Sensors and Actuators A 161 (2010) 278–287

Contents lists available at ScienceDirect

Sensors and Actuators A: Physical

journal homepage: www.elsevier.com/locate/sna

Fluid level measurement in dynamic environments using a single ultrasonic

sensor and Support Vector Machine (SVM)
Jenny Terzic a,∗ , C.R. Nagarajah b,1 , Muhammad Alamgir a,2
a
Delphi Corporation, 86 Fairbank Road, Clayton Sth, VIC 3169, Melbourne, Australia
b
Swinburne University, Burwood Hwy, Hawthorn, VIC 3122, Melbourne, Australia

a r t i c l e i n f o a b s t r a c t

Article history: A fluid level measurement system to accurately determine fluid levels in dynamic environments has been
Received 13 August 2009 described. The measurement system is based on a single ultrasonic sensor and Support Vector Machine
Received in revised form 19 March 2010 (SVM) based signal processing and classification scheme. For exemplification of the measurement system
Accepted 11 May 2010
in dynamic environments, the novel measurement system is experimented and verified on a fuel tank of
Available online 1 June 2010
a running vehicle. The effects of slosh and temperature variations on the acoustic sensor based measure-
ment system are reduced using the novel approach. The novel approach is based on ␯-SVM classification
Keywords:
method with the Radial Basis Function (RBF) to compensate for the measurement error induced by the
Intelligent level measurement
Liquid slosh
sloshing effects in the tank due to the motion of the moving vehicle. In this approach, raw sensor sig-
Radial Basis Function nals are differentiated after smoothing with some selected pre-processing filters, namely, Moving Mean,
Support Vector Machine Moving Median, and Wavelet filter. The derivative signal is then transformed into Frequency Domain to
reduce the size of input features before performing the signal classification with SVM. Field trials were
performed on actual vehicle under normal driving conditions at various fuel volumes ranging from 5 L to
50 L to acquire sample data from the ultrasonic sensor for the training of SVM model. Further drive trials
were conducted to obtain data to verify the SVM results. A comparison of the accuracy of the predicted
fluid level obtained using SVM and the pre-processing filters is provided. It is demonstrated that the ␯-
SVM model using the RBF kernel function and the Moving Median filter has produced the most accurate
outcome compared with the other signal filtration methods in terms of fluid level measurement.
© 2010 Elsevier B.V. All rights reserved.

1. Introduction
Modern automotive vehicles are equipped with digital gauges
as well as with additional functionalities that inform drivers about
their vehicle’s fuel consumption and the remaining distance that
the vehicle can travel without refueling. However, the high preci-
sion digital displays and these additional utilities have to rely on the
accuracy of the level sensor itself. The reliability and accuracy of the
fluid level measurement system in a dynamic environment, which
primarily depends on the level sensor, is increasingly becoming a
concern for automotive industries as well as the everyday vehicle
user.
Conventional fluid level measurement systems determine the
fluid level with the use of float that is linked with a variable resistor
whose resistance is a function of the fluid level. These conventional
∗ Corresponding author. Tel.: +61 3 9239 2148; fax: +61 3 9551 8764.
E-mail addresses: jenny.terzic@delphi.com (J. Terzic), rnagarajah@swin.edu.au
(C.R. Nagarajah), muhammad.alamgir@delphi.com (M. Alamgir).
1
Tel.: +61 3 9214 8530.
2
Tel.: +61 4 2301 7656.
mechanical level sensors need to occupy a large space and suffer
from the frictional wear-out over a period of time [1]. The impor-
tance of level sensor reliability in hostile environments over long
periods of time has lead to the introduction of various forms of
motionless level sensors. Ultrasonic devices can be used in con-
tainers with pressures up to 2 mega Pascal (MPa), temperatures up
to 100 ◦ C, and depths up to 30 m, with an accuracy of approximately
2% [2]. The ultrasonic sensor is one such example of a compact as
well as contact-less proximity sensor that is being investigated to
determine the fluid level in automotive fuel tanks. The ultrasonic
sensor determines the fluid level by transmitting echo pulses and
measuring the return time of the reflected echoes. If the speed of
sound in the medium is known then the fluid level can be calculated
using the following equation:
1
level = levelref − v· (1)
2
where levelref is the height of the tank, v is the speed of the sound
and  is the time-of-flight of the ultrasonic echo (Fig. 1).
However, the speed of sound is influenced by the temperature of
the medium through which it travels [3]. Therefore, changes in the
ambient temperature will create incorrect fluid level readings. The

