Professional Documents
Culture Documents
NOMENCLATURE 1. INTRODUCTION
v : longitudinal velocity of ATV The growing demand for enhanced reliability and safety in
vdes : desired longitudinal velocity autonomous control systems has drawn significant pro-
vGPS : longitudinal velocity measured by GPS gress in fault detection and diagnosis (FDD) over several
ωw : angular velocity of wheel decades. In particular, as the control systems for vehicles,
ωe : angular velocity of engine industrial robots and aircrafts have become more com-
h : effective wheel radius plicated and distributed, rigorous research on FDD has paid
α : angle of throttle control motor much attention in these areas of application (Zhang and Li,
β : angle of brake control motor 1998; Yi et al., 2000; Fantuzzi et al., 2003).
ωα : angular velocity of α Most model-based FDD methods in the literature use
ωβ : angular velocity of β residuals (Isermann, 1995, 2006). Residuals are quanti-
Vti : input DC voltage to throttle motor tative differences between estimates based on a nominal
Vbi : input DC voltage to brake motor model and measurements from sensors. The residuals
Iti : input DC current to throttle motor should be robust to avoid a false alarm due to noise and
Ibi : input DC current to brake motor disturbances as well as sensitive to the occurrence of faults
KE : back-EMF constant of DC motor to enhance reliability. In general, the performance of FDD
Ra : armature resistance of DC motor depends on the accuracy of model-based estimates and the
Vtd : input DC voltage to throttle motor driver optimal choice of threshold. Most estimation techniques to
Vbd : input DC voltage to brake motor driver generate these residuals for FDD are classified as the
Dti : input PWM duty ratio of throttle motor driver observer, the detection filter or the parity equation approach
Dto : output PWM duty ratio of throttle motor driver (Chow and Willsky, 1984; Gertler, 1998; El Brouji et al.,
Dbi : input PWM duty ratio of brake motor driver 2006). For determination of the threshold, the automatic
Dbo : output PWM duty ratio of brake motor driver calculation method was recently introduced in the
literature. For instance, there is a threshold determination
by analyzing the stability of a system using linear matrix
*Corresponding author. e-mail: bsong@ajou.ac.kr inequalities (Ding et al., 2003) and another method called
505
506 S. KIM, B. SONG and H. SONG
the double-threshold-statistical-testing method, which sets for an individual DSP to detect and isolate faults of the
the second threshold for the function statically by calculat- corresponding sensors and actuators, depending on charac-
ing the case that a residue exceeds the first threshold (Fang teristics of each hardware model, simultaneous CFDD based
et al. , 1994). However, since there are always noise and on kinematic and kinetic relationships among the measure-
disturbances in dynamic systems in reality, “robustness” ment information communicated via CAN is proposed.
still remains one of the most important issues for FDD Then, the integrated FDD algorithm, which is a two-
(Fang , 1994).
et al. layered FDD structure that is a combination of DFDD and
A vehicle is one of the most well-known distributed CFDD, is proposed to enhance the robustness of fault
systems and contains many processors, sensors and actuators. detection and more specific fault isolation by taking the
The multiple processors communicate via so called ECU benefits of both. Finally, Section 4 will present experi-
and share information via in-vehicle networks, e.g , the . mental results to show the effectiveness of IFDD.
controller area network (CAN). In the literature, appro-
aches of FDD are classified into two groups: decentralized 2. DISTRIBUTED CONTROL SYSTEM
and centralized FDD. The former (DFDD) aims to detect
and identify faults in a single processor that connects 2.1. System Hardware Layout
sensors and actuators. This is performed by analyzing The hardware layout of the autonomous ATV is shown in
residuals and has been applied to many vehicle appli- Figure 1. Additional hardware, such as a laser scanner, an
cations, e.g., a motor (Moseler and Isermann, 2000; Kim, inertial measurement unit (IMU) and a global positioning
2007; Benbouzid et al. , 2007), a throttle and brake valve system (GPS), are implemented. Furthermore, rotary en-
system (Pfeuer, 1997; Isermann, 2006) and sensors (Ding coders are used to measure the engine and wheel speeds for
at al. , 2004). In general, the DFDD enables us to detect the longitudinal control, and two DC motors and rotary en-
fault more immediately and identify it in terms of more coders are implemented to control the throttle and brake
specific hardware components due to the simplicity of the valves.
