Você está na página 1de 13

Finite Element and Probabilistic Analysis for Simple Truss Bridge

Chen-Min Huang Research Experiences for Undergraduates in Japan in Advanced Technology (REUJAT) University of California, Irvine Department of Civil and Environmental Engineering

ABSTRACT In this study, the Finite Element Method was combined with probabilistic concepts from the Monte Carlo Simulation (MCS) to compute the failure probability of a steel truss bridge. The steel bridge consisted of 17 elements and 10 nodes with multiple loading conditions at the deck. Random cross-section areas with normal distribution were analyzed by the FEM to determine the stress intensity factors of each element. The results of this study showed the computing power and flexibility of computer algorithms in identifying the structural properties and failure probability of the bridge.

INTRODUCTION Structural reliability analysis is an increasingly important tool that assists engineers to consider uncertainties during the design, construction, and life of a structure in order to calculate its probability of failure. The reliability of a structure is an essential factor in fulfilling a successful structural design. Advances in structural analysis by the Finite Element Method allow engineers to take into account most details of a complex object or structure in the finite element model such as coordinates, material properties, loads, etc. Numerical simulation is now used in conjunction with Finite Element Method to predict the behavior and response of these finite element systems. However, most real-life values in the finite element model differ from assumed values due to tolerances or scattering [3]. Therefore, the question of how safe the structure is in reality can be answered by considering the stochastic nature of uncertain structural parameters such as element properties, loads, ands geometry. These uncertain parameters are all represented by stochastic or random variables that is not of a fixed value, but by a probability distribution function that characterizes the behavior and safety of a structure [4]. This paper focuses on probabilistic finite element analysis as a means of quantifying the reliability of a truss bridge subject to external loading and inherent randomness in geometric imperfections. Coupling the probabilistic analysis and finite element method is a computer program written at Keio University, which integrates a finite element solver and the Monte Carlo simulation. An example problem consisting of a truss bridge is used to demonstrate concepts and assess the structures probability of failure. CONCEPTS OF FINITE ELEMENT METHOD

Finite element analysis is becoming a routinely used method for obtaining numerical solutions to very complicated stress problems, particularly in the engineering field. The power of the Finite Element Method (FEM) allows the understanding of how various elements behave with arbitrary shape, loads, and boundary conditions. It can handle complex restraints, which allow indeterminate structures to be solved. Lastly, all this information can be contained within a single computer program for users to input data such as an objects geometry, boundary conditions, and element selection. FEM is a numerical technique for solving problems which are described by partial differential equations. The first step is to divide a solution region of the mathematical model into several or many finite elements connected at discrete points called nodes as shown in step 1 and step 2 of Figure 1. The response of each element is expressed in terms of a finite number of degrees of freedom characterized as the value of an unknown function, or functions, at a set of nodal point [2, Chap. 7]. The response of the mathematical model is approximated by connecting all elements as if nodes were pins that hold elements together. Fabricating complex objects into simpler components can be easily visualized in bridges and buildings.
Element

Node

Figure 1. Graphical representation of FEM for a circle [2] Matrix Analysis of Trusses The Direct Stiffness Method (DSM) of structural analysis is by far the most common implementation of FEM programs and is applied to a simple plane truss structure [2, Chap 2]. DSM is applied to pin-jointed plane trusses particularly for short-span bridges and buildings in the field of civil engineering. The stiffness of each truss element is considered one at a time to determine the forces that are set up in the truss elements by the displacements of the nodes. As a result, a set of simultaneous algebraic equations are written in matrix form as shown with nodal displacements as unknowns and applied nodal forces as known quantities:

(Eq.1) Here, vector quantities ui and fj indicate the deflection and force at the ith and jth node respectively. The Kij coefficient is called the global stiffness matrix with the influence of the deflection from the applied force. The matrix equations can also be written as

Kijuj or Ku = f

(Eq, 2)

