Ozcan Samlioglu

NAVAL POSTGRADUATE SCHOOL
Monterey, California
THESIS
A NEURAL NETWORK APPROACH FOR HELICOPTER
AIRSPEED PREDICTION
by
Ozcan Samlioglu
March 2002
Thesis Advisor: Russell W. Duren

Assoc. Advisor: Monique P. Fargues
Approved for public release; distribution is unlimited.

THIS PAGE INTENTIONALLY LEFT BLANK
REPORT DOCUMENTATION PAGE Form Approved OMB No. 0704-0188
Public reporting burden for this collection of information is estimated to average 1 hour per response, including
the time for reviewing instruction, searching existing data sources, gathering and maintaining the data needed, and
completing and reviewing the collection of information. Send comments regarding this burden estimate or any
other aspect of this collection of information, including suggestions for reducing this burden, to Washington
headquarters Services, Directorate for Information Operations and Reports, 1215 Jefferson Davis Highway, Suite
1204, Arlington, VA 22202-4302, and to the Office of Management and Budget, Paperwork Reduction Project
(0704-0188) Washington DC 20503.
1. AGENCY USE ONLY (Leave blank) 2. REPORT DATE 3. REPORT TYPE AND DATES COVERED
March 2002 Master’s Thesis
4. TITLE AND SUBTITLE: Title (Mix case letters) 5. FUNDING NUMBERS
A Neural Network Approach for Helicopter Airspeed Prediction
6. AUTHOR(S) Ozcan Samlioglu
7. PERFORMING ORGANIZATION NAME(S) AND ADDRESS(ES) 8. PERFORMING
Naval Postgraduate School ORGANIZATION REPORT
Monterey, CA 93943-5000 NUMBER
9. SPONSORING /MONITORING AGENCY NAME(S) AND ADDRESS(ES) 10. SPONSORING/MONITORING
N/A AGENCY REPORT NUMBER
11. SUPPLEMENTARY NOTES The views expressed in this thesis are those of the author and do not reflect the official
policy or position of the Department of Defense or the U.S. Government.
12a. DISTRIBUTION / AVAILABILITY STATEMENT 12b. DISTRIBUTION CODE
Approved for public release; distribution is unlimited.
13. ABSTRACT (maximum 200 words)
Conventional pitot-static airspeed measurement systems do not yield accurate measurements when aircraft
speed is below 40 knots. Recent studies have demonstrated that neural network approaches for predicting airspeed
are quite promising. In this thesis, a back-propagation neural network is used to predict the airspeed of UH-60A
and OH-6A helicopters in the low speed environment. The input data to the neural networks were obtained using
the FLIGHTLAB flight simulator. The results obtained by flight simulation were validated by comparison to
results of a previous study of the UH-60A helicopter based on actual flight data. The results of the work performed
for this thesis show that at sea level the UH-60A low airspeed can be predicted with an accuracy of ± 0.71 knots
and ± 0.88 knots for out of ground effect and in ground effect conditions respectively. OH-6A analyses were
performed at two pressure altitudes. At sea level the OH-6A airspeed can be predicted with an accuracy ± 0.75
knots when the aircraft is out of ground effect and ± 0.88 knots when the helicopter is in ground effect. At a
pressure altitude of 6000 feet OH-6A airspeed can be predicted with an accuracy of ± 0.64 knots for both flight
conditions.
14. SUBJECT TERMS 15. NUMBER OF

Airspeed measurement, Neural Networks, Back-propagation PAGES
137
16. PRICE CODE
17. SECURITY 18. SECURITY 19. SECURITY 20. LIMITATION
CLASSIFICATION OF CLASSIFICATION OF THIS CLASSIFICATION OF OF ABSTRACT
REPORT PAGE ABSTRACT
Unclassified Unclassified Unclassified UL
NSN 7540-01-280-5500 Standard Form 298 (Rev. 2-89)
Prescribed by ANSI Std. 239-18
i
ii
Approved for public release; distribution is unlimited
A NEURAL NETWORK APPROACH FOR HELICOPTER AIRSPEED

PREDICTION
Ozcan Samlioglu
1st Lieutenant, Turkish Army
B.S., Turkish Army Academy, 1994
Submitted in partial fulfillment of the

requirements for the degree of
MASTER OF SCIENCE IN AERONAUTICAL ENGINEERING
from the
NAVAL POSTGRADUATE SCHOOL

March 2002
Author: Ozcan Samlioglu
Approved by: R. W. Duren, Thesis Advisor
M. P. Fargues, Assoc. Advisor
Max F. Platzer, Chairman

Department of Aeronautics and Astronautics
iii
iv
ABSTRACT
Conventional pitot-static airspeed measurement systems do not yield accurate

measurements when aircraft speed is below 40 knots. Recent studies have demonstrated
that neural network approaches for predicting airspeed are quite promising. In this thesis,
a back-propagation neural network is used to predict the airspeed of UH-60A and OH-6A
helicopters in the low speed environment. The input data to the neural networks were
obtained using the FLIGHTLAB flight simulator. The results obtained by flight
simulation were validated by comparison to results of a previous study of the UH-60A
helicopter based on actual flight data. The results of the work performed for this thesis
show that at sea level the UH-60A low airspeed can be predicted with an accuracy of
± 0.71 knots and ± 0.88 knots for out of ground effect and in ground effect conditions
respectively. OH-6A analyses were performed at two pressure altitudes. At sea level the
OH-6A airspeed can be predicted with an accuracy ± 0.75 knots when the aircraft is out
of ground effect and ± 0.88 knots when the helicopter is in ground effect. At a pressure
altitude of 6000 feet OH-6A airspeed can be predicted with an accuracy of ± 0.64 knots
for both flight conditions.
v
vi
TABLE OF CONTENTS
I. INTRODUCTION........................................................................................................1
A. BACKGROUND ..............................................................................................1
B. TECHNOLOGICAL PROBLEM ..................................................................2
II. NEURAL NETWORKS ..............................................................................................7
A. INTRODUCTION TO NEURAL NETWORKS ..........................................7
B. NEURAL NETWORKS BACKGROUND..................................................11
C. INTRODUCTION TO THE BACK-PROPAGATION NEURAL
NETWORK ....................................................................................................13
III. DEVELOPMENT OF NEURAL NETWORK MODEL .......................................17
A. NEURALWORKS PROFESSIONAL II/PLUS SOFTWARE..................17
B. FLIGHTLAB SIMULATOR ........................................................................18
C. SELECTING DATA AND BUILDING THE MODEL .............................20
IV. ANALYSIS AND RESULTS ....................................................................................25
A. ANALYSIS OF THE UH-60A MODEL ......................................................27
1. Out of Ground Effect (OGE) Analysis.............................................27
a. 2-Hidden Layer BPNN............................................................27
b. 1-Hidden Layer BPNN............................................................32
2. RBFN Networks .................................................................................34
3. Pruned BPNN .....................................................................................35
4. In-Ground Effect Analysis ................................................................37
a. 15-25-25-1 BPNN NCD ..........................................................37
b. 15-25-25-1 BPNN Ext. DBD ..................................................38
c. 15-25-25-1 BPNN NCD (Pruned) ..........................................39
d. 15-18-25-1 BPNN NCD ..........................................................40
5. Baseline Data Set Analysis ................................................................42
a. 15-18-25-1 BPNN NCD ..........................................................42
b. 15-25-25-1 BPNN NCD ..........................................................43
c. 15-25-25-1 BPNN Ext. DBD ..................................................44
d. 15-25-25-1 BPNN NCD (Pruned) ..........................................45
e. Baseline Data Set IGE Analysis .............................................46
f. Baseline Data Set OGE Analysis............................................49
6. Simplifying The Data Set Using Eigenvalues and Eigenvectors....52
a. 6-18-18-1 BPNN Ext. DBD ....................................................53
b. 6-18-18-1 BPNN NCD (Pruned) ............................................54
c. 6-18-18-1 BPNN NCD ............................................................56
B. ANALYSIS OF THE OH-6A MODEL........................................................57
1. Out of Ground Effect Analysis at Sea Level....................................57
a. 14-25-25-1 BPNN NCD ..........................................................57
b. 14-25-25-1 BPNN Ext. DBD ..................................................58
c. 14-25-25-1 BPNN NCD (Pruned) ..........................................59
vii
2. In-Ground Effect Analysis at Sea Level...........................................60
a. 14-25-25-1 BPNN NCD ..........................................................60
b. 14-25-25-1 BPNN Ext. DBD ..................................................61
c. 14-25-25-1 BPNN NCD (Pruned) ..........................................62
3. OH-6A Baseline Data Analysis at Sea Level....................................63
a. 14-25-25-1 BPNN NCD ..........................................................64
b. 14-25-25-1 BPNN Ext. DBD ..................................................65
c. 14-25-25-1 BPNN NCD (Pruned) ..........................................66
d. OH-6A SL Baseline Data IGE Analysis.................................67
e. OH-6A SL Baseline Data OGE Analysis ...............................70
4. OH-6A Out of Ground Effect Analysis at High Altitude ...............73
a. 14-25-25-1 BPNN NCD ..........................................................73
b. 14-25-25-1 BPNN Ext. DBD ..................................................74
c. 14-25-25-1 BPNN NCD (Pruned) ..........................................75
5. OH-6A In-Ground Effect Analysis at High Altitude ......................76
a. 14-25-25-1 BPNN NCD ..........................................................76
b. 14-25-25-1 BPNN Ext. DBD ..................................................77
c. 14-25-25-1 BPNN NCD (Pruned) ..........................................79
6. OH-6A Baseline Data Analysis at High Level .................................80
a. 14-25-25-1 BPNN NCD ..........................................................80
b. 14-25-25-1 BPNN Ext. DBD ..................................................81
c. 14-25-25-1 BPNN Ext. DBD (Pruned)...................................82
7. OH-6A HL Baseline Data IGE Analysis ..........................................83
a. 14-25-25-1 NCD IGE ..............................................................83
b. 14-25-25-1 Ext. DBD IGE ......................................................84
c. 14-25-25-1 NCD (Pruned) IGE ..............................................85
8. OH-6A HL Baseline Data OGE Analysis ........................................86
a. 14-25-25-1 NCD OGE.............................................................86
b. 14-25-25-1 Ext. DBD OGE.....................................................87
c. 14-25-25-1 Ext. DBD (Pruned) OGE .....................................88
C. NETWORK PERFORMANCES SUMMARY ...........................................89
V. CONCLUSIONS AND RECOMMENDATIONS...................................................95
A. SUMMARY ....................................................................................................95
B. RECOMMENDATIONS FOR FURTHER RESEARCH .........................96
LIST OF REFERENCES ......................................................................................................97
APPENDIX A. NEURALWORKS PROFESSIONAL PLUS/II PROGRAM SETUP ..99
APPENDIX B. MATLAB® M-FILES ................................................................................103
INITIAL DISTRIBUTION LIST .......................................................................................117
viii
LIST OF FIGURES
Figure 1. Structure of a Neuron [From Ref. 10]................................................................8

Figure 2. System Diagram of NN [From Ref.2]................................................................9
Figure 3. Layer structure of NN [From Ref. 10]. ..............................................................9
Figure 4. Structure of a layer [from Ref.2]......................................................................10
Figure 5. A Typical Back-Propagation Network [From Ref. 2]. ....................................13
Figure 6. Typical Window of Neural Works Professional II/Plus Software...................17
Figure 7. Xanalysis Window [From Ref. 7]....................................................................20
Figure 8. Results for the UH-60A helicopter at 85 ft.; network configuration 14-25-
25-1; Ext. DBD learning rule...........................................................................28
25-1; NCD learning rule. .................................................................................29
25-1; DBD learning rule. .................................................................................30
25-1; Quickprop learning rule..........................................................................31
1; NCD learning rule........................................................................................32
1; Ext. DBD learning rule. ...............................................................................33
Figure 14. Results for the UH-60A helicopter at 85 ft.; network configuration 14-
200-1; RBFN NCD learning rule.....................................................................35
25-1; NCD (Pruned) learning rule. ..................................................................36
Figure 20. Results for the UH-60A helicopter with baseline data; network
configuration 15-18-25-1; NCD learning rule. ................................................42
configuration 15-25-25-1; Ext. DBD learning rule..........................................45
configuration 15-25-25-1; NCD (Pruned) learning rule. .................................46
ix
Figure 24. Results for the UH-60A helicopter at 20 ft with baseline data; network
Figure 30. Results for the UH-60A helicopter at 85 ft with simplified data; network
configuration 6-18-18-1; Ext. DBD learning rule............................................54
configuration 6-18-18-1; NCD (Pruned) learning rule. ...................................55
configuration 6-18-18-1; NCD learning rule ...................................................56
Figure 33. Results for the OH-6A helicopter at 100 ft (SL); network configuration
14-25-25-1; NCD learning rule........................................................................58
14-25-25-1; Ext. DBD learning rule. ...............................................................59
14-25-25-1; NCD (Pruned) learning rule.........................................................60
Figure 39. Results for the OH-6A helicopter with baseline data (SL); network
Figure 42. Results for the OH-6A helicopter at 12 ft with baseline data (SL); network
Figure 45. Results for the OH-6A helicopter at 100 ft with baseline data (SL);
network configuration 14-25-25-1; Ext. DBD learning rule. ..........................70
x
network configuration 14-25-25-1; NCD learning rule. ..................................71
network configuration 14-25-25-1; NCD (Pruned) learning rule. ...................72
Figure 48. Results for the OH-6A helicopter at 100 ft (HL); network configuration
14-25-25-1; NDC (Pruned) learning rule.........................................................76
14-25-25-1; NDC learning rule........................................................................77
Figure 54. Results for the OH-6A helicopter with baseline data (HL); network
configuration 14-25-25-1; NCD learning rule. ...............................................80
configuration 14-25-25-1; Ext. DBD (Pruned) learning rule...........................83
Figure 57. Results for the OH-6A helicopter at 12 ft with baseline data (HL); network
configuration 14-25-25-1; NCD (Pruned) learning rule ..................................86
Figure 60. Results for the OH-6A helicopter at 100 ft with baseline data (HL);
network configuration 14-25-25-1; Ext. DBD (Pruned) learning rule. ...........89
Figure 63. Back-propagation network setup window .....................................................100
Figure 64. Instrument /Create menu................................................................................101
Figure 65. SaveBest command window ..........................................................................102
Figure 66. Test command window ..................................................................................102
xi
xii
LIST OF TABLES
Table 1. Neural Network Input Parameters. ..................................................................21

Table 2. Commonly Used Transfer Functions. ..............................................................23
Table 3. Results for the UH-60A helicopter at 85 ft.; network configuration 14-25-
25-1; DBD learning rule. .................................................................................30
25-1; Quickprop learning rule..........................................................................31
1; NCD learning rule........................................................................................33
1; Ext. DBD learning rule. ...............................................................................34
Table 9. Results for the UH-60A helicopter at 85 ft.; network configuration 14-
200-1; RBFN NCD learning rule.....................................................................34
25-1; NCD (Pruned) learning rule ...................................................................37
25-1; NCD learning rule. Results of UH-60A for 15-18-25-1 NCD. ..............41
Table 15. Results for the UH-60A helicopter with baseline data; network
Table 19. Results for the UH-60A helicopter at 20 ft with baseline data; network
xiii
Table 25. Results for the UH-60A helicopter at 85 ft with simplified data; network
configuration 6-18-18-1; Ext. DBD learning rule............................................54
configuration 6-18-18-1; NCD (Pruned) learning rule. ...................................55
configuration 6-18-18-1; NCD learning rule. ..................................................56
Table 28. Results for the OH-6A helicopter at 100 ft (SL); network configuration
14-25-25-1; NCD learning...............................................................................61
Table 34. Results for the OH-6A helicopter with baseline data (SL); network
Table 37. Results for the OH-6A helicopter at 12 ft with baseline data (SL); network
Table 40. Results for the OH-6A helicopter at 100 ft with baseline data (SL);
network configuration 14-25-25-1; NCD (Pruned) learning rule. ...................72
Table 43. Results for the OH-6A helicopter at 100 ft (HL); network configuration
xiv
14-25-25-1; NDC (Pruned) learning rule.........................................................75
14-25-25-1; NDC learning rule........................................................................77
Table 49. Results for the OH-6A helicopter with baseline data (HL); network
configuration 14-25-25-1; Ext. DBD (Pruned) learning rule...........................82
Table 52. Results for the OH-6A helicopter at 12 ft with baseline data (HL); network
Table 55. Results for the OH-6A helicopter at 100 ft with baseline data (HL);
network configuration 14-25-25-1; Ext. DBD (Pruned) learning rule. ...........88
Table 58. Overall Results for the UH-60A Helicopter. ...................................................90
Table 59. Overall Results for OH-6A Helicopter at Sea Level. ......................................92
Table 60. Overall Results of OH-6A Helicopter at High Altitude...................................93
xv
xvi
ACKNOWLEDGMENTS
My special thanks goes to Dr. Russell Duren for his assistance, endless effort and
patience throughout the whole process of this thesis. I am truly grateful to Dr. Duren for
his friendship and generosity during my education in NPS. I also wish to thank Dr.
Monique Farques for her valuable guidance, time and patience. Her insight and
knowledge helped me to facilitate this effort.
I owe many to thanks to LT. Gregory Ouellette, USN for his endless help, support
and valuable friendship. His help was crucial to the completion of this thesis.
I would like to extend my appreciation to Jan Goericke and Dr. Dean Hoff for
their kindness and help.
I am greatly appreciative of the opportunity provided by the Turkish Armed
Forces Command to study for this degree.
Most of all, I would like to thank my family. I dedicate this work to my parents
and my brother Ercan, for their infinite love and support over the years and to my dear
wife Hamra and my lovely daughter Bersan for their never-ending love and
understanding. Without their support none of this would have been achieved.
xvii
xviii
I. INTRODUCTION
Since the 1970s there has been a significant concern to increase the accuracy of
airspeed measurement systems because current measurement systems are ineffective at
low airspeeds. The focus of this thesis is to develop a neural network (NN) model to
estimate helicopter airspeed in the low speed environment using the NeuralWorks
Professional Plus/II software. A flight simulator provides the required data that is used as
inputs to the neural network model.
The thesis is organized into five chapters. Chapter I is the introduction chapter,
which presents research results currently available in the area of neural network and
aircraft speed measurement systems. Chapter II introduces the main concepts behind NN
implementations and the basic structure of the Back-Propagation Neural Network
(BPNN). Chapter III describes a specific implementation using the NeuralWare
Professional Plus/II software and presents the FLIGHTLAB simulation tool. The
simulations performed as part of this thesis are discussed in Chapter IV. Finally, Chapter
V summarizes results obtained and discusses avenues for further research.
A. BACKGROUND
Since 1995, the Naval Postgraduate School (NPS) has conducted several research
projects using OH-6A “Cayuse” helicopters after receiving two such helicopters from the
Massachusetts Army National Guard. One of these projects involves the development of
a Vortex-Ring State Warning System (VRSWS). The vortex-ring state is a condition of
powered flight where the helicopter settles into its own downwash. The helicopter will
increase its rate of descent very rapidly as the lift is lost when entering the vortex-ring
state, and any further application of collective, a flight control mechanism for helicopters,
tend to reduce rotor efficiency. In this state the rotor experiences a very high vibration
level and loss of control. The consequences of the vortex-ring state when the helicopter is
close to the ground might be extremely dangerous because loss of control at low altitudes
often results in aircraft crash. Development of VRSWS requires better airspeed
measuring systems than these currently available with most avionics systems.
1
Last year, NPS signed a Cooperative Research and Development Agreement
(CREDA) with Advance Rotorcraft Technology, Inc. (ART) to reciprocally develop
advancements in rotorcraft technology. Based on this agreement, NPS provided ART
with equipment for a flight simulator, including the sticks and grips, seats and avionics
suite of an OH-6A for integration into the control loader platform. ART provided their
helicopter modeling suite with advanced visual rendering equipment to produce a fully
functional stationary open platform flight simulator. In return for this equipment, ART
provided the flight modeling software to NPS to develop rotorcraft models for future use
which includes the OH-6A and V-22 “Osprey” models. NPS was able to obtain flight
parameters such as rotor RPM, engine torque, roll rate, etc., using simulation techniques
through its cooperation with ART. LT Gregory Ouellette is currently conducting research
about FLIGHTLAB, a flight simulation tool provided by ART. The required flight
parameters for the NN model of this thesis were provided by LT. Ouellette, USN
[Ref.19].
B. TECHNOLOGICAL PROBLEM
The aerodynamic velocity in the plane of a helicopter’s rotor disc significantly

