Você está na página 1de 13

Computers and Geotechnics 30 (2003) 255267

www.elsevier.com/locate/compgeo

Location of critical failure surface and some further studies


on slope stability analysis
Y.M. Cheng
Department of Civil and Structural Engineering, Hong Kong Polytechnic University, Hong Kong
Received 27 November 2001; received in revised form 6 January 2002; accepted 17 January 2003

Abstract
The search for critical failure surface of a general soil slope is dicult as the objective function of the factor of safety is nonconvex and multiple minima exist in general. For this problem, the author proposes to use simulated annealing method and transforms the constraints of the problem to the determination of the dynamic upper and lower bounds of the control variables. The
critical failure surface can then be located with high precision with reasonable computer time under the present proposal. The
author has demonstrated that the proposed method is eective and ecient in analysis. Furthermore, the author has also proposed
a double QR factorization method in the evaluation of factor of safety and has demonstrated that this new approach is much better
in convergence.
# 2003 Published by Elsevier Science Ltd.
Keywords: Critical failure surface; Simulated annealing; Circular; Non-circular; Control variables

1. Introduction
For a proper slope stability analysis, it is required to
determine the critical failure surface and the corresponding factor of safety. In most of the commercial
programs, only systematic pattern search for critical
circular failure surface is available to the engineers. The
centres of rotation are dened over a grid and a contour
of factor of safety will be drawn [1]. While this geometric approach is acceptable for circular failure mode
where there are only three control variables, this
approach cannot be applied to non-circular failure
mode. Circular failure mode is a special case of noncircular failure mode and it is usually not the most critical case. Due to the limitations of the commonly used
slope stability analysis programs which cannot locate
the critical non-circular failure surface of a slope under
general conditions with general constraints, most of the
engineers are forced to perform the search on a trial and
error basis. Engineers have to dene several non-circular failure surfaces according to their experience and
determine the corresponding factors of safety. The

failure surface with the minimum factor of safety is then


chosen as the critical case. This approach is however
unsatisfactory as usually only limited number of failure
surfaces are tried in practice. Furthermore, the shape of
the failure surface that are considered by the engineers
are usually smooth concave. For highly irregular nonhomogeneous problems, it is dicult to locate the
critical solution purely by experience.
The location of the critical failure surface can be
viewed as a form of nonlinear non-smooth global optimization problem and the objective function to be
minimized is the factor of safety function. Some of the
diculties in the location of critical failure surface are:

E-mail address: ceymchen@polyu.edu.hk (Y.M. Cheng).


0266-352X/03/$ - see front matter # 2003 Published by Elsevier Science Ltd.
doi:10.1016/S0266-352X(03)00012-0

1. The objective function of the safety factor f is


non-smooth and can be non-convex in nature.
The constraints which include kinematically
acceptable shapes of failure surfaces, rock and
soil prole etc. may also be non-smooth, nonconvex functions.
2. Chen and Shao [8] have found that multiple
minima exist in a feasible solution domain. In
fact, the author has also tried the gradient type
optimization method and found that dierent
initial trial will give dierent local minima even

256

Y.M. Cheng / Computers and Geotechnics 30 (2003) 255267

for a homogeneous slope with simple geometry.


It appears that the existence of multiple local
minima is the fundamental feature of slope stability problem. Furthermore, the global minimum is not necessarily given by the condition of
f=0. Some of the control variables may lie
along the boundary of the feasible solution
domain so that f 6 0. This will be demonstrated
by the example in Fig. 3 in the following section.
3. A good initial trial for general ground conditions
with arbitrary loadings can be dicult to develop
for optimization analysis.
Baker and Garber [2] have applied calculus of variation to this problem and an Eulers dierential equation
is obtained. This approach is complicated even for a
relatively simple problem and is not practical to real
problems. The internal stress distribution from their
analysis may not be reasonable and later Baker reverted
to the dynamic programming method using Spencers
analysis. Revilla and Castillo [15] have applied similar
approach using Janbus simplied method while
Boutrup and Lovell [4] and later Chen [8] have adopted
a random generation approach. This approach is extremely inecient if a high accuracy is required and is seldom used for real problems. In fact, most engineers still
rely on their experience at present. Celestino and Duncan [5] have adopted the method of unidirectional
search while Nguyen has adopted the simplex approach
in this problem. Even though these methods may work
for relatively simple problems, there are many limitations associated with these methods which have been
addressed by Li and White [10]. The compound iteration scheme by Li and White [10] is an improvement of
the method by Celestino and Duncan and it appears to
be eective and fast in relatively simple cases. The most
critical limitation to all these methods is that there is no
guarantee that the critical failure surface as obtained is
the global minimum.