Stiffness Matrix for Single Truss Element A set of matrix equations that describe truss systems can be developed with a relationship between the forces and displacements at each node of a single truss element A simple structural finite element is the 2-node bar element in the x-y plane, attached to nodes i and j and include at an angle from the horizontal as illustrated in Figure 2.
P

Figure 2. Individual 2-node bar element Because elongation vector is directed along and transverse to the element, the elongation in the truss element can be written in terms of the differences in the displacements of its two nodes in Equation #:

e = (u j u i ) cos + (v j vi ) sin

(Eq. 3)

where u and v represent the first and second nodes of a single truss element respectively and i and j represent the horizontal and vertical displacement direction respectively. A matrix form of this relation can be expressed as:

(Eq. 4) where c = cos and s = sin. Taking the Mechanics of Materials approach by treating the truss member in Figure #(a) as a linear spring with uniform member properties along its length as depicted in Figure #(b), the bar theory describes stiffness as: EA ks = (Eq. 5) L The axial force P that accompanies the elongation is given by Hookes law for linear elastic bodies as P = (AE/L).

-P

Figure 3. (a) idealization as bar element, (b) interpretation as equivalent spring [2] The horizontal and vertical nodal forces in Fig. 3 can be written in terms of the total axial force as: EA F = ks d = d or L (Eq. 6) ui f xi c v f yi s AE [ c s c s ] i = u j f xj c L v j f yj s

Carrying out the matrix multiplication:

(Eq. 7) The quantity in the brackets, multiplied by AE/L, is known as the element stiffness matrix.
CONCEPTS OF MONTE CARLO SIMULATION

A major part of this study is dedicated to the concept of probability, particularly within the Monte Carlo Simulation (MCS). The objective of MCS is to generate a sequence of pseudo-random numbers that satisfy essential properties of actual random numbers. These numbers are randomly generated between 0 and 1 from a program or table, and then read from the quantile associated with that number [4]. Because probabilistic analysis methods are approximate, errors in computed probabilistic results will always be present due to characteristics of the numerical solution employed, lack of data, and human error. The MCS method is often applied when the analytical solution is not possible and the failure domain cannot be quantified or approximated by an analytical form [9]. A sufficient large number of randomly distributed samples is evaluated and the ratio of failed to the total number of samples is determined. The mathematical

formulation of the MCS has the capability of handling practically every possible case regardless of its complexity. However, calculating the low or high probabilities generally requires hundreds of thousands of simulations, and the effort to perform such simulations involves excessive finite element computations if there are more than a few basic variables.
EXPERIMENTAL APPROACH

Before a truss bridge can be constructed, important specifications about its shape, dimensions, member forces, strain, and stress, are meticulously calculated by design engineers that allow the bridge to maintain its structural integrity. However, not everything goes according to plan during the actual construction phase. For example, although the length and cross sectional area of each truss member should be manufactured to design specifications and dimensions on paper, marginal errors still exist among some members due to poor measurement equipments, human carelessness, or rushed production deadlines. These marginal errors reduce the structural integrity of the bridge, increase the likelihood of a collapse, and endanger human lives.
Problem of Study

An investigation is conducted in the Noguchi laboratory at Keio University to determine the probability of a steel truss bridge failing by combining and programming the probabilistic concepts from the Monte Carlo simulation with FEM. In this experiment, assume truss beam members produced by a particular manufacturer for bridges are always precise lengthwise according to design. Because of poor measurement equipment, the width and height of a few members are not accurately shaped, which cause the members cross sectional area to be slightly bigger or smaller than the design value. Before the failure probability can be determined, the maximum stress intensity must be determined and plotted on a histogram by using the FEM and MCS algorithm.
Finite Element Algorithm

