Professional Documents
Culture Documents
Abstract
This paper deals with the possibility of the probabilistic analysis of the reliability of the machine foundation
considering the soil-foundation-machine interaction. The requirements to design of the foundation under rotating
machines increased due to development of calculation method and computer tools. The advantages and
disadvantages of the deterministic and probabilistic analysis of the machine foundation resistance are discussed.
The sensitivity of the machine foundation to the uncertainties of the soil properties due to long-time rotating
movement of machine is not negligible for design engineers. The Response Surface Method (RSM) for the
analysis of the compressor foundation reliability was used on program ANSYS. The probabilistic analysis gives
us more complex information about the soil-foundation-machine interaction as the deterministic analysis.
Key words: soil-foundation-machine interaction, probability, sensitivity, ANSYS, FEM, RSM
Introduction
The vibration of the machines in the industrial enterprises impacts to the building structures
and technology as well as to the human comfort [2, 3, 5, 7, 9, 10 and 11]. The requirements to
design of the foundation under rotating machines increased in the last years. The Eurocodes
and national standard define much of these requirements [4, 7, 9 and 11]. The soil-foundation
and foundation-machine interaction are the principal problems what determine the effect of
the industry performance to the environs. The designers have to consider these effects in the
point of view of the safety, reliability and durability of the structures.
Behaviour of soil and structures depends on character and intensity of dynamic load. The
velocity of strain and stress intensity in the subsoil has important influences to the material
characteristics of the soil. Randomness in the loading and the environmental effects, the
variability of the material and geometric characteristics of structures and many other
-1-
"uncertainties" affecting errors in the computing model lead to a situation where the actual
behaviour of a structure is different from the modelled one [1, 6, 7 and 8].
During the structural design process, an engineer has to consider problems of the safety,
reliability and durability of machine foundations from the point of view of its planned life
cycle. Much attention should be paid to using the probabilistic approach in an analysis of the
reliability of structures [8]. Most problems concerning the reliability of building structures are
defined today as a comparison of two stochastic values, loading effects E and the resistance R,
depending on the variable material and geometric characteristics of the structural element.
The variability of those parameters is characterized by the corresponding functions of the
probability density fR(r) and fE(e). In the case of a deterministic approach to a design, the
deterministic (nominal) attributes of those parameters Rd and Ed are compared.
The deterministic definition of the reliability condition has the form
Rd Ed
(1)
and in the case of the probabilistic approach, it has the form
RF = R E 0
(2)
where RF is the reliability function, which can be expressed generally as a function of the
stochastic parameters X1, X2 to Xn, used in the calculation of R and E.
RF = g ( X 1 , X 2 ,..., X n )
(3)
The failure function g(X) represents the condition (reserve) of the reliability, which can either
be an explicit or implicit function of the stochastic parameters and can be single (defined on
one cross-section) or complex (defined on several cross-sections, e.g., on a complex finite
element model).
The most general form of the probabilistic reliability condition is given as follows:
(4)
p f = P ( R E < 0) P ( RF < 0) < pd
where pd is the so-called design (allowed or acceptable) value of the probability of
failure.
The reliability criteria are defined in the Eurocode in dependency on reliability index, what
is adequate to target level of failure probability (Table 1).
Table 1: Target reliability index and probability of failure by Eurocode 1990
Limit state
Target reliability index d
50 years
1 year
Ultimate
3,8 (pf 10-4)
4,7 (pf 10-6)
Fatigue
1,5-3,8*) (pf 10-110-4)
Serviceability
1,5 (pf 10-1)
3,0 (pf 10-3)
In the case of the stochastic approach, various forms of analyses (statistical analysis,
sensitivity analysis, probabilistic analysis) can be performed.
The reliability assessment criteria according to the reliability index were mentioned below.
Most well-known is the modified Monte Carlo method and Latin Hypercube Sampling (LHS)
simulations. The simulation methods on the base of Monte Carlo method are more
advantageous for the estimation of the failure probability. The probability of failure is
calculated as best estimation of the statistical parameters and theoretical model of the
probability distribution of the reliability function Z = g(X).
The probability of failure is defined as best estimation of the numerical simulations in the
form
-2-
1 N
(5)
I g ( X i ) 0
N i =1
where N is the simulation number, g(.) is the failure function, I[.] is the function with value
1, if the condition in the square bracket is fulfilled, otherwise is equal to 0.
