You are on page 1of 8

2008 ASME International Mechanical EngineeringProceedings

Congress and
of Exposition
IMECE2008
IMECE2008
2008 ASME International Mechanical Engineering Congress and Exposition
October 31 –November 6, 2008, Boston, Massachusetts, USA
October 31-November 6, 2008, Boston, Massachusetts, USA

IMECE2008-67260

MODEL BASED SYSTEM MONITOR USING ARTIFICIAL NEURAL NETWORKS (ANN)


Albert Albers1 Jiangang Wang

Sascha Ott Tobias Düser Sarawut Lerspalungsanti

Institute of Product Development


Department of Mechanical Engineering
University of Karlsruhe,
76131 Karlsruhe, Germany

ABSTRACT One of the most frequently used way for fault detection in
Nowadays, with the integration of diverse mechatronic this case is threshold checking. This method supports the
components, a mechatronic system is capable of performing system with basic functional safety, and requires little
more difficult and complex tasks. At the same time, it calculation effort. The targets of the threshold checking are
becomes more and more difficult to develop a system to measureable system parameters, like force F, torque T,
ensure the safety of these complex mechatronic systems. rotational speed n, temperature T, pressure P, etc. The state of
the system is monitored with the method. Furthermore, in an
This paper outlines a novel method for fault detection, automation system it is necessary to ensure the functionality of
which uses a neural network monitor based on a feed forward the mechatronic components. By means of timing constraint a
back propagation (FFBP) algorithm. The goal of this study is monitor can determine, whether a required action is
to develop a real-time capable method for detection of accomplished within a defined time interval. In order to have
abnormal or faulty behavior of an automation system. An an overview of fault-detection methods, a summarization of
example is used to illuminate the capability of the neural the practical fault-detection methods is shown in Figure 1.
network in modeling and real-time data in this work.
Furthermore this method is validated on a test bed, which is FAULT‐DETECTION METHODS
developed at the institute of product development Karlsruhe
(IPEK). This method will allow the monitor to detect critical detection with detection with multiple 
single signals signals and models
state of the system due to unexpected influences in real-time,
so that safety hazard to personnel and equipment caused by limit trend signal process multi‐variant 
checking checking models used models used data analysis
the fault can be prevented or mitigated.
parameter neural principal
Keywords: ANN, model based system monitor, reliability fixed
threshold
adaptive 
thresholds
correl‐
ation
spectrum
analysis estimation networks component
analysis
engineering, change state limit
wavelet
detection observers checking
analysis
methods
INTRODUCTION parity
equations
Due to the development of automation systems,
tremendous progress has been achieved in manufacture,
transportation and energy systems. With the increasing Figure 1. Survey on fault-detection methods [1]
integration of mechatronic components in technical systems,
the correct behaviors of all actuators, sensors and control units The methods for fault detection can be divided into the
are more and more critical to the environment, business and following parts:
the safety of operators and equipments. Hence, the • fault detection with single signals
functionality and the reliability of the automation system - through limit checking
should be assured during the operation process. - through trend checking

1
Professor and author of correspondence, Phone: +49 (721) 608 2371, Email: albers@ipek.uni-karlsruhe.de

1 Copyright © 2008 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 06/24/2018 Terms of Use: http://www.asme.org/about-asme/terms-of-use


- using signal models of the complex mechatronic systems. The development on
- through timing constraint monitoring system shows that model-based monitor can better
• fault detection with multiple signals and models manage the requirement of complex systems.
- using signal models By building a model-based monitor it is necessary to have
- using process models a holistic analysis of the system. In order to determine the
- through multi-variant data analysis vehicle longitudinal dynamics, a simplified vehicle for
In a complex mechatronic system the interaction between demonstration of the tractive resistance [5] is shown in Figure
the different components must be considered within the 3.
monitor. Consequently the method of the fault detection with
single signals will not be sufficient to ensure the functionality FA
of the whole system. Hence, a model based fault detection v
method is studied in this work. FG s
A model based system monitor determines the system
status through a model describing the relations between the
input signals U and output signals Y. According to Isermann FR
[2] concept of the model based system monitor is shown in α m·g
Figure 2.

