1318 IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS, VOL. 34, NO.
6, NOVEMBER/DECEMBER 1998
Fault Detection and Isolation on a PWM
Inverter by Knowledge-Based Model
Raphael Peuget, Stéphane Courtine, and Jean-Pierre Rognon
Abstract— This paper presents an approach based on knowl-
edge models to detect and isolate faults in a pulsewidth mod-
ulation inverter supplying a synchronous machine. These faults
do not affect the system protections. A diagnosis system which
uses only the input variables of the drive is presented. It is
based on the analysis of the current-vector trajectory and of the
instantaneous frequency in faulty mode. These two methods have
been successfully applied to an experimental system.
Index Terms— Knowledge-based fault detection and isolation,
pulsewidth modulation inverter.
NOMENCLATURE
Electric rotor position.
Rotor speed.
Inverter drive.
Capacitor voltage.
Current reference.
Speed reference.
Transistor .
Phase current measure.
Phase current measure.
Components of current vector in the Concordia
frame.
I. INTRODUCTION
T HE expansion of automation in industrial processes has
led to very complex systems. The costs of unplanned
stops are so high that fault detection, isolation, and even
prediction are more and more economically viable. Several
fault detection methods [1] have been introduced in recent
years, including the following:
1) analytical redundancy [2] (parallel average models [3],
observer structures [4], frequency-domain approaches
[5]);
2) knowledge-based approaches [6]–[8] (fault tree [9]–[11],
expert system, fuzzy logic, neural network).
Paper IPCSD 98–38, presented at the 1997 Industry Applications Society
Annual Meeting, New Orleans, LA, October 5–9, and approved for publication
in the IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS by the Industrial Power
Converter Committee of the IEEE Industry Applications Society. Manuscript
released for publication June 4, 1998.
The authors are with the Laboratoire d’Electrotechnique de Grenoble,
Institut National Polytechnique de Grenoble, 38402 Saint Martin d’Hères,
Cedex, France (e-mail: Jean-Pierre.Rognon@leg.ensieg.inpg.fr).
Publisher Item Identifier S 0093-9994(98)08106-7.
0093–9994/98$10.00  1998 IEEE
Authorized licensed use limited to: IEEE Xplore. Downloaded on October 15, 2008 at 02:12 from IEEE Xplore. Restrictions apply.
This paper investigates the case of faults when the drive
continues to operate in faulty conditions. A diagnosis system
which uses only the input variables of the drive is presented.
It is based on the analysis of the current vector; it is applied to
a synchronous motor drive. This method has been validated in
simulation and successfully implemented on an experimental
system. The paper is organized as follows.
1) In Sections II and III, the system is described and the
faults and their consequences are studied.
2) Two fault diagnosis approaches are analyzed in Sections
IV and V.
3) In Section VI, the way the two detection algorithms are
implemented is described.
4) Simulation and experimental results are presented and
discussed in VII–IX.
II. SYSTEM DESCRIPTION
As shown in Fig. 1, the experimental system is composed
of the following:
1) a dc supply;
2) a pulsewidth modulation (PWM) inverter (inverter fre-
quency 20 kHz);
3) a synchronous motor and its mechanical load;
4) speed and current analog controllers;
5) a PC equipped with a real-time data acquisition card;
the data treatments are performed off line.
III. FAULTS UNDER STUDY
Various types of faults are likely to occur on the different
elements of the drive, including the following:
1) network—one phase missing, supply line to ground;
2) dc bus—a diode of the rectifier off, capacitor breakdown,
voltage sensor fault, no turning off or on of the transistor
used for voltage limitation;
3) PWM inverter—no turning off or on of one transistor;
4) Current sensors faults;
5) Faults in the controllers and the motor (short circuits,
rotor eccentricity, ).
In this paper, only transistor and current sensor faults are
considered. The other defects will be studied in another paper.
When a transistor of the inverter does not turn on, the
phase currents of the machine are no longer sinusoidal. For
example, if phase current ( ) is positive ( ) and the
transistor T2 is inoperative, is connected to the negative bus
by the diode D5, instead of being connected to the positive
PEUGET et al.: FAULT DETECTION AND ISOLATION ON A PWM INVERTER BY KNOWLEDGE-BASED MODEL 1319

Fig. 1. Experimental system.

