You are on page 1of 10

ARTICLE IN PRESS

Control Engineering Practice 15 (2007) 315–324


www.elsevier.com/locate/conengprac

Fault detection for lateral and vertical vehicle dynamics


D. Fischer, M. Börner, J. Schmitt, R. Isermann
Institute of Automatic Control, TU Darmstadt, Landgraf-Georg-Str. 4, 64283 Darmstadt, Germany

Abstract

Mechatronic systems for vehicles have received increased importance for improving automotive safety and comfort. These systems are
designed to aid the driver by preventing unstable or unpredictable vehicle behavior and to stabilize the horizontal and vertical motion of
the vehicle. This is achieved by the integration of actuators, sensors and data processing. However, the attained benefits are paralleled by
an increase in the complexity of the system requiring enhanced methods for fault detection and diagnosis. Therefore, concepts for model-
based fault detection and diagnosis along with sensor fault tolerance are presented and realized for both a vehicle lateral dynamics system
and an active suspension system.
r 2006 Published by Elsevier Ltd.

Keywords: Fault detection; Vehicle dynamics; Mechatronic systems

1. Introduction symptoms and the faults is established. The task at hand is


to find significant symptoms, which are robust against
The introduction of mechatronic systems in today’s noise, disturbances, and changes of the setpoint. This
vehicles aims at increasing the performance, functional requires application of both signal- and model-based
range, quality and dependability of automotive systems by concepts. Signal methods are used to detect directly
the integration of actuators, sensors and data processing ascertainable and mostly major faults such as noise,
(Alleyne & Hedrick, 1995; BuXhardt, 1993; Chen & Guo, outliers, or sensor breakdowns. Signal-based fault detec-
2001; Fischer, 2004; Gerdes, Rossetter, & Switkes, 2004; tion has progressed to maturity leaving the focus of current
Kiencke & Nielsen, 2000; Pyper, Schiffer, & Schneider, research on model-based methods, which are necessary in
2003; Rajamani and Hedrick, 1995). Due to high reliability order to detect sensor offsets, gains, and process faults.
demands of such mechatronic systems, recent fault detec- One approach toward model-based methods is parameter
tion and diagnosis concepts for these systems are of monitoring, where online estimated physical parameters
particular importance (Isermann, 2003; Isermann, of the process are compared to reference values (Isermann,
Schwarz, & Stölzl, 2002; Patton, 1998). 1997). Alternatively, the parity equation approach
The objective of fault detection and isolation (FDI) is to (Gertler, 1998; Patton, 1998), based on the deviations of
determine if a fault is present in a system (fault detection), estimated process states or outputs from measured ones,
as well as to determine the kind, the location (fault the so-called residuals, may also be employed. The
isolation), the size, and the time-varying behavior (fault parameter and state estimation is based on process models,
identification) of the fault. Most schemes consist of two causing non-linear processes to be difficult to handle. In the
levels, divided into symptom generation and fault diag- case of parameter estimation, the models are linearized due
nostic (compare Fig. 1 and Fig. 6). In the first part, to an easy online realization of the estimation algorithm. In
symptoms are generated that indicate the state of the the case of state estimation, semi-physical models using
process. In the second part, the relationship between the LOLIMOT (Nelles, 2000) are applied.
These methods are now used for sensor and process fault
Corresponding author. Tel.: +49 (0) 6151 16 7407; detection in vehicles. Due to structural conditions, the
fax: +49 (0) 6151 16 7421. topic of vehicle dynamics is typically divided into
E-mail address: dfischer@iat.tu-darmstadt.de (D. Fischer). horizontal and vertical dynamics (Kiencke and Nielsen,

0967-0661/$ - see front matter r 2006 Published by Elsevier Ltd.


doi:10.1016/j.conengprac.2006.05.007
ARTICLE IN PRESS
316 D. Fischer et al. / Control Engineering Practice 15 (2007) 315–324