0924-4247/$ – see front matter © 2010 Elsevier B.V. All rights reserved.
doi:10.1016/j.sna.2010.05.005
 J. Terzic et al. / Sensors and Actuators A 161 (2010) 278–287
Fig. 1. Fluid level measurement using ultrasonic transducer.
speed of sound in air can be approximated in terms of temperature
as [4]:
v(T ) = 331.3 + kT m/s
where T is the ambient temperature in ◦ C, k is the rate at which the
speed changes with respect to the temperature, which is approxi-
mately 0.607 m/s at every change of 1 ◦ C in temperature.
Ultrasonic sensors are normally combined with a temperature
sensor to compensate for the effects of temperature variations
[5–8]. Apart from the accuracy of the level sensor itself, the fluid
level measurement system in dynamic environments, such as in
automotive fuel tanks, is influenced by the sloshing effects caused
by acceleration. In automotive fuel tanks, vehicle acceleration
induces slosh waves having natural frequencies whose wave pat-
tern is dependent on the magnitude of the acceleration, geometry
of the tank and the amount of fluid contained in the tank [9]. Theo-
retical studies and numerical analysis have been carried out in the
past to describe various sloshing phenomena [10–15].
To compensate for the effects of sloshing in fluid level measure-
ment systems, various mechanical dampening methods consisting
of baffles, electrical dampening techniques, and statistical aver-
aging methods have been used in the past. However, all these
methods follow approaches that lead to higher production cost.
The accuracy of these measurement systems under sloshing con-
ditions is also very limited. The electrical dampening techniques
and the statistical averaging methods primarily perform averag-
ing on the raw sensor signals over some period of time. Averaging
over a variable time frame has also been used in the past [16–18]
to improve the level sensor accuracy under sloshing conditions
by determining the running state of the vehicle using the vehicle
speed data from the speed sensor. The fluid measurement systems
described by Kobayashi and Kita [17] employ a vehicle speed sen-
sor to determine the running state of the vehicle. When the vehicle
is operating at low speed (i.e. near static condition), the averaging
period is reduced to small values typically in the range from 5 to
25 s, and when the vehicle is operating at a higher speed, the aver-
aging period is prolonged up to 90 s. Despite the dependence of
Fig. 2. Fuel level signal observed by the level sensor and the calculated average signal in a sample drive trial.
279
the measurement system on the speed sensor, after analyzing the
raw sensor data from a resistive type fuel level sensor in a moving
vehicle, it can be observed that the averaging method still pro-
duces significant errors after averaging the raw sensor signal over
a longer period of time. Fig. 2 illustrates the raw volume signal
obtained from a driven vehicle, and two averaged signals calcu-
lated after averaging the raw signal over 20 s, which is the typical
averaging time used in an automotive instrument cluster; and the
second curve is the averaged signal averaged over 90 s, which is a
reasonably long period of time.
Support Vector Machine (SVM) is a newly developed machine
learning algorithm [19]. SVM is based on Statistical Learning The-
ory and has the ability to recognise patterns [20]. SVM represents
novel learning techniques in the framework of structural risk mini-
mization (SRM) and in the theory of Vapnik–Chervonenkis [21] (VC)
bounds [22–24]. SVM has been successfully used in various appli-
(2) cations for solving classification, regression, time series prediction
and function estimation problems [25]. SVM has quickly found its
place in many applications of pattern recognition such as hand-
written character recognition [26], image classification [20,27], face
detection [28] and signal processing [29].
Additionally, prior to classifying the sensor signals with SVM, the
system approach described in this paper performs signal filtration
on the raw sensor signals. Three signal filters have been investi-
gated through experimentation. These investigated filters consist
of Moving Mean, Moving Median, and Wavelet filters. These filters
provide the following enhancements [30]:
• these filters remove impulse noise,
• these filters smooth the signal,
• they can be taken over a wider interval so that the filter removes
transients longer than a time instant in duration, and
• these filters preserve sharp edges of the signal curve.
SVM can also be used to predict the fluid level in a dynamic
environment, especially considering the variations in tempera-
ture and the vehicle movement creating slosh waves. This paper
describes an SVM approach developed to determine the fluid level
within a dynamic environment without compromising accuracy.
The approach described here is also applicable to non-acoustic
sensors such as capacitive and Hall-effect sensors. The existing sta-
tistical slosh compensation methods are compared with the results
obtained using the SVM approach.
2. SVM based measurement system
2.1. Concise overview of SVM
The Support Vector Machines (SVMs) construct a system model
using a set of given training samples for the classification and
280 J. Terzic et al. / Sensors and Actuators A 161 (2010) 278–287

Fig. 3. Classification of a set of two-class data using hyperplanes.

prediction of the output based on the training samples and input

samples. The basic idea behind the SVM is to create a distinction
between two or more data classes. In principle, SVM constructs a Fig. 4. Maximum-margin hyperplane between two classes of data.
hyperplane between two classes of data and then optimizes the
separation distance between the two classes to provided optimal
classification. For a given data points D containing N number of data or
points such that:
w · x − b ≥ +1 for Class B (8)
 d
 N
D= (x1 , y1 ), (x2 , y2 ), . . . , (xN , yN ) |xi ∈  , yi ∈ −1, +1 For the optimal SVM classification solution, a maximum-margin
i=1