Assuming that and are given by

(1)

The current is periodically sampled. The rates which repre-

Fig. 2. Concordia transform.
sent the slope of the current trajectory are considered

bus. Consequently, decreases to zero. During this step, the

(2)
output of the current controller ( ) increases to counteract
downfall; as T2 is inoperative, the control is impossible and
reaches its maximum positive value. remains null as long as where are the present and the previous samples.
the current reference is positive. When the current reference In the fault-free case, the motor currents are sinusoidal and,
changes sign, the transistor T2 is no longer used and is consequently, the current vector trajectory is a circle. is
normally controlled. always varying in this case.
As a consequence, this kind of fault leads to the cancellation
of the current in the faulty phase during half of the current
A. Cases Under Study
period, which causes a dc current offset in the faulty phase.
This offset is equally divided between the healthy phases. 1) One Phase-B Transistor is Open: If either T2 or T5 is
The interaction between the current dc component and the open, the phase current is null during half of the current
rotor field generates a pulsating torque at the stator current period, as explained in Section III. Equations (1) and (2) imply
frequency. Moreover, the current dc component generates and
unequal current stresses in the upper and lower transistors
of the inverter, which may cause thermal defects in these (3)
transistors [12]. Two methods are proposed to detect and
isolate the faults. In case of a measurement offset, the measured current is no
1) The first method is based on the analysis of the current longer equal to zero, but constant ( ); then, the relation
vector trajectory in the Concordia frame. between and is
2) The second method uses the instantaneous frequency of
the current vector. (4)
It must be noted that the current sensors are already necessary
to control the drive. 2) One Phase-A Transistor is Open: With the same rea-
soning during the faulty mode, (1) and (2) give
IV. METHOD 1: ANALYSIS OF THE and
CURRENT-VECTOR TRAJECTORY
The Concordia transform is applied to the phase currents and without sensor offset (5)
(Fig. 2). It transforms a three-phase system ( ) into a and with a sensor offset (6)
two-phase system ( ).

Authorized licensed use limited to: IEEE Xplore. Downloaded on October 15, 2008 at 02:12 from IEEE Xplore. Restrictions apply.
1320 IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS, VOL. 34, NO. 6, NOVEMBER/DECEMBER 1998

(a) (b)

(c)
Fig. 3. Current-vector trajectories in faulty modes.

3) One Phase-C Transistor is Open: In this case,

and

(7)

The lissajous figures (Fig. 3), show the different current

trajectories in the different faulty cases: T1, T2, T3, T4, T5,
or T6 open.

B. Fault Diagnosis Fig. 4. Derivative of the current-vector phase s .

1) The observation of the current trajectories (Fig. 3)

shows that (2) is constant during half of the current the derivative of the current vector phase (Fig. 4)
period and equal to one of the constants mentioned before
( ). Therefore, the faulty leg in the PWM
inverter can be localized by only knowing the value of this
constant. The faulty transistor is isolated by determining (9)
which alternation of the current is missing in the faulty phase.
A measurement offset on the current sensor of the faulty phase The current sampling frequency is 20 the current fre-
can also be detected. For example, if T2 is open with a sensor quency. The derivative of the phase current vector is estimated
offset, is equal to , but the relation given by (3) is not (Fig. 5) by
verified; therefore, the sensor offset can be estimated by

(8)

2) If a sensor fault occurs, the measured current is zero and

the sensor fault can be detected because the slope is always
constant and the current is always zero on the current period. The instantaneous frequency is then estimated by

(10)
V. METHOD 2: INSTANTANEOUS FREQUENCY ANALYSIS
In the second approach, the instantaneous frequency of the where is the current sampling period.
current vector is used to detect a transistor fault in the inverter. 1) In faulty conditions, according to (10) and to the fact that
The current vector instantaneous frequency is calculated by is proportional to

Authorized licensed use limited to: IEEE Xplore. Downloaded on October 15, 2008 at 02:12 from IEEE Xplore. Restrictions apply.
PEUGET et al.: FAULT DETECTION AND ISOLATION ON A PWM INVERTER BY KNOWLEDGE-BASED MODEL 1321

Fig. 5. Derivative of the phase current vector.

Fig. 7. Detection of the missing alternation.

Fig. 8. Fault isolation.