Nomenclature iM motor current (A)


ist steering system gear ratio (dimensionless)
dst steering wheel angle (rad) J moment of inertia motor (Nms2)
CM motor torque constant (Nm/A) l distance front-rear axle (m)
C motor field flux linkage (Vs) M0 static motor moment (Nm)
c yaw angle (rad) mB, mW body mass, wheel mass (kg)
j0 mean pump position (1) r dynamic tire radius (m)
oi wheel speed, i ¼ FL, FR, RL, RR; front Left/ Rhyd laminar hydraulic resistance (Pa s/m3)
right, rear left/right (1/s) RM armature resistance (O)
oM motor speed (1/s) uM motor voltage (V)
_
c yaw rate (rad/s) v longitudinal velocity (m/s)
A plunger cross sectional area (m2) vch characteristic velocity (m/s)
b track width (m) Vz0 volume constant of pump (m3/s)
cB spring coefficient (N/m) ÿ lateral acceleration (m/s2)
cS stiffness of hydraulic accumulator (Pa/m3) zB body height (m)
dB damper coefficient (Ns/m) zP plunger position (m)
FC Coulomb force (N) zWB suspension deflection (m)

signal, a virtual value is calculated to enable a reconfigura-


tion. The entire scheme is shown in Fig. 1.

2.1. Parity equations

The developed parity equations are based on simple


kinematical equations and on models of the lateral vehicle
dynamics (e.g., Isermann, Schmitt, Börner, & Fischer, 2004).
For a detailed description of the model, accuracies and
environment conditions see Börner and Isermann (2005).
Four different models calculate each sensor signal with
different inputs using analytical redundancies, see Table 1.
The outputs of the 12 resulting models zmod,i(t) are compared
with the measured sensor signal zmeas,i(t), yielding the
residuals ri(t), see Fig. 2. The 12 calculated residuals ri(t)
are then used for fuzzy logic fault diagnosis. Residuals r1yr4
are related to the yaw rate, r5yr8 to the lateral acceleration,
and r9yr12 to the steering wheel angle models, (Table 2).

2.2. Weighting of ABS wheel speed signals


Fig. 1. Fault detection and diagnosis of the lateral behavior based on
signal and process models. The ground velocity v can be calculated by averaging the
four wheel speed signals as
vFL þ vFR þ vRL þ vRR
v¼ , (1)
2000). Within this report therefore, fault detection for 4
mechatronic systems is discussed in both horizontal
dynamics and vertical dynamics. v ¼ o  r, (2)
with o as the rotational speed and r as the dynamic tire
2. Fault detection for lateral vehicle dynamics radius. After a fault in the front left wheel speed sensor vFL,
all residuals, except residual r2, exceed their nominal values
The input signals for the presented fault detection and (see Table 1). Hence, this method cannot distinguish
diagnosis system are obtained by the standard sensors of between faults in the left-hand or right-hand side of one
the ESP system (electronic stability program): the steering axle. In order to distinguish between sensor faults of the
wheel angle dst, the ABS-wheel speeds vij, the lateral left and right sensor signals, the deviations of all ABS
_ The sensors and the
acceleration ÿ, and the yaw rate c. signals must be taken into account. Therefore, a weighted
lateral driving behavior are monitored by applying signal- factor voter (Broen, 1975), based on a weighted mean
and model-based methods. In the case of a faulty sensor calculation from four measured sensor signals, is used.
ARTICLE IN PRESS
D. Fischer et al. / Control Engineering Practice 15 (2007) 315–324 317

Table 1
Models for fault detection

Yaw rate c_ Lateral acceleration ÿ Steering wheel angle dst


vFR  vFL  
^_ ¼ vFR  vFL
c y^€ 1 ¼ v l  ist v2 vFR  vFL
1
bF bF d^ st;1 ¼  1þ 2 
v vch bF

vRR  vRL  
^_ ¼ vRR  vRL
c y^€ 2 ¼ v l  ist v2 vRR  vRL
2
bR bR d^ st;2 ¼  1þ 2 
v vch bR
 
^_ ¼ y€ y^€ 3 ¼ c_  v l  ist v2
c 3 d^ st;3 ¼ 2  1 þ 2  y€
v v vch

^_ ¼ v v2 l  ist _
c 4
l  ist
 dst y^€ 4 ¼  dst d^ st;4 ¼ c
l  ist v

The weighting of one fault can be influenced by parameter


a, which is chosen experimentally (Stölzl and Isermann,
1999). Eq. (3) directly yields the following statements:

 if the ABS wheel speed signal information vi are nearly


equal, it follows wFL ! wFR ! wRL ! wRR ! 1,
 if only vFL, vFR, and vRR are nearly equal, and vRL is
diverging from the common value, it follows wFL !
wFR ! wRR ! 1, and wRL ! 0.

Two new residuals are defined:


)
r13 ¼ wFL  wFR
r13;14 2 ½1; þ1. (4)
r14 ¼ wRL  wRR

Fig. 2. Model-based fault detection. For instance, a fault of the front left wheel speed sensor
yields residual r13o0.