failure Figure 3. Driving resistance


n The driving torque provided by the combustion engine
U Y will be expended to overcome the tractive resistance FTR and
system + in the meantime to accelerate the vehicle. A simplified relation
between tire tractive (braking) force Ft, vehicle speed, and
tractive resistance is described in equation (1), where nW
denotes number of driving wheels (during acceleration) or the
total number of wheels (during braking).
model based
failure detection mv&(t ) = nW ⋅ Ft (t ) − FTR (t ) (1)
The tractive resistance is composed of air resistance FA,
symptom rolling resistance FR and gradient resistance FG.
FTR (t ) = FA (t ) + FR (t ) + FG (t ) (2)

Figure 2. Concept of model based system monitoring [2] The FA, FR and FG can be calculated with equations (3) [5]:
v2
In practical application it is required that the monitor is FA = cw ⋅ A ⋅ ρ ⋅
capable of online data processing. The development of a real- 2 (3)
time capable system model for the model based monitor FR = f R ⋅ m ⋅ g
becomes a focus of this study. This work presents feed FG = m ⋅ g ⋅ sin α
forward back propagation Network (FFBP) to build the
system. The object of the monitor in this study is a limousine Where the cW, A, ρ, fR, g and α denote drag coefficient,
with front-engine rear-wheel-drive, 2.5 liter 6 cylinder cross-sectional area, atmospheric density, rolling resistance
gasoline engine, 5-speed manual transmission and friction coefficient, acceleration due to gravity and road gradient.
clutch. The goal of the actual work is to monitor the vehicle’s The tire tractive (braking) force Ft in equation (1) is the
start up process. Simple descriptions of vehicle dynamic and force on the contact patch of tires. It is the average friction
test process are introduced in section 2. Section 3 describes the force of the driving wheels for acceleration and the average
method of monitoring vehicle longitudinal dynamic and the friction force of all wheels for deceleration.
ANN based vehicle model. The validation of the model is
described in Section 4, followed by conclusions. Considering the functions of the clutch, brake and
transmission systems in the drive train, it is difficult to
VEHICLE MODEL AND TEST PROCEDURE describe the vehicle dynamic with simple equations. An
Vehicle Longitudinal Dynamics Model alternative method to describe vehicle longitudinal dynamic is
The integration of diverse mechatronic components in the to use computer aided simulation technique. Figure 4 shows
vehicle has enhanced the driving safety and ride comfort. the demonstration of a model structure for simulation of
Concerning the complexity of the mechatronic components in vehicle longitudinal dynamic. The difficulties by modeling a
the vehicle, a monitor system is necessary to ensure the vehicle system with non-linear physical based models are:
functionally and reliability of these components. Methods for • it needs great effort to determine parameters in the
monitoring or checking software and hardware performance vehicle model, and
[3], [4] can hardly meet the increasing reliability requirements

2 Copyright © 2008 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 06/24/2018 Terms of Use: http://www.asme.org/about-asme/terms-of-use


• the simulation of interface, where an energy and link also has a numeric weight Wj,i associated with it, which
material transport is possible, are still not clearly determines the strength and sign of the connection [8]. ANN
defined or standardized, for instance, the models of based model of the vehicle longitudinal dynamic will be
clutch and tire. introduced in section 3.

ai = g (∑ j =0 W j ,i a j )
On the other hand, because of the complexity of the n
model, the vehicle model requires great calculation effort. It Bias Weight
can be difficult to satisfy the real-time requirement of an on- a0= -1 W0,i
g
line monitoring system. ini

