Вы находитесь на странице: 1из 4

USEFULNESS OF FAULT TREE TO COMPLEX SYSTEM FOR RELIABILITY ANALYSIS

In the MFOSM and AFOSM methods just described, there is no theoretical limit to the number
of basic variables, x1, that can be considered. However, in a complex engineering system there
may exist several independent or dependent failure modes, each being a function of many basic
variables of different natures. Often the different failure modes are analysed by different teams
belonging to different technical disciplines. For such a system it is often helpful to utilize a fault
tree to sort out the manner and sequence in handling the individual basic variables. Thus, it
allows a division of the work to be carried out separately by teams of different specialities and
it provides a formal framework to combine the risk of different failure modes or components to
yield the total risk of failure.

A fault tree can be defined as a branched network representation of the relationship between
the defined failure situation and all the possible failure causes, including technical failure of the
system elements, extremal geophysical events, operational and management faults, and
aggressive human actions. An example fault tree for failure of a dam is shown in Figure 7. There
are two types of branch nodes in a fault tree. The OR node indicates that the occurrence of any
of the branches will result in a failure. This node corresponds to a union in probability theory.
An AND node indicates the failure will occur if all the branch events under the node occur
simultaneously, corresponding to an intersection in probability theory.

In application of fault tree and reliability analysis to a complex system, first all the possible failure
causes and mechanisms are identified and their relationships are established using the fault tree
as a tool. Next the probability of failure for each cause or combination of causes (as identified in
the fault tree) is evaluated using a risk evaluation method such as AFO-T or MFO-Z1 or Z3 as
appropriate. Finally, the total system risk is calculated according to the probabilistic relationship
identified from the fault tree.

Example applications of fault tree to first-order second-moment risk analysis can be found in
Cheng (1982), Ang and Tang (1984), and Vrijling (1982). Vrijling also gave an interesting
comparison between a deterministic design and an AFO-T design of a storm-surge barrier.

COMPARISON BETWEEN MEAN-VALUE AND ADVANCED FIRST-ORDER METHODS

The common advantages of MFOSM and AFOSM methods in comparison with other risk
evaluation methods have been discussed on p. 3. In this section a comparison between MFOSM
and AFOSM is presented. In general AFOSM is more complicated, requiring more computations
and more experiences or knowledge on the basic variables xi’s than MFOSM, but it also offers
potential to be more accurate. For the sake of clarity in discussion without losing generality,
numerical results of a simple system consisting of two basic random variables, x1 and x2, are
obtained and presented in Figures 8, 9, 10, and 11, respectively for four cases of different vales
of the coefficient of variation of X1and X2,. In all cases the expected value of X2 is kept constant
with 2. The expected value of X1 varies as necessary to establish the curves.

As described previously, there are two approaches for AFOSM: (1) solving Eq. 53 directly for xi*
without equivalent normal distribution transformation of the basic variables Xi, hereafter called
AFO-D, and (ii) solving Eq. 63 for xi* with equivalent normal transformation for the basic
variables, referred to as AFO-T. For either AFO-D or AFO-T, it has been found that the results are
independent of the definition of the performance variable Z given in Eqs. 3, 4, or 5. Conversely,
the results of MFOSM depends on how Z is defined. For the simple examples shown in Figures 7
to 10, X1 can be regarded as the resistance R and X2 as the load L. Accordingly, the ratio u1/u2
may be regarded as the central safety factor (Yen, 1979). All the three cases of definition of Z,
namely, X1-X2, (X1/X2)-1 and ln(X1/X2), are investigated, and they are referred to hereafter as
cases MFO-Z1, MFO-Z2 and MFO-Z3, respectively.

For AFO-D and for all three Z cases of MFOSM, in order to evaluate the risk no knowledge of the
probability distribution of the basic variables is required. Only the values of the mean and
coefficient of variation of the basic variables and the probability distribution of Z are needed. In
Figures 8 to 11 the results for two different distributions of Z, normal and uniform, are shown.
These two distributions of Z may be regarded as two extremes since in reality the Z distribution
of most cases probably falls in between. For AFO-T, some probabilistic knowledge or assumption
of each of the basic variables is required in addition to knowing the values of the mean and thus
not a prerequisite to assign a distribution for Z. This additional probabilistic information needed
can be the distributions of the basic variables, or merely the CDF and pdf values of the basic
variables around the failure point. In Figures 8 to 11 the results for three different basic variable
distributions are shown, namely, normal, lognormal or uniform distribution for both X1 and X2,
respectively.