affects the control characteristics of that helicopter. The pilot easily feels this effect
during flight and especially in hover, low speed and transition conditions. Small changes
in these flight regimes are rapidly followed by substantial changes in control sensitivities
and trim positions. Therefore, the pilot’s job is very difficult when performing tasks
during these specified conditions, and it is possible for these conditions to occur in any
flight mission. Pilots are capable of successfully handling the control helicopter with
training but in IFR flight, where ground references are no longer accessible, the pilot’s
job becomes increasingly more difficult. The need to extend military operations during
poor weather and at night has been inevitable for military effectiveness throughout
history. Therefore, many military missions have been performed during poor visibility
and in a low speed environment. However, operations under these conditions endanger
flight safety. The roles of search and rescue, submarine detection, mine countermeasures,
front line supply, air-to-ground attack, and reconnaissance all demand a low speed, poor
weather, day or night capability. Thus, several technological improvements have been
2
made to aircraft to conduct missions without endangering the lives of the aircrew. Today,
military helicopters perform a wide variety of tasks in conditions ranging from hot and
dry to cold and wet, windy and low visibility weather. Accurate low speed velocity
sensing devices are essential because aircraft velocity and position information are what
pilots need to perform the aforementioned tasks.
Conventional methods to measure airspeed for aircrafts have been used for over
60 years. Pitot-static systems are still commonly used since they offer simple, low cost
and reliable enough solutions to measure airspeed. In this system, airspeed is derived by
measuring the difference between the total pressure occurring at the forward facing pitot-
probe and the static pressure measured at a static vent [Ref. 12]. Since helicopters fly at
relatively lower speeds than aircrafts, and the cruise speed of a helicopter is less than 0.3
Mach, helicopters are used to fly in incompressible subsonic flow conditions where
Bernoulli principles pertain. In such flow regimes, airspeed is obtained simply by taking
the square root of this pressure difference and multiplying the scale factor.
The flight of a helicopter occurs in the two distinct regimes of hover/low speed
flight up to 45 knots, including vertical maneuvering, and mid/high speed flight up to Vne
–never exceed velocity- where Vne is the maximum airspeed permitted under any
circumstances [Ref. 17]. It is defined as a function of altitude, temperature and gross
weight. For example, using flight manual of UH-60A, Vne is found to be 186 knots at -20o
C, 4000 feet and when the gross weight of helicopter is18000 lb. The low speed regime is
very much unique to the helicopter as an operationally useful regime. No other flight
vehicles are as flexible and efficient at maneuvering slowly close to the ground and at
avoiding obstacles. Therefore, the low speed regime is a significant portion of helicopter
flight time. In fact, the maneuvers done in this regime make the helicopter invaluable.
Although low speed is very critical for helicopters during take-off, landing or hovering,
measuring airspeed and wind direction is generally lacking in this regime [Ref. 3].
During low speed flight, the current airspeed measurements systems are inaccurate due to
the rotor downwash and limitations of the pitot-static system. The conventional pitot-
static sensor is ineffective at airspeeds below 40 knots and does not function at all during
rearward flight [Ref. 8]. John Carter explains in Ref. 12 why it does not function
properly.
3
The pressure difference in pitot-static system becomes very small at low
airspeeds. This creates a practical difficulty in that presently available
pressure measuring techniques which are suitable for aircraft application
are poor at measuring pressure differentials of less than 1/1200
atmospheres which corresponds to an airspeed of 20 knots.
Moreover, rotor downwash inevitably enters the pitot probe, which causes a fast
error or enters the static port resulting in a slow error. Flow patterns developing around
the aircraft during sideward and rearward flight may bring erroneous airspeed indications
[Ref. 13].
In low speed flight, required engine power increases due to the difficulty of
maneuvers. Due to these high engine and tail rotor anti-torque requirements, extra
attention must be paid to directional control margins. Moreover, vibrator loads can occur
in some maneuvers, which can result in fatigue damage accumulation in flight critical
components [Ref. 4]. The pilot is often required to fly this type of maneuver, and ground
references, if available, are mostly used instead of instruments. However, in a combat
environment, ground references might not be available which increases the need for
accurate measurement systems.
Many developments have been completed since the 1950’s to increase the
accuracy of measurement systems. One study suggested moving the probes above the
rotor hub in order to protect the pitot system from the down-wash effect. Another study
suggested using a swivel device mounted above the rotor hub that can measure true
airspeed magnitude and direction [Ref. 3]. In this design, two venturi tubes were mounted
on opposite ends of a rotating arm to provide a differential measurement between the two
sensors. These sensors were used to calculate airspeed and wind direction [Ref. 4].
However, such a system requires a slip ring assembly or a similar means of transferring
the data from the rotating system to the body of the helicopter. One other approach is
based on the study of wake under the rotor and using a sensor mounted under the rotor to
determine the airspeed of the helicopter. In this approach, a 3600 rotating pitot-static
probe was used to measure the true airspeed and wind direction [Ref. 4].
In the late 1970s and early 1980s, Faulkner and Buchner proposed a study based
on the idea that the measurement of the control position can be used to estimate airspeed
4
when airspeed substantially affects the control trim position. These researches built a
model using simplified thrust and flapping equations to obtain longitudinal and lateral
velocity components. However, their analysis requires knowledge of the main rotor flap
angle, which is difficult to measure [Ref. 8].
Although some progress has been made because of these studies, none of them
have received worldwide acceptance due to system complexity, reliability, and economic
and aerodynamic issues.
McCool and Haas describe some other efforts to determine the airspeed
analytically [Ref. 3]. One of these efforts is a neural network-based approach, which is
the core of this thesis. In that study, first the flight parameters, such as rotor RPM, cyclic
position, etc. were obtained during test flights of HH-60J helicopter and CH-46E
helicopter and then these parameters were used as inputs into a NN model to predict
airspeed and wind direction. By using fuselage parameters, the problem of transmitting a
signal from the rotating system to the fuselage is eliminated [Ref. 3]. The results obtained
from this study proved that NN could be used to predict airspeed with reasonable
accuracy.
A NN implementation is chosen for several reasons. First, it is easy to use. Input

parameters can be chosen as quantities, which are commonly measured on a flight data
recorder. Second, the combined influence of inputs can be investigated by using multiple
hidden layers. Third, a more flexible and empirical estimation can be obtained, as new
data becomes available to retrain the network [Ref. 16]. Based on the aforementioned
benefits, McCool and Haas showed that the neural network approach provides a
mechanically simple and inexpensive alternative to current low airspeed measurement
technology [Ref.3].
The primary objective of this thesis is to determine the airspeed of an OH-6A

helicopter using an NN implementation with the input data provided from the helicopter
simulator model FLIGHTLAB. Our intention is to increase the accuracy of the
measurement systems in a cost-effective way by using the NN and a simulator model.
Since simulator data is used as inputs to the system, first UH-60A data were obtained
from the simulator and used as inputs to the NN. HH-60J and UH-60A helicopters are
5
very similar and they are both variant of Skorsky basic Hawk series helicopters.
Therefore a comparison of HH-60J helicopter and UH-60A helicopter results would
provide the verification of using simulator data for the NN model. Once the model is
validated for the UH-60A helicopter, the next step is to predict the speed of the OH-6A
helicopter using the same approach.
6
II. NEURAL NETWORKS
A. INTRODUCTION TO NEURAL NETWORKS

The human brain is the world’s most complex computing device. Thus, the
brain’s powerful thinking, remembering, learning and problem solving capabilities have
been studied and modeled for digital computing. The brain is so complex that scientists
have only just begun to understand how it works. Although little is known about the brain
and the neural system, modeling the functionality of the brain in a very fundamental
manner leads to the creation of neural networks.
The neuron is the fundamental cellular unit of the nervous system and, in
particular, the brain [Ref. 6]. The neuron functions like a microprocessing unit, which
receives and combines signals from many other neurons. The brain consists of 1011
number of neurons. They are connected to each other by approximately 104 connections
per element or 60 trillion connections total [Ref. 1]. The input path of a cell body is
called “dendrites”. The output path is called “axon”. The axon of a neuron splits up and
connects to the dendrites of other neurons through a connection called a “synapse”. The
cell body, which is also called “soma”, sums incoming signals when sufficient inputs are
received. It is generally thought that all functions are stored in the neurons and in the
connections between them. Learning occurs when new connections are established or
existing connections are modified. The structure of a neuron is depicted in Figure 1 [Ref.
10].
In an artificial neural network (NN), the same principles are used to simulate and
capture some of the power of the brain and the neural system. In a NN, the unit
corresponding to a neuron is called the “processing element (PE)”. A PE has many input
paths and combines the values of these input paths. The combined input is then modified
by a transfer function.
7
A Pwts of *
Tvpio»! Nerv« C«ll
Dendrites : Accept inputs
So ma : Process the inputs
Axon : Turn the processed inputs

into outputs
Synapses : The electrochemical

contact between neurons
Figure 1. Structure of a Neuron [From Ref. 10].
The transfer function is also referred to as an “activation function”. This transfer

function can be either a threshold function or a continuous function of the combined input
[Ref. 6]. When the transfer function is a threshold function, as when the combined
activity level of PE reaches a certain level, the information passes or otherwise is filtered.
The output path of the transfer function is passed directly to the output path of PE. The
PE output is connected to the other PE input paths through connection weights, as in a
neural system.
The NN is trained to perform a particular function by adjusting the values of

strength of connections (weights) between elements. The procedure to modify weights
and biases of a network is called a learning rule, where the bias corresponds to a weight
with a constant input. Note that, biases are sometimes added to add more flexibility to the
network configuration. There are two types of learning:
1) Supervised Learning, where sets of inputs and desired outputs (target values)
are presented to the network, and the NN configures itself to achieve desired input/output
mapping; 2) Unsupervised Learning, where only inputs are shown to the network and NN
organizes itself internally so that each PE responds strongly to different sets of inputs
In this thesis, only supervised learning schemes are considered. In that type of
learning, a NN generates its own rules by learning from shown examples. Learning is
achieved through a learning rule that makes necessary modifications to weights and
8
biases in response to network output and target values. Figure 2 shows how such a NN
system works.
Figure 2. System Diagram of NN [From Ref.2].
A NN consists of many PEs grouped together called layers. A layer is defined as a

group of neurons having connections to the same inputs and sending outputs to the same
destinations [Ref. 2]. Figure 3 depicts the layer structure of a NN.
INPUT
LAYER
0) 0,
HIDDEN
LAYER
(there may be several
hidden layers)
o o o OUTPUT
LAYER
Figure 3. Layer structure of NN [From Ref. 10].
A general NN usually consists of an input layer, one or two hidden layers, and
one output layer. The typical structure of a layer with one neuron is presented in Figure 4.
9
Figure 4. Structure of a layer [from Ref.2].
A layer consists of the following elements:
• An input vector (p),

• Weights (w) represent connection strength,
• Summation (∑),
• Bias vector (b) represents a column vector of bias values for a layer of
neurons,
• Transfer function (f),
• Output vector (a) is the output of the network.
The architecture of a NN depends on the number of layers the network possesses,
the number of neurons in each layer, each layer’s transfer function, and how the layers
are connected to each other. Note that there is no unique architecture for any given
problem, and the best architecture heavily depends on the data presented to the network,
the number of neurons for each layer selected by the user, and the selected activation
function. Most practical NNs have just two or three layers. It is certain that the greater
number of neurons in the hidden layer, the more powerful the network may potentially
be. Note that, however, adding more neurons makes the network more complex and
complexity should be minimized since it increases training time. Furthermore, the NN
may memorize the training pattern set, but not perform well on data outside of the
training set when the ratio of neurons to training patterns is too large. This problem is
referred to as overspecialization or overfitting.
10
B. NEURAL NETWORKS BACKGROUND
Although background studies of NN, extend back to the late 19th and early 20th
centuries, modern NN implementations first appeared in 1943 with the works of Warren
McCulloch and Walter Pitts. Their researches considered the concept of artificial
neurons, which have the capability to compute arithmetic or logical functions. McCulloch
and Pitts published watershed paper entitled “A Logical Calculus of Ideas Imminent in
Nervous Activity” [Ref. 6].
In the late 1950s, Frank Rosenblatt proposed the perceptron network and the
associated learning rule. In 1957, Rosenblatt published the first major research project in
neural computing which included the development of the perceptron element. The
perceptron is a pattern classification system, which could identify both abstract and
geometric patterns. In addition, the perceptron can make limited generalizations and can
properly categorize patterns despite noise in the input [Ref. 5]. This study showed the
first practical application of NN by demonstrating how NNs can perform pattern
recognition. In 1959, Bernard Widraw and Tedd Hoff proposed the Adaline (Adaptive
Linear Element), based on simple neuron-like elements and used it to train adaptive linear
networks. The Adaline and the two-layer Madaline version were used for a variety of
applications including speech recognition, character recognition, weather prediction, and
adaptive control. Widraw used the adaptive linear element algorithm to develop adaptive
filters that eliminate phone line echoes, in the first real life NN application [Ref. 6].
In the mid-1960s, Marvin Minsky and Seymour Papert, considered the NN

potential limitations. They showed that these already known networks could handle
linearly separable problems only and were usually not appropriate for real life
applications. As a result, NN research faded for a while.
In the 1970s, Kohonen, Grossberg and Anderson proposed the Kohonen Network
and the Self–organizing Network. Kohonen introduced the concept of the competitive
learning rule in which PEs compete to respond to an input stimulus and the winner adapts
itself to respond more strongly to that stimulus. This type constitutes an unsupervised
learning process and the internal organization of the network is governed only by input
stimuli [Ref. 5]. Grossberg’s contribution was a wealth of research towards the design
11
and construction of neural model as he used neurological data to build neural computing
models. Anderson developed a linear model, called a linear associator. Which is based on
models of memory storage, retrieval and recognition. In addition, Anderson improved the
model by combining it with a nonlinear post-processing algorithm, which is used to clean
up spurious responses. This model is called Brain-State-in-a –Box [Ref. 5].
In the 1980’s, NN became popular again with the back-propagation algorithm for
training multilayer perceptron networks. The concept of back-propagation algorithm was
presented by several researchers, such as David Parker, Yaun LeCun, David Rumelhart,
James McClelland, and Geoffrey Hinton. While a perceptron network is only capable of
solving linear problems, back-propagation network can solve more complex nonlinear
problems. This significant capability made the back-propagation networks the most
widely used networks.
In 1982, John Hopfield presented a paper describing his neural computing system
called the “crossbar associative network” or known as the “Hopfield Model”. This model
represented a neuron operation as a thresholding operation and illustrated memory as
information stored in the interconnections between neuron units. He also illustrated and
modeled the brain’s ability to call up responses from many locations in response to a
stimulus. Thus, this model represents how a NN associates information from many
storage sites for a given input [Ref. 6]. In the 1980s, the Bi-directional Associative
Memory (BAM) network, Boltzman Machine, the General Regression NN, and the
Learning Vector Quantization Network were developed, in addition to the back-
propagation and Hopfield models.
Although the concept of NN has been around for about 50 to 60 years, most
applications have appeared in the last fifteen years and the field is still developing very
rapidly. NNs can be found in many fields ranging from aerospace to medicine, banking
and robotics. Given the work done and range of applications, NN will most likely be a
permanent fixture not only as a solution to everyday problems but also as a tool to be
used in appropriate situations. It is certain that the more the structure of the brain is
understood, the more advances there will be in NN.
12
C. INTRODUCTION TO THE BACK-PROPAGATION NEURAL
NETWORK
Although in 1974 Paul Werbos described a similar algorithm to train multilayer

network, Back-propagation (BP) was invented independently by David Parker, Yaun
LeCun, David Rumelhart, Geoffrey Hinton and Ronald William in the mid 1980s.
Back-propagation is a very popular NN learning algorithm to train multilayer

networks with differentiable transfer functions to perform function approximation,
pattern association and classification. The algorithm is named for how it handles errors in
the network. The derivatives of the network error with respect to the network bias and
weights are computed starting from the very last layer up to the first layer in order to
modify the weights and biases of the network. The process starts from the output layer
and goes back to the first layer and is therefore called the “Back-Propagation Network”.
Typically, back-propagation network has an input layer, one or two hidden layers
and one output layer. Although there is no limit on the number of hidden layers, generally
one or two hidden layered networks are selected due to the complexity of the resulting
system. An example of a typical network is shown in Figure 5.
Figure 5. A Typical Back-Propagation Network [From Ref. 2].
A BPNN is typically represented with the following notation: R-S1-S2-S3, where

