You are on page 1of 12

IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 6, NO.

5, SEPTEMBER 1998 651

Brief Papers
Frequency-Domain Identification of Gas Turbine Dynamics
Ceri Evans, David Rees, and Dave Hill

Abstract— The identification of the fuel flow to shaft speed It is therefore common practice to linearize the thermody-
dynamics of a twin-shaft gas turbine is addressed, with the aim namic models around a series of operating points and then
of validating thermodynamic engine models. A measured input carry out a model-order reduction in order to arrive at models
signal must be used in estimation in order to exclude the fuel
feed dynamics from the model. This has been shown to present which are suitable for control system design. Since these
problems when fitting discrete models to engine data, and this models are based on a priori assumptions about the engine
paper examines the direct estimation of s-domain models in the physics, it is then important to validate their performance
frequency domain. A number of different multisine test signals against real engine data.
were applied to the engine for the purposes of model estimation This paper deals with the linear identification of the fuel
and nonlinear detection. The use of frequency-domain techniques
is shown to produce high-quality models, and the tests also feed to shaft speed dynamics of a Rolls Royce Spey Mk
yield information on the levels of noise and nonlinearity and the 202 engine at a single operating point. This is a typical
length of the pure time delay. This work illustrates the potential military twin-shaft turbofan, with a low bypass ratio and a
of frequency-domain techniques for modeling systems where a variable reheat nozzle. Although no longer in service, it has
physical interpretation is to be made of the model and where the the same basic architecture, for control purposes, as that of
need for accuracy requires that a measured input signal be used
in estimation. more modern engines such as the EJ200, which is used to
power the Eurofighter [3].
Index Terms—Frequency domain, gas turbines, multisine test A simplified diagram of a Spey engine is shown in Fig. 1,
signals, system identification, system testing.
where concentric shafts are seen to connect the compressors
at the intake end to the turbines at the nozzle end. Modern
I. INTRODUCTION gas turbines usually have two shafts, one connecting the
high-pressure (HP) compressor to the HP turbine and the
G AS TURBINE engines were originally designed for
aircraft propulsion, but are now extensively used in aero,
marine, and industrial applications. There is also considerable
other connecting the low-pressure (LP) compressor to the LP
turbine. These rotate at different speeds, which are denoted
crossfertilization between steam and gas turbine technolo- and . The shaft speeds are the primary outputs of a
gies, such as the collaboration between Westinghouse and gas turbine from which the internal engine pressures and the
Rolls Royce, resulting in the development of small bespoke thrust can be calculated.
power stations based on aeroengines [1]. Gas turbines are The aim of this study was to estimate models which could be
even finding application in a new generation of electric cars, used to validate existing thermodynamic models derived from
currently being tested by Volvo [2]. With such widespread the engine physics. The poles and zeros of the estimated model
and increasing applications, the modeling of these engines is are thus to be given a physical interpretation and directly
an issue of some importance. related to the -domain poles of the linearized thermodynamic
Engine models are required both in the development and models.
operational stages of the life of a gas turbine. Thermodynamic
models are derived during the development stage, based on II. BACKGROUND
knowledge of the engine physics, and provide important At the initial stages of an engine design, a thermodynamic
insights into the engine behavior. Such models are both model is derived from the nonlinear differential equations
complex and nonlinear, making them unsuitable for use in which govern energy transfer within the engine. The nonlinear
the design of engine control systems. model is then numerically linearized about a set of operating
points, using small perturbations and evaluating the partial
derivatives. In the case of the Spey, this results in multivariable
Manuscript received March 1, 1997; revised February 1, 1998. This work models with five inputs, ten outputs and fifteen states.
was supported by Rolls Royce plc and the U.K. Defense Evaluation and
Research Agency at Pyestock. All but two of these states are associated with rates of
C. Evans and D. Rees are with the School of Electronics, University of change of pressure within the compressor and turbine volumes.
Glamorgan, Pontypridd, CF37 1DL, Wales, U.K. These states have very fast time constants, in the order of
D. Hill is with Rolls Royce plc, Advanced Controls Division, Derby, DE24
8BJ, U.K. 20 ms or less, and can be eliminated by a model reduction
Publisher Item Identifier S 1063-6536(98)06133-8. process. The remaining two states are the shaft speeds and