Table 2
Relationship of residuals 2.3. Diagnosis
_
c ÿ dst
The 14 calculated residuals ri(t) are used for the
Model 1 r1 r5 r9 fuzzy logic fault diagnosis (Füssel, 1999; Isermann, 2003).
Model 2 r2 r6 r10 Table 3 shows how different sensor faults affect the
Model 3 r3 r7 r11 symptoms. Here, an additional case is defined with all
Model 4 r4 r8 r12 residuals set to zero to refer to a faultless driving condition.
The processing of Table 3 is performed by a fuzzy logic
Faulty sensor signals are then weighted based on amplitude. system. Three trapezoid membership functions, such as
If a certain threshold is exceeded, a soft mask, depending on ‘‘residual positive’’, ‘‘residual zero’’, and ‘‘residual nega-
the fault size, is used (similar to the principle of fuzzy logic). tive’’, are applied to the residuals of the yaw rate models,
The weights wFL, wFR, wRL and wRR of their respective ABS the lateral acceleration models, and the steering wheel
wheel speed sensor vFL, vFR, vRL and vRR may now be angle models. Two symmetric membership functions are
calculated in the range between 0 and 1 according to fault applied for the residuals r13 and r14, Fig. 3. The rules of the
size. The deviation of one signal from the others is inversely fuzzy system are given by the rows of Table 3. Each row
proportional to its corresponding weight. refers to a fuzzy rule with AND operations. For example,
the fuzzy rule for a negative offset fault in the rear right
1 ABS wheel speed sensor (F14) is
wi ¼ Q vi vj 2 ; i ¼ FL; FR; RL; RR.
1þ a
jfFL;FR;RL;RRg\fig IF r1 zero AND r2 negative AND r3 zero
(3) AND . . . AND r15 positive THEN F vRR  DvRR ¼ 1:
ARTICLE IN PRESS
318 D. Fischer et al. / Control Engineering Practice 15 (2007) 315–324

Table 3
Fault symptom relationships, Influences of different sensor faults on the estimated symptoms

Symptoms

Yaw rate residuals Lat. accel. Residuals Steering wheel res.


r1yr4 r5yr8 r9yr12 r13 R14

Lateral accel. ÿ+Dÿ 0 0 + 0     0 0 + 0 d d


Lateral accel. ÿDÿ 0 0  0 + + + + 0 0  0 d d
Yaw rate dc/dt+dc/dt     0 0  0 0 0 0 + d d
Yaw rate dc/dtdc/dt + + + + 0 0 0 0 0 0 0  d d
St. w. angle dL+DdL 0 0 0 + 0 0 0 +     d d
St. w. angle dL+DdL 0 0 0  0 0 0  + + + + d d
ABS signal vFL+DvFL  0 0 0  0 0 0  0 0 0  0
ABS signal vFLDvFL + 0 0 0 + 0 0 0 + 0 0 0  0
ABS signal vFR+DvFR + 0 0 0 + 0 0 0 + 0 0 0 + 0
ABS signal vFRDvFR  0 0 0  0 0 0  0 0 0 + 0
ABS signal vRL+DvRL 0  0 0 0  0 0 0  0 0 0 
ABS signal vRLDvRL 0 + 0 0 0 + 0 0 0 + 0 0 0 
ABS signal vRR+DvRR 0 + 0 0 0 + 0 0 0 + 0 0 0 +
ABS signal vRRDvRR 0  0 0 0  0 0 0  0 0 0 +
Normal driving situation 0 0 0 0 0 0 0 0 0 0 0 0 0 0

+, increase; , decrease; d, don’t care; 0, no deflection; 7D, positive or negative offset of sensor signal.

Yaw rate [rad/s] As each rule fires for an individual output (fault), no
explicit accumulation and defuzzification is required.
1 zer
Finally, the output of a rule describes the possibility for
o
 the incidence of the specific fault. Thus, a fault can be
0.5
detected by exceeding a specific threshold.
Table 3 shows that 14 different sensor faults and a
0
normal driving condition can be classified. Each fault has a
-0.5 -0.3 0 0.3 0.5
different pattern, a requirement for an isolation of all
Lateral acceleration [m/s2] faults. Robustness to false alarms is one of the major
problems of fault detection and diagnosis systems. For this
1 zer application, the robustness to false alarms is achieved
o

by taking into account sensor faults greater than 10%
0.5 during normal operation and non-critical driving situa-
Membership function μ [-]

tions. The threshold of 10% was chosen arbitrarily.