R is the number of inputs and Si is the number of neurons at layer i. In Figure 5, the input
vector to the network is shown by Pi. Transfer function is represented by fi and bi
represents bias. Depending on the selected transfer function, the BPNN can be used to
13
solve linear or nonlinear problems. In order to solve a nonlinear problem, a nonlinear
transfer function should be selected.
The first step in BPNN is to transfer the input forward through the network. The
output of one layer becomes the input to the following layer. Thus, the output of a
network can be defined as
aM = f M
( w M * a ( M −1) + b M ) , (2.1)
where M = 1,…….., R and R is the number of layers in the network. In Figure 5, for the
first layer, a(M-1) is a(0), which is the input of the network, and a(3) is the output of the last
layer.
The second step is to propagate the errors, or in other words, sensitivities

backward through the network from the last layer all the way to the first layer. The
BPNN algorithm is a generalization of the LMS algorithm, which uses the mean square
error as a performance index. As each input is applied to the network, the output, a, is
compared with the associated target value, t. The algorithm then adjusts the network
parameters in order to minimize the mean square error [Ref. 1].
Thus, given the error is defined as:
e = [t − a ] , (2.2)
and the expectation of mean square error is defined as
f ( x) = E ([eT e]) = E[(t − a)T (t − a)] , (2.3)
the expectation of the squared error at iteration k becomes [Ref. 1]
f ( x) = E[(t (k ) − a(k ))T (t (k ) − a(k ))] . (2.4)
The algorithm adjusts weights and biases as follows [Ref. 1]
∂f
w i,j (k+1) = w i,j (k) - α
M M
M
, (2.5)
∂w i,j
∂f
bi (k+1) = bi (k) - α
M M
M
,
∂bi
14
where α is the learning rate. This shows that the weights at any given iteration are equal
to the weights at the previous iteration adjusted by some fraction, α, of the sensitivity of
the error to that weight. In other words, the weights at each iteration are adjusted in a way
that reduces the error at the previous iteration.
The selection of α is usually done by trial and error. Too large an α often leads to
divergence of the learning algorithm while too small an α results in a slow learning
process [Ref. 9].
Note that the above equations include partial derivatives. Since the error is an
indirect function of the weights in the hidden layer, a chain rule is used to compute these
partial derivatives. These partial derivatives are called sensitivities and defined as:
∂f
siM = . (2.6)
∂niM
In the above equation si is called the sensitivity of f to changes at layer M in the ith
element of the net input.
The next step is to propagate the sensitivities backward
• M −1
M −1
s = f n M −1 ( wM )T s M , for M = R,….,1 . (2.7)
The final step is to adjust weight and biases with respect to these sensitivities, which
leads to
wM (k + 1) = wM (k ) − α s M (a M −1 )T , (2.8)
b M (k + 1) = b M (k ) − α s M .
Although the BP network is a very powerful technique, it usually requires a long

training time. However, using some tricks and heuristics can often help to obtain an
efficient BPNN implementation [Ref. 9].
In summary, the BPNN is a commonly used technique for solving nonlinear

estimation problems, provided that a nonlinear transfer function is selected. The BP
algorithm is based on a LMS algorithm that minimizes the squared error, where the chain
rule is used to compute the derivatives of the error with respect to weights and biases in
15
the hidden layers. The methodology in the BP network is a three-step process. First, the
input is propagated forward through the network and then the sensitivities are computed.
Second, starting from the last layer, these sensitivities are propagated backward through
the network. Finally, weights and biases are updated using these sensitivities.
16
III. DEVELOPMENT OF NEURAL NETWORK MODEL
A. NEURALWORKS PROFESSIONAL II/PLUS SOFTWARE
The model built for this analysis is derived from the NeuralWorks simulation
software program presented by NeuralWare, Inc. NeuralWorks is a user-friendly program
that makes it possible to not only select network parameters easily and quickly but also to
present network results effectively. Many NN algorithms such as BPNN, Radial Basis
Function NN, and LVQ, are included in the software. Once the algorithm type is chosen,
the next step is to define the network architecture, which includes specifying the number
of layers and the number of PEs associated with each layer, etc. Several types of learning
rules and PE transfer functions are embedded in the program. NeuralWorks also allows
the user to select learning rates and momentum terms. In addition, the extensive and
powerful instrumentation and diagnostic package allows the user not only to monitor and
adjust network parameters but also to display weights, errors, classification rates and
confusion matrices in graphical formats. Moreover, users can display the networks
through network or Hinton diagrams. These specifications make NeuralWorks
Professional II/Plus useful in designing, building, training, testing and deploying neural
networks to solve complex, real-world problems [Ref. 5].
The following picture shows a typical window of the Neural Works Professional
II/Plus software:
Figure 6. Typical Window of Neural Works Professional II/Plus Software.

17
B. FLIGHTLAB SIMULATOR
Flight simulation tools have been used extensively not only for training and
evaluation of aircrew members but also for the design and analysis of aerospace vehicles.
The need to use flight simulation systems instead of real aerospace vehicles depends on
the following reasons:
• The complexity of aircraft systems,
• High operating costs of aircrafts,
• The limitations of operating environment of aircrafts,
• Technological improvements in flight simulators.
In addition, flight simulators provide safe and effective conditions for training
purposes as instruction, demonstration and practice of certain maneuvers and procedures
that cannot be done during real flight conditions may easily be performed in a simulated
environment. Finally, longer training periods can be tolerated due to the low operating
cost of simulators. Therefore, modern training procedures benefit from simulation tools
extensively.
Simulators also play an important role in engineering processes such as design

and analysis of aircraft. It is certain that the dynamics of rotary-wing aircraft is much
more complex than those of fixed-wing aircraft. Consequently, helicopter simulation
systems require high computational power to include a complex set of computer
programs, very powerful and expensive computer systems and full motion based
simulation devices. Recent improvements in computer and simulation technology make
simulators able to produce data as correct as real helicopters. These improvements make
it possible to make use of simulators for design, test and analysis purposes.
FLIGHTLAB is a commercial-off-the-shelf type of software product developed

by Advanced Rotorcraft Technologies, Inc (ART) for the development and operation of
flight vehicle dynamics models in simulation applications especially for helicopters.
FLIGHTLAB is a tool that allows a user to build each section of a model separately and
then combine the pieces under the same framework similar to building a model by using
a finite element approach [Ref. 14]. This approach is very useful for engineering studies.
In order to provide a desktop pilot interface for FLIGHTLAB models, ART has
18
developed PilotStation, a turnkey executive for real time simulation operations that
couples image generation and pilot control inputs with the FLIGHTLAB flight dynamics
model to provide a low cost, pilot-in-the-loop simulation. PilotStation can also be used
with FLIGHTLAB code generated models to provide a low cost, real time simulation
capability using PCs [Ref. 7]. PilotStation, which can be run on UNIX or LINUX
operated computers, processes the high fidelity rotorcraft model and generates the cockpit
gauges and the window displays.
FLIGHTLAB uses a simulation language called Scope. This language is an

interpretive language that uses the industry standard MATLAB syntax, coupled with new
language constructs and a combination of C and Fortran computer languages to make the
building, testing and solving of nonlinear dynamic rotorcraft simulation models easy
[Ref. 7].
Another feature of the FLIGHTLAB Simulation System is Xanalysis. This tool

makes it possible to design, test and analyze rotorcraft models by allowing the user to
modify model parameters and perform a wide range of analyses on design alternatives. A
set of predefined test scenarios carry specific rotorcraft analyses such as performance,
stability and control, handling qualities and aerodynamic and structural loads. In addition
to these features, the simulation can be automatically configured to interact with the test
vehicle’s configuration and test conditions while using the time history of the test flight’s
control inputs to implement the simulation [Ref. 7]. These features are the required
characteristics necessary to implement the model built for this thesis and make use of
data provided by a simulator instead of a real test flight.
The following figure depicts the X-Analysis ``Flight Test Utility'' [Ref. 7]:
19
Figure 7. Xanalysis Window [From Ref. 7].
C. SELECTING DATA AND BUILDING THE MODEL

In the early 1990s, Kelly M. McCool and David J. Haas at Naval Surface Warfare
Center, Bethesda, MD, developed a study of UH-60 helicopter airspeed estimations using
NN in a low speed environment. McCool and Haas built and implemented a BPNN with
two hidden layers and 25 processing elements in each layer. The data used as input for
the NN were obtained from actual test flights performed at the Naval Air Warfare Center,
Aircraft Division, Patuxent River, MD. [Ref. 4]. Their study was used as a starting point
for this thesis since similar analysis was performed but with different data sets. They used
real flight test measurements as a source of data for their model. FLIGHTLAB simulator
outputs were used for NN input for this thesis. In order to evaluate how well the
simulator performs and how well the relationship between the simulator parameters and
low airspeed is developed, the simulator was first run for the UH-60A helicopter to
compare the results of the present study and the results of the actual test flights. The
simulation data was analyzed using a similar network architecture and network
parameters, to make a fair comparison. The goal of this approach is to determine whether
using a simulator instead of a real test flight is feasible or not.
The simulator was run for UH-60A helicopter from hover to 50 knots with 5 knots
intervals at various gross weights ranging from 16000 to 24000 lbs with 1000 lbs
intervals and at various sideslip angles from 0 to 3600 with 300 intervals to obtain the
input data for the NN model. However, sideslip angles were varied from 3000 to 600 for
velocities 35 knots and above because of the limitations of the UH-60A helicopter since
20
the helicopter cannot fly rearward or sideward when the speed is above 30 knots. The
altimeter was set to 85 feet and AGL to obtain the results when the helicopter is out of the
ground effect and at level flight. In a ground effect, analysis was performed by setting the
altimeter to 20 feet AGL for the UH-60A helicopter. The wind was assumed to be zero
for this analysis. Fifteen parameters were chosen as inputs for the neural network model
based on the variables described above, and are shown in Table 1.
MODEL VARIABLES FOR UH-60A HELICOPTER

Airspeed 0 - 50 Knots (5 Knots intervals)
Gross weight 16000 - 24000 lbs (1000 lbs intervals)
Sideslip angle 0- 360 degrees and 300 - 60 degrees (30 degrees intervals)
Altimeter (AGL) 20 ft (for in ground effect), 85 and (for out of ground effect)
Wind Speed 0
Pressure Altitude Sea level
NN MODEL INPUT PARAMETERS
1. Cyclic position (Lateral)
2. Cyclic position (Longitudinal)
3. Collective position
4. Pedal position
5. Roll rate
6. Pitch rate
7. Yaw rate
8. Pitch attitude
9. Roll attitude
10. Altimeter
11. Climb rate
12. Main Rotor Blade RPM
13. Engine torque
14. Gross weight
15. Sideslip angle
Table 1. Neural Network Input Parameters.
21
These fifteen parameters, which define the condition of the related part of the
helicopter at certain settings, are the outputs of the FLIGHTLAB simulator. The data was
split into two sets, one for training the network (training data set) and the other one for
evaluating the network performance (test data set). The first set used for training includes
the data related to airspeeds at 0, 10, 20, 30, 40, 50 knots at all gross weights and sideslip
angles. The second set used for testing was formed with the remaining input data.
The type of network was selected, after setting up training and test data sets. A
Back-propagation approach was selected as it was used before with success [Ref. 3]. A
Radial Basis Function Network (RBFN) was selected as an alternative network because it
can be used in most applications where back-propagation approaches may be used.
However the RBFN trains faster and leads to better decision boundaries than a BPNN in
many classification and decision problems. The learning rule used in RBFN is an
unsupervised learning rule. The results of these two models are presented in Chapter IV.
The next step in the BPNN scheme involved the selection of number of layers and
neurons in each layer. First the model in Reference 3 was used as a starting point, since
the first goal of this study was to compare the results of the model implemented by real
flight data and the model by simulator data. Thus, we selected two hidden layers and 25
neurons in each hidden layer, resulting in a 15-25-25-1 BPNN structure. Several models,
were also studied; 14-25-25-1, 15-18-25-1, 14-14-14-1, 14-14-12-1, and 14-14-10-1.
Next, the learning coefficient and momentum term were determined. Learning
coefficients control the changes in size of weights and biases during learning. Setting an
appropriate learning rate is significant because a small learning coefficient leads to very
little learning, which increases the training time. However, a large learning coefficient
may cause the performance index to diverge, meaning that no learning occurs. Therefore,
a small learning rate is generally used to avoid divergence. Next, the momentum term is
selected. The momentum term helps to obtain faster learning when using a low learning
rate. In this study, the appropriate values for these two terms were determined by trial and
error.
Finally, the learning rule and transfer function type were selected. There are
several learning rules available in NeuralWorks, such as, the Extended Delta-Bar-Delta
22
(Ext. DBD), Normalized Cumulative Delta (NCD), Quickprop, Delta Bar Delta (DBD)
and Delta Rule (DR). In this study, several types of learning rules were analyzed for the
UH-60A helicopter and the results are presented in Chapter IV. Based on the results of
the UH-60A helicopter model, only Ext. DBD and NCD learning rules were used for the
analysis of the OH-6A helicopter. NCD is a learning rule, which is immune to changes in
the epoch length, where an epoch is the number of sets of training data presented to the
network between weight updates. The Ext. DBD rule takes momentum term into
consideration. The transfer function is used to map the output of a neuron or a layer to its
actual output [Ref. 2] and may be linear or nonlinear. The most commonly used transfer
functions are depicted in Table 2 [Ref. 1].
NAME INPUT/OUTPUT RELATION
Hard Limit a=0 n<0

a =1 n ≥ 0
Linear a=n
Log-Sigmoid 1
a=
1 + e− n
Hyperbolic Tangent Sigmoid en − e− n

a=
en + e− n
Competitive a = 1 neuron with max n
a = 0 all other neurons
Table 2. Commonly Used Transfer Functions.
The sigmoid transfer function is commonly used in BPNN implementations

because it is differentiable. NeuralWorks provides the hyperbolic tangent and the sine
function as alternative functions. The hyperbolic tangent function is a bipolar version of
the sigmoid function. The sigmoid function maps the output between 0 and 1 smoothly
23
whereas the hyperbolic tangent function places it between –1 and 1. In this work, the
hyperbolic tangent transfer function (Tanh) is used.
Once the feasibility of this approach had been verified, the same type of
architectures and NN model parameters, were used for the OH-6A helicopter. These
parameters were the same as those given in Table 1, except for the engine torque
parameter. The only difference utilized for the second part of the analysis was the
altimeter setting. The altimeter was set to 12 feet for in ground effect data and 100 feet
for out of ground effect data for the OH-6A helicopter. Besides the altimeter setting,
other parameters, such as gross weight and pressure and altitude, were adjusted for the
OH-6A helicopter. Gross weight for OH-6A was defined ranging from 1500 lb. to 2500
lb. with 100 lb. increments. Pressure altitude was set to 90 ft for sea level and 6000 ft for
high-level altitude. For both pressure altitudes, in ground effect analyses were performed
at 12 ft, while out of ground effect analyses were performed at 100 ft.
24
IV. ANALYSIS AND RESULTS
The NN-based approach to determine helicopter low airspeed developed by Haas

and McCool has proven to be quite promising. In their study, a BPNN configuration with
the Extended Delta Bar Delta learning rule, 2 hidden layers and 25 PEs per layer was
selected [Ref. 3]. The data used as input to the NN were obtained from actual test flights
performed at the Naval Air Warfare Center, Aircraft Division, Patuxent River, MD.
Based on the model they implemented, and using the methodology described in the
previous section, in this thesis the NN model was first developed for the UH-60A
helicopter by using simulator data. The first model architecture developed for this study
was approximately the same model as that described in Ref. 3, since most of the network
parameters such as number of layers and number of neurons are determined by trial and
error. The first network selected was a 14-25-25-1 BPNN with the Ext. DBD learning
rule, and did not include the sideslip angle parameter as input to the network. Various
architectures were subsequently tried with different learning rules. The momentum
function was varied from 0.2 to 0.6 with the best results obtained at 0.4. Therefore, all
OH-6A analyses used a momentum function to 0.4. The NN output error airspeed was
measured in knots at 1 σ and in terms of the root mean square (RMS) error, as reported
by the NeuralWare software. The number of learning iterations was set to 50000 since the
network stabilized after approximately 20000 iterations for most of the implementations.
A helicopter flying close to the ground requires less power than when it is flying
far from the ground. The proximity of the ground to the rotor disk constrains the rotor
wake and reduces the induced velocity at the rotor, which causes a reduction in the power
required for a given thrust. This is called ground effect. A helicopter can hover in ground
effect at a higher gross weight or altitude than when it is out of ground effect. However,
in forward flight, the effect of the ground diminishes with the forward speed. Data sets of
both helicopters were obtained by running the simulator for two flight conditions: in-
ground effect (IGE) and out-of-ground effect (OGE). These data sets are called single
condition data sets. After training and testing with the single condition data sets
25
separately, the sets were combined for a baseline data set. The networks were retrained
and tested with the combined data set.
The OGE single condition data set has a dimension of [1144x14] for the UH-60A.
The IGE single condition data set has a dimension of [1144x15]. The difference is due to
the fact that the UH-60A OGE data did not include the sideslip angle. Sideslip angle was
included for the UH-60A IGE and combined data sets. For the OH-6A helicopter type,
the sideslip angle was included in all data sets, but the engine torque was not available, so
all OH-6A single condition data sets have dimensions of [1144x14]. The single condition
data sets were split into training and testing sets. Each training set is a subset of its single
condition data set with a dimension of [638x14] or [638x15]. Each testing set is made up
of the remaining elements of the single condition set and has a dimension of [506x14] or
[506x15]. Various networks were trained separately using these single condition data
sets. Later, the single condition data sets were combined forming a baseline data set with
a dimension of [2288x15] for the UH-60A and [2288x14] for the OH-6A. A baseline
training set was formed as a subset of the combined set with a dimension of [1276x15 or
14]. The remaining elements of the baseline set formed the baseline testing set with
dimensions of [1012x15 or 14]. The networks were retrained using these combined
training sets. Combined and single condition, OGE and IGE, results were obtained again
from each network that was trained with these baseline data sets. Finally, NN model
results were exported to MATLAB to develop the following tables and figures.
Each figure consists of four windows. The first window shows the relationship
between gross weight and the predicted speed related to that particular gross weight. The
second window shows the range of the predicted speed for each actual speed. The third
and fourth windows display the relationship between predicted speed and the sideslip
angle. In the third window, the speed range is from 0 knots to 30 knots, whereas in the
last window, the speed varies from 35 knots to 50 knots. In order to make the figures
more readable and easier to distinguish the NN predictions from one another, adjacent
speeds are illustrated with a different marker. NN outputs for 5 knots, 15 knots and 45
knots are presented with circles while outputs for 15 knots and 35 knots are demonstrated
with triangles. The structure of the NN, learning rule, helicopter type, RMS error and
altitude information about that figure is provided in the figure label.
26
Each table corresponds to a different NN type and consists of five columns. The
first column presents actual airspeed and target values for the NN model. The second
column gives the mean value of the NN outputs related to each target speed. The third
column shows airspeed errors at 1 σ in terms of knots. This error shows the standard
deviation (SD) around the mean and it is computed by the MATLAB function “std”. The
next column demonstrates the error percentage at 1 σ relative to the target speed. It is
computed by the following formula:
σi
Percent Error = * 100 , (4.1)
Vi
where σi is the error SD related to that speed and Vi is the actual speed. The last column
shows the absolute maximum error of the NN predicted speeds.
Results obtained for analyzed architectures follow in the next section.

