Академический Документы
Профессиональный Документы
Культура Документы
Abstract
Risk analysis of concrete dams and quantification of the failure probability are important tasks in dam safety assessment.
The conditional probability of demand and capacity is usually estimated by numerical simulation and Monte Carlo tech-
nique. However, the estimated failure probability (or the reliability index) is dam-dependent which makes its application
limited to some case studies. This article proposes an analytical failure model for generic gravity dam classes which is
optimized based on large number of nonlinear finite element analyses. A hybrid parametric–probabilistic–statistical
approach is used to estimate the failure probability as a function of dam size, material distributional models and external
hydrological hazard. The proposed model can be used for preliminary design and evaluation of two-dimensional gravity
dam models.
Keywords
Dams, fragility, risk, reliability, finite element, sigmoid function
2. Figure 2(d) for the models with (known) pre- be applied to a system with pre-existing initial (model
defined final failure probabilities; 1) and pre-defined final (model 2) failure probabilities).
3. Figure 2(f) for the models with non-uniformly Fitting this seven-parameter model to the empirical
increasing failure probabilities; data points obtained from numerical simulations is dif-
4. Figure 2(h) for the models with sudden increase in ficult but possible. However, finding the relationship
failure probability; between any two sigmoid curves and presenting a gen-
5. Figure 2(i) for the models with steady failure prob- eralized equation is practically impossible. Thus, the
ability at low intensities; following two- to four-parameter simplified models are
6. Figure 2(j) for the models with instant time-variant considered.
failure probability;
7. Figure 2(m) for the models with cumulative time-
variant failure probability; Simplified functions
8. Figure 2(q) for the models with delaying uniform
Logistic function is the simplified model of generalized
growth rate;
logistic function and is presented with three parameters
9. Figure 2(t) for the models with delaying non-
uniform growth rate. c1
SðxÞ ¼ ð2Þ
1 + e 2 ðxc3 Þ
c
Also note that the rest of sub-figures in Figure 2 can
be interpreted as combination of two (or more) of the where c1 is the amplitude, c2 is the steepness of the
nine above-mentioned base models (e.g. Figure 2(e) can curve, and c3 is the x-value of the sigmoid’s midpoint.
Figure 2. Sample parametric sigmoid plots based on generalized logistic function: (a) c1 = var.; c2 = 0.1; c3 = c4 = c5 = c6 = c7 = 1.0, (b)
c1 = var.; c2 = 10; c3 = c4 = c5 = c6 = c7 = 1.0, (c) c1 = var.; c7 = 10; c2 = c3 = c4 = c5 = c6 = 1.0, (d) c2 = var.; c1 = 0.0; c3 = c4 = c5 = c6 = c7 = 1.0,
(e) c2 = var.; c1 = 0.0; c7 = 10; c3 = c4 = c5 = c6 = 1.0, (f) c3 = var.; c1 = 0.0; c2 = c4 = c5 = c6 = c7 = 1.0, (g) c3 = var.; c1 = 0.0; c4 = 0.1;
c2 = c5 = c6 = c7 = 1.0, (h) c3 = var.; c1 = 0.0; c5 = 10; c2 = c4 = c6 = c7 = 1.0, (i) c3 = var.; c1 = 0.0; c7 = 10; c3 = c4 = c5 = c6 = c7 = 1.0, (j)
c4 = var.; c1 = 0.0; c2 = c3 = c5 = c6 = c7 = 1.0, (k) c4 = var.; c1 = 0.0; c3 = 0.1; c2 = c5 = c6 = c7 = 1.0, (l) c4 = var.; c1 = 0.0; c7 = 10;
c2 = c3 = c5 = c6 = 1.0, (m) c5 = var.; c1 = 0.0; c2 = c3 = c4 = c6 = c7 = 1.0, (n) c5 = var.; c1 = 0.0; c3 = 0.1; c2 = c4 = c6 = c7 = 1.0, (o) c5 = var.;
c1 = 0.0; c3 = 10; c2 = c4 = c6 = c7 = 1.0, (p) c5 = var.; c1 = 0.0; c7 = 10; c2 = c3 = c4 = c6 = 1.0, (q) c6 = var.; c1 = 0.0; c2 = c3 = c4 = c5 = c7 = 1.0,
(r) c6 = var.; c1 = 0.0; c2 = c3 = c4 = c5 = c7 = 1.0, (s) c7 = var.; c1 = 0.0; c2 = c3 = c4 = c5 = c6 = 1.0, and (t) c7 = var.; c1 = 0.0; c4 = 10;
c2 = c3 = c5 = c6 = 1.0.
