10.1016@s1474 66701751984 9

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

C"I" !"Ig'" cD 1F.

\C 11,,, Tril'lllli;d \,'"rld


(:()Ilgn:~ ... Ltllillll. F,,\ollia. t'SSR. Iqql)

DYNAMIC MODEL BASED INCIPIENT FAULT


DETECTION CONCEPT FOR ROBOTS
J. Wl1nnenberg and P. M. Frank
L'II/"I'I\//Y '1 n/li.,blllg. Fllcligl'hlfi ,\IRT. FE 'Y. Bi.'/Ill1l'ck,'/m.\\(' 81 IW.
D- -I/ ()() DII/.,fllng I, FR(;

A bstract. The paper discusses two approaches to fault detection in robotic systems. The first ap-
proach consists of the use of a linear Luenberger state observer. This observer generates a robust
residual (innovation) that is maximally sensitive to the faults that have to be detected while remaining
unsensit.ive to the neglected non-lineal~ties. It therefore serves as a fault indicator and allows a fault
detection and unique fault isolation. The computation ally simple design and application procedme
is outlined and illustrated by an example. The second part of the paper shows how a fully nonlinear
observer for fault detection in robots can be designed and applied. Because of its non-linear observer
structure the estimation error dynamics of this fault detection observer is linear and t.he choice of the
observer feedback is particularily simple. Its residual is only a function of the faults and the model
uncertainties. The fault detection abilities are significantly improved compared to the linear residual
generation procedure. Simulation results show the successful implementation.

Keywords. Failure Detection; Nonlinear Filtering; Observers; Redundancy ; Robustness; Robots; Sen-
sor Failures; Analytical Redundancy;

1 INTRODUCTION The second part of the paper presents a fully non-linear


design that considers all system non-linearities. Because
In highly automated systems the demand for powerful su- of the structure of tile non-linearities in robots the choice
pervision concepts is rapiclly increasing (see e.g. Frank of the observer feedback is particularily simple and the
(1989) and Patton et al.(1989)). An appropriate tool to estimation error dynamics are linear. The procedure is
achieve reliable on-line supervision in dynamic systems therefore applicable under each operating condition of the
is the use of analytical redundancy. The analytical re- robot and the generated residuals depend only on mod -
dundancy concept considered in the proposed approach is elling errors and faults while being unaffected by the sys-
based on the nonlinear dynamic model of the physical sys- tem non-linearities. Both approaches will be outlined in
tem. Under consideration of this model a Luenberger state the following.
observer is designed that generates an error signal (resid-
ual) that indicates deviations between the dynamic model
and the physical system. Assuming a sufficiently exact 2 LINEAR FAULT DETECTION
model it serves as a fault indicator and allows an on-line
system supervision independent of the trajectories driven It is assumed that the system under consideration can be
by the system. The use of a Luenberger state observer described by a discrete linear state space model
is of particular advantage if incipient faults in the system (DIn - A)Xk = BUk + Edd k + KdJb Xk=O = Xo (1)
shall be detected rapidly. They offer the advantage of low
computational expenditure together with a fast response to (2)
the faults which allows on-line supervision with low need Xk represents the n-dimensional state vector , A the system
in computer power. dynamics matrix, Uk the r -dimensional known input sig-
This fault detection procedure is applied to a model nal with the corresponding input distribution matrix B.
of the robotic manipulator Manutec R3. The simulation The term Edd k describes an s - dimensional unknown input
results show that the procedure is applicable to the detec- signal dk with known input distribution matrix Ed acting
tion of sensor faults, actuator faults as well as component directly onto the system dynamics equation (1). The ex-
faults. pression Kdh is a representation of the faults acting on the
The paper is divided into two parts. The first part system dynamics such as actuator or component faults and
shows an approach of fault detection with linear fault is described by the known input distribution matrix Kd and
detection filters that are robust wi th respect to the sys- the I-dimensional unknown fault signal Jk. C is the mea-
tem non-linearities as well as disturbances while remaining surement matrix of the system and Yk the m-dimensional
maximally sensitive to the faults that are to be detected. measurement signal that is available for further treatment.
A perfomance index is introduced that considers the ro- The disturbances acting as unknown inputs onto the mea-
bustness against deviations from the model as well as the surement matrix are described by the expression Emdk
fault detection sensitivity simultaneously. The resulting where the s· - dimensional signal dk is again assumed to be
observer algorithm is computationally simple and requires unknown. The faults that alter the measurements, called
low need in on-line computer power. instrument faults, are described by the term Km with un-