Fig. 6. Slope test.

electrical position is equal to (20 measures
per current period).
the instantaneous frequency becomes zero. To detect an
2) The second task is presented in Fig. 6; it calculates the
inverter fault, it is sufficient to test whether is lower
ratio on the last two measures and returns six Boolean
than a threshold .
signatures (S1–S6). To prevent false alarms, must
2) If a measurement offset is present on the current sensor,
remain unchanged during five successive measures (a
the estimated frequency is not zero and depends on the
quarter of the current period) so that symptoms S2, S4,
difference between and ( )
and S6 are validated.
(11) 3) If either S2, S4, or S6 is true, a procedure tests which al-
ternation of the phase current is missing and returns nine
where and are constants. Boolean signatures. An example is given in Fig. 7 for
phase- faults. Fault isolation is performed as explained
By comparing the instantaneous frequency and the current
in Fig. 8.
fundamental frequency obtained by the rotor position sensor
(it is used for the control of the drive), an inverter fault,
with or without current offset, can be diagnosed: if B. Method 2
or then an inverter fault is present (the The second detection algorithm is composed of the follow-
threshold is equal to 10 Hz). ing two tasks.
This approach (instantaneous frequency) is easier to imple- 1) The first is the measurement of filtered and currents
ment than the first one (current-vector trajectory). It allows us and calculation of the Concordia components and
to detect an inverter fault, but not to isolate which transistor . These components are calculated when the rotor
is inoperative. electrical position is equal to (20 measures
per current period).
VI. IMPLEMENTATION 2) The second is the estimation of the instantaneous fre-
quency of the current vector and comparison with .
A. Method 1
The detection algorithm is composed of the following three VII. SIMULATION RESULTS
tasks. The whole system (supply-inverter-machine-control-
1) The first one is the measurement of filtered and detection system) has been simulated using the
currents and calculation of the Concordia components Xmath System build package. The parameters of the
and . These components are calculated when the rotor experimental system are given in Table I.

Authorized licensed use limited to: IEEE Xplore. Downloaded on October 15, 2008 at 02:12 from IEEE Xplore. Restrictions apply.
1322
TABLE I
PARAMETERS OF THE EXPERIMENTAL SYSTEM
Fig. 9. T5 open, sensor offset on phase b: simulation results.
The first detection procedure (Method 1) is illustrated in
Figs. 9 and 10. Transistor T5 fails at ms. The offset
on sensor is equal to one ampere and the speed reference
is 150 rad/s.
Fig. 9(a) and (b) shows the motor current and the result
of the detection, respectively. As shown in Fig. 9(b), the
diagnosis system detects and isolates the fault one current
period after its occurrence. This is due to the fact that one
current period is necessary to detect which alternation of the
phase current is missing.
Authorized licensed use limited to: IEEE Xplore. Downloaded on October 15, 2008 at 02:12 from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS, VOL. 34, NO. 6, NOVEMBER/DECEMBER 1998
Fig. 10. Simulated current trajectory for T5 open and sensor offset on phase
b.
Fig. 11. T5 open and sensor offset on phase b: estimation of the offset
(simulation).
Fig. 10 shows the current-vector trajectory; the trajectory is
characteristic of a problem on transistor T5 (Fig. 3). The offset
is correctly estimated as shown in Fig. 11.
VIII. EXPERIMENTAL RESULTS
In order to illustrate that the two diagnosis methods are in-
dependent from the PWM inverter frequency, the experiments
have been realized with two different PWM frequencies, 2
and 20 kHz. The command of a transistor is forced to zero
to simulate the faults to be detected (the transistor does not
turn on).
A. 2-kHz Inverter Frequency: Method 1
The case considered in Figs. 12–14 corresponds to a fault
on transistor T2 occurring at ms, with a 3-A offset on
the phase- current sensor, a 2-kHz inverter frequency, and a
70 rad/s speed reference.
Fig. 12(a) and (b) shows the motor currents and ,
respectively; Fig. 13(a) and (b) shows the filtered and sampled
current and the result of the detection, respectively. The
fault is detected at 270 ms; its influence is observed at 250 ms.
In accordance with the simulation results (Section VII), the
fault detection system detects and isolates the fault one current
period after its occurrence.
Fig. 14 shows the current vector trajectory; it is charac-
teristic of a transistor fault on leg . The influence of the
PEUGET et al.: FAULT DETECTION AND ISOLATION ON A PWM INVERTER BY KNOWLEDGE-BASED MODEL 1323

Fig. 14. T2 open; sensor offset on phase b: experimental current vector

trajectory.

Fig. 12. T2 open; sensor offset on phase b: experimental current waveforms.

Fig. 15. T2 open; sensor offset on phase b: estimation of the offset (exper-
imental result).

