2004-01-1418

Instantaneous Exhaust Temperature Measurements Using

Thermocouple Compensation Techniques
Kenneth Kar, Stephen Roberts, Richard Stone, Martin Oldfield
University of Oxford

Boyd French
Ford Motor Company
Copyright 2003 SAE International

ABSTRACT
This paper discusses a method of measuring the
instantaneous
exhaust
gas
temperature
by
thermocouples. Measuring the exhaust gas temperature
is useful for a better understanding of engine processes.
Thermocouples do not measure the instantaneous
exhaust gas temperature because of their limited
dynamic response. A thermocouple compensation
technique has been developed to estimate the time
constant in situ. This method has been commissioned in
a simulation study and a controlled experiment with a
reference temperature. The studies have shown that the
signal bandwidth has to be restricted, since noise will be
amplified in the temperature reconstruction. The
technique has been successfully applied to some engine
exhaust measurements. A comparison between two
independent pairs of thermocouples has shown that
temperature variations at frequencies up to 80Hz can be
recovered. The medium load results agree with a
previous study, which used fast response thermometers
with a bandwidth of about 50 Hz. However, the results at
low load and two different speeds have highlighted the
need to do some 1-D unsteady flow simulations, in order
to gain more insight into the exhaust process.

INTRODUCTION
With the introduction of even more stringent emission
limits (EURO IV, SULEV), engine developers are striving
to better understand engine processes. Exhaust gas
temperature (EGT) can be very useful in this respect. For
instance, the heat transfer across the exhaust port can
be estimated with EGT measurements, and this would
enable a better heat transfer model to be developed. In
addition, EGT forms the basic calibration data for
computational fluid dynamic models (CFD), which are
commonly used in engine design.
Knowing the exhaust gas temperature can immediately
help many other applications as well. Following the
engine cold start, the catalytic converter has not yet
reached its operating temperature, so most of the

hydrocarbon emissions are untreated. A study has

suggested that more than 80% of the hydrocarbon
emission measured over the entire FTP-75 drive cycle
are emitted within the first 20 seconds after the engine
has started [1]. The performance of the catalyst can be
optimized if the EGT is known. This should have an
immediate improvement on hydrocarbon emissions.
Performance wise, using EGT as the control input, there
is a scope for active exhaust control as the propagation
velocity of the pressure waves is closely related to EGT
[2]. Lastly, the performance of a turbocharger may be
enhanced since its operating point can be accurately
determined with instantaneous exhaust temperature
measurements, especially when it is subject to a
fluctuating exhaust flow.
There are a range of thermometers available to measure
exhaust gas temperature. The most commonly used is
the thermocouple because of its low cost, simplicity and
the wealth of experience in their application. However,
because of their slow response (typically 0.1 Hz to 5 Hz),
they are only used for measuring time-averaged
temperatures. Ideally the instantaneous temperature is
needed, so that transient phenomenon can be observed.
Other thermometers have been applied in engine
diagnostics such as the acoustic technique [3], optical
(laser) diagnostics [4, 5] and resistance thermometry
[6, 7]. Their responses certainly meet the requirements
and may exceed them in many cases. However, they
tend to be complicated, difficult to set up and maintain,
and so are generally expensive. Only a limited number of
people and laboratories would have access to them. It is
unlikely that such systems will become so economical
that they are as popular as some well established
thermometers, like thermocouples.
For exhaust gas temperature measurement, the
frequencies of interest are mostly below 100 Hz. In fact
this is not a very demanding situation. Opting for the
above fast response systems may not be a cost effective
solution. Instead, trying to improve the performance of
thermocouples would be desirable. Now it is possible
that the response of thermocouples may be improved by

some novel compensation techniques. In this case, the

total cost of the measuring system would not increase
much, as no additional hardware is required. All the
technology and equipment developed for thermocouples
can still be used. This technology will have much wider
applications than the existing system, and yet have an
acceptable response (~100 Hz).
The response of thermocouples along with the various
compensation techniques will be presented here. A new
method will be proposed to overcome the previous
shortcomings. This method will be validated by
simulation and a controlled experiment. Finally, the
application of this method is demonstrated by showing
exhaust temperature measurements from a sparkignition engine.

DYNAMIC RESPONSES
By performing an energy balance, the dynamic response
of a thermocouple can be described by Eq. 1.

Tg Tw =

w c w d Tw
4h

k w d 2Tw w 4
4
(1)
+
Tw Tsurr
4h x 2
h

The above analysis has assumed that the temperature

distribution of any cross section of the thermocouple wire
is uniform at any instant. In this paper, the Biot number
of the largest thermocouple is in the order of 3e-3, which
is much smaller than 0.1, so this assumption is justified.
Eq. 1 states that the difference between the gas
temperature (Tg) and the wire temperature (Tw) is due to
three effects, as shown by the three terms on the RHS.
The first term is the dynamic error. Because the
thermocouple wire has a finite mass, its temperature
cannot follow the changes in gas temperature
instantaneously. The second term is the conduction
error: the probe support will generally respond to the
change in gas temperature more slowly than the wire, so
a temperature gradient will exist between the wire and
the support, and heat will conduct from or to the wire.
The last term is the radiation error. The surroundings are
normally much colder than the thermocouple, so heat will
radiate from the thermocouple to the surrounding. If
conduction and radiation are negligible, Eq. 1 can be
reduced to

Tg Tw =
where,

= w c w d 4h .

d Tw
dt

(2)

Eq. 2 is a first-order ordinary

differential equation, completely characterized by . Even

in the presence of conduction and radiation, Scadron
and Warshawshy [8] have shown that the thermocouple
response still behaves like a first-order system, which
can be described by

~
~ ~
* d Tw
Tg Tw =
dt

(3)

where: the tilde denotes the fluctuation about the mean

temperature, is the ratio of equilibrium thermocouple
reading to the reading it would give if it were at the true
temperature, and * is the time constant which has the
conduction and radiation effect taken into account.

TEMPERATURE RECONSTRUCTION
RESPONSE EQUATION
In exhaust gas measurements, what is of interest is the
instantaneous gas temperature (Tg). If * and are
known, Eq. 2 or Eq. 3 can be solved for Tg; may be
calculated from a know radiation and conduction
correction factor [8]. Otherwise the thermocouple is
designed to minimize conduction and radiation errors, so
is close to unity. The Conduction and Radiation section
will show that is close to unity, so the main task here is
to find the time constant.
PREVIOUS WORK
It is very difficult to find the time constant analytically for
the following reasons:

Nusselt correlations are derived for steady flow

conditions.
Some fluid and material properties are strong
functions of temperature
The time constant is a function of radiation and
conduction parameters.

In an engine exhaust, the flow is unsteady with large

temperature variations (of the order 500 K). In addition,
chemical reactions (post oxidation of the burnt gas)
might still be taking place. Therefore, early research
concentrated on empirical methods. The time constant
can be measured by exposing the thermocouple to the
test flow. If an electric current is then passed through the
thermocouple, creating a step change in temperature,
then by observing the temperature decay, the time
constant is the time required for the junction temperature
to decrease to 63.2% of the difference between the initial
and the final temperatures. This method can only give an
average time constant for a particular flow condition. So,
the accuracy of the reconstruction will depend on how
much the flow changes. In an engine exhaust, the flow
will be almost stagnant when the exhaust valve is closed,
but it will be at very high speed during blowdown.
Therefore, this method is unsuitable.
Later developments have been with 2-thermocouple
techniques. Many techniques have been proposed, but
they are all based on the same principle. If two
thermocouples are placed in close proximity, they will be
exposed to the same temperature and velocity fields. It
follows that by utilizing this fact, with additional
assumptions, the time constants can be estimated in

situ. Denoting subscript 1 and 2 for thermocouples 1 and

2 respectively, the response equations are:

d Tw1

= Tg

dt

dT
Tg2 = Tw2 + 2 w2 = Tg

dt
Tg1 = Tw1 + 1

(4)

There are 3 unknowns with 2 equations. Cambray [9]

introduced a constant time-constant ratio () to solve
Eq. 4. If the heat transfer between the gas and the
thermocouple is dominated by convection and the
temperature dependence of the material properties of
the thermocouple is weak, it can be shown that:

d
= 2 = 2
1 d1

2 m

(5)

where: d is the diameter of the junction and m is the

coefficient of the Reynolds number in the Nusselt
correlation. Eq. 4 can then be rearranged and solved for
1 as