2. Use of simulated annealing method and generation


of failure surfaces
In view of the limitations of the classical methods as
described above, the author has adopted the simulated
annealing method in location of the global minimum of
the safety factor objective function. Simulated annealing
method is a combinatorial global optimization technique based on a random evaluation of an objective
function in such a way that transitions out of a local
maximum/minimum are possible and it is a useful technique for N-P types optimization problems [3]. The
concept of this method is similar to heating a solid to a
high temperature and then slowly cooling the molten
material in a controlled manner until it crystallizes

which is the minimum energy level. The solution starts


with a high temperature and a sequence of trial vectors
are generated until equilibrium (minimum energy of the
system) is reached. The step vector is adjusted periodically to follow the behaviour of the objective function.
Once thermal equilibrium is reached at a particular
temperature, the temperature is reduced and a new
sequence of moves will start. This process is continued
until a suciently low temperature is reached at which
no further improvement in the objective function can be
achieved. The readers should refer to Refs. [3,13,14] for
details of this method as this method is relatively new
and is seldom used in civil engineering discipline. Simulated annealing method can perform well even for nonconvex function with the presence of many local minima
and is very suitable for the present study [2,13,14,16].
Another advantage of the simulated annealing technique is that a good initial trial which is dicult to be
established for a general problem is not required. Other
possible techniques include the genetic algorithm and
tuba search [3] have not been adopted in the present
study but the approaches as proposed by the author are
actually suitable to be used for these techniques as well.
The diculties in the adoption of classical simulated
annealing, genetic algorithm or similar techniques in the
location of critical non-circular failure surface are: (1)
constraints from slope geometry; (2) failure surface
should be kinematically acceptable. The use of static
bounds to the control variables are easy to implement
but has a low eciency in the analysis. The author proposes to transform the various constraints and the
requirement of kinematically acceptable mechanism to
the evaluation of upper and lower bounds of the control
variables. Consider a typical failure surface ACDEFB
as shown in Fig. 1. The x-ordinates of the two exit ends
A and B are taken as the control variables of the objective function and the upper and lower bounds of these
two variables are specied by the user. Once the two exit
ends A and B of the failure surface are dened, the
requirements on kinematically acceptable mechanism
can be implemented as:
1. The x-ordinates of the interior points C, D, E
and F of the failure surface can be obtained by
uniform division of the horizontal distance
between A and B which is Xright-Xleft. The
x-ordinates of C, D, E and F are hence not control
variables. Alternatively, the division can be made
to follow the slope prole and the x-ordinates of
the interior points are also not control variables.
2. Joins points A and B and determine C1 which is a
point located vertically above C. The y-ordinate
of C1 is the lower value of either: (1) y-ordinate of
the ground prole as determine by x-ordinate of C,
(2) y-ordinate of the point lying along the line joining points A and B and determined by x-ordinate of

Y.M. Cheng / Computers and Geotechnics 30 (2003) 255267

C. C1 is the upper bound to the y-ordinate of


the rst inter-slice. The lower bound of the yordinate of C (third control variable) is set by
the author as C1-AB/4. In fact, such a lower
bound can allow for deep seated failure surface
and is more than enough for all the cases that
the author has encountered. Chen and Morgenstern [6] has pointed out that the left exit end
of the failure surface will make an angle of
45 /2 with the ground prole. The lower
bound of this third control variable will be
further checked with this criterion which is not
controlling in most cases. The lower bound of
the y-ordinate of C can be set to C1-AB/5 in
actual case without aecting the solution. The yordinate of point C is a control variable of the
objective function and it is conned within the
upper and lower bound as determined in step 2.
3. Once a y-ordinate of C is chosen in the simulated
annealing analysis, joins A and C and extrapolate
the line to G which is dened by the x-ordinate of
point D. The lower bound of the y-ordinate of
point D will be point G in order to maintain a
concave failure shape. The upper bound of D
which is D1 is determined in the same way as for
point C1. If part the ground prole lies below the
line joining B and C and aect the determination
of D1 (for example point J in Fig. 1), joins C and
J instead of B and C and determine the upper
bound as D2 instead of D1.
4. Perform step 3 for the remaining points until all
the upper and lower bounds of the control variables are dened.

257

5. To allow for non-concave failure surface which is


unlikely to occur in reality, the author allows an
option where the lower bound of point E will be
set to the lower of the value as determined in step
3 above or the y-ordinate of point D. The yordinate of point E cannot be lower than that of
D or else there will be a kink at the failure surface
which prevent failure to occur.
For Fig. 1, the control variables are the x-ordinates of
A, B and y-ordinates of point C, D, E and F. A control
variable vector X is used to store these control variables
and the order of the control variables must be (XA, XB,
YC, YD, YE, YF). For the location of the global minimum of the objective function, the user needs to dene
only the upper bound and lower bounds of the rst two
control variables. An initial trial will be determined in a
way similar to the approaches as shown above. The
upper and lower bounds of the other control variables
will then be calculated according to steps 2 and 3 above.
If the number of slice is No_Slice, then the number of
control variables will be No_Slice+1. If rock is present
in the problem, the lower bound determination as
shown above has to be modied slightly. In steps 2 and
3 above, the lower bound will either be the y-ordinate of
point G or the y-ordinate of the rock prole as determined
by the x-ordinate of D.
For a circular failure surface, there are only three
control variables which are the x and y coordinates of
the centre of rotation and the radius of the failure surface. The author however adopts the x-ordinates of the
two exit ends and the radius of the failure surface as the
three control variables in analysis as it is easier to

Fig. 1. Generation of a non-circular failure surface.