1063–6536/98$10.00  1998 IEEE


652 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 6, NO. 5, SEPTEMBER 1998

Fig. 1. Simplified schematic of a Rolls Royce Spey engine.

transfer function models can be obtained for each of the


shafts by evaluating the transfer function matrix. These models
have one zero and two poles, with a pole-zero pair modeling
the interaction between the shafts. There is a near pole-zero
cancellation in the HP shaft model.
Initial work on validating small signal engine models was
carried out in the 1950’s on single shaft and twin shaft engines Fig. 2. Engine measurement setup.
[4], [5]. Single sine and step inputs were used to excite an
engine and -domain models calculated. The main problems
were the length of time required for the single sine tests and techniques were reported by Hill in a recent paper [8]. Un-
the lack of a systematic way of fitting models to the data. This der certain circumstances, discrete models can be estimated
made it difficult to validate the one zero, two pole models of which have a continuous -domain equivalent which cannot
twin shaft engines, since the close pole-zero pairs could not be realized by a physical system. This is due to the presence
be estimated accurately. of a single real negative -domain pole, which maps to a single
This problem was later addressed by Cottington and Pease complex pole in the -domain.
[6], who presented the results of testing a number of engines Such models clearly cannot be used to validate the linearized
thermodynamic models. Evans [9] has shown that this effect
with maximum length binary sequences (MLBS’s). The am-
can be caused by using a measured bandlimited input signal in
plitude of the perturbation signals was restricted to produce
the estimation of discrete models, or by the presence of a pure
output speed variations of no more than 2% of the maximum
time delay which is not an integer multiple of sampling period.
in an effort to ensure a linear response and minimize
If a physical interpretation is to be made of a -domain
engine wear. Cottington and Pease stressed the importance
model, the input signal should be applied under zero-order-
of excluding the dynamics of the fuel feed system from the
hold (ZOH) conditions [10]. In this application, the measured
estimated model by using the measured fuel flow as the input
value of the fuel flow is used for estimation in order to
signal.
eliminate the fuel feed dynamics from the model. The fuel
The development of increased computing power, along with
feed bandlimits the input signal and, hence, violates the ZOH
advances in system identification techniques, now make it
condition.
possible to apply a systematic approach to fitting parametric
models to engine data. Work conducted by Hill [7] exam-
ined the application of a range of time-domain methods to III. THE MEASUREMENT PROBLEM
estimating discrete -domain engine models, with the aim of There are a number of problems associated with the mea-
reducing engine test times, while improving the accuracy of surement of the fuel flow and shaft speeds, which will be
the estimated models and quantifying the model uncertainty. discussed in turn. Some of these are inherent to gas turbines,
Particular importance was placed on the identification of the others are a feature of the measurement transducers. A sim-
second-order effects, since they had not been adequately dealt plified block diagram of the measurement setup is shown in
with in past work. Fig. 2.
This work showed that discrete models with good in- 1) Fuel Feed System: The fuel feed system is composed of
put–output properties could be estimated, making them suit- a valve driven by a stepper motor, which exhibit both linear
able for engine simulation and control system design. How- and nonlinear dynamics and affect the actual fuel flow applied
ever, some problems with the application of time-domain to the gas turbine. It is important to eliminate these effects
IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 6, NO. 5, SEPTEMBER 1998 653

(a)

(b) (c)
Fig. 3. Circular covariance of input signal. (a) Across complete data length, (b) first lag, and (c) fifth lag.

