6th World Congresses of Structural and Multidisciplinary Optimization

Rio de Janeiro, 30 May - 03 June 2005, Brazil

Experimental Force Identification by using Natural Optimization

Felipe A. Chegury Viana, Jhojan E. Rojas Flores, Domingos A. Rade, Valder Steffen, Jr.
Federal University of Uberlndia, School of Mechanical Engineering Av. Joo Naves de vila 2121 Campus Santa Mnica P.O. Box 593 CEP 38400-902 Uberlndia, MG - Brazil 1. Abstract In various applications related to structural engineering, it is important to determine the external loading under real service conditions to evaluate, for example, the level of safety of the structure. However, the determination of external loading is not simple from the experimental point of view. On the other hand, the influence exerted by the external load on the dynamic response of the system through the so-called stress-stiffening effect can be taken into account in order to obtain information about the load distribution, by exploring the dynamic responses, through an inverse problem approach. In the present paper, optimization methods are used to solve the inverse problems, involving both classical and heuristic approaches. The authors explore an identification procedure based on the LifeCycle Model. This heuristic is inspired on the biological concept of life cycle, which defines the passage through various phases during the life of an individual. Genetic Algorithm and Particle Swarm Optimization are used as search methods and the LifeCycle Model intends to combine the advantageous characteristics found in both methods. Applications are performed based on numerical simulations and experimental tests on a column loaded in compression with various boundary condition configurations, for which a set of natural frequencies is used to form the cost function to be minimized. The results obtained demonstrate the feasibility of the identification procedure. 2. Keywords: Inverse problems, experimental identification, LifeCycle model, Particle Swarm Optimization 3. Introduction In Structural Engineering, it becomes of paramount importance to determine the external loading under real operational conditions, aiming at evaluating the level of security of the structure, to verify the design configurations that were adopted at the design stage, or for redesigning structural elements for new operating conditions. However, the determination of external loading is not simple from the experimental point of view because, in general, transducers cannot be easily introduced in the structure during its construction and/or assembling. On the other hand, it is well known the fact that external loads can influence the static and dynamic behavior of structural systems, through the so-called stress-stiffening effect [1], [2]. Reference [3] was the first one to put in evidence the effect of axial loads on the natural frequencies of structural components. In [4] it was recognized a common theoretical foundation underlying free vibration and stability analyses. In [5] it was demonstrated the existence of a linear relation between the axial load and the natural frequencies corresponding to the lateral motion of a simply supported column. The changing of natural frequencies as related to stability problems was also discussed in the works reported in [6] and [7]. More recently other authors investigated analytically and experimentally the influence of axial loads on the vibration of beams under various boundary condition configurations [8], [9]. In [10] the authors demonstrated the possibility of introducing residual stresses as a mean to improve the mechanical behavior of thin plates. It has been demonstrated elsewhere [11] that, by taking into account the influence exerted by the external loading on the dynamic response of the system through the so-called stress-stiffening effect, it is possible to obtain information about the loading distribution, given the dynamic responses, by solving an inverse problem. Such procedure presents a number of advantages but also some difficulties that have to be dealt with, as pointed-out by [12]. In the context of inverse problems, reference [13] used modal parameters combined with Least Squares to estimate the axial loads of Euler-Bernoulli beams having elastic supports. In [14] it was studied the effect of the application of an axial load to one of the bars in a truss structure by using experimental dynamic responses in a model fitting approach, in which the axial loads were considered as parameters to be adjusted. The results were then compared to the static loads as calculated from experimental measurements by using strain gages. The sensitivity analysis of the parameters to be adjusted has also been carried out. Besides, through experimental tests in a similar structure, reference [15] analyzed the effect of residual stresses due to the construction process on the modal characteristics of the structure. There exist various techniques to solve inverse problems by using optimization methods, involving either classical as well as heuristic approaches. Once the parameters are identified, the mathematical model becomes an effective tool to analyze and predict the dynamics of the structure under different operating conditions. It is broadly recognized that the solution of inverse problems by using classical gradient-based optimization methods is a difficult task due to the existence of local minima in the design space. Moreover, such methods require an initial guess to the solution and it is not possible to assure global convergence. These aspects have motivated the authors of this paper to explore a hybrid approach for the determination of external loading in structures, based on two heuristic methods, namely Genetic Algorithms (GA) and Particle Swarm Optimization (PSO) combined in an strategy known as LifeCycle model, which has been introduced in reference [16]. In this paper the optimization procedure is used for the identification of axial loads applied to a straight column subjected to different boundary condition configurations. Both numerical simulations and experimental tests are considered in such applications, in which a set of natural frequencies and vibration mode shapes of the loaded structure as well as a finite element model of the structural system are used to constructed an objective function that is minimized to determine the optimal values of the external loads.