mathematical model for state estimation of the single If the control system is centralized, i.e., all data and
module. However, since it depends on single residual and information are transmitted to single main processor, the
threshold, it is sensitive to noise and uncertainties in the rate of data loss and time delays increase due to the limits
system and may result in a false alarm (Fang et al. , 1994; of the computational capability of the processor and data
Isermman, 1995). communication. To overcome this disadvantage of centrali-
On the other hand, the centralized FDD (CFDD) focuses zed control systems, seven DSPs are distributed over the
on performance degradation resulting from faults on the system, as shown in Figure 2. The DSPs send either pro-
level of a system or vehicle, not an individual component. cessed or measurement data via the CAN bus and share
That is, it monitors the performances of multiple pieces of information among them. For instance, while DSP 3 is
hardware and identifies faults using kinematic, kinetic and/ connected with the engine and wheel speed sensors and
or heuristic relationships among individual modules. For transmits the message ω and ω via the CAN bus, infor-
e w
instance, CFDD has been applied to longitudinal and/or mation measured by the IMU is transmitted from DSP 5.
lateral control for automated vehicle systems (Yi ,
et al. The velocity and absolute position of the ATV are received
2000; Song and Hedrick, 2005; Rajamani et al. , 2007). from GPS, and DSP 7 receives obstacle information from a
Since it uses multiple measurements coming from distri- laser scanner. DSP 2 receives all messages necessary for
buted processors, the robustness for FDD can be enhanced, longitudinal control through the CAN bus and calculates
and this is why CFDD is applied to safety critical system, both the desired throttle and brake control motor angles .
such as automated highway systems (AHS). However, it is DSP 1 includes a fault manager, which gathers all CAN
still limited to isolating the fault in the terms of the specific
component or part where the fault occurred (Isermann,
2006).
The contribution of this paper is to propose an integrated
FDD (IFDD) using the combination of the two approaches
above to achieve more robust fault detection and more
specific fault isolation. Furthermore, the proposed FDD
will be applied to a CAN-based distributed system for the
longitudinal control of an autonomous all-terrain vehicle
(ATV). The remainder of this paper is divided into three
sections; Section 2 will describe the CAN-based distributed
control system consisting of multiple DSPs, sensors and
actuators and define the possible faults considered in this
study. In Section 3, three approaches will be explained in
detail: DFDD, CFDD and IFDD. While DFDD is designed Figure 1. Hardware layout for autonomous ATV.
INTEGRATED FAULT DETECTION AND DIAGNOSIS SYSTEM FOR LONGITUDINAL CONTROL 507
1 DC motor Additive
- abrupt
H/W (electrical)
- change inner resistance
H/W (electrical)
2 Encoder Additive
- abrupt - exchange any bits of
digital signals
H/W (mechanical)
3 Actuator Additive
- abrupt - impose additional fric-
tion to motor shaft
4 CAN Additive
- abrupt
H/W (electrical)
- take CAN plug out
Figure 2. Schematics of data communication among DSPs
via CAN. 5 Wheel speed Multiplicative S/W (emulator)
encoder - abrupt - add multiplicative error
DSP 2 DC motor BI
V different DSPs (or modules). The last method is called
Brake encoder β IFDD, which combines the two methods to gain the bene-
actuator fits of both.
motor driver BO
D
respectively, as follows.
R1 V t V̂ t R2 V t V̂ t
= ti ( )– ti ( ), = (2) bi ( )– bi ( )
where all duty ratios of the PWM are monitored for FDD.
comes relatively easier than that of CFDD. Third, the faults of encoders used to measure the output
angles of the throttle and brake motors are considered. In
3.1.1. Residual generation for DFDD this study, an absolute type of rotary encoder is used to
In order that faults in the throttle and brake actuators are overcome the fluctuation problem of the zero point of
considered in the framework of DFDD, the residual is encoders due to vibration of the ATV. Therefore, the out-
generated based on an independent model of each hard- puts of encoders are given as gray BCD codes of 10 bits,
ware component. As actuators to control throttle and brake which are generated by digital signals from 10 wires for
valves, two DC motors including a motor driver and an each encoder. Accordingly, if any fault occurs in one or
encoder are used and connected to DSP 2 (in Figure 2). As more wire, then the output value of the encoder changes
shown in Figure 3, there is a position controller to track the discontinuously. Using the fact that the amount of change
desired position angle resulting from a longitudinal con- in the output angle value is limited to some boundary under
troller in DSP 2, and the pulse width modulation (PWM) the no fault condition, the residual for each encoder is
signal corresponding to the command input calculating defined as follows.