In order to provide a measure of accuracy for the comparison of the MFOSM and AFOSM
methods, the exact solutions through direct integration are obtained and plotted in Figures 8
and 11 for the simple example system for the following cases of the distributions of the basic
variables

(a) Normal (N) distribution for both X1 and X2, and the risk computed from Eqs. 11 and 14,
identified as Exact Nxi in Figures 8 to 11.

(b) Lognormal (LN) distribution for both X1 and X2, and the risk computed from Eqs. 11 and 15,
identified as Exact LNxi in Figures 8 to 11.

(c) Lognormal distribution (LN-1) for both X1 and X2 and the risk computed from Eqs. 11 and the
direct integration first-order approximation, Eq. 16, identified as 1st-order LNxi in Figures 8 to
11.

(d) Uniform (U) distribution for both X1 and X2, for which the working equations are given in the
paper entitled "Sensitivity of Risk Evaluation to Coefficient of Variation" by Cheng et al. in this
volume, the solution is identified as Exact Uxi in Figures 8 to 11.

In this example since there are only two basic variables, the values of mean and coefficient of
variation used in Eqs. 14 to 16 are precise and hence the solutions are exact. Comparison of the
different cases of MFOSM and AFOSM using these exact solutions are summarized in Table 3.

For MFOSM, among the three definitions of the performance variable Z, MFO-Z2 consistently
performs less satisfactory than MFO-Z1 or -Z3, whereas the preference of Z1 or Z3 varies with
situations.

It can be seen from Figures 8 to 11 that the risk of a system is a function of the probability
distributions of the basic variables Xi, especially when the risk is low. Without any information
of the distributions of Xi and Z, usually the risk can be anywhere within the domain between the
exact curves of normal Xi and uniform Xi. Cases of uniform distribution of Z are shown merely
for illustration and rarely occur in reality. In practice, usually some information on the nature of
the distributions of the basic variables is available, though imperfect and imprecise. Such limited
information provides great help in yielding risk evaluation with acceptable accuracy. For
example, if the distribution of the basic variables are uni-modal and the mode is near the
beginning of the range, MFO-Z3 with normal distribution of Z or AFO-T assuming lognormal Xi
would provide reasonable estimates of the risk. If the distributions of Xi are uni-modal with small
skewness, MFO-Z1 and AFO-D assuming normal distribution of Z as well as AFO-T assuming
normal Xi would give reasonable estimates, AFO-D, inspite of its more computations than MFO-
Z1, yields a solution the same or close to that of MFO-Z1 unless for highly nonlinear performance
function with many basic variables. If the information on Xi is precise, AFO-T gives estimates
identical or close to the exact solutions. Therefore, when some helpful information is available
on Xi such that their equivalent normal transformation can reasonably be performed, the use of
AFO-t is preferred to AFO-D or MFOSM. On the other hand, if no information is available in the
basic variables except the values of their first two solutions xxxx, MFO-T or MFD-Z3, assuming
normal Z are preferred because of their simplicity in providing rough estimates of the risk for
comparison of different engineering alternatives.

The simple example cited above consists of only two independents basic random variables. But
the comparison based on it is valid for a performance function of many basic variables. The
superiority of AFO-T to MFOSM when some information on the distribution of the basic variables
is available has also been shown by Cheng (1982) for a system of more than two basic variables.
The MFOSM and AFO-T results of an example with four basic parameters given by Cheng is
summarized in Table 4 and plotted in Figure 12 and compared with the corresponding exact
solutions obtained from direct integration. The performance function is defined according to
Eqs. 3, 4 or 5 where the loading, L, is a linear combination of two independent random variables,

L=X1+X2

and the capacity or resistance, R, is the product of two independent random variables