Fig. 13. T2 open; sensor offset on phase b: experimental sampled current

and detection results.

offset on the current sensor can be observed, since the origin

( ) does not belong to the straight line. As shown
in Fig. 15, when a transistor fault is present, the value of the
sensor offset is correctly estimated (about 3 A).

B. 2-kHz Inverter Frequency: Method 2

The second approach (instantaneous frequency) has been
applied to the previous fault situation (fault on transistor T2
at 250 ms). Fig. 16(a) shows the filtered and sampled phase- Fig. 16. T2 open; sensor offset on phase b: experimental result.
current. In Fig. 16(b), the instantaneous frequency of the
current vector is presented with the two upper and lower In the fault-free case, the instantaneous frequency is equal
bounds and ; the result of the detection is to ; if a fault occurs, is outside and an
presented in Fig. 16(c). inverter fault is diagnosed.

Authorized licensed use limited to: IEEE Xplore. Downloaded on October 15, 2008 at 02:12 from IEEE Xplore. Restrictions apply.
1324 IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS, VOL. 34, NO. 6, NOVEMBER/DECEMBER 1998

Fig. 19. Fault T3 open: experimental results.

Fig. 17. T3 open: experimental result with 20-kHz inverter frequency.

Fig. 18. T3 open: experimental current-vector trajectory.

Fig. 20. Reply on load torque step.

C. 20-kHz Inverter Frequency: Method 1 D. 20-kHz Inverter Frequency: Method 2

The case considered in Figs. 17 and 18 corresponds to a Fig. 19(a)–(c) shows and motor currents, the instanta-
fault on transistor T3 occurring at ms, without offset neous frequency of the current vector, and the detection result,
on the current sensor, a 20-kHz inverter frequency, and a 70 respectively.
rad/s speed reference. When the fault occurs, the instantaneous frequency of the
Fig. 17(a)–(c) shows the motor currents and , the current vector is periodically equal to zero and the inverter
filtered and sampled and motor currents, and the result of fault is diagnosed.
the detection, respectively. The influence of transistor T3 fault
can be observed at ms; it is isolated at 270 ms. The E. Robustness
current-vector trajectory is presented in Fig. 18 with a ratio To test the robustness of these approaches, we have applied
characteristic of a problem on leg . a load torque step from 7 to 14 N m. The inverter frequency is

Authorized licensed use limited to: IEEE Xplore. Downloaded on October 15, 2008 at 02:12 from IEEE Xplore. Restrictions apply.
PEUGET et al.: FAULT DETECTION AND ISOLATION ON A PWM INVERTER BY KNOWLEDGE-BASED MODEL 1325

The current-vector trajectory (Fig. 22) is a circle and is not

affected by the speed step. In this case, the first approach based
on current-vector trajectory is more robust than the second
approach based on instantaneous frequency.

IX. CONCLUSION
In this paper, we have proposed two approaches based
on the analysis of the current vector. The first one is based
on the analysis of the current-vector trajectory to detect and
isolate faults in a PWM inverter and current sensors. All the
faults (transistor and sensor) have been detected and isolated.
The second one estimates the instantaneous frequency of
the current vector; the isolation of the faulty transistors is
impossible. These knowledge-based methods have been tested
in simulation and applied successfully to an experimental
system.