0 Finally, the automated fault diagnosis showed very good
-3 -2 -1 0 1 2 3
results during several test drives (Börner and Isermann,
Steeringwheel angle [rad] 2003).
In addition to sensor fault diagnosis, a monitoring
1 zer system for critical driving situations was developed based
o
on the characteristic velocity (Börner, Andréani, Albertos,
0.5 
& Isermann, 2002).

0 2.4. Analytic sensor fault tolerance


-3 -2 -1 0 1 2 3

Residuals r13 and r14 The fault-tolerant sensor configuration should be at least
fail-operational for one sensor fault. This can be obtained
1 by applying analytical redundancy with different sensors
neg (physical sensor) and process models (software sensor).
0.5
pos
A reconfiguration will be necessary if a sensor is in a fault
0 condition (Van Zanten, Erhardt, Landesfeind, & Pfaff,
2000). The reconfiguration is determined by the fault
-1 0 1
detection and diagnosis system. The faulty sensor has to be
Fig. 3. Membership functions of the fuzzy logic system. excluded from driver assistance systems, e.g. within 250 ms
ARTICLE IN PRESS
D. Fischer et al. / Control Engineering Practice 15 (2007) 315–324 319

(Van Zanten et al., 2000). Due to the exclusion of the faulty 3. Fault diagnosis for vertical vehicle dynamic systems
sensor, a fault-free sensor configuration is restored.
The decision when a sensor has to be excluded and the First results on fault detection methods for passive and
system has to be reconfigured is done by the fuzzy logic semi-active suspensions were published by e.g., BuXhardt
fault diagnosis system and a fault counter. A stationary (1993); Majjad (1997) and Weispfenning and Isermann
circular test drive, Fig. 4, shows the reconfiguration of the (1997). Although a variety of publications dealing with
sensor system during an offset fault in the lateral control strategies for active suspension systems exists (e.g.
accelerometer or the yaw rate sensor. A non-linear two- Alleyne and Hedrick, 1995), fault detection for these
track vehicle dynamics model is used to deliver an estimate systems has not been investigated widely (Rajamani and
of the faulty sensor (Börner and Isermann, 2003). Fig. 4 Hedrick, 1995). In the following text, model-based fault
shows the lateral acceleration signal and the yaw rate signal detection methods are applied to an active vehicle
(grey curves). These signals contain an offset fault. With suspension system on a test rig.
the help of the reconfiguration system, a correct working
driver assistance system can be guaranteed. 3.1. Example of an active suspension system

The investigated active suspension is a fully loaded


neg. lateral accelerometer hydraulic system, Fig. 5. It consists of a hydraulic cylinder,
fault (10%) which is constructed as a plunger and is connected in series
Lateral acceleration ÿ

6
with a steel spring. The damper directly connects the body
4 and the wheel mass. This assembly is equivalent to the
[m/s2]

2 Faultysensorsignal active body control system of DaimlerChrysler (e.g.,


Signal after Mercedes CL-Class). The plunger is connected to a
0
reconfiguration motor-pump unit, which provides a controlled hydraulic
-2 flow from/to a hydraulic accumulator. The test rig
(a) demonstrates a quarter of a real car with suspension arms,
pos. yaw rate
Zoom
sensor fault (30%) a real tire, and an appropriate body mass. A realistic road
1.5 110 ms excitation (measured road profile of a rough road at a
speed of 20 m/s) is simulated by the vertical movement of
Yaw rated /dt

1
the base plate. The test rig is equipped with a variety of
0.5 500 ms
sensors, whereas the presented algorithms are based on a
0 minimal sensor setup due to automotive requirements. The
-0.5 sensor setup includes sensors for the suspension deflection
0 10 20 30 40 50 zWB, for the body acceleration z€B , as well as for the motor
(b) time t [s] speed oM, voltage uM, and current iM. The applied
Fig. 4. Reconfiguration of a sensor system during (a) a fault in the lateral controller is a state space controller with pole placement
_
acceleration sensor ÿ (b) a fault in the yaw rate sensor c. design.