A. ANALYSIS OF THE UH-60A MODEL
1. Out of Ground Effect (OGE) Analysis

This section shows the results of OGE analyses by setting the altimeter of the
simulator to 85 feet AGL. The network was trained using this single condition data set
obtained for this altitude. The results of OGE analysis using the combined data set are
shown in the UH-60A Baseline Analysis section.
a. 2-Hidden Layer BPNN
(1) 14-25-25-1 Ext. DBD. Results are shown in Figure 8 and in
Table 3. The RMS error is 0.0593. Although the network performance on the training
data is quite good, the test results are not close enough to the target speeds. Note that, the
NN produced a maximum speed error of 3.51 and 3.86 knots at 5 knots and 45 knots
respectively. A 0.7017 knots airspeed error SD was achieved by this setup. The
percentage of error is worst at 5 knots, where it is 8.5%, while it is about 3 % for other
speeds.
27
NEURAL NETWORK RESULTS
Actual Total SD = 0.7017
Airspeed (kts) Mean of Airspeed Error Percent Error Abs. Maximum
Airspeed (kts) at 1 σ (kts) at 1 σ Error (kts)
5 2.3322 0.4249 8.4979% 3.5125
15 16.1533 0.4601 3.0675% 1.8307
25 23.8216 0.5045 2.0179% 2.2057
35 36.2752 1.0928 3.1222% 3.1262
45 46.2789 1.4228 3.1617% 3.8625
Table 3. Results for the UH-60A helicopter at 85 ft.; network configuration 14-25-25-1;
Ext. DBD learning rule.
Figure 8. Results for the UH-60A helicopter at 85 ft.; network configuration 14-25-25-1;
(2) 14-25-25-1 NCD. Results for this configuration are shown

in Table 4 and in Figure 9. The RMS error is 0.0658 with the NCD learning rule. The
maximum value of the predicted speed error is ± 3.8584 knots. The speed prediction
accuracy within ± 1 σ is 0.7506 knots. The airspeed error SD at speeds 25, 35 and 45
28
knots is about 1 knot or 3 %. For speeds equal to 5 and 15 knots, the error percentage
increases to 6 %.
5 2.7932 0.3221 6.4425% 2.7872
15 17.0481 0.8209 5.4726% 3.0500
25 23.3043 0.6470 2.5878% 2.5837
35 36.8160 1.0996 3.1417% 3.8584
45 46.2666 1.1273 2.5051% 3.0720
NCD learning rule.
NCD learning rule.
29
(3) 14-25-25-1 DBD. Figure 5 and Table 10 summarize the
findings for this configuration. Results show the RMS error for this learning rule is
0.0721. This model predicts the speed very accurately when the actual speed is 45 knots
with a mean value of 45.19 knots, corresponding to a speed error prediction rate of
1.42%. However, when the speed is 15 knots and 35 knots the maximum error in the
prediction increases significantly to 4.9681 knots, while the error SD goes up to 1.1142
knots. The total airspeed error for the DBD configuration is equal to 0.7544 knots.
DBD learning rule.
5 4.8389 0.1628 3.2559% 0.4681
15 11.3654 0.5467 3.4647% 4.5417
25 24.5698 1.0986 4.3944% 2.2028
35 38.0388 1.1142 3.1834% 4.9681
45 45.1962 0.6442 1.4227% 1.2996
DBD learning rule.
30
(4) 14-25-25-1 QUICKPROP. Figure 11 and Table 6 present
results obtained for this configuration. The RMS error for the quickprop learning rule is
0.0711. The maximum predicted speed absolute value error is 4.6189 knots. Note that the
results are very close to those obtained for the DBD configuration. Although the mean
airspeed value at 15 knots is significantly lower than that of the target speed, the error SD
is the largest at 25 knots. The overall network airspeed error SD is 0.7326 knots.
5 4.5007 0.1677 3.3546% 0.8191
15 11.4068 0.6067 4.0444% 4.6189
25 24.8524 1.1168 4.4673% 2.3644
35 37.9663 0.8684 2.4811% 4.4915
45 44.6973 0.5294 1.1764 1.1524
Quickprop learning rule.
Quickprop learning rule.
31
b. 1-Hidden Layer BPNN
Several one hidden layer BPNNs were implemented; the configurations
considered were 14-8-1, 14-10-1, 14-12-1, 14-20-1, 14-22-1, 14-25-1, and 14-50-1.
Results for best performing networks are shown below.
(1) 14-20-1 NCD. Table 7 and Figure 12 summarize the
findings for this configuration. The 14-20-1 NCD model performs quite well for the
velocity of 25 knots with the error SD of 0.5 knots and percent error of 2 %. For all other
velocities, the error percentage is 4 %. The RMS error is found to be 0.0407. The
maximum error is 3.3 knots at the speed of 5 knots and 35 knots. However, the accuracy
of prediction is better than many 2-hidden layer BPNN. The overall error SD is ± 0.8394
knots.
Figure 12. Results for the UH-60A helicopter at 85 ft.; network configuration 14-20-1; NCD
learning rule.
32
5 3.6992 0.2008 4.0170% 3.3712
15 13.7729 0.6762 4.5079% 1.6789
25 25.0478 0.5373 2.1494% 1.2362
35 35.6197 1.5618 4.4623% 3.3712
45 45.4842 1.5915 3.5366% 2.8769
Table 7. Results for the UH-60A helicopter at 85 ft.; network configuration 14-20-1; NCD
learning rule.
(2) 14-20-1 Ext. DBD. Table 8 and Figure 13 summarize the
findings obtained for this configuration. Results show the RMS error is 0.0657 and that
the error SD increases along with the speed. The maximum error is 3.7970 knots for a 35
knots speed. The error SD is also the highest at this speed. However, results show a
significant degradation of the airspeed prediction in % at 5 knots. The overall error SD is
0.7002, which is quite better than that obtained with the other networks.
Figure 13. Results for the UH-60A helicopter at 85 ft.; network configuration 14-20-1; Ext.
DBD learning rule.
33
5 2.8849 0.4173 8.3461% 2.9556
15 17.4373 0.7170 4.7798% 3.2811
25 23.7263 0.6597 2.6389% 2.3582
35 36.9723 1.0187 2.9105% 3.7970
45 46.1386 0.9508 2.1129% 2.6505
Table 8. Results for the UH-60A helicopter at 85 ft.; network configuration 14-20-1; Ext.
DBD learning rule.
2. RBFN Networks
Figure 14 and Table 9 present findings obtained for the 14-200-1 RBFN
configuration. Several RBFN network configurations were implemented. The best result
was obtained with the 14-200-1 NCD configuration; where the RMS error is 0.0733.
However, results also show a larger velocity variance and maximum error than those
obtained with BPNN implementations. In addition, the maximum speed error is 9.2719
knots, which is much larger than that observed with BPNN implementations. The error
SD is around 1 knot at all speeds, except at 45 knots where it significantly increases to
about 4 knots. The overall error SD is 1.634. Therefore, results showed that BPNN
implementations are better suited than RBFN configurations for this problem.

5 2.1368 1.0915 21.8301% 5.4400
15 14.3318 1.0551 7.0338% 2.7856
25 25.6402 1.1621 4.6485% 2.9813
35 35.9359 1.4807 4.2307% 3.8438
45 44.0553 3.8565 8.5699% 9.2719
Table 9. Results for the UH-60A helicopter at 85 ft.; network configuration 14-200-1;
RBFN NCD learning rule.
34
Figure 14. Results for the UH-60A helicopter at 85 ft.; network configuration 14-200-1;
RBFN NCD learning rule.
3. Pruned BPNN
Figure 15 and Table 10 present the findings for this configuration. The BPNN was
implemented with a pruning capability embedded in the NeuralWare software to increase
the prediction efficiency, where pruning is used to remove connections from the network
when their contributions are very small [Ref. 6].
In light of the previous analyses, the 14-25-25-1 BPNN configuration with the
NCD learning rule was chosen. In this scheme, data were presented to the network as
before but the network performance was monitored against the previous ones during the
training process after each 1000 iterations. The RMS error was compared at each
“checkpoint”, and PE contributions checked. At that point the training stopped when the
RMS was larger than at the last check, otherwise, it continued for the next 1000
iterations. During this process, PEs with contributions smaller than a given tolerance
35
were disabled, which reduces network complexity and increases generalization efficiency
[Ref. 4].
In this analysis, the process went up to 46000 iterations and 7 PEs in the first
hidden layer were disabled. Results show that the RMS error is 0.0355 which is best
among all architectures investigated. The maximum predicted speed error is ± 3.35 knots,
which occurs when the actual speed is 35 knots. Errors for low velocities are below ± 1.8
knots. The error percentage for 5, 15 and 35 knots is about 4 % and about 2.5 % for other
speeds. Results show that the predicted speed mean values are very close to the target
values when compared with previous networks. Finally, the overall error SD is 0.7167,
which is quite good.
NCD (Pruned) learning rule.
36
5 3.6789 0.2113 4.2253% 1.7089
15 14.5970 0.6322 4.2147% 1.6131
25 24.9281 0.6610 2.6440% 1.2897
35 36.0136 1.3438 3.8395% 3.3496
45 46.0334 1.0039 2.2309% 2.6618
NCD (Pruned) learning rule
The above results showed that the networks with NCD and Ext. DBD learning
rules and pruned network have better performance than those obtained with other
networks. Therefore, all other analyses, UH-60A in-ground effect, baseline data and OH-
6A analyses, were performed using these configurations only.
4. In-Ground Effect Analysis
In-ground effect analysis of the UH-60A helicopter is performed when the
pressure altitude is at sea level and the altimeter above ground is at 20 feet. The sideslip
angle parameter was added to the NN inputs. Three network architectures were used for
the in-ground effect analysis; 15-25-25-1 BPNN with NCD and Ext. DBD learning rules,
and 15-18-25-1 BPNN with the NCD learning rule.
a. 15-25-25-1 BPNN NCD
Figure 16 and Table 11 summarize the findings for this configuration. The
RMS error is 0.0374 and the maximum error is 3 knots (at 35 knots). The error SD is also
the highest at this speed. The network performance at low speeds and at 45 knots is quite
good, with the error SD of less than 1.2, but the error SD goes up to 1.8 knots for a 35
knots speed, which corresponds to about 5 %. The overall airspeed error SD is 0.8469
knots.
37
5 4.1567 0.2334 4.6686% 1.3234
15 14.0652 0.5405 3.6033% 2.1345
25 25.8750 0.6741 2.6963% 2.5196
35 35.5357 1.8206 5.2017% 3.0036
45 45.2548 1.2123 2.6941% 2.2981
NCD learning rule.
NCD learning rule.
b. 15-25-25-1 BPNN Ext. DBD
Figure 17 and in Table 12 present the results for this configuration. The
RMS error obtained is 0.0637. Note that the maximum error occurs when the helicopter is
moving at 45 knots with a -600 angle, unlike other networks. Moreover, the maximum
error, 4.81 knots, is higher than that obtained in most networks. Apparently, the error SD
is also larger at 45 knots. Although the error SD at low speeds is less than 0.88 knots, the
error percentage is about 6%. Overall, the error SD is 0.8308 knots.
38
5 2.5383 0.3051 6.1027% 3.1938
15 16.7466 0.8875 5.9166% 3.7167
25 23.4740 0.5188 2.0750% 2.1472
35 35.6272 0.9530 2.7229% 2.1010
45 46.5186 1.6658 3.7018% 4.8112
c. 15-25-25-1 BPNN NCD (Pruned)
Table 13 and Figure 18 present the results for this configuration. Seven
PEs were disabled in the first hidden layer. The RMS error is about 0.0724. The
maximum error is 5.2 knots at 35 knots. The error SD is also larger at this speed. The
39
overall error SD is 0.8884 knots which is slightly larger than the NCD 15-25-25-1
network.
5 4.7761 0.2380 4.7603% 0.7986
15 11.4320 0.4591 3.0609% 4.8662
25 24.8531 1.3065 5.2260% 2.8201
35 37.9958 1.4152 4.0435% 5.2195
45 45.4203 0.7768 1.7263% 1.9679
d. 15-18-25-1 BPNN NCD
Table 14 and Figure 19 present the results for this configuration. This
setup is implemented as an alternative to the pruned network to reduce the network
40
training time and complexity. The RMS error is about 0.0384. The maximum error is
2.27 knots at 35 knots. We note the error SD is larger at this speed. The overall error SD
is 0.8853 knots, which is slightly higher than the NCD 15-25-25-1 network.
5 3.7156 0.2597 5.1943% 1.7481
15 15.0498 0.6162 4.1081% 1.3976
25 24.1842 0.9682 3.8726% 1.7528
35 34.6029 1.5548 4.4424% 2.2788
45 45.5511 1.2541 2.7870% 2.2318
NCD learning rule. Results of UH-60A for 15-18-25-1 NCD.
NCD learning rule.
41
5. Baseline Data Set Analysis
The baseline data set was formed combining both 20 ft. and 85 ft. data sets. First,
the network was trained with this data set, and the network performance was measured
using the baseline test set. Then, using single condition test data sets separately, the
network performance was evaluated for IGE and OGE conditions. As a result, the
network-training time increased, due to the increase in the number of data points in the
baseline data set.
a. 15-18-25-1 BPNN NCD
Results for this setup are shown in Table 15 and in Figure 20. They show
that the RMS error is 0.0755. The maximum error is about 6.3 knots when the speed is 35
knots. The error SD of 1.4224 knots is higher than that observed with other baseline
networks and the single condition data networks. The error percentage for all speeds at 1
σ is about 6 %. The error SD is higher at fast speed, especially when the sideslip angle is
–60 and 60 degrees.
Figure 20. Results for the UH-60A helicopter with baseline data; network configuration 15-
18-25-1; NCD learning rule.
42
5 2.9564 0.3565 7.1292% 2.8415
15 16.4937 0.9016 6.0108% 4.2300
25 26.0262 1.5856 6.3426% 5.1133
35 38.8061 2.3985 6.8529% 6.2953
45 45.4173 2.1685 4.8188% 4.1447
Table 15. Results for the UH-60A helicopter with baseline data; network configuration 15-
b. 15-25-25-1 BPNN NCD
Results are depicted in Figure 21 and Table 16. They show that the RMS
error is 0.0762, and the maximum error is 6.2424 knots at 45 knots. The airspeed error
SD is larger when the speed is 25 knots. The largest error percentage is equal to 9 % (at 5
knots). The overall network error SD is 1.5664 knots.
43
5 3.7732 0.4005 8.0108% 1.8981
15 13.4327 1.0494 6.5960% 3.1877
25 26.4841 2.2515 9.0058% 6.1337
35 38.4891 1.9183 5.4810% 5.1834
45 43.3588 1.9560 4.3466% 6.2424
c. 15-25-25-1 BPNN Ext. DBD
Table 17 and in Figure 22 present the findings for this configuration.
Results show that the RMS error is 0.0501, which is better than that of the NCD scheme.
The maximum error is about 4.5 knots (at 35 knots). We note the predicted speed mean
values are close to the target speeds. The error percentage at the speed of 5 knots is about
10 %. The NN prediction is quite good at the speed of 45 knots. The overall network
error SD is 0.7320, which is the best obtained with the baseline data set.

5 3.2079 0.5382 10.7647% 2.8105
15 16.2084 0.6293 4.1953% 2.5304
25 25.0000 0.7696 3.0782% 2.3555
35 37.3300 1.0832 3.0949% 4.5313
45 45.9893 0.8478 1.8839% 1.9112
25-25-1; Ext. DBD learning rule.
44
25-25-1; Ext. DBD learning rule.
d. 15-25-25-1 BPNN NCD (Pruned)
Table 18 and Figure 23 present the findings for this configuration. Results
show that one PE was pruned in the first hidden layer. The resulting RMS error is 0.1213,
and the maximum error at 45 knots is 8.6414 knots. The largest error SD and error
percentage (11.5 %) are obtained for a 25 knots speed. The overall network error SD is
2.0830 knots.
5 2.9069 0.3420 6.8404% 2.8877
15 9.8217 1.1116 7.4108% 7.2736
25 23.5765 2.8864 11.5454% 6.5795
35 38.3384 2.8346 6.6262% 7.1468
45 43.8399 2.9818 1.8839% 8.6414
25-25-1; NCD (Pruned) learning rule.
45
25-25-1; NCD (Pruned) learning rule.
e. Baseline Data Set IGE Analysis
(1) 15-25-25-1 BPNN NCD. Results are shown in Table 19
and Figure 24. They show the RMS error is 0.0759. The maximum error is 6.2 knots,
which is almost the same for speeds 25 knots and higher. The network performance at
high speeds is better than that of at low speeds since the percent error at 35 and 45 knots
is about 5 %, whereas at lower speeds it is twice that number. The overall error SD is
1.5516 knots.
5 3.7699 0.5063 10.1268% 1.8981
15 13.7003 1.1242 7.4947% 2.7778
25 26.5671 2.1650 8.6599% 6.1337
35 38.3892 1.7970 5.1342% 5.1715
45 42.8593 2.0669 4.5931% 6.2424
configuration 15-25-25-1; NCD learning rule.
46
(2) 15-25-25-1 BPNN Ext. DBD. Results are shown in Table
20 and Figure 25. They show the RMS error is 0.0499. The maximum error of 4.5 knots
is obtained at a 35 knots speed. Note that the network performance is quite good when
compared with the NCD scheme. The overall error SD is 0.6505 knots, which is
significantly less than it is in the NCD scheme.
5 3.5127 0.4383 8.7656% 2.2475
15 16.4063 0.6193 4.1287% 2.5304
25 25.1852 0.7545 3.0179% 2.3505
35 37.6802 0.9720 2.7772% 4.5313
45 46.0822 0.4932 1.0961% 1.8867
configuration 15-25-25-1; Ext. DBD learning rule.
47
(3) 15-25-25-1 BPNN NCD (Pruned). Results are shown in