1 =

(T1 T2 )
d T1 /d t d T2 /d t

(6)

This method gives the instantaneous time constant, but it

can be shown to be highly susceptible to noise. Consider
Eq. 7 below:

1 =

(T1 T2 ) + enum
d T1 /d t d T2 /d t + eden

(7)

In a noiseless case when the denominator is

approaching zero, the numerator is also approaching
zero, so there is no singularity. However, if noise enum
and eden are present, enum will be amplified when the
denominator is approaching zero. Because the
denominator also contains noise, the equation is ill
defined over a region. The result is erratic with
unphysical time constants, rendering the temperature
reconstruction erroneous. Forney and Fralick [10] tried to
tackle the noise problem by working in the frequency
domain, such that

Tg1 = Tg 2 =

Figure 1 Comparison of the true and the reconstructed gas

temperature using frequency domain technique (OReilly, et. al. [11])

( 1)T1T2
T2 T1

(8)

where, the overbar denotes a discrete Fourier transform

(DFT) of the respective quantities. In this method, the
gas temperature is found directly, bypassing the step of
finding the time constants. Unfortunately, OReilly, et. al.
[11] later discovered that noise can drive Eq. 8 to a
singularity. When this occurs, the reconstructed gas
temperature exhibits large amplitude oscillations as
shown in Figure 1.

Recall that Eq. 4 can be solved if either an additional

criterion or information is introduced. So far the
techniques have been based on assuming a constant
time-constant ratio, which is susceptible to noise. Other
criteria are possible, such as the Least-Squares method
first proposed by Tagawa and Ohta [12]. They argued
that the two thermocouples in practice may not sense the
same temperature because of dust contamination and
spatial errors. Also, in the process of reconstruction,
noise in the thermocouple signals will prevent the two
gas temperatures being equal. Instead they sought to
minimize the difference over a time window of N
samples, which is given by:

1 N (i )
Tg1 Tg2(i )

N i =1

(9)
th

where the superscript i indicates the i

digitized data. Eq. 9 is minimized when

= 0;
1

=0
2

sample of the

(10)

The two extra equations enable the time constants to be

determined without a priori a time-constant ratio. With
noisy signals, this method may sometimes give negative
time constants, which are physically impossible. This
may be explained by considering Eq. 4. The 1dTw1/dt
and 2dTw2/dt terms represent the dynamic errors of the
thermocouples at any time. The least-squares criterion
only considers a difference in Tg (Eq. 9), and this can be
influenced by the time constants and/or the gradients. In
other words, it is a conditional minimization procedure,
dependent upon what gradients are given. Unfortunately,
it can be shown that noise in the signals can be overamplified when numerical gradients are evaluated. This
will be discussed further later. Therefore if the gradients
have been amplified, the least-squares method would try
to make a correction by giving a negative time constant,
since negative time constants have the effect of
attenuation.

Several improvements have been proposed by various

researchers such as:

Correlation coefficient [13] this method is based

on the assumption that noise has no correlation. By
using the correlation coefficient as the criterion, it
should better discriminate the signal from the noise.
Frequency domain [14] this method equates the
two reconstructed temperatures in the frequency
domain. This would allow filtering to be applied
easily. At the same time this method restricts the
bandwidth of the reconstruction by only considering
the spectrum up to a certain frequency.
Constant time-constant ratio [2] the least squares
method is prone to noise and can give negative time
constants.
By reintroducing the constant timeconstant ratio assumption, it reduces one order of
freedom in the equations, making this method less
likely to give negative time constants.
Total Least Squares [15] the response equations
(Eq. 4) are described by difference equations to form
an autoregressive (ARX) model with noise added in
the input as well as in the output. The parameters of
the ARX model are related to the time constants,
and can be estimated by the Total Least Squares.
Theoretically the Total Least Squares will give
unbiased estimates when the noise variances on the
input and output are equal.

All these improvements aim to handle the noise better,

so as to get an unbiased estimate of the time constants.
However, one serious limitation remains, that is the least
squares method processes the data in a time window.
Within each window, the time constants are assumed to
be constant. The size of the window (N in Eq. 9) is found
to be a compromise between temporal resolution and
accuracy [16]. Tagawa and Ohta [12] recommended that
the window should have a duration of at least 1.5 times
longer than the mean time constant of the smaller
thermocouple. Yet, ideally the window size should
accommodate the dynamics of the system. For example
during blowdown, a short window size should be used
compared to the time when the exhaust valve is closed.
Above all, without knowing the gas temperature, it is
difficult to find the optimal window duration for the
reconstruction. OReilly [11] overcame this limitation by a
nonlinear state estimator the Extended Kalman Filter
(EKF). The thermocouple model is written in a statespace model with the constant time-constant ratio
assumption. The gas temperatures (Tg) and the time
constant () being the states, are estimated
simultaneously and recursively, so there is no time
window to be set. The model itself is stochastic in nature,
allowing explicitly for noise on the signals. However it
also means that the dynamic of the states, in particular
Tg and are governed by the noise covariances, which
have to be specified a priori. Because the EKF works by
linearizing the states about the current states estimate, it
is a perturbation model. Unsuitable initial conditions and
covariances may lead to stability problems. Simulation
experience has shown the covariance on has a great
impact on stability. There are a total of six parameters;

without a reference thermometer, there is no guideline as

to how to tune these six parameters, since their effects
on accuracy are not known. An alternative Kalman filter
has been developed here. It has all the advantages of
the EKF, plus the following:

No a priori assumptions for the time constant ratio

The covariances are to be inferred from the data,
thereby eliminating any subjective tuning (see the
Estimation of noise parameters section).
The two reconstructed temperatures may differ. This
flexibility makes it less susceptible to noise (unlike
the constant time-constant method).
The accuracy of the estimation is quantified, and
provided to the user.

This is to the authors knowledge the most versatile

method available. Table 1 compares the various
methods in terms of the assumptions made.
Table 1
Method
Cambray
Forney and
Fralick
Tagawa and
Ohta
OReilly et. al.
Kalman filter

Assumptions in the different methods

for time constant estimation
Tg1 = Tg2
Yes
Yes

Constant
No
Yes

Constant
Yes
Yes

No

Yes

No

No
No

No
No

Yes
No

KALMAN FILTER
The Kalman filter is a recursive procedure for estimating
the latent variables, , in the following linear stochastic
equation [17].

(i ) = G (i ) (i 1) + w (i )
y (i ) = F ( i ) (i ) + v ( i )

(11)

where:
G is a flow matrix
w is the state noise, normally distributed with zero mean
and covariance W.
y are the multivariate observations
F is a transformation matrix
v is the observation noise, normally distributed with zero
mean and covariance matrix V.

TIME CONSTANT FORMULATION

The latent variables of interest are the time constants 1
and 2. To use the Kalman filter, the flow matrix, which
describes how 1 and 2 vary with time, has to be defined.
This requires knowledge of the system dynamics of the
time constants (a mathematical model). As discussed
before, this would be very difficult. The best alternative is
to assume G = I, which means that the most likely values
in the next time step are the same as the present values.
It would be the noise, w that allows the time constants to
change over time. It will be shown later on how this noise
can be estimated. This assumption will simplify the
application of the Kalman filter to a Dynamic Linear
Model [17] as,
( i 1)

(i )

(i )

= F + v (i )
(i )

+w

(12)

d Tw1
dT
+ 2 w2
dt
dt

(i )

, is the Kalman gain matrix,

2 (i )
y

R (i ) = (i1) + W (i ) ,

( )

2 (i )
y

( ) + ( )

= 2

(i )

2 (i )

variance due to uncertainty in the latent variables,

( )

(13)

(14)

By comparing Eq. 14 against Eq. 12, it can be seen that

Eq. 14 is a Dynamic Linear Model with:

A univariate observation, y = Tw1 Tw2

The two time constants (1 and 2) as the latent
variables,
The gradients defining the transformation matrix, F
2

Because the observation is univariate, only a scalar is

required to describe the observation noise, v.
RECURSIVE ESTIMATION

(i ) = R (i ) K ( i ) F ( i ) R (i )
where:

( i ) is the covariance of the latent variables,

= F (i ) R (i ) F T

(i )

2 . It

e ( i ) = y (i ) y (i ) , is the prediction error, and

A useful quantity is the likelihood of an observation,

given the model parameter before they are updated is
given by:

( )

( )

p y (i ) = N y ( i ) , y2

(i )

(16)

In Bayesian terminology this likelihood is known as the

evidence for the data point. It will be used for estimation
2
of the observation noise, .
ESTIMATION OF NOISE PARAMETERS
State noise
The noise parameter is estimated by Jazwinskis method
[18]. The state noise covariance matrix is assumed to be
2
an isotropic matrix, W = qI. The prediction variance (q0 )
is calculated, assuming no state noise, i.e. q = 0 by:

( )

2 (i )
q0

( )

= 2

(i )

( )

+ F (i ) (i 1) F T

(i )

(17)

The latent variables can then be estimated with the

following recursive formulae

(i ) = (i1) + K (i ) e (i )

( )

2 (i )

th

d Tw(i2) 1(i )

d t 2(i )

, is the estimated prediction

variance, and is composed of two terms: the observation

2
noise variance, and the component of prediction

Eq. 13 can be written into Matrix-Vector form for the i

time step as

d T (i )
Tw( i1) Tw( i2) = w1
dt

is the covariance of the latent

(i)
variables prior to y
being
observed,

y (i ) = F (i ) ( i1) , is the prediction.