4. Finite Element Modeling of the Dynamic Behavior of Beam-Columns Including the Stress-stiffening Effect In this section it is briefly reviewed the finite element modeling of two-dimensional beam-like structures, according to the theory of Euler-Bernoulli, including the effect of the axial load, as illustrated in Fig. 1.

y viL ( t ) Ni L uiL ( t )

qi (x,t)

iL ( t )
pi (x,t) li Ei, Ai, Ii,i R

viR ( t )

iR ( t )
Ni uiR ( t ) x

Figure 1. Two-dimensional beam element. In Fig.1 li is the length of the element; Ai is the area of the cross section; Ii is the area moment of inertia; pi is the distributed longitudinal load; qi is the distributed transversal load; Ni is the nodal load applied in the axial direction; Ei is the modulus of elasticity of the material; i is the density of the material; i is the nodal cross section rotation, ui and vi are the longitudinal and transversal nodal displacements, respectively. The indexes L and R indicate, respectively, the displacements and rotations at the left hand and right hand nodes of the element. Using a linear interpolation function to represent the longitudinal displacement and a cubical function for the transversal displacement, the following expressions for the element stiffness and mass matrices are obtained [17]: Ei Ai l i Ki = sim 0 12 Ei I i 6 N i + 5 li li3 0 6 Ei I i 1 + Ni 10 li2 4 Ei I i 2 + N i li 15 li Ei Ai li 0 0 Ei Ai li 0 12 Ei I i 6 N i 5 li li3 6 Ei I i 1 Ni 10 li2 0 12 Ei I i 6 N i + li3 5 li 6 Ei I i 1 + Ni 2 10 li 2 Ei I i 1 N i li 30 li 0 6 Ei I i 1 2 Ni li 10 4 Ei I i 2 + N i li li 15 0

(1)

0 140 0 156 22li 4li2 mi Mi = 420 sim

70 0 0 70

0 54 13li 0 156

0 13li 3li2 0 22li 4li2

(2)

where mi = i Ai li . The effect of the axial load can be observed in the stiffness matrix at the elements corresponding to the bending stiffness, representing, therefore, the so-called stress-stiffening effect. The element matrices are mounted by using the standard finite element (FE) matrix assembling procedure based on the connectivity of elements. Figure 2 illustrates a column elastically supported at both ends for different boundary conditions. Such boundary conditions were modeled in this work to account for the flexibilities of the supports, which lead to deviations from the ideal perfectly rigid conditions. Then, pinned-pinned (P-P), pinned-clamped (P-C), and clamped-clamped (C-C) boundary conditions were substituted by appropriate flexible supports. The flexibility of the supports can be introduced in the model by a proper modification of the stiffness matrix, in which the values of the translational and rotational spring coefficients are added to the diagonal terms corresponding to the coordinates to which they are attached.

k1x

k1y

k1x

k1y

k1x

k1y k1t

k2t k2y k2y

k2t k2y

k2x (a)

k2x (b)

k2x (c)

Figure 2. Flexible support for various boundary conditions - (a): P-P, (b): P-C, (c): C-C. Equation (3) represents the global equations of motion in the matrix form:
M X (t ) + K ( p) X (t ) = Q (t )

(3)

where p is the vector of the axial loads applied to the beam elements that form the finite element model of the structure. From the equations of motion, the following eigenvalue problem can be derived:
K ( p) M X = 0

(4)

where = is an eigenvalue (natural frequency) and X is an eigenvector (mode shape). The matrix form of the frequency response functions (FRFs) is calculated as:
2