(3) hyperplane is solved using the mathematical programming solu-

tion by minimizing w and b parameters. The optimization can be
where xi is the feature vector containing d number of features (d- expressed in terms of Lagrange multipliers ˛i as [22]:
dimensional), yi is the corresponding class of xi or expected output  
value. For example, Class A can be assigned the value of yi as ‘−1’ 1  N
min ||w||2 − ˛i [yi (w · xi − b) − 1] (9)
and Class B as ‘+1’. SVM classifies the above mentioned data points w,b,˛ 2
i=1
for the two-class solution (Class A and B) by first plotting the data
points D into the feature space. Then, SVM constructs a hyperplane This problem can now be solved using any standard quadratic
that separates the data points of one class (i.e. Class A) from other programming method. The solutions for w and b can be expressed
class (i.e. Class B). in the following forms [22]:
Fig. 3 shows an illustration of SVM based pattern classification
between two classes of data using three hyperplanes (H1 , H2 , H3 ). 
N

There are infinite number of ways to create a hyperplane. However, w= ˛i yi xi (10)

the goal in SVM is to create an optimal hyperplane that creates the i=1
best distinction line between the two classes of data. In Fig. 3, H1
1 
NSV
does not separate the two classes of data in any way, whereas, H2
b= (w · xi − yi ) (11)
and H3 do separate the two groups of data. However, H3 provides NSV
i=1
the best separation between the two class when compared with
H2 . Any hyperplane can be described as set of points x satisfying: where vector xi is the support vector points that lie exactly on the
margin hyperplanes, and NSV is the total number of support vectors
w·x−b=0 (4) (SV).
The ␯-Support Vector Machine (␯-SVM) [31] is an improved
where vector w is a normal vector perpendicular to the hyperplane
form of the soft-margin SVM that is also known as C-SVM [32,23].
and b is the bias value of the hyperplane. The offset of the hyper-
The ␯-SVM [31] includes two additional cost factors  and  [33].
plane from the origin along the normal vector w can be described
The role of these two additional parameters is to improve the clas-
by parameter b/ w.
sification accuracy. The optimization problem for ␯-SVM can be
The goal in SVM is to maximize the margin or separation
expressed as [31,32]:
between the two classes of data as far apart as possible while still
separating the two groups of data. Maximum-margin hyperplanes
1
N
1
can be described by the following equations (Fig. 4): min ||w||2 −  + i (12)
w,b,, 2 N
i=1
w · x − b = −1 (Class A) (5)
Subject to
w · x − b = +1 (Class B) (6)
yi ((w · xi ) + b) ≥  − i , i = 1, 2, . . . , N;
Hence, each data point in both classes has to satisfy the following i ≥ 0; and  ≥ 0
conditions:
where , , and  are tolerance and cost factors that improve
w · xi − b ≤ −1 for Class A (7) the accuracy of the classification system. The above optimization
 J. Terzic et al. / Sensors and Actuators A 161 (2010) 278–287 281

Fig. 5. Mapping of nonlinearly separable data sets into higher dimensional feature
space.

problem can be expressed in the dual form using the Lagrange

multipliers (˛, ˇ) as [32]:

1
N
1
L(w, , b, , ˛, ˇ, ı) = ||w||2 −  + i
2 N
i=1


N
− (˛i (yi (w · xi + b) −  + i ) + ˇi i − ı)
i=1
Some variants of the ␯-SVM have been described in Refs.
[34–36]. The description of the SVM classification and optimization
provided so far was related to linearly separable or non-separable
data sets only. In many real-life applications, the data set may
not be linearly separable, especial due to the effects of measure-
ment errors and noise [37]. To accurately classify data sets that
are nonlinear, the training data sets are mapped onto higher (or
infinite) dimensional space called feature space (Hilbert Space or
Inner Product Space). The intuition of mapping data sets into higher
dimensional space is that data can be clearly separated and hence
better classified in higher dimensional space as illustrated in Fig. 5.
SVM optimization for the data sets that are not linearly separa-
ble can be carried out simply by mapping each feature vector x with
a mapping function such that x → (x) and then carrying out the
SVM optimization described earlier [38]. For a set of training sam-
ples T = {(x1 ,y1 ), (x2 ,y2 ), . . ., (xn ,yn ) }, where x is an d-dimensional
feature vector, the mapping of the training data set T is expressed
in the feature space as:
(T ) = {((x1 ), y1 ), ((x2 ), y2 ), . . . , ((xn ), yn )}
Hence, the decision function D(xk ) for the prediction of the out-
put can be given as [37]:
 n  at time t and temperature T can be calculated as:

D(xk )=sign[w · ␾(xk ) + b] = sign ˛i yi ␾(xi ) · ␾(xk ) + b
i=1
A famous Radial Basis Function (RBF) Kernel function can be
expressed as [39]:
(xi , xj ) = exp(−||xi − xj ||2 ), >0
Fig. 6. Relationship between the (a) observed level sensor signal and (b) vehicle
acceleration.
(13)
where L0 is the tank level at still, and f is the unknown sloshing
function that depends on the acceleration effects exhibited on the
tank, the existing fluid level, and the tank geometry. The goal is
focused on determining the actual level L0 using the sensor output
L(t) and the function f. The output of the level sensor was observed
(Fig. 6) to have direct relationship with the vehicle acceleration,
when the vehicle was driven under normal driving conditions.
After observing the relationship between the acceleration of the
vehicle and the output signal L(t) from the level sensor, the effects
of slosh can be minimized by predicting the slosh wave once the
sloshing function f has been determined. Assuming the amount of
fluid in the storage tank to be constant, the actual fuel level L0 in
the storage tank under the effects of sloshing can be given as:
L(t)
L0 = = constant (18)
f
The unknown function f is solved by experimentation with the
aid of the SVM based approach. An SVM model was constructed
and trained with the actual driving data obtained through several
field trials. Fig. 7 demonstrates the method adopted to develop an
(14) accurate fluid level measurement system.
For the ultrasonic transducer mounted at a height levelref on top
of the tank, the instantaneous output of the ultrasonic level sensor
(15) (t)
Level(t, T ) = levelref − v(T ) (19)
2
where (t) is the time-of-flight at instant t of the ultrasonic echo,
and v(T ) is the speed of ultrasonic echo at temperature T. The
expression v(T ) can be obtained using Eq. (2).
(16)

