Towards Gaussian Process Models of Complex Rotorcraft Dynamics

Ryan D Jackson Michael Jump Peter L Green

School Of Engineering School of Engineering School of Engineering
Institute for Risk and Uncertainty University of Liverpool Institute for Risk and Uncertainty
University of Liverpool Liverpool, UK University of Liverpool
Liverpool, UK L69 3GQ Liverpool, UK
L69 3GQ L69 3GQ
ABSTRACT
Physical law based models (also known as white box models) are widely applied in the aerospace industry, providing
models for dynamic systems such as helicopter flight simulators. To meet the criteria of real-time simulation, simplifi-
cations to the underlying physics sometimes have to be applied, leading to errors in the model’s predictions. Grey-box
models use both physics-based and data-based models. They have potential to reduce the difference between a simula-
tor’s and real rotorcraft’s response. In the current work, a preliminary step to the grey-box approach, a machine learnt
data-based, i.e ‘black box’ model is applied to the dynamic response of a helicopter. The machine learning methods
used are probabilistic and can capture uncertainties associated with the model’s prediction. In the current paper, ma-
chine learning is used to create a Gaussian Process (GP) non-linear autoregressive (NARX) model that predicts pitch,
roll and yaw rate. The predictions are compared to a physical law based model created using FLIGHTLAB software.
The GP outperforms the FLIGHTLAB model in terms of root mean squared error, when predicting the pitch, roll and
yaw rate of a Bo105 helicopter.

NOTATION y∗ Vector of training observations including new predic-

tion y = [y1 , y2 , . . . , yN , y∗ ]T
T Transpose
δo Collective lever
α Hyperparameter
δp Pedal position
β Precision of noise
δx Longitudinal stick position
Φ Design Matrix
δy Lateral stick position
φ Vector of ’basis functions’ ε Noise term goverened by ε ∼ N (ε | 0, β −1 )
C Covariance matrix with additional noise parameter, γ w) = N (w
Precision of the distribution p(w w | 0 , γ −1 I
such as C = K + I β −1
E Expected value
C N+1 (N + 1) × (N + 1) covariance matrix
GP Gaussian Process
f Vector of latent function
N Normal Distribution
I Identity Matrix
µ Mean
K Covariance Matrix (Noise Free)
σ2 Variance
k Vector, short for k (xxn , x∗ ) cov Covariance
w Vector of parameters θ Hyperparameters to be optimised, α and β
x Input Vector c Short for β −1 + k(xx∗ , x ∗ )
x∗ New Input Vector Cnm C)
Elements of Covariance matrix (C

y Vector of training observations y = [y1 , y2 , . . . , yN ]T f Latent function values