61
known transient of the Z"-dimensional fault event vector and
k
The matrices A, B, C, E d , K d , Em, and Km are real val-
ued and of appropiate dimension. In the following they
will be named nominal system matrices. In represents a (12)
n-dimensional unity matrix. The operator D denotes the
shift operator (D<Xk := Xk+d with the constant sampling
time T. The initial condition Xk=O of the System is abbre-
viated by Xo. Based on this description the residual generation will be
Note that the System equations ( 1) and (2) also reflect outlined.
all possible parameter variations in the system. In order to
2.1 Residual Generation
show this it is assumed that the actuaZ system is described
by In order to generate a scalar residual signal rk one has
(3) to check if the state equation (8) holds for the available
input and output data using the parity space approach
Here Aa and Ba represent the actual matrices currently de-
(see e.g. Chow and Willsky (1984) ). This is accomplished
scribing the system. The deviations from the nominal ma-
by calculating the equation
trices which are caused e.g. by the system non-linearities
are expressed by
Aa = A + 6A (4 )
Ba = B + 6B. (5)
The System dynamics equation (3) can, therefore, be
rewri t ten as on-line at each sampling time k. The vector v T that has to
be found is called the residual generator and has to meet
the following requirements. In order to make the residual
independent of any initial condition Xk - s the vector v T has
or to be chosen such that the condition
(7)
which is equal to the description chosen in eqn.(l). An
equivalent procedure is applicable for deviations in the C
and D matrices that leads to the matrix Em.
v
T
[ g:, 1 = 0 (14)

In the following this system desription will be used for


CA'
the design of a state observer for residual generation. The
state eqn. (1) and (2) can be rewritten as is fulfilled. Since the residual has to be affected by the
fault
(15)

HQ~'" H, ( "':j:;' ), must hold. If almost any fault event signal shall be de-
tected the above requirement must be strengthened to
v T H 3 ,; of- 0 (16)
with
dk_S)
dk-,+I (1k_')
1k-,+1

K~.l
H H
2 ... + 3 ... (8)
o
(
dk 1k o
i " ,, ... , ,'
where s is the considered time horizon and
o
(17)

HQ = [ i:, 1'
CA'
(9)
with Km.; the i-th column of the matrix Km and

H 3 ,; =
C~d"
CAKd , CK d ,;
0
o 0] i=l, ... ,s

CAs~IK
[
D o ~ C~d,;

J
d,' .. 0
CB D o (18)
CAB CB D o ( 10) with K d ,; the i-th column of the matrix K d . This guar-
antees that almost any fault acts onto the residual. The
CB remaining once that do not show up in the residual be-
cause they fulfil a certain time function are called detec-
tion concealed (see Wiinnenberg (1989)). This shows which
conditions must be fulfilled to generate a residual that is
o sensitive to faults. The ideal solution for a robust residual

:.1
(11)
that is not affected by any unknown input vector dk is
v T H2 = O. (19)
Necessary and sufficient conditions for the existence of this
solution in terms of a direct observer formulation avoid-
ing the parity space approach can be found in Frank and
Wiinnenberg (1989).