The first step is to write a finite element code in Fortran 77 language. Most of the finite element code was written by and borrowed from Yuuichi Tadano, a PhD student in the Noguchi laboratory. The Direct Stiffness Method is applied for this FEM program and left be to code in the subroutine name element. The element subroutine algorithm creates the stiffness matrix for each truss element, which is used to compute the displacement at each node as described by Eq. 7. The next step is to code another subroutine that calculates and outputs the reaction forces, strain, and stress. This subroutine first computes the elongation e of a plane truss bar as shown in Eq. 3. After e is computed, the axial stress can be described below: e = E = E (Eq. 8) L where axial strain = e/L. Lastly, a particular reaction force, Ri at each node and its support reaction can be computed if any exist from Eq. 9:

K
j

ij

D j = Ri or

k
m

ij

d j = Ri

(Eq. 9)

Physical and Material Properties of Truss Bridge

Once the FEM source code is completed, the design of the truss bridge is inputted into the sample file. The sample data file allows users to easily input new or existing nodal coordinate, external load, and material property values for each element. In this research, the nodal coordinates and elements are inputted to resemble the two-dimensional truss bridge shown in Figure 4. The truss bridge has seventeen elements and ten nodes with two degrees of freedom. There is one fixed support in the horizontal and vertical direction at the bottom left node, and one fixed support in the vertical direction only at the bottom right node. The total span of the bridge is 20 meters with each unit on the graph being equivalent to 1 meter.
Node 9 Node 8 Node 7 Node 6

Element 8 Element 12

Element 7

Element 6 Element 16

9 9

13

n n nt en em Ellll

El Ele em me en ntt

El em en t

t en m m m em Ellll

t t t t e e e en em E E E El 1 1

Element 10

Element 14

5 5 5 5 1

Element 17

Element 1

Element 2

Element 3

Element 4

Node 10 (Fixed)

Node 1

Node 2

Node 3

Node 4

Node 5 (Rolling)

Figure 4. Illustration of inputted values in sample data file Lastly, property values for the bridge are assigned to complete the information needed to perform the finite element analysis. Each truss bridge element is comprised of steel materials with a modulus of elasticity of 200 GPa. Vertical external forces of 4000 N are applied at the four free nodes to represent dead loads felt throughout the deck. No external loads are at any of the fixed nodes. A horizontal force of 1000 N is applied at the top-left node of the bridge as shown in Fig. 5 to represent the design wind load acting on the bridge. The design value of the cross-sectional area for each member is calculated to be 0.01 m2. Once the dimensions and properties of the truss bridge are inputted into the sample data, the new displacements, stresses, and reaction forces can be easily computed from the finite element code. A scaled down simulation is conducted using the Virtual Laboratory: Bridge Designer program by the Johns Hopkins University that computes the internal force of each member after the external loads are entered into the program.

Figure 5. 2-D steel truss bridge with external loads used in FEM code The largest internal forces are felt at elements 2 and 3 with 124 N and elements 7 and 8 with 126 N as shown in Fig. 5. From these results, it is concluded that the greatest stress intensity would also occur at these four members, therefore, any changes to these members can compromise the structural integrity of the bridge.

Figure 6. Analysis report after running FEM program Fig. 6 shows the displacement, reaction force, stress and strain values after completing the full scale finite element analysis from the sample data file. All results are shown in SI units. It is noted that elements 2, 3, 7, and 8 are the highest stress intensities as previously observed from the earlier simulation. Thus, these four elements are the most sensitive in the truss bridge, and therefore changing the properties of any of the four elements may cause the bridge to sustain structural instability and damage.
Approach to Probabilistic FEM Analysis

Random Number Generator With the finite element code and truss bridge complete, the next step is to apply probabilistic FEM. The random number generator algorithm by LEcuyer et al is used to run the Monte Carlo Simulation [4]. This code allows users to generate random numbers

between 0 to 1 in parallel or several streams at once. A 10,000 sample is generated and plotted on a histogram in Figure 7.