Presented at the AHS International 74th Annual Forum & Inm Elements of identity matrix (II )
Technology Display, Phoenix, Arizona, USA, May 14–17, 2018.
Copyright c 2018 by AHS International, Inc. All rights reserved. k Kernel Function
1
n Time step approach ensures that the ‘black box’ (machine-learnt) model
will step-in only when the simplified physical-law model de-
R Number of Monte Carlo runs viates from the ideal. In the current study, as a preliminary
step towards this goal, the authors aim to develop and vali-
v Generic system input
date a black box model using machine learning methods that
y∗ Prediction from the Gaussian Process given x ∗ can emulate the dynamic behaviour of a rotorcraft. Crucially,
the methods used are probabilistic and are therefore able to
capture the uncertainties associated with such an approach.
INTRODUCTION
In this paper, machine learning is used to create a Gaussian
Flight simulators are a vital part of the aircraft life cycle. They process (GP) non-linear autoregressive model that predicts
are used in design and development phases, testing and qual- pitch, roll and yaw rate of the Bo105 helicopter trained only
ification activities as well as in training and research (Ref. 1, upon a longitudinal cyclic input. The autoregressive model
2). Due to their availability and cost compared to the corre- makes a prediction using the longitudinal cyclic position as
sponding real actual aircraft, their use continues to increase. well as previous observations of the relevant output.
Simulators can also help to address the increased demand for
new pilots, who are needed to replace the current ageing popu- The Gaussian process model for pitch, roll and yaw rate
lation of pilots. Moreover, the military is increasing the use of are compared to physical law based models, which are im-
simulators for mission rehearsal in land, sea and air contexts. plemented using Advanced Rotorcraft Technology’s (ART)
FLIGHTLAB software (Ref. 9). The comparison provides
The heart of any flight simulation facility is the flight dynam-
an excellent basis for future work where our approach will
ics model. Techniques to design and develop such models
predict model error.
are well known and documented (Ref. 3, 4). However, there
is a requirement for the entire simulation system to run in
real-time and this can lead to simplifications to the underlying NUMERICAL METHODS
physics having to be made, particularly for more complex air-
craft such as rotorcraft. These simplifications that are applied Gaussian Processes
mean that the flight model cannot necessarily capture all of the
complex dynamics that would be present during the equivalent
real scenario. They can lead to significant differences between Gaussian Processes (GPs) have been widely used in recent
the model and the real aircraft. These differences can, in the years for many different applications. In the current work,
worst case, have a negative impact, for example, on training GPs are used to perform regression (they can also perform
for the crew using the simulator. other tasks such as classification (Ref. 10)). An advantageous
property of GPs is that they can be used to quantify the un-
The quality of the flight dynamics model speaks to the ‘fi- certainties in one’s predictions which, here, produces a worst
delity’ of the simulation device. The ‘engineering fidelity’ of case scenario given the model uncertainty. This property of
such a device is typically measured using a series of quan- being able to quantify errors is useful. For example, in the
titative requirements contained within simulator qualification current context whereby the authors wish to extend the work
documents such as (Ref. 5, 6). It is recognized that examining to predict model error, GPs could be used to quantify the
the response of the simulator in this way only partially serves uncertainties involved in capturing the discrepancies between
to characterize its utility. While efforts are underway to seek physical-law based simulators and reality. Another beneficial
methods that can better meet this need (Ref. 7), this paper property of GPs is that, once trained, they can produce
seeks to explore techniques whereby the accuracy of the flight very fast emulators of complex models - this is beneficial
dynamics element of the simulation device can be improved, when dealing with non-linear behaviour such as helicopter
even when the modelled physics can no longer accurately rep- dynamics.
resent the situation.
The aim, then, of the research presented in this paper, was to
(start to) develop simulation methods that can more accurately
capture the complex dynamics of rotorcraft, while still being Linear Regression To aid the understanding of the Gaussian
able to be run in real-time. The research is based on the hy- process, consider a model which is defined as a linear combi-
pothesis that current flight dynamics models can be improved nation of fixed basis functions (Ref. 10, 11):
using machine learning; data-based models that, once trained,
can predict the model error 1 that is present in current simula- f (xx) = w T φ (xx) (1)
tors. This approach aims to generate a ‘grey box’ model that
combines physical-law based simulations (i.e. ‘white box’) where x is the input vector, w is a vector of parameters to be
with data based simulations (i.e. ‘black box’). The ‘grey box’ identified, φ is a vector of ‘basis functions’ and f is a latent
function. One can choose the prior distribution over w to be:
1 Model error captures the inaccuracy of the physical law based model

compared to that of the flight test data. For more information on model error
see the study by Kennedy and O’Hagan (Ref. 8) p(w w | 0 , γ −1 I )
w) = N (w (2)
2
 where γ is the precision of the distribution, and I is an iden- discussed in a later Section).
tity matrix. The function values are given by vector f :

f (xx1 )
 The distribution over y , conditional on f , is given by:
 f (xx2 ) 
f = . 
 
(3) p(yy | f ) = N (yy | f , β −1 I ). (11)
 .. 
f (xxN ) The marginal distribution of y is defined as:
Z
where N is the number of of training points. Using equation p(yy) = p(yy | f )p( f )d f (12)
(1), the vector f is given by:

f = Φw (4) which, given equations (9) and (11), allows us to write

where Φ is a design matrix (for more information see Bishop p(yy) = p(yy | 0 ,C
C) (13)
(Ref. 10)) . Given the prior (equation (2)), one can then show
that f is Gaussian, with mean and covariance matrix: where the elements of C are given by:

w] = 0
E[ f ] = Φ E[w (5) Cnm = C(xxn , x m ) = k(xxn , x m ) + β −1 Inm (14)

where Inm is an element of identity matrix I .

T 1
Φ = ΦΦT = K
ww ]Φ
cov[ f ] = E[ f f ] = Φ E[w T T
(6)
β
Prediction The main aim of GP regression is to make pre-
where K is a covariance matrix with elements : dictions for new input data that are not in the training data.
1 Given a new input vector, x ∗ , one can then estimate the proba-
Knm = φ (xxn )T φ (xxm ). (7) bility of a new point y∗ given previous observations y . To find
γ
y∗ the predictive distribution is evaluated, which is given by
where n, m = 1, . . . , N. p(y∗ | y ). Defining y ∗ = [y1 , y2 , . . . , yN , y∗ ]T , the joint distribu-
tion over y ∗ is
Gaussian Process for regression Regression problems of-
ten include noise on the observed training data. To apply p(yy∗ ) = N (yy∗ | 0 ,C
C N+1 ) (15)
GPs to regression the defined training data should also include
noise: where C N+1 is a (N + 1) × (N + 1) covariance matrix. C N+1
y n = f n + εn (8) can be shown to be:
 
