10.

3: Vehicle Fault Diagnostics Using a Sensor Fusion Approach

Steven E. Muldoon Mark Kowalczyk John Shen
Delphi Cosworth Technology University of Michigan - Dearbom
Techcal Center Brighton Dearbom, MI, USA Dearbom, MI 48128, USA
Brighton, MI USA mkowalc [email protected] [email protected]
[email protected]

Abstract Vehicle data communication infiastructures and protocols

Vehicle electronics and mechanical systems continue to were created to perform information exchange between the
become more complex and interdependent. Automotive vehicle and the service community (i.e. OBD parameter
electronic control units (ECUs) used in vehicle sub-systems monitoring). The creation of this communication irhastruc-
execute high pegormance algorithms requiring robust fault ture was (in part) driven by the mandates and regulations
detection and diagnostics. Our technical approach begins by created by the California Air Quality Board (CARB) and the
presenting an overview of several techniques used for United States Environmental Protection Agency (EPA). To-
processing information)om multiple input sources (sensors day, in addition to providing the means for exchanging de-
in particular) in the context of ECU fault detection and tailed ECU information with the service community, this data
diagnostic schemes. We assert that inter-relationships bus is also utilized for exchanging particular data between
between groups of sensors can be exploited through sensor ECUs, including information from and related to vehicle sen-
fusion for signal integrity estimation, and present two vehicle sors. Any straightforward effort to design automotive diag-
application examples (chassis and powertrain). We propose nostics must attempt to determine the constantly changing
a virtual Jirsion / estimation technique which provides basic states of both the ECU hardware as well as the application
signal redundancy and fault tolerance. Dynamic vehicle s o h a r e utilized in these systems.
sensor information was used to develop an uncompounded
jiision-processing algorithm, expressed through a A. Vehicle Fault Diagnostics
Matlab/Simulink model. The results of these modeling Fault Diagnosis can be generally defined as the process that
experiments are presented to the reader and subsequent identifies the root cause of a failure. Vehicle level fault
conclusions are drawn regarding the implementation of a diagnostics should be considered an important area of study
sensor Jirsion-based or sensor jiision-enhanced ECU for several reasons, including the following:
algorithm as part of a more comprehensive diagnostic, Failure diagnosis adds directly to the cost of vehicle
control and service methodologv. The potential increase in parts - software is not “fie,” and warranty part
vehicle system functionality and utility realized )om sensor returns cost money.
jiision-based or sensor fusion-enhanced ECU fault Failure diagnosis is a large contributor to the labor
diagnostics has signlJcant implications for the sensor time involved in vehicle servicing - once a problem
designer, control algorithm architect, vehicle systems is identified, a solution is typically swift.
integration engineer, and automotive service community.
Improper fault diagnosis can lead to an incorrect re-
Keywords pair action - th~sleads to opportunity costs and fos-
Sensor Fusion, Vehicle Diagnostics, Automotive Sensors. ters the perpetuation of inefficient problem solving
(unsatisfied customers tend not to return).
There are several levels of vehicle diagnostics. The first level
1. INTRODUCTION of diagnostics may be a simple reminder to the driver that
The current use of automotive diagnostic tools and sensors standard maintenance (i.e., replacing filters and fluids) is
can be directly traced to the introduction of microprocessors required. The next level of diagnostics is targeted for the
in vehicle applications, and more specifically the introduction service technician, using more sophisticated tools and
of emissions control systems for the engine. Proliferation of techniques. And, there is a level or levels of tools and
microprocessor based vehicle sub-systems such as Electronic methods used by the system architects and engineers who
Engine Control (EEC), Anti-Lock Brakes (ABS) and develop the sub-system hardware and software. A well-
Electronic Suspension Control has led to a situation where a planned diagnostic strategy hlly comprehends the
multitude of sensors, actuators, inputs, outputs and relationships between these levels. [ 1,2]
application specific diagnostic schemes are found in
automobiles. The demands for increased performance and
robustness have resulted in algorithm improvements,
optimization and intra-vehicle system integration.