62
If the ideal solution does not exist the most desired is with
to find an optimal solution which meets a certain perfor-
mance index. This index must contain a measure for the
effect of the unknown input d k and for the effect of the
VIC + vzCA + V3 CAz + . + V,CA'-l
VzC + v 3C A + + v,C A z S
-
1
fault s fk. Since no assumptions or restrictions are made T =- V3C + + v,CAs- 3 (29)
about the time history of the disturbance and the faults,
[
the performance index can only take into account the dis-
v,C
tribution of the signals into the system. As a proper choice
we therefore state the problem: and the observer output matrices by
Find a vector WT such that the performance index
wTVoHzCdH!vl w LI (0,0, ... , 0, 1) (30)
P- (20)
- w T VoH 3C I Hlvl w Lz Vs (31)
becomes minimal. Here Vo is a base for all possible solu- - v, D. (32)
tions to eqn.(14) . Hence, WT singles out the best vector
In the fault and disturbance free case and neglecting the
(2 1)
initial conditions the transformation between the system
from all possible solutions Vo. The performance index and the observer state variables is described by
was introduced by the authors in Wi.innenberg and Frank
( 1988 ) for the special case that Cd and Clare uni ty ma- (33)
trices. Clearly, the choice of the filter degree s has an
influence onto this optimal solution. The least square char- From the dynamics matrix Fo it is visible that the res ulting
acter of the performance index shows that an increase of s observer is a dead-beat observer for fault detection. The
generally reduces the value of the performance index. An matrix LI shows that the observer is in observer canonical
appropriate value for s has to be selec ted by the designer. form. For a pole placement for the observer the feedback
In addition, the choice of the weighting matrices Cl and Cd law
are of fundamental importance for the result of the design . zk+1 = Fozk + GoYk + J ollk - Hrk (34)
One physical interpretation of these weighting matrices will
be given at the en d of this section.
Tk = L1z k + LZYk + L 311k (35)
The solution to the optimization problem according to is introduced which allows an arbitrary pole placement of
eqn.(20) is derived by a differentiation of the performance the observer, as shown in Wiinnenberg (1989) . The final
index with resp ect to wT. The necessary condition for a observer is then d escribed by
solution

(22)
(37)
shows that the problem reduces to a generalized eigenvalue and the choice of the feedback matrix H becomes particu-
- eigenvector problem. The minimal eigenvalue is the opti- larily simple. Denoting the coefficients of the characteristic
mal value of the performance index and the corresponding polynomial of the observer
eigenvector w T is the selector for the optimal residual gen-
erator v T
The residual generator according to eqn.(13) can be
rewritten in state space form with ai the matrix H must be chosen to be

(23) (39)

(24) which is equivalent to the well known observer design for


a single output system in observer canonical form.
where the conversion is described as follows: The vector
With consideration of the pole placement abilities it be-
UT = (VO,Vl ,v s ) (25) comes visible that this parity space approach is also appli-
1 • •

cable for continuous time systems when a continuous time


is divided into s single row vectors Vi. By a direct refor- observer has to be designed. The difference is a simple ex-
mulation of the residual generator the observer matrices change of t he shift operator D by the differential operator
dldt and an appropriately chosen characteristic equation
for the observer.
° ° The remaining problem is the appropriate interpreta-
° (26 )
tion of the weighting matrices Cl and Cd. One possible

° ° way is to assume the disturbance d k and the fault fk to be


a stochastic signal with normal distribution and zero mean.
An appropriate performance index is then the ratio of the
° ° variance of the residual Td affected only by the unknown
input (Jk = 0) to the variance of the residual TI affected
(27) only by the faults (d k = 0) expressed by

E[T~l
(40)
p = E[rJ]
J (28)
where E detones the expectation operator. As shown in
Frank and Wi.innenberg (1989) the pole placement of the
observer acts like a low-pass filtering via a single input
filter onto the residual. The transfer function of this low

63
pass filter is completely determined by the characteristic
polynomial of the observer and can be described with help (48)
of the z-transfonnation by