258

Y.M. Cheng / Computers and Geotechnics 30 (2003) 255267

dened the upper and lower bounds for the two exit
ends (see Fig. 2). The control variable vector X will be
(XA, XB, r). For the lower and upper bounds of the
radius, the lower bound is set to half of the length of
line AB which is the minimum possible radius. For the
upper bound of the radius, it is set to 50 times AB
(default is 50 but other values can be chosen) which is
sucient for the generation of shallow slip surface.
Kinematically unacceptable failure surface should not
be generated in the analysis and the constraints will
modify the lower and upper bounds of the radius when
the two exit ends are dened. The constraints include:
1. Failure surface cannot cut the ground prole at
more than two points within the two exit ends.
As seen in Fig. 2, point C may control the upper
bound of the radius instead of the default value
50 times AB.
2. Failure surface cannot cut into rock stratum
which may modify the lower bound of the radius.
3. The y-ordinate of the centre of rotation is higher
than the y-ordinate of the right exit end. For this
case, the last slice cannot be dened. This constraint may also modify the lower bound of the
radius.
In the simulated annealing analysis of this optimization problem, the variables are varied sequentially. The
rst two variables are varied within the user dened
lower and upper bounds which are static in nature.
Once these two variables are dened, the bounds for the
remaining variables are computed sequentially according to the guidelines as shown above for circular and
non-circular failure surfaces. This is dierent from classical simulated annealing method where the upper and
lower bounds for the control variables remain unchanged during the analysis. The special features of the

authors proposal is that the upper and lower bounds of


the control variables (except the rst two variables) will
keep on changing sequentially during each trial to
achieve kinematically acceptable failure surface. The
generation of trial failure surface and the search direction will then proceed in accordance with the normal
simulated annealing procedure and global minimum can
be located easily with a very high accuracy under the
present proposal.
Under the present simulated annealing analysis, there
is no dierence between optimization analysis for circular or non-circular failure surfaces. The dierences
between the two failure modes are the criteria on kinematically acceptable mechanism during the failure surface generation and the number of control variables.
Theoretically, simulated annealing analysis will be able
to give the global minimum if the number of evaluation
tends to innity. Practically, a termination criterion can
be set to end the search in a particular direction if the
reduction in the factor of safety is below a user specied
limit. Through this termination criterion, the search for
global minimum can be terminated at an acceptable
number of evaluations.
The location of critical failure surface will now
depends on the lower and upper bounds of the left and
right exit ends which can be decided easily. For experienced engineers, limited ranges for these two bounds
can be dened easily. For inexperienced engineers,
wider ranges can be dened for the lower and upper
bounds. The number of trials required for analysis will
only increase gently with the increase in the left and
right ranges which is another major advantage of the
approach by the author. For example, the author nds
that when the ranges for the left and right exit ends are
increased by two times, the number of trials required
will remain unchanged in most cases and may increase
only by less than 10% in some rare cases. The precision

Fig. 2. Kinematically acceptable circular failure surface.

Y.M. Cheng / Computers and Geotechnics 30 (2003) 255267

used by the author for the location of global minimum


is 0.0001 which is suciently good for most purposes.
For circular failure mode, the number of trials required
to achieve this precision is around 40006000 using the
simulated annealing technique. If the precision is lower
by 10 times to 0.00001, the number of trials required
will rise up to between 6000 and 7000. The actual
improvement in the solution is however very small and
the precision of 0.0001 is adopted by the author in the
later studies. For non-circular failure surface with 10
slices, the number of trials required by the present
method is usually around 14 00017 000. The computer
time required for 2D optimization analysis is usually
within 3 min for circular failure mode and 5 min for
non-circular mode with a PII 300 PC and is considered
to be acceptable. The author has varied the precision in
locating the global minimum and found that the analysis is not sensitive to this precision if it is not greater
than 0.001. The global minima for a precision of 0.0001
and 0.001 dier only in their third or fourth decimal
places and such a dierence is about the order as the
precision. The number of trials can hence be reduced by
around 50% if the ground condition is not complicated
or the accuracy required is not high (in the order of
102). In general, with the use of the modern computer,
the author would recommend a precision of 0.0001
which is acceptable in terms of computer time.
The present approach is similar to the use of Monte
Carlo technique by Malkawi, Hassan and Sarmas [18]
in some respects which requires random generation of

259

slip surfaces and renement of the search path. The


advantages of the present approach as compared with
Malkawis approach are:
1. Kinematically unacceptable shapes will not be
generated.
2. Each control variable is bounded by a dynamic
upper and lower range so that the solution
domain and the number of trials required will be
reduced.
3. No mechanism is available in Malkawais
approach to escape from local minimum in the
analysis as Malkawais approach relies on a good
estimate of the failure surface from Monte Carlo
analysis before the renement analysis. Simulated annealing technique can overcome this
limitation as escape from any local minimum is
crucial in the analysis. This limitation may be
important in cases where there is a soft soil band
and this will be illustrated in the next section.