0-7803-7454-1/02/$17.00 02002 IEEE 1591

Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on April 10,2024 at 17:34:22 UTC from IEEE Xplore. Restrictions apply.
 Table 1. General Dynamic System Variables [2]
B. Sensor Fusion
In the case of multiple sensor faults, highly interactive I Variable 1
I
Description
components or subsystems, information from a single sensor U" I Input vector
does not provide enough intelligence to accurately diagnose
I Au I Input fault vector I
all of the possible faults. A better approach is to exploit the
0, Nominal parameter vector
use of multiple sensors to provide additional information to
the diagnostic algorithm. Using this approach, observers can A0 Parameter (component) fault vector
be selected such that they provide complementary and/or K State vector
redundant information about the component or subsystem
behavior. The additional information provided by this scheme 1 I Output measurement function
can be used to broaden the number and scope of potential
h(.,.,-.) - I
faults that the diagnostic algorithm detects and reduces the I Y 1 Actual output vector I
AY Output fault vector
probability of misdiagnosing a failure. [3]
Y' Measured output vector
By fusing two raw signals together and processing the
resulting information, improved characterization of the input
signals can be obtained. Further, by combining a number of
fused input signals with a simple voting or decision making
B. Neural Networks
algorithm, improvements in fault classification accuracy can
A neural network is a (sometimes very large) parallel
also be obtained. 131
computational structure. Neural networks consist of a very
large number of simple processing elements called neurons,
with each neuron connected to a large number of other
II. TECHNICAL APPROACH neurons.
A. Diagnostic Methodologies
Vehicle diagnostics will play a significant role in the
automotive markets of 21'' century. Therefore, there is a
strong need for the development of strategies that can
diagnose, predict, and perhaps prevent system failures. A P
U -
bt . 7
brief overview of some of the basic concepts and methods
used in fault detection are presented here.
A general dynamic system block diagram and fault
representation is shown below in Figure 1 [2]. Figure 2. Diagram of a Simple Neuron [4]

Figure 2 shows a simple neuron that receives inputs fiom the

external inputs, or from other neurons. The inputs are multi-
plied by connection weights, and summed to form the activa-
tion of the neuron. The neural network has an input side

L A -
Figure 1. General Dynamic System Block Model.[2]
where stimuli are presented to the network, and an output side
where the results of neural network's results can be obtained.
These results are commonly referred to "residuals", particu-
larly when used as feedback for control and diagnostic pur-
poses. [4] The neural network can be effectively used as the
basis of a multi-sensor hsion implementation, which will be
The relationship between the variables of this system can be shown in greater detail using our model example.
described as follows: [2]
C. Fuzzy Logic
.i = f ( x ,n.8 Fuzzy logic is a process designed to deal with ambiguous or
y =h(-r.u,8)
non-crisp situations. It was developed to represent and reason
with knowledge, expressed in terms of linguistic quantities,
e.g., "hot", " W m " ,"cold", rather than in terms of crisp
where ii = u o +Arc, 0 =e, + A 6 , y . = y + A y . numbers. A fuzzy logic controller attempts to emulate the
qualitative knowledge of an operator or expert of the process
The variables for the block diagram and expressions shown concemed. An expert of a particular procedure can easily
are defined in Table 1 . express knowledge about the process in terms of linguistic
rules, rather than in crisp rules. [5] In the context of this pa-

1592

Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on April 10,2024 at 17:34:22 UTC from IEEE Xplore. Restrictions apply.
per, fuzzy logic can be used as a complementary or altemate physical laws, or descriptions obtained using system
approach in designing sensor fusion algorithms. Figure 3 identification techniques. With these models, the user can
shows an example of how fuzzy logic can implemented for generate sensor signal residuals. Fault detection then takes
fault detection. place by evaluation of residual signals combined with a logic
decision. It is widely accepted that the main difficulty of
using the model-based fault detection schemes is the model
uncertainty. [6] This problem becomes accentuated in model
based fault detection used for automotive based systems,
since the process of "car driving" is influenced by many
unknown factors. Some of these factors can only be partially
modeled, modeled with uncertain fidelity or, in some cases,
cannot be mathematically described.
Since many new sensors will be introduced into passenger
.) cars in the near fhture due to the desire for additional features
and functions, the model based redundancy concept offers a
&lidof fault
general purpose approach to sensor monitoring. Figure 4
shows a generic example of a model based diagnostic
Figure 3 - Fuzzy Model for Fault Diagnosis [5] scheme. In this example, the inputs to the system (sensors for
our purpose) are processed in parallel through the control
system and the associated model of the control system. Then,
C. Observer Based Residual Generation a comparison is made between the system and system model
The use of the Dedicated and General Observer schemes for outputs, generating multiple residuals. Fault conditioning and
sensor fault diagnosis are explored and discussed in detail by logic (which may be implemented using fuzzy logic)
Kim in [l]. In these methods, each observer uses determine the final system output and error signals.
measurements fkom different sensors to provide an estimate
of the system output. Again, the results of these observers are
also typically referred to as "residuals". These schemes allow
Fault; I
Noise
Yisturbances System
OUtDUt
.-.r--

for individual or multiple sensor faults to be isolated and

identified when used in conjunction with threshold or logic
algorithms. A diagram of this approach is shown in figure 4.