4 Proc IMechE Part O: J Risk and Reliability 00(0)
local failure modes (e.g. sliding along the lift joints and
where c1 and c2 . 0 are location and scale parameters, smeared cracking of mass concrete) are neglected.1,25
respectively. Such a distribution has a mean and med- Note that the cross-correlation of all failure modes is
ian equal to c1 and variance of ðc22 p2 Þ=3. not the objective of this article.26
Gompertz function is also a special case of the gener- Since only the sliding is accounted for the predomi-
alized logistic function in which the right-hand asymp- nant failure mode, the tangential contact between the
tote of the function is approached much more concrete and rock is modeled by Coulomb friction
gradually than the left-hand asymptote (note that in law.27 This model couples the maximum allowable
the logistic function both asymptotes are approached shear stress at the dam–foundation interface and the
symmetrically)21 applied pressure perpendicular to the surfaces. The
c3 x
interface joint can carry shear stresses up to the ulti-
SðxÞ ¼ c1 ec2 e ð5Þ mate value before failure (which depends on friction
angle, cohesion, normal stresses, and effective base
where c1 is the asymptote, c2 ; c3 . 0 set the displace- length). Figure 3(a) shows the dam and interface joint
ment along the x-axis and the growth rate along y-axis, at different stages of sliding until failure. It is note-
respectively. worthy that only static analyses are performed in this
Hariri function was originally proposed to model article; however, the loads are applied in the incre-
capacity function of concrete dams yet can be used for mental mode to prevent (or minimize) the structural
estimation of failure probability.1 Originally, it includes instability.
six parameters; however, only four (except c5 and c6 ) Magnitude and distribution of the uplift pressure are
are enough to be used in regression analysis of failure computed based on the drainage location, pre-defined
probability base crack length, and elevation of the reservoir water
in upstream face (headwater). Figure 3(b) shows all the
1 ec1 x + c2 applied loads on the dam as well as the dimensions of
SðxÞ ¼ c6 + c5 ð6Þ
1 + ec 3 x + c 4 the dam. The model parameters and material properties
where ci (i = 1, 2, . 6) are the model constants are summarized in Table 1. Only a limited number of
obtained from nonlinear least-squares regression. parameters are assumed to be RVs (i.e. those highly
Morgan–Mercer–Flodin (MMF) function presents a affect the demand and capacity).
four-parameter model originally proposed for nutri- The step-by-step strategy to perform the combined
tional responses of higher organism22 parametric–probabilistic simulations can be summar-
ized as follows:
c1 c2
SðxÞ ¼ c1 ð7Þ
1 + ðc3 xÞc4 A total of 11 (50:5:100 divisions on B1 ) 3 21
(50:5:150 divisions on H1 ) = 231 different generic
where c1 and c2 are upper and lower asymptotes, gravity dam models are constructed which covers a
respectively; c3 is growth rate; and c4 is a parameter wide range of possible dam classes (Figure 3(b)).
that controls the inflection point and can be explicitly Five parameters are taken to be the most effective/
computed as sensitive RVs in calculation of the failure probabil-
1=c4 ity: Lcr , Hw , crc , frc , and effD (Table 1). The
c4 1 importance of these parameters is verified based on
xinflection ¼ ð8Þ
c4 + 1 first-order reliability method (FORM) and simplified
Hariri-Ardebili 5
Figure 3. Finite element analysis of gravity dam: (a) crack propagation at the base joint from sound condition (left) to ultimate
failure (right) and (b) loads and dimension on pilot dam (left), uplift pressure distribution (middle), and matrix of different dam
classes (right).
Table 1. Model and material parameters for probabilistic finite element simulations.
analytical limit state functions. See ‘‘Fundamentals distributional model (dam class-independent), for
of FORM’’ in Appendix 1 for more explanation. example, silt internal friction angle (Table 1)
These five parameters are varying in both para- Impact of those five major RVs is studied
metric and probabilistic manners. It means that the independent of each other. For example, when a1
possible range of these parameters is divided into in Table 1 takes six different values, the quantities
some finite sub-spaces and then, samples are drawn of a2–a5 are assumed to be at ‘‘reference value.’’
randomly from each of these specific sub-spaces. The same analogy is applied to other four major
All other parameters (beside those five most effec- RVs.
tive RVs) are either deterministic, for example, con- The ‘‘reference value’’ for those five major RVs are
crete mass density, or randomly selected from fixed a1 ¼ 0, a2 ¼ 0:9, a3 ¼ 0:5, a4 ¼ 30, and a5 ¼ 0:8.