H ( ) = K ( p ) 2 M

(5)

where is the excitation frequency. The equations above show that the dynamic responses depend on the axial loads (p), which depend directly on the external load applied to the structure. Before performing the dynamic analysis of the structure, a static analysis must be carried out to determine the axial loads for each element, as explained in [11].
5. Load Identification as an Optimization Problem

5.1. Problem Formulation Load identification is dealt with by formulating a constrained optimization problem in which the design variables are the parameters that characterize the external loads and the boundary conditions. In this paper, it was adopted a cost function representing the dimensionless difference between the values of the experimental natural frequencies of the loaded beam and those predicted by the Finite Element Model, as described in Section 4. Thus, the optimization problem is formulated as follows:
min J =

W
p =1

(pm ) (pc ) ({ p})

(6)

1 n (m) i n i =1

(7)

where {p}designates, generically, the set of unknown load parameters, n is the number of natural frequencies used for identification, Wp are user-defined weighting factors and p(m) and p(c) designate the measured and experimental values of the natural frequencies, respectively. Side constraints are introduced to limit the values of the design variables within a feasible design space, avoiding the possibility of buckling or yielding due to extreme external load levels. In the applications considered in this work, using the first six natural frequencies and limiting the value of the total load between zero and approximately 75% of the buckling load value of the structure, the cost function was constructed. Obviously, when both the position and the direction of the load are to be identified, the design space becomes discrete and its dimension depends on the maximum number of nodes of the finite element model (considered as candidate positions).

5.2. LifeCycle Model an Overview From the biology point of view, the term refers to the passage through the phases during the life of an individual. As examples of life phases, can be cited the sexual maturity and the mating seasons. LifeCycle Model is inserted in the natural optimization context, following other heuristics such as Genetic Algorithms (GA) and Particle Swarm Optimization (PSO). As in nature, the ability of an individual to actively change its own phase or stage in response to its success to the environment is the main inspiration for LifeCycle. In fact, the idea behind LifeCycle is to use the transitions to handle the mechanism of self-adaptation to the optimization problem. The fitness value offers a criterion used by each individual to shift from one life stage to another and vice-versa. To close the definition, LifeCycle stages must be defined. In the present work, two heuristics are used as stages, namely the GA and the PSO. Other versions of the LifeCycle can be proposed by considering other heuristics and even a mix of them, as shown in [16]. To obtain details about GA see [18] and [19]. To learn more about PSO, see [20] and [21]. For more detailed information about LifeCycle Model check [16] and [12]. The outline of a basic LifeCycle algorithm is as follows:
Initialize the algorithm parameters for the PSO and GA Create an initial swarm of PSO particles and/or initial population of GA individuals

Evaluate the objective function for all PSO particles For all particles No Is there recent improvement? Switch the LifeCycle stage (change from PSO to GA)

Evaluate the objective function for all GA individuals. For all individuals No Is there recent improvement? Switch the LifeCycle stage (change from GA to PSO)

For all PSO particles, run the PSO algorithm

For all GA individuals, run the GA algorithm

No

Yes Stop criterion Results

Figure 3. Outline of a basic LifeCycle algorithm.

6. Numerical Simulation Figure 4 shows a finite element model of two columns used in the identification, which are submitted to different load scenarios.
Scenario 1 Scenario 2 k1x 11 10 9 F1 L 8 7 L 6 5 4 A 3 2 1 A A 11 10 9 8 7 6 5 4 3 2 1 k2x A k2t k2y F1 = 267.6317 N F2 = 535 N h k1t F2 Finite element model 10 elements 33 degrees of freedom L = 0.650 [m] b = 0.02 [m] h = 0.004 [m] E = 2.1x1011 [N/m2] = 7800 [Kg/m3] Degree of freedom 2 3 Node i b 1 k1y

Figure 4. Finite element model of the beam-like column.

According to Fig. 4, the following scenarios are considered: 1- identification of the magnitude, position and direction of F1; 2 identification of the magnitude of F2 and the values of the each spring that determine the boundary condition. From the optimization point of view, scenario 1 illustrates an optimization problem containing continuous and discrete variables simultaneously. Table 1 shows the setup for LifeCycle. Table 2 presents the identification results for the various loading scenarios. As in real applications, identification methods have to be robust enough to deal with experimental errors. Consequently, it is also considered a situation in which the values of the natural frequencies are corrupted with 5% of random error, as also presented in Tab. 2. The results show that the optimization approach used in the force identification procedure was efficient for all scenarios analyzed. Table 1. LifeCycle parameters used in numerical simulation.