R=X3*X4

The mean values of X1, X2, and X3 are 0.5, 1.5, and 1.0, respectively. The mean value of X4
varies from 1.5 to 4.75 in order to yield different risk values for different values of uR/uL to
establish the risk curves. The values of the coefficient of variation are 0.2, 0.4, 0.005, and 0.1
for X1, X2, X3, and X4, respectively. The distributions of X3 and X4 are assumed to be
lognormal so that R is also lognormal. The distributions of X1 and X2 are either normal or
uniform. Normal distributions of the basic variables, whereas MFO-Z1, Z2, and Z3 depend on
the definition of Z but independent of the distributions of Xi. In this example AFO-T and MFO-
Z1 give acceptable results whereas MFO-Z2 and MFO-Z3 involve significant error.
The risk, Pf, computed from Eq. 10 using E(Z) and oz given by Eqs. 41 and 44 of AFOSM at
the failure point provides an improvement over that of MFOSM corresponding to Eqs. 31 and
32 if the basic variables Xi of the performance function Z=G(Xi) exhibit nonlinear behaviour.
For linear performance functions where all the basic variables are normally distributed, the risk
evaluated using Eq. 11 with B given by either Eq. 46 or Eq. 64 is exact. For non-normal
variables, the risk evaluated using Eq. 11 with B given by Eq. 46 is only an approximation
while the accuracy of the first-order theory (Rackwitz and Fiessler, 1978). For convex
performance functions, where all the basic variables are normally distributed or equivalently
transformed normal, the risk is bound as
(68)
in which xm2(.) is a chi-square distribution with m degrees of freedom. Rackwitz (1975)
pointed out that if the distance function from system mean to the failure surface shows a distinct
minimum, then the lower bound will generally serve as a good risk estimation within the realm
of first-order accuracy. Furthermore, Ang and Tang (1984, p. 383) offer a comparison of AFO-
T (lower bound) and Monte Carlo estimations of risk for four nonlinear example, problems,
concluding that the AFO-T yields valid and accurate risk estimation for typical, practical
engineering problems. Moreover, Garen and Burges (1981) found that hydrograph error bounds
estimated by first-order uncertainly analysis compare favourably with those generated from the
Monte Carlo method. As an illustration of the upper bound, risk curves for the examples two-
basic-variable system are calculated and plotted in Figures 8 to 11 for AFO-D assuming chi-
square distribution of Z.
Despite the good results obtained using FOSM described above, the wide disparity between the
various curves in Figures 8 to 11 may lead the reader to question the utility of FOSM analysis.
Especially, for cases where information and experience regarding the basic variables and the
performance of the system is limited. However, Cornell (1972) pointed out a significant point
in any reliability analysis:
“It is important to engineering applications that we avoid the tendency to model only those
probabilistic aspects that we think we know how to analyse. It is far better to have an
approximate model of the whole problem that an exact model of only a portion of it.”

Without a doubt, the true risk of any engineering system design or the true reliability of any
model of a natural system will never be known exactly. Nevertheless, FOSM analysis provides a
simple, comparative, objective, consistent, and systematic procedure of evaluating
uncertainties and estimating relative risks for various design or decision alternatives. In
engineering practice often such relative reliability estimates for engineering systems or
alternatives are of more interest in decision making than the true value of the reliability.
Furthermore, the risk estimates may be quite close to the truth if adequate information on the
basic variables and system performance is available.

Thus, MFO-Z1 and MFO-Z3 may be regarded as rough risk evaluation tools when statistical
information consisting of only the values of the first two moments is available. AFO-D can be
regarded as a first-order alternative to direct integration with assumed or approximated
distribution for the performance variable Z. AFO-T can be considered as a “one-and-half” order
approximation to make possible an improved but transformed numerical substitution of the
direct integration. The source of errors for AFO-T is mainly on the fidelity of the equivalent
normal transformation of the basic variables Xi. For AFO-D and MFOSM there is no direct and
explicit means to account for the distributional information of the basic variables beyond the
first two statistical moments. The main sources of errors for AFO-D are from the assumption on
the distribution of the performance parameter Z and from the second or higher order error in
locating the failure point xi*. The errors involved in MFOSM comes from the assumption on Z
and the first-order approximation of nonlinear effect between the mean -xi and failure point xi*.

Вам также может понравиться