G(z)= l+a,_l+ .. . +ao where the coefficients a and b reflect how strong the fault
(41) event signal are correlated either already in the model or
z' + a,_lz,-l + ... + ao
by the low pass filtering in the observer. Therefore the
Because of the li nearity of the system this low pass filtering poles are not chosen yet.
can be assumed to act onto the system inputs and the The design result is a performance index p = 0 inde-
variance of r d can be expressed by pendent of the values of a and b except a = b = l. The
solution is therefore independent of the chosen observer
(42) poles. The calculated residual generator described by
The matrix Cd is t hen defined by rk = (Yk - Yk-l) - (Yk- l - Yk - 2) - (Uk_2 - Uk-l) (49)
corresponds in the case of an analog system description to a
double differentiator. Th erefore it becomes visible that the
residual generator is insensitive to constant disturbances
where the process die is generated from the process dk by
acting onto the system input while remaining sensitive to
low pass filtering of each component di of d according to
all fault signals except if Jk can be described by a first
d:(z) = G( z)di(z ). (44) order polynomial in k. Rewriting the observer ill state
space form results in
An equivalent consideration holds for the characteristics of
the fault event signal which shows that the performance (50)
index fulfills the desired properties. The design procedure
and
is summarized by:
(51)
l. Choice of an appropriate system description includ- If the observer shall contain a double pole at Ai = 0.5 the
ing the stochastic characteristics of the unknown in- feedback matrix H results in
puts and the faults
H = (0.25 , -If (52)
2. Choice of the order s of the observer such that the observer is finally described by
3. Choice of the observer poles
(~ -0{25) Zk +( 0~715 ) Yk +( ~l ) Uk (53)
4. Calculation of the weighting matrices Cd and Cl un-
and
der consideration of the observer poles
rk = (0, l )zk + Yk (54)
5. Solution of the eigenvalue-eigenvector problem which completes the design.
6. Determination of the observer equations.
3 NONLINEAR FAULT DETEC-
2.2 Application Example TION
A single axis of a robot can be described by a single inte- Many results obtained in the area of fault detection in lin-
grator if the torque is considered as input and the velocity ear systems can be transfered to nonlinear syst.ems . This
as state variable and output. The gravitational, Corio- may become a very complicated procedure if a most gen-
lis and centripetal forces and the friction can be modelled eral nonlinear system description is assumed. In this ap-
as one disturbance acting like an input torque. The fault proach we limit ourselves to the special case that the sys-
that shall be detected is a fault in the velocity measure- tem nonlinearities acting onto the dynamics equation can
ment. The scaled discretized model reduces to a first order be expressed as a function of the measured output signal
system described by and follow thereby a way outlined by Birk and Zeitz ( 1988).
The continuous time system dynamics equation is asswned
(45)
to be of the form
(46)
Dx(t) = Ax(t )+F(y(t), u(t))+Ed(t )+ /{b(t))J(t) (55)
The choice of the weighting matrices Cd and Cl is of par-
ticular importance because a complete decoupling of the and the system output equation of the Form
residual from the unknown input dk for arbitrary time y(t ) = Cx(t) + K m(x(t))J(t). (56)
functions in dk is impossible in the case of single output
systems. The order of the parity space was chosen to be It is easily seen that the measurement is only corrupted
s = 2. In this example it was assumed that the unknown by the faults and not by unknown inputs . The unknown
input is constant with unknown amplitude. The unknown inputs acting onto the system equation enter linearily into
input signals are therefore completely correlated such that the dynamics equation. The operator D denotes the dif-
the weighting matrix Cd for an arbitrary pole placement in ferentiation w.r.t. time.
the observer is described by
3.1 Residual Generation
Cd = (~111
1 ~). (47) The fault detection observer is described by

Dz(t) = Foz(t) + F(y (t), u(t)) + Gy(t) (57)