Structure of a simplified vehicle model aj


Wj,i
∑ ai

-C-
Clutch
Input Input Activation Output Output
Pedal n_Gearbox_in
Links Function Function Links
-C- Acc Pedal Clutch Pedal T_Clutch
Accelerator n_Engine n_DMF_prim.
Pedal T_clutch
DMF & Clutch
Figure 5. A simple mathematical model for a neuron.
Engine
The unit’s output is ai, and aj is the output activation of
unit j and Wj,i is the weight on the link from unit j to this
-C- N_of_ratio n_Gearbox_in unit [8]
Number of
Ratio steps T_Clutch n_Gearbox_out Test procedure
Gearbox The development of an ANN based vehicle model follows
the process of VDI 2206 [9], which is shown in Figure 6.

requirements product
n_G_out n_Drivetrain

Drive train

v_vehicle

n_Drivetrain F_Drive assurance of


Brake Pedal
properties
-C-
Brake Wheels & Tires
Pedal Brake control Unit

F_Drive
v_vehicle
-C- Slope
Slope
Vehicle Mass
domain‐specific design
Front-engine real -wheel -drive , gasoline engine, mechanical engineering
manual transmission and friction clutch, electronical engineering
information technology
Figure 4. Demonstration of simplified vehicle model

The investigation shows that the vehicle velocity is a Figure 6. V-model as macro cycle [9]
function of accelerator pedal travel PA, brake pedal travel PB,
clutch pedal travel PC, engine speed nEngine, road gradient α, With the analysis of the vehicle system, as shown in
number of ratio steps i, etc. section 2.1, it is possible:
v = f ( PA , PB , PC , n Engine ,α , i,...) (4) • to obtain an overview of the system,
The preliminary studies [6] [7], show that among the • to determine parameters which are relevant to the
methods for modeling the vehicle longitudinal dynamics, network,
ANN based models are more efficient compared to other
models, such as non-linear physical based models or • to determine characteristics of the parameters. For
mathematical approximate models. instance, a state parameter like number of ratio steps i,
describes actual selected gear ratio, and parameter like
ANN is a mathematical model, which is based on brake pedal travel PB, influences the system
biological neural network. It is composed of nodes, or units progressively. The influence can be described as
connected by links, as shown in Figure 5. A link from unit j to integration of PB.
unit i serves to propagate the activation aj from j to i. Each

3 Copyright © 2008 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 06/24/2018 Terms of Use: http://www.asme.org/about-asme/terms-of-use


t
E Influrence = const . ⋅ ∫ P dt
t =t0
B
(5) MONITOR CONCEPT AND VEHICLE MODEL
To have an overview of the working principle of the
After the determination of the input, output parameters monitor, Figure 7 describes a structure of the model based
and the structure of the network, it is necessary to train the vehicle longitudinal dynamic monitor. The driver perceives
network with measured datum. the environment and the vehicle behavior and controls the
vehicle dynamic. The model based monitor predicts the
For the network training process, vehicle start up tests are vehicle dynamic with a real-time capable vehicle model, and
performed. The sample frequency of measured datum is 100 compares the predicted dynamic with the measured dynamic.
Hz. During the tests the 1st gear is permanently selected. By After the comparison process the monitor can identify,
designing the test procedure, there are combinations of five whether the vehicle is in a critical state, which can be caused
parameters to be considered, namely by conditions such as:
• accelerator pedal travel PA or • defective components in the vehicle
• engine speed nEngine, • faulty operation of the components in the vehicle
• brake pedal travel PB, • unanticipated states in the vehicle
• clutch pedal travel PC and • fault of sensors in the vehicle
• rate of change of the three parameters.
Vehicle
In consideration of variation of the pedal travel PB, and
PC, the values of these parameters are divided into five grades:
0%, 25%, 50%, 75% and 100% of pedal travel. The engine
speed at the beginning of each test are 800 rpm (idle state), Comparing
1600 rpm, 2400 rpm or 3200 rpm. The rate of change of the Driver Environment
model
parameters can be described as “very slow”, “slow”, “fast”,
“very fast”, as shown in Table 1. In this study it is assumed
that the accelerator pedal and the brake pedal cannot be Vehicle model
performed at one time. The possible combinations of the (longitudinal)
factors are shown in Table 2. There are more than 100
variations of test segments in the test procedure. In order to
Figure 7. Working principle of the model based vehicle
design the tests with more efficiency and little requirements on
longitudinal dynamic monitor [6]
resource, the method of design of experiments (DoE)
introduced in [10], [11] is considerable.
Name Grade
Accelerator pedal travel PA 0%, 25%, 50%, 75%, 100%
Brake pedal travel PB 0%, 25%, 50%, 75%, 100% monitor
environment
Clutch pedal travel PC 0%, 25%, 50%, 75%, 100%
Engine speed nEngine 800 rpm, 1600 rpm, 2400 rpm, 3200 rpm
Rate of change Hold, very slow, slow, fast, very fast
z0
y0
Table 1. Parameter grades in the test procedure o x0