Eliminating Tg in Eq. 4 and rearranging gives

Tw1 Tw2 = 1

( )
( )

R (i ) F T

is given by

(i )

(i )

(i )

If q0 is smaller than the prediction error (e ), the state

noise is inferred to be non-zero. This logic can be
implemented using a ramp function in Eq. 18.

x Kif x 0
h( x ) =
0Kotherwise

(15)

(18)

Therefore, q can be updated according to:

(i )

( ) ( )
( )

e 2 (i ) 2
q0
= h
F (i ) F T ( i )

(i )

(19)

Observation noise

AIR TEST RIG

Eq. 15 shows that the estimated prediction variance is

made up of the observation noise and the component
due to the uncertainty in the latent variables. Thus the
observation noise can be estimated by subtracting the
second component from the prediction error as

The thermocouples were tested in a well-controlled

environment before the engine exhaust temperature
experiments. A special air rig was set up as shown in
Figure 3. There are two hot air guns supplying hot and
cold air to a mixing valve. Within the valve, a shutter is
revolving, and admitting hot and cold air alternately. The
air is then mixed and accelerated through the nozzle at
the outlet. The air rig is able to create temperature
fluctuation from 90 to 120C, at up to 200 Hz.

( )

2 (i )

= h (e 2 ) F ( i ) R ( i 1) (F T )
(i )

(i )

(20)

It can be proved that the observation noise estimated by

Eq. 20 will maximize the likelihood of the observation in
the next sample (Eq. 16).

3-WAY
VALVE

SHUTTER

EQUIPMENT
THERMOCOUPLE PROBE

ADAPTOR
NOZZLE

A schematic diagram showing the main features of the

probe design is given in Figure 2. The thermocouple
probe is 100 mm long, and 3 mm in diameter. It
comprises of three main components:
1. Unsheathed fine gauge K type thermocouples,
2. Ceramic and stainless steel tubes,
3. Electrical socket and collet sets.
In this project, three pairs of bead-type thermocouples
with wire diameters ranging from 50 to 127 m (0.002 to
0.005) were used. This gives 3 independent
thermocouple pairs for comparison. With their different
sensor sizes, the effects of: the signal-to-noise ratio,
conduction and radiation, and the time-constant ratio can
be investigated. For the compensation technique to be
valid, the three thermocouples have to be close enough
so that the measured volume is sufficiently small
compared to the characteristic length scales of the
thermal fields being investigated [12]. Therefore, the
three thermocouples were densely packed, being about
500 m apart.
The thermocouples were soldered into a seven-way
LEMO socket, and this forms additional thermocouple
junctions. An independent thermocouple is therefore
needed to measure the junction temperature for coldjunction compensation.
100

3
4

STAINLESS
STEEL TUBE

CERAMIC 6BORE TUBE

KEYS:
CHROMEL
ALUMEL
COPPER

LEMO 7-WAY
SOCKET
NOT TO SCALE

Figure 2 Schematic diagram of the thermocouple probe

REFERENCE
JUNCTION
THERMOCOUPLE

COLD STREAM

HOT AIR
GUN

HOT STREAM

Figure 3 Schematic diagram of the air rig

To validate the Kalman filter model, the instantaneous air

temperature at the outlet was measured by a reference
thermometer. The reference thermometer is a hot wire
anemometer system using a two-wire technique in
constant temperature mode [19]. With dual wires,
velocity
and
temperature
can
be
measured
simultaneously. For the test conditions, the bandwidth of
the reference system was in excess of 30 kHz, which far
exceeded the temperature frequency that the air rig can
produce. Unfortunately, the hot wire system is too fragile
to be used in the engine exhaust directly.
ENGINE AND EXHAUST
The thermocouple probe was installed in the exhaust
pipe of a production-type, spark ignition engine. It has 4
3
cylinders with a swept volume of 1396 cm . The engine
specification is summarized in Table 2 [20].
Table 2

Specification of the test engine

Bore (mm)
Stroke (mm)
Compression ratio
Maximum bmep (bar)
Inlet valve opens (btdc)
Inlet valve closes (abdc)
Exhaust valve opens (bbdc)
Exhaust valve closes (atdc)

75.0
79.0
10.5
11.2 @ 3500 rpm
12
52
52
12

This study is concentrated on the 4th cylinder of the

engine. The positions of various instruments are given in
Figure 4. The thermocouple junctions are along the
exhaust pipes centerline. There is a piezoresistive
pressure transducer, about 27 mm upstream of the
thermocouple. Both of them are about 100 mm away
from the mean centre of the exhaust valves. The cylinder

and the inlet manifold pressures were also measured. To

gauge the effect of radiation, a thermocouple was
mounted on the external surface of exhaust pipe. Since
the largest thermal resistance from the exhaust gas to
the surrounding is the convection to and from the metal,
there should be little difference between the internal and
external surface temperatures. Moreover, a previous
study on the instantaneous in-cylinder temperature
suggested that there is only a 10 K temperature swing in
a 4-stroke cycle of a firing engine [21]. Hence the
temperature swing in the surface temperature of the
exhaust would be even less; therefore the dynamic error
in the exhaust pipe surface thermocouple can be
ignored.

Figure 5 Power spectral density of a raw thermocouple signal

measuring exhaust temperature. Sampling frequency is 300 kHz.

Figure 4 Schematic diagram showing how the instruments are placed

relative to the exhaust valves. Left hand diagram is a plan view. Right
hand diagram is a cross-sectional elevation. All dimensions are in
millimeters.

SIGNAL CONDITIONING AND DATA ACQUISITION

The thermocouple signals were amplified 200 times by
fast-settling instrumentation amplifiers. To reduce the
high frequency electrical noise, the amplified signals
were passed through a passive low-pass filter with a
-3 dB point at 13 kHz. Since the maximum frequency of
interest would be at least 10 times smaller, the phase
distortion and attenuation introduced by the filter should
be minimal. As a further assurance, a high-speed
spectral analysis was conducted on the thermocouple
signals. The results are plotted in Figure 5. The power
spectrum diminishes quickly and settles to -100 dB at
about 500 Hz Above this frequency, the spectrum
contains little power, indicating that no extraneous signal
has been added to the thermocouple signal. Signals from
other thermocouples show similar spectra.

Lastly, the signals were digitized by a computer-based

data acquisition system and saved in a standard ASCII
file for post analysis. The data was collected by multichannel scanning, in which the clock rate was controlled
by the TTL pulse of a shaft encoder which was coupled
to the crankshaft. Samples were collected every 1
degree crank angle.

MODEL VALIDATION
To ensure the integrity of the temperature reconstruction,
the Kalman filter was commissioned in a simulation study
and a well-controlled experiment before being applied to
the exhaust temperature measurements.
SIMULATION STUDY
A sinusoidal variation of the gas temperature was
defined. It was then used for solving the two
thermocouple temperatures by Eq. 4 with a common
initial temperature. The time constants, which have mean
values of 30 and 70 ms, were also defined to vary
sinusoidally at 1/5 of the temperature frequency. The two
synthetic thermocouple time series were contaminated
by Gaussian noise with very small variance of 1e-8. The
resulting gas and thermocouple temperature traces are
plotted in Figure 6.