Table 21 and Figure 26. They show that the RMS error is 0.0901. The maximum speed
error of 7.9 knots is obtained for a 45 knots speed. We note that the network performance
is not good when the pruning facility is enabled. The overall error SD is 1.7173 knots.
5 4.3654 0.2933 5.8653% 1.1256
15 11.5338 1.1787 7.8578% 4.9166
25 23.9066 2.2817 9.1268% 4.0707
35 37.3058 2.1850 6.2429% 4.4207
45 42.1901 2.5806 5.7346% 7.9128
configuration 15-25-25-1; NCD (Pruned) learning rule.
48
f. Baseline Data Set OGE Analysis
(1) 15-25-25-1 BPNN NCD. Findings for this configuration
are presented in Table 22 and Figure 27. Results show that the RMS error is 0.0766. The
maximum error is 5.86 knots. The network performance at a speed of 25 knots is not
good, as the error SD is 2.3 knots at that speed. The error percentage is about 5 % for
speeds other than 25 knots but it increases up to 9 % for 25 knots. The overall error SD is
1.5516 knots.
5 3.7779 0.2569 5.1390% 1.6144
15 13.1631 0.8966 5.9771% 3.1890
25 26.4013 2.3416 9.3664% 5.8619
35 38.5880 2.0476 5.8502% 5.1844
45 43.8585 1.7192 3.8204% 4.3961
49
(2) 15-25-25-1 BPNN Ext. DBD. Table 23 and in Figure 28
present the results for this configuration. Results show that the RMS error is 0.0482. The
maximum error observed at 35 knots is equal to 3.8 knots. Note that this error is smaller
than that observed in the NCD scheme, and the network performance is quite good when
compared with the NCD scheme for the OGE baseline setup. At low speeds, the error SD
is about 0.5 knots. The network performance is quite good for all speeds except for 5
knots where the error percentage increases to 9 %. The overall error SD is 0.6850 knots.
50

5 2.9227 0.4746 9.5519% 2.8377
15 15.8606 0.5962 3.9746% 1.9213
25 24.8085 0.6801 2.7205% 1.5601
35 36.8469 1.0210 2.9171% 3.8006
45 45.9612 0.9043 2.0096% 1.9007
(3) 15-25-25-1 BPNN NCD (Pruned). Table 24 and in
Figure 29 present the results for this configuration. Results show that the RMS error is
0.1087. The maximum error observed for a 35 knots speed is equal to 7.14 knots. Note
that the network performance degrades when the pruning facility is enabled. The overall
error SD is 1.8642 knots.
51

5 3.7243 0.1567 3.1338% 1.6318
15 10.3127 0.5639 3.7592% 5.7803
25 24.6236 2.3566 9.4263% 4.6493
35 39.1423 3.2081 9.1660% 7.1416
45 45.3547 2.8266 6.2814% 3.5760
6. Simplifying The Data Set Using Eigenvalues and Eigenvectors
Preprocessing the data to obtain a simplified NN is often useful. Goff developed a

NN helicopter airspeed prediction study similar to the Haas and McCool study [Ref. 20].
Goff analyzed the each input contribution to the network performance and determined
that 11 of the 16 inputs played an important role. Results obtained by preprocessing the
52
data and simplifying the network inputs were quite successful. His study on the Lynx
MK9 showed that airspeed in the low speed environment could be predicted with an error
of +/-3.1 knots at 1 σ [Ref. 20].
Another way to decrease the NN input dimension and simplify the network
structure by projecting the high-dimensional data onto a lower dimensional input space
using principal component analysis (PCA). PCA is applied by considering the input data
set as a matrix, A, and translating this matrix to a diagonal or upper triangular form to
compute its eigenvalues and eigenvectors. Eigenvectors are the normal modes of the
system and they act independently. The beauty of eigenvalues and eigenvectors is that
they capture the characteristics and behavior of the whole system. The key equation for
eigenvalues and eigenvectors is:
Ax = λ x , (4.1)
where λ is an eigenvalue of the matrix A and the vector x is the associated eigenvector.
Most vectors x will not satisfy this equation. A typical x changes direction when
multiplied by A, so that Ax is not a multiple of x. Thus, only certain special numbers are
eigenvalues and only certain special vectors are eigenvectors [Ref. 18].
After obtaining the eigenvalues and eigenvectors of the single condition data set,
dominating eigenvalues were selected. It was observed that six of the eigenvalues were
the dominating eigenvalues, leading to a input data sub-matrix with a dimension of
[14x6]. This sub-matrix was multiplied by the whole data set again. The purpose of this
procedure was to reduce the number of inputs to 6 while still capturing the properties
from all inputs. Several network architectures were implemented using this simplified
data set. Network configurations with the best performances are described in the
following sections. Note that none of these performed as well as the best network trained
with the full data sets.
a. 6-18-18-1 BPNN Ext. DBD
Results of this setup are shown in Figure 30 and in Table 25. They show
that the RMS error is 0.0565. Although the maximum error of 5.4 knots is at 45 knots, the
11% error percentage is higher for a 5 knots speed. The overall error SD is 1.1692 knots.
53
The network performance at low speeds is not as satisfactory as it is at fast speeds since
the percent error is quite large at low speeds.
Airspeed (kts)Mean of Airspeed Error Percent Error Abs. Maximum
5 2.9375 0.5879 11.7571% 3.0269
15 15.3870 0.7861 5.2444% 2.2.1139
25 23.8447 1.2453 4.9813% 3.5517
35 35.4789 1.4587 4.1677% 2.5422
45 43.5555 2.2160 4.9244% 5.3797
b. 6-18-18-1 BPNN NCD (Pruned)
Results are depicted in Figure 31 and Table 26. One PE in the first layer
was disabled by the NN. Results show that the RMS error is 0.0648. The maximum error
is over 5.0 knots at speeds of 25 and 35 knots. The error SD is also larger at those speeds.
We note that, the maximum error percentage is 7.95 % at a speed of 5 knots, and the
54
overall error SD is 1.1942 knots. In conclusion, results showed that the network
performance is not as good as that obtained when using the NCD learning rule without
the pruning facility.

5 3.5084 0.3977 7.9568% 1.9809
15 14.7297 0.5445 3.6297% 1.1050
25 26.7339 1.8057 7.2229% 5.4495
35 38.2674 1.8897 5.3992% 5.3390
45 43.6582 0.9459 2.1020% 2.6491
55
c. 6-18-18-1 BPNN NCD
Figure 32 and Table 27 present the results for this configuration. Results
show that the RMS error equal to 0.06561 is the best among all the configurations
investigated with the simplified data. The maximum error is about 4 knots when the
helicopter is moving at a speed of 35 knots with a 300 sideslip angle to the left. The error
SD is above 1.5 knots at that speed. The error percentage for each speed is about 5 %,
except for a 45 knots speed, and the overall error SD is equal to 1.0032 knots.
configuration 6-18-18-1; NCD learning rule
5 2.7558 0.2888 5.5765% 2.7471
15 15.3714 0.6022 4.0149% 1.9253
25 25.0220 1.3322 5.3287% 2.7581
35 37.1657 1.5913 4.5467% 3.8912
45 43.5339 1.2950 2.8777% 3.1240
56
Results showed that the network with simplified data performs worse than
the network with the pruning facility enabled. Therefore, the simplified data was not used
for OH-6A analyses.
B. ANALYSIS OF THE OH-6A MODEL
1. Out of Ground Effect Analysis at Sea Level
This section presents results of the OGE analysis for the OH-6A helicopter with
the simulator altimeter set at 100 feet AGL. The network was trained using the single
condition data set obtained for this altitude similarly to the UH-60A analysis. Results for
the OGE analysis using the combined data set are shown in the OH-6A Baseline Analysis
Section. Note that the engine torque input parameter for the NN model was removed due
to the limitation of the engine model parameters for the OH-6A helicopter in the
FLIGHTLAB simulator. Therefore, a 14-25-25-1 setup was used for all OH-6A model
analyses. Finally, in light of the results obtained from the UH-60A helicopter analysis,
only the NCD and Ext. DBD learning rules were investigated.
a. 14-25-25-1 BPNN NCD
Results for this setup are given in Figure 33 and Table 28. They show that
the RMS error is 0.0609. The 5 knots maximum speed error is observed at 35 knots,
while the absolute maximum error is less than 3 knots at other speeds. The airspeed error
SD is less than 1 knot except for 45 knots. The overall error SD is 0.7921 knots.

5 4.2398 0.4797 9.5947% 1.5363
15 12.7387 0.7130 4.7531% 3.3062
25 25.6227 0.8655 3.4618% 2.0925
35 38.6028 0.8703 2.4865% 5.2048
45 44.8812 1.2253 2.7229% 1.9604
14-25-25-1; NCD learning rule.
57
Results are shown in Figure 34 and Table 29. They show that the RMS
error is 0.0735. Note that the absolute maximum error is about 5 knots at the speed of 35
knots, but the overall error SD of 1.3207 knots is larger than that observed for the NCD
setup. For all speeds, except for 45 knots, the error percentage is 5 % and over.
5 3.6946 0.4724 9.4478% 2.3008
15 13.3800 1.2846 8.5642% 3.8498
25 25.3109 1.2145 4.8578% 3.2181
35 36.3450 2.6079 7.7543% 5.0472
45 43.8989 1.1780 2.6179% 3.2799
14-25-25-1; Ext. DBD learning rule.
58
Figure 35 and Table 30 present the findings for this configuration. Results
show that the RMS error is 0.0669. The maximum speed error is about 4.6 knots at a
speed of 35 knots. The error SD is less than 0.88 knots except for a 45 knots speed. The
network performance is quite good except for 5 knots as the error percentage is equal to 9
%, which is very large at that speed. The overall error SD is 0.759 knots.
5 3.9667 0.4976 9.9521% 1.9224
15 12.0952 0.6239 4.1595% 4.0568
25 25.1773 0.8877 3.5509% 1.7791
35 38.4750 0.6479 1.8510% 4.6989
45 44.9067 1.2291 2.7914% 1.9686
14-25-25-1; NCD (Pruned) learning rule.
59
2. In-Ground Effect Analysis at Sea Level
The OH-6A helicopter in-ground effect analysis was performed at 12 ft AGL.
Same network architectures as those considered with out-of-ground analyses were
considered. Results are presented below.
a. 14-25-25-1 BPNN NCD
Results are given in Table 31 and Figure 36. They show that the RMS
error for this architecture is 0.0592. The maximum error is 5 knots for a speed equal to 35
knots. The percentage of error at 1 σ is about 3% for all speeds except for 5 knots. The
speed error SD gets larger at speeds above 25 knots. The overall error SD is 0.8465
knots.
60

5 4.4013 0.4882 9.7633% 1.4300
15 13.0112 0.3802 2.5348% 2.7769
25 25.9480 1.1424 4.5695% 2.7853
35 38.5135 0.9581 2.7373% 5.1048
45 45.3026 1.2993 2.8872% 2.1848
14-25-25-1; NCD learning.

show that the RMS error is 0.062. The maximum error is equal to 6 knots (at 35 knots).
The maximum error percentage is 12 % and occurs at 5 knots. The network performance
is quite good at 45 knots with an error SD of 0.8 knots. We note that the overall error SD
of 0.9678 is slightly higher than that of obtained with the NCD scheme.
61
5 4.9451 0.6122 12.2449% 1.1009
15 15.4143 0.6925 4.6168% 1.6649
25 27.1087 1.1644 4.6577% 3.7000
35 38.7739 1.6571 4.7345% 5.9939
45 45.7621 0.8066 1.7952% 2.0848
Figure 38 and Table 33 present the results for this setup. Results show that
the RMS error is 0.0674 and the maximum error of 4.4 knots occurs at 35 knots. Note
that the maximum error is about 2 knots for all other speeds. The error SD is the largest at
25 and 45 knots. The overall network error SD is 0.8803 knots.
62
5 4.1006 0.6088 12.1766% 1.9135
15 12.2608 0.3937 2.6246% 3.4638
25 25.5791 1.1830 4.7320% 2.5694
35 38.5929 0.5935 1.6956% 4.4157
45 45.1430 1.4917 3.3148% 2.0239
3. OH-6A Baseline Data Analysis at Sea Level
The baseline data set was formed by combining both 12 ft. and 100 ft. data sets.
First the network was trained with this combined data set, and the network performance
was measured with the baseline test set. Then, using single condition test data sets
separately, the network performance was evaluated for IGE and OGE conditions.
63
a. 14-25-25-1 BPNN NCD
Results for this configuration are shown in Figure 39 and Table 34. They
show that the RMS error is 0.0622. The absolute maximum speed error occurs at 35 knots
and is equal to 4.82 knots. The airspeed error SD at 1 σ is more than 1 knot at the speed
of 25 knots and over. The overall error SD is 1.2 knots.
Figure 39. Results for the OH-6A helicopter with baseline data (SL); network configuration
5 3.6199 0.4348 8.6964% 2.3155
15 13.2606 0.4590 3.0599% 2.9471
25 25.8467 1.6476 6.5904% 4.3737
35 37.8501 1.1617 3.3190% 4.8224
45 44.8255 2.0805 4.6234% 2.8179
Table 34. Results for the OH-6A helicopter with baseline data (SL); network configuration
64
Table 35 and Figure 40 present the results for this setup. Results show that
the RMS error is 0.0687. The maximum speed error observed is 7.5 knots for 35 knots.
At most speeds, the error SD at 1 σ is more than 1.2 knots. Hence, the error percentages
are also larger when compared with other network performances. The overall error SD is
1.4071 knots.
5 2.9860 0.7204 14.4071% 3.5396
15 14.7369 1.3806 9.2038% 4.0839
25 26.0708 1.2023 4.8092% 4.9530
35 38.3260 2.6810 7.6601% 7.4491
45 45.6813 1.3554 3.0120% 2.9237
65
Findings for this configuration are presented in Table 36 and Figure 41.
The pruning disabled 3 PEs in the first hidden layer. The RMS error is 0.0611 and the
maximum speed error is 4.98 knots at 25 knots. The error SD is greater than 1.2 knots for
the speeds 25 knots and higher. The overall error SD is 1.1689 knots.

5 3.9731 0.3442 6.8035% 1.7741
15 13.4709 0.3862 2.5744% 2.3807
25 26.1026 1.6586 6.6343% 5.0413
35 38.1568 1.2776 3.6504% 4.9842
45 44.9858 1.9399 4.3109% 3.1513
66
d. OH-6A SL Baseline Data IGE Analysis
(1) 14-25-25-1 NCD IGE. Table 37 and Figure 42 present
detailed results for this configuration. Results show that the RMS error is 0.0593. The
maximum speed error is 4.5 knots observed for a 35 knots speed. The largest error SD is
2.14 knots at 45 knots. The overall error SD is 1.1501 knots.

5 3.7499 0.4647 9.2946% 2.2505
15 13.4738 0.3264 2.1762% 2.1856
25 26.0936 1.5397 6.1587% 4.2902
35 37.6622 1.1159 3.1883% 4.4943
45 44.6424 2.1425 4.7612% 2.8106
67
(2) 14-25-25-1 NCD (Pruned) IGE. Results are given in
Table 38 and Figure 43. They show that the RMS error is 0.0592. The maximum speed
error is 4.5 knots observed for a 35 knots speed. The overall error SD is 1.155 knots. We
note that no significant improvement was obtained when compared with the un-pruned
NCD scheme.
Airspeed (kts) at 1 σ (kts) At 1 σ Error (kts)
5 3.8734 0.4085 8.1709% 2.0251
15 13.5302 0.3172 2.1143% 2.0447
25 26.1493 1.5478 6.1914% 4.1953
35 37.7847 1.0744 3.0696% 4.5171
45 44.7929 2.2042 4.8981% 2.8103
68
(3) 14-25-25-1 Ext. DBD IGE. Results are illustrated in
Figure 44 and Table 39. They show that the RMS error is 0.0686. The maximum speed
error is 7.4491 knots observed for a 35 knots speed. We note that neither the prediction of
low speed nor the prediction of fast speed is as good as that obtained with the NCD setup.
The overall error SD is 1.3293 knots.
5 3.4696 0.5957 11.9134% 2.7024
15 15.3847 1.1166 7.4493% 2.5874
25 26.5757 1.1507 4.6029% 4.9330
35 38.6624 2.7151 7.7574% 7.4491
45 45.8415 1.4335 3.1855% 2.9237
69
e. OH-6A SL Baseline Data OGE Analysis
(1) 14-25-25-1 Ext. DBD OGE. Results for this network are
depicted in Figure 45 and Table 40. They show that the RMS error is 0.0687. The
maximum speed error is 6.6 knots observed for a 35 knots speed. The error SD is over 1
knot almost at all speeds. The overall error SD is 1.3015 knots.
5 2.5071 0.4741 9.4826% 3.5396
15 14.0892 1.3168 8.7787% 4.0839
25 25.5659 1.0316 4.1264% 3.7463
35 37.9896 2.6280 7.5087% 6.6395
45 45.5210 1.2654 2.8119% 2.2839
70
(2) 14-25-25-1 BPNN NCD OGE. Figure 46 and Table 41
present the results for this setup. Results show that the RMS error is 0.0671. The
maximum speed error is 4.8 knots observed for a 35 knots speed. The error SD is above 1
knot for speeds 25 knots and higher. The overall error SD is 1.2022 knots.

5 3.4898 0.3599 7.1982% 2.3155
15 13.0474 0.4741 3.1618% 2.9471
25 25.5998 1.7194 6.8786% 4.3737
35 38.038 1.1861 3.3887% 4.8224
45 45.0086 2.0195 4.4874% 2.8176
71
(3) 14-25-25-1 BPNN NCD (Pruned) OGE. Results are
depicted in Figure 47 and in Table 42. Results show that the RMS error is 0.0650. The
maximum speed error is about 5 knots at 35 knots. The maximum error is about 2 knots
for all other speeds. One PE in the first hidden layer was disabled. The overall error SD is
1.2 knots. Note that performance of this setup is better than that of the Ext. DBD
configuration.
5 3.8439 0.3031 6.0621% 1.8782
15 13.3969 0.5177 3.4513% 2.5735
25 25.9133 1.7329 6.9317% 4.5386
35 38.2975 1.184 3.3828% 5.1005
45 45.2226 2.0303 4.5118% 2.8686
72
4. OH-6A Out of Ground Effect Analysis at High Altitude
The simulator was run at a pressure altitude of 6000 feet for high altitude analysis.
At that pressure altitude, the altitude AGL was set to 100 ft. for the OGE condition,
which is the same AGL altitude, that the sea level OGE analysis was performed at.
a. 14-25-25-1 BPNN NCD
Results are presented in Figure 48 and Table 43. They show that the RMS
error for this setup is 0.0485. The maximum speed error, which is 2.47 knots, and the
maximum error percentage equal to 9.4 % are both observed for a 5 knots speed. The
maximum airspeed error SD of 1.1356 knots occurs at 35 knots. The overall error SD is
0.6637 knots.
73
5 3.5101 0.4738 9.4794% 2.4762
15 16.209 0.6861 4.5744% 2.2914
25 23.272 0.5601 2.2405% 2.3443
35 35.288 1.1356 3.2444% 1.8547
45 45.985 0.6179 1.3732% 1.8831

show that the RMS error is 0.0958. This is one of the largest errors observed for the OH-
6A analysis. The maximum error is 6.36 knots, which is more than twice the error of the
NCD setup. The error percentage at 5 knots is better than that of the NCD configuration.