where yn represents the nth observation of the system’s re- CN k
C N+1 = T (16)
sponse, fn = f (xn ) and εn ∼ N (εn | 0, β −1 ), where β is the k c
precision of noise. Instead of defining basis functions, as in
equation (1), with a GP one can simply define a prior of the where:
form:
k = k(xxn , x ∗ ), n = 1, . . . , N (17)
p( f ) = N ( f | 0 , K ) (9)
where the covariance matrix, K is given by: and
c = β −1 + k(xx∗ , x ∗ ). (18)
 
k(xx1 , x 1 ) k(xx2 , x 1 ) . . . k(xxN , x 1 )
 k(xx1 , x 2 ) k(xx2 , x 2 ) . . . k(xxN , x 2 )  The mean and variance of y∗ given y can be shown to be (Ref.
K =

.. .. .. ..

 10):
 . . . .  µ(y∗ | y ) = k T C −1
N y
k(xx1 , x N ) k(xx2 , x N ) . . . k(xxN , x N )
σ 2 (y∗ | y) = c − k T C N k
which is produced by a user-defined kernel function (k). An
example of such a kernel function is the squared exponential such that:
(Ref. 10): p(y∗ | y ) = N (y∗ | µ, σ 2 ) (19)

K nm = k(xxn , x m ) = exp − α2 (xxn − x m )T (xxn − x m )

(10) where µ is the mean prediction and σ 2 is the variance which
is used to measure the uncertainty in the predictions of y∗ .
where the ‘hyperparameter’ (α), induces correlations that This can be written more compactly using the notation
depend on the ‘closeness’ of x n and x m . Choosing different
hyperparameters can affect how accurate the GP model is. y∗ ∼ GP(xx∗ ) (20)
To achieve ‘optimum’ values for the hyperparameters, an
optimisation technique is required (these techniques are
3
Choice of Kernel One Step ahead predictions

The kernel chosen for the current study was used by Higdon One step ahead predictions (OSAP) use the previously ob-
et al (Ref. 12). The kernel takes the form of: served data to predict a single step into the future. Using GP
N to represent the prediction made by a Gaussian Process then,
i i 2)
k(xxn , x m ) = ∏ α (4(xxn −xxm ) (21) in our specific case, a OSAP is defined as:
i
y∗n = GP (δnx , yn−1 ) (25)
where x i is used to represent the ith element of the vector x .
The kernel of equation (21) allows the hyperparameter (α) to where yn−1 is the previous observation of the relevant quantity
always be between zero and one. This is beneficial when us- (i.e. roll, pitch or yaw). During ‘training’ of the GP, OSAP
ing a property of GPs that allows the relevance of inputs to are used to quantify the fidelity of the emulator. The obvious
be determined (known as ‘Automatic Relevance Determina- disadvantage of this approach is that the resulting model can
tion’). The relevance property is not explored in the present only predict a single step into the future. Predicting further
paper, however it is of future interest for the authors. into the future requires ‘full model predictions’.

Gaussian Process NARX models

Full Model Predictions
Depending on their input structure, GPs can be used to em-
ulate static or dynamic relationships. The non-linear autore- To make predictions beyond a single step, the GP-NARX
gressive with exogenous inputs (NARX) structure has format: framework must utilise previous predictions as part of the
model input. To illustrate this, we consider the situation where
a single prediction, at time n, has already been made ac-
yn = f (yn−1 , yn−2 , yn−3 , . . . , vn , vn−1 , vn−2 , vn−3 , . . . ) + εn cording to equation (25). Following this, predictions of y∗n+1
(22) would be realised according to
where v is a generic system input and, as before, y represents
system observations, n is the time step and f is the function
y∗n+1 = GP δn+1
x
, y∗n .