6 Proc IMechE Part O: J Risk and Reliability 00(0)
Figure 4. Parametric computed failure probabilities for different dam classes based on finite element simulations; in each plot, the
provided range is the bottom-to-top order of the surfaces: (a) Lcr (0.0, 0.1, 0.2, 0.3, 0.4, 0.5), (b) Hw (0.6, 0.7, 0.8, 0.9, 1.0, 1.1), (c) crc
(0.9, 0.7, 0.5, 0.3, 0.1), (d) frc (50, 40, 30, 20, 10), and (e) effD (0.8, 0.6, 0.4, 0.2, 0.0).
Table 2. Comparison between the sigmoid functions in terms of GOF and the optimized coefficients (large Pf scenario).
MMF: Morgan–Mercer–Flodin; SSE: sum of squares due to error; RMSE: root-mean-square error.
Figure 5. Estimation of failure probability with simulation-based method: (a) single function regression, (b) residuals for large Pf
scenario, and (c) Residuals for medium Pf scenario.
at a time. The box plots present the median of the fail- (c), respectively. Logistic distribution leads to largest
ure probability and outliers in each case.28 residual error ð;8%Þ, while Gompertz function gener-
These surfaces can be directly used to predict the fail- ates only ;3% error.
ure probability in the similar cases. However, the use of Residual plots can be helpful to assess the quality of
these multiple 3D plots is difficult in practice especially the regression. The scatter plot of the residuals is the
when the interpolation is needed. Thus, an attempt is ideal condition showing that there is no dependence
made to provide appropriate analytical models. between residuals and a value of the dam height (inde-
pendent parameter). This can be seen for ‘‘Hariri func-
tion’’ in Figure 5(b). Other simplified functions show a
Comparing the sigmoid functions dependency between the residuals and the order of the
First, accuracy of regression only for one single sigmoid function, for example, a harmonic trend in Logistic dis-
curve is investigated. For this purpose, two sets of data tribution. This dependency might be eliminated by
points are considered as follows: increasing the order of function (from two-parameter
logistic distribution to four-parameter Hariri function).
Large Pf . This group presents data in which the sys- Furthermore, the histograms of the residuals corre-
tem reaches nearly full failure. spond to a symmetric bell-shaped distribution, which
Medium Pf . This group presents data in which the indicates that the variance is normally distributed (not
system nearly reaches half the way to be failed. shown here).
Table 2 compares all five sigmoid functions in terms
In each case, all the five simplified sigmoid growth of coefficient of determination ðR2 Þ, adjusted R2 , the
functions are examined through the regression on the sum of squares due to error (SSE), and root-mean-
data points obtained from finite element simulations. square error (RMSE).29 Also, shown in this table are
Figure 5(a) shows the fitted curves for large and the coefficients ðci Þ for each sigmoid function. Since
medium Pf data points. Weighted nonlinear least- residuals are practically negligible for ‘‘medium Pf sce-
squares optimization technique is used in each case.29 nario’’ (Figure 5(c)), Table 2 presents GOF results only
Relatively higher weights are assigned to the data for ‘‘large Pf scenario.’’ As seen, Gompertz three-
points at the beginning of the set to reduce the possibil- parameter function provides the best estimation fol-
ity of negative Pf . The corresponding residuals for large lowed by Hariri and MMF four-parameter functions.
and medium Pf scenarios are shown in Figure 5(b) and The worst case belongs to two-parameter Logistic
8 Proc IMechE Part O: J Risk and Reliability 00(0)
Figure 6. Trend seeking among the five simplified model coefficients obtained from full set optimization only for Lcr (Figure 4(a)):
(a) logistic function, (b) logistic distribution, (c) Gompertz function, (d) Hariri function, and (e) MMF function.