Fig. 5. Active suspension test rig with measured signals ( ) and control scheme.
ARTICLE IN PRESS
320 D. Fischer et al. / Control Engineering Practice 15 (2007) 315–324

The components of the electro-hydraulic suspension suspension. In the case of a detected sensor fault, the
system, such as the hydraulic accumulator, the hydraulic system can be reconfigured and the physical sensor is
lines, the motor, the pump, the plunger and the suspension replaced by an estimated sensor signal, based on analytical
system are modeled according to Isermann (2003). Con- models. Signal-based symptoms sS are generated with low-
sidering a linear spring, a piecewise linear damper with pass and high-pass filters in order to detect obvious faults
Coulomb friction, laminar hydraulic resistances, the motor such as outliers, noise, and breakdowns (Fischer et al.,
as DC drive, the hydraulic accumulator as an air spring, 2004). The generation of model-based symptoms is
and a positive displacement pump, the following equations described in the following text.
are derived (for details see table of symbols and Fischer, The fault detection scheme was applied to test rig data.
Schöner, & Isermann, 2004): For this purpose, the base plate of the wheel at the test rig
Z was excited according to a measured street profile and a
cB cB  V z0
z€B ðtÞ ¼  zWB ðtÞ þ oðtÞ dt speed of 15 m/s. The power density spectra of the wheel
mB mB  A
movement and of the wheel and body mass acceleration
d Bi FC show typical frequencies for a realistically configured
þ  z_WB ðtÞ þ  signðz_WB ðtÞÞ
mB mB vehicle.
cB  V z0  j0
þ , ð5Þ
mB  A 3.3. Parameter monitoring

uM ðtÞ ¼ RM  iM ðtÞ þ C  oM ðtÞ, (6) Based on measurements at the test rig, the parameters of
Eqs. (5–7) are estimated with the Discrete Square Root
CM V z0  cB Filter in Information Form (DSFI) (Isermann, 1992), where
_ M ðtÞ ¼
o iM ðtÞ þ M 0  zWB
J AJ the forgetting factors were set to l ¼ 0.9999 and l ¼ 0.99
Z
Rhyd  V 2z0 cS  V z0 for the motor equation. DSFI is based on the recursive
 oðtÞ  oM ðtÞ dt. ð7Þ
J J least squares (RLS) algorithm and, additionally, uses a
QR-decomposition with its numeric advantages. The
required derivatives of the suspension deflection and the
3.2. Fault detection scheme motor speed were determined using state variable filters
(Butterworth characteristic with third order). The identifi-
The entire fault detection scheme is presented in Fig. 6. cation results show a fast and converging estimation for all
The measured signals of the process are used to generate 13 parameters (A and Vz0 have to be known a priori).
symptoms using signal- and model-based methods. These Results are presented for the body spring stiffness cB, the
symptoms are then used to classify faults of the active hydraulic resistance Rhyd, the electric motor resistance RM,
and the field flux linkage CM in Fig. 7. The estimations
show very constant values after a short regulation period.

×105
2.5

2
cB [N/m]
1.5

0.5 Rhyd [Pa·s/m3]


0
(a)
2

1.5
RM [Ω]
1

0.5
 [V·s/U]
0
0 2 4 6 8
(b) t [s]
Fig. 6. Fault detection, isolation and reconfiguration concept for the
active suspension. Fig. 7. Parameter estimation of the suspension and electric motor.
ARTICLE IN PRESS
D. Fischer et al. / Control Engineering Practice 15 (2007) 315–324 321

Fig. 8. Model-based fault detection with parity equations.

Fig. 9. Online estimation of the motor current and body acceleration.

In general, the comparison of all estimated values to The LOLIMOT models are trained with test rig data and
reference values indicates an accurate estimation and verified with generalization data. The models show relative
substantiates correct modeling of the process. errors of 15% (rms). Examples for the motor current iM
For fault detection, the online estimated parameters y^ and the body acceleration z€B are shown in Fig. 9. Both
are compared to reference values yref and the differences estimations are close to the measured signals and show
are used as symptoms for fault detection. minor deviations despite different values and dynamics of
^  yref . the signals. Hence, the models are used for residuals
DyðtÞ ¼ yðtÞ (8)
according to Fig. 10.

