Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
To cite this article: P. Thoft-Christensen & J. D. S⊘rensen (1987) Optimal strategy for
inspection and repair of structural systems, Civil Engineering Systems, 4:2, 94-100, DOI:
10.1080/02630258708970464
Article views: 59
Download by: [IIT Indian Institute of Technology - Mumbai] Date: 23 October 2017, At: 07:37
Optimal strategy for inspection and
-
0263-0257/X7/02094-07/$03.00
94 Civ. Engng Syst. 1987. Vol. 4. J u n e 0 1987 Rurterworth 6r Co (Publishers) Lttl
Optimal maintenance strategy: P. Thoft-Christensen and J. D. SGrensen
dependent on 11. Further, 11 is an integer variable and it is must be based on a systems approach. For statically
therefore complicated to deal with in relation to the other determinate (non-redundant) structures failure in any
continuous variables. The influence of iz can be taken into member will result in failure of the total system (structure).
account by solving the optimization problem for a However, failure in a single elcment in a structural system
number of different values of 11 and comparing the optimal will not always result in failure of the total system, because
costs C . The n-value with the smallest C-value is then the the remainingclements may be able tosustain theexternal
optimal value. load by redistribution of the internal load effects (stati-
A n alternative formulation of the strategy presented cally indeterminate o r redundant structures). Further, a
here is the following: Detertnitze u strategj~for itlspectioil structural system will in general have a large number of
urld repuir of u given strircture so tlrot a yivet~ur?~ountof potential failure modes and the most important modes
r~loizcyis used optitpzalli~,i.e. the rnii7itnurn 17ultteof fl in the must be taken into account in an estimate of the reliability
liJi.tinze T i s a.shig11a.spossih1e.This optimization problem of a structure. Identification of the most important
can be solved in a similar way. (significant) failure modes can be performed by e.g. the
The strategy presented is purposefully simplified as fl-unzipping method.
much as possible without neglectihg the main effects. It is Let a structural systcm consist of a number m of
rather straightforward to include more details so that the single failure elements (failure modes) such as failure in
strategy becomes more applicable. bending, fatigue etc. A simple definition of systems failure
Downloaded by [IIT Indian Institute of Technology - Mumbai] at 07:37 23 October 2017
Updating of the reliability of single elements after would be failure in a single failure element, i.e. the
inspection has been treated by several authors, see e.g. structure is considered to be in a state of failure when a
Madsen and Tallin' and Yang and Chen2 for the special single failure element fai1s:Thc probability of failure Pf is
case of fatigue failure. Similar updating techniques can be then calculated as the probability of having failure in
incorporated in the strategy presented here. failure element 1 and/or in failure element 2 . . . and/or
in failure element m, and it can be shown that a good
cstimatc of P, is:
Structural reliability theory
As mcntioned above an essential assumption behind the
optimal strategy is the existence of a reliability measure where @, is the m-dimensional standardized normal
for the reliability of structural elements and/or the P
distribution function, = (/I,, . . . , /I,) and P is the
structural system itsclf. A natural choice for such a correlation matrix for the safety margins. This definition
reliability measure is the so-called reliability index /I. In is called a failure modelling at level 1.
this section the reliability index /? will be briefly intro- For redundant structures failure in a single failure
duced. More detailed presentations are given by Thoft- element will in general not be considered as failure of the
Christenscn and Baker3 and Thoft-Christensen and Mur- complete structure. For some elasto-plastic structures it
otsu4. The presentation here is based on a papcr by Thoft- may be morc relevant to define failure of the structure as
Christensen5. formation of a mechanism. For other structures it could
Estimation of the probability of failure of single be morc natural to define failure of the structure as failure
structural elements is now considered a rather trivial task in two failure elements (level 2 modelling). Independently
although there is still a need for data concerning the of the definition chosen it is important to have available
probability distributions of material propertics (e.g. yield a method by which the most significant failure elements,
stresses), load parameters (e.g. wind loads), and geo- pairs of failure elements o r mechanisms can be identified
metrical quantities (e.g. cross-sectional areas). because the total number of failure elements, pairs of
Let these so-called basic variables be X = (X ,, . . . , failure elements o r mechanisms are usually too high to
X,). By a suitable transformation X is transformed into a include all in the reliability analysis. By the /?-unzipping
set of independent standard normal variables 2 = ( Z , , method the significant failure modes can be identified.