The thermocouple signals were inputted to the Kalman

filter to estimate the time constant, and reconstruct the
gas temperatures. The estimated time constants are
plotted against the true defined time constants in
Figure 7. The variances of the estimates (), estimated
prediction, ( y ) and the evidence (
2

Figure 6 Gas and the two thermocouple temperatures in a simulation

study.

p( y ) ) are shown in

Figure 8. For the first 0.1 second, there is a start-up

transient as demonstrated by large run-in errors in the
time constant, their variances and their prediction
variances. It is also clearly shown that the start-up has
very low evidence (about an order of magnitude lower
than the steady sate), indicating the results should not be
trusted. Even for steady state the quality of the estimate
varies. The time constant traces exhibit some kind of
sawtooth structure compared to the true values. The
Kalman filter reflects such inaccuracies in the estimate
variances in Figure 8, where the local maximum
variances coincide with the peaks in the time constants,
e.g. at t = 0.54 s. Finally, the reconstructed temperatures
are shown in Figure 9. The two reconstructed
temperatures (R1 and R2) are almost identical
(indistinguishable in the figure). The start-up, as
expected has the greatest difference compared to the
true gas temperature. The sawtooth errors in the time
constants manifest in over and under-compensation at
the maximum and minimum gas temperatures.

Figure 7 Time constants estimated by Kalman filter, noise variance is

1e-8.

Figure 9 Comparison of the reconstructed temperatures and the true

gas temperature. R1 and R2 are indistinguishable in the figure

The Kalman filter performs well in virtually noiseless

conditions the reconstructions are respectable, and
the qualitative measures correctly indicate the estimation
errors. In reality measurements always have noise, so
the simulation was performed again with a realistic noise
level (noise variance is 0.1). In Figure 10, the time
constant estimates oscillate around zero, and show no
sign of tracking the true values, and of course the
reconstruction will fail.
Figure 8 Top is the variances of the time constants; middle is the
estimated prediction variance; bottom is the evidence against time. All
graphs are for the data corresponding to Figure 7.

matter how small the signal has become, it is still

possible to recognize the signal, and apply
compensation. This suggests that the bandwidth of the
reconstruction signal will be, in theory infinite or at least
as high as the Nyquist frequency. In the real world, all
signals contain noise. This implies at a certain frequency,
the signal is too small to be resolved from the noise.
Trying to recover the full bandwidth including the noise
spectrum, will invariably amplify the noise and get
erroneous results. Therefore, some kind of bandwidth
restriction must be implemented, and this will set the
bandwidth improvement of the reconstruction. For
example, if the thermocouple bandwidth is restricted to
100 Hz, the reconstructed temperature would be similar
to the true gas temperature up to 100 Hz. The extent of
any improvement will thus depend on the signal-to-noise
ratio.
Figure 10 Time constant estimates, noise variance = 0.1

Filtering and gradient estimation

It has been argued that for a successful reconstruction,
the bandwidth of the thermocouple signals must be
restricted. In the context of a Kalman filter, the gradients
of the thermocouple output need to be filtered. Tagawa
et. al. [13] used polynomial smoothing (a Savitzky-Golay
filter) because the gradients can be evaluated readily.
The polynomial fit requires a short data length to improve
its frequency response, however if the data length is
made too short, the accuracy of the gradient cannot be
assured. In this study, use of a Marr-Hildredth filter is
proposed. The Marr-Hildredth filter suppresses the high
frequency noise by convolving the signal with a Gaussian
kernel (Eq. 22). The Gaussian kernel is defined by:

g (t ) = exp( t 2 )
Figure 11 Time constant estimates with noise variance = 0.1. The
transformation matrix is derived from the noiseless data.

where is a smoothing parameter that sets the cutoff

frequency. Let the thermocouple signal be x(t); the
filtered signal is then:

y (t ) = x(t ) g (t )
When the simulation is rerun for the same conditions,
except that the transformation matrix (F), (which is
formed by the gradients), is calculated from the noiseless
data, then the result is shown in Figure 11. With accurate
gradients, the Kalman filter is able to estimate the time
constants. So it can be concluded that the Kalman filter
cannot handle noise in F. This is similar to the leastsquares method when it gives a negative time constant.
The Kalman filter works by maximizing the likelihood of
the prediction (y). The prediction is however calculated
through the transformation matrix, F. A noisy F is less
deterministic, so it cannot transform the time constants
to the prediction. Therefore the estimations fail.
Noise and reconstruction
The process of reconstruction is to recover (amplify) the
high frequency component of the thermocouple signals,
which have been attenuated due to the thermal inertia of
the thermocouple junction. In a noiseless signal, no

(21)

(22)

The time derivatives can easily be found by using the

commutative law of convolution as follows:

d y (t ) d x(t )
=
g (t )
dt
dt
d g (t )
= x(t )
dt

(23)

where dg(t)/dt = -2texp(-t ).

Spectral analysis was performed on the synthetic signal
data of the previous section. The power spectral density
(PSD) relative to DC is plotted in Figure 11. The peak at
about 4 Hz is the fundamental frequency of the signals.
The power of the signal reduces to the white noise level
(-80dB) at about 100 Hz.

illustrated in Figure 14. The bigger the value of , the

lesser the smoothing, and hence the higher the cutoff
frequency.

Figure 12 Power spectral density of T1 with noise variance of 0.1

Figure 14 Frequency response of Marr-Hildredth filter with various . A

Hamming window was applied to all traces.

The Marr-Hildredth filter is applied to the transformation

matrix in the simulation. Compared with the best case in
Figure 11, this time the estimations are almost identical
and so can be considered to be satisfactory, as in
Figure 15. The only notable difference is the longer startup transient.

Figure 13 The power spectral density of the gradient over its signal
(the transfer function of the gradient). A Hamming window was applied
to the Marr trace.

The gradient of T1 was calculated by the central

difference and Marr-Hildredth filter. Their frequency
responses are given in Figure 13. The trace for central
difference shows a straight line sloping upwards. It can
be proved that the spectrum of a perfect differentiator will
have a 20 dB / decade slope. In that respect, the central
difference method is not far from ideal. However after 20
Hz, the spectrum becomes increasingly noisy. This is the
onset of noise dominating over the signal. For time
constant estimation, it would be prudent to restrict the
bandwidth at that point, as any information in the signal
is corrupted. This is implemented by a Marr-Hildredth
filter as shown in Figure 13. The spectrum is tapered at
low frequencies. This is a consequence of the
windowing, so as to avoid spectrum leakage. Beyond 2
Hz, the spectrum has a 20 dB/decade slope, which is
identical to the central difference one. At around 20 Hz
(-3dB point), the filter now has a significant effect,
suppressing the noise. The cut-off frequency can be
customized by adjusting the smoothing parameter (), as

Figure 15 Time constant estimates with noise variance = 0.1,

transformation matrix processed by Marr-Hildredth filter with = 2e-3

AIR RIG STUDY

The second phase of the model validation is to apply the
Kalman filter to a well-controlled experiment. Simulation
is often too simplistic, and may miss out important
features from the real world. The object in the air rig
study is to investigate the practicality of the technique.

the assumption that they sense the same temperature. If

the measurement is offset by some external means, the
assumption will break down, and the techniques will
certainly fail. As a demonstration, the simulation can be
re-run with offset intentionally added to one
thermocouple signal. When this is done, the time
constant estimates also show large swings between the
true values at the temperature frequency; see Figure 17.

Experiments were conducted with the air test rig at

various temperature frequencies. The fundamental
frequencies (fT) tested were from 2 Hz, up to 50 Hz. This
is equivalent to a 4-stroke engine running from 240 to
6000 rpm. The shutter in the mixing valve was designed
to minimize velocity variation. Measurements from the
hot wire have shown the mean velocity at the outlet was
nearly constant at 21 m/s for all fT. The turbulence
intensities were also roughly constant at 6%. Since the
temperature only fluctuated from 90 to 120C and the
flow velocity did not change significantly, the time
constants should have small variations only. The MarrHildredth filter was applied in the same manner as in the
simulation study. The time constant estimates in 5
temperature cycles are given in Figure 16. Both of them
have unusual fluctuations. For example, 2 can oscillate
from 0.02 s to 0.06 s at the same frequency as the air
temperature. These results do not seem to be plausible.
Figure 17 Adverse effect of offset errors demonstrated by a simulation

The thermocouple signals were therefore normalized to

have a common mean value. The time constants were
re-estimated as shown in Figure 18. Both time constants
now have much smaller variations. Their high frequency
components may indicate the turbulence in the flow.