The weighting matrix corresponding to the fault event vec-
tor was chosen to be

64
For this example the matrix T , necessary for the ob-
and the observer output equation by
server design, can be chosen to be the unity matrix such
(58) that the observer is represented in matrix form via

under the assumption that the observer state z(t) results Dz(t) = F(y(t), u(t» - Hr(t) (67)
in the fault-free case from a linear transformation of the
system state and the observer output equation
z(t) = Tx (t) (59)
r(t) = y(t) - z(t) (68)
the resulting estimation error equation
which has to be implemented on-line using an appropriate
De(t ) D(z(t) - Tx(t » integration routine. The elements of the residual are sen-
Foz(t) + F(y(t), u(t» + Gy(i) sitive to disturbances and fault s in actuators, components
- TA1·(t) - TF(y(t) , u(t» as well as all sensors. This will be demonstrated by some
simulation results.
-TEd(t) - TKd(X(t»f(t) (60)

leads to the determination of the observer matrices 3.3 Simulation Results


The following figures show simulation results when the
TA-FT GC (61)
residual generation is applied to the complete nonlinear
TE o (62)
robot model. The manoeuvre performed by the robot is a
F(y(t), u(t)) TF(y(t) , u(t» (63) change of axis 1 from 90 to - 90 degrees, axis 2 from -45
to 45 degrees and axis 3 from 0 to 45 degrees. The model
and from the observer output equation
mismatch between the simulated system and the observer
(64) is an extra hand-load on the simulated system of 2 kg .
This mismatch was not considered in the observer design.
The problem is to find the matrix T that fulfills the above The figures always show the first three residuals, namely
equations. Th e difference to the linear case is reflected the differences between the measured and est im ated posi-
by eqn. (63) which solution is completely determined by tions, in the upper part of the figure. The lower section
T. This problem is known as unknown input observer and of the figures show the differences between the measured
solved for linear systems. A solution can be found in Frank and the estimated velocities. All figures show a devi ation
and Wunnenberg (1989). In addition T must be chosen in the beginning of the simulation caused by deviations in
such that the faults act onto the residual which then serves the initial conditions between observer and system . These
as a fault indicator. In the following chapter this design errors decrease with a time constant of 100 ms that was
pocedure is applied to the model of a robot. chosen via the observer poles in H .
The simulation result of a fault in the velocity sensor of
axis 1 is shown in fig. 2. The measured velocity drops to
3.2 Application Example o from its current value at t=l see. The simulation shows
If one considers a robot as a st iff system and measures a significant response of the residuals r I and r 4. The fault
the velocities and positions of each arm the procedure be- can be detected and isolated. Similar simulation results
comes simple because the dynamic system state is com- were obtained for the other two axes.
pletely measurable . Furtheron, faults in the sensors shall More difficulties occur if the position sensors fail be-
be detected. In this case the robot can be described by the cause these measurements do not have such a significant
model (see e.g. Otter and Turk (1988» impact onto the residual generation. Figure 3 shows a sim-
ulation result for a sticking of the position sensor of axis
2 at t= 1 sec. such that the current position sensor read-
Dx(t) =
ing remains at its current value while the physical system
is still moving. The simulation shows a deviation of the
residual even though no fault is present. This deviation is
caused by the model mismatch between observer and sys-
tem. The residuals still show a significant response and the
effect of the fault remains visible in the residuals.
F(x(t),u(t» (65)
y(t) (qT(t), DqT(tW + K mfm(t) (66)