distribution. However, note that R2 is more than 0.99 Based on this figure, functions with less parameters
in all cases. (i.e. logistic function and logistic distribution) exhibit a
Thus, at the first glance, Gompertz function is the significant trend for multiple 3D failure surfaces. Hariri
target model which is appropriate for analytical repre- function has also a meaningful trend with high fluctua-
sentation of single failure curve. However, the failure tion. MMF function is practically turned into a three-
probabilities in Figure 4 contain multiple curves (which parameter model (since c2 is zero). Note that c2 in
forms a surface) and multiple surfaces. Therefore, it is MMF function corresponds to lower asymptote of the
important to find an appropriate trend among different sigmoid curve and thus its value is determined even
curves. Figure 6 presents the relationship between each before the optimization. However, the most important
of the coefficients only for base crack ratio, Lcr observation is that Gompertz function (so far the best
(Figure 4(a)). Since this 3D plot is discretized to multi- model for single 2D failure curve) does not show any
ple 2D curves, \ H1 ; Pf . , the coefficients are pre- significant trend. Thus, based on the overall results
sented as a function of dam width, \ B1 ; ci . . from Figures 5 and 6 and Table 2, it is decided to select
Hariri-Ardebili 9
Table 3. Detailed constants of equation (11) controlling the slope and intercept of the logistic function–based failure surfaces.
f
RV c1 c2 c3
Mi1 NI1 MI2 Ni2 bi2 Mi3 Ni3
Figure 7. Trend seeking among the model coefficients obtained from full set of optimization for five basic RVs using logistic
function: (a) Lcr, (b) Hw, (c) crc, (d) frc, and (e) effD.
the ‘‘logistic function’’ as the final model to be used for c3 presents the midpoint of the 2D failure curves:
analytical equation. c3 has a linearly increasing trend bounded
8 between 100 (maximum of B1 ) and 150 (maxi-
mum of H1 ).
Analytical failure surfaces The slope of the idealized lines is approximately
So far, the logistic function has been selected as the best
8 + 70%.
model (both for single 2D curve and the multiple 3D The idealized lines have varying intercept.
surfaces). Thus, a set of nonlinear least-squares optimi-
8 The idealized lines are uniformly distributed.
zations are performed on data points in Figure 4, and
8
the resulted coefficients are shown in Figure 7 for five Hence, final form of the proposed analytical model
base RVs. The major observations are as follows: for the failure probability of generic gravity dam classes
can be written as
c1 presents the amplitude of the 2D failure curves:
c1 has a linearly decreasing trend bounded ci1
8 between 0.0 and 1.0. Pif ¼ ð10Þ
ci2 ðH1 ci3 Þ
1+e
The slope of the idealized lines is approximately
8 21.25%. where the superscript i refers to one of the five base
The idealized lines have varying intercept. RVs and coefficients cij are defined as follows
8 The idealized lines are uniformly distributed.
8 8
c2 presents the steepness of the 2D failure curves: > i
¼ g + Ni
a1 3 B1 + Mi1 3 RV
> c1
>
< 1
c2 has a linearly increasing trend bounded
8 between 0.05 and 0.30. ci2 ¼ g + Ni 3 ðB1 25Þ + bi
Mi2 3 RV ð11Þ
>
>
> 2 2
The slope of the idealized lines is varying; how- : ci ¼ i g
a3 3 B1 + M3 3 RV + N3i
8 ever, they have identical intercept. 3
10 Proc IMechE Part O: J Risk and Reliability 00(0)
Figure 8. Estimation of failure probability with simulation-based methods: (a) intercept of c1, (b) slope of c2, and (c) intercept of c3.
where a1 ¼ 0:0125, a3 ¼ + 0:7, bi2 is RV-dependent unknown parameters is discussed. Then, they are con-
g is normalized RV and refers to one
slope (Table 3), RV trasted in terms of accuracy and pattern. Finally, a
of the following items (with specific range of applica- model was optimized based on the LHS-based simula-
tion). Note that normalized RVs are used to combine tion results obtained from nonlinear finite element anal-
all five base RVs in a single equation. Each RV is sim- yses. Five most sensitive RVs are studied, that is, crack
ply normalized by dividing the pre-defined ranges in length ratio, headwater elevation, drain efficiency, base
Table 1 (a1 to a5 ) to its maximum absolute value joint coherency, and friction angle.
(therefore, the upper bound is always unit) The proposed model provides a good accuracy over
8 the wide range of gravity dam classes and can be used
> Lcr Lcr at preliminary stages of design or analysis to estimate
>
> ¼ ; ½0:00; 1:00
>
> L max 0:5 the failure probability of the system. Due to complexity
>
> Hw Hw
>
> of the analytical model, the proposed equation is in
>
> ¼ ; ½0:55; 1:00
>
> H 1:1
< wmaxcrc crc capable of providing the failure probability for five
g¼
RV ¼ ; ½0:10; 1:00 ð12Þ base RVs one at a time. For the complex combinations
>
> c rcmax 0:9
>
> of the input parameters, the use of surrogate models
> frc ¼ frc ;
> ½0:20; 1:00
>
> frcmax 50 (e.g. support vector machine) is recommended (which
>
>
>
> eff eff do not provide an analytical form).
