Jtam.dvi JOURNAL OF THEORETICAL AND APPLIED MECHANICS 46, 2, pp. 413-434, Warsaw 2008 IDENTIFICATION AND SIMULATION OF INITIAL GEOMETRICAL IMPERFECTIONS OF STEEL CYLINDRICAL TANKS Jarosław Górski Tomasz Mikulski Gdańsk University of Technology, Poland e-mail: jgorski@pg.gda.pl Critical loadsof shell structures canbeproperlyapproximatedonly thro- ugh including randomness in their geometry.As it is difficult and expen- sive tomeasure the initial structure imperfections in situ or in laborato- ry, a methodology of identification and description of the available data should be provided. The presented procedure provides an opportunity for the reproduction of measured maps of steel cylindrical tank geo- metrical imperfections. Simulations of nonhomogeneous random fields of imperfections, based on the original conditional-rejection method of simulation, are applied. Using the measured data, an envelope of the imperfections is also estimated. It allows for simulation of extreme but still realistic fields of imperfections. Additionally, nonlinear numerical analyses of tanks with andwithout initial geometrical imperfections are performed. The results indicate that the initial imperfections influence the solutions. Key words: cylindrical tanks, geometrical imperfections, random fields, identification and simulation 1. Introduction Assessment of the reliability, safety and stability of structures with initial ma- terial and geometrical imperfections belong to the most complex problems in applied mechanics. The response of a large class of imperfection-sensitive structures (thin shells, thin-walled beams, arches and others) exhibit some features of chaotic systems. Many structural models deal mainly with elastic 414 J. Górski, T. Mikulski stability and include imperfections in a deterministic way, but alternativeme- thods are also implemented. The random nature of the structure behaviour has initiated the use of probabilistic methods (Augusti et al., 1984). Unfortu- nately, exact analytical solutions to stochastic problems exist only for simple models of structures and uncomplicated cases of loading. To obtain approxi- mate results, a great variety of perturbation techniques and stochastic finite element methods have been developed (see Surdet and Der Kiureghian, 2002; Matthies et al., 1997; Schuëller, 2001). Structure reliabilities are also estima- ted. For these purposes, the following techniques are in use: theMonte Carlo method, first and second order reliability methods (FORM and SORM), re- sponse surface techniques, artificial neural network methods and others (see Hurtado and Barbat, 1998; Marek et al., 1996; Melchers, 1999; Raizer, 2004). When nonlinear geometrical andmaterial effects are taken into consideration, the structure reliability can be evaluated only numerically (Anders and Hori, 1999). But the probabilistic analysis shouldbeused cautiously. For example, it has been demonstrated by Ferson and Ginzburg (1996) what spurious results could be obtained by a stochastic method applied inappropriately. It is well known that the effect of structural geometrical imperfections can dramatically decrease the nominal load carrying capacity of engineering struc- tures (Arbocz and Starnes, 2002; Khamlichi et al., 2004; Papadopoulos and Papadrakakis, 2004). An analysis of these imperfections can improve the de- sign process (Arbocz, 1998). On the basis ofmeasured and experimental data, the initial structure imperfections can be realistically described by randomva- riables or randomfields (Arbocz and Starnes, 2002). An important part of the structural imperfection modelled in a reliability context is the representation of random fields describing the statistical variation of properties or structure parameters. Various attempts have been made to find out ways of doing this (see: SurdetandDerKiureghian, 2002;Matthies et al., 1997;Vanmarcke, 1983; Vanmarcke et al., 1986; Li and Der Kiureghian, 1993; Zhang and Ellinwood, 1995; Mignolet and Spanos, 1992; Spanos andMignolet, 1992). Themeasured response of a structure, whenever available, should be used to update its reliability. For example, Papadimitriou et al. (2001) showed that the structural reliability computed before and after using additional dynami- cal data could differ significantly. The degree of structure sensitivity to the imperfections depends not only on their magnitude but also on their shape (Arbocz and Starnes, 2002). It is not sufficient to check the shell surface for the maximum imperfection amplitude by carrying out selected circumference or axialmeasurement scans.Acomplete surfacemapof themeasured structure should be provided. Identification and simulation of initial geometrical... 415 The developed methods of the structural data identification are related mainly to dynamical problems (Bendat and Piersol, 1993; Brandt, 1997), and their application to static analysis needs further investigation. Comprehensive numerical analysis of these problems can be found, for example, in Schenk and Schuëller (2003), where buckling behaviour of cylin- drical shells with random geometrical imperfections was examined. A concept of numerical prediction of a large scatter in the limit load observed in experi- ments using the directMonteCarlomethod, the simulation technique, and the nonlinear finite element method was introduced. The geometric imperfections were modelled as two-dimensional, the Gaussian stochastic process with pre- scribed secondmoment characteristics was based on a data bank of measured imperfections. The tests were aided by the Karhunen-Loéve expansion me- thod.Owing to this, it was possible to give the secondmoment characteristics of the limit load. Stefanou and Papadrakakis (2004) presented a stochastic triangular shell element in the case of a combined uncertain material (Youngs modulus and thePoisson ratio) and geometrical (thickness) properties. The properties were described by uncorrelated two-dimensional homogeneous fields. The spectral description of the random fields in conjunction with theMonte Carlo simula- tion was used for the computation variability. In Papadopoulos and Papadrakakis (2004) by means of the same metho- dology a parametric study was