3. Demonstration of search for global minimum


To illustrate the eectiveness of the authors proposal,
eight test problems (Nc1.datNc8.dat) are conducted.
For testing, a demo version of Slope 2000 together with
the user guide and the test data can be found from the
web page http://www.cse.polyu.edu.hk/  ceymcheng/.
The details of these problems are:

Fig. 3. Example 4 in ACAD study.

260

Y.M. Cheng / Computers and Geotechnics 30 (2003) 255267

1. Nc1.dat: 2 soil and a water table


2. Nc2.dat: 2 soil, a water table and a horizontal
pressure
3. Nc3.dat: 2 soil, a water table and a vertical
pressure
4. Nc4.dat: 2 soil, a water table and horizontal and
vertical pressure
5. Nc5.dat: 3 soil, a water table and a perch water
table
6. Nc6.dat: 3 soil, a water table and a perch water
table, horizontal pressure
7. Nc7.dat: 3 soil, a water table and a perch water
table, vertical pressure
8. Nc8.dat: 3 soil, a water table and a perch water
table, vertical and horizontal pressure
The eight test problems above can be grouped into
two main cases. Nc1.dat belongs to case 1 and is shown
in Fig. A1. The soil properties of the two types of soil
are unit weight=18.5 and 19 kN/m3, cohesive
strength=5 kPa and 2 kPa while =35 and 32 . Problems 2, 3 and 4 are similar to problem 1 except that
external loads are present in the problem. Nc5.dat
belongs to case 2 and is shown in Fig. A2. The soil
properties for the three types of soil are unit weight=19
kN/m3 for all the soil, cohesive strength=5, 2 and 5 kPa
while =32 , 30 and 35 . Problems 6, 7 and 8 are
similar to problem 2 except that external loads are
present in the problem.
A foolish but robust way to demonstrate the capability of the present proposal in locating critical failure
surface is to use pattern search approach. To limit the
amount of computer time used, the number of slices is
limited to ve in these studies and the slices are divided
evenly. The x-ordinates of the left and right of the failure surfaces are xed so that only the y-ordinates of the
four inter-slices are variables. There are hence actually
four control variables in the present study. Slope 2000 is
used to locate the critical failure surface. Based on the
critical result obtained, a pattern search for 0.5 m above
and below the critical failure surface as obtained by
simulated annealing analysis is tried. The y-ordinates
will vary with an increment of 0.01m so that there will
be 101101101101 or 10 406 041 combinations for
each test problem under the pattern search (more than
10 million combinations). Slope 2000 is specially modied to perform this pattern search. Using several 733
MHz PIII computers and several days in computer
analysis, all these combinations are tried and the yordinates of the critical failure surfaces are grouped in
Tables 14. The critical results all turn out to be kinematically acceptable and all kinematically unacceptable
mechanisms are associated with high factors of safety or
failure to converge. The method of analysis used in the
present comparison is Janbus simplied method as this
method is fast in convergence [1].

The x- and y-ordinates of the eight test problems are


shown in Tables 14. The results as shown in Tables 14
are clear and denite proof that the present technique is
able to locate the critical non-circular failure surfaces
for relatively complicated cases with the presence of
external loads. The minimum factors of safety from
more than 10 million combinations and several days
computer time are actually the same as that using Slope
2000 with less than 20 000 trials and 5 min analysis. The
slight dierences in the location of the critical failure
surface can be neglected in actual practice but such
dierences can be explained in the following way:
1. The default precision for locating the critical
failure surface is 0.0001 in Slope 2000. That
means, in the various search paths of the solution
domain vector spaces, the search along a particular path will stop and a new path will begin if
the reduction of the factor of safety is less than
0.0001. The results from the exhaustive search
and the simulated annealing search give solutions
which are equal up to the four decimal places
which is the default tolerance in the simulated
annealing search.
2. In the pattern search, the increment is 10 mm
each which is ne enough for all practical purposes. The critical failure surface may not lie
exactly on the grid points and hence theoretically
the solution from pattern search can still be
further improved.
3. The change of the factor of safety is usually very
small for a minor change in the prole of the
failure surface if the location of the failure surface is close to the critical failure surface (critical
band). In fact, this behaviour is well known and
is conrmed once again in the present study.
In the famous ACAD study [19] example 4, there is a
400 mm thick soft band of soil which has been studied
by Baker, Giam and Donald and others and a referee
value of 0.78 is recommended as the minimum factor of
safety. The unit weight of the soil is 18.4 kN/m3 for the
four layers of soil. The cohesive strengths are 28.5, 28.5,
0 and 28.5 kPa while  are 20, 20, 10 and 20 . The
author has studied this problem and has obtained a
lower factor of safety 0.688 which is shown in Fig. 3.
The diculty of this problem is the presence of a soft
band of soil and the thickness if this soft band is relatively small so that it may be dicult to escape from any
local minimum during the optimization search. Furthermore, the right exit end lies exactly on the right
hand side of the right exit domain and f60. These are
the diculties in location of critical failure surface for
this problem.
The author has also studied all the problems as given
by Malkawi and has obtained results similar to that by