that we wish to model. The NARX structure uses information (26)
from ‘lagged’ terms (observations and inputs) and the current The key aspect to note with regard to equation (26) is that y∗n
input to help predict yn . The NARX structure has been used - the uncertain prediction made by the GP at time n - is now
as the input structure when modelling non-linear dynamical part of the model input.
system such as the Duffing oscillator (Ref. 13, 14). It has also
been used to develop models of helicopter dynamics (Ref.
A Monte Carlo analysis can be used to address the additional
15, 16, 17).
uncertainty that is introduced by including y∗n as a model
input. By definition, y∗n is a Gaussian random variable and, as
The helicopter investigated in the current paper is treated as such, samples of y∗n can be generated easily.
having four inputs; longitudinal stick position (δ x ), lateral
stick position (δ y ), pedal position (δ p ) and collective lever
(δ o ). The outputs correspond to the helicopter’s three axes: A simple algorithm for generating an ensemble of predictions
pitch rate, roll rate and yaw rate. The NARX structure is re- for full model predictions (FMP) is shown in ‘Algorithm 1’
quired to capture the behaviour of the helicopter’s dynamics
(Ref. 15, 16, 17). Considering an example where the goal is Algorithm 1 Full Model Predictions algorithm
to map the longitudinal stick position to pitch rate the NARX 1: for r = 1 : R do
input structure for the GP, is: 2: ∗
Y(n),r ∼ GP(y(n−1) , δn )
∗ ∗ ,δ
∼ GP(Y(n),r
3: Y(n+1),r n+1 )
 x 
δn
xn = (23) 4: end for
yn−1
where y is the relevant observation (pitch, roll or yaw rate).
When using the GP NARX model the kernel from equation where R is the number of Monte Carlo samples and Y denotes
(21) takes the form: a sample from taken from the GP.
x x 2 2 For a more detailed algorithm for FMP, see the thesis by Gi-
k(xxn , x m ) = α 4(δn −δm ) × α 4(yn−1 −ym−1 ) . (24) rard (Ref. 18).
In the following, the inputs and outputs were normalised to
ensure that all values are between 0 and 1. The model gener- Optimisation of hyperparameters for Gaussian Processes
ated here will be used to produce what is known as ‘one-step- training
at-a-time predictions’ and ‘full model predictions’. These two
types of predictions are discussed, in detail, in the following The GP process requires the optimisation of hyperparameters
two Sections. to create an optimal model. To facilitate this, one can generate
4
samples from the posterior parameter distribution, which is Propagating hyperparameter uncertainty MCMC creates
given by Bayes’ rule: samples of the hyperparameters (θ ). The GP also has uncer-
tainty associated with its predictions, which can be combined
p(θ | y ) ∝ p(yy | θ )p(θ ) (27) with the hyperparameter uncertainty. Including both sources
of uncertainty, the uncertainty in one’s predictions can be writ-
where θ = [α, β ]T .The techniques used in the current paper ten as:
are discussed in the next section. y∗ ∼ N (µ(θ ), σ 2 (θ )) (29)
where:
Markov Chain Monte Carlo (MCMC) MCMC can be used θ ∼ p(θ | y). (30)
to generate samples from the posterior hyperparameter distri-
bution given in equation (27) , thus quantifying uncertainties Sampling from equation (29), using samples from equation
in the hyperparameter selection. It also provides a more global (30), is known as ‘Ancestral sampling’ where the posterior
search, unlike the local search provided by methods such as samples are generated using MCMC. For more information
gradient-based ascent. Utilising MCMC to generate samples on Ancestral sample see Bishop (Ref. 10).
from the posterior hyperparameter distribution is of interest
for the authors. This enables the incorporation of hyperpa- RESULTS
rameter uncertainty into the GP prediction.
The best known MCMC method is the Metropolis algorithm In this Section, the 3-axis rotational responses of the FLIGHT-
(MA) (Ref. 10). The target distribution, π(θ ), is equal to LAB Bo105 model to a longitudinal 3-2-1-1 cyclic input are
p(θ | y ). The first step in each iteration of MA is to propose compared with the GP model’s output. Both are compared
a new state θ 0 , where the current state of the Markov chain with data from equivalent Bo105 flight test data. The flight
is θ (r) (Ref. 19). The proposal is taken from a probability test control input from the Bo105 helicopter has been applied
density function (PDF) q(θ 0 | θ (r) ) which is conditional on to the physical-law based FLIGHTLAB model. The input data
the current state. The proposal is accepted as the new state of is shown in Figure 1:
the Markov chain with probability:
62
π(θ 0 )) π ∗ (θ 0 ))
n o n o
min 1, π(θ (r) ) = min 1, π ∗ (θ (r) ) (28) 60
Longitudinal stick position (%)

58

where π ∗ (θ )is the un-normalised target distribution. If ac- 56

cepted, the new state of the Markov chain is θ (r+1) = θ 0 , oth- 54

erwise θ (r+1) = θ (r) . This process is run for a user-defined 52

number of samples. The Markov chain will then reach a sta- 50

tionary distribution producing posterior samples of the hyper- 48
parameters. 46
0 1 2 3 4 5 6 7 8
Time (s)