Residual
Generator
Y"
Observer (m)

Figure 4 - Model Based Diagnostic Scheme [SI

Figure 4 - Dedicated Observer Scheme [i]

If all sensors are operating in a normal condition, the 111. SENSOR FUSION APPLICATION EXAMPLES
estimated output value of the observers will be the same as A. Chassis Sensor Fusion / Diagnostic
the sensor being evaluated. Similarly, by evaluation of the Since the beginning of chassis control systems, it has been
individual observer outputs, a sensor fault can be identified necessary to implement condition monitoring of the electrical
since only the observers using that particular sensor input will and electronic components being used. For sensors and
deviate fiom the measurement made by the collective actuators, the use of signal reasonability criteria and electrical
observer system. By further comparison of the deviant circuit monitoring is common practice. For this purpose,
observer, sensor fault isolation can be obtained. several vehicle sensors and control system inputs were
monitored and captured, including steering angle, vehicle
D. Model Based Diagnostics
speed, yaw rate, and lateral acceleration.
The basic concept of model based fault detection is the
design of theoretical redundancy using system or process In a generic stability control system, one popular method of
models, which may be mathematical descriptions of known providing input to the control algorithm is through the use of

1593

Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on April 10,2024 at 17:34:22 UTC from IEEE Xplore. Restrictions apply.
 yaw rate, lateral acceleration and steering angle information. compared with a calibratable max and min, and an error code
As the vehlcle is driven, the driver indicates the desired is set if the signal is out-of-range for a given number of
direction of travel through positioning of the steering sampling periods. Once an error code is set, the vehicle must
handwheel. The static and dynamic relationships between the either be serviced to clear the DTC, or (if the fault is deemed
various sensor signals determine the need and/or level for less severe) the DTC will automatically clear from the ECU
controlled brake intervention to physically correct the memory after a preset number of driving cycles are
directional path of the vehicle. In a simplified example, if the completed without the fault condition reoccurring.
difference between the yaw rate signal and steering angle Using the data collected from the vehicle, we later performed
signal become greater than a predetermined threshold level, sensor fision diagnostic modeling using three engine
stability system intervention takes place. Therefore, the parameters, MAF, throttle position, and engine RPM. These
importance of the signal integrity from various sensor inputs signals were selected based on the observation that they have
to the system, in both magnitude and phase, becomes some degree of systematic inter-relationship.
obvious. The control system can set a Diagnostic Trouble
Code (DTC) based on the results of sensor electrical and
correlation testing. Once a DTC is set, the features provided IV. SENSOR FUSION MODELING
by the system are typically degraded or disabled, and may A conceptual model of a multi-sensor fusion fault diagnosis
require vehicle service depending on the severity of the algorithm for a vehicle sensor application was inspired by
sensor failure detected. work being pursued in the areas of control system algorithm
Using the data collected fiom the vehicle, we later performed development and fault detection and identification (FDI) re-
sensor fusion diagnostic modeling using three chassis control search. The implementation discussed here was modeled
parameters, yaw rate:, lateral acceleration, and steering angle. using the Matlab / Simulink software package. Figure 5
These signals were selected based on the observation that shows a high level view of the Simulink model created. The
they demonstrate systematic inter-relationship. data is processed through what the authors call a direct, em-
pirical system identification transfer function. These transfer
B. Powertrain Sensor Fusion / Diagnostic functions were developed using the ARMAX (Auto Regres-
The intake air system provides clean air to the engine, sive Moving Average) system identification algorithm found
optimizes air flow and reduces unwanted induction noise. The in the Simulink System Identification Toolbox. By process-
intake air system consists of an air cleaner assembly, ing each sensor input as a system output of the other two
resonator assemblies and hoses. The mass air flow (MAF) available sensor inputs, six unique 3d degree transfer h c -
sensor is attached internally or externally to the air cleaner tions describing the mathematical relationship between the
assembly and measures the quantity of air delivered to the sensor signals were established.
engine combustion chamber. The MAF sensor can be
serviced or replaced as an individual component. The air By using two transfer function outputs derived from two sen-
intake subsystem is used to provide increased intake airflow. sors that describe the input signal of a third sensor, an ele-
This improves torque, emissions and performance. The mentary sensor fusion estimation signal (representative of the
third sensor input) was obtained. The combination of these
throttle body controls the overall quantity of air delivered to
transfer functions was obtained by a single, equally weighted,
the engine.
summing neural node. The output of the neural node is then
We collected real vehicle data from several sensor and engine filtered again using the ARMAX algorithm to provide a final
control inputs, including RPM, LAMBSE, throttle position, forcing function for each “estimated” sensor signal.
mass air flow, and other powertrain diagnostic information.
This dormation would later be used as input data for the Simultaneous to the signal fusion operation, a fuzzy logic
sensor fusion modeling activity. control algorithm implements a fault detection feature in the
model. Based on a limited rule set, the fuzzy logic algorithm
One engine parameter that can give insight into the overall monitored the sensor input signals for deviation from a prede-
condition of the engine system is the desired air/fuel flow fined dynamic operating range. This threshold level fault
“Lambda” (A). There are many factors that can influence the detection scheme is discussed in many of the reference
airhe1 ratio in the engine. If there is MAF sensor sources. When no faults are detected by the fuzzy processing
malfunction, air entering an engine is not being measured unit, the sensor inputs are passed straight through to the out-
correctly. If the sensor fails, a false indication of larger or put ports.
smaller amount of air entering the intake will cause an
imbalance to the air / fuel ratio, resulting in the engine
running leaner or richer. [7] The LAMBSE signal is