Figure 16 Time constant estimates at fT = 2.6 Hz. Subscript 2 and 3

denotes 0.002 and 0.003 thermocouples respectively.

Offset Errors and Normalization

Further investigation showed that the mean values of the
thermocouples did not agree. There was about a 2.1 K
difference. The mean values represent the DC
component of the signals, and so should be unaffected
by the thermal inertia of the thermocouples. It turned out
that the amplifiers of each thermocouple had different
offsets from ground, and hence the difference in the
mean values. In measurement terms, 2 K difference is
quite acceptable, but given that the raw temperature
ranges measured by the thermocouples were only 25
30 K, this could seriously impair the temperature
reconstruction. All two-thermocouple techniques rely on

Figure 18 Time constant estimates at fT = 2.6 Hz using the normalized

signals.

With these time constant estimates, the thermocouple

temperatures are reconstructed and plotted against the
true gas temperature in Figure 19. Unlike the
reconstructed temperature, the waveform of the true gas
temperature is not well-defined. This could be the effect

of electrical noise in the signal and turbulence. It was

determined that the temperature error due to electrical
noise was only 0.4 K at most. Furthermore, if the true
gas temperature signal is purely contaminated by
electrical noise, the noise should be approximately
Gaussian distributed and contain no systematic patterns.
Figure 20 clearly illustrates that the temperature variation
(measured by the standard deviation) has structure.
Since the shutter generates 3 temperature pulses in one
revolution, the 3 cycles in Figure 20 must be related to
the air rig system itself. Thus the noisy appearance of
the true gas signal must be the consequence of
turbulence.

comparison, a cross plot for the raw thermocouple

temperature is provided in Figure 22. Only the signal
from the 0.002 thermocouple is shown. The raw
thermocouple signal suffers from attenuation, and this
will translate into the data points sloping at a steeper
angle than unity. This is shown as the dashed line in
Figure 22. Also, Figure 22 seems to have a void at the
centre of the data points; this is an indication of a phase
error. In the ideal reconstruction, all data points will lie on
the black line with unity slope. Unfortunately the cross
plot shows a scattering around the black line with no sign
of any phase error. It can be concluded that the
reconstruction can only recover the main features of the
true gas signal. The high frequency features such as
turbulence cannot be recovered. This is probably
because of the bandwidth restriction imposed by the
Marr-Hildredth filter.

Figure 19 Comparison of the reconstructed temperatures against the

true temperature at fT = 2.6 Hz. R2 and R3 are indistinguishable in the
figure. For clarity, only part of the whole time series are shown.
Figure 21 Cross plot of the true gas temperature against the
reconstructed temperature at fT = 2.6 Hz. Black line has a unity slope.

Figure 20 Standard deviation of the true gas temperature with valve

rotation, an ensemble average over 100 cycles.

Because of turbulence, it is hard to evaluate the

accuracy of the reconstruction objectively. An alternative
is to plot the true gas temperature against the
reconstructed temperature a cross plot is an
informative way of doing this (Figure 21). For

Figure 22 Cross plot of the true gas temperature against the raw
thermocouple temperature at fT = 2.6 Hz. Black solid line has a unity
slope. Black dashed line is the best fit line for the data.

Limit of Temperature Reconstruction

At a 10 Hz fundamental frequency, the thermocouple
signals become heavily attenuated and phase delayed
(Figure 23). The temperature signal spectra suggest that
there should be a usable bandwidth of 185 Hz. However
the time constant estimates after filtering at that
frequency are found to be too small; 2 is close to 0 and
3 has a mean of 0.004 s. As mentioned, the time
constants are independent to the temperature frequency
in the air rig. A comparison of the true temperature and
the thermocouple temperatures has indicated that the
mean values of 2 and 3 are around 30 and 60 ms
respectively. The very small estimates are signs of
insufficient filtering on the gradients. This leads to the
question of what is an appropriate level of filtering. Since
the mean time constants are known, they can be used as
a rigorous test for the accuracy of the estimation. The
time constants are estimated again with various cutoff
frequencies. Their means are plotted against the cutoff
frequency in Figure 24. Three regions can be depicted
from the plots. The first region is beyond 110 Hz where
the mean time constants are virtually zero. This is the
noise dominant region as described in the noise and
reconstruction section. Between 30 Hz and 100 Hz is the
transition region in which the Marr-Hildredth filter is
becoming effective, yet the noise in the signal still has a
noticeable effect. Below 30 Hz, the mean time constants
settle to the true values. This is the region in which the
Kalman filter can give unbiased estimates of the time
constants. Therefore 30 Hz could be regarded as the
upper limit of temperature reconstruction. Beyond that,
noise compromises the accuracy of the estimation.

Figure 23 Cross plot of the true gas temperature against the raw
thermocouple temperature at fT = 10 Hz. Black solid line has a unity
slope. Black dashed line is the best fit line for the data.

Figure 24 The mean time constant estimates obtained from filtered

data with different cutoff frequencies.

Filtering and Reconstructed Temperatures

By restricting the bandwidth to 30 Hz, the reconstructed
temperatures in Figure 25 are obtained. Because of the
filtering required, many higher harmonics cannot be
recovered. It is now difficult to judge the accuracy of the
reconstruction by direct comparison against the true
temperature trace, simply because the true temperature
contains too much high frequency information. Figure 26
may offer more insight. Instead of plotting the true
temperature directly, it has been filtered in the same way
as the thermocouple signals to 30 Hz. Now the
reconstructed and the true temperature traces resemble
each other closely. The worst part of the reconstruction
can be easily pinpointed, e.g. between i = 2000 and
2150. A bandwidth of 30 Hz seems rather narrow, but it
is already an improvement of 6 to 11 times over the
original thermocouples. A higher bandwidth can be
enforced by having a higher cutoff frequency, such as
the case in Figure 27 and Figure 28. Certainly more high
frequency harmonics can be seen, but at the expense of
some serious flaws at i = 2200 and 2550. The cross plot
reveals more flaws. The odd peaks in the time series
appear as triangles protruding from the data points.
Furthermore, the data points are inclined at a steeper
slope than the unity line. This is a sign of undercompensation of the fundamental frequency. The
alternative is to have different levels of filtering for the
time constant estimation and the temperature
reconstruction. In the time constant estimation, gradients
are required to form the transformation matrix (Eq. 14).
Gradients are also required in the temperature
reconstruction (Eq. 4). Gradients evaluated with different
cutoff frequencies may be used. For example, the
temperatures in Figure 29 have been reconstructed with
gradients of 85 Hz bandwidth. However the time
constants are estimated from signals with a 30 Hz
bandwidth only. By comparing Figure 26 and 29, there is
a dramatic improvement in terms of revealing the high
frequency features. However, users must be warned that
now there is a more noticeable difference between the
two reconstructed temperatures. The accuracy of the

reconstruction has deteriorated. At the peaks, an error of

5 K is common. The root of the problems is that
components at 85 Hz were not considered by the
Kalman filter, and the reconstruction temperatures
possess no maximum likelihood (evidence), so the
accuracy can no longer be guaranteed. Therefore it is
concluded that this method of bandwidth improvement is
only suitable for a qualitative study, in which the absolute
value of the temperature is unimportant.

Figure 27 Comparison of the reconstructed temperatures against the

true temperature at fT = 10 Hz. All signals have their bandwidth
restricted to 54 Hz.

Figure 25 Comparison of the reconstructed temperatures against the

true temperature at fT = 10 Hz. R2 and R3 are indistinguishable in the
figure. For clarity, only part of the whole time series is shown.

Figure 28 Cross plot of the true gas temperature against the raw
thermocouple temperature at fT = 10 Hz. All signals have their
bandwidth restricted to 54 Hz.

Figure 26 Comparison of the reconstructed temperatures against the

true temperature at fT = 10 Hz. R2 and R3 are indistinguishable in the
figure. All signals have their bandwidth restricted to 30 Hz.

Figure 29 Effect of different filtering, an 85 Hz cutoff is used for

reconstruction, while a 30 Hz cutoff is used for time constant
estimation.