The variation of this failure estimation can be described by Melcher in the form
2
1 N
1 1 N 2
2
s pf =
I g ( X i 0 ) I g ( X i 0 )
(6)
( N 1) N
i =1
N i =1
The various forms of analyses (statistical analysis, sensitivity analysis, probabilistic analysis)
can be performed. Most of these methods are based on the integration of Monte Carlo (MC)
simulations [8]. Three categories of methods have been presently realized:
Direct methods
(Importance Sampling - IS, Adaptive Sampling - AS,
Direct Sampling - DS)
Modified methods
(Conditional, Latin Hypercube Sampling - LHS)
Approximation methods (Response Surface Method - RSM)
pf =
A) The direct Monte Carlo Simulation (MCS) method is based on generation of sets of
realizations of the random variables in the limit state function (with the assumed known
probability distributions) and to record the number of times the resulting limit state function is
less than zero (i.e., failure). The estimate of the probability of failure (Pf) then is simply the
number of failures divided by the total number of simulations (N). Clearly, the accuracy of
this estimate increases as N increases, and a larger number of simulations are required to
reliably estimate smaller failure probabilities. The generation of random variables is a
relatively simple task (provided the random variables may be assumed independent) and
requires only (1) that the relevant CDF is invertible (or in the case of normal and lognormal
variants, numerical approximations exist for the inverse CDF), and (2) that a uniform random
number generator is available. The generation of correlated varieties is not described here, but
information may be found in the literature [8].
The accuracy of this method is depended on the number of the simulations and it is defined by
the variation coefficient
1
(7)
pf =
Np f
where N is the number of simulations. When the target probability of failure is pf = 10-4, thus
the variation factor is equal to 10% for the number of simulations N=106, what is acceptable.
Advantages of the method:
the final values of the reliability reserves can be continuously displayed in the form of
a histogram or cumulative function; the simulations are independent,
the method is easily understandable and transparent,
the method enables the estimation of the statistical discrepancy of the estimation on a
particular relevance level.
Drawbacks of the method:
large number of simulations for small probability values,
slow calculations of complex problems (for Finite Element Method models, the
calculations are expensive and ineffective).
-3-
B) The Modified LHS method is based on the simulations of the function g(X) so thus MC
method, but the definition domain of the distribution function (Xj) is divided to N intervals
with the identical probability 1/N. Characteristic values of the simulations as calculated
randomly on the base of the permutation integer number 1,2,...N. The reliability function g(X)
must be determined from the N simulation. The interval of the random simple is used one
time only. This method gives us the best estimation of the statistical parameters of the
structure in the comparison with MC method. The result of the LHS analysis is the parameter
of the failure function mean value Z , standard deviation z, skew coefficient z, kurtosis
coefficient ez, versus empirical cumulative distribution function.
C) Approximation methods - Response Surface Methods are based on the assumption that it
is possible to define the dependency between the variable input and the output data through
the approximation functions in the following form:
N
N 1 N
i =1
i =1
i =1 j >i
Y = co + ci X i + cii X i2 + cij X i X j
(8)
where co is the index of the constant member; ci are the indices of the linear member and cij
the indices of the quadratic member, which are given for predetermined schemes for the
optimal distribution of the variables or for using regression analysis after calculating the
response [8].
Approximate polynomial coefficients are given from the condition of the error minimum,
usually by the "Central Composite Design Sampling" (CCD) method or the "Box-Behnken
Matrix Sampling" (BBM) method.
Advantages of the method:
considerably less number of simulations than with the straight Monte Carlo method,
it is possible to define dependencies using the "design experiments" method or
regression analysis from the defined points in the case of improper approximation
functions,
particular simulations are independent from each other parallel calculations can be
used here.
Drawbacks of the method:
the number of simulations depends on the number of variable input parameters; in the
case of a large number of input parameters, the method is ineffective,
the method is unsuitable in the case of discontinuous changes in the dependencies
between the input and output values (e.g., the method is not suitable for resolving the
stability of ideal elasto-plastic materials beyond the failure limit...).
The ANSYS Program belongs among the complex programs for solving potential problems
[8]. It contains a postprocessor, which enables the execution of the probabilistic analysis of
structures. In Figure 3, the procedural diagram sequence is presented from the structure of the
model through the calculations, up to an evaluation of the probability of structural failure. The
postprocessor for the probabilistic design of structures enables the definition of random
variables using standard distribution functions (normal, lognormal, exponential, beta, gamma,
weibull, etc.), or externally (user-defined sampling) using other statistical programs. like
AntHILL or FReET. The probabilistic calculation procedures are based on Monte Carlo
simulations (DS, LHS, user-defined sampling) and "Response Surface Analysis Methods
(RSM)" (CCD, BBM, user-defined sampling). The statistical postprocessor compiles the
results numerically and graphically in the form of histograms and CDF.