nEngine PA PB PC Rate
[rpm] [%] [%] [%] [-]
1600 - 0 - 0
Inputs Layer 1 Layer 2

… … … … … p1(t) a1(t) a 2(t)

1600 - 25 slow 0
T
D IW1,1 n1(t) LW2,1 n2 (t)
L S1 x 1 S2 x 1
S1 x f1 S 2 x S1 f2
d (R1d)

… … … … … R1
1
S1 x
b1
1
S1 x 1

S1
1 b2
S2 x 1
S2 x 1

S2

1600 25 - slow PC & 25


hold PA 0.1

0.08
Error in Validation

Error

monitor
… … … … … 0.06

0.04

4000 - 25 0 slow
0.02
Error [-]

-0.02

… … … … … -0.04

-0.06

Table 2. Combination of the parameters at the start point -0.08

-0.1
284 285 286 287 288 289
t [s]
290 291 292 293 294

of each test segment


Figure 8. Monitor and monitor environment

4 Copyright © 2008 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 06/24/2018 Terms of Use: http://www.asme.org/about-asme/terms-of-use


As shown in Figure 8 monitor environment is composed
of the driver, vehicle, vehicle environment, and the interplay u(t)
between these elements. The monitor obtains the information
from the sensors which are mainly installed in the vehicle. The Feed
model in the monitor describes relation between the monitor u’(t) Forward y(t)
environment and the vehicle longitudinal dynamic. For network
instance, the following influences are considered: y’(t)
• the influence of vehicle speed v on tractive resistance
FTR, Figure 10. Network structure of the vehicle model [7]
• the influence of the clutch pedal travel PC on the
vehicle dynamic, EXPERIMENTAL RESULTS
Training results
• the influence of gas pedal travel PA and brake pedal After the designing of the network, the initialized network is
travel PB on tire tractive (braking) force Ft. trained with the measured datum from drive tests. Training
The monitor processes the input information with its process is necessary for the network to update its weights. The
ANN model, and returns simulated vehicle dynamic. In the data are normalized before they are sent to the network. To
mean time the measured vehicle real dynamic is delivered to obtain optimal results the training epoch of this network is set
the monitor. The difference between the simulated and the to 900 epochs, and in order to reduce the memory cost during
measured dynamic during the training process indicates the the training process parameter “net.trainParam.mem_reduc” is
simulation quality of the model. In practical application if the set to 10. An introduction about commonly used methods for
difference exceeds safety thresholds which are denoted as red training of the networks are presented by Demuth [13]. The
lines in Figure 8, an alarm will be triggered to indicate the result of the training process is illustrated in Figure 11. The
abnormal working state of the vehicle. blue dotted line denotes the simulated shaft speed on the
gearbox output shaft, the green line denotes the measured
The model is a multilayer feedforward backpropagation shaft speed of gearbox output shaft, the red slashed line
network (FFBP), as shown in Figure 9. There are 12 Inputs, 3 denotes measured pressure in the hydraulic clutch control, and
layers of neurons in the net. According to the Huang’s the black slashed line denotes the measured rotational speed of
network [12] the neuron distribution in each layer is 16-8-2. gearbox output shaft with a delay time of 0.1 second. In order
The transfer functions between each layer are Tan-Sigmoid to recognize the estimated and measured shaft speed, the
transfer functions [13]. second diagram shows magnified section of diagram (a). The
different between these two values during the whole training
IW{1,1} LW{2,1} LW{3,2} process is shown in diagram (c). It shows that this trained
ANN is capable of accurately emulating the target model.
Training Results
b{1} b{2} b{3}
12 1
n
16 8 2 Brake Sim
n
Layer 3 0.8 Target
Inputs Layer 1 Layer 2
Clutch
Brake
0.6
Accelerator Pedal
Figure 9. Structure of the network used for modeling of Clutch n
Engine
vehicle longitudinal dynamics [7] 0.4
n
O,GearboxD
Normalized Value [-]