2.2. SVM classification and measurement system

The observation and analysis of the slosh pattern, produced

under the effects of acceleration in a closed container, instigated an
approach that can eliminate the sloshing effects, whereby accurate
fluid level measurements would be possible in dynamic environ-
ments. If the fluid quantity in a storage container remains constant,
the instantaneous fluid level, L(t), in a dynamic environment can be
defined as:

L(t) = L0 · f (17) Fig. 7. Block diagram of the proposed system.

282 J. Terzic et al. / Sensors and Actuators A 161 (2010) 278–287

In a dynamic environment, the term (t) will exhibit variation

that reflects an inverted image of the slosh wave produced in the
liquid tank. The term (t) will be inverted since the ultrasonic sensor
is facing down measuring time-of-flight of ultrasonic echo from the
top of the tank. Expression / t describes the change in fluid level
over time. In a dynamic environment / t will exhibit the fluctu-
ation of fluid surface (slosh), however during static conditions, the
expression / t will be equal to zero. Fig. 8 shows the variation
in (t) and the actual slosh wave produced in the liquid tank.
The ultrasonic level sensor signal, denoted as s(t), is typically a
voltage signal in the range of 0.5–4.5 V, which represents the min-
imum and maximum of the level range, respectively. The sensor
signal s(t) is sampled at 100 Hz. The sampled signal is accumu-
lated in a ␻´ -second window frame (wi ). After collecting the sensor
data over ␻ ´ seconds, the ␻
´ second data is filtered using the three
investigated filters: Moving Mean, Moving Median and Wavelet
Transform. Then the signal features are extracted using the famous
Fast Fourier Transform (FFT) based feature extraction method. The
coefficients (coef) obtained from the FFT based feature extraction
function, the median value (med) of the ␻ ´ -second ultrasonic sensor
signal, and the temperature readings T are all contained in a vector
forming input features for the SVM model. The median value of the
raw signal describes the middle (bias) factor of the slosh wave and
hence was decisively added into the input feature vector. The SVM
input vector xi can be represented as:

xi = {coef1 , coef2 , . . . , coefn , med, T } (4.4)

Signal classification is performed using the ␯-SVM signal clas-

Fig. 8. Illustration of (a) the time-of-flight signal, (b) the exhibited slosh wave, and
sification technique having the Radial Basis Function (RBF) as the (c) the frequency spectrum of the level signal.
kernel function described earlier.

Fig. 9. Illustration of the Moving Mean and Moving Median filters.

 J. Terzic et al. / Sensors and Actuators A 161 (2010) 278–287
Fig. 10. Wavelet filter applied on the raw signal.
2.3. Signal enhancement
The raw sensor signals were smoothened or filtered to remove
high-frequency noise and spikes before processing the sensor sig-
nals through the SVM based signal processing method described
previously. The Moving Mean and Moving Median filters slide
across the raw signal and calculate the mean/median values in
the neighbouring sampled points. If x is the sampled raw signal of
length N, and w is size of the moving window of the filter, then the
filtered output y using mean and median can be obtained using Eqs.
(20) and (21), respectively. The width of the moving window w is
set to 20. Therefore, the sliding window takes 20 sampled values of
the raw signal and produces a mean or median value at the output:
y[i] = mean(x[i − 1], x[i − 2], . . . , x[i − w]), for w ≤ i ≤ N
(20)
y[i] = mean(x[1], x[2], . . . , x[i]), for 1 ≤ i < w
y[i] = median(x[i − 1], x[i − 2], . . . , x[i − w]), w≤i≤N
(21)
y[i] = median(x[1], x[2], . . . , x[i]), for 1 ≤ i < w
Fig. 12. Block diagram of the SVM system for training and verification of the experimental data.
283
Fig. 11. Typical speed and observed during the experiments for each field trials.
The value of N for the 20 s signal at 100 Hz is calculated as:
N = 100 samples/s × 20 s = 2000 samples (22)
Fig. 9 illustrates the signal output when the Moving Mean and
Moving Median filters are applied to the raw signal data. The filtered
versions of the raw signal do not contain high-frequency noise.
The third filter investigated in the experiment is the Wavelet
Transform (WT) filter that analyses signals at different frequency
bands by de-composing them into coarse information and detailed
information sets. The coarse information set contains the low-
frequency components, and the detailed information set contains
the high-frequency components of the input signal. Only low-
frequency components, which reflect a smoothened version of the
raw signal, are used and the high-frequency components of the raw
signal, which usually contain noise, are eliminated. Fig. 10 shows
the high frequency signal (b) and the low-pass filtered signal (c)
when the raw sensor signal (a) was processed with the Discrete
Wavelet Transform (DWT) function. The Wavelet Transformation
was processed through the MATLAB Wavelet Toolbox [40] using
the dwt function with Daubechies [41] Wavelet (db1).
3. Experimental setup
A fuel tank was fitted with an ultrasonic sensor near the top
center of the tank. The tank can be approximated as a rectangular
cuboid with dimensions 34 cm × 34 cm × 81 cm. The fuel tank was
284 J. Terzic et al. / Sensors and Actuators A 161 (2010) 278–287