3.4. Parity equations with semi-physical models 3.5. Diagnosis

One goal in designing parity equations is to find a set of The 13 parameter symptoms and the 6 residuals are
isolating residuals that shows an unambiguous pattern of combined to the symptom vector sM, which represents the
deflections for each fault (Fig. 8). As this is not fulfilled by input of the fuzzy diagnostic system. The rule basis of the
the models of Eqs. (5–7), further models are obtained by fuzzy diagnostic system is the fault–symptom relationship
combining these equations. Experiments showed that six between the symptoms and the process, as well as the
models provide the best diagnostic information, although sensor faults. Thus, process and sensor faults can be
five models would be sufficient for isolating faults of the distinguished for differing symptom patterns. Sensor faults,
involved five sensors. However, the resulting six models sensor offsets, and sensor gains are both considered due to
showed insufficient accuracy for residuals due to unmo- their different symptom patterns. The fuzzy logic system is
deled non-linearities and due to effects of the spring, realized equally to Section 2.3 with three trapezoidal
hydraulic accumulator, pump, damper, and the suspension membership functions for each symptom. An excerpt of
transmission ratio. As a consequence, the semi-physical the fault symptom relationship is demonstrated in Table 4.
approach LOLIMOT is applied, which is based on a The whole fault symptom relationship has an isolating
combination of local linear models and neuronal networks structure (Fischer et al., 2004). The diagnosis focuses on
with basis functions (for details see Nelles, 2000). The the smallest intermountable part in a vehicle workshop.
inputs and outputs of the local linear network are defined Consequently, the outputs of the fuzzy diagnostic system
by the structure of the six models according to the are the fault possibilities of the six main components
following equations: of the process (motor-pump unit, hydraulic accumulator,

z€B ðkÞ ¼ f 1 ðzWB ðk; . . . ; k  4Þ; oM ðk; . . . ; k  4Þ; iM ðk; . . . ; k  4Þ; z€B ðk  1ÞÞ,
z€B ðkÞ ¼ f 2 ðzWB ðk; . . . ; k  3Þ; iM ðk; . . . ; k  3Þ; oM ðk; . . . ; k  3Þ; z€B ðk  1ÞÞ,
z€WB ðkÞ ¼ f 3 ðz€B ðk; . . . ; k  3Þ; oM ðk; . . . ; k  3Þ; uM ðk; . . . ; k  3Þ; z€B ðk  1ÞÞ,
uM ðkÞ ¼ f 4 ðiM ðk; k  1Þ; oM ðkÞÞ,
iM ðkÞ ¼ f 5 ðzWB ðk; . . . ; k  2Þ; oM ðk; . . . ; k  3Þ; iM ðk  1ÞÞ,
oM ðkÞ ¼ f 6 ðiM ðk; . . . ; k  2Þ; uM ðk; . . . ; k  2Þ; oM ðk; . . . ; k  1ÞÞ. ð9Þ
ARTICLE IN PRESS
322 D. Fischer et al. / Control Engineering Practice 15 (2007) 315–324

suspension strut, etc.) and the five sensors. The fault diagnosis requires that the maxima lie on the matrix
possibility for a specific component is calculated by diagonal and that a reliable diagnosis is characterized by
applying a fuzzy-or-operation (algebraic sum) to all small values in the non-diagonal elements in each column.
relevant faults. For instance, the fault possibility for the In the case of a certain fault, only the associated value in a
body acceleration sensor fault is given by the algebraic sum column exhibits a clearly increased fault possibility. These
of the fault possibilities for the offset and the gain. conditions are clearly met with the results of Table 5,
indicating an unambiguous and reliable diagnosis. Only
3.6. Sensor fault detection faults in the sensors for body acceleration and suspension
deflection show reduced diagnosis sharpness. This stems
For the demonstration of the correct diagnosis of all from the model equations, where these variables are
sensor faults, offsets and gains are applied to all sensor strongly linked with each other. This phenomenon is
signals. The offsets refer to 6% of the total sensor range already known from passive and semi-active suspension
and the gain faults refer to a multiplication of the signal by systems (Weispfenning and Isermann, 1997).
1.3. The results for all faults are summarized in Table 5 in For instance, Fig. 11 demonstrates the diagnosis of an
which each column corresponds to one of the outputs of offset and a gain fault of the body acceleration sensor.
the fuzzy diagnosis system, fM. Each row refers to one Fig. 11a shows the real and the faulty body acceleration
present fault and the values are the maximum of the fault with an offset (t ¼ 9–11 s) and a gain fault (t ¼ 13–15 s) of
possibilities during this fault. Hence, an unambiguous the sensor. The small deviations between the faulty and the
real signal indicate that such faults cannot be detected with
signal-based methods. Fig. 11b illustrates the evaluated
fault possibilities of the offset rule f M;€zB offset and the gain
rule f M;€zB gain of the fault–symptom relationship in Table 4.
Both signals show clear deflections during the present
faults, but only the fault possibility of the offset rule clearly
identifies the associated fault. Due to high amplitudes and
small dynamics of the body acceleration for t ¼ 9–11 s,
offset and gain fault show similar symptom deflections.
Finally, the combination of these single fault possi
bilities results in the total fault possibility for the body

Table 5
Sensor fault detection results

Present fault Maximal fault possibilities fM during the present faults

f M , uM f M , iM f M , oM f M ; z€B fM, zWB

uM Offset 1 0 0 0 0
Gain 1 0 0 0 0
iM Offset 0.005 1 0 0 0
Gain 0 1 0 0 0
oM Offset 0.003 0 0.981 0 0
Gain 0.002 0 1 0 0
z€B Offset 0 0 0.002 0.9 0.582
Gain. 0 0 0 0.2 0,018
zWB Offset 0 0 0.003 0.03 1
Gain 0 0.004 0 0 0.856
Fig. 10. Model-based fault detection with parity equations.