from the estimated engine model, and this is achieved by input noise cannot be considered as simply measurement noise,
measuring the actual fuel flow downstream of the fuel feed, since the turbulence in the input fuel flow also influences the
using a turbine flow meter. shaft speed outputs.
2) Speed Measurement: The speed of the low pressure 5) Nonlinearities: The dynamics of a gas turbine vary
shaft is measured by counting the rotations of the turbine considerably within its operating range. A series of linear
blades and the speed of the HP shaft by measuring the rotation models are thus estimated around a set of operating points. The
of a gear linked to the shaft itself. A series of pulses are nonlinear behavior of the fuel feed system of a Pegasus engine
generated, the frequency of which depends on the shaft speed. was studied by Godfrey and Moore [11], who proposed the use
The number of pulses in a given interval are counted and the of an inverse repeat binary sequence to reduce the influence
resulting speed provided as the transducer output. of the nonlinearity on the estimated model. It was thus of
The engine control computer then polls this output at each interest to see if significant nonlinear effects were encountered
sampling instant. It should be clear that it is not possible to in these small signal tests, both with the fuel feed system and
incorporate any antialiasing (AA) filters into this setup, since the engine itself.
the conversion is from a pulse frequency to a digital value (F- 6) Combustion Delay: There is a pure time delay in the
D) and no true analogue signal exists before digitizing. This system, which is due to the delay involved in vaporizing and
also applies to the turbine flow meter on the input. igniting the fuel in the engine. Previous tests on this engine
estimated the delay at 35 ms or less, using a cross-correlation
3) Engine Control Computer: The number of samples
method. A semiempirical model of the combustion dynamics
which can be logged in one test is limited to 15 000, and
was proposed by Bauerfeind [12] and later Thomson [13],
for this reason only the demanded fuel flow, measured fuel
which predicted time delays in the order of 15 ms for a typical
flow, and the two shaft speeds were recorded. Recording
engine.
the demanded fuel flow allows the generation of the test
signals to be checked and also allows the fuel feed dynamics
IV. FREQUENCY-DOMAIN ESTIMATION
to be studied. The digital-to-analog converter (DAC) at the
computer output and the samplers at its input can be clocked The frequency-domain techniques employed in this study
at different speeds, allowing signal reconstruction and data and the test signals used will now be described.
sampling at different rates.
4) Noise: The dominant noise sources are the turbulence A. Estimation
in the fuel flow at the input and the vibration of the turbine Since periodic input signals were used the frequency re-
blades and jitter in the gears at the outputs. This means that the sponse function (FRF) was estimated using ratio of the mean
654 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 6, NO. 5, SEPTEMBER 1998

TABLE I
SIGNAL-TO-NOISE RATIOS (dB)

output and input Fourier coefficients Frequency-domain identification has a number of clear
advantages if a physical interpretation of the model is to be
made, since an -domain model can be directly identified,
whose poles and zeros can be related to physical parameters of
(1) the system. The estimator is now available as part of a Matlab
toolbox, written by Kollár [16].

B. Test Signals
where and are the input and output spectra
at the test frequencies measured across periods. This Frequency-domain estimation is greatly simplified if peri-
was termed the estimator by Guillaume [14], who odic input signals are used and the measurements are per-
showed that it gives an asymtotically unbiased estimate of the formed under dynamic steady state conditions. The measured
FRF, if the measurements are synchronized, in the presence signals must be bandlimited and sampled at more than twice
of normally distributed noise on the real and imaginary parts the highest signal frequency in order to prevent aliasing. The
of the Fourier coefficients. lack of AA filters in the turbine measurement setup might thus
Parametric frequency-domain identification involves select- seem to preclude the use of frequency-domain techniques.
ing the parameters of an -domain model, with pure time delay This would be true if binary signals were used, since they
do not have bandlimited spectra. This is not the case if
multisine signals are used, since the total signal power can
be concentrated in the frequency range of interest. The only
higher frequency components in the signals would be those
(2) generated at the signal reconstruction frequency by the DAC.
It is, however, possible to utilize the low-pass characteristic
Employing an error-in-variables model, Pintelon et al. [15] of the fuel feed system to filter out these components, if a suf-
derived a maximum likelihood estimator of , with the ficiently high-reconstruction frequency is used. The multisines
cost function as shown in (3), given at the bottom of the page. can thus be strictly bandlimited without the use of AA filters
The variances of the input and output noises are denoted and the only aliasing present will be due to high-frequency
by and , and is their covariance. The noise.
summation is across the excitation frequencies, and is a A multisines signal is an arbitrary ensemble of harmonically
vector of the model parameters. related cosines
Double the value of is distributed with
degrees of freedom, where is the number of free parameters.
(5)
Consequently, if there are no modeling errors, the cost function
should reach a value of
where is a vector of amplitudes, a vector of harmonic
(4) numbers, the fundamental frequency, and a vector
of phases. The relative phases of the harmonics must be
The estimates are asymptotically unbiased, efficient, and carefully selected to minimize the crest factor (CF) of the
robust to various time-domain disturbances. The variances of signal. By minimizing the CF, it is possible to ensure that
the input and output noise, and their covariance, are required as the maximum possible power is injected for a given input
a priori information, but these can be estimated while making amplitude constraint, which improves the signal-to-noise ratio
the measurements if the Fourier coefficients are averaged. (SNR) and thus shortens measurement times
The nonparametric noise model is particularly attractive in
applications such as the gas turbine, since the noise is known CF (6)
to be colored, with a number of peaks in its spectrum. rms