LifeCycle

Number of Individuals 100

Number of Iterations 75

Stage Interval 5

Table 2. Identification results for the beam-like column.

Without corrupted data Optimum Error [%] 270.2345 0.9725 9 2 535.18 -0.0161 1.0434 x 1011 4.3385 With corrupted data Optimum Error [%] 259.5107 -3.0344 9 2 565.15 5.5835 9.6086e+010 -3.9144

Scenarios F1 [N] 1 P D F2 [N] k1 x

Exact 267.6317 9 2 535.26 1.0 x 1011

k1 y
2

1.0 x 1011 1.0 x 10 1.0 x 10 1.0 x 10 1.0 x 10

11 11 11 11

1.0276 x 1011 1.0356 x 10 9.8459 x 10 9.498 x 10

11 10

2.7648 3.5629 -1.5412 -5.02 -3.3653

1.0731e+011 9.9948e+010 1.0304e+011 1.0415e+011 9.3371e+010

7.3099 -0.05209 3.0378 4.1474 -6.6293

k k

2 x

2 y
1 t

k k

10 10

2 t

9.6635 x 10

The evolution of the LifeCycle along the iterations can be observed in Fig. 5 for scenario 1, for the case in which the effect of simulated random noise was included. Figure 5-(a) shows the transitions due to its self-adaptation skills. Figure 5-(b) shows which heuristics is conducting the optimization process at a given iteration.

Figure 5. Evolution and performance of Lifecycle

7. Experimental Results 7.1. Experimental Setup The load identification procedure was performed using experimental data obtained from laboratory tests as performed on steel-made columns subjected to an instrumented test apparatus, which is capable of applying compressive loads. Such device is originally intended for the demonstration of the buckling phenomenon. Figure 6 shows the experimental setup in which, by choosing proper boundary conditions various scenarios can be studied. In this way, three different boundary conditions could be tested, namely, C-C, P-C, and P-P. For each value of the load, a set of frequency response functions (FRF) was obtained by processing the Fouriertransforms of the input (impact forces) and output (transverse accelerations).

5 6 1

2 3 4

(a) 1 Hammer (with a force transducer) 2 Accelerometer 3 Load cell

(b) 4 Force transducer signal conditioning 5 Accelerometer signal conditioning 6 Computer Figure 6. Scheme of the experimental test rig

7.2. Demonstration of the Stress-stiffening Effect In order to illustrate the stress-stiffening effect, the FRFs for a clamped-pinned column subjected to two different loads are presented in Fig. 7. It can be seen that as the compression load increases the natural frequencies decrease. In fact, the stress-stiffening generated by the external load leads to a smaller column bending stiffness.

Load

Figure 7. Experimental verification of the stress-stiffening effect. Table 3 provides the values of the first seven natural frequencies for both load conditions (see Fig. 7), as well as the estimator errors caused by experimental noise, as obtained from Eq. (8):

( f ) =
where:

2 1 xy ( f )

xy ( f ) 2nd

(8)

: is the normalized random error, 2 xy ( f ) : is the coherence function taken under resonance conditions, and
nd : is the number of averages considered.

Table 3. Values of the natural frequencies of the beam-like column (P-C)

Load [N] Natural Frequency Frequency [Hz] Estimator Error [%] Frequency [Hz] Estimator Error [%] 1 31.24 3.50 37.35 18.56 2 106.01 4.96 98.07 2.86 3 227.30 2.79 218.88 3.00 4 391.57 2.56 383.34 2.47 5 602.72 2.53 594.82 2.34 6 854.09 2.44 852.06 1.60 7 1164.70 2.47 1158.50 1.77

250 750

It should be noticed that the high value of the error obtained for the first natural frequency preclude any conclusion about its variation when the external load is increased. 7.3. Load Identification The identification procedure was applied by using the experimental data obtained from tests performed on the beam-like column to check whether residual stress, initial strain, possibly introduced by manufacturing or previous tests, could have some significant influence on the identification results. Table 4 shows the setup for LifeCycle used in experimental tests. Table 4. LifeCycle parameters used in experimental load identification.