REFERENCES
[1] R. J. Patton, “Robust model-based fault diagnosis: The state of the art,”
in Proc. IFAC Symp. SAFEPROCESS, Espoo, Finland, 1994, pp. 69–74.
[2] P. M. Franck, “Fault diagnosis in dynamic systems using analytical
and knowledge-based redundancy—A survey and some new results,”
Fig. 21. Response to a speed reference step.
Automatica, vol. 26, no. 3, pp. 459–474, 1990.
[3] C. S. Berendsen, G. Rostaing, and G. Champenois, “How to detect and
localize a fault in a DC/DC converter,” in Proc. IECON92, 1992, vol.
1, pp. 536–541.
[4] C. S. Berendsen, G. Rostaing, G. Champenois, G. Obrecht, and J. Saadi,
“Detection of sensor faults with observer structures in control loops,”
in Proc. IECON’93, 1993, vol. 1, pp. 344–348.
[5] S. Courtine, A. Benamara, and J. P. Rognon, “Perfect and H - 1
optimized fault detection in a DC motor loop,” in Proc. Elelectri-
macs’96, Saint-Nazaire, France, 1996, pp. 599–604.
[6] B. Freyermuth, “Knowledge based incipient fault diagnosis of industrial
robots,” in Proc. IFAC Symp. SAFEPROCESS, Baden-Baden, Germany,
1991, vol. 1, pp. 31–32.
[7] R. Iserman, “Fault diagnosis of machines via parameter estimation and
knowledge processing,” Automatica, vol. 29, no. 4, pp. 815–835, 1993.
[8] L. Zhang, I. B. Aris, and L. N. Hulley, “A knowledge-based system
for on-line fault diagnosis of power inverter circuits for AC machine
drive,” in Proc. EPE’92, 1992, vol. 3, pp. 334–339.
[9] W. S. Lee, D. L. Grosh, F. A. Tillman, and C. H. Lie, “Fault tree
analysis, methods, and applications,” IEEE Trans. Rel., vol. R-34, pp.
194–203, Sept. 1985.
[10] R. Peuget, S. Courtine, and J. P. Rognon, “Fault diagnosis in
Fig. 22. Current-vector trajectory for a speed step. DC/DC converter using fault tree analysis,” in Proc. Int. Symp.
Diagnostics for Electrical Machines, Power Electronics and Drives
IEEE-SDEMPED’97, Carry-le Rouet, France, Sept. 1997, pp. 132–139.
[11] P. Calonnec, T. Derrey, E. Destobbeleer, and L. Protin, “Induction
equal to 20 kHz and the speed reference is equal to 70 rad/s. motors reliability: Use of fault tree,” in Proc. EPE’95, Sevilla, Spain,
In Fig. 20(a), the motor phase current and the estimated 1995, vol. 3, pp. 281–285.
[12] D. Kastha and B. K. Bose, “Investigation of fault modes of voltage-fed
instantaneous frequency of the motor current are presented. inverter system for induction motor,” IEEE Trans. Ind. Applicat., vol.
The first approach with the trajectory of the current vector 30, pp. 1028–1037, July/Aug. 1994.
is not affected by this load torque step. Concerning the second
approach, the estimated frequency of the current vector re-
mains equal to the fundamental frequency during this transient
and there are no false alarms. In a second test, a step change
in the speed reference from 50 to 70 rad/s is applied. The data
acquisition sequence is synchronized to the step. Raphael Peuget received the engineer diploma in
electrical engineering in 1994 and the doctorate
Fig. 21(a) shows , motor currents and current reference degree in 1997 from the Institut National Polytech-
and the estimated frequency of the current vector; Fig. 21(b) nique de Grenoble, Saint Martin d’Hères, France.
He has been an active Member of the Control Re-
shows the fundamental frequency . search Group of the Laboratoire d’Electrotechnique
In Fig. 21, the frequency is well estimated during the de Grenoble, Institut National Polytechnique de
transient, except when the motor current is near zero. To Grenoble. His major field of interest is fault
diagnosis in power electronics converters using
prevent any false alarms, the frequency is estimated only if knowledge-based methods. He currently has an
the absolute value of the current is greater than a threshold. engineering position at Staubli, France.

Authorized licensed use limited to: IEEE Xplore. Downloaded on October 15, 2008 at 02:12 from IEEE Xplore. Restrictions apply.
1326 IEEE TRANSACTIONS ON INDUSTRY APPLICATIONS, VOL. 34, NO. 6, NOVEMBER/DECEMBER 1998

Stéphane Courtine received the engineer diploma Jean-Pierre Rognon received the engineer diploma
in electrical engineering in 1993 and the doctorate in electrical engineering in 1972, the aggregation in
degree in 1997 from the Institut National Polytech- applied physics in 1974, and the habilitation degree
nique de Grenoble, Saint Martin d’Hères, France. in 1988 from the Institut National Polytechnique de
He currently holds a post-doctorate position in Grenoble, Saint Martin d’Hères, France.
the Control Research Group of the Laboratoire He has been with the Electrical Engineering
d’Electrotechnique de Grenoble, Institut National School of the Institut National Polytechnique de
Polytechnique de Grenoble. His major field of in- Grenoble since 1974 and is currently a Professor
terest is safety and reliability of electromechanical of Control and Electrical Engineering. He is also
systems. in charge of the Laboratoire d’Electrotechnique
de Grenoble, Institut National Polytechnique de
Grenoble, and is an active member of the Control Research Group, which
mainly works on vector control of ac machines, robust digital control, and
supervision and fault detection on electromechanical systems.

Authorized licensed use limited to: IEEE Xplore. Downloaded on October 15, 2008 at 02:12 from IEEE Xplore. Restrictions apply.