. . . . Z,). Further, let the so-called failure function (limit In Example 2 in this paper only the systems reliability
state function) fdivide the z-space intoa failure region (f(Z) index at level 1 is used for illustration. In a real practical
< 0) and a safe region (f(Z)> 0). The reliablity index /I is application estimation at level 2 or perhaps at mechan-
then defined as the smallest distance from the origin to ism level is more relevant. However, this will increase the
the failure surface (f(Z)= 0) in the standard normal amount of computer work considerably.
z-system. It can then be shown that
Assumptions
whcrc 4, is the standard normal distribution function. M For the sake of simplicity a number of assumptions and
= f(Z) is called the safety margin.
simplifications are made in this paper. Several of these
The reliability index P can be cstirnated for any assumptions can be releascd if necessary. However, it is
fa~luremode of a structural element if the corresponding believed that the modelling used here incorporates the
failure function is known. It is much more complicated to essential features.
estimate the probability of failure for the complete The design of the structure is assumed to be given.
(redundant) structure. Howcver, in the last decade several The structure can be modelled by a finite number of
heuristic tcchniqucs have been devcloped. Two methods- failure elements m. The reliability of the failure
the /?-unzipping method and the branch-and-bound elements is assumed to be non-increasing functions
method - are presented in detail by Thoft-Christensen of time
and Murotsu4.
It has ofcourse for many years been recognized that a
fully satisfactory estimate of the reliability of a structure if no repair is performed.
If the structural system is modelled by a system of The reliablity of the elements and the system has to
failure elements then it is assumed that the systems fulfil the following inequalities
reliability index P, is a non-increasing function of time
if no repair is performed.
The damage of a failure element is assumed to be where pin and Kinare minimum acceptable reliabil-
measured by an increasing function of time ity indices for element i and the system, respectively.
and complete.
Not all critical damages are discovered by an inspec- repair courses (branches) is 2" (see Fig 3), where 0 and 1
tion. Let F be the event that D > Dinand let Q be the signify non-repair and repair.
event that F is discovered given F. Then it is assumed Let r j be the number of repairs in branch j a n d 3,the
that event that branch j occurs. Then the following optimiz-
ation problcm can now be formulated
P(Q) = P(F discovered 1 F) = p(q) (7)
where q is a measure of the quality of inspection,
0 < q ,< I and p is an S-shaped increasing function,
2"
min C = x" - co
+ jC C,rjP(Bj)
for example I,.....I n , I - q =,
41. . . . .4"
-
- branch 1
branch 2
- branch j
- branch 2"
0.2 0.4 0.6 0.8 1.0
P
Fig 2 Inspection quality Fig 3 Repair realizations for a single element
If the correlation between the failure elements before and Alternatively the optimization problem could be
after reparation is neglected then at T, we have formulated (see the introduction) as
I,
rnin
.....I,,
{ i=l.
max
....n + l
{-fi(~,))}
41. ....9"
2"
C= " - Co
+
s.t. 1
i - 1 1.1 - q i ,=,1 CRrjP(Bj) < Cm""
li<T
If no repair has been performed before the time '7;,, i= 1
we have p i n < ti< tmax, i = 1 , . . ., n
P(-RIA ... A- R , - , A + R,) q m i n < q i < q m P X , i = I, ..., 11
Downloaded by [IIT Indian Institute of Technology - Mumbai] at 07:37 23 October 2017
{ rnin
i:::::::,
-B(TI)
-scT,+ 1) 1
-P(-R,A . . . A - R , - , A + R,) (26) instead of (33) we have a so-called multiobjective optimi-
In general, if correlation between failure elements zation problem which can be solved in a subjective way
bcforc and after repairs is ncglectcd and repair is pcr- by choosing the solution from the Pareto optimum set".
formed at times T,,, . . . , T,lR(7;,, = To= 0) In this paper the formulation (13)-(17) is used. The
optimization problem is solved using the N L P Q L algo-
rithm implemented by Schittkowski7. The algorithm is
based on the method by Hans, Powell9 and Wilson1'.
Generally, it is a very effective method where each
iteration consists of two steps. The first step is determina-
tion of a search direction by solving a quadratic optimiza-
tion problem formed by a quadratic approximation of
where the Lagrangian function of the non-linear problem and
P i = P ( - R n , - l + l .~. .A - R , , - , A + R , , ) , ~ =1 ,..., R a linearization of the constraints at the current design
point. The second stcp is a linc search with a n augmented
(28) Lagrangian merit function.
P R + l= P ( - R n R f l ~. . . A - R n ) (29)
are calculated using (19)-(26) successively. Example 1
If p is the correlation coefficient between repaired Consider a single failure element which corresponds to
elements then an improved approximation can be calcu- fatigue failure of a structural element. Let the fatigue
lated by damage D be estimated by Miner's rule (see Wirshing")
Let Ino + lnc = 31.10, n l , = 3, E[lnX] = -0.0196, element the optimization variables are ti = K - T- i ,,
a[lnXJ = 0.198, E[lnB] = -0.468, o[lnB] = 0.472, = 1 : . .. , n and q,, i = 1, . . . , n. The total number of
E[ln K] = 36.06, ~ [ l n K ]= 0.572, and let T be measured in d i f f e ~ n trepair courses (branches) is 2"". The following
years. Further, the parameters in the optimization optimization problem can then be formulated (see
problcm are qmin= 0.2: qWX= 0.95, Co = 1, C, = 10, ( 1 3H 17))
/Pin = 3, T = I0 years, t m i n
= 0.25 years and tmax = 2 years.