261

Y.M. Cheng / Computers and Geotechnics 30 (2003) 255267

Malkawi. Since the number of trials required for each


problem is not given by Malkawi, the author cannot
compare the eciency of the two computational methods. In view of the fact that the bounds to each control
variable is static by Malkawi, the author suspects that
the number of trials required by Malkawis approach is
more than the present approach if the same precision is
adopted. Example 3 by Malkawai provides another
good illustration of the diculty in locating the global
minimum. The minimum factor of safety obtained by
the author is 0.394 with a failure surface specied by
ABCDEF in Fig. 4 which is very close to that by
Fletcher, Sridevi and Greco [18]. Malkawi obtained a
factor of safety of 0.401 (only slightly larger than that
by the author or Fletcher) with the critical failure surface specied by ABCDGH as shown in Fig. 4. It
appears that Malkawis approach fails to escape from
this local minimum in the analysis as the dierence

between Malkawais result and the authors result is


very small. The author suspects that the Monte Carlo
analysis by Malkawai may generate a good trial similar
to failure surface ABCDGH while another trial with
a slightly higher factor of safety is generated similar
to ABCDEF. Under Malkawais approach, the one
close to ABCDGH with a slightly lower factor of
safety will be chosen for further renement while for
simulated annealing analysis, escape from ABCDGH is
considered according to the requirement of cooling
process.
Finally, the author would like to use a very complicated problem as shown in Fig. 5 for illustration of the
technique as proposed. As shown in Fig. 5, there is a
very thin zigzag soft band with a thickness of 1 mm. It is
obvious that the majority of the critical failure surface
should lie along this zigzag soft band. It is however dicult to capture this critical failure surface automatically

Table 1
Comparison between optimization search and pattern search for
nc1.dat and nc2.dat

Table 3
Comparison between optimization search and pattern search for
nc5.dat and nc6.dat

4
6
8
10
12
14
FOS

NC1.data

NC2.datb

Optimization

Pattern

Optimization

Pattern

13.5
13.046
13.232
14.153
15.727
22.0
0.7279

13.5
13.05
13.24
14.16
15.74
22.0
0.7279

13.5
12.785
12.8
13.63
15.219
22.0
0.8872

13.5
12.78
12.79
13.62
15.22
22.0
0.8872

a
NC1.dat: No. of trial=10 081, critical solution is found at trial
9757.
b
NC2.dat: No. of trial=10 585, critical solution is found at trial
9428.

Table 2
Comparison between optimization search and pattern search for
nc3.dat and nc4.dat
X

4
6
8
10
12
14
FOS

NC3.data

NC5.data

NC4.datb

5
6.8
8.6
10.4
12.2
14.0
FOS
a

Pattern

Optimization

Pattern

13.5
12.928
13.227
14.267
15.996
22.0
0.7684

13.5
12.93
13.23
14.27
16.0
22.0
0.7685

13.5
12.677
12.831
13.784
15.539
22.0
0.9243

13.5
12.67
12.82
13.78
15.54
22.0
0.9243

a
NC3.dat: No. of trial=9577, critical solution is found at trial
9279.
b
NC4.dat: No. of trial=10 585, critical solution is found at trial 10
296.

Optimization

Pattern

Optimization

Pattern

0
0
0.92
2.813
4.707
6.6
0.7727

0
0
0.89
2.81
4.72
6.6
0.7726

0
0.964
0.026
1.566
3.653
6.6
1.1072

0
0.96
0.03
1.57
3.66
6.6
1.1072

NC5.dat: No. of trial=12 097, critical solution is found at trial 12

044.
b

NC6.dat: No. of trial=13 105, critical solution is found at trial 12

605.

Table 4
Comparisons between optimization search and pattern search for
nc7.dat and nc8.dat
X

Optimization

NC6.datb

5
6.8
8.6
10.4
12.2
14.0
FOS

NC7.data

NC8.datb

Optimization

Pattern

Optimization

Pattern

0
0
0.852
2.768
4.684
6.6
0.7494

0
0
0.85
2.8
4.78
6.6
0.7492

0
0.869
0.081
1.563
3.735
6.6
1.0327

0
0.87
0.08
1.56
3.73
6.6
1.0327

a
NC7.dat: No. of trial=11 593, critical solution is found at trial
11.545.
b
NC8.dat: No. of trial=12 601, critical solution is found at trial 12
090.