5 3.1676 0.2838 5.6769% 2.3937
15 19.875 0.7859 5.2395% 6.3677
25 23.339 1.2915 5.1659% 5.3069
35 33.995 1.9872 5.6777% 3.4085
45 45.699 0.3127 0.6917% 1.1757
74
Figure 50 and Table 45 present the findings for this configuration. The
pruning disabled 1 PE of the first hidden layer. Results show that the RMS error is
0.0475. The maximum speed error is 2.85 knots observed for 5 knots speed. The results
for fast speed predictions are better than the those at low speeds. The overall error SD is
0.6401 knots.
5 2.8147 0.3546 7.0932% 2.8415
15 14.812 0.7660 5.1070% 1.4709
25 24.0 0.7102 2.8410% 2.2145
35 36.208 0.7016 2.0047% 2.0621
45 46.040 0.6219 1.3821% 2.0527
14-25-25-1; NDC (Pruned) learning rule.
75
14-25-25-1; NDC (Pruned) learning rule.
5. OH-6A In-Ground Effect Analysis at High Altitude

In-ground effect analyses of the OH-6A at high altitude were implemented by
setting the simulator pressure altitude to 6000 feet. At that pressure altitude, the altitude
AGL was set to 12 ft. for the IGE condition, which is the same AGL altitude at which sea
level IGE analyses were performed.
a. 14-25-25-1 BPNN NCD
show that the RMS error is 0.0516, and the maximum speed error is 2.596 knots for a 5
knots speed. The maximum error percentage equal to 9.12 % is also observed at that
speed. The maximum airspeed error SD of 1.20 knots occurs at 35 knots. The overall
error SD is 0.6855 knots.
76
14-25-25-1; NDC learning rule.

5 3.4367 0.4811 9.1210% 2.5962
15 16.3790 0.6755 4.503% 2.4700
25 23.1736 0.5637 2.2547% 2.5948
35 35.3076 1.2062 3.4462% 2.0407
45 46.0330 0.7075 1.5722% 2.0734
14-25-25-1; NDC learning rule.

Results for this configuration are presented in Figure 52 and Table 47.
They show that the RMS error is 0.0865, which is larger than that of the NCD setup.
Although the maximum error is 5.84 knots for a 15 knots speed, the maximum error
percentage equal to 5.53 % is significantly smaller than that of the NCD scheme. The
77
maximum airspeed error SD of 1.71 knots occurs at 35 knots. Note that the airspeed
estimation is poor at low speeds and it is best at high speeds when the gross weight of the
helicopter is less than 2100 lb. The overall error SD is 0.9556 knots.

5 3.5658 0.2767 5.5336% 1.9728
15 19.5361 0.8258 5.5054% 5.8437
25 23.7345 1.1813 4.7250% 3.7400
35 34.4205 1.7182 4.9093% 2.8189
45 45.5856 0.6026 1.3391% 1.6418
78
Results are presented in Figure 53 and Table 48. One PE in the first
hidden layer was disabled. Results show that the RMS error is 0.0443, which is the best
result obtained for the analysis of this altitude. The maximum speed error is 2.58 knots
(at 5 knots) and the maximum error percentage is 5.59 %. The maximum airspeed error
SD equal to 0.5883 knots occurs at 15 knots speed. The overall error SD is 0.6401 knots.
5 2.9434 0.2797 5.5945% 2.5801
15 14.7437 0.6863 4.5756% 1.4916
25 24.2145 0.6529 2.6118% 2.0764
35 36.3517 0.6982 1.9949% 2.1910
45 46.8740 0.6187 1.3749% 1.8320
79
6. OH-6A Baseline Data Analysis at High Level
A baseline training data set was formed by combining both 12 ft. and 100 ft. set of
high level altitude data. First, the network was trained with this combined data set, and
then the network performance was measured using the baseline testing set. Finally, the
network performance was evaluated for IGE and OGE conditions, using single condition
test data sets separately. Results are given below.
a. 14-25-25-1 BPNN NCD
error is 0.0507, and the absolute maximum speed error is 3.15 knots (at 25 knots). The
airspeed error SD at 1 σ is less than 1 knot at all speeds. The overall error SD is 0.7139
knots.
Figure 54. Results for the OH-6A helicopter with baseline data (HL); network configuration
80
5 3.1503 0.4366 8.7318% 2.7569
15 15.9246 0.8179 5.4525% 2.1210
25 23.3005 0.7533 3.0132% 3.1582
35 35.7437 0.9455 2.7014% 2.5005
45 45.8256 0.6244 1.3876% 2.1869
Table 49. Results for the OH-6A helicopter with baseline data (HL); network configuration

Results are presented in Figure 55 and Table 50. They show that the RMS
error is 0.0772, and the absolute maximum speed error is 5.39 knots, which occurs at 15
knots speed. The maximum airspeed error SD at 1 σ is 1.61 knots (at 25 knots). The
prediction accuracy of low speeds is not as good as that of fast speeds. The overall error
SD is 1.0209 knots.

5 3.7873 0.4398 8.7951% 2.0195
15 19.0707 0.7415 4.9436% 5.3927
25 24.8423 1.6157 6.4628% 3.7711
35 35.460 0.9531 2.7230% 2.0827
45 45.6283 0.8166 1.8147% 1.8147
81
c. 14-25-25-1 BPNN Ext. DBD (Pruned)
The Ext. DBD pruned network results are presented because they were
better than those of the NCD pruned network. The RMS error for this configuration is
found to be 0.0549. The absolute maximum speed error is 3.43 knots (at 5 knots). The
maximum airspeed error SD at 1 σ is less than 1 knot at all speeds. The overall error SD
is 0.6505 knots.
5 2.3968 0.3774 7.5484% 3.4378
15 15.6340 0.5457 3.6383% 1.4552
25 23.7417 0.8713 3.4852% 3.3192
35 36.3556 0.6385 1.8243% 2.4002
45 45.9307 0.7895 1.7544% 2.1794
14-25-25-1; Ext. DBD (Pruned) learning rule.
82
14-25-25-1; Ext. DBD (Pruned) learning rule.
7. OH-6A HL Baseline Data IGE Analysis

a. 14-25-25-1 NCD IGE
The findings for this configuration are presented in Table 57 and Figure
57. Results show that the RMS error is 0.0446 and the absolute maximum speed error
is2.49 knots (at5 knots). The airspeed error SD at 1 σ is less than 1 knot at all speeds. The
overall error SD is 0.6296 knots.
5 3.2718 0.4172 8.8436% 2.4936
15 15.7052 0.7943 5.2950% 1.9305
25 23.5417 0.5634 2.2538% 2.2911
35 35.6807 0.7965 2.2758% 1.8238
45 45.5805 0.5918 1.3150% 1.4307
83
b. 14-25-25-1 Ext. DBD IGE
show that the RMS error is 0.0682, and the absolute maximum speed error is 4.55 knots,
which occurs at 5 knots speed. The maximum airspeed error SD at 1 σ is 1.47 knots (at
25 knots). The overall error SD is 0.9027 knots.

5 3.8334 0.4224 8.4479% 1.8070
15 18.5289 0.4925 3.2834% 4.5586
25 24.4931 1.7556 5.9025% 3.3707
35 35.5342 0.8230 2.3513% 1.7742
45 45.4751 0.7995 1.7767% 1.8916
84
configuration 14-25-25-1; Ext. DBD learning rule
c. 14-25-25-1 NCD (Pruned) IGE
show that the RMS error is 0.0732. The absolute maximum speed error equal to 2.09
knots occurs at 25 knots speed. The maximum airspeed error SD at 1 σ is less than 1 knot
at all speeds. The overall error SD is 0.6655 knots.

5 3.6158 0.3581 7.1625% 2.0768
15 15.4433 0.8653 5.7688% 1.9782
25 24.1375 0.6593 2.6372% 2.0969
35 35.1451 0.6882 1.9693% 1.3938
45 44.5488 0.6975 1.5501% 1.4001
85
configuration 14-25-25-1; NCD (Pruned) learning rule
8. OH-6A HL Baseline Data OGE Analysis
a. 14-25-25-1 NCD OGE
Results for this setup are shown in Figure 60 and Table 55. They show
that the RMS error is 0.0561. The absolute maximum error equal to 3.15 knots occurs at
25 knots. The maximum airspeed error SD at 1 σ is 1.07 knots, which is observed for 35
knots speed. The overall error SD is 0.7393 knots.
5 3.0288 0.4230 8.4591% 2.7569
15 16.1439 0.7843 5.2287% 2.1210
25 23.0594 0.8394 3.3576% 3.1582
35 35.8068 1.0779 3.0798% 2.5005
45 46.0707 0.5608 1.2461% 2.1869
86
b. 14-25-25-1 Ext. DBD OGE
error is 0.0853 and the absolute maximum speed error is 5.39 knots (at 15 knots). The
maximum airspeed error SD at 1 σ equal to 1.67 knots is observed for 25 knots speed.

5 3.7411 0.4534 9.0672% 2.0195
15 19.6126 0.5194 3.4625% 5.3927
25 25.1916 1.6786 6.7143% 3.7711
35 35.3858 1.070 3.0573% 2.0827
45 45.7815 0.8118 1.8040% 1.9277
87
c. 14-25-25-1 Ext. DBD (Pruned) OGE
show that the RMS error is 0.0565, which is slightly greater than that of the NCD setup.
The absolute maximum speed error equal to 4.71 knots occurs at 35 knots speed. The
maximum airspeed error SD at 1 σ is 1.25 knots (at 35 knots). The overall error SD is
0.8344 knots.

5 2.9652 0.3962 7.9235% 2.7769
15 16.1362 0.6021 4.0141% 2.1404
25 24.9846 1.1310 4.5241% 2.8046
35 37.5287 1.2565 3.5785% 4.7113
45 46.6852 0.7609 1.6906% 3.0221
configuration 14-25-25-1; Ext. DBD (Pruned) learning rule.
88
configuration 14-25-25-1; Ext. DBD (Pruned) learning rule.
C. NETWORK PERFORMANCES SUMMARY
Results for all experiments are summarized in Table 58 to Table 60. Analysis
results for the UH-60A helicopter and OH-6A helicopter at sea level and high level
analyses results are shown separately in Tables 58, 59 and 60 in terms of airspeed error
SD, absolute maximum error, RMS error and maximum percent error of each NN
architecture.
Table 58 displays results obtained for the UH-60A helicopter model. For the UH-
60A helicopter at 85 ft (out of ground effect condition) a 2-layer BPNN network with the
NCD learning rule with pruning yielded the best results. This configuration shows a
predicted airspeed with 0.7 knots error SD, while the maximum error is 3.34 knots, the
maximum error percentage error is 4.22 % and the RMS error is 0.0355. When the
helicopter is at 20 ft. (in ground effect condition), the network with NCD rule produced
an estimate with 0.84 knots error SD, a RMS error equal to 0.0374, a 3 knots maximum
error and a maximum error percentage within 5.2 %. These results are significantly better
than those obtained by Haas and McCool, with real flight data. Their study showed that
89
using real flight data as input to the NN, UH-60A airspeed can be predicted with an
accuracy of ± 5 knots when the aircraft is in ground effect. However they performed that
analysis with a reference airspeed uncertainty of ± 2 knots. Note that prediction accuracy
improved considerably with the simulator data, which has no wind effect nor any other
uncertainties.
UH-60A Helicopter RMS Error of Abs. Max. Error Max. Percent
BPNN Models Error SD at 1 σ (knots) Error (%)
OGE NCD 0.0658 0.7506 3.85 6.4
(at 35 knots) (at 5 knots)
OGE Ext. DBD 0.0593 0.7017 3.86 8.49
OGE NCD Prune 0.0355 0.7167 3.34 4.22
Simplified Data OGE NCD 0.0656 1.0032 3.89 5.5
Simplified Data OGE Ext. DBD 0.0565 1.1692 5.4 11.75
Simplified Data NCD Prune 0.0648 1.1942 5.4 7.95
(at 25 knots) (at 5 knots
One Layer NCD 0.0407 0.8394 3.37 4.5
One Layer Ext. DBD 0.0657 0.7002 3.79 8.34
IGE NCD 0.0374 0.8469 3.0 5.2
IGE Ext. DBD 0.0637 0.8308 4.8 6.0
IGE NCD Prune 0.0724 0.8864 4.86 5.22
(15 knots) (at 25 knots)
Baseline Data NCD 0.0762 1.5664 6.24 8.0
Baseline Data Ext. DBD 0.0501 0.7320 4.5 10.7
Baseline Data NCD Prune 0.1213 2.0830 8.64 11.5
Baseline Data IGE NCD 0.0759 1.5516 6.24 10.12
Baseline Data IGE Ext. DBD 0.0499 0.6505 4.5 8.7
Baseline Data IGE NCD Prune 0.0901 1.7173 7.91 9.12
Baseline Data OGE NCD 0.0766 1.5516 5.86 9.3
Baseline Data OGE Ext. DBD 0.0482 0.6850 3.8 9.5
Baseline Data OGE NCD Prune 0.1087 1.8642 7.14 9.42
Table 58. Overall Results for the UH-60A Helicopter.
90
Results showed that the network with the Ext. DBD rule produced the best results
for the baseline data analysis. This network predicted airspeed with a 0.73 knots error SD
and a 0.05 RMS error. The maximum error was estimated at 4.5 knots at 35 knots speed,
while the maximum error percentage was 10.7 % observed for a 5 knots speed. Using the
baseline data for the IGE condition airspeed prediction was good with the Ext. DBD
network since it produced an error SD of 0.65 knots, a RMS error of 0.05, maximum
error of 4.5 knots at the speed of 35 knots and a maximum percent error of 8.7%. The
baseline data OGE condition results were very close to those of the baseline data IGE
condition results. Using this data, the network with Ext. DBD rule predicted the airspeed
with a 0.048 RMS error, a 0.68 knots error SD, a 3.8 knots maximum error at the speed
of 35 knots and a 9.5% maximum percent error. The single condition data prediction was
slightly better than that observed with the baseline data.
We note that the maximum error and maximum error SD occurred mostly at 35
knots speed for all networks. A potential explanation might be that the hover to forward
flight translational lift was set to 30 knots for all the FLIGHTLAB simulator models. The
network was trained with the test data set and tested with the training data set in order to
explore this idea. Results obtained with the switched data showed that the maximum error
and maximum error SD occurs when the helicopter is moving at 30 knots. While not
conclusive, this set indicates the difficulty is associated with the simulated helicopter
performance near these speeds.
OH-6A analyses were conducted at sea level and at high altitude (at 6000 feet
pressure altitude) using a similar methodology for the UH-60A model. Table 59 shows
the results of the OH-6A analyses at sea level, and Table 60 shows the results for high
altitude analyses.
At sea level, the OGE condition using the single condition data NCD network
with the pruning function predicted an airspeed with a 0.76 knots error SD and a 0.067
RMS error. The maximum error was about 4.69 knots for a 35 knots speed while the
maximum error was observed at 9.9 %. At sea level, the IGE condition NCD network
produced a 0.84 knots error SD. The RMS error was about 0.059, the maximum error was
5.1 knots, and the maximum percent error was 9.7%. Results showed that using baseline
91
data reduced the network performance, as the error SD was 1.169 knots, the RMS error
was 0.0611, the maximum error was 5 knots and the maximum percent error was 6.8 %.
The NCD network with the pruning function yielded the best results for this condition.
For baseline data, the IGE condition NCD pruned network predicted the airspeed
with a 1.15 knots error SD, while the RMS error was about 0.059 and the maximum error
was 4.5 for a 35 knots speed. The maximum error percentage was found to be 8.19% for
this condition. For the baseline data OGE condition, airspeed was estimated with a 1.20
knots error SD and a 0.065 RMS error with the pruned NCD network. A 5.1 knots
maximum error occurred for a 35 knots speed and the maximum error percentage was
6.9%.
OH-6A Helicopter RMS Error of Abs. Max. Max. Percent

Error (%)
BPNN models (at sea level) Error SD at 1σ Error (knots)
OGE NCD 0.06 0.79 5.2(at 35 knots) 9.5 (at 5knots)
OGE Ext. DBD 0.0735 1.32 5 (at 35 knots) 9.4 (at 5knots)
OGE NCD Prune 0.067 0.759 4.69(at 35 knots) 9.9 (at 5knots)
IGE NCD 0.059 0.8465 5.1(at 35 knots) 9.7 (at 5knots)
IGE Ext. DBD 0.062 0.9678 5.9(at 35 knots) 12.2 (at 5knots)
IGE NCD Prune 0.0674 0.88 4.41(at 35 knots) 12.1(at 5knots)
Baseline Data NCD 0.062 1.192 4.82(at 35 knots) 8.69(at 5knots)
Baseline Data Ext. DBD 0.0687 1.407 7.44(at 35 knots) 14(at 5knots)
Baseline Data NCD prune 0.0611 1.16 5(at 35 knots) 6.8(at 5knots)
Baseline Data IGE NCD 0.0593 1.15 4.49(at 35 knots) 9.29(at 5knots)
Baseline Data IGE Ext. DBD 0.0686 1.32 7.44(at 35 knots) 11.9(at 5knots)
Baseline Data IGE NCD Prune 0.0592 1.15 4.51(at 35 knots) 8.19(at 5knots)
Baseline Data OGE NCD 0.0671 1.20 4.82(at 35 knots) 7.2(at 5knots)
Baseline Data OGE Ext. DBD 0.0687 1.3 6.6(at 35 knots) 9.4(at 5knots)
Baseline Data OGE NCD Prune 0.0650 1.2 5.1(at 35 knots) 6.9(at 5knots)
Table 59. Overall Results for OH-6A Helicopter at Sea Level.
The pruned NCD network predicted the airspeed with a 0.64 knots error SD for
the OH-6A high level OGE condition. The RMS error was found to be equal to 0.0475,
while the maximum error was 2.8 knots at 5 knots speed and the maximum error was
about 7%. Finally, the airspeed was predicted with a 0.64 knots error SD, a 0.0433 RMS
92
error, a 2.58 knots maximum error and a 5.6% maximum error percentage for the IGE
condition using the single condition data with the pruned NCD network.
The following results were obtained for the baseline data. The Ext. DBD network
performed best with a 0.65 knots error SD, a 0.055 RMS error, a 3.4 knots maximum
error for a 5 knots speed and a 7.5 % maximum error percentage. The NCD pruned
network yielded the best results for the baseline data IGE condition. For this setup, the
RMS error was 0.0352 while the error SD was 0.66 and the maximum error was 2 knots
at the speed of 25 knots. The maximum percent error was 7.1% for this condition. The
baseline data OGE condition results were close to the IGE condition results. The NCD
network estimated the airspeed with an error SD of 0.74 knots, a RMS error of 0.056, a
maximum error of 3.15 knots and a maximum percent error of 8.4%.
OH-6A Helicopter RMS Error of Abs. Max. Max. Percent

