NN in Aviation
NN in Aviation
1 Introduction
Calculations and mathematical modeling are essential for the aircraft development
and determination of its operating limitations, including estimation of aircraft behav-
ior safety limits. The recent development of aircraft modeling applied for real time
analysis of flight data enables to prevent accidents [1,2] and demand the reliable
mathematical presentation of an aircraft and its systems behavior. However, applica-
tion of computational methods requires compliance between the computation (math
modeling) results and the experimental data, i.e. it is necessary to identify the math
model parameters from the experimental data of real object behavior.
The experience in development and application of the procedure for the flight dy-
namics math model parameters estimation according to flight tests data [3, 4] has
shown that the most complicated element of practical identification tasks is the ad-
justment of identification results obtained from different samples of initial data. The
effort to solve this problem was made in the procedure [3] by identification of correc-
tions for aerodynamic coefficients with “parallel” optimization of disagreement crite-
ria in two (some) flight test data fragments obtained in similar conditions at similar
speed and altitudes (for example: maneuvers with stick “to the left” and “to the right”,
“forwards” and “backwards”). However, when solving the problem [4], main efforts
were made to adjust the corrections determined by identification procedure on differ-
ent data samples.
D. Palmer-Brown et al. (Eds.): EANN 2009, CCIS 43, pp. 246–256, 2009.
© Springer-Verlag Berlin Heidelberg 2009
Using of ANN for Aircraft Motion Parameters Identification 247
Artificial neural networks (ANN) allow determination of the required relations be-
tween input and output parameters of the object. Moreover, unlike the traditional
identification methods, neural networks have a memory: it means that the results
could be verified and accumulated during repeated “training” cycles (during the proc-
essing of new samples of initial data). Thus, neural networks allow getting the re-
quired relations for wide range of conditions at once and, as a result, to align random
factors, which are unavoidable for experimental data.
This paper presents the use of neural network to solve the problem of identification
(during the aircraft takeoff or landing run on a concrete runway) of rolling resistance
and wheels braking coefficients. In recent years international aviation community has
paid much attention to the analysis of aircraft behavior during its motion on runway,
in particular during take-off and landing on precipitation-covered runway. For in-
stance, since 1996 NASA, FAA and Transport Canada have been performed
JW/RFMP program (The Joint Winter Runway Friction Measurement Program).
Present-day knowledge about aircraft behavior on the contaminated runway are gen-
eralized in the amendment to European certification requirements (NPA No.
14/2004), developed by JAA, but the conclusion on the necessity of further investiga-
tion of this problem was made. Reports on flight accidents, related to aircraft overrun
the runway indicate the necessity of such research.
At this stage we considered to estimate the efficiency of neural networks as the
means of identification of the aircraft motion math model parameters. For this pur-
pose the simple enough but practically important task was selected: evaluation of
wheels compression and their rotational speed influence on the friction coefficient
value (for dry concrete runway). This task is presented here in details.
A more complicated procedure based on ANN application, which is presented here
in main results, is intended for identification of dependencies describing wheels resis-
tance and braking performances on precipitation-covered runway (see [5]).
G – aircraft weight; i – runway slope angle (uphill >0); P – engines thrust; Xaero = CD *q*S; q – dynamic
pressure; S – wing area; FR = µ R* Fy; Fy = G*cos(i) - Yaero; Yaero = CL*q*S.
Nature of rolling friction coefficient µ R is shown in Fig. 1 taken from [6]. At this
figure, parameter “а” determines the resultant normal force displacement from the