ENGINE EXHAUST MEASUREMENTS

The thermocouple compensation technique has been
applied to some engine exhaust measurements. In all
the tests, the engine was running at a stoichiometric
condition with a fixed ignition timing of 15 btdc. The fuel
used was a standard unleaded gasoline with 95 RON.
The engine was run at 3 different loads (2, 4 and 6 bar
bmep) and two speeds (1000 and 2000 rpm). The engine
was allowed to warm up before the exhaust
temperatures were measured by the thermocouple
probe; signals were logged by the data acquisition
system as described in the Equipment section.

All the two-thermocouple techniques discussed here are

based on the response equations (Eq. 4). These
equations have assumed that the radiation and
conduction terms are negligible. This assumption is
acceptable for the air rig experiment because:

The length-to-diameter ratios are more than 100 for

all the thermocouples, so the conduction loss to the
probe body would be small.
The operating temperature is around 100C, which is
only an 80 K temperature difference (Tw-Tsurr) from
the surroundings. Thus the radiative loss would be
4
4
insignificant as it is proportional to Tw -Tsurr .

In an engine exhaust, the time-averaged thermocouple

temperatures were typically at 600C, and the conduction
and radiation losses may become substantial, and
cannot be ignored. This section will investigate this issue
by means of theory and experiments.
Considering the AC component only, Eq. 4 is the same
as Eq. 3 if:

1Tg = Tg1
2Tg = Tg 2

(24)

For the two-thermocouple techniques to work,

~
~
Tg1 = Tg 2

CONDUCTION AND RADIATION

radiate to the exhaust pipe and conduct heat to the probe

body. Thus their time-averaged temperatures will
indicate the relative extent of radiation and conduction
losses. The time-averaged cycle temperatures of the
three thermocouples are listed in Table 3. At most the
difference is only 6.5 K, which is within the tolerance of
the thermocouples at these temperatures. Also had
conduction and radiation been significant, the largest
thermocouple would have had the lowest time-averaged
temperature, and vice versa. This is because of the
following reasons:

(25)

Eq. 25 is satisfied either when 1 and 2 are close to

unity (negligible conduction and radiation), or when 1
and 2 are approximately the same. The conclusion is
that the reconstruction can still proceed if the conduction
and radiation characteristics of the two thermocouples
are similar. In this case, the reconstructed temperature
will have the dynamic errors corrected, and the
conduction and radiation can be corrected on a steadystate basis if 1 and 2 are known.
The time-averaged temperatures of the thermocouples
represent their DC component, and should thus have no
dynamic errors. However, the thermocouple will still

The largest thermocouple has the lowest length-todiameter ratio.

The thermocouple always conducts and radiates
heat to the surrounding.

No such trend can be seen from Table 3. The exhaust

pipe temperature was monitored. Depending on the load
and speed, it varied from 300 to 400C, so there was
only 150-250 K temperature differences between the
thermocouples and the pipe wall. Therefore, from the
above reasons, it was concluded that conduction and
radiation may be ignored in the engine exhaust
measurement.
Table 3

Time-averaged
thermocouple
temperatures. Engine was running at 2
bar bmep, 1000 rpm.

Thermocouple
0.002, T2
0.003, T3
0.005, T5

Time-averaged temperature
[C]
582.3
587.5
581.0

METHOD OF ANALYSIS
The thermocouple signals were compared cycle by cycle.
Even though the engine was running in steady state, the
thermocouple temperatures were subject to large cycleby-cycle variations as depicted in Figure 30. For this
reason, each thermocouple signal is ensemble averaged
over 200 cycles to get a representative mean. This is
primarily to ensure fair comparison across different loads
and speeds, but it also improves the signal to noise ratio
of the signal. For fast tracking of the time constants, the
signals are upsampled 10 times. Finally these processed
signals are used for compensating the thermocouple
temperatures in the same way as they were in the air-rig
experiment.

Kalman filter cannot react instantaneously to a change in

system dynamics. It needs a finite amount of sample to
adapt to the new dynamics, so the time constants could
have decreased at EVO. Point D and point E can be
explained likewise. All this information is to warn the
user, the quality of the estimation varies, especially at the
time when the system dynamics changes. Users should
thus be cautious about the accuracy of the reconstructed
temperatures within low evidence regions.

Figure 30 Thermocouple temperature, T2 , for 10 successive engine

cycles. The engine was running at 4 bar bmep, 1000 rpm.

ACCURACY AND LIMITATIONS

In the absence of the reference temperatures, the
accuracy of the temperature compensation can only be
evaluated indirectly from: firstly the quality measures,
and secondly by comparing results from different
thermocouple pairings. The thermocouple probe has
three thermocouples, forming three independent pairs. If
the compensation technique works, the reconstructed
temperature from different pairs should give the same
exhaust temperature.
The results for 4 bar bmep at 1000 rpm are used to
illustrate the quality of the compensation. The bandwidth
of the temperature reconstruction has been determined
to be 80 Hz. Figure 31 shows the time constant
estimates (2 and 3) for the 0.002 and 0.003
thermocouple pair. Both time constants have responses
with step changes. Before EVO, the time constants stay
almost constant at the highest level in the cycle. This is
the stagnant period when there is no flow. As they
approach EVO, they decrease slightly. During blowdown,
there are sharp falls in the time constants until about
420ca. They then increase sharply to some mid level,
and return back to the stagnant level after EVC. These
characteristic changes in time constants are seen at
other loads and speeds. 2 sometimes appears to be
erratic, e.g. the odd peak just after EVO and the ripples
after EVC. Figure 32 may be used to judge if these are
artifacts. A very low likelihood is registered five times.
This indicates when the Kalman filter fails to estimate the
time constants or when the system dynamics are
changing. The first low likelihood (Point A) coincides with
3 falling from the stagnant level (0.06 s) before EVO.
This might mean that the decrease in the time constant
has a low confidence, and so should not be trusted. The
second low evidence point (Point B) occurs at EVO.
Point C follows closely when both time constants
decrease sharply. The Kalman filter is able to detect the
change in system dynamics, i.e. from the stagnant
regime to the blowdown regime. Unfortunately, the

Figure 31 Time constant estimates of the 0.002 (2) and 0.003

thermocouples (3) when the engine was running at 4 bar bmep, 1000
rpm. EVO and EVC indicate when the exhaust valve opens and closes
respectively.

Figure 32 Logarithm of the likelihood of the estimation. Letters A to E

point out regions of low likelihood.

Another method to evaluate the accuracy of the

technique is to compare the results from different pairs of
thermocouples in one test. Figure 33 compares the
reconstructed temperatures using the 0.002/0.003 pair
(R23) and the 0.002/0.005 pair (R25). Because the two
reconstructed temperatures from each pairing are almost
identical, only one from each pair is shown. The
bandwidth of the reconstruction has been restricted to 80
Hz. The two reconstructed temperatures have
remarkably similar waveforms, so no artifact has been

introduced. Comparing their magnitude, the peak

temperatures have a largest difference of 30 K. Apart
from this, R23 and R25 typically differ by 10 K only, which
is quite acceptable, given that the fluctuations in T2, T3
and T5 are so heavily attenuated. Moreover, the R23 and
R25 traces are in phase, for example, both show an
immediate cooling after EVO. In summary, the above
findings have confirmed that the technique is robust, and
suitable for applying to exhaust temperature
measurements.

Figure 34 Comparison of the reconstructed exhaust temperature at 6

bar, 1000 rpm by extending the bandwidth of the gradients. R2 and R3
are the temperature reconstructed from thermocouples T2 and T3
respectively.

Although the 6-bar result is less than ideal, the

comparison of the reconstructed exhaust temperatures
can still reveal many interesting features, as seen in
Figure 35.
Figure 33 Comparison of the reconstructed temperatures (R23, R25)
from different pairings. The uncompensated temperatures are also
shown. The engine was running at 2 bar bmep, 1000 rpm.

DIFFERENT LOADS
The exhaust measurements at various engine loads
were analyzed. The results are listed in Table 4. It was
found that the 6-bar case was much more difficult to
compensate, and the bandwidth achieved was only 60
Hz. In the Filtering and Reconstructed Temperatures
section, the high frequency features can be revealed by
having different levels of filtering on the gradients. This
method was applied to the 6 bar results as shown in
Figure 34. A bandwidth of 60 Hz is just insufficient. The
initial cooling and the first peak in the 500 Hz traces are
absent in the 60 Hz trace. However, it must be stressed
that the absolute magnitude of these high frequency
features may be subject to large errors as proven in the
air rig experiment.
Table 4

Results of time constant estimation

under various loads

Nominal load, bmep

Bandwidth [Hz]
Mean 2 [s]
Mean 3 [s]

2 bar
80
0.037
0.082

4 bar
80
0.024
0.048

6 bar
60
0.020
0.032

Figure 35 Reconstructed temperatures for different engine loads at

1000 rpm.