262

Y.M. Cheng / Computers and Geotechnics 30 (2003) 255267

by classical methods as the soft band is similar to a


Dirac function within the factor of safety objective
function f. Using the technique as suggested in this
paper, the critical failure surface is obtained without
diculty. To the authors knowledge, a soft band
with 1mm thickness has never been studied in the
past due to the dicult shape of Dirac function in
optimization.

4. Double QR method in evaluation of factor of safety


In the past, the factor of safety is commonly determined by iterative method using an initial trial factor of
safety. This approach is fast and robust in most cases.
The author has however come across many cases during
optimization analysis that dierent initial factor of
safety can give dierent converged result. This situation

Fig. 4. Example 4 by Malkawis study.

Fig. 5. Location of critical failure surface with 1 mm soft band.

Y.M. Cheng / Computers and Geotechnics 30 (2003) 255267

is particularly serious for irregular failure surfaces using


rigorous methods of analysis. This problem is experienced by many engineers in Hong Kong as the slopes in
Hong Kong are usually not regular and non-homogeneous in nature. For Morgenstern-Prices analysis, it
is also well known among many engineers in Hong
Kong that the choice of the point for the moment equilibrium consideration may aect the results of analysis.
Incorrect factor of safety can be obtained with an associated incorrect internal force distribution. Sarma [17]
has also pointed out the problem of multi-solution in
the evaluation of the factor of safety. To overcome this
problem, the author perform a check on the acceptability of the internal force distribution and the factor
of safety will be calculated again using a dierent initial
trial factor of safety. While this approach will work for
Janbus simplied method, the author has found that
there are many cases where the iterative method appears
to break down irrespectively of the initial factor of
safety for those rigorous methods. In this section, the
author proposes a new concept in the evaluation of the
factor of safety.
Refer to Fig. 6, the following assumptions are used in
the analysis:
1. for the rst slice (0,1) and the last slice (n,n+1)
and R for the rst slice (R0,1) and the last slice
(Rn,n+1) are set to 0 as the internal forces at the
two ends are 0.
2. X for the rst slice (XR0,1) and the last slice
(XRn,n+1) are also 0.

263

3. Based on force equilibrium, the following equations can be derived easily


K  W1 Q1 sin1 H1 cos1
R12

f1 W1 Q1 cos1 H1 sin1  U1 c1 L1
K  sin12 1  1 f1 cos12 1  1
1


R23

K W2 Q2 sin2 H2 cos2


f2 W2 Q2 cos2 H2 sin2  U2 c2 L2

K  sin23 2  2 f2 cos23 2  2
R12 K  sin12 1  2 f2 cos12 1  2
K  sin23 2  2 f2 cos23 2  2

2


K Wi Qi sini Hi cosi


Ri;i1

fi Wi Qi cosi Hi sini  Ui ci Li


K  sin i;i1 i  i fi cos i;i1 i  i


Ri1;i K  sin i1;i i1  i

fi cos i1;i i1  i



K sin i;i1 i  i fi cos i;i1 i  i
3
K  Wn Qn sinn Hn cosn

Rn1;n

fn Wn Qn cosn Hn sinn  Un cn Ln

K  sin n1;n n1  n

fn cos n1;n n1  n

Fig. 6. Internal forces in a failure mass.

264

Y.M. Cheng / Computers and Geotechnics 30 (2003) 255267

where k=factor of safety, Wi=weight of slice i,


Ui=pore pressure at base i, Qi=vertical load at slice i,
Hi =horizontal load at slice i, i =base angle at slice i,
fi=tan at base i, Ri,j=interslice internal force,
i=angle between interslice direction and horizontal,
set to 90 in present study, i,j=angle of inclination of
interslice force with horizontal and tan=lf(x), ci=c at
base i, Li=base length at slice i.
If we substitute Eq. (1) into (2), Eq. (2) into (3), and
so on until Rn,n+1, and using the fact that Rn,n+1=0, a
polynomial of order n for K will be established. The n
possible values for the factor of safety based on force
equilibrium can be found by the solution of the polynomial equation. Although the above formulation is
based on force equilibrium, moment equilibrium is not
neglected in the analysis. Any internal force system
which satises force equilibrium or both force and
moment equilibrium will have the same equations
above. For method which does not satisfy horizontal
force equilibrium (like Bishops method), the above
method cannot be used in the determination of the factor of safety and the classical iteration method must be
used.
For moment equilibrium: (take moment about the
midpoint of each slice base)

W1 Xw1 Q1 XQ1 H1 XH1  R12
XR12
0


W2 Xw2 Q2 XQ2 H2 XH2  R23
XR23




R12 XR12 L1 cos12 1  1

0

XRi;i1
Wi Xwi Qi XQi Hi XHi  Ri;i1




Ri1;i XRi1;i Li1 cos i1;i i1  i1

0
Wn Xwn Qn XQn Hn XHn




Rn1;n
XRn1;n Ln1
cos n1;n n1  n1
0

Slope stability methods, such as Janbus simplied


Method, Corps of Engineers Method, Lowe and Karaath Method, Janbus Rigorous Method, Spencer
Method and Sarmas Method [1] have various assumptions on in order to render the problem statically
determinate which are:
1. The angle  of each interface with the horizontal
direction. Except Sarma method for non-vertical
slice, all the other methods assume  to be equal
to 90 or vertical slices are considered.