LifeCycle

Number of Individuals 20

Number of Iterations 50

Stage Interval 5

A set of identification results for the different boundary conditions of the system is given in Tab. 5. It can be concluded that the algorithm satisfactorily identified the experimental loads. Table 5. Identification results for the beam-columns.
Boundary Force [N] Condition Experimental Identified 150 147.32 P-P 350 310.42 250 211.81 P-C 750 831.84 1000 1132.93 C-C 1500 1580.35 Error [%] -1.79 -11.31 -15.28 10.91 13.29 5.36

The behavior of the LifeCycle along the iterations can be observed in Fig. 8 for the identification of a 750 N load for the P-C boundary condition. As for the numerical simulation, the transitions due to its self-adaptation skills can be seen in Fig. 8-(a); Fig. 8(b) shows which heuristics is conducting the optimization process at a given iteration.

Figure 8. Evolution and performance of Lifecycle for the experimental tests. Table 6 summarizes the values of the experimental natural frequencies used in the optimization procedure and the values of the identified natural frequencies, corresponding to the identified values of the load parameter shown in Tab. 5. It can be seen that, in general the variation between natural and identified frequencies are very small. In some cases one of the natural frequency had to

be discarded in the identification process due to unacceptable frequency estimation error. Table 6. Natural frequencies and percentage errors for the loaded structure.
Boundary Force Condition (N) Mode Shape Frequencies 1 2 3 4 5 6 7 Objective Function J

150
P-P

350

250
P-C

750

1000
C-C

1500

Experimental Identified Error (%) Experimental Identified Error (%) Experimental Identified Error (%) Experimental Identified Error (%) Experimental Identified Error (%) Experimental Identified Error (%)

31.24 30.55 -2.21 34.54 36.78 6.50 25.31 27.22 7.55

85.91 85.86 -0.06 82.24 81.32 -1.11 106.01 108.12 2.00 98.07 98.13 0.06 115.42 122.24 5.91 106.96 112.72 5.38

197.91 197.37 -0.28 194.86 193.00 -0.96 227.30 230.61 1.46 218.88 220.60 0.78 239.89 255.14 6.36 232.42 245.64 5.68

351.86 353.79 0.55 351.21 349.48 -0.49 391.57 398.25 1.71 383.34 388.18 1.26 403.23 433.65 7.54 399.43 424.03 6.16

559.17 555.89 -0.59 555.27 551.61 -0.66 602.72 611.90 1.52 594.82 601.78 1.17 616.72 658.68 6.80 613.89 648.97 5.71

804.97 805.10 0.02 802.19 800.85 -0.17 854.09 873.17 2.23 852.06 863.02 1.29 885.99 931.96 5.19 884.93 922.18 4.21

1096.90 1103.80 0.63 1093.00 1099.60 0.61 1164.70 1184.50 1.70 1158.50 1174.40 1.38 1171.20 1255.90 7.23 1181.20 1246.00 5.49

0.02

0.04

0.13

0.55

0.34

0.30

8. Conclusions This paper presented an identification procedure to determine external forces applied to a beam-like column for various boundary conditions. The inverse problem solver is based on the LifeCycle method, a heuristics that considers different stages along the evolution, to mimic nature in the passage through the phases experienced by an individual along life. In the present contribution, GA and PSO represented these phases. Load identification was performed both for numerical FE data and real experimental data. The experimental data were obtained by using a buckling test apparatus that was adapted to perform vibration tests. The goal of the numerical investigation was to test the procedure for the case in which discrete and continuous design variables are considered simultaneously and to check efficiency for various load scenarios and boundary conditions. Finally, the experimental investigation illustrated the possibility of using the present technique in real engineering environment. The results are very encouraging in the sense that more complex inverse problem will be analyzed in further research. 9. Acknowledgments Mr. Viana and Mr. Rojas are thankful to CNPq and CAPES, respectively, for their PhD scholarship. Dr Rade and Dr Steffen, they both acknowledge CNPq for their research grant. 10. References