from the position controller is supplied to the motor driver. R5 α· R6 β· (4)
Finally, the amplifying PWM is sent to the DC motor, and = , =
the motor angle measured by the encoder is fed into the Finally, if other faults which are not considered above
position controller. Among these components, faults of the may occur in components of throttle and brake actuators, it
DC motor, motor driver, encoder and position controller is expected that the desired performance of actuators is not
will be considered here. achieved. So, the residual for each of actuators using errors
First, to generate the residual related with the DC motor, between desired and measured angles of DC motors can be
the following mathematical model is used to estimate the defined as
electrical characteristics of the DC motor (Moseler and R7 α α R 8 β β (5)
Isermann, 2000). = – des , = – des
estimated using least squares estimation based on experi- 3.1.2. Determination of threshold
mental data. Then, based on the average model of the It is challenging to set an optimal threshold of a residual for
motor driver, ti( ) is approximated as V t V D t ,
V t ti ( ) td ⋅ to ( ) FDD satisfying both sensitivity and robustness. In general,
where td and to( ) are measured. Finally, the time
V D t while more robustness can be achieved by setting a higher
responses of V̂ t and ti( ) are compared in Figure 4. If
ti ( ) V t threshold, sensitivity may be degraded. In order to over-
there are either electrical or mechanical faults in the DC come this drawback, a double-threshold-statistical-testing
a discrepancy between V̂ t and ti( ). Therefore, the ti ( ) V t 1994). This method uses two separate tests with two levels
residuals for the throttle and brake DC motors are defined, of thresholds to facilitate the trade-off between detection
power and false alarm probability. The fault is confirmed
by the second threshold for the function statically calcu-
lating the case that a residual exceeds the first threshold
established ahead. Mathematically, this can be expressed as
follows.
⎧
R ε 1 ε 2 F R H for = 1, ..., 8
if Pr ( i > i )> i ( i
(6)
)=
i
FR L
⎨
⎩ Otherwise, ( i )=
Figure 4. Time responses of V̂ t and V t under no fault ti ( ) ti ( ) where εi1 and εi2 are the first and second thresholds,
condition. respectively, and ( i) is the fault detection flag.
F R
INTEGRATED FAULT DETECTION AND DIAGNOSIS SYSTEM FOR LONGITUDINAL CONTROL 509
Throttle L L H L L L H L Driver ( 2)
F D
actuator H L L L H L H L Encoder ( 3)
F D
L L L L L L H L Controller ( 4)
F D
L H L L L L L H Motor ( 5)
F D
Brake L L L H L L L H Driver ( 6)
F D
actuator H L L L L H L H Encoder ( 7)
F D
L L L L L L L H Controller ( 8)
F D
isolated by fault signature, as shown in Table 2. That is, value in Figure 6(b). Similarly, ( 1) also turns to at 2
H F R H t
Rotary
Wheel speed L H L L encoder ( 2) f C
encoder
H H H L CAN ( 3) f C
GPS L H L H CAN ( 4) f C
cation period is used to generate the residual. That is, since CAN, and the corresponding fault can be isolated by the
the communication period of a CAN message can be fault signature (Yi et al., 2000; Rajamani , 2007).
et al.
measured using a timer in DSP 2 for longitudinal control, Based on the fault signature in Table 3, five possible faults
the residuals for the CAN are defined as corresponding to fault flags are detected and isolated.
r3 ∆ T ∆ T r4 ∆ T ∆ T (9) The performance of CFDD is tested experimentally for
= ref – w , = ref – GPS
the case of continuous packet loss of a CAN message, as
where ∆ is the reference communication period of CAN
Tref described in Table 2 (No. 4). After intentionally blocking a
message, and the communication periods of the wheel CAN message of ω for about 5 seconds starting at 1 in
w t
speed and GPS messages are defined respectively as follows: Figure 7(a), the corresponding residual is observed. The
∆T t[ ω (k ) ]−t[ ω ( k ) ] for = , residuals 1, 2, and 3 increase and go over each threshold
r r r
j j j –1 j w GPS
after 1 second, as shown in Figure 7(b)~(d). Then, each of
Next, as done in DFDD, a similar threshold to detect a the probability functions for 1, 2 and 3 exceeds the second
r r r
fault is determined based on the double-threshold-stati- threshold, as shown in Figure 8(a). Finally, three fault
stical-testing method, and fault detection flags are express- signals, ( ), become , as shown in Figure 8(b), and
f ri H
⎧ if Pr ( ri > δ i1 ) > δi2 Therefore, the fault in the CAN message of the wheel
f ( ri ) = H
⎨ for = 1, ..., 4
i (10) speed is detected and diagnosed successfully.
⎩ Otherwise, f (ri )=L
It is necessary to remark that IFDD can be expanded to ment of throttle motor changes suddenly with large
apply to sensors if more residuals are defined for the more amplitude. Then, the fault flag ( 3) coming from DFDD
F D
other is to show how the robustness of FDD can be enhan- 9(d). Therefore, the fault of the encoder in the throttle
512 S. KIM, B. SONG and H. SONG
control motor system is doubly detected by both DFDD IEEE Trans. Automatic Contol, AC-29, 7, 603−614.
and CFDD. Moreover, IFDD diagnoses the fault specifical- Ding, E. L., Fennel, H. and Ding, S. X. (2004). Model-
ly as a throttle motor encoder problem based on Table 4. based diagnosis of sensor faults for ESP systems. Control
In the second scenario, it is supposed that temporary Engineering Practice, 847−856.
friction is imposed to the throttle motor shaft from t1 to t2 in Ding, S. X., Zhang, P. and Frank, P. M. (2003). Threshold
Figure 10(b) (refer to Table 2~No. 3). The fault is detected calculation using LMI-technique and its integration in
and isolated as one of the throttle actuator by DFDD as the design of fault detection systems. Proc. IEEE Conf.
fault flag F(D1) turns to H at t3 in Figure 10(c) (refer to Decision and Control, Maui, Hawaii, USA.