(3)
IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 6, NO. 5, SEPTEMBER 1998 655

The approach proposed by Guillaume et al. [17] was


employed for the test signals, as it generates the lowest CF
for band limited multisines.
A great advantage of multisine signals is the flexibility
afforded in the signal design. The multisine should be designed
to concentrate the signal power within the system bandwidth,
with sufficient harmonics to allow CF minimization, give the
required spectral coverage, and allow for detection of modeling
errors. Including too many harmonics will however reduce
the power at each test frequency and consequently affect the
accuracy of the FRF, which is to be used for model validation.
It is also possible to omit certain harmonics from the mul-
tisine signal. If all the even harmonics are omitted, the signal
will be immune to even-order nonlinear distortions, since
(a)
even sums of odd harmonics will always result in nonlinear
contributions at even harmonics. If some of the odd harmonics
are also omitted it is possible to detect nonlinear effects by the
presence of output power at the omitted harmonics within the
signal bandwidth. Output power at the even harmonics will
indicate the presence of even-order nonlinearities and output
power at the omitted odd harmonics will indicate odd-order
effects [9].
Three signal designs were employed in these tests, with an
equal fundamental frequency of 0.01 Hz and roughly equal
bandwidths. The first being a signal composed of 30 odd
harmonics, termed an odd multisine
Signal 1 (7)
(b) A second signal was employed for cross-validation pur-
poses, composed of 15 odd harmonics with the same overall
bandwidth, but with its power concentrated at low frequencies.
This will be termed a quasi-log multisine since it has a
spectrum which is similar in appearance to that of a log-tone
signal
Signal 2 (8)
The third signal was designed with all the even harmonics
omitted, along with every other odd harmonic, and termed
an odd–odd multisine. This allowed the degree of nonlinear
distortion present within the linearization range of the model
to be assessed
Signal 3 (9)
A criterion is required to assess whether any of the ad-
ditional output components are indeed periodic nonlinear
contributions, or simply noise harmonics. McCormack et al.
[18] proposed using the squared coherence function, defined
for systems with a noise free input, which depends only on
the output signal

(10)

(c)
Fig. 4. Noise variances of (a) input, (b) output, and (c) absolute value of where is the output spectrum at both the excited and
covariance for HP shaft. omitted harmonics measured across periods. The quantity
656 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 6, NO. 5, SEPTEMBER 1998

(a) (b)
Fig. 5. Odd–odd multisine spectral amplitude at (a) HP and (b) LP shaft outputs.