2. The interslice force angle (), where tan is equal


to lf(x). For example, is assumed to be 0 in
Janbus simplied method and is assumed to be
constant in Spencers method.
4.1. Solution procedure
First, the factor of safety (K) is computed using Eqs.
(1)(4). The complete solution of all the possible factors
of safety can be obtained by double QR method, which
is a useful numerical method to calculate all the roots
associated with the Hessenberg matrix arising from Eqs.
(1)(4). It should be noted that solutions with imaginary
numbers may come out from the computations and the
use of double QR method instead of classical QR
method is necessary. Next, the interslice force (using
force equilibrium) and its location (using moment equilibrium) can be obtained from the above equations
directly without using iterative analysis. The factor of
safety will then be examined with respect to the following
conditions:
1. tan=l*f(x). However, tanj should be less than
the maximum angle of friction of the interface
and must be greater than 0.
2. R will resist slippage of the slope, hence it must
be positive.
3. The base and interface normal forces computed
must be greater than 0.
For those methods which do not satisfy moment
equilibrium, the requirements on moment equilibrium
need not be considered. If the above requirements are
satised, the computed K value is acceptable and the
analysis nish. If a K value from the QR analysis cannot
satisfy the above requirement, the next K value will be
computed. The process continues until all the possible K
values are examined. Beyond that, the analysis is
assumed to fail in convergence.
The advantage of the present method is that the factor of safety and internal forces with respect to force
equilibrium is obtained directly without any iteration
analysis. The authors procedures have also demonstrated that there can be at most n possible factor of
safety (including negative value and imaginary number)
from the double QR analysis for a failure mass with n
slices. The actual factor of safety can then be obtained
from the force and moment balance at a particular l
value. This is important in analysis as there are many
cases where the use of iterative method fails to give a
converged result or an incorrect converged result is
obtained. The time required for double QR computation is not excessively long as interslice normal and
shear forces are not required to be determined in
obtaining a factor of safety. In general, if the number of
slices used for analysis is less than 20, double QR

Y.M. Cheng / Computers and Geotechnics 30 (2003) 255267

method is only 50100% longer than iteration


method.
Consider the slope as shown in Fig. 7, if an initial trial
factor of 1.0 is used in the iteration analysis using Janbus simplied analysis, no converged result can be
found. However, if an initial factor of safety of 2.0 is
used, a converged result of 2.8708 (without the correction factor) is found. Using the double QR method
considering only the force equilibrium, the same result is
found without the need of an initial factor of safety.
Another example is a 45 slope specied by coordinates
(0,0), (5,0), (10,5) and (20,5). The slip surface is specied
by centre of rotation (1.989,11.616) and the radius of
rotation is 12.0. The soil parameters are unit weight=18
kN/m3, c=5 kPa and =30 . The method of analysis
adopted is MorgensternPrices method and f(x) is
taken as sin(x). If moment equilibrium is considered
about the centre of rotation, a factor of safety is
obtained as 0.747 using iterative analysis or double QR
method while Bishops analysis gives 0.742. If moment
equilibrium is considered about the centroid of the failure mass, iterative method fails to converge irrespective
of the initial factor of safety while double QR analysis is
not aected by this change. For the same slope, if the
centre of rotation is (0.09,15.49) while the radius of
rotation is 16.25 and f(x)=1.0, iterative method also
fails to converge if the centroid of the failure mass is
used in the analysis while double QR method can still
perform successfully. A further interesting case is a
slope denoted by (0,0), (5,0), (12,5), (20,14) and (30,27)
with a pore pressure ratio 0.3. The soil parameters are

265

unit weight=19 kN/m3, c=10 kPa and =36 . If the


left exit end is conned within x=25.3 while the right
exit end is conned within x=12.1420, iterative
method fails to converge for all the possible circular
failure surfaces within these two bounds when f(x)=1
or sin(x) and the centroid of the failure mass is used in
the moment equilibrium analysis. If the point for
moment equilibrium consideration is raised up by about
1 m, iterative method can converge for about 50% of
the cases in the optimization analysis. Those surfaces
which fail to converge in the iterative analysis can
usually get converged when double QR method is used.
It is interesting to nd that if the point for moment
equilibrium consideration is raised up by a great
amount, iterative method can get converged in most
cases. The author has tried many other cases and found
that in general up to 30% of the cases which failed to
converge with iterative method (using dierent initial
trial factor of safety) can actually get converged with
double QR method (excluding some special cases where
iterative method may totally break down). The main
reason is that in the rst step of iterative approach, the
interslice shear forces are assumed to be 0 and this may
leads to the wrong path in the solution of the nonlinear
factor of safety equation (even if the correct factor of
safety is used as the initial trial). Double QR method
can totally eliminate this problem by determining the
factor of safety directly.
The author has also found that sometimes a converged result with an unacceptable internal force distribution may be obtained from iteration analysis. The

Fig. 7. A deep seated failure surface to illustrate the use QR method.