Error (%)
BPNN models (at high altitude) Error SD at 1σ Error (knots)
OGE NCD 0.0485 0.6637 2.47 (at 5knots) 9.47(at 5knots)
OGE Ext. DBD 0.0958 1.022 6.36(at 15knots) 5.67(at 5knots)
OGE NCD Prune 0.0475 0.64 2.85(at 5knots) 7.1(at 5knots)
IGE NCD 0.0516 1.20 2.59(at 5knots) 9.12(at 5knots)
IGE Ext. DBD 0.0865 0.95 5.8(at 15knots) 5.53(at 5knots)
IGE NCD Prune 0.0443 0.64 2.58(at 5knots) 5.59(at 5knots)
Baseline Data NCD 0.05 0.71 3.15(at 25knots) 8.73(at 5knots)
Baseline Data Ext. DBD 0.077 1.02 5.4(at 15knots) 8.8(at 5knots)
Baseline Data Ext. DBD prune 0.055 0.65 3.43(at 5knots) 7.5(at 5knots)
Baseline Data IGE NCD 0.0446 0.63 2.49(at 5knots) 8.84(at 5knots)
Baseline Data IGE Ext. DBD 0.0687 0.9 4.55(at 5knots) 8.44(at 5knots)
Baseline Data IGE NCD Prune 0.0352 0.66 2.09(at 25knots) 7.1(at 5knots)
Baseline Data OGE NCD 0.056 0.74 3.15(at 25knots) 8.45(at 5knots)
Baseline Data OGE Ext. DBD 0.0853 1.02 5.39(at 15knots) 9(at 5knots)
Baseline Data OGE Ext. DBD Prune 0.0565 0.83 4.71(at 35knots) 7.9(at 5knots)
Table 60. Overall Results of OH-6A Helicopter at High Altitude.
In summary, these results show that the NN approach to predict airspeed using
simulation data is quite promising. The BPNN with two hidden layers and 25 PEs in each
layer performs the best among all studied architectures. We note that different learning
93
rules yielded different results and that enabling the pruning facility improved the network
performance in most cases.
94
V. CONCLUSIONS AND RECOMMENDATIONS
A. SUMMARY
Today, military helicopters perform a wide variety of tasks in conditions ranging
from hot and dry to cold and wet, windy and low visibility weather. Accurate low speed
velocity sensing devices are essential because aircraft velocity and position information
are what pilots need to perform safely in these regimes. However, conventional speed
measuring systems do not work accurately when the aircraft speed is below 40 knots.
NN-based airspeed prediction studies developed by McCool, Haas and others

showed that NN is a good approach to solve this problem. The objective of this thesis is
to build a NN model in order to predict OH-6A helicopter airspeed in the low speed
environment using a flight simulator to obtain the parameters required for the NN. In our
study the NN model is developed with Neural Works Professional Plus/II software.
First a NN model of the UH-60A helicopter was built and implemented, and
several NN configurations analyzed. The reason for building a NN model for the UH-
60A helicopter was to lay out a background to make a comparison of NN predictions
using simulator data and NN predictions using real flight data. The results showed using
simulator data potentially improves the accuracy of prediction significantly.
Three different methods were investigated to select the NN training data. The first
one is a single condition data set in which the data belongs to one altitude only. The
second one is called the baseline data set, and is formed by combining the data of two
single condition data sets of different altitudes. The third set is obtained by applying
principal component analyses to decrease the input space dimension and it is called a
simplified data set. Results showed that the network trained by using a single condition
data set proved to be the most successful and performance degraded only slightly with the
baseline data. Moreover, the BPNN network produced more successful predictions than
the RBFN implementation.
Among all BPNN architecture types considered, a two-hidden layer BPNN with
an enabled pruning facility for the NCD learning rule showed the best performance. At
sea level pressure altitude, the UH-60A low airspeed was predicted with one-sigma
95
accuracy of ± 0.71 knots when the aircraft was out of ground effect. The accuracy of
prediction was ± 0.88 knots when the aircraft was in ground effect.
The OH-6A low speed was predicted using a similar methodology based on the
results obtained from the UH-60A model. The OH-6A analysis was performed at high
pressure altitude as well as at seal level altitude. Results showed that at sea level, the OH-
6A airspeed could be predicted with one-sigma accuracy of ± 0.75 knots when the
aircraft is out of ground effect. The best performance for all high altitude analyses
obtained from the pruned NCD network was with single condition data. For IGE
conditions, the prediction accuracy was about ± 0.88 knots. High altitude analysis was
performed at 6000 feet. The results showed that at high altitude, the OH-6A airspeed
could be predicted with an accuracy of ± 0.64 knots when the aircraft is out of ground
effect. For IGE conditions, the prediction accuracy was about ± 0.64 knots.
This study showed that a NN based approach to determine OH-6A helicopter

airspeeds using a flight simulator is quite promising. The approach considered presents a
mechanically simple alternative to current low airspeed measurement systems, and as a
result contributes to increase the flight safety and combat effectiveness.
B. RECOMMENDATIONS FOR FURTHER RESEARCH
Several avenues are available to extend the work presented in this study. First,
results obtained using real flight data are needed. These results should be compared to
the results obtained here to evaluate the performance of our approach. Second, more
sophisticated maneuvers need to be simulated and analyzed. The NN model might be
further improved by using different network models, model parameters, etc. Another type
of network to be investigated may be NN-based sideslip angle estimator, which would
provide accurate wind direction for pilot-aid systems. Finally, the effects of measurement
noise, sensor errors, and sensor failure should be investigated.
96
LIST OF REFERENCES
1. Neural Network Design, M. Hagan, H., Demuth, M. Beale, PWS Publishing

Company, 1996
2. Neural Network Toolbox For Use with Matlab, The Math Works, Inc., 2001
3. “A NN Based Approach To Helicopter Low Airspeed And Sideslip

Angle Estimation,” K.M. McCool, D.J. Haas, Proceedings of the AIAA Flight Simulation
Technologies Conference, 1996
4. “A NN Based Approach To Helicopter Low Airspeed And Sideslip

Angle Estimation,” K.M. McCool, D.J. Haas, Proceedings of the AIAA Flight
Simulation Technologies Conference, 1996
5. Using Neural Works, Neural Ware, Inc., 1993
6. Neural Computing, Neural Ware, Inc., 1993
7. FLIGHTLAB Manual Vol. II, Scope, Gscope, PilotStation, CSGE, Model Editor,
Xanalysis, Advance Rotorcraft Technology, June 2001
8. “Flight Investigations of a Helicopter Low Airspeed Estimation System Based on

Measurement of Control Parameters,” A. J. Faulkner, F. Buchner, Sixth European
Rotorcraft and Powered Lift Aircraft Forum, Sep. 1980.
9. Neural Networks: Tricks of the Trade, K-R. Muller (Editor), G. Orr (Editor), 1st
Edition, May 1999
10. Artificial Neural Networks, D. Klerfors, St.Louis Uni., 1998

(http://www.hj.se/~de96klda/NeuralNetworks.htm), last accessed December 2001.
97
11. Neural Network FAQ, Warren S. Sarle, 2001
(ftp://ftp.sas.com/pub/neural/FAQ2.html), last accessed January 2002.
12. “The Measurement of Helicopter Air Data Using a Swivelling Pitot-Static

Pressure Probe,” J. Carter, Air Data Symposium, Monterey, CA, June 1976
13. “Omni-directional Low Range Airspeed, Display and System Requirements for
Helicopters and V/STOL Aircraft,” D. L. Green, Air Data Symposium, Monterey, CA,
June 1976
14. FLIGHTLAB Manual Vol. I, Advance Rotorcraft Technology, April 2001
15. “An Electro-Optic Airspeed Sensor,” G. A. Dubro, D. G. Kim, M. J. Rudd, Air

Data Symposium, Monterey, CA, June 1976
16. “Practical NN Applications in the Mining Industry,” L. Miller-Tait, R. Pakalnis,

Proceedings of the Second International Conference, Vol.2, 1999
17. Helicopter Flight Dynamics: The Theory and application of Flying Qualities and
Simulation Modeling, G. D. Padfield, AAIA Education Series, 1999
18. Linear Algebra and Its Applications, Third Edition, G. Strang, Saunders HBJ
Publishing, 1988
19. Modeling the OH-6A Using FLIGHTLAB and Helicopter Simulator

Considerations, Master’s Thesis, Naval Postgraduate School, G. A. Ouellette, to be
completed March 2002
20. “Predicting Airspeed and Sideslip Angle using an Artificial Neural Network”, D.
Goff, S. Thomas, P. Jones, C. Massey, AHS 55th Annual Forum, Montreal, Canada, 1999
98
APPENDIX A. NEURALWORKS PROFESSIONAL PLUS/II
PROGRAM SETUP
Neural Works Professional II/PLUS is a powerful and flexible development tool.

It has over 20 different algorithms including common network types such as Back
Propagation, Kohonen and Radical Basis Functions. The product includes a variety of
diagnostic tools as well as options for quick and easy building networks.
Before creating the network, data files must be put under the directory of
Neuralware Professional directory. The test and training input data files must be given
different names but both of them must have extension of “.nna”. The data files used in
this thesis may be obtained from Professor R. W. Duren, Naval Post Graduate School
Department of Aerospace and Aeronautical Engineering, by request.
Neural network development process begins with collecting and preprocessing
input data. In this study data is collected from FLIGHTLAB simulation and prepared by
using MATLAB. Preparation involves encoding data to a format that NN can deal with.
For this problem data set was converted to a matrix with a dimension of 1114x14, where
columns represent the inputs and rows represent the samples of each inputs. MATLAB is
also used to separate the data into to sets, training set and test set, as well as to display the
results. MATLAB codes for preprocessing and post processing the data are included in
Appendix B.
After starting NeuralWorks program, Back-propagation command, under the
InstaNet menu, must be selected to create a BP network. Selecting this command opens
up BP dialog box, which allows user to build the main frame of the network by entering
number of layers and number of PEs per layer. In addition to these, learning coefficient
of each layer, momentum term, learning rule, transfer function, test data and training data
of the network can be selected. Also by selecting “minmax” radio button, inputs can be
mapped from the data file to a desired range, such as –1 to 1.
Activating this box opens up another window, which asks user how to display the
performance of the network, such as RMS error, classification rate, etc. After these
selections network is created which is presented in a window like in Figure 5. Created
99
network can be saved either in binary format or in ASCII format by selecting the Save
command under the File menu.
Selecting the Learn command under the Run menu starts training. At this point,
number of learning iterations can be entered by the user. As each training example is
presented to the network, the network produces an output, which is used to evaluate the
training performance of the network. Another way to train the network is to use Savebest
command under the Run menu. This commands opens up Run/Check dialog box, which
makes pruning facility accessible. Based on a decision criteria specified by the user,
pruning facility disables connections in a network as the network is training.
Network is tested using the test data sets and by selecting the Test command
under the Run menu. The desired outputs, along with the actual network results are
written to the results file which has “.nnr” extension.
Based on the above explanations and after preparing the data, OH-6A helicopter
14-25-25-1 BPNN NCD (Prune) network model created using the following steps:
Start Neural Works on the computer
Select Back-Propagation command from the InstaNet menu. This command pops
up the following window.
Figure 63. Back-propagation network setup window
100
1. Select “trainfile_trn” from the training input scroll window.
2. Select “testfile_tst” from the testing input scroll window.
3. In the number of PEs section, enter the following numbers:
-Input: 14
-Hid1: 25
-Hid2: 25
-Output: 1
4. Enter 0.4 for momentum and 0.5 for LCoef ratio.
5. Select Norm-Cum-Delta for the learning rule.
6. Select TanH for the transfer function.
7. Check MinMax table box.
8. Click the OK button.
9. After clicking OK the following window opens automatically. Select RMS
Error.
Figure 64. Instrument /Create menu

10. Click OK.
11. Start training by selecting the SaveBest command under the Run menu.
SaveBest command setup is shown in Figure #. Training can also be started by
selecting Learn command under the same menu. SaveBest command allows
user to use Pruning facility.
12. Type the name of the file that results are to be written.
13. Enter 50000 in the For field.
14. Enter 1000 for the Test Interval.
101
15. Check the pruning radio button.
16. Enter 0.975 for the Tolerance field.
17. Select Classification Rate for the Objective Function list.
18. Click OK. to start training.
Figure 65. SaveBest command window

19. After training is completed select Test in the Run menu.
Figure 66. Test command window

20. Select One Pass/All.
21. Click OK.
After test process has been completed the performance of the network is displayed
on the main window in terms of selected options, such as RMS error, network weights,
etc. and the results are stored in the test file with .nnr extension. In this work, the result
files were saved as text files in order to be exported into the MATLAB.
102
APPENDIX B. MATLAB® M-FILES
MATLAB m-files were developed to prepare the simulation outputs for NN and
to process the NN results.
a. FILE Name: Gregcode

The first code included in this appendix is the FLIGHLAB scope language, an
interpretive language that uses the industry standard MATLAB syntax. This routine
loads the specified model and runs the simulator. The outputs of this routine are used as
inputs to the NN model.
/***********************************************************************
This code was developed by LT. Gregory OUELLETTE, USN. for the Naval
Postgraduate School in partial fulfillment of the requirements for a Masters Degree in
Aeronautical Engineering.
September 2001.
/***********************************************************************
exec("$FL_DIR/flme/models/articulated/arti-rgd-3iv-qs-simeng.def",1)
world_model_airframe_cpg_testcond_poszic = -85;
exec("xatestcond.exc",1)
exec("xamodeltrim.exc",1)
goto world
group test
gw = [16000:1000:24000]';
hdg = [30:30:360]';
vel = [0:5:30]';
column = 0;
utrim = @trimvariable;
statesave = savestates(world_topsolve);
for ngw = 1:prod(size(gw))

for nhdg = 1: prod(size(hdg))
for nvel = 1: prod(size(vel))
world_model_airframe_cpg_testcond_veq = vel(nvel);
world_model_airframe_cpg_testcond_gamh = hdg(nhdg);
103
world_model_data_vweight = gw(ngw);
outputs(column+nvel,1) = world_model_control_data_xatrm;
outputs(column+nvel,2) = world_model_control_data_xbtrm;
outputs(column+nvel,3) = world_model_control_data_xctrm;
outputs(column+nvel,4) = world_model_control_data_xptrm;
outputs(column+nvel,5) = world_model_airframe_cpg_xaout_p;
outputs(column+nvel,6) = world_model_airframe_cpg_xaout_q;
outputs(column+nvel,7) = world_model_airframe_cpg_xaout_r;
outputs(column+nvel,8) = world_model_airframe_cpg_xaout_phi;
outputs(column+nvel,9) = world_model_airframe_cpg_xaout_psi;
outputs(column+nvel,10) = world_model_airframe_cpg_xaout_radralt;
outputs(column+nvel,11) = world_model_airframe_cpg_xaout_vclimb;
outputs(column+nvel,12) = world_model_rotor1_rotor_cpg_xaout_omega;
outputs(column+nvel,13) = world_model_propulsion_cpg_xaout_etorq;
outputs(column+nvel,14) = world_model_data_vweight;
outputs(column+nvel,15) = world_model_airframe_cpg_xaout_tas;
outputs(column+nvel,16) = world_model_airframe_cpg_testcond_gamh;
outputs(column+nvel,17) = world_model_airframe_cpg_testcond_veq;
outputs(column+nvel,18) = 0;
end
column = column + prod(size(vel));

savestates(world_topsolve,statesave);
@trimvariable = utrim;
end
end
trainset = outputs';
save("velhead.sav",trainset)
exec("$FL_DIR/flme/models/articulated/arti-rgd-3iv-qs-simeng.def",1)
goto world
group test
104
gw = [16000:1000:24000]';
hdg = [-60:30:60]';
vel = [35:5:40]';
column = 0;
utrim = @trimvariable;
statesave = savestates(world_topsolve);
for ngw = 1:prod(size(gw))

for nhdg = 1: prod(size(hdg))
for nvel = 1: prod(size(vel))
world_model_airframe_cpg_testcond_veq = vel(nvel);
world_model_airframe_cpg_testcond_gamh = hdg(nhdg);
world_model_data_vweight = gw(ngw);
outputs(column+nvel,1) = world_model_control_data_xatrm;
outputs(column+nvel,2) = world_model_control_data_xbtrm;
outputs(column+nvel,3) = world_model_control_data_xctrm;
outputs(column+nvel,4) = world_model_control_data_xptrm;
outputs(column+nvel,5) = world_model_airframe_cpg_xaout_p;
outputs(column+nvel,6) = world_model_airframe_cpg_xaout_q;
outputs(column+nvel,7) = world_model_airframe_cpg_xaout_r;
outputs(column+nvel,8) = world_model_airframe_cpg_xaout_phi;
outputs(column+nvel,9) = world_model_airframe_cpg_xaout_psi;
outputs(column+nvel,10) = world_model_airframe_cpg_xaout_radralt;
outputs(column+nvel,11) = world_model_airframe_cpg_xaout_vclimb;
outputs(column+nvel,12) = world_model_rotor1_rotor_cpg_xaout_omega;
outputs(column+nvel,13) = world_model_propulsion_cpg_xaout_etorq;
outputs(column+nvel,14) = world_model_data_vweight;
outputs(column+nvel,15) = world_model_airframe_cpg_xaout_tas;
outputs(column+nvel,16) = world_model_airframe_cpg_testcond_gamh;
outputs(column+nvel,17) = world_model_airframe_cpg_testcond_veq;
outputs(column+nvel,18) = 0;
end
column = column + prod(size(vel));

savestates(world_topsolve,statesave);
@trimvariable = utrim;
end
end
105
trainset = outputs';
save("velheadfast.sav",trainset)
b. FILE Name: Ozcan.m