Simulated Annealing Simulated annealing is a form of Fig. 1: 3-2-1-1 Longitudinal stick input
MCMC which uses MA. The difference is that the method
slowly increases the influence of the likelihood, via a variable
ζ (Ref. 20). Increasing the influence of the likelihood Using the theory, cited in the previous Section, a GP model
increases the influence of the data and this assists algorithm was also created to predict the Bo105 pitch, roll and yaw rates
convergence. As with the MA, a prior is chosen to generate from the 3-2-1-1 longitudinal cyclic input. We note that, here,
a candidate hyperparameter θ (s) , a small initial value of ζ the GP models do not use all of the available training data..
is also chosen. For each ζ value, a full run of the MA is The GPs only use 32 ‘training’ points from Figure 1, which
conducted using a user-defined number of samples. The rate are denoted by diamonds. All three GP models (pitch, roll and
at which ζ is increased is called the annealing schedule. yaw) are compared to the Bo105 flight model predictions and
In the current paper an adaptive annealing schedule is used the flight test data in the following sections.
to ensure constant change in Shannon entropy of the target
distribution (Ref. 20). The optimisation stops when ζ reaches
Pitch Rate
one, in which case the whole likelihood has been introduced.
The pitch rate response is investigated first. Figure 2 shows
The advantage of using this method over the standard MA is the OSAP predictions made by the GP. As was discussed pre-
that is it easier to tune and often demonstrates better conver- viously, it would be expected that the predictions are very
gence, as the proposal width is being updated after each run close the Bo105 flight test data. The RMS error between the
of the MA. GP model and the Flight Test data is 0.0012.
5
 a Gaussian. In the case of FMP, the predictions have come
0.15
from a mixture of Gaussian’s.. Another possible reason is that
0.1
the GP assume the likelihood is also a Gaussian, this may not
0.05
be the case for a real system such as the Bo105 helicopter. For
Pitch Rate (rad/s)

0
more information on non Gaussian likelihoods see the study
-0.05 by Saul et al. (Ref. 21).
-0.1

-0.15 Confidence bounds

Flight Test data
-0.2 GP Model
Training Points 0.15
-0.25
0 1 2 3 4 5 6 7 8
0.1
Time (s)
0.05

Pitch Rate (rad/s)

Fig. 2: Gaussian process one step ahead pitch rate predictions 0
using hyperparameters located by simulated annealing -0.05

-0.1

-0.15
Figure 3 shows the predicted helicopter response from the Confidence bounds
Flight Test data
FMP. This is a better test of the GP model. It is clear to see -0.2 GP Model
Training Points
that the predictions are not as accurate as for the OSAP model. -0.25
0 1 2 3 4 5 6 7 8
However, this is to be expected given that, for this model, the Time (s)

previous prediction becomes an input to the next prediction

Fig. 4: Gaussian process full model pitch rate predictions with
and, hence, predictive uncertainty is carried through the sim-
incorporated hyperparameter uncertainty from the simulated
ulation.
annealing results
0.15

0.1

0.05
Pitch Rate (rad/s)

-0.05

-0.1

-0.15 Confidence bounds

Flight Test data
-0.2 GP Model
Training Points
-0.25
0 1 2 3 4 5 6 7 8
Time (s)

Fig. 3: Gaussian process full model pitch rate predictions us-

ing hyperparameters located by simulated annealing

Note that, in Figures 2 and 3, hyperparameter uncertainty is

Fig. 6: Pitch rate Simulated annealing results
not considered. The hyperparameter and the full model pre-
dictions uncertainty are included in Figure 4. Note that be-
tween Figure 3 and 4 that the confidence bounds are very sim-
ilar. This implies that the hyperparameter uncertainty does
not have a significant affect compared to the use of the FMP.
To calculate hyperparameter uncertainty, MCMC samples are
The comparison of the flight test data, FLIGHTLAB and GP
required. The posterior hyperparameter samples for the GP
models, are shown in Figure 5. Both models capture the
predicting pitch rate are shown in Figure 6.
essence of the real aircraft response. However, it is apparent
It can be seen in Figures 3 and 4 that the confidence bounds that the GP model outperforms the FLIGHTLAB model in
do not encompass all of the flight test data. It is currently this case; in general, the GP predictions remain much closer
unclear why this occurs, therefore it is of future interest for to the flight test truth data throughout the maneuver but partic-
the authors. One reason could be, when using Monte Carlo ularly in its latter stages. To try to quantify this improvement,
samples for the FMP, the mean and variance are taken from the root mean squared errors between prediction and truth data
a mixture of Gaussian’s (Ref. 18). The confidence bounds for both models were computed (RMSE). The results of this
are calculated for three standard deviations. The belief that exercise are given in Table 1. These confirm what is easily
99.7% of the data falls within the three standard deviations is observable in the Figure; that, for the primary axis response,
based on the assumptions that the predictions have come from the GP model outperforms the FLIGHTLAB model.
6
 0.2

0.1
Pitch Rate (rad/s)

-0.1

-0.2
Confidence bound
Flight Test data
-0.3
FLIGHTLAB Model
GP Model
-0.4
0 1 2 3 4 5 6 7 8
Time (s)
Fig. 5: Comparison of the pitch rate response of the real helicopter to the FLIGHTLAB and GP model