will be termed the nonlinear coherence to distinguish where is the number of samples, the sampling period,
it from the more commonly used definition of coherence. As and corresponds to a circular shift of samples to
becomes large, the nonlinear coherence will express the the right. This can be calculated via the inverse fast Fourier
ratio of the periodic power over the total power at each output transform (FFT) of the signal power spectrum, with the dc
frequency. If the periodic component is zero, it will assume component suppressed [19]. The circular covariance of the 30
a value of , which gives an useful bound with which to odd-harmonic input signal is plotted in Fig. 3. The periodicity
judge the significance of the nonlinear coherence values [9]. of the input is clear, and the detailed plots of the first and
fifth lags indicate that the synchronization is very good. The
maximum deviation between the clocks must be less than
V. DATA ANALYSIS
0.025% (0.2/2 in 4 500 0.2).
Tests were conducted at an operating point of 75% of
the maximum HP shaft speed and the signals imposed with
amplitudes of 10% of the steady-state fuel flow. Six periods B. Noise Analysis
of each signal were measured with a sampling period of 0.2 s, The SNR’s were also calculated, beginning with an estimate
generating 500 samples per period. A detailed analysis was of the raw SNR from the ratio of the total power at the excited
made of the gathered data before proceeding with the para- harmonics over the total signal power. Estimates of the noise
metric model estimation. variances and covariance were then made during the averaging
of the Fourier coefficients. These allowed the improvements in
A. Synchronization SNR obtained by excluding the noise lines from the data, and
by averaging across six periods, to be calculated. The results
It is important that the signal reconstruction and sampling
for the 30 odd-harmonic signal are listed in Table I, where it
clocks are synchronized when conducting measurements for
can be seen that the SNR’s after averaging are very good, with
frequency-domain identification, since a deviation in either
values of at least 40 dB in each case.
clock will result in incomplete signal periods being sam-
The noise variances and covariance are plotted in Fig. 4
pled. The quality of the synchronization can be checked by
for the 30 odd-harmonic signal, HP shaft data, which shows
examining the circular covariance of the measured signals
stronger noise components at low frequency and also large
covariance at these frequencies. The large covariance is to
be expected since the dominant noise source at the input is
process noise, which also passes through the system and is
(11) hence correlated with the output noise.
IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 6, NO. 5, SEPTEMBER 1998 657

(a) (b)
Fig. 6. Nonlinear coherence of odd–odd multisine at (a) input and (b) HP shaft output. Input harmonics (solid), omitted even harmonics (dash–dot),
and omitted odd (dotted). With 1=M bound shown dashed.

C. Nonlinearities of the HP and LP shafts were estimated using the


The influence of nonlinearities was investigated by exam- estimator defined in (1) and their amplitudes are shown in
ining the FFT of the odd–odd signal at the HP and LP shaft Fig. 7. It can be seen that the bandwidth of the signals was well
outputs, shown in Fig. 5. The presence of colored noise is chosen, covering a drop of 17 dB on the HP shaft and 13 dB
clearly seen, but it is difficult to see if there are any additional on the LP shaft. The corresponding phase shifts vary between
periodic components rising above the noise floor, which would 10 and 85 . The results of the parametric estimation for
indicate the presence of nonlinear effects. each of the shafts will now be discussed in turn.
A clearer picture is obtained from the nonlinear coherence,
which is plotted in Fig. 6 for the input and the HP shaft A. High-Pressure Shaft
output, along with the bound. It is seen that only one The variation of the cost function with model order is shown
omitted frequency within the signal bandwidth has a significant in Table II. The theoretical minimum of the cost function is
coherence in the input spectrum, while a series of omitted also given, as this varies with the number of parameters, in
even harmonics have a significant coherence in the output. accordance with (4). The major drop in the cost function occurs
The coherence of the omitted odd harmonics is consistently with the one-zero two-pole (1/2) model, suggesting that this
lower on both the input and output. is the best structure.
This suggests that the fuel feed can be considered as linear Table III shows the pole and zero positions of the first three
over this input range, while a weak even-order nonlinearity of these models. The 1/2 model has a pole-zero pair close
may be present in the gas turbine. A similar result was obtained to the origin, whose regions of uncertainty do overlap.
with the LP shaft. The influence of this even-order nonlinearity Adding another pole and zero results in an unstable model
on the estimated linear models is eliminated simply by using being estimated. The frequency responses of the 0/1 and 1/2
odd harmonic signals. A detailed analysis of the gas turbine models are shown in Figs. 8 and 9, where the additional pole
nonlinearities has been presented in a previous paper by Evans and zero are seen to be modeling a low-frequency effect.
et al. [20].