1. Greening P D and Lieven N A J. Modeling Dynamic Response of Stressed Structures. Proceedings of the 17th International Modal Analysis Conference, 1999, Florida, 103-108. 2. Lurie H. Lateral Vibrations as Related to Structural Stability. Journal of Applied Mechanics, 19, 1952, 195-204. 3. Rayleigh Lord. Theory of Sound. 2 (2nd edition), Dover, New York, 1877, 1945 re-issue. 4. Stephens B C. Natural Vibration Frequencies of Structural Members as an Indication of end Fixity and Magnitude of Stress. Journal of the Aeronautical Sciences, 1936, 4: 54-56. 5. Chu T H. Determination of Buckling Loads by Frequency Measurements. Thesis at the California Institute of Technology, 1949. 6. Wittrick W H. Rates of Changes of Eigenvalues, with Reference to Buckling and Vibrations Problems. Journal of the Royal Aeronautical Society, 1962, 66: 590-591. 7. Baruch M. Integral Equations for Nondestructive Determination of Buckling Loads fos Elastic Plate and Bars. Israel Journal of Technology, 1973, 11: 1-8. 8. Virgin L N and Plaut R H. Effect of Axial Load on Forced Vibrations of Beams. Journal of Sound and Vibration, 1993, 168 (3): 395-405. 9. Go C G and Liou C D. Load-response Determination for Imperfect Column Using Vibratory Data. Journal of Sound and Vibration, acepted 20 December 2002, in press. 10. Almeida S F M and Hansen J S. Enhanced Elastic Buckling Loads of Composite Plates with Tailored Thermal Residual Stresses. Journal of Applied Mechanics, 1997, 64 (4): 772-780. 11. Rojas J E. Characterization of Stress-stiffening Effect and Identification of Loads in Structures from Dynamic Responses (in portuguese). M. Sc. Dissertation, Federal University of Uberlndia, Uberlndia-MG, Brazil, 2004. 12. Rojas J E, Viana F A C, Rade D A and Steffen Jr V. Identification of External Loads in Mechanical Systems Through Heuristicbased Optimization Methods and Dynamic Responses. Latin American Journal of Solids and Structures, 2004, 1 (3): 297-318. 13. Livingston T, Bliveau J G and Huston D. R. Estimation of Axial Load in Prismatic Members Using Flexural Vibrations.

Submitted to Journal of Sound and Vibration, 1993. 14. Greening P D and Lieven N A J. Identification and Updating of Loading in Frameworks Using Dynamic Measurements. Journal of Sound and Vibration, 2003, 260 (1): 101-115. 15. Lieven N A J and Greening P D. Effect of Experimental Pre-stress and Residual Stress on Modal Behavior. Philosophical Transactions of Royal Society London, 2000, A 359: 97-11. 16. Krink T and Lvberg M. The LifeCycle Model: Combining Particle Swarm Optimisation. Genetic Algorithms and HillClimbers, Proceedings of the 7th International Conference on Parallel Problem Solving from Nature, Granada, Spain, 2002 September 7-11, Granada, 2002, 621-630. 17. Craig R R Jr. Structural dynamics: An Introduction to Computer Methods, Wiley-Interscience, New York , 1981, 544. 18. Michalewicz Z. Genetic Algorithms + Data Structures = Evolution Programs. 2nd ed., Springer-Verlag, New York, 1994. 19. Haupt R L and Haupt S E. Practical Genetic Algorithm. 1st ed, Wiley-Interscience, New York, 1998. 20. Kennedy J and Eberhart R C. Particle Swarm Optimization. Proceedings of the 1995 IEEE International Conference on Neural Networks, Perth, Australia, Perth, 1995, 1942-1948. 21. Venter G and Sobieszczanski-Sobieski J. Particle Swarm Optimization. Proceedings of the 43rd AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics and Materials Conference, Denver, CO - USA, 2002 April 22-25, Denver, 2002 (AIAA): 20021235. 22. Bendat J S and Piersol A G. Random Data Analysis and Measurement Procedures. (2nd edition). Wiley-Interscience, New York, 1986, 561.