Table 4). However, the error between the desired and El Brouji, H., Poure, P. and Saadate, S. (2006). A fast and
measured ωw remains small over the period that the fault reliable fault diagnosis method for fault tolerant shunt
occurs in Figure 10(a); thus, the fault flag f (C1) of CFDD three phase active filter. IEEE ISIE, Montreal, Canada.
does not become H in Figure 10(d). This means that the Fang, X., Gertler, J., Kunwer, M., Heron, J. and Barkana, T.
motor fault detected by DFDD does not degrade the per- (1994). A double-threshold-testing robust method for
formance of the longitudinal controller much. Therefore, fault detection & isolation in dynamic systems. Proc.
the fault is classified as tolerable and isolatable according ACC, Baltimore, Maryland. 1979−1983.
to Table 4. In this scenario, the imposed friction is not the Fantuzzi, C., Secchi, C. and Visioli, A. (2003). On the fault
fault but a disturbance. This experimental result implies detection and isolation of industrial robot manipulators.
that a false alarm may be generated due to noise and dis- Proc. IFAC Symp. Fault Detection, Supervision and Safety
turbance if only DFDD is used and/or its sensitivity is high. for Technological Processes, Washington.
However, the two-layered structure of IFDD including Gertler, J. J. (1998). Survey of model-based failure detection
DFDD and CFDD allows us to minimize the false alarms and isolation in complex plants. IEEE Control Systems
due to uncertainties and disturbances. Magazine, 3, 11.
Isermann, R. (2006). Fault Diagnosis Systems. Springer.
5. CONCLUSIONS Berlin.
Isermman, R. (1995). Model-based fault detection and
The integrated FDD combining DFDD with CFDD was diagnosis methods. Proc. ACC, Seattle, Washington,
applied to a CAN-based distributed system for an auto- 1605−1609.
nomous ATV. While the DFDD was designed to detect a Kim, N. (2007). Rotor fault detection system for inverter
fault in hardware connected to a single DPS, the CFDD driven induction motors using current signals and an
was proposed to the monitor performance degradation encoder. J. Power Electronics 7, 4, 870−881.
resulting from faults at the level of a system or vehicle, not Moseler, O. and Isermann, R. (2000). Application of model-
an individual component. By integrating both methods, the based fault detection to a brushless DC motor. IEEE
proposed IFDD algorithm allows us to perform more robust Trans. Industrial Electronics 47, 5, 1015−1020.
fault detection as well as more specific fault isolation. The Pfeuer, T. (1997). Application of model-based fault detec-
performance and effectiveness of IFDD was validated tion and diagnosis to the quality assurance of an auto-
experimentally for the given possible faulty scenarios. motive actuator. Control Engineering Practice 5, 5,
703−708.
ACKNOWLEDGEMENT−This work was supported by grant Rajamani, R., Howell, A. S., Chen, C., Hedrick, J. K. and
No. R01-2006-000-11373-0 from the Basic Research Program of Tomizuka, M. (2007). A complete diagnostic system for
the Korea Science & Engineering Foundation. automated vehicles operating in a platoon. IEEE Trans.
Control Systems Technology 9, 4, 553−564.
REFERENCES Song, B. and Hedrick, J. K. (2005). Fault tolerant nonlinear
control with applications to an automated transit bus.
Benbouzid, M. E. H., Diallo, D., Zeraoulia, M. and Zidani, Vehicle System Dynamics 43, 5, 331−350.
F. (2006). Active fault-tolerant control of induction motor Yi, J., Alvarez, L., Howell, A., Horowitz, R. and Hedrick,
drives in EV and HEV against sensor failures using a J. K. (2000). A fault management system for longitudinal
fuzzy decision system. Int. J. Automotive Technology 7, vehicle control in AHS. Proc. ACC, 1514−1518.
6, 729−739. Zhang, Y. and Li, X. R. (1998). Detection and diagnosis of
Blanke, M., Kinnaert, M., Lunze, J. and Staroswiecki, M. aircraft sensor and actuator failure. IEEE Trans. Aero-
(2003). Diagnosis and Fault-tolerant Control. Springer. space and Electronic Systems 34, 4, 1302−1312.
Chow, E. Y. and Willsky, A. S. (1984). Analytical redun-
dancy and the design of robust failure detection systems.