B. Low-Pressure Shaft
VI. ESTIMATION RESULTS The variation of the cost function with model order for
Frequency-domain estimation was carried out using the the LP shaft is presented in Table IV. The large drop in the
Fourier coefficients of the 30 odd-harmonic signal. The FRF’s cost function between the 0/1 and 1/2 models shows that the
658 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 6, NO. 5, SEPTEMBER 1998

(a) (b)
Fig. 7. Amplitude of FRF for (a) HP and (b) LP shafts.

TABLE II
COST FUNCTION FOR HP SHAFT MODELS, 30 ODD HARMONICS

TABLE III
HP SHAFT MODELS, 30 ODD HARMONICS

dynamics are at least second order. There is even a case for pole and zero do however overlap. This close pole-zero pair is
selecting a 2/3 model, which conflicts with the structure of the once again modeling a low frequency effect in a similar way
derived thermodynamic model. to the 1/2 model for the HP shaft. The difference between the
Table V shows that the 2/3 model has a very close pole- frequency response of the 1/2 and 2/3 models is, however, so
zero pair near to the origin. The uncertainty regions of the small as to be almost indistinguishable in Fig. 10.
IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 6, NO. 5, SEPTEMBER 1998 659

(a) (b)
Fig. 8. Frequency response of HP shaft model (solid) and FRF (crosses): for 0/1 model.

(a) (b)
Fig. 9. Frequency response of HP shaft model (solid) and FRF (crosses): for 1/2 model.
660 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 6, NO. 5, SEPTEMBER 1998

TABLE IV
COST FUNCTION FOR LP SHAFT MODELS, 30 ODD HARMONICS

TABLE V
LP SHAFT MODELS, 30 ODD HARMONICS

TABLE VI
LP SHAFT MODELS, 15 QUASI-LOG HARMONICS

C. Cross Validation in the time-domain estimation of discrete models, since faster


Cross validation was performed using the quasi-log mul- sampling is required in order to exactly specify the delay. This
tisine and the estimated models confirmed the above results. can cause numerical problems, as the poles and zeros of the
The LP shaft models are shown in Table VI and can be seen discrete models will tend to cluster around the point .
to match closely those shown in Table V. A similar pattern
was found with the HP shaft. VII. COMPARISON WITH THERMODYNAMIC MODELS
Models of the HP and LP shaft have been estimated with
D. Combustion Delay a high degree of confidence from the test data and it is now
The pure time delay was included as an estimation parame- possible to compare these estimated models with the linearized
ter and Tables III, V, and VI show that values of between 13 thermodynamic models derived from the engine physics.
and 21 ms were obtained for the selected models on each shaft, In Table VII, the poles and zero of the HP shaft model
which match the typical values of combustion delay previously estimated with the 30 odd-harmonic data are compared to
proposed [12], [13]. This small delay can present a problem those of the thermodynamic model. Each of the models has a
IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 6, NO. 5, SEPTEMBER 1998 661