Prediction of pitch rate, corresponding to Figure 5 0.1

Model RMSE 0
Gaussian Process 0.019 -0.1
FLIGHTLAB Model 0.074
Pitch Rate (rad/s)

-0.2

-0.3
Table 1: RMSE of the predictions of pitch rate for the GP and
FLIGHTLAB models compared to the flight test data of the -0.4

Bo105. -0.5 Confidence bound

Flight Test data
-0.6 FLIGHTLAB Model
GP Model
-0.7
0 1 2 3 4 5 6 7 8
Time (s)

Fig. 7: Gaussian process full model pitch rate predictions

with ‘unseen’ data using the incorporated hyperparameter un-
The GP was trained using only a subset of the training data. certainty from the simulated annealing results and including
In the previous figures, the GP is interpolating between points a comparison between the FLIGHTLAB, GP model and the
in the training set. This could justify why the GP in Figure flight test data
2 shows a very accurate fit. A more challenging test for the
GP model is to predict data that has not been ‘seen’ by it be-
fore i.e. data that it has not been trained on. Figure 7 shows
the GP-predicted response to another 3-2-1-1 flight test ma-
neuver. The GP was trained on a 3-2-1-1 maneuver where
the initial longitudinal cyclic input was in a positive input di-
rection. The next, validation case in Figure 7 is the opposite
i.e. the initial longitudinal cyclic input was in the negative
direction. As such, this prediction is based upon previously
‘unseen’ data.
In a qualitative sense, Figure 7 shows a reasonably good
match between the GP model prediction and the flight test Table 2 shows the RMS error for the FLIGHTLAB and GP
truth data. The global features of the response are captured. It data to the ‘unseen’ validation data. The GP model is not as
is arguably less good than the prediction of Figure 4 and does accurate in terms of RMS error compared to Table 1, however
not capture some of the higher frequency behaviour observ- this is expected as this is data that the GP has not been trained
able in the flight test data. on. The GP model far outperforms the FLIGHTLAB model.
7
 Prediction of pitch rate, corresponding to Figure 7
Model RMSE
Gaussian Process 0.022
FLIGHTLAB Model 0.357

Table 2: RMSE of the predictions of the validation ‘unseen’

pitch rate for the GP and FLIGHTLAB models compared to
the flight test data of the Bo105.

Fig. 9: Roll rate Simulated annealing results

Roll Rate

Figure 10 shows the comparison of the GP predictions of roll

The helicopter roll response to a longitudinal cyclic input is an rate to that of the FLIGHTLAB model and the Bo105 heli-
off-axis response and is known to be hard to capture for real- copter data. The FLIGHTLAB model captures the initial part
time simulation using existing physics-based modelling tech- of the response very well but then becomes less accurate as
niques. For the sake of brevity, the OSAP and FMP predic- the maneuver progresses. Conversely, the GP predictions do
tion without HP uncertainty are not shown for roll (and yaw) less well at the start of the maneuver but become increasingly
rate. The conclusions for these are the same as those drawn accurate towards the final stages. The RMS error analysis of
for the pitch rate predictions. Figure 8 shows the prediction the maneuver indicates that, on average, the GP outperforms
of roll rate response made by GPs FMP with hyperparameter the FLIGHTLAB model. These data are shown in Table 3.
uncertainty included. The main character of the response is This shows that RMS error analysis, while of some value, is
captured, particularly towards the end of the maneuver. The somewhat of a blunt tool in this case as it does not capture
wider GP confidence bounds compared to the pitch rate re- the nuances of the response predictions as the maneuver pro-
sponse are indicative a reduced confidence of the model in its gresses.
predictions; presumably a function of the off-axis nature of
the response. It is interesting to note, however, that the fight
test data does lie, to a greater extent within those confidence Prediction of roll rate, corresponding to Figure 10
bounds. Model RMSE
Gaussian Process 0.047
FLIGHTLAB Model 0.059
0.2

Table 3: RMSE of the predictions of roll rate for the GP and

0.1
FLIGHTLAB models compared to the flight test data of the
Bo105.
Roll Rate (rad/s)

-0.1

Confidence bounds
-0.2 Flight Test data
GP Model
Training Points
-0.3
0 1 2 3 4 5 6 7 8
Yaw Rate
Time (s)