Figure 7. Histogram of 10000 random numbers between 0 and 1 Figure 7 shows a uniformly distributed histogram, which indicates that this code effectively generates pseudo random numbers that satisfy the essential properties of actual random numbers [1]. Random numbers uniformly distributed between 0 and 1 allow simulation from any distribution. Therefore, this source code is recommended for use because of its flexibility to generate random samples from any distribution to run the Monte Carlo simulation. For this research paper, the random generator algorithm is altered so the 10,000 random samples follow a normal distribution between -4 and 4 instead of a uniform distribution between 0 and 1. The number of random samples desired can be easily increased by the user if so desired. The values of the random numbers can now be adjusted to solve any unknown basic variables such as load factors, element properties, and material data within this continuous distribution. In this paper, the focus is on the cross-section areas of elements 2, 3, 7, and 8, which are assumed to be uncertain and follow a normal distribution similar to Fig. 8 below.

Figure 8. Histogram of 10,000 normally distributed random samples between -4 and 4.

The cross-section areas of the most stress intensive elements (element 2, 3, 7, 8) are assumed to be uncertain with a normal distribution. As previously mentioned, any slight change to the properties of one of the four elements may cause the bridge to sustain damage and increase its chance of failure. Assume that all 17 members are designed with a cross section area of 0.01 m2. Therefore, if random area values are drawn with a normal distribution, the expected value, , of the cross-section area is 0.01 m2. Three cases are considered to observe how the probability of failure changes as values of the cross-section area deviate farther away from the expected value. The first case is when the members cross-section area has a small margin of error within = 0.001. Second, a member has a moderate margin of error when = 0.002. Lastly, a large margin of error is when the range of random area values fall between = 0.003.

Figure 9. Random cross-section area values with different boundary conditions The random number generator algorithm with normal distribution is scaled to fit the crosssection area model to be sampled. 10,000 random area values are generated and plotted on a histogram for each of the three cases as shown in Fig. 9. With 10,000 randomly generated values per case, the Monte Carlo simulation and FEM program work concurrently to determine the maximum stress value in the truss structure. First, the Monte Carlo simulation randomly draws 10,000 area values and replaces the first random value in element 2, while the cross-section area remains as 0.01 m2 for the other elements. The FEM is run to analyze the new displacement, reaction force, and stress values. The program loops and replaces the first random value with element 3, 7, and lastly 8. After the program calculates the FEM results from all four elements, the maximum stress value from all four sets of stress values is recorded for that random number. The program runs in this manner until all 10,000 values are analyzed. Therefore, each of the three cases will have an output of 10,000 maximum stress values. The structure fails if the maximum stress on any of the 17 members exceeds a limit value which by itself is a stochastic value with a normal

distribution. The maximum stress limit value is set at the third deviation from the mean. Any number that exceeds that value contributes to increasing the probability of failure.
RESULTS AND ANALYSIS After the FEM analysis is completed and all the maximum stresses are found, the data can finally be evaluated to determine the probability of failure. The values of the maximum stress are graphed on a different histogram for each case as shown below in Fig. 10 a-c.

Figure 10 (a). Case 1: Area range of 0.009 0.011 m2

( 1)

Figure 10 (b). Case 3.Area range of 0.008 0.012 m2

( 2)

Figure 10 (c). Case 3: Area range of 0.007 0.013 m2 ( 3) Note that maximum stress is a stochastic value and follows a normal distribution when it is graphed. Next, the mean, and standard deviation, are computed for each histogram. The standard deviation is multiplied by 3 and its result is added to the mean to arrive at the maximum stress failure limit. For case 1, the maximum stress limit is 1.21 x 108 Pa. For case 2, the maximum stress limit is 1.31 x 108 Pa. Lastly, for case 3, the maximum stress that is allowed for the bridge before failure is 1.41 x 108 Pa. The last step is to solve for the probability of failure of the bridge, which is the ratio of failed to the total number of samples. Any stress value over the maximum limit is added to the failure counter and divided by 10,000 for each case. These values can be read directly from the stress and frequency column in Figures #-#. The probability of failure for case 1 is 9.80 x 10-3 or 0.0098 % when members have a random cross-section area between 0.0090.011 m2. For case 2, the failure probability is 6.60 x 10-3 or 0.0066%, and increases to 1.06 x 10-2 or 0.0106 for case 3.
CONCLUSION

