
noted here that during the prediction of response, the lift and
TABLE III. ESTIMATED PARAMETERS FROM DIGITIZED REAL FLIGHT
moment coefficient required for the input vector to be feed to DATA OF STABLE AIRCRAFT
FFNN were computed by plugging numerical value of the
estimated parameters in Eqs. (3) and (4). A comparison Parameters Estimated parameters via
MLM[7] DM [7] MDM
between the predicted and measured response is graphically 4.91 4.84 5.011
presented in Fig. 4. Although for no noise case, the matching CLα
(0.03)* (0.837) (0.023)
between predicted and measured response are exact, however CLq 6.33 10.92 8.563
for flight data with 5% noise and 10% off from the true values, (0.33) (2.91) (0.049)
the matching between predicted and measured response CLδ 1.14 0.600 0.866
deteriorated progressively. At this point, it can only be said (0.05) (0.230) (0.004)
that validation using FFNN mapping as depicted in Fig. 4 CL0 - 0.4032 0.381
shows lots of promise for further investigation. *
Standard deviation

VI. PARAMETER ESTIMATION FROM DISCRETIZED ANALOG


PLOTS OF REAL FLIGHT DATA
Due to non availability of real data for the longitudinal
response of aircraft, the next possible way of testing the
proposed method on real flight data was attempted. The
analog plot as given in [10] was digitized and used as
measured flight data shown in Fig. 5. Using variation in
motion and control variables ( Δα , Δq , Δδ e ) as the network
inputs and ΔCL as the output, the FFNN was trained.
Modified Delta method was used to estimate longitudinal
aerodynamic parameters. The results have been given in Table
III. A comparison among estimates obtain using the Delta and
Modified Delta method revealed that the estimates obtaining
using Modified Delta method were in closer agreement with
the estimates obtained using ML method. Further, the
estimates obtained using the Modified Delta method had
consistently higher values of confidence level (lesser value of Fig.4. Digitized real flight data of [10] for parameter estimation
standard deviation). To see how good the estimated values are,
we compared the estimated CL , obtained by substituting the
estimated values in the right hand side of Eq. (7), with the true
CL being analyzed. Fig. 6 compares the true CL , the
estimated CL (via the ML method, [7]), the estimated CL
(via the Delta method, [7]), and the estimated CL (via the

2013 3rd IEEE International Advance Computing Conference (IACC) 779


parameters obtained through the Modified Delta method are
one order less in magnitude when compared with the estimates
obtained through the Delta method. Next, the estimated
derivatives ∂q ∂α , ∂q ∂q , and ∂q ∂δ e are used to compute
the estimated
q − q* = (∂q ∂α )(α − α *) + (∂q ∂q)( q − q*)
+ (∂q ∂δ e )(δ e − δ e *)

The estimated q − q * via the Delta method and Modified


Delta method show reasonable match with the true (digitized)
q − q * as shown in Fig. 8.

TABLE IV. ESTIMATED PARAMETERS FROM DIGITIZED FLIGHT DATA OF


Fig.5. Comparison of true and estimated lift coefficient from real flight data of X-31 UN STABLE AIRCRAFT
[10]
Para-
meters
∂q ∂α ∂q ∂q ∂q ∂δ e Cmα Cmq Cmδ
e
The ultimate test of ability of a method for estimating From
parameters of an unstable aircraft would come from its [11]
---- ---- ---- 0.119 -1.650 -0.57
validation on real flight data. The analog plots of real flight 2.545 -1.273 -10.21
DM
(0.251)*
---- ---- ----
data of the X-31A aircraft in unstable regime, given in [11], (0.101) (0.099)
were digitized and interpolated to yield data at an interval of 2.385 -1.312 -10.04
MDM
(0.061) (0.073) (0.043)
---- ---- ----
0.1s for α , q , δ e , and q . Digitized measured flight data is *
Standard deviation
shown in Fig. 7. Because no information about the geometric,
moment of inertia, and reference flight condition is given [11],
one can not compute Cm using Eq. (2). This difficulty was
circumvented by choosing to estimate derivatives ∂q ∂α ,
∂q ∂q , and ∂q ∂δ e instead of Cmα , Cmq , and Cmδ . It was
e