Fig. 8: Gaussian process full model roll rate predictions with

incorporated hyperparameter uncertainty from the simulated Like roll rate, the yaw rate induced by a longitudinal cyclic
annealing results input is an off-axis response and this, again, can be difficult to
model accurately using physics-based modelling techniques.
The predicted yaw rate response of the Bo105 for FMP for
the GP with hyperparameters is given in Figure 11. It is clear
that the GP struggles to make an accurate prediction, with a
Figure 9 shows the accepted samples for the hyperparameters significant portion of the maneuver being predicted to be in
using simulated annealing for the prediction of roll rate. The the opposite direction to the flight test data. Once again, how-
accepted samples are used to generate the additional hyperpa- ever, much of the flight test response is encompassed within
rameters uncertainty in the FMP. this uncertainty.
8
 0.2

0.1
Roll Rate (rad/s)

-0.1

Confidence Bounds
-0.2 Flight Test data
FLIGHTLAB Model
GP Model
-0.3
0 1 2 3 4 5 6 7 8
Time (s)
Fig. 10: Comparison of the roll rate response of the real helicopter to the FLIGHTLAB and GP model

0.2
The comparison of the GP model for the prediction of yaw rate
Confidence bounds
Flight Test data
to the FLIGHTLAB model and the Bo105 helicopter data is
0.15 GP Model
Training Points
shown in Figure 13. Both models make quite poor predictions
0.1 overall with the FLIGHTLAB model over predicting the yaw
Yaw Rate (rad/s)

0.05 response quite significantly in the later stages of the maneuver.

0
This is confirmed in the RMSE values shown in Table 4. The
GP has a slightly smaller value of RMS error.
-0.05

-0.1
Prediction of yaw rate, corresponding to Figure 13
-0.15
0 1 2 3 4 5 6 7 8
Model RMSE
Time (s) Gaussian Process 0.043
FLIGHTLAB Model 0.052
Fig. 11: Gaussian process full model yaw rate predictions
with incorporated hyperparameter uncertainty from the sim-
Table 4: RMSE of the predictions of yaw rate for the GP and
ulated annealing results
FLIGHTLAB models compared to the flight test data of the
Bo105.
The accepted samples of the GP model for yaw rate are shown
in Figure 9. The accepted samples are used in the hyperpa-
rameter uncertainty.
DISCUSSION
This investigation has presented some early results relating to
the potential use of Gaussian Process models to the applica-
tion of real-time helicopter response prediction. The on-axis
predictions of Figures 5 and 7 show promise for the methods.
The method works well on the data that it was trained on, as
might be expected. However, it also works well on unseen
data. The slightly surprising result is how well the GP model
works for off-axis response rates. In this regard, the GP model
performed no less well than the physics-based FLIGHTLAB
model and, in a global sense at least, introduced a lower over-
all magnitude of error into the response. That said, the GP
Fig. 12: Yaw rate Simulated annealing results models do not yet predict the finer points of the responses
9
 0.2
Confidence Bounds
0.15 Flight Test data
FLIGHTLAB Model
GP Model
0.1
Yaw Rate (rad/s)

0.05

-0.05

-0.1

-0.15
0 1 2 3 4 5 6 7 8
Time (s)
Fig. 13: Comparison of the yaw rate response of the real helicopter to the FLIGHTLAB and GP model

shown in the flight test data. It is hoped that this might be im- FUTURE WORK
proved when the models are trained using lateral and rudder
pedal response data. As noted above, the on-axis predictions from the GP model
are very good, the off-axis predictions less so. It is anticipated
Where this method shows promise is in its ability to indicate
that the prediction of roll and yaw rate could be improved by
estimates of the uncertainty in its predictions. A key advan-
incorporating other inputs into the NARX structure, i.e. re-
tage over GPs relative to other methods, is that GPs can pro-
sponse data from flight test points that use lateral cyclic and
vide confidence bounds on their predictions. It can be seen
rudder pedal position. This will form an early part of the fu-
from Figures 5 and 13 that the level of confidence in the
ture work.
pitch rate prediction is higher than the yaw rate, for exam-
ple, as might be expected from the source of the training data. As this work is only a preliminary step towards the author’s
This ability to provide the degree of uncertainty might, in the main aim which is the prediction of model error, this would
longer term, be able to inform differences in, for example, pi- be one of the next steps. Further future work would include
lot opinion or workload/handling qualities rating between the implementing our data-based model into e.g. FLIGHTLAB
real vehicle and simulator experiments. software for simulations purposes.
The GP was only trained on a limited amount of data; these
CONCLUSION training points were chosen by selecting every 25th point. One
of the problems associated with GPs is their computational
Gaussian Process models, using only a small amount of
cost of training (O(N 3 ), where N is the number of training
training data, can produce excellent predictions of on-axis
points). It is therefore beneficial to choose a subset of the
longitudinal helicopter response data. In both a qualitative
available data to use for training. A method called ‘sparse
and quantitative sense (lower RMSE), the GP-predicted
Gaussian Processes’ (Ref. 22) can be used on the training
responses are an improvement on a physics-based model
data to automatically select points which are more ‘informa-
developed for real-time operation. The GPs also produce, to
tion rich’. A useful property of GPs is that they can determine
some degree, ‘better’ predictions for the off-axis responses
the relative relevance of the input data; this information can be
of roll and yaw rate using training data based only upon
used to analyse whether or not the correct NARX structure has
longitudinal axis inputs. The results shown in the paper
been used. This property could be utilized to establish which
provide an excellent basis for the prediction of model error in
other helicopter inputs to the NARX structure should be in-
future work.
cluded to improve the accuracy of the predicted response.