Table 1
List of tank volumes investigated in the experiment.

Investigated tank levels

5 L, 6 L, 7 L, 8 L, 9 L, 15 L, 20 L, 25 L, 30 L, 35 L, 36 L, 37 L, 38 L, 39 L, 40 L, 45 L,
46 L, 47 L, 48 L, 49 L, 50 L

Table 2
Techinical details of the ultrasonic transducer used in the experiment.

Ultrasonic sensor specifications

Accuracy ±0.32 cm
Output 0.5–4.5 V (min–max)
Resolution 0.18 cm
Operating Temperature −40 to 80 ◦ C
Designed for gasoline and diesel liquids

filled with fuel levels ranging from 5 to 50 L in the experiment,

which corresponds to 6–70% of the tank capacity. The fuel tank
was mounted in latitudinal direction, where the longest length of
the tank was in parallel to the direction of the vehicle. Table 1 lists
all the fuel levels investigated in the experiment.
Each experiment was conducted by driving a vehicle containing
the instrumented fuel tank for 3 km in a suburban residential area,
where occasional stops were made at some road intersections. Fur-
thermore, all field trials were carried out on the same travel route
(Fig. 11).
Table 2 lists the technical specifications of the ultrasonic sensor
used in the experiments. The level signal from the ultrasonic sensor
was acquired using the LabVIEW software and a Data Acquisition
Card, which was connected to the ultrasonic sensor in the vehicle.
The fuel level signal indicated by the ultrasonic sensor output was
sampled and recorded at 100 Hz. Fig. 12 illustrates an overview of
the experimental setup.

Table 3
Training parameters used by the LIBSVM train tool.

Parameter Value Fig. 13. SVM training and validation flowchart.

SVM type =1 (␯-SVM)

Kernel function =2 (RBF) 4. Validation results
Gamma () =0.125
Nu () =0.05 Each investigated filter was applied on a separate SVM model,
Tolerance () =10−5
where the SVM parameters were the same. One half of the data

Fig. 14. Graph of the frequency coefficients obtained from the training data.
 J. Terzic et al. / Sensors and Actuators A 161 (2010) 278–287 285

Fig. 15. Validation result generated after training and validating the SVM models for each investigated drive trial.

from the first field trials was used to train the SVM model, and half
of the samples obtained from the second field trials were used for
verification of the network performance.
The training and validation process is shown in Fig. 13, which
was carried out using MATLAB and LIBSVM [42] software appli-
cations. LIBSVM [42] is an integrated software for support vector
classification, regression and distribution estimation. It supports
multi-class classification, which is required for training the ultra-
sonic signals at multiple volume levels. The signal features were
scaled between an optimum range (0–1) using the LIBSVM scale
tool.
Table 3 lists the training parameters used to train the SVM mod-
els using the LIBSVM train tool. These parameters were selected
experimentally using the LIBSVM tool until a higher classification
rate was achieved.
Fig. 14 shows the frequency coefficients of the unfiltered sig-
nals obtained using the MATLAB built-in fft function. Trunk [43] has
demonstrated that data can be detrimental to classification, espe-
286 J. Terzic et al. / Sensors and Actuators A 161 (2010) 278–287

Table 4
Relative average error using statistical averaging methods and SVM based system.

Actual volume Statistical averaging Support Vector Machines

Moving Moving ␯-SVM ␯-SVM ␯-SVM ␯-SVM

Mean Median (unfiltered) (Moving (Moving (Wavelet
(without (without Mean) Median) filter)
SVM) SVM)