%*********************************************************************
% This routine takes simulation output files as input and makes them applicable for NN
% model by forming a matrix, where columns represent the input parameters to NN and
% rows represent the samples.
% mat1 is the output matrix of the routine where all the data are stored and to be used as
% input data to the NN.
% The last part of the routine may be used to obtain the simplified data using
% eigenvalues and eigenvectors. Dataeig is the resultant matrix for simplified data.
% September 2001
%*********************************************************************
clear all;
format short e
load velhead.txt; % loading ascii file (simulation output file)
load velheadfast.txt; %loading ascii file
m=1;
for i = 1:length(velhead)/6
slat_stick(i) = velhead(m,1);
slong_stick(i) = velhead(m,2);
scoll_pos(i) = velhead(m,3);
sped_pos(i) = velhead(m+1,1);
sroll_rate(i) = velhead(m+1,2);
spitch_rate(i) = velhead(m+1,3);
syaw_rate(i) = velhead(m+2,1);
spitch_att(i) = velhead(m+2,2);
sroll_att(i) = velhead(m+2,3);
salt(i) = velhead(m+3,1);
sclimb_rate(i) = velhead(m+3,2);
smrb_rpm(i) = velhead(m+3,3);
seng_torque(i) = velhead(m+4,1);
sgw(i) = velhead(m+4,2);
stas_trim(i) = velhead(m+4,3)*(360/608); % ft/sec is converted into knots
sheading(i) = velhead(m+5,1);
stas_trgt(i) = velhead(m+5,2);
106
szero(i) = velhead(m+5,3);
m=m+6;
end
m=1;
for i = 1:length(velheadfast)/6
flat_stick(i) = velheadfast(m,1);
flong_stick(i) = velheadfast(m,2);
fcoll_pos(i) = velheadfast(m,3);
fped_pos(i) = velheadfast(m+1,1);
froll_rate(i) = velheadfast(m+1,2);
fpitch_rate(i) = velheadfast(m+1,3);
fyaw_rate(i) = velheadfast(m+2,1);
fpitch_att(i) = velheadfast(m+2,2);
froll_att(i) = velheadfast(m+2,3);
falt(i) = velheadfast(m+3,1);
fclimb_rate(i) = velheadfast(m+3,2);
fmrb_rpm(i) = velheadfast(m+3,3);
feng_torque(i) = velheadfast(m+4,1);
fgw(i) = velheadfast(m+4,2);
ftas_trim(i) = velheadfast(m+4,3)*(360/608); % ft/sec is converted into knots
fheading(i) = velheadfast(m+5,1);
ftas_trgt(i) = velheadfast(m+5,2);
fzero(i) = velheadfast(m+5,3);
m=m+6;
end
lat_stick=[slat_stick,flat_stick];
long_stick=[slong_stick,flong_stick];
coll_pos=[scoll_pos,fcoll_pos];
ped_pos=[sped_pos,fped_pos];
roll_rate=[sroll_rate,froll_rate];
pitch_rate=[spitch_rate,fpitch_rate];
yaw_rate=[syaw_rate,fyaw_rate];
pitch_att=[spitch_att,fpitch_att];
roll_att=[sroll_att,froll_att];
alt=[salt,falt];
climb_rate=[sclimb_rate,fclimb_rate];
mrb_rpm=[smrb_rpm,fmrb_rpm];
eng_torque=[seng_torque,feng_torque];
gw=[sgw,fgw];
tas_trim=[stas_trim,ftas_trim];
heading=[sheading,fheading];
tas_trgt=[stas_trgt,ftas_trgt];
%***for UH-60A helicopter add parameter “eng_torque” to the matrix below****
107
mat1=[lat_stick;long_stick;coll_pos;ped_pos;roll_rate;pitch_rate;yaw_rate;pitch_
att;roll_att;alt;climb_rate;mrb_rpm;gw;heading;tas_trgt]'
% To simplify the data by eigenvalues and eigenvectors use the following part
% mat2=mat1'*mat1;
% [v,d]=eig(mat2);
% After examining the eigenvalues create submatrix u
% u=[mat1(:,9),mat1(:,10),mat1(:,11),mat1(:,12),mat1(:,13),mat1(:,14)];
% dataeig=mat1'*u;
c. FILE Name: Train.m

%*********************************************************************
% This routine takes the output matrix of Ozcan.m code, mat1 or dataeig, by loading the
% ascii file and filters it so that train data set can be obtained from the whole data.
% September 2001
%**********************************************************************
clear all;
format short e
load oh6sl_100ft.txt; % loading ascii file
[rmax,cmax]=size(oh6sl_100ft);
dmax=924 ; % Enter the # of rows associated with slow velocity data
s=1;
x=0;
y=8;
for d=1:dmax/14
for i=1:2:7
trndt(s,:)=oh6sl_100ft(i+x,:);
s=s+1;
end
for j=0:2:6
trndt(s,:)=oh6sl_100ft(j+y,:);
s=s+1;
end
x=x+14;
y=y+14;
end
for d1=(dmax+2):2:rmax
trndt(s,:)=oh6sl_100ft(d1,:);
s=s+1;
end
trndt % training data set
108
d. FILE Name: Test.m
%**********************************************************************
% This routine takes the output matrix of Ozcan.m code, mat1 or dataeig, by loading the
% ascii file and filters it so that test data set can be obtained from the whole data.
% September 2001
%**********************************************************************
clear all;
format short e
load oh6sl_100ft.txt; % loading ascii file
[rmax,cmax]=size(oh6sl_100ft);
dmax=924 ; % Enter the #of rows associated with slow velocity data
s1=1;
x1=0;
y1=9;
for d2=1:dmax/14
for l=2:2:6
tstdt(s1,:)=oh6sl_100ft(l+x1,:);
s1=s1+1;
end
for j1=0:2:4
tstdt(s1,:)=oh6sl_100ft(j1+y1,:);
s1=s1+1;
end
x1=x1+14;
y1=y1+14;
end
for d3=(dmax+1):2:rmax
tstdt(s1,:)=oh6sl_100ft(d3,:);
s1=s1+1;
end
tstdt % test data set
109
e. FILE Name: Bersan.m
%**********************************************************************
% This program used to process the outputs of the NN. In this routine the output file of
% the NN is taken as input and vectors of NN predicted speeds are created related to each
% gross weight and sideslip angle of the helicopter. This file also produces the figures
% and the evaluation of the NN results.
% October 2001
% Note: For baseline data and single data use the specified sections of the program. Also
% depending on the helicopter type activate and deactivate the stated lines of the code.
%**********************************************************************
clear all;
load oh6_sl_ncdprune.txt; % loading ascii file
a=oh6_sl_ncdprune(:,2); % NN outputs(predicted speeds)
ilk=oh6_sl_ncdprune(:,1); % target values(actual speeds)
%set initial variables

g=1;
j=1;
k=1;
l=1;
m=1;
n=length(a);
aci=[30:30:360]; % angles for slow speed
aci1=[-60:30:60]; % angles for fast speed
ai=length(aci);
% x=1.6:0.1:2.4; % gross weight range of uh-60a helicopter(10^4 lb)

x=1.500:.100:2.550; % gross weight range of oh-6a helicopter(10^3 lb)
%(for UH-60A:16000:1000:24000 - for OH-6A: 1500:100:2550 )
gw=length(x);
na=(ai*gw*6); % slow velocity vector dimension (for baseline data
% multiply by 6, for single data multiply by 3)
% Classify the speeds and get each speed vectors

for i=1:3:na
a1(j,1)=a(i,1); %vector of 5 kts
fark5(j)= abs(a1(j,1)-ilk(i,1));
j=j+1;
end
110
for i=2:3:na
a2(k,1)=a(i,1); %vector 15 kts
fark15(k)= abs(a2(k,1)-ilk(i,1));
k=k+1;
end
for i=3:3:na
a3(l,1)=a(i,1); %vector 25 kts
fark25(l)= abs(a3(l,1)-ilk(i,1));
l=l+1;
end
for t=(na+1):2:n
a4(m,1)=a(t,1); % vector 35 kts
fark35(m)= abs(a4(m,1)-ilk(t,1));
m=m+1;
end
for t=(na+2):2:n
a5(g,1)=a(t,1); %vector 45 kts
fark45(g)= abs(a5(g,1)-ilk(t,1));
g=g+1;
end
% classify speeds according to gross weights
y=length(a1);
b=y/gw;
% for speed = 5 kts

c=1;
for i=1:gw
if i>=2
z=1;
for j=(((i-1)*b)+1):i*b
a11(z,i)=a1(j,1);
z=z+1;
end
else
for v=1:b
a11(c,1)=a1(v,1);
c=c+1;
end
end
end
figure(1)
111
subplot(2,2,1);plot(x,a11,'ro')
%for speed = 15 kts

c=1;
for i=1:gw
if i>=2
z=1;
for j=(((i-1)*b)+1):i*b
a22(z,i)=a2(j,1);
z=z+1;
end
else
for v=1:b
a22(c,1)=a2(v,1);
c=c+1;
end
end
end
hold on
subplot(2,2,1);plot(x,a22,'g<')
%for speed = 25 kts

c=1;
for i=1:gw
if i>=2
z=1;
for j=(((i-1)*b)+1):i*b
a33(z,i)=a3(j,1);
z=z+1;
end
else
for v=1:b
a33(c,1)=a3(v,1);
c=c+1;
end
end
end
hold on
subplot(2,2,1);plot(x,a33,'bo')
% for airspeed = 35 kts

y1=length(a4);
b1=y1/gw;
c=1;
for i=1:gw
112
if i>=2
z=1;
for j=(((i-1)*b1)+1):i*b1
a44(z,i)=a4(j,1);
z=z+1;
end
else
for v=1:b1
a44(c,1)=a4(v,1);
c=c+1;
end
end
end
hold on
subplot(2,2,1);plot(x,a44,'m>')
% for airspeed 45 kts

c=1;
for i=1:gw
if i>=2
z=1;
for j=(((i-1)*b1)+1):i*b1
a55(z,i)=a5(j,1);
z=z+1;
end
else
for v=1:b1
a55(c,1)=a5(v,1);
c=c+1;
end
end
end
hold on
subplot(2,2,1);plot(x,a55,'ko')
axis([1.4 2.6 0 50])
title('Predicted Speed vs Grossweight','FontSize',10)
xlabel('Gross Weight (lb*10000)','FontSize',7) %for UH-60A
% xlabel('Gross Weight (lb*1000)','FontSize',7) % for OH-6A
ylabel('Predicted Speed (Kt)','FontSize',7)
set(gca,'xtick',[1.6 1.8 2.0 2.2 2.4],'ytick',[0 5 15 25 35 45]) %for UH-60A
% set(gca,'xtick',[1.5 1.8 2.0 2.2 2.5],'ytick',[0 5 15 25 35 45]) % for OH-6A
hold off
% Plotting predicted speed vs actual speed
subplot(2,2,2);plot(ilk,a,'o')
113
axis([0 55 0 55])
title('Predicted Speed vs Actual Speed','FontSize',10)
xlabel('Actual Speed (Kt)','FontSize',7)
set(gca,'xtick',[0 5 15 25 35 45],'ytick',[0 5 15 25 35 45])
hold off
% plotting airspeed vs. angle (for single data only)
% subplot(2,2,3);plot(aci,a11,'ro')
% hold on
% subplot(2,2,3);plot(aci,a22,'g<')
% hold on
% subplot(2,2,3);plot(aci,a33,'bo')
% axis([0 400 0 35])
% title('Slow Speed vs Angle','FontSize',10)
% xlabel('Side Slip Angle (Deg)','FontSize',7)
% ylabel('Predicted Speed (Kt)','FontSize',7)
% set(gca,'xtick',[0 60 120 180 240 300 360],'ytick',[0 5 15 25 30])
% hold off
% subplot(2,2,4);plot(aci1,a44,'m>')
% hold on
% subplot(2,2,4);plot(aci1,a55,'ko')
% axis([-100 100 30 50])
% title('Fast Speed vs Angle','FontSize',10)
% xlabel('Side Slip Angle (Deg)','FontSize',7)
% ylabel('Predicted Speed (Kt)','FontSize',7)
% set(gca,'xtick',[-60 -30 0 30 60],'ytick',[30 35 40 45 50])
% hold off
% For Baseline data plotting speed vs. angle use the following section of the routine
for df=1:12
a1m1(df,:)=a11(df,:);
a2m1(df,:)=a22(df,:);
a3m1(df,:)=a33(df,:);
end
for vc=1:5
a4m1(vc,:)=a44(vc,:);
a5m1(vc,:)=a55(vc,:);
end
nd=1;
for gf=13:24
a1m2(nd,:)=a11(gf,:);
a2m2(nd,:)=a22(gf,:);
a3m2(nd,:)=a33(gf,:);
nd=nd+1;
114
end
nd1=1;
for gf1=6:10
a4m2(nd1,:)=a44(gf1,:);
a5m2(nd1,:)=a55(gf1,:);
nd1=nd1+1;
end
subplot(2,2,3);plot(aci,a1m1,'ro')
hold on
subplot(2,2,3);plot(aci,a1m2,'ro')
hold on
subplot(2,2,3);plot(aci,a2m1,'g<')
hold on
subplot(2,2,3);plot(aci,a2m2,'g<')
hold on
subplot(2,2,3);plot(aci,a3m1,'bo')
hold on
subplot(2,2,3);plot(aci,a3m2,'bo')
axis([0 400 0 35])
title('Slow Speed vs Angle','FontSize',10)
xlabel('Side Slip Angle (Deg)','FontSize',7)
set(gca,'xtick',[0 60 120 180 240 300 360],'ytick',[0 5 15 25 30])
hold off
subplot(2,2,4);plot(aci1,a4m1,'m>')
hold on
subplot(2,2,4);plot(aci1,a4m2,'m>')
hold on
subplot(2,2,4);plot(aci1,a5m1,'ko')
hold on
subplot(2,2,4);plot(aci1,a5m2,'ko')
axis([-100 100 30 50])
title('Fast Speed vs Angle','FontSize',10)
xlabel('Side Slip Angle (Deg)','FontSize',7)
set(gca,'xtick',[-60 -30 0 30 60],'ytick',[30 35 40 45 50])
hold off
% Finding max errors for the worst case

for i=1:n
fark(i)=(abs(ilk(i)-a(i)));
end
maxfark=max(fark)
maxfark5=max(fark5)
maxfark15=max(fark15)
115
% Computation of SD and percent error for each speed:

mn5=mean(a1)
sd5=std(a1)
per5=(sd5/5)*100
mn15=mean(a2)
sd15=std(a2)
per15=(sd15/15)*100
mn25=mean(a3)
sd25=std(a3)
per25=(sd25/25)*100
mn35=mean(a4)
sd35=std(a4)
per35=(sd35/35)*100
mn45=mean(a5)
sd45=std(a5)
per45=(sd45/45)*100
% SD for whole data set
sdw=sqrt((((na/3)-1)*(sd5^2+sd15^2+sd25^2)+(((n-na)/2)-1)* sd35^2+sd45^2)) /(n-1))
116
INITIAL DISTRIBUTION LIST
1. Defense Technical Information Center

Ft. Belvoir, Virginia
2. Dudley Knox Library

Naval Postgraduate School
3. Chairman
Department of Aerospace and Aeronautics, Code AA
Naval Postgraduate School,
4. Dr. Russell Duren

Department of Aerospace and Aeronautics, Code AA/Dr
5. Dr. Monique P. Fargues

Department of Electrical and Computer Engineering, Code EC/Fa
6. Ozcan Samlioglu
7nci. Kor. Tk. Hv. Gr. K.ligi, Hlkp. Tb. K.ligi
Diyarbakir, TURKEY
7. K K.K.ligi. Kara Harp Okulu Kutuphanesi

Kara Harp Okulu, Bakanliklar,
Ankara, 06200 TURKEY
8. Kara Kuvvetleri K.ligi Kutuphanesi

Bakanliklar,
Ankara, 06200 TURKEY
9. Hasan Akkoc
Kara Harp Okulu Ogr. Bsk.ligi,
Bakanliklar, Ankara, 06200 TURKEY
117

Ozcan Samlioglu

Uploaded by

Copyright:

Available Formats

Ozcan Samlioglu

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Ozcan Samlioglu

Uploaded by

Copyright:

Available Formats

NAVAL POSTGRADUATE SCHOOL

Thesis Advisor: Russell W. Duren

Approved for public release; distribution is unlimited.

14. SUBJECT TERMS 15. NUMBER OF

A NEURAL NETWORK APPROACH FOR HELICOPTER AIRSPEED

Submitted in partial fulfillment of the

MASTER OF SCIENCE IN AERONAUTICAL ENGINEERING

NAVAL POSTGRADUATE SCHOOL

Author: Ozcan Samlioglu

Approved by: R. W. Duren, Thesis Advisor

M. P. Fargues, Assoc. Advisor

Max F. Platzer, Chairman

Conventional pitot-static airspeed measurement systems do not yield accurate

Figure 1. Structure of a Neuron [From Ref. 10]................................................................8

Table 1. Neural Network Input Parameters. ..................................................................21

The aerodynamic velocity in the plane of a helicopter’s rotor disc significantly

A NN implementation is chosen for several reasons. First, it is easy to use. Input

The primary objective of this thesis is to determine the airspeed of an OH-6A

A. INTRODUCTION TO NEURAL NETWORKS

Dendrites : Accept inputs

So ma : Process the inputs

Axon : Turn the processed inputs

Synapses : The electrochemical

Figure 1. Structure of a Neuron [From Ref. 10].

The transfer function is also referred to as an “activation function”. This transfer

The NN is trained to perform a particular function by adjusting the values of

Figure 2. System Diagram of NN [From Ref.2].

A NN consists of many PEs grouped together called layers. A layer is defined as a

Figure 3. Layer structure of NN [From Ref. 10].

A layer consists of the following elements:

• An input vector (p),

In the mid-1960s, Marvin Minsky and Seymour Papert, considered the NN

Although in 1974 Paul Werbos described a similar algorithm to train multilayer

Back-propagation is a very popular NN learning algorithm to train multilayer

Figure 5. A Typical Back-Propagation Network [From Ref. 2].

A BPNN is typically represented with the following notation: R-S1-S2-S3, where

The second step is to propagate the errors, or in other words, sensitivities

Thus, given the error is defined as:

and the expectation of mean square error is defined as

f ( x) = E ([eT e]) = E[(t − a)T (t − a)] , (2.3)

the expectation of the squared error at iteration k becomes [Ref. 1]

f ( x) = E[(t (k ) − a(k ))T (t (k ) − a(k ))] . (2.4)

The algorithm adjusts weights and biases as follows [Ref. 1]

The next step is to propagate the sensitivities backward

Although the BP network is a very powerful technique, it usually requires a long

In summary, the BPNN is a commonly used technique for solving nonlinear

A. NEURALWORKS PROFESSIONAL II/PLUS SOFTWARE

Figure 6. Typical Window of Neural Works Professional II/Plus Software.

Simulators also play an important role in engineering processes such as design

FLIGHTLAB is a commercial-off-the-shelf type of software product developed

FLIGHTLAB uses a simulation language called Scope. This language is an

Another feature of the FLIGHTLAB Simulation System is Xanalysis. This tool

C. SELECTING DATA AND BUILDING THE MODEL

MODEL VARIABLES FOR UH-60A HELICOPTER

Table 1. Neural Network Input Parameters.

NAME INPUT/OUTPUT RELATION

Hard Limit a=0 n<0

Hyperbolic Tangent Sigmoid en − e− n

Competitive a = 1 neuron with max n

a = 0 all other neurons

%*for UH-60A helicopter add parameter “eng_torque” to the matrix below**