The inclusion of the hyperparameter uncertainty into the full ACKNOWLEDGEMENTS

model predictions did not add any significant uncertainty to
the model predictions. For future use, it should not be neces- The authors gratefully acknowledge Deutsches Zentrum für
sary to include this in the model. Luft- und Raumfahrt e.V. (DLR) Institut für Flugsystemtech-
10
nik for the provision of the flight test data and the permission 16. Kumar, MV, Omkar, S, Ganguli, R, Sampath, P, and
to use it for this study. Suresh, S. Identification of helicopter dynamics using
recurrent neural networks and flight data. Journal of the
REFERENCES American Helicopter Society 2006;51:164–174.
1. Shy, KS, Hageman, JJ, and Le, JH. The Role of Aircraft 17. Omkar, S, Mudigere, D, Senthilnath, J, and Kumar, MV.
Simulation in Improving Flight Safety Through Control Identification of Helicopter Dynamics based on Flight
Training. National Aeronautics and Space Administra- Data using Nature Inspired Techniques. arXiv preprint
tion, Dryden Flight Research Center, 2002. arXiv:1411.3251 2014.
2. Bell, HH and Waag, WL. Evaluating the effectiveness 18. Girard, A. Approximate methods for propagation of un-
of flight simulators for training combat skills: A re- certainty with Gaussian process models. PhD thesis.
view. The international journal of aviation psychology University of Glasgow, 2004.
1998;8:223–242. 19. Green, P and Worden, K. Bayesian and Markov Chain
3. Padfield, GD. Helicopter flight dynamics. John Wiley & Monte Carlo methods for identifying nonlinear systems
Sons, 2008. in the presence of uncertainty. Phil. Trans. R. Soc. A
4. Punjani, A and Abbeel, P. Deep learning heli- 2015;373:20140405.
copter dynamics models. In: Robotics and Automation 20. Green, P. Bayesian system identification of a nonlin-
(ICRA), 2015 IEEE International Conference on. IEEE. ear dynamical system using a novel variant of simulated
2015:3223–3230. annealing. Mechanical Systems and Signal Processing
5. Anon. JAR-FSTD H, Helicopter Flight Simulation 2015;52:133–146.
Training Devices. May, 2008. 21. Saul, AD, Hensman, J, Vehtari, A, and Lawrence, ND.
6. Anon. FAA Advisory Circular AC120-63 Helicopter Chained gaussian processes. In: Artificial Intelligence
Simulator Qualification. November, 1994. and Statistics. 2016:1431–1440.
7. Perfect, P, Timson, E, White, MD, Padfield, GD, Erdos, 22. Titsias, MK. Variational Model Selection for Sparse
R, and Gubbels, AW. A rating scale for the subjective as- Gaussian Process Regression.
sessment of simulation fidelity. The Aeronautical Jour-
nal 2014;118:953–974.
8. Kennedy, MC and O’Hagan, A. Bayesian calibration of
computer models. Journal of the Royal Statistical Soci-
ety. Series B, Statistical Methodology 2001:425–464.
9. Anon. FLIGHTLAB Development Software, Advanced
Rotorcraft Tehcnology Inc, Sunnydale, California.
http : / / www . flightlab . com / flightlab .
html.
10. Bishop, C. Pattern recognition and machine learning.
springer, 2006.
11. Rasmussen, CE and Williams, C. Gaussian processes for
machine learning. 2006. Cited on 2006:1–31.
12. Higdon, D, Gattiker, J, Williams, B, and Rightley,
M. Computer model calibration using high-dimensional
output. Journal of the American Statistical Association
2008;103.
13. Worden, K, Manson, G, and Cross, E. On Gaussian Pro-
cess NARX Models and Their Higher-Order Frequency
Response Functions. In: Solving Computationally Ex-
pensive Engineering Problems. Springer, 2014:315–
335.
14. Worden, K, Surace, C, and Becker, W. Uncertainty
Bounds on Higher-Order FRFs from Gaussian Process
NARX Models. Procedia Engineering 2017;199:1994–
2000.
15. Manso, S. Simulation and system identification of he-
licopter dynamics using support vector regression. The
Aeronautical Journal 2015;119:1541–1560.
11