-4-
Dynamic soil characteristics can be determined from the refraction and reflex investigation of
the locality. The material characteristics of the soil are dependent on the shear and longitude
wave velocity. The wave velocities are measured between two bores for various levels under
free field. The shear velocity of the wave is calculated from the strain in the soil due to
experimental results or comparison with the similar type of soil.
The fundamental material parameter characteristic for the dynamic of soil [3] is dynamic
shear modulus Gdyn (versus Young elastic modulus Edyn)
v 2p 2v 2s
2
2
Gdyn = vs .,
(9)
Edyn = vs . .2 (1 + dyn ) ,
dyn =
2 ( v 2p v 2s )
where is density, vs shear velocity of the soil, vp is longitude velocity, vdyn dynamic
Poisson constant. The stiffness and damping characteristics of the soil depends on the ground
peak acceleration by Eurocode 8.
Dynamic soil modulus measured in situ is adequate for the small dynamic action. These
values are correct for the calculation of SSI in the case of machine foundation and small
seismic loads. The dynamic stiffness of the calculation model of the subsoil depends not only
on physical and mechanical properties of the soil but on the shape and depth of the foundation
[8 and 10] also.
Table 2: Ratio of the dynamic and static modulus [10]
Ground type
Incohesive ground
Cohesive ground
Rock
Edyn/Estat
2,54,0
4,010,0
6,060,0
Estat [MPa]
30-120
6-30
60-700
The dynamic response is other in the case of stiff and soft soil [2, 3, 5, 6, 9 and 10] due to
soil-foundation interaction effects. There are following aspects:
Soil move can affect the rotation of foundation about its horizontal axis,
First period of foundation under soft soil will be longer as in the case of stiff soil,
Eigenvalues and a participation factors will be different in the case of soft and stiff
soil,
No proportional damping is depend on the radial and reflex damping of soil under
foundation and different damping of foundation structure
The consideration of SSI effects is very important. The influence of stiffness and damping
characteristic of the soil to the structure are not negligible.
From the point of view of Eurocode [4] the engineer-designer has take into account following
influences
Impact of machine vibration to structures
Impact of machine vibration to the people and operation (mechanic, acoustic, optic)
Impact of machine vibration to the technology (requirements of manufacturer)
On the base of the evaluation of all influences it is necessary to check following assessment:
-5-
Design values of the displacements, velocities, accelerations and forces have be obtained from
the combination of the static and dynamic load for the normal and extreme machine
performance. The practical experiences prove the fact that the disregard of the soil machine
interaction effects can to cause the malfunction of the machine. The foundation shape
machine design has to respect the designing principles as follow:
The foundation masses must be higher than the machine device or equipment,
The eccentricity between the machine mass centre and the foundation contact area
centre have be less than 5% of the foundation dimension,
In the case of the vertical excitation forces the horizontal foundation dimension has be
lower than the vertical distance between the machine centre and contact area otherwise
in the case of the horizontal excitation forces the horizontal foundation dimension has
be higher,
The foundation dimensions ensure against its collapse, the stability of the foundation in
the rotation has be guarantee,
The optimal design of the machine foundation pursues the minimization of the dynamic
effects of the machine to the structure. This task can be achieved to respect following
requirements:
To change the frequency tuning of the structure (to change the ratio between the impact
and structure frequencies) by changing the stiffness, masses of the system, etc.
To change the localisation of the structure and equipment or the dynamical effect of this
system,
To reduce the effect of the excitation forces,
To improve the stiffness and damping characteristics of the subsoil,
To design the efficiency damping system between the machine and foundation,
-6-
Machine
Natural
frequency
f p / f n < 0 .6 or 1 .5 < f p / f n
3 < f p / fn
The efficiency of the dynamic characteristics of the machine foundation can be improved by
avoid the resonance effects. The recommended ratio of the natural frequency of the structure
fn and the frequency of the machine fp are presented in Table 5.
Soil
13K401L
13K401M
13K401H
Low
Medium
High
Direction X
Direction Y
Direction Z
Frequency Prop. ratio Frequency Prop. ratio Frequency Prop. ratio
[Hz]
[Hz]
[Hz]
[%]
[%]
[%]
15,02
48,9
12,06
53,9
14,66
65,8
18,18
51,7
14,69
55,1
16,47
54,8
22,04
52,5
17,91
55,4
17,78
41,8
-7-
On the base of measured data three soil models low, medium and high were incorporated in
the FEM model (Table 6). The stiffness of soil has the considerable influence to the modal
characteristic and the eigenvalues of entire structure (Figure 2).