>
: D
¼
D
; ½0:00; 1:00
effDmax 0:8 Last but not least, the proposed simplified approach
can be used for preliminary design of different concrete
Other RVg dependent constants are summarized in gravity dam classes. It provides a reliable starting point
Table 3. Note that instead of using this table, the value for the managers and engineers. Such a model can be
of three parentheses in equation (11) can be directly eval- further combined with information from construction
uated from Figure 8. Dash line in this figure indicates the cost of new dams or repair cost of the existing damaged
cross-correlation between different models to ensure that dams for prioritizing of the national infra-structural
the identical Pf is obtained from different equations when projects. In addition, the project managers can estimate
the similar base quantities are used. the benefit-over-cost of different dam classes (account-
Equation (10) provides a good analytical model for ing for hydrological hazard) with this simplified model
failure probability of generic gravity dam classes. without referring the project to structural engineers for
Overall, the GOF is more than 95% for different RVs time-consuming simulations.
and relatively wide range of input distributional model
which is acceptable for engineering purposes. A simple Acknowledgements
MATLAB-based script is also developed which facili-
tates visualization and estimation of Pf for any desired The author would like to express his sincere apprecia-
dam class and input RVs. tion to his former advisor (and current mentor),
Professor Victor E Saouma at the University of
Colorado Boulder for his enthusiastic guidance and
Conclusion advice throughout this research. Last but not least, the
author would like to thank the reviewers for their help-
Reliability analysis of concrete dams is a crucial task in ful and constructive comments that greatly contributed
dam safety management. It is usually performed in the to improving the final version of the paper.
context of the probabilistic approaches accounting for
the epistemic and aleatory uncertainties.
This article proposed a sigmoid-type analytical fail- Declaration of Conflicting Interests
ure model for a wide range of concrete gravity dams. The author(s) declared no potential conflicts of interest
First, different types of sigmoid growth functions are with respect to the research, authorship, and/or publi-
identified, and the mathematical model as well as the cation of this article.
Hariri-Ardebili 11
iterative search to find the ‘‘design point.’’ The general optimization problem can be solved using the algo-
solution to find the safety index was proposed by rithm introduced first by Rackwitz and Flessler31 where
Hasofer and Lind.30 The first step to define the Hasofer– the iterative linearized solution continues around the
Lind reliability index, bHL , is to normalize the RVs, X, current point until it converges with desired accuracy.
by transferring into a standard normal vector, U Detailed formulation of this procedure is beyond the
scope of this article; however, the reliability index in
Xi mXi
Ui ¼ ð13Þ the optimal fashion can be formulated as32
s Xi
bHL ¼ g U ð15Þ
where mXi and sXi are the mean and standard deviation
of the RV Xi , respectively. Note that in this new trans- where g 2i (independently for each of the ith RV) is a
ferred space mUi ¼ 0 and sUi ¼ 1, X RM is a random metric to evaluate the local sensitivities.
vector of N basic variables X ¼ X1 ; X2 ; . . . ; XN . In the current example, the following RVs contrib-
Note that the limit state function GðXÞ in the origi- ute about 98% of the total dam response: Hw ¼ 56%,
nal space will take the form GðUÞ in the standard nor- crc ¼ 14%, Lcr ¼ 13%, frc ¼ 10%, effD ¼ 5%. Thus,
mal space. The Hasofer–Lind reliability index is defined these five RVs are selected as base ones.
as the shortest distance from the origin in the standard
normal space to the new transferred failure surface Detailed comparison of Pf
GðUÞ (this point is called design point, U )
Detailed failure probability of the dam classes as a
U ¼ argmin kuk; GðuÞ40 ð14Þ function of dam height, dam width, and one of the five
uRM
base RVs is shown (Figure 9). In each case, statistics of
For the general nonlinear case, equation (14) may the data points are presented using box plots including
be reformulated as an optimization problem with the upper and lower quantiles and the outliers.
Lagrangian multiplier. This nonlinear constrained
Figure 9. Box plots showing dependency of failure probability on dam class and varying parameters (left to right: crack length,
water height, coherency, friction angle, and drain efficiency): (a) function of dam width, (b) function of dam height, and (c) function of
random variable.