TABLE VII
COMPARISON OF MODELS, HP SHAFT

TABLE VIII
COMPARISON OF MODELS, LP SHAFT

dominant pole close to the same point (around 0.1 Hz) and, such as blade expansion. Blade expansion time constants of 5 s
in addition, a close pole-zero pair, which indicates that the were quoted by Hennecke and Trappman [21] and blade and
HP shaft dynamics are predominantly first order. The major casing time constants of 2 and 10 s, respectively, by Pilidis
difference is that the pole-zero pair of the estimated model is and Maccallum [22]. The estimated low-frequency modes have
much closer to the origin. This was shown to model a real time constants of between 3 and 10 s.
feature of the data, in Fig. 9, but it is clearly modeling a
different effect to that of the higher frequency pole-zero pair VIII. CONCLUSIONS
in the thermodynamic model. The frequency-domain identification of the fuel flow to shaft
For the LP shaft, the 1/2 thermodynamic model is compared speed dynamics of a twin-shaft gas turbine has been studied,
to the 2/3 model estimated using the 30 odd-harmonic data with the aim of validating thermodynamic engine models. This
in Table VIII. The separation of the zero and poles of the work was motivated by the problems previously encountered
thermodynamic model is greater than with the HP shaft, which when fitting discrete-time models to turbine data. These prob-
suggests that the dynamics are at least second order. lems are caused by the need to use the measured fuel flow
Neglecting for a moment the close pole-zero pair in the in the estimation in order to exclude the fuel feed dynamics.
estimated model, a direct comparison can be made between This can result in time-domain models with no continuous
the models. The first pole of the estimated model is at a counterparts, thus precluding the physical interpretation of
lower frequency than that of the thermodynamic model, at their parameters.
a frequency of 0.076 Hz compared to 0.11 Hz. The zero and The results presented in this paper clearly show that one
second pole are also at a lower frequency, which suggests that solution is to use frequency-domain techniques, employing
the LP shaft response is actually different to that predicted by multisines as test signals. Such signals can be designed to
the thermodynamic model. concentrate their whole power within the system bandwidth.
Comparing the estimated models of the HP and LP shafts A sufficiently high-reconstruction frequency can then be used,
it is seen that they share a close, but not identical, dominant so that the fuel feed system filters out the reconstruction
mode at a frequency of around 0.09 Hz. The results suggest components. This overcomes the problems caused by the lack
that the HP shaft dynamics are first order and the LP shaft of AA filters in the measurement setup. The flexibility of
dynamics second order. The additional pole-zero pairs are multisine signals was illustrated by the three signal designs
too slow to be associated with these shaft dynamics and are applied to the gas turbine.
probably modeling a heat soakage effect. The tests revealed that the influence of nonlinearities was
Such effects become prominent during large transient ma- smaller than expected across the linearization range of the
neuvers of an engine, and while the tests were carried out model. They also revealed that the noise on the input and
around a fixed operating point, there may have been sufficient output signals was highly correlated at low frequencies. High-
variation of the engine operation during the tests to excite some quality models of each shaft were estimated and their fit to
heat transfer dynamics. It is possible that the low-frequency the estimated frequency response was illustrated. The pure
modes are modeling one of the faster heat-soakage effects, time delay was included as a parameter for estimation and
662 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 6, NO. 5, SEPTEMBER 1998

interpretation of the model parameters is to be made. This is


particularly true if the input signal has to be measured, due to
the adverse effects of the input actuator dynamics.

ACKNOWLEDGMENT
The authors would like to thank the staff at Rolls Royce
plc and the U.K. Defense Evaluation and Research Agency at
Pyestock.