When the exhaust valve is closed, the reconstructed

temperatures are very similar, decreasing at about the
same rate. In fact, comparing back to the
uncompensated temperatures, there is always a 30 K
difference, regardless of the engine load. This probably
suggests that the exhaust gas is losing heat to the
exhaust pipe, which is a process not directly affected by
the engine. At EVO, the 2-bar and 4-bar traces decrease
rapidly. As discussed earlier with Figure 34, the same
should apply to the 6-bar case. This rapid cooling may be
attributed to 2 effects:

1. Before EVO, the valve has been in contact with the

seat for a while, so the valve surface in contact with
the valve seat could be a few hundred Kelvin cooler
than the gas at EVO. During the initial phase of the
blowdown, the flow across the valve is a choked
flow, which has very heat transfer coefficient. With
the combination of a large temperature difference
and a high heat transfer coefficient, the exhaust gas
can lose a substantial amount of heat and hence the
decrease in its temperature.
2. The measuring point is about 100 mm away from the
exhaust port exit. There are some exhaust gases
which left the cylinder at the end of the previous
exhaust cycle and will have been stationary in the
exhaust port while the valve has been closed. They
have been cooled substantially because the cylinder
head is maintained at about 100C by the coolant. At
the start of the blowdown, they are mixed with the
hotter gas which has just left the cylinder [22]. Hence
the exhaust temperature does not increase
immediately after EVO.

Table 5

Results of engine cycle analysis under

various loads, all are averages over 200
cycles

Nominal load, bmep

bmep [bar]
IMEP [bar]
Inlet manifold pressure,
gauge [bar]
Crank angle at which
maximum cylinder
pressure occurs [atdc]
Mass fraction burnt, 010% [ca]
Crank angle at which
10% mass fraction burnt
[atdc]

2 bar
1.9
2.35

4 bar
4.2
4.71

6 bar
6.1
6.46

-0.56

-0.35

-0.19

29

27.1

22.7

24.3

21.4

18.7

12

DIFFERENT SPEEDS

For 4 bar and 6 bar bmep, what follows the rapid cooling
is an increase in exhaust temperature to the first peak
just before bdc. The gas temperatures decrease from
this peak to a minimum at about 440ca, then increase
again until 480ca. These general changes in exhaust
temperatures agree with Caton & Heywoods study [23],
in which the exhaust temperature was measured by a
resistance wire thermometer [7]. Heywood [22] pointed
out that the local minimum at 440ca is the transition
from blowdown flow to displacement flow. Referring back
to the time constant estimates in Figure 31, the time
constants suddenly increase at this point, whilst the
likelihood of the estimation in Figure 32 indicates there is
a change in the system dynamics. The 2-bar result does
not follow the trend. Perhaps the mass flux becomes so
low at 2 bar bmep that it forms only a small proportion of
the mixed gas (Reason 2), and so delays the
temperature rise.

Two different engine speeds (1000 and 2000 rpm) have

been investigated with a fixed bmep of 2 bar. It was
found that the bandwidth of the compensation doubles
when the engine speed doubles. This is probably
because the mean exhaust flow rate is roughly double,
so the responses of the thermocouples are improved.
This is reflected by the much lower mean time constants
in Table 6. The bandwidth improvement could also be
related to the sampling frequency. The sampling rate is
synchronized to each crank angle, so by doubling the
engine speed, the sampling rate doubles. Practically, the
Kalman filter has double the amount of data to estimate
the time constants. This is one of the reasons why the
signals were upsampled 10 times in the data processing
stage.

One of the obvious differences in the exhaust gas

temperatures at different engine loads is in the peak
temperatures or temperature ranges. Increasing the load
should increase the peak temperatures and the
temperature range. However, this does not occur in the
experiments reported here, and this is largely the
consequence of a fixed ignition timing. A lower engine
load leads to relatively more residuals trapped. Residuals
slow down the combustion, as evident by longer 0-10%
mass fraction burn duration, and later peak pressures;
see Table 5. Therefore, with a fixed ignition timing, the
2 bar bmep has the most retarded ignition. This implies
less work is extracted in the expansion stroke, giving a
hotter exhaust gas.

Nominal engine speed

Bandwidth [Hz]
Mean 2 [s]
Mean 3 [s]

Table 6

Results of time constant estimation at

2 bar bmep at different engine speeds.
1000 rpm
80
0.037
0.082

2000 rpm
160
0.017
0.035

The reconstructed temperatures are plotted in Figure 36.

It is clear that increasing the engine speed, raises the
exhaust temperatures throughout the exhaust process.
There are some other similarities:

When the exhaust valves are closed, the exhaust

gas cools down steadily (approximately constant
gradients).
The initial cooling upon EVO is evident.
The highest and lowest temperatures occur at the
same crank angle.
After the cooling, there is a region of constant
temperature before the exhaust gas reaches its peak
temperatures. This region arrives earlier (bdc vs.

30abdc) and lasts longer (48 ca vs. 33 ca) at the

higher speed.

This suppresses the high frequency noise, and the time

derivatives can be found easily.

This is still yet to be explained. What is required is to

perform a 1-D unsteady flow simulation, so that
measurements and a model can be compared directly.
This should give some new insights into the underlying
physics that governs the exhaust process.

In the air rig experiment, with the aid of a reference

thermometer, several practical issues were highlighted.
First, offset errors exist between thermocouple signals,
and this leads to the time constant estimates showing
large departures from their true values. To solve this, the
thermocouple signals have to be normalized. Second,
with the bandwidth restriction (because of noise), the
temperature reconstruction can only recover the main
features of the true gas signal. The high frequency
features such as turbulence cannot be recovered. The
frequency limit of the temperature reconstruction has
been established by looking at the mean time constant
estimates at various cut-off frequencies. The unbiased
estimates can be obtained when the mean values settle
below a certain frequency. This sets the possible
bandwidth improvement. Further improvement can be
partially achieved by having different filtering levels for
the time constant estimation and the temperature
reconstruction, but it is only suitable for a qualitative
study, since the accuracy is no longer guaranteed.

Figure 36 Comparison of the reconstructed temperature at 2 bar

bmep, different engine speeds.

CONCLUSIONS
The dynamic response of a thermocouple can be
described by a time constant. Because it is too difficult to
predict analytically, a thermocouple compensation
technique has been developed to estimate the time
constant in situ. By this means, the instantaneous gas
temperature may be found, once the time constant is
known. This technique is one of many 2-thermocouple
techniques, which are based on the assumption that two
thermocouples are exposed to the same temperature
and velocity fields. It is based on a Kalman filter, and has
the following advantages over other methods:

No a priori assumptions for the time constant ratio

The data is analyzed recursively and the time
constants are estimated at each time step. There is
no issue of window size.
The two reconstructed temperatures may differ. This
flexibility makes it less susceptible to noise.
Qualitative measures of the estimation are provided
to the user.

The Kalman filter has been commissioned in a

simulation study and an air rig experiment. The
simulation has shown that the Kalman filter performs well
in virtually noiseless conditions, but returns unphysical
estimates when a realistic level of noise is present. Noise
will get amplified in the process of reconstruction, giving
erroneous results. For successful estimation, the
bandwidth of the signals has to be restricted by a filter.

Finally, the technique has been applied to some engine

exhaust measurements. By comparing the timeaveraged temperature of the three thermocouples, it has
been concluded that the relative extent of radiation and
conduction losses are negligible. The likelihood of the
estimation shows that the Kalman filter is able to detect
the changes in system dynamics, e.g. from the stagnant
regime to the blowdown regime. However, the Kalman
filter needs a finite amount of sample to adapt to the new
dynamics, hence the results of the transitional period
should be used cautiously. The accuracy of the
reconstruction is further examined by comparing the
results from different pairs of thermocouples in one test.
They are remarkably similar with the largest difference of
30 K at the peak temperatures which occur with high
temperature gradients.
After reconstruction, the exhaust gas temperatures at all
engine loads investigated, are seen to cool down at
EVO. This is due to the large heat loss to the valve, and
the mixing of the cooled gas in the exhaust pipe. At 4 bar
and 6 bar bmep, the gas temperatures decreases from
the first peak to a minimum and then increase again.
These general changes agree with previous studies,
which had a lower bandwidth [23]. It is found that
increasing the engine load reduces the peak
temperatures and the temperature ranges. This is due to
a fixed ignition timing, resulting in retarded ignition for the
lower engine loads. A higher engine speed raises the
exhaust temperature throughout the exhaust process.
The initial cooling upon EVO still exists. A region of
constant temperature exists following the initial cooling.
This remains unexplained. A 1-D unsteady flow
simulation may give some insight into these exhaust
processes.