0.2 Accelerator Pedal

The inputs of the model are composed of three parts, as shown 0


in Figure 10.
-0.2
• measured and calculated system parameters in real-time, nEngine
-0.4
u(t)
-0.6
• measured and calculated system parameters with defined
-0.8
delay time, u’(t)

-1
measured system output with defined delay time, y’(t) 221 222 223 224 225 226 227 228 229 230 231 232
t [s]

The real-time inputs u(t) chosen for this work are the
accelerator pedal travel PA_RT, brake pedal travel PB_RT, clutch (a)
pedal travel PC_RT and engine speed nEnginer_RT, etc. The input
parameters with defined delay time are the travel of three
pedals, PA_TD, PB_TD, PC_TD, and mean value of the wheel
speeds nWheel_TD, etc. Regarding reaction time of a normal
driver and response time of a vehicle system, the delay time is
set to 0.1 second. In order to improve the simulation quality of
the networks, further methods for the data manipulation are
considered in this work.

5 Copyright © 2008 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 06/24/2018 Terms of Use: http://www.asme.org/about-asme/terms-of-use


Training Results - magnified with the drift. Another test to verify the behavior of the model
n
Sim by defect sensor is shown in the prior work [7].
0.34 n
Target
nO,GearboxD
Clutch Validation results - Brake drift
Brake
Sim
Accelerator Pedal
0.32 0.6 Target
n
Engine Clutch
n Brake
O,GearboxD
Normalized Value [-]

nTarget n
0.4 Clutch Target Accelerator Pedal
0.3 n nEngine
Sim

nO,GearboxD
0.2
nSim Accelerator
0.28 Pedal
0

Y [-]
0.26 -0.2
nEngine

Brake -0.4 n
Engine
0.24
-0.6

226.2 226.4 226.6 226.8 227 227.2


t [s] -0.8
Brake

(b) -1
317 318 319 320 321 322 323 324 325 326
Training result - error t [s]
0.1

0.08
(a)
Validation results - Brake drift error
0.06 0.1
Error
0.04
0.08

0.02
0.06
Error [-]

0
0.04
-0.02
0.02
-0.04
Y [-]

0
-0.06
-0.02
-0.08

-0.04
-0.1
0 100 200 300 400 500 600
t [s]
-0.06

(c) -0.08

Figure 11. Results of training process (a) Diagram of -0.1


317 318 319 320 321 322 323 324 325 326
predicted and actual shaft speed (nSim and nTarget). (b) t [s]

magnified section 226.1 – 227.3 s of (a). (c) Error of the (b)


trained network
Validation results - original data

Sim
0.6 Target
Validation of the ANN model Clutch
The validation of the model proceeds under regarding of 0.4
Brake
Accelerator Pedal
Clutch
the following points: nEngine

0.2 nO,GearboxD
• ability of the model to detect fault by defect sensor Accelerator
(drift, total defect) 0 Pedal
Y [-]