266

Y.M. Cheng / Computers and Geotechnics 30 (2003) 255267

author nds that double QR method seldom gives rise


to this condition as compared with iterative method and
this situation can be avoided easily. In the double QR
analysis, all the n possible factors of safety together with
the internal forces can be evaluated. Those results which
are not kinematically acceptable can be eliminated
before the moment equilibrium consideration. In conclusion, the author nds that double QR method is a
very powerful technique in determining the factor of
safety of highly irregular slopes.

5. Discussion and conclusion


Based on the authors proposal on the treatment of
constraints, simulated annealing technique has been
used successfully for the location of critical circular and
non-circular failure surface under general conditions.
Actually there is no dierence in the optimization process between circular and non-circular modes under the
present formulation. The dierences between the two
failure modes are the number of control variables and
the criteria on kinematically acceptable mechanism (or

equivalently the dynamic bounds to the control variables). The present approach has been demonstrated to
be ecient and eective under relatively complicated
cases. The precision of the global minimum can actually
be specied by the user which is a feature not available
in other methods. The author has also demonstrated
that the present technique is also applicable even when a
very thin soft band is present. The solution time
required is acceptable and the present method is useful
for both research as well as real engineering problems.
The author has also proposed a new method for the
determination of the factor of safety which does not
require an initial trial factor of safety. This method
requires more computation as compared with the
classical iteration analysis but possesses the advantage
of reducing fail to converge situation. It is useful
for deep seated non-circular failure surface or similar
problems which are dicult to get convergence but may
be inecient for ordinary failure surfaces as the computation time is 1.52 times longer than the classical iterative method. Combining all the proposed techniques as
suggested in this paper, the analysis of slope stability
problem can be automated and evaluated easily.
Appendix

Appendix

Fig. A1. Critical failure surface for data NC1.dat.

Y.M. Cheng / Computers and Geotechnics 30 (2003) 255267

267

Fig. A2. Critical failure surface for data NC5.dat.

References
[1] Abramson LW, Lee TS, Sharma S, Boyce GM. Slope stability
and stabilization methods. 2nd ed. John Wiley; 2002.
[2] Baker R, Garber M. Theoretical analysis of the stability of
slopes. Geotechnique 1978;28:34195.
[3] Belegundu AD, Chandrupatla TR. Optimization concepts and
applications in engineering, Prentice Hall; 1999.
[4] Boutrup E, Lovell CW. Searching techniques in slope stability
analysis. Engineering Geology 1980;16:5161.
[5] Celestino TB, Duncan JM. Simplied search for non-circular slip
surface. In: Proceeding 10th International Conference on Soil
Mechanics and Foundation Engineering, Stockholm, Sweden,
1981. p. 3914.
[6] Chen Z, Morgenstern NR. Extension to the generalized method
of slices for stability analysis. Canadian Geotechnical Journal
1983;20(1):1049.
[8] Chen Z, Shao C. Evaluation of minimum factor of safety in slope
stability analysis. Canadian Geotechnical Journal 1983;25(4):
73548.

[10] Li KS, White W. Rapid evaluation of the critical slip surface in


slope stability analysis. Report no. 9. Australia: University College, Australian Defence Force Army, University of New South
Wales; 1986.
[13] Pham DT, Karaboga D. Intelligent optimisation technique.
Springer Verlag; 2000.
[14] Press WH, Teukolsky SA, Vetterling WT, Flannery BP., Numerical recipes in Fortran. 2nd ed. Cambridge University Press; 1992.
[15] Revilla J, Castillo E. The calculus of variations applied to stability of slopes. Geotechnique 1977;27:111.
[16] Sait MS, Youssef H. Iterative computer algorithms with applications in engineering. IEEE Computer Society; 1999.
[17] Sarma SK. A note on the stability of slopes. Geotechnique 1987;
37(1):10711.
[18] Husein Malkawi A. I., Hassen W.F. and Sarma S. K., A global
search method for locating general slip surface using Monte
Carlo techniques. Geotechnical and Geoenvironmental Journal,
ASCE 2002, 127(8) 688-98.
[19] ACADS. Soil slope stability programs reviews [publication no.
U255]. Melbourne: ACADS; 1989.

Você também pode gostar