where q represents the angular positions of the robot


arms, Dq (t) the angular velocities and D2q(t ) the accel- 4 CONCLUSIONS
erations. J( q) represents the moment of inertia matrix,
The presented results show that the analytical redundancy
x d ( q( t), Dq( t» the dynamic forces applied to the system
concept may become a powerfull tool to detect faults in
by t he nonlinear coupling term among different axes and
robots. The first procedure presented is a systematic ap-
x 9 ( q( t» the gravitational forces. The term u represents
proach for the design of linear observers especially for fault
the normed input torques. Kmfm(t ) represents the to be
detection purposes. It uses all design freedom available
detected faults. The chosen application system is the robot
such as decoupling, eigenvalue assignment as well as the ob-
Manutec R3. A dynamic system description for this robot,
server order. The numerical treatment is simple, straight-
neglecting the frictional forces, is available in Otter and
forward and solvable with commonly available software
Tiirk (1988) . The Manutec R3 is scetched in fig. 1. In
tools. It is a consequent continuation of the work presented
this specific example the coupling terms from axes 4, 5,
by Chow and Willsky (1984) and Lou et al. (1986).
and 6 are neglected and only the first 3 axes are consid-
ered dynamically. The model is, therefore, of order 6 with
6 available measurements.

65
--rill - - rl21 - -rl31
The use of the nonlinear observer shows some very chal-
lenging simulation results. The advantage of the chosen 1. 750

system description is the applicabi lity of the theory devel-


1.250
oped for linear systems to nonlinear systems. The unknown
input observer approach is applicable and a transfer ation of
•. 750
the optimal fault detection observer design is possible. An
application of the residual generation to a physical system
• . 250
is currently in progress and first results will be available
soon. ~ . 25O
\ ..... -
~ . 750
tt .. l.. c)

- -rl<l r [5) - - rI61

• . 300

0 . \50

- -- -- -- -

-0 . 150

-0.300

~.<50 1 tt •• (n e)

Figure 3: Fault in posi t ion sensor 2

REFERENCES
Birk, J. , Zeitz, M.(1988 ), Extended Luenberger Observer
for Non-Linear Multivariab le Systems, Intern. Jour -
nal of Control, Vol 47 , No. 6, 1823-1836
Chow , E .Y. , Willsky, A.S. (1 984), Analytical Redundancy
and the Design of Robust Failure Detection Systems,
Figure 1: Manutec R3 ,(Otter, Turk (1988))
I EEE Tranj . on Automatic Control, Vo!. AC-29,603-
614.
--rIO - - rl21 - - rI31 Frank, P.M.(1989), Evaluation of Analytical Redundancy
for Fault Diagnosis in Dynamic Systems, P roceed·
ingj of the I FAC- IMA CS·IFORS International Sym-
pOjium on Advanced Information Procejjing in A uto-
matic Control· A IPAC '89, Nancy, France 3-5 July
1989.
Frank , P.M., Wunnenberg, J.(1989) , Robust Fault Diagno-
sis using Unknown Input Observer Schemes in: Pat-
ton , R.J. , Clark, R.N. , Frank , P.M. , Fault Diagnojij
in Dynamic SYjtemj, Prentice Hall , Dodrecht, NL .
Lou , X.C. , Willsky, A.S. , Verghese, G.C.(1986 ), Optimally
Robust Redundancy Relations for Failure Detection
tt.lllt)
in Uncertain Systems, A utomatica, Vol 22 , 333-344.
~I., - - rl51 --rI61
Otter , M. , Turk , S.(1988) , The DFVLR Models 1 and 2 of
3. 500
the Manutec R3 Robot, DFVLR Injtitutfiir Dynamik
der FlugjYjteme , Oberpfaffenhofen , Germany.
2.500 Patton, R. J ., Clark, R.N. , Frank , P.M.(1989), Fault Diag·
nOji3 in Dynamic SY3tem3 , Prentice Hall, Dodrecht ,
1.500
NL.
W iinnenberg, J., Frank, P.M. (1988), Model-Based Resid-
ual Generation fo r Dynamic Systems with Unknown
• •500
Inputs, Proc. 12th IMACS World Congrejj an Sci-
entific Computation , Paris , Vo!. 2, 435-437.
~ . 500 Wiinnenberg, J .(1989), Fault Detection in Dynamic SY3 .
tt.l.. tl
temj, Dissertation , University of Duisburg, Duisburg,
F igure 2: Fault in velocity sensor 1 Germany.

66

You might also like