50 11.5% 10.4% 0.0% 0.0% 0.0% 0.0%

49 10.1% 9.3% −1.7% −1.7% −1.7% −1.7%
48 2.7% 2.3% −2.1% −1.8% −2.8% −2.2%
47 −5.3% −5.3% 0.9% 0.6% 0.8% 0.9%
46 −0.7% −0.9% −0.6% −0.5% −0.5% −0.6%
45 −2.0% −2.6% −0.5% −0.5% −0.5% −0.5%
40 0.1% −0.1% −0.2% −1.1% 0.0% −0.2%
39 −4.5% −4.6% 0.5% 0.4% 0.5% 0.4%
38 0.5% −0.7% 0.9% 0.9% 0.9% 0.9%
37 0.9% 0.0% 4.8% 7.1% 4.8% 5.8%
36 −2.0% −2.5% −1.0% −0.8% −0.6% −1.0%
35 1.1% 0.1% 0.6% 0.8% 0.2% 0.6%
30 0.4% −0.6% 1.2% 3.6% 2.4% 1.2%
25 8.9% 6.0% 6.9% 5.4% −1.4% 6.9%
20 14.5% 10.6% 16.1% 16.1% 3.2% 16.1%
9 40.6% 35.4% −0.8% −0.8% −0.8% −0.8%
8 22.9% 17.1% −1.8% −1.8% −2.7% −2.7%
7 4.5% 0.7% 1.0% 1.0% 0.0% 1.0%
6 −0.1% −1.6% 0.0% 0.0% 0.0% 0.0%
5 6.4% 1.6% 1.4% 2.9% 1.4% 1.4%

Abs. average error 7.0% (1.53 L) 5.6% (1.33 L) 2.2% (0.57 L) 2.4% (0.63 L) 1.3% (0.39 L) 2.2% (0.59 L)
Max. error 40.6% (5.77 L) 35.4% (5.19 L) 16.1% (3.21 L) 16.1% (3.21 L) 4.8% (1.79 L) 16.1% (3.21 L)