Table 4
Fault–symptom-relationship for the active vehicle suspension system

Faults (excerpt) Symptoms

Parameter estimation Parity equations

cB dB+, FC Vz0 C RM cS J CM Rhyd M0 j0 r1 r2 r3 r4 r5 r6

Clogging 0 0 0 0 0 0 0 0 0 + 0 0 0 0 7 0 7 7
z€B Offset 0 0 0 0 0 0 0 0 0 0 0 + + + 7 0 0 0
Gain + + + + 0 0 0 0 0 0 0 0 7 7 7 0 0 0
ARTICLE IN PRESS
D. Fischer et al. / Control Engineering Practice 15 (2007) 315–324 323

acceleration sensor f M;€zB total . This fault possibility clearly sensitive and intensive performance. This leads to a clearer
indicates the present faults. Table 5 contains the maximal differentiation between faults of different components.
values of the fault possibility during the offset (0.9) and the Diagnosis results for process faults at the test rig are
gain fault (0.2), whereas the columns are normalized with summarized in Table 6. It can be seen that all faults are
respect to the diagonal elements. clearly detected and correctly identified. Fault F refers to
mass changes, which does not influence the process fault
3.7. Process fault detection detection.
For example, clogging in the oil line was generated at the
Process fault detection is based on the same symptoms, test rig by partial closing of a valve in the hydraulic line
sM, as the sensor fault detection. Thus, all possible process between plunger and pump. Fault detection of clogging in
faults are added to the fault–symptom relationship with the hydraulic lines is essentially based on the estimated
their individual symptom patterns. The resulting fault- hydraulic resistances and on the related parity equations.
symptom-relationship shows characteristic patterns for The online estimated hydraulic resistance is presented in
each process fault, which are also different from the sensor Fig. 12a. The estimated resistance shows an increased
fault patterns, allowing all faults to be isolated. value. Due to the low-pass characteristics of the recursive
According to a close-to-production diagnosis approach, parameter estimation, this deflection is observed approx.
the diagnosis should focus on the smallest exchangeable 4 s after the fault’s occurrence. The fault possibility
unit. Different process faults of a component are combined f M; MoPu , comprising all faults of the motor-pump unit, is
to accomplish this goal. As different process faults of a illustrated in Fig. 12b. It shows a clear deflection and
component have often related symptom patterns, the demonstrates that this fault can be classified.
combination of their fault possibility shows a more These results and also the sensor fault detection results
come from a test stand on laboratory conditions. There-
fore, it is to be still tested in experiments with a driving car
and in long-term tests, where variations in temperature and
attrition affect the fault detection.

0.8
Clogging
Rhyd [Pa·s/m3]

0.6

0.4

0.2

(a) 0
1
fM, MoPu [1]

0.5

-0.5
8 10 12 14 16 18 20 22
(b) t [s]

Fig. 11. Fault detection for body acceleration signal. Fig. 12. Fault detection of clogging in the line between plunger and pump.

Table 6
Process fault detection results

Present fault Maximal fault possibilities fM during the present faults

f M;A f M;B fM,C FM,D fM,E fM,F

A: MoPu Short 1 0 0 0.001 0 0


inner leakage 0.21 0 0 0 0 0
B: Clogging Accum. line 0.08 1 0 0.91 0 0
Plunger line 0 0.073 0 0.54 0 0
C: Air Trapped air 0.78 0 1 0 0 0
D: Accum. Gas leakage 0 0 0 1 0 0
E: Oil leakage 0.004 0 0.05 0.38 1 0
F: Mass change 0.002 0.8 0.2 0.6 0.21 1
ARTICLE IN PRESS
324 D. Fischer et al. / Control Engineering Practice 15 (2007) 315–324