In this study, the Monte Carlo simulation and Finite Element Method were employed to compute the maximum stress limit and probability of failure of a truss bridge. The crosssection area values were unknown and 10,000 randomly values were chosen within different boundary conditions. Three cases were considered in this study, and its results plotted on a histogram. From the data results shown above in Figure 10a and Figure 10c, it is important to note that the probability of failure increased by 7.5% as the boundaries of the cross-section area values increase away from the mean. In addition, the failure probability increased by 62.3% from 6.60 x 10-3 Pa to 1.06 x 10-2 Pa in Figures 10b and 10c. Therefore, the failure probability shows an overall increase as the boundary conditions

for the cross-sectional area widens. With more faster and powerful computers, the Monte Carlo and FEM simulation has proven to be a convenient and powerful method for reliability analysis of bridges and structures.
FUTURE WORK

As discussed previously, the probabilistic FEM analysis used to determine the probability of failure from a normal distribution curve gave an approximate value only. Due to the short duration of time on this project, more time will be spent to confirm the accuracy and credibility of the results. The first is to find the exact area under the curve by computing the integral of the normal distribution function for all three cases. The actual failure probabilistic value will be compared with the approximate value for error and accuracy analysis. These values will help to address the errors that may be present in this study such as insufficient number of samples, human error, or algorithm error.
ACKNOWLEDGEMENT

I would like to thank Dr. Shirley Dyke and Dr. Makola Abdullah for their help and dedication to the 8 undergraduates and 1 graduate students in the REUJAT program. I am very grateful for the support of the National Science Foundation for funding the REUJAT program every summer and providing students with the opportunity to conduct research in Japan. Deep gratitude is extended to Dr. Yozo Fujino at Tokyo University for hosting the program in Japan. I sincerely thank my Japanese advisor, Dr. Hirohisa Noguchi at Keio University for providing an excellent facility to conduct research as well as his guidance during my stay in Japan. A great appreciation is extended to Dr. Kawai who helped me thoroughly understand the technical and theoretical aspects of this research subject.
REFERENCES

[1] Deutsch, Clayton V. et al., The Principles of Monte Carlo Simulation, Lecture Four: Monte Carlo Simulation. Centre for Computational Geostatistics, University of Alberta. [2] Felippa, Carlos A., Introduction to Finite Element Methods. Department of Engineering Sciences and Center for Aerospace Structures. University of Colorado at Boulder, 2003, http://caswww.colorado.edu/courses.d/IFEM.d/Home.html. [3] Fischer, Rolf, Reliability Analysis and Optimization in PERMAS. Stuttgart, Germany. 2003 May 7-8. [4] L'Ecuyer, R. Simard, E. J. Chen, and W. D. Kelton, An Objected-Oriented RandomNumber Package with Many Long Streams and Substreams, Operations Research, 50, 6 (2002), 1073-1075. [5] Riha, David S. et al., Capabilities and Applications of Probabilistic Methods in Finite Element Analysis. 5th ISSAT International Conference on Reliability and Quality in Design, Las Vegas, Nevada, 1999 August 11-13. [6] Robert Cook et al., Concepts and Applications of Finite Element Analysis, John Wiley & Sons, 1989

[7] Roylance, David, Finite Element Analysis, Massachusetts Institute of Technology. Cambridge, Massachusetts. 2001 February 28 [8] Thacker, Ben H. et al., Errors and Uncertainties in Probabilistic Engineering Analysis. Southwest Research Institute. San Antonio, Texas. 2001 [9] Yiannis Tsompanakis et al., Reliability Analysis of Structures Under Seismic Loading. 5th World Congress on Computational Mechanics, Vienna, Austria, 2002 July 7-12.

Você também pode gostar