ACKNOWLEDGMENTS
This work was supported by the Ford Motor Company.
The first author (KK) would also like to acknowledge the
financial support of Foundation for Research, Science
and Technology of New Zealand.

13.

REFERENCES
14.
1. Koehlen, C., Holder, E., and Vent, G., Investigation
of Post Oxidation and Its Dependency on Engine
Combustion and Exhaust Manifold Design. SAE
Paper 2002-01-0744, 2002.
2. Kee, R., O'Reilly, P.G., Fleck, R., and McEntee, P.T.,
Measurement of Exhaust Gas Temperatures in a
High Performance Two-Stroke Engine. SAE Paper
983072, 1998.
3. Bauer, W., Tam, C., Heywood, J.B., and Ziegler, C.,
Fast Gas Temperature Measurement by Velocity of
Sound for IC Engine Applications. SAE Paper
972826, 1997.
4. Andresen, P., Meijer, G., Schlter, H., Voges, A.,
and Hentschel, W., Fluorescence Imaging Inside an
Internal Combustion Engine using Tunable Excimer
Lasers. Appl. Opt., 1990. 29: p. 2392-404.
5. Koch, A., Voges, H., Andresen, P., Schlter, H.,
Wolff, D., Hentschel, W., Oppermann, W., and
Rother, E., Planar Imaging of a Laboratory Flame
and of Internal Combustion using UV Rayleigh and
Fluorence Light. Appl. Phys. B, 1993. 56: p. 177-84.
6. Benson, R.S. and Brundrett, G.W., Development of a
resistance wire thermometer for measuring transient
temperatures in exhaust systems of internal
combustion engines, in Temperature - Its
Measurement and Control in Science and Industry.
1962, Reinhold: New York. p. 631-653.
7. Benson, R.S., Measurement of transient exhaust
temperatures in I.C. engines. The Engineer, 1964.
217: p. 376-383.
8. Scadron, M.D. and Warshawsky, I., Experimental
Determination of Time Constants and Nusselt
Numbers for Bare-Wire Thermocouples in HighVelocity Air Streams and Analytic Approximation of
Conduction and Radiation Errors. NACA TN 2599,
1952. p. 5-6, 20-22, 40-45.
9. Cambray, P., Measuring Thermocouple Time
Constants: A New Method. Combustion Science and
Technology, 1986. 45: p. 221-224.
10. Forney, L.J. and Fralick, G.C., Two wire
thermocouple: Frequency response in constant flow.
Review of Scientific Instruments, 1994. 65(10): p.
3252-3256.
11. O'Reilly, P.G., Kee, R.J., Fleck, R., and McEntee,
P.T., Two-wire thermocouples: A nonlinear state
estimation approach to temperature reconstruction.
Review of Scientific Instruments, 2001. 72(8): p.
3449-3457.
12. Tagawa, M. and Ohta, Y., Two-Thermocouple Probe
for Fluctuating Temperature Measurement in
Combustion- Rational Estimation of Mean and

15.

16.

17.

18.
19.
20.
21.

22.
23.

Fluctuating Time Constants. Combustion and Flame,

1997. 109: p. 540-560.
Tagawa, M., Shimoji, T., and Ohta, Y., A twothermocouple probe technique for estimating
thermocouple time constants in flows with
combustion: In situ parameter identification of a firstorder lag system. Review of Scientific Instruments,
1998. 69(9): p. 3370-3378.
Tagawa, M., Kato, K., and Ohta, Y., Response
compensation of temperature sensors: Frequencydomain estimation of thermal time constants. Review
of Scientific Instruments, 2003. 74(6): p. 3171-3174.
Hung, P.C.F., McLoone, S., Irwin, G., and Kee, R.J.
A Total Least Squares Approach to Sensor
Characterisation. 13th IFAC Symposium on System
Identification,
SYSID-2003,
Rotterdam,
The
Netherlands, August 27-29, 2003.
Kar, K., Three-Thermocouple Technique for
Fluctuating Temperature Measurement. Internal
Report, Department of Engineering Science,
University of Oxford, 2002. p. 3-15.
Penny, W.D. and Roberts, S., Dynamic Linear
Models, Recursive Least Squares and SteepestDescent Learning. Technical Report, Department of
Electrical and Electronic Engineering, Imperial
College of Science and Technology and Medicine,
1998. p. 1-3.
Jazwinski, A.H., Adaptive filtering. Automatica, 1969.
5: p. 475-485.
Bruun, H.H., Hot-Wire Anemometry. 1995, Oxford:
Oxford University Press.
Stone, C.R., Introduction to Internal Combustion
Engines. Third ed. 1999, Basingstoke: Macmillan
Press Ltd.
Oude Nijeweme, D.J., Kok, J.B.W., Stone, C.R., and
Wyszynski, L., Unsteady in-cylinder heat transfer in a
spark ignition engine: experiments and modelling.
Proceeding Institution of Mechanical Engineers,
2001. 215(D): p. 747-760.
Heywood, J.B., Internal Combustion Engine
Fundamentals. 1988, New York: McGraw-Hill Book
Company.
Caton, J.A. and Heywood, J.B., An Experimental and
Analytical Study of Heat Transfer in an Engine
Exhaust Port. Int. J. Heat Mass Transfer, 1981.
24(4): p. 581-595.

CONTACT
Any comments and questions should be directed to
Kenneth Kar (kenneth.kar@eng.ox.ac.uk) or Richard
Stone (richard.stone@eng.ox.ac.uk). The corresponding
address is: Department of Engineering Science,
University of Oxford, Parks Road, Oxford OX1 3PJ

DEFINITIONS, ACRONYMS, ABBREVIATIONS

Temperature, K/ C

AC

Alternating current, the fluctuating

component of a signal

Time, s

Observation noise

bdcc

Bottom dead centre at the start of the

compression stroke

Observation noise covariance matrix

bmep

Brake mean effective pressure, bar

Static noise

btdc

Before top dead centre

State noise covariance matrix

DC

Direct current, zero frequency

EGT

Exhaust gas temperature

Coordinate system in the axial

direction of the thermocouple wire

EVC

Exhaust Valve Closure

Multivariate observations

EVO

Exhaust Valve Opening

FTP-75

Federal Test Procedure

Time constant ratio

imep

Indicated mean effective pressure,

bar

Emissivity

Radiation and conduction correction

factor

Smoothing parameter of the MarrHildredth filter

PSD

Power spectral density

Specific heat, kJ/kgK

Diameter of thermocouple junction, m

Latent variable vector

Noise; prediction error

Density, kg/m

Transformation matrix

Covariance of the latent variables

Frequency, Hz

Stefan-Boltzmann constant; variance

Flow matrix

Time constant, s

g(t)

Gaussian kernel

Mean squared difference between two

2
reconstructed temperatures, K

Heat transfer coefficient, W/m K

h(x)

Ramp function

i sample of the digitized data

Superscript

Identity matrix

(overbar)

Discrete Fourier transform

Thermal conductivity, W/mK

Estimated value

Kalman gain matrix

Fluctuation about the mean value

Coefficient of Reynolds number in the

Nusselt correlation

Transpose of a matrix

Covariance of the latent variables

prior the observation

Reconstructed temperature, C

th

Subscript
1

Thermocouple 1

Thermocouple 2; Thermocouple with

a nominal diameter of 0.002

23

0.002 and 0.003 thermocouple pair

25

0.002 and 0.005 thermocouple pair

Thermocouple
with
diameter of 0.003

nominal

Thermocouple
with
diameter of 0.005

nominal

den

Denominator

Gas

num

Numerator

q0

No state noise condition

surr

Surroundings

Temperature

Wire

Prediction
Latent variables