REFERENCES
[1] “Aero power comes down to earth,” IEE Review, p. 2, Jan. 1994.
(a) [2] D. Vernet, “Electric cars and the quality of life,” Electro Technol., vol.
5, no. 2, pp. 8–13, 1994.
[3] G. J. Dadd, A. E. Sutton, and A. W. M. Greig, “Multivariable control
of military engines,” in AGARD Conf. Proc. no. 572—Advanced Aero-
Engine Concepts and Controls, 1996, paper 28.
[4] J. O. N. Lawrence and R. D. Powell, “The application of servo-
mechanism analysis to fuel control problems,” in Proc. Instit. Mech.
Eng., vol. 172, pp. 439–469, 1957.
[5] J. W. Fitchie, S. W. White, R. H. Cronshaw, and H. J. Willcocks, “An
experimental and theoretical study of the response of a two shaft turbojet
subjected to small disturbances,” National Gas Turbine Establishment
(U.K.) Note NT 420, Oct. 1959.
[6] R. V. Cottington and C. B. Pease, “Dynamic response testing of gas
turbines,” ASME J. Eng. Power, vol. 101, pp. 95–100, 1979.
[7] D. C. Hill, “System identification of gas turbine engines,” Ph.D. disser-
tation, Birmingham Univ., Birmingham, U.K., 1994.
[8] , “Identification of gas turbine dynamics: Time-domain estimation
problems,” in ASME Gas Turbine Conf., 1997, paper 97–GT-31.
(b) [9] C. Evans, “Identification of linear and nonlinear systems using multi-
sine signals, with a gas turbine application,” Ph.D. dissertation, Univ.
Glamorgan, Pontypridd, U.K., 1998.
[10] J. Schoukens, R. Pintelon, and H. Van Hamme, “Identification of linear
dynamic systems using piecewise constant excitations: Use, misuse and
alternatives,” Automatica, vol. 30, no. 7, 1994.
[11] K. R. Godfrey and D. J. Moore, “Identification of processes having
direction dependent responses, with gas-turbine engine applications,”
Automatica, vol. 10, pp. 469–481, 1974.
[12] K. Bauerfeind, “A new method for the determination of transient
jet engine performance based on the nonstationary characteristics of
the components,” in AGARD Conf. Proc., 1968, no. 34, Advanced
Component for Turbojet Engines, pt. 2, paper 32.
[13] B. Thomson, “Basic transient effects of aero gas turbines,” in AGARD
Conf. Proc. No. 151—Power Plant Controls for Aero-Gas Turbine En-
gines, 1974, paper 2.
[14] P. Guillaume, “Identification of multi-input multi-output systems using
frequency-domain models,” Ph.D. dissertation, Univ. Brussel, Brussels,
Belgium, 1992.
(c) [15] R. Pintelon, P. Guillaume, Y. Rolain, and F. Verbeyst, “Identification
Fig. 10. Amplitude response of LP shaft model (solid) and FRF (crosses): of linear systems captured in a feedback loop,” IEEE Trans. Instrum.
for model orders (a) 0/1, (b) 1/2, and (c) 2/3. Meas., vol. 41, no. 6, pp. 747–754, 1992.
[16] I. Kollár, Frequency Domain System Identification Toolbox for Use with
Matlab. MA: Mathworks, 1994.
the results were close to those obtained on other engines of [17] P. Guillaume, J. Schoukens, R. Pintelon, and I. Kollár, “Crest factor min-
imization using nonlinear Chebyshev approximation methods,” IEEE
a similar type. Trans. Instrum. Meas., vol. 40, no. 6, pp. 982–989, 1991.
The estimated models show the HP shaft dynamics to be [18] A. S. McCormack, K. R. Godfrey, and J. O. Flower, “The design of
predominantly first order and those of the LP shaft to be second multilevel multiharmonic signals for system identification,” in IEE Proc.
Control Theory and Applications, 1995, vol. 142, pp. 247–252.
order. However, low-frequency effects were detected on both [19] I. Kollár, R. Pintelon, and J. Schoukens, “Frequency domain system
shafts which could be modeled by the addition of a further identification toolbox for Matlab: A complex application example,” in
Preprints 10th IFAC Symp. System Identification, Denmark, 1994, vol.
pole-zero pair. The time constants of these additional pole- 4, pp. 23–28.
zero pairs suggest that they may be modeling thermal effects, [20] C. Evans, D. Rees, and L. Jones, “Identifying linear models of systems
which were not incorporated in the original thermodynamic suffering nonlinear distortions, with a gas turbine application,” in IEE
Proc. Control Theory and Applications, 1995, vol. 142, pp. 229–240.
models. Further work needs to be conducted in order to study [21] D. K. Hennecke and K. Trappmann, “Turbine tip clearance control in
these effects in more detail, using test signals with a lower gas turbine engines,” in AGARD Conf. Proc. no. 324—Engine Handling,
fundamental frequency. 1982, paper 16.
[22] P. Pilidis and N. R. L. Maccallum, “Models for predicting tip clearance
This work illustrates the potential of frequency-domain changes in gas turbines,” in AGARD Conf. Proc. no. 324—Engine
techniques for modeling industrial systems, where a physical Handling, 1982, paper 17.

You might also like