P(Bj)is estimated using (27)where the reliability indices
are calculated from (40) with D, = Din= 0.005. The relia- min C = " --1 c o- 2""
i = ' 1.1 - q i
+ jC c~rjP(Bj) (41)
bility index in (32) is calculated from (40) with D, = I r 1 . .... I n =,
q,. ....4"
which corresponds to failure.
For n = 9 the optimal values of T, q and C arc C
= 21.7,T=(2.00,2.00,2.00,1.03,0.74,0.65,0.54,0.43,0.34)
and ?j =(0.2, 0.2, 0.26, 0.30, 0.31, 0.31, 0.31, 0.31, 0.31).
In Fig 4 the variation of the reliability index /3 with
time is shown. It is seen that the time intervals between
inspections are decreasing with time and that apart from
the first two inspections thc inspection quality should be
Downloaded by [IIT Indian Institute of Technology - Mumbai] at 07:37 23 October 2017
nearly constant.
In Fig 5 the optimum value of thc cost C is shown
as a function of the number of inspections. The minimum where flS(T)is thc systems reliability index at time T .
value of C occurs for n = 9.
fiS(T)can be calculated from
Example 2
In this example we considcr a scries system with rn = 3
clemcnts each modelled as in Example 1 except that
Inv + Inc = 30.80 is used. For thc sake of simplicity it is
assumed that the elements are independent.
Fig 4 Optimal inspection strategy for a single element For n = 8 the numbcr of different branches is Z3 " ' =
(Example 1 ) 1.7. lo7. Most of thcse branches d o not have significant
influcnce on thcobjective function (41)and the constraints
(42). In order to reduce the computational costs it is
therefore important for the applicability of this strategy
to be able to identify only thc most significant branches.
In this exarnplc the identification of these branches is
simply based on thc relative magnitudes of P(B,) at each
of the times T,, . . . , T,. Branches with probabilities less
than I % of the maximum P(R,) at each inspection time
are neglcctcd.
For 17 = 7 the solution of the optimizat~onproblem
(4 1)-(45)using N L P Q L becomes C = 17.4,T = (2.00,2.00,
2.00, 1.45, 0.72, 0.66, 0.61) and ?j = (0.22,0.29, 0.30, 0.31,
0.30, 0.29, 0.28). In Fig 6 thc vitriation of the systems
reliability index with time is shown. As in Example 1 the
time in~crvalsbetwccn inspections decrease with time.
The inspection qualities arc sccn to be of the same order
Fig 5 Optimal inspection and repair cost C as a function of nlagnitudc.
of the number of inspections for a single element In Fig 7 the optimum value of C is shown as a
(Example 1 ) function of the nuniber of inspections rl. As in the exitmpl~
2 Yang, J. N. and Chm, S Fatigue Reliability of Structural Compon- 7 Schittkow%ki, K. NLPQL: A FORTRAN subroutine solving
ents under Scheduled Inspection and Repair ~Maintenanoe.Proc. constrained non-linear programming problems, To appear in
IUTAlW .$?nip. Srockholm. Sweden 1984 (eds. Eggn-errr and Lind), Anna1.s of Operarions Research.
Springer- Verlag, Berlin. pp. 559-568 8 Han, S.-P. A globally convergent method for non-linear program-
3 Tboft-Christensen, P. and Baker, M. .I. Structural reliability theory ming J. 0primi:arion Theory and Applicarionr. 1977. 22. 277
and its applications, Springer- Verlug. 1982 9 Powell, M. J. D. A fast algorithm Tor non-linearly constrained
4 Thoft-Christensen, P. and &lurotsu, Y. Application of structural optimization calculationst In ~VumericalAnaly.~ir (ed. G. A. War-
systems reliability theory, Springer-Verlag. 1986 son), Lzcrure Nores in Marhemarics, 1978, 630, Springer- Verlag
5 Thoft-Christensen, P. R a n t advances in structural systems relia- 10 Wilson, R. B. A simplicia1 algorithm for concave programming,
bility theory, IABSE Symp. Safery and Qualiry Assur. Civil Engng Ph.D. Thesis. Harvard University. Boston. 1963
Srruct.. Tokjo. 1986, IABSE Report. 51, 101-108 1 1 Wirshing, P. H. Fatigue reliability Tor offshore structures, ASCE
6 Carmichael, D. G . Computation of Pareto optima in structural J. of Srrucrurul Engng, 1 lO(10). 1984. 2340-2356
design, Internur. J. ~Vunzericallblerhods in Engng, 1980.15.925-929
Downloaded by [IIT Indian Institute of Technology - Mumbai] at 07:37 23 October 2017