4. Conclusion Fischer, D., Schöner, H. P., & Isermann, R. (2004). Model-based fault
detection for an active vehicle suspension. In FISITA world automotive
congress. Barcelona.
The detection and diagnosis of sensor and process faults,
Füssel, D. (1999). Fuzzy logic for fault detection and diagnosis. In Sixth
as well as sensor reconfiguration, were presented for lateral UK workshop on fuzzy systems. 8–9 September 1999, Brunel
and vertical vehicle dynamics. As a result of a model-based University, Uxbridge (pp. 79–90).
approach, parameter estimation and parity equation design Gerdes, J. C., Rossetter, E. J., & Switkes, J. P. (2004). Experimental
were used for symptom generation. By considering the validation of the potential field lanekeeping system. International
different signs of the symptom’s deflection and by applying Journal of Automotive Technology, 5(2), 95–108.
Gertler, J. (1998). Fault detection and diagnosis in engineering systems.
a diagnosis with fuzzy logic, small faults could be detected New York: Marcel Dekker Inc.
with a good isolation performance. It could be shown that Isermann, R. (1997). Supervision, fault-detection and fault-diagnosis
diagnosis robustness can be increased by combining several methods—an introduction. Control Engineering Practice, 5(10),
fault modes for isolation of the smallest faulty exchange- 1355–1361.
able unit. In general, the presented methods are based on Isermann, R. (2003). Mechatronic systems: fundamentals. London:
Springer.
series sensors and form a basis for the implementation of Isermann, R., Schmitt, J., Börner, M., & Fischer, D. (2004). Control of
actuator, process, and sensor diagnostic systems in auto- vehicle dynamic systems. In IFAC symposium on mechatronic systems.
motive mechatronics. Syndney, Australia.
Isermann, R., Schwarz, R., & Stölzl, S. (2002). Fault-tolerant drive-by-
wire systems. IEEE Control Systems Magazine, 22(5), 64–81.
References Kiencke, U., & Nielsen, L. (2000). Automotive control systems. Berlin,
Heidelberg: Springer.
Alleyne, A., & Hedrick, J. K. (1995). Non-linear adaptive control of active Majjad, R. (1997). Estimation of Suspension Parameters. In IEEE
suspensions. IEEE Transactions on Control Systems Technology, 3(1), international conference on control applications. Hartford, CT.
94–101. Nelles, O. (2000). Nonlinear system identification. Berlin: Springer.
Börner, M., Andréani, L., Albertos, P., & Isermann, R. (2002). Detection Patton, R. (1998). Fault diagnosis in dynamic systems. Cambridge: Prentice
of lateral vehicle driving conditions based on the characteristic Hall.
velocity. In IFAC worldcongress 2002. Barcelona, Spain. Pyper, M., Schiffer, W., & Schneider, W. (2003). ABC—active body
Börner, M., & Isermann, R. (2003). Supervision, fault detection, and control. Verlag Moderne Industrie, Landsberg/Lech.
sensor fault tolerance of passenger cars. In Fifth IFAC symposium on Rajamani, R., & Hedrick, J. K. (1995). Adaptive observers for active
fault detection, supervision and safety of technical processes, SAFE- automotive suspensions: theory and experiment. IEEE Transactions on
PROCESS. Washington, DC, USA. Control Systems Technology, 3(1), 86–93.
Börner, M., & Isermann, R. (2005). Model-based detection of critical Stölzl, S., & Isermann, R. (1999). Online supervision of fault-tolerant
driving situations with fuzzy logic decision making. Control Engineer- systems for safety-related applications. In ESREL0 99, European safety
ing Practice, 14(5), 527–536. and reliability conference, 13–17 September, Munich.
BuXhardt, J., (1993). Parameter adaptive semi-active shock absorbers Van Zanten, A. T., Erhardt, R., Landesfeind, K., & Pfaff, G.
based on nonlinear models. In Second European control conference (2000). Vehicle stabilization by the vehicle dynamics control system
ECC ‘93, 28.06.-01-07.1993. Groningen (NL). ESP. IN First IFAC conference on mechatronic systems. Darmstadt,
Chen, H., & Guo, K. (2001). An LMI approach to multiobjective RMS Germany.
gain control for active suspensions. In American control conference. Weispfenning, T., & Isermann, R. (1997). Fault detection of
Arlington, VA, USA. vehicle suspensions. In IFAC symposium on fault detection, super
Fischer, D. (2004). Mechatronic semi-active and active vehicle suspen- vision and safety of technical processes SAFEPROCESS.
sions. Control Engineering Practice, 12(11), 1353–1367. Hull, GB.

You might also like