( x xo ) k
( y yo ) k
k ( x, y ) = k z.var + 2
(10)
yy.var + 2
xx.var k z ,k
Lx
Ly
where kz.k is a characteristic value of soil stiffness xo, yo are coordinates of foundation
structure gravity centre, Lx and Ly are the plane dimensions of the slabs in directions x and y.
The variability of geometric characteristics is defined with h.var (column dimension), d1.var
(foundation plate thickness), d2.var (compressor plate thickness). The stiffness of the structure
is determined with the characteristic value of Youngs modulus Ek and variable factor evar.
A load is taken with characteristic values Gk, Fk, Fr.k and variable factors gvar, fvar and fr.var
(Table 7). The uncertainties of the calculation model are considered by variable model factor
R and variable load factor E for Gausss normal distribution. The results of the probability
analysis of the foundation model present that the principal frequencies are variable in the
direction X (from 4,32HZ to 6,37Hz), Y (from 13,05HZ to 17,61Hz) and Z (from 16,84HZ to
21,68Hz). These frequency intervals have the important influence to response from the
harmonic compressor excitation.
-8-
Name
Soil
Material
Load
Geometric
Model
Min.
value
0,67
-1
-1
0,868
0,719
0,752
0,719
0,860
0,972
0,972
0,719
0,719
Max.
value
1,5
1
1
1,149
1,281
1,317
1,281
1,140
1,028
1,028
1,281
1,281
-9-
Sensitivity analysis
Sensitivity analysis of the influence of the variable input parameters to the reliability of the
structures depends on the statistical independency between input and output parameters.
Matrix of correlation coefficients of the input and output parameters is defined by Spearman
in the form [8]
n
rs =
( Ri R ) ( Si S )
i =1
( Ri R ) ( Si S )
2
i =1
(13)
2
i =1
where Ri is rank of input parameters within the set of observations [xi]T, Si is rank of output
parameters within the set of observations [yi]T, R, S are average ranks of the parameters Ri
and Si respectively.
Figure 5: Sensitivity analysis of the vertical displacement for normal and extreme
performance
- 10 -
Variability of three input quantities (velocity of the turbine rotor, load amplitudes, foundation
mass and stiffness) is important to the displacement of compressor foundation (Figure 4) due
to normal performance of rotor. The frequency of rotor movement is lower in the case of
extreme performance than the normal performance. It is the reason of the higher sensitivity of
foundation to the variability of the model input parameters. The sensitivity of the vertical
displacement over the compressor operation frequencies is demonstrated in the Figure 5a for
normal performance and in the Figure 5b for the extreme performance. The horizontal
displacements of the compressor foundation are higher for the lower frequency as 5Hz. In the
case of vertical displacements their peaks are about the frequency 15Hz for both performances
normal and extreme (Figure 5).
The comparison of deterministic and probabilistic solution of the safety and reliability of the
compressor foundation is documented in the table 8.
Table 8: Comparison of deterministic and probabilistic analyses
Method
Deterministic
Probabilistic
Deterministic
Probabilistic
Deterministic
Probabilistic
Deterministic
Probabilistic
Deterministic
Probabilistic
Deterministic
Probabilistic
Model
Z4L
Z4M
Z4H
Z4L
Z4M
Z4H
The differences between deterministic and probabilistic results are equal about to 5-33% (or
50-100%) for mean (or maximum) displacement amplitude values. In the case of normal
forces and bending moment these differences are lower.
Conclusion
This paper deals with the possibility of the sensitivity and probabilistic analysis of the
reliability of the compressor foundation depending on variability of the soil stiffness, structure
geometry and machine operation. The sensitivity of the machine foundation to the
uncertainties of the soil properties due to long-time rotating movement of machine is not
negligible for design engineers. On the example of compressor foundation 13K401 and
- 11 -
turbine GK22/28 fy. SIEMENS AG the affectivity of the probabilistic design methodology
was presented. The simulation method RSM for the analysis of the compressor foundation
reliability was used on program ANSYS. The 151 simulations for five load cases were
calculated in the real time on PC (CPU=626sec). The probabilistic analysis gives us more
complex information about the soil-foundation-machine interaction than the deterministic
analysis.
Acknowledgements
This project was realized with the financial support of the Grant Agency of the Slovak
Republic (VEGA 1/0849/08).
References
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]