1594

Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on April 10,2024 at 17:34:22 UTC from IEEE Xplore. Restrictions apply.
 Figure 5. Simulink Sensor Fusion Model

A switchmg network was developed that allows the estimated and which may have discontinuities across their dynamic
sensor signal to be passed to the output port of the model range that invalidate this modeling algorithm. Methods to
when a fault on the original signal is detected. This is an on- address these concems are discussed in other research
line process that operates in real time. If the fuzzy controller literature and are not presented here. With that said, our data
no longer detects the fault, the original sensor signal is again showed positive results supporting the feasibility of this
passed through to the output of the model. In this sense, a sensor fusion concept as applied to an automotive sensor set.
fault tolerant sensor module is represented by the model, Figure 6 shows an example of the results fi-om the Simulink
providing both a sensor fusion estimate of the original signal, sensor fusion model described in the previous section. The
and a means to detect signal faults and switch signal paths upper graph in the figure shows the actual signal and the
dynamically in response to those faults using fuzzy logic. modeled signal. The lower graph shown in figure 6 and
Generally, it is proposed by the authors that the end user of Figure 7 shows the relative error between the actual and
the output fiom this type of sensor fusion processing modeled signals. As can generally be seen in this figure, the
algorithm take into consideration the lowered fidelity of the sensor fusion algorithm employed in this model was very
estimated signal provided during a fault condition and take responsive and followed the actual input signal well. Initial
appropriate actions to adjust the system control or diagnostic calculations show that error was about 10 to 20% for the
response, with the advantage being the ability to continue signal estimation with some local maximums and minimums
operation, rather than disabling a function due to the faulted performing slightly better or worst.
sensor condition. Additionally, the authors suggest that the
fuzzy logic cross correlation features could ultimately be
refined to provide automated “parameter of interest” selection
Actual Output (Red Line) ys. The Predicted Model output (slue Line)
and triggering for an enhanced diagnostic datalogging,
providing a source of valuable information to the service I
0.5
technician. a
L

0 0

Ian
-0.5
V. RESULTS AND DISCUSSION
1.2 1.25 1.3 1.35 1.4 1.45 1.5 1.55
Several assumptions were made in developing this simulation Time secs)
Error In Pre6icted Model
model: First, the model and methodology used are
acknowledged to have limitations in application and practical
use. Second, it is beyond the scope of this paper to Mly
define mathematical relationships between the sensor signals,
our intent is to demonstrate that the interrelationships exist,
and can be exploited to enhance system robustness and fault -0.3
1.2 1.25 1.3 1.35 1.4 145 1.5 1.55
i
diagnostics. Last, the assumption is made that the sensor Time (secs)
signals used have a mathematical relationship, which may or
may not be linear
Figure 6 - Chassis Fusion Results