cially if the additional data is highly correlated with previous data

[44]. Apart from the correlation of data, the size of input feature data
is also important in signal classification systems. Thus, an increase
in the input feature dimension ultimately causes a decrease of per-
formance [45]. Hence, the correlation of the input data and the
number of input features were decisively selected during the devel-
opment of the SVM model. It was observed through experiments
that only the first sixty-three frequency coefficients that approxi-
mately correspond to the slosh frequency 0–6.5 Hz has significant
magnitude and hence only this range of data was used for classifica-
tion. The frequency coefficients, the median value of the signals, and
the temperature values were all bundled in an array of sixty-five
elements, which were finally fed into the ␯-SVM.
After training the SVM models, the models were validated using
the test samples obtained from the second field trial. Fig. 15 shows
the output results for selected (lower and higher) tank volumes. The
output results were obtained after processing the ultrasonic sensor Fig. 16. Maximum and average error figures of the overall averaged error for differ-
signals with different processing methods. The time length of each ent signal processing methods.
trial is indicated as 280 s. The graphs in Fig. 15 show fuel volumes
averaged over the whole drive period of 280 s, after processing the 5. Conclusion
signals through different processing methods. To describe the steps
undertaken to obtain the overall averaged volume, a closer look at The Support Vector Machines based signal processing and clas-
the investigated 49 L trial is also shown in Fig. 15. The raw signal sification approach coupled with a signal ultrasonic sensor has been
illustrated in Fig. 15(A) was divided into 20-s long signals, as shown used to accurately determine the fuel level in an automotive fuel
in Fig. 15(B), which were then filtered and processed through the tank under dynamic conditions. Four identical SVM models were
SVM. The overall averaged volume as shown in Fig. 15(C) was cal- developed and an investigation was carried out by applying three
culated by averaging the SVM model outputs for each trial over the filtration methods and keeping one unfiltered raw signal to analyze
whole 280 s period. the performance of the ␯-SVM model in improving the accuracy of
Table 4 shows the relative average error figures for all investi- the level sensor in the presence of liquid slosh. The SVM model
gated volumes computed using the statistical averaging methods, applied with the Moving Median filter produced a maximum aver-
and the SVM based signal classification method having different aged error of 1.8 L, which is significantly better than the results
pre-processing filters. obtained using the statistical and non-SVM network Moving Mean,
A comparison of the accuracy of different processing techniques and Moving Median functions that produced a maximum averaged
investigated in the experiment is shown in Fig. 16. The overall error of 5.8 L and 5.2 L, respectively. The increased accuracy of the
results obtained from the SVM based system indicate significantly fuel level measurement system in dynamic environment will pro-
less error in fuel volume measurement compared with the simple vide more confidence to drivers regarding the actual amount of
averaging methods under dynamic conditions. The ␯-SVM model fuel indicated by the instrument panel. With the accurate fuel level
with the RBF kernel function and applied Moving Median filter pro- measurement system, the distance-to-empty figures can be accu-
duced the most accurate results compared with the other methods. rately computed. In particular the SVM based method is suitable
 J. Terzic et al. / Sensors and Actuators A 161 (2010) 278–287
for use in a professional car racing where vehicle is subjected to
a highly dynamic maneuvers. Drivers of cars equipped with this
measurement method can confidently drive higher number of laps
without fear of running out of fuel in situations where fuel level in
the tank is low.
6. Future work
An ultrasonic sensor coupled with the Support Vector Machine
(SVM) approach to signal processing will be used to address other
influencing factors such as atmospheric pressure and the tilt that
causes liquid to shift to one side. With the rapid improvements
in microprocessor technology, it will be possible to automatically
train the SVM model in real time, which will further increase
the effectiveness of the measurement system in dynamic environ-
ments.
References
[1] A.C. Fischer-Cripps, Force, pressure and flow, in: Newnes Interfacing Compan-
ion, Newnes, Oxford, Boston, 2002, pp. 54–70.
[2] W.C. Dunn, Level measurement, in: Introduction to Instrumentation, Sensors
and Process Control, Artech House, Boston, 2005, pp. 115–126.
[3] H. Kuttruff, Ultrasonics—Fundamentals and Applications, Elsevier Applied Sci-
ence, 1991.
[4] R.A. Serway, J.W. Jewett, Sound waves, in: Physics for Scientists and Engineers,
6th ed., Thomson-Brooks/Cole, Belmont, CA, 2004, pp. 512–542.
[5] Z.-Y. Song, C.-Y. Liu, X.-L. Song, Application research of information fusion tech-
nology of multi-sensor in level measurement, in: International Conference on
Machine Learning and Cybernetics, vol. 6, 2004, pp. 3511–3514.
[6] Z. Song, C. Liu, X. Song, Y. Zhao, J. Wang, A virtual level temperature compen-
sation system based on information fusion technology, in: IEEE International
Conference on Robotics and Biomimetics, 2007, pp. 1529–1533.
[7] D.C. Gazis, W.F. Kane, R.J. von Gutfeld, inventors, International Business
Machines Corporation, assignee, Ultrasonic liquid level gauge for tanks subject
to movement and vibration, Patent No. 5,793,705, 1996.
[8] C.M. Combs, P.H. Goodwin Jr., inventors, Robertshaw Controls Company,
assignee, Adjustable ultrasonic level measurement device, Patent No. 4221004,
1978.
[9] R.A. Ibrahim, Liquid Sloshing Dynamics: Theory and Applications, Cambridge
University Press, Cambridge, UK, New York, 2005.
[10] K. Modaressi-Tehrani, S. Rakheja, R. Sedaghati, Analysis of the overturning
moment caused by transient liquid slosh inside a partly filled moving tank,
Proceedings of the Institution of Mechanical Engineers, Part D: Journal of Auto-
mobile Engineering 220 (3) (2006) 289–301.
[11] L. Dai, L. Xu, A numerical scheme for dynamic liquid sloshing in horizontal
cylindrical containers, Proceedings of the Institution of Mechanical Engineers,
Part D: Journal of Automobile Engineering 220 (7) (2006) 901–918.
[12] S. Wiesche, Computational slosh dynamics: theory and industrial application,