also demonstrated [14] that the ratio Cmα Cmδ would be


e
same as the ratio (∂q ∂α ) (∂q ∂δ e ) . The flight data
containing the information about α , q , δ e , and q (as used
in [14]), was used, while obtaining parameters using Modified
Delta method. The aim of the exercise was to investigate
whether the Modified Delta method would help in getting
estimates with lower standard deviation. During the procedure
to estimate parameters ( ∂q ∂α , ∂q ∂q , and ∂q ∂δ e ), the
input vector for FFNN training for the Delta method consisted
of α , q , and δ , whereas the output vector had q as its Fig.6. Digitized real flight data of X-31A unstable aircraft for parameter
element [14]. However, for the application of the Modified estimation
Delta method, the input vector to FFNN had variations of α ,
q , and δ , ( Δα , Δq , Δδ ) and the output vector consisted of
Δq only. The estimated parameters, ∂q ∂α , ∂q ∂q , and
∂q ∂δ e , obtained using Modified Delta method and the Delta
method are listed in Table IV. Referring Table IV, it is readily
seen that the ratio (∂q ∂α ) (∂q ∂δ e ) = − 0.237 compares
reasonably well with the ratio Cmα Cmδ = − 0.208 despite
e
the uncertainties caused by discretization of analog plots of
real-flight data. The ratio (∂q ∂α ) (∂q ∂δ e ) obtained using
the estimates through the Delta method is − 0.249 . It can be
appreciated that, as expected, the ratio (∂q ∂α ) (∂q ∂δ e )
obtained through Modified Delta method is in closer
agreement with the numerical value of Cmα Cmδ . Further,
e
the standard deviation of derivatives of the estimated

780 2013 3rd IEEE International Advance Computing Conference (IACC)


REFERENCES
[1] R. A. Hess, “On the Use of back propagation with feed forward neural
networks for the Aerodynamic Estimation Problem,” AIAA Paper 93-
3639, August 1993.
[2] K. Basappa, and R. V. Jategaonkar, “Aspects of feed forward feural
fetwork modeling and its application to lateral-directional flight-data,”
DLR IB-111-95/30, September 1995.
[3] D. J. Linse, and R. F. Stengel, “Identification of aerodynamic
coefficients using computational neural networks,” Journal of Guidance
Control and Dynamics, Vol. 16, No. 6 , pp. 1018–1025, November-
December 1993.
[4] H.M. Youseff, “Estimation of aerodynamic coefficients using neural
networks,” AIAA Paper 93-3639, August 1993.
[5] K. Hornik, M. Stinchcombe, and H. White, “ Multi layer feed forward
neural networks are universal approximator,” Neural Networks, Vol. 2,
No. 5, pp. 359-366, 1989.
[6] J.R. Raol, R.V. Jategaonkar, “Aircraft parameter estimation using
recurrent neural networks -- a critical appraisal,” AIAA Paper 95-3004,
1995.
Fig.7. Comparison of true and estimated accelerations from real flight data of [7] S. C. Raisinghani, A. K. Ghosh, and P. K. Kalra, “Two new techniques
[11] for aircraft parameter estimation using neural networks,” The
Aeronautical Journal, Vol. 102, No. 1011, pp. 25-29, January 1998.
[8] A. K. Ghosh, S. C. Raisinghani, and S. Khubchandani, “Estimation of
VII. CONCLUSION aircraft lateral-directional parameters using nural networks,” Journal of
Aircraft, Vol. 35, No. 6, pp. 876-881, November-December 1993.
The Modified Delta method (by improving the existing [9] S. Singh, and A. K. Ghosh, “Longitudinal parameter estimation using
Delta method), has been proposed for estimating the aircraft wind tunnel and simulated flight data of missile configuration,” Journal
parameters from flight data using the feed forward neural of Aerospace Sciences and Technologies, Vol. 59, No. 2, pp. 102-120
networks. The results obtained for simulated flight data and May 2007.
real flight data have shown the success and potential of the [10] M. J. Queijo, W. R. Wells, and D. A. Keskar, “Inclusion of unsteady
aerodynamics in longitudinal parameter estimation from flight data,”
proposed methods. As compared to the Delta method, the NASA TP-1536, 1979.
proposed Modified Delta method yields estimates with lesser [11] R. V. Jategaonkar, and F. Thielecke, “Evalution of parameter estimation
standard deviation. The proposed Modified Delta method is a methods for unstable aircraft,” Journal of Aircraft, Vol. 31, No. 3, pp.
straightforward method to apply; the FFNN is trained on the 510-519, 1994,.
given flight data containing differential variation of motion [12] M. R. Waszak, D. K. Schmidt, “Flight dynamics of aeroelastic vehicles,”
and control variables and the parameters estimated in one try. Journal of Aircraft, Vol. 25, No. 6, pp. 563-571, 1988.
The applicability of the proposed Modified Delta method is [13] J. Roskam, “Flight dynamics of rigid and elastic airplane,” University
shown on simulated as well as digitized analog plots of real of Kansas Press, 1972.
flight data. The results suggest that the Modified Delta method [14] A. K. Ghosh, and S. C. Raisinghani, “Parameter estimation from flight
can be used advantageously to estimate parameters of a stable data of an unstable aircraft using neural networks,” Journal of Aircraft,
Vol. 35, No. 5, pp. 892-894, 2002.
aircraft in general and unstable aircraft in particular. The
proposed method does not require integration of system
equations as required by earlier techniques being used for
parameter estimation.

2013 3rd IEEE International Advance Computing Conference (IACC) 781

Vous aimerez peut-être aussi