• ability of the model to detect fault by partly defect -0.2

system components
-0.4
n

Engine
real time capability of the model Brake
-0.6

In order to assure the functionality of the sensors in the


-0.8
system, the model should have a proper ability to detect sensor
fault. In this study the drift effect and a total defect of the -1
sensors are concerned. In the first experiment it is assumed 317 318 319 320 321
t [s]
322 323 324 325 326

that there is drift effect on the sensor for measure of the brake
pedal travel. In this case measured value has drift effect in the (c)
time interval between 317.0s and 328.8s. Figure 12 illustrates
the behavior of the model when receiving input signal afflicted

6 Copyright © 2008 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 06/24/2018 Terms of Use: http://www.asme.org/about-asme/terms-of-use


0.1
Validation results - original error In the first step the model, as shown in Figure 14, is built
Error to process every single data packet. Each data packet is
0.08 composed of all input values, which are measured at the same
0.06
time. The input data are fed to the real time capable model on
the SIRIUS-Box at a frequency of 100 Hz, each data packet is
0.04 sent to the model for calculation. The translation of the data
packet uses UDP/IP transport protocol layered on top of the
0.02
internet protocol (IP). The result of the test is shown in Figure
15 (a). The average task execution time (TET) in this case is
Y [-]

0
0.000141 second. Because of the limitation of data transport
-0.02
technique using UDP/IP, the transport of input data through
-0.04 TCP/IP connection at a frequency of 100 Hz or higher is not
real-time capable. Thus the processing of data with the length
-0.06
of 20 seconds needs in this experiment 33.6480 seconds.
-0.08
-C- K*u K*u
-0.1 Network Inputs IW{1,1} netsum tansig LW{2,1}
317 318 319 320 321 322 323 324 325 326
t [s] -C-
Layer 1
b{1}
(d)
Figure 12. Validation of the vehicle model with abnormal
inputs signals -C- K*u
(a) Diagram of validation results with abnormal brake b{2}
netsum1 tansig1 LW{3,2} Layer 2
signal as input. (b) Absolute error of the simulation result
shown in (a). (c) Diagram of validation results with
original brake signal as input. (d) Absolute error of the
simulation result shown in (c) -C- *K
netsum2 tansig2 Matrix Layer 3
b{3}
reshape
The validation of the model in case of partly defect
system component is similar to the first test. Both of the
experiments use model-in-the-Loop (MiL) technique to verify Target Scope
Id: 1
the functional capability of the model. Scope (xPC)

The following experiments verify real-time capability of


Figure 14. Structure of the Network for SIRIUS-Box
the model. In the experiments the model is implemented on a
SIRIUS-Box which is developed in the IPEK, as shown in
Figure 13. The SIRIUS-Box is a mobile PC-System with In the next experiment the model is built to process
LAN- and CAN-Bus-Connection. The essential elements of colligated data packets, which are composed of 10 sequential
this box are: single data packets. The colligated packets are sent to the
SIRIUS-Box at a frequency of 10 Hz. The result of this test is
• Pentium III processor with 266 MHz shown in Figure 15 (b). The average task execution time
(TET) in this case is 0.001022 second, and the data are
• 128 MB SRAM processed in real time.
• 16 MB flash disk, etc.

(a) (b)
Figure 15. Validation of the vehicle model on SIRIUS-
Box

Figure 13. Sirius Box Further experiments are performed on a universal drive
train test bench which is developed in the IPEK(Figure 16).
By using rapid-prototyping software xPC-TargetTM from Goal of these experiments is to validate the monitor on a
the Mathworks, the box can validate the vehicle model in real- general real system.
time. Hence the influences on the validation test deriving from
the operating system and software can be ignored.