Computational Mechanics 30 (5) (2003) 374–387.
[13] N.C. Sinha, S.K. Pal, R.K. Bhattacharyya, Experimental investigation of slosh
dynamics of liquid-filled containers, Experimental Mechanics 41 (2001) 63–69.
[14] H.-S. Kim, Y.-S. Lee, Optimization design technique for reduction of sloshing
by evolutionary methods, Journal of Mechanical Science and Technology 22
(2008) 25–33.
[15] D. Reed, J. Yu, H. Yeh, S. Gardarsson, Investigation of tuned liquid dampers
under large amplitude excitation, Journal of Engineering Mechanics 124 (4)
(1998) 405–413.
[16] H. Kobayashi, H. Obayashi, inventors, Nissan Motor Company, Limited,
assignee, Fuel volume measuring system for automotive vehicle. Patent No.
4611287, 1983.
[17] H. Kobayashi, T. Kita, inventors, Nissan Motor Company, Limited assignee, Fuel
gauge for an automotive vehicle, Patent No. 4470296, 1982.
[18] T. Guertler, M. Hartmann, K. Land, A. Weinschenk, inventors, DAIMLER BENZ AG
(DE) assignee, Process for determining a liquid quantity, particularly an engine
oil quantity in a motor vehicle, Patent No. 5831154, 1997.
[19] S. Zhao, Remote sensing data fusion using support vector machine, in: IEEE
International Geoscience and Remote Sensing Symposium Proceedings, vol. 4,
2004, pp. 2575–2578.
[20] R.A. Reyna, D. Esteve, D. Houzet, M.-F. Albenge, Implementation of the SVM neu-
ral network generalization function for image processing, in: IEEE International
Workshop on Computer Architectures for Machine Perception (CAMP’00),
2000.
[21] V.N. Vapnik, Statistical Learning Theory, Wiley, New York, 1998.
[22] V.N. Vapnik, The Nature of Statistical Learning Theory, Springer, New York,
London, 1995.
287
[23] Y.H. Hu, J.-N. Hwang, Handbook of Neural Network Signal Processing, CRC Press,
Boca Raton, 2002.
[24] V. Kecman, Learning and Soft Computing: Support Vector Machines, Neural
Networks, and Fuzzy Logic Models, MIT Press, Cambridge, MA, 2001.
[25] J. Zhang, X. Liu, J. Liu, F. Peng, J. Tian, Y. Wang, W. Zhang, M. Xie, SVM-based ultra-
sonic medicine image diagnosis, Medical Image Acquisition and Processing
4549 (2001) 92–95.
[26] A.R. Ahmad, C. Viard-Gaudin, M. Khalid, E. Poisson, Online handwriting recog-
nition using support vector machine, in: IEEE Region 10 Conference 2004,
2004.
[27] K.M. Rajpoot, N.M. Rajpoot, Wavelets and support vector machines for texture
classification, in: Proceedings of International Multitopic Conference (INMIC),
2004, pp. 328–333.
[28] J. Huang, X. Shao, H. Wechsler, Face pose discrimination using support vector
machines (SVM), in: International Conference on Pattern Recognition, vol. 1,
1998, pp. 154–156.
[29] A. Boni, L. Gasparini, R. Pianegiani, D. Petri, Low-power and low-cost implemen-
tation of SVMs for smart sensors, in: IEEE Instrumentationand Measurement
Technology Conference Proceedings, vol. 1, 2005, pp. 603–607.
[30] R.L. Allen, D.W. Mills, Time-domain signal analysis, in: Signal Analysis: Time,
Frequency, Scale, and Structure, IEEE Press, Wiley–Interscience, Piscataway, NJ,
2004, p. 322.
[31] B. Schölkopf, A. Smola, R. Williamson, P. Bartlett, New support vector algo-
rithms, Neural Computation 12 (5) (2000) 1207–1245.
[32] P.-H. Chen, C.-J. Lin, B. Schölkopf, A tutorial on v-support vector machines,
Applied Stochastic Models in Business and Industry 21 (2) (2005) 111–136.
[33] S. Abe, Two-class support vector machines, in: Support Vector Machines for
Pattern Classification, Springer, London, 2005, pp. 15–82.
[34] C.-C. Chang, C.-J. Lin, Training v-support vector classifiers: theory and algo-
rithms, Neural Computation 13 (9) (2001) 2119–2147.
[35] M.A. Davenport, The 2nu-SVM: A Cost-Sensitive Extension of the
nu-SVM, Journal [serial on the Internet], 2005 Date, Available from:
http://dspace.rice.edu/handle/1911/19831.
[36] A new nu-support vector machine for training sets with duplicate samples, in:
Y.-S. Jia, C.-Y. Jia, H.-W Qi (Eds.), 2005 International Conference on Machine
Learning and Cybernetics, Guangzhou, 18–21 August, 2005.
[37] O. Ivanciuc, Applications of support vector machines in chemistry, Reviews in
Computational Chemistry 23 (2007) 291–400.
[38] B. SchNlkopf, A.J. Smola, Learning with Kernels: Support Vector Machines, Reg-
ularization, Optimization, and Beyond, MIT Press, Cambridge, MA, 2002.
[39] C.-W. Hsu, C.-C. Chang, C.-J. Lin, A Practical Guide to Support Vector Classifica-
tion, 2008.
[40] M. Misiti, Y. Misiti, G. Oppenheim, J.-M. Poggi, Wavelet Toolbox 4—Users Guide,
MathWorks, 2009.
[41] I. Daubechies (Ed.), Ten Lectures on Wavelets, Society for Industrial and Applied
Mathematics, Philadelphia, PA, 1992.
[42] C.-C. Chang, C.-J. Lin, LIBSVM: a library for support vector machines. Software
available at http://www.csie.ntu.edu.tw/∼cjlin/libsvm, 2001.
[43] G.V. Trunk, A problem of dimensionality: a simple example, Pattern Analysis
and Machine Intelligence PAMI-1 (3) (1979) 306–307.
[44] O. Bousquet, U. Luxburg, G. Rätsch, Machine learning summer school, in: O.
Bousquet, U. von Luxburg, G. Rätsch (Eds.), Advanced Lectures on Machine
Learning: ML Summer Schools 2003, Canberra, Australia, February 2–14, 2003,
Tübingen, Germany, August 4–16, 2003, revised lectures, Springer, Berlin, New
York, 2004.
[45] F. van der Heijden, R.P.W. Duin, D. de Ridder, D.M.J. Tax, Feature extraction
and selection, in: Classification, Parameter Estimation, and State Estimation:
An Engineering Approach using MATLAB, Wiley, Chichester, West Sussex, Eng.,
Hoboken, NJ, 2004, pp. 183–14.
Biographies
Jenny Terzic Quality/Six Sigma Manager at Delphi Automotive Systems Australia
– Powertrain Division. Holds Bachelor of Mechanical Engineering (with honours)
and Master of Engineering in Computer Integrated Manufacturing. Has extensive
experience in vehicle fuel system design, development and testing. Area of research
include: neural networks, support vector machines, intelligent sensors and applica-
tions of artificial intelligence in automotive vehicle engineering.
Romesh Nagarajah Professor of Mechanical Engineering at Swinburne University
of Technology. Graduated with an Honours degree in Mechanical Engineering from
the University of Ceylon in 1973, M.Phil. in Robotics from the University of Notting-
ham in 1985 and a Ph.D. from Swinburne University of Technology in 1994. He has
received several national competitive grants, named in several international patents
and published 75 papers in international journals and conferences.
Muhammad Alamgir Electrical Engineer at Delphi Automotive Systems Australia.
Graduated in Computer Systems Engineering from RMIT University in 2008. Areas
of research include smart sensors, artificial neural networks and support vector
machines.