1595

Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on April 10,2024 at 17:34:22 UTC from IEEE Xplore. Restrictions apply.
 Actual Output (Red Line) y5. m e Predicted Model output (Blue Line) The number of sensors in automobiles has risen dramatically
in the last decade. Current vehicles can contain fiom 40 to 50
sensors costing in excess of $1000. [8] The technique of
sensor fusion has been presented here as a potential tool for
automotive sensor parameter estimation problems, and
optimization of sensing resources. The benefits of sensor
0 0.02 0.04 O.Oti 0.08 0.1 0.12 0.14 0.16 0.18
fusion for vehicle diagnostics will include new functions that
Error i?%%%)Model
enable a cleaner environment, increased fuel economy, and
greater vehicle safety and reliability. Automobiles of the
future will require unique, complex and reliable diagnostic
schemes, as they become more integrated with advanced
technologies for safety, features, and performance.
4 0 5 b V , , , I I I , 4 Continued study and development of sensor fusion techniques
0 0.02 0.04 0.M; 008 0.1 0.12 0.14 0.16 0.18
Tlme (secs) (especially as applied to automotive applications) will
provide additional technical momentum towards
Figure 7 - Powertrain Fusion Results accomplishing these goals.

Figure 7 shows an example of processing the powertrain sen- ACKNOWLEDGMENTS

sor signals in the same fashion as the chassis sensors with the The authors wish to thank Delphi, Cosworth Technology,
only difference being the transform function equations and Ford Motor Company, and the University of Michigan -
threshold levels used. Note that the same algorithm and Dearbom for generously allowing access to resources used to
model was used to process both the chassis and powertrain conduct research and experiments related to this paper.
data. The differences were dealt with in the transform func-
tions and threshold levels, which in a more sophisticated im-
plementation of the model could be made calibratable. It can REFERENCES
easily be seen that the concept of sensor fusion for signal [l] Y.W. Kim, "A Methodology for Integration of Control
estimation andor fault detection in a vehicle application is a and Fault Diagnostics", Ohio State University, PhD
valid concept worthy of further exploration. Dissertation, UMI Microfilm 993 1629, 1999
[2] Ahmed A. Soliman, "The Application of Fuzzy Logic to
the Diagnosis of Automotive System", Ohio State
VI. CONCLUSION University, PhD. Dissertation, 1997
This paper reviews the technology related to sensor fusion [3] Claudio A. Hernandez, Alois J. Adams, "Application of
and vehicle fault diagnostics to provide the reader with a Mathematical Models to Detect and Diagnose Reciprocating
better understanding of the methodologies and concepts used Compressor Valve Leakage", SAE 200 1-01-1724
for correlation of vehicle sensor / diagnostic data, as well as
[4] Sameer M. Prabhu, "An Overview of Neural Network and
to demonstrate the trends and needs for future vehicle
Fuzzy Logic Applications to Earthmoving Machine System
diagnostic strategies. Improvements and optimization of
Control", SAE! 1999-01-1880
existing systems, system integration, and the demands for
more performance and features fiom future vehicle system [5] Carlo N. Grimaldi and Francesco Mariani, "OBD Engine
and diagnostic schemes are inevitable. Fusing two raw signals Fault Detection Using a Neural Approach", S A E 2001-01-
together and processing the information with a fusion 0559
algorithm can improve characterization of the input. [6] Fredrik Gustafsson, Stefan Ahlqvist, Urban Forssell and
Furthermore, by combining a number of fused input signals Niclas Persson, "Sensor Fusion for Accurate Computation of
and adding a simple voting or decision making algorithm, Yaw Rate and Absolute Velocity", SAE 2001-01-1064
improvements to fault classification accuracy can also be [7] Yi Lu Murphey, "Fuzzy Diagnostic Systems for Automo-
obtained. tive Engineering", University of Michigan-Dearborn, Techni-
One area of research directly related to this topic is the cal Project Report 1995/9, 1999
application of sensor fusion techniques to enhance the [8] D. Sparks, T. Noll, D. Agrotis, T. Betzner and K.
capabilities of the automotive service community. The Gschwend, "Multi-Sensor Modules with Data Bus
authors allude to this by stating that the fuzzy voting or Communication Capability", SAE 1999-01-1277
diagnostic algorithm could be easily modified to
accommodate an output provision to be used for fault
isolation datalogging. This selective captured information
would be available to the service t e c h " for more
effective fault analysis.

1596

Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on April 10,2024 at 17:34:22 UTC from IEEE Xplore. Restrictions apply.