7 Copyright © 2008 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 06/24/2018 Terms of Use: http://www.asme.org/about-asme/terms-of-use


nW number of driving wheels (during acceleration) or
the total number of wheels (during braking)
nWheel wheel speed
T torque
Wj,i numeric weight

REFERENCES
[1] Isermann, R., 2006, Fault-Diagnosis Systems, Springer-
Verlag, Berlin Heidelberg New York.
[2] Isermann, R., 1994, Überwachung und Fehlerdiagnose:
Moderne Methode und ihre Anwendung bei technischen
Figure 16. Universal drive train test bench with drive Systemen, VDI-Verlag, Düsseldorf.
train of upper middle class vehicle and electric clutch [3] Burkhart, H. and Millen, R., 1999, “Performance-
actuation device measurement tools in a multiprocessor environment”,
IEEE transaction on computers, Vol. 38, No. 5.
CONCLUSIONS AND FUTURE WORKS [4] Ray, F., 1990, “Monitoring distributed embedded
This work it proves that the model based monitor using systems”, Proceedings of the 1990 Symposium on
ANN is one of the promising methods to supervise the Applied Computing.
mechatronic systems [6] [7]. The FFBP network introduced in [5] Grote, K.-H. and Feldhusen, J., 2005, “Dubbel -
this work is validated regarding the real-time capability and Taschenbuch für den Maschinenbau”, Springer-Verlag
the simulation accuracy. This monitoring technique can be
applied on complex nonlinear mechatronic systems, in which [6] Albers, A.; Ott, S.; Wang, J., 2007, “Attempt of
the requirements of an online model based monitor can be Prediction of Vehicle Longitudinal Dynamic
difficult satisfied by physical based system models or Performance using Artificial Neural Networks (ANN) ”,
mathematical approximate models. ASME IDETC: 9th International Conference on
Advanced Vehicle and Tire Technologies (AVTT), 04.-07.
Further works will concentrate in September 2007, Las Vegas, Nevada, USA.
• standardization of the processes by building ANN [7] Albers, A.; Wang, J.; Ott, S.; Düser, T.; Lerspalungsanti,
model and S., 2008, “Application of Artificial Neural Networks
• design of real-time fault diagnosis technique using (ANN) Based Model for Monitoring of Vehicle
online FMEA. Longitudinal Dynamic Performance”, ASME IDETC:
10th International Conference on Advanced Vehicle and
ACKNOWLEDGMENTS Tire Technologies (AVTT), August 3-6, 2008, Brooklyn,
This paper is part supported by project of “model based New York, USA.
monitor for measuring equipments and test benches” under the [8] Russell, S.; Norvig, P., 2003, Artificial Intelligence: A
grant of PRO INNO II1 KF0398803MB7. Modern Approach 2/E, Prentice Hall
[9] VDI 2206, 2004, Design methodology for mechatronic
systems, Verein Deutscher Ingenieure, Düsseldorf
APPENDIX NOMENCLATURE
α road gradient [10] Kleppmann, W., 2001, Taschenbuch Versuchsplanung,
A vehicle cross-sectional area Carl Hanser Verlag, München, Wien.
cW drag coefficient [11] Klein, B., 2004, Versuchsplannung – DoE, Oldenbourg
FA air resistance Wissenschaftsverlag GmbH.
FG gradient resistance [12] Huang, G.-B., Zhu, Q.-Y. and Siew C.-K., 2006, “Real-
fR rolling resistance coefficient Time Learning Capability of Neural Networks”, IEEE
FR rolling resistance Transactions on Neural Networks. Vol. 17, No. 4, pp.
Ft tire tractive (braking) force 863 – 878
FTR tractive resistance
g acceleration due to gravity [13] Demuth, H., Beale, M. and Hagan, M., 2006, “Neural
i number of ratio steps Network Toolbox User’s Guide”, The MathWorks
ρ atmospheric density
PA accelerator pedal travel
PB brake pedal travel
PC clutch pedal travel
nEngine engine speed

1
PROgramm „Förderung der Erhöhung der INNOvationskompetenz
mittelständischer Unternehmen“ (PRO INNO II) des Bundesministeriums für
Wirtschaft und Technologie (BMWi)

8 Copyright © 2008 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 06/24/2018 Terms of Use: http://www.asme.org/about-asme/terms-of-use

You might also like