performed to evaluate the sensitivity of the buckling load of a cylindrical panel to the amplitude and shape of the initial imperfections. One- and two-dimensional stochastic imperfections of shapes, the modulus of elasticity and the shell thickness were introduced individually or in combination. The influence of the shape and magnitude of the imper- fections on the form and magnitude of the panel buckling load distributions was investigated. In the case of one-dimensional combined imperfections, a drastical reduction of about 50% of the mean value of the buckling load was observed,whereas in two-dimensional combined imperfections sucha reduction was not proved. TheMonte Carlo method combined with the finite element program ana- lysis was also employed by: Bielewicz et al. (1994), Walukiewicz et al. (1997), Bielewicz andGórski (2002) andGórski (2006). Structural initial imperfections were assumed to be random fields described in terms of a set of simulated re- alizations and the resulting statistic estimators. This lead to solutions of a set of deterministic problems for the assessment of the structure model re- sponses. The critical load histograms obtained numerically made it possible to estimate the structure reliability. Special attention was paid to discussion 416 J. Górski, T. Mikulski of reductionmethods concerning the number of theMonte Carlo realizations. Two methods were considered: the reduction of the initial sets of data, and the accuracy analysis of the output results. An assessment of the structure reliability interval was also proposed. With regard to compressed cylindrical shells, a reliability based knock- down factor was derived to describe the allowable load applied byArbocz and Starnes (2002). It can replace the empirical knockdown factor so chosen that whenmultiplied by the calculated perfect shell buckling load, a ”lower bound” relating to all existing experimental data is obtained. But despite some achie- vements in accurate prediction of the load carrying capacity of an imperfect shell, the problem is still an open question (Papadopoulos and Papadrakakis, 2004). In this paper, the measured imperfections obtained for steel cylindrical vertical tanks of 5000-50000m3 capacity (Orlik, 1976) are considered. An identification procedure and numerical simulation of the initial geometrical imperfections of tanks is proposed. For this purpose a comprehensive statistic analysis of themeasured imperfection is provided. To accomplish the numeri- cal simulation, a correlation functiondescribing the imperfection randomfields is introduced. In calculations, the conditional-rejection simulation method is used (see Bielewicz et al., 1994; Walukiewicz et al., 1997; Górski, 2006). As the field of initial imperfections is an example of circular data (Fisher, 1993) the simulation method is appropriately modified. The proposed simulation method makes it possible to implement the entire measured data in the cal- culation process. On this basis a formula which can be used in the prediction of any extreme imperfection in the tank is developed. Additionally, nonlinear numerical analyses of tanks with and without initial geometric imperfections are performed.The results indicate that the initial imperfections influence the solutions. 2. Stochastic description of tank geometric imperfections A detailed description of in a imperfections steel cylindrical vertical tank can be found inOrlik (1976).Nine sets ofmeasured in situ imperfectionsof tanks of V = 5000-50000m3 capacity were analysed (see Table 1, and Fig.1b-e). The basic statistic parameters of the initial imperfections, i.e. the mean values, standard deviations, and the maximal values are presented in Table 2. It is easy to notice that the discrepancies between the parameters are significant. Examining the shapes of the measured imperfection fields one can describe Identification and simulation of initial geometrical... 417 them as accidental or even chaotic. Only the first two tank imperfectionmaps are similar. Even in the case of tanks no. 3 and 4, the dimensions of which are the same as in the first two, the initial imperfections differ evidently. The differences result from tank dimensions, construction method used and precisionof the tankassembly.The last reason seems tobe themost important. Using the measured data, a procedure of the identification and simulation of the steel tank geometrical imperfectionmaps is proposed (see alsoGórski and Mikulski, 2005). Probabilistic methods are applied. Table 1.Cylindrical vertical tank data (Orlik, 1976) Dia- Tank assem- bly method Mesh dimen- sion [mm] Nos. of measu- reing point Nos. of Nos. of Tank V meter Height points points no. [m3] d [m] h [m] along along perim. height 1 5000 23.75 11.98 rolled strike 396×750 3232 202 17 2 5000 23.75 11.98 rolled strike 396×750 3232 202 17 3 5000 23.75 11.98 rolled strike 396×750 3030 202 15 4 5000 24.50 11.80 sheets, hand 4810×1500 128 16 9 welding 5 13000 36.58 13.45 sheets, hand 442×750 168 12 14 welding 6 30000 49.50 16.44 sheets, hand 370×1500 3780 420 9 welding 7 30000 49.50 16.44 sheets, hand 5981×750 572 26 22 welding 8 32000 54.10 14.40 sheets, hand 6060×800 448 28 16 welding 9 50000 65.00 18.01 sheets, 400×1500 5100 510 10 automatic 2.1. Identification of the measured geometrical imperfections The tank side surface imperfections can be considered in terms of a two- dimensional scalar random field described by a probability density function. In Wilde (1981) and Orlik (1976), the essential features of the measured im- perfection fields were considered. The following hypotheses were formulated and proved: • the stochastic process is stationaryandergodic along thehorizontal lines, • the randomvariables can bedescribed by aGaussian probability density function. 418 J. Górski, T. Mikulski Fig. 1. Measured and simulated geometrical imperfections of petrol tanks. Notes: tanks are in different scales; in tanks no. 6 and no. 9 only every second vertical line is drawn.Measured imperfections: (b) tank no. 1 – V =5000m3, d=23.75m, h=11.98m; (c) tank no. 3 – V =5000m3, d=23.75m, h=11.98m; (d) tank no. 6 – V =30000m3, d=49.50m, h=16.44m; (e) tank no. 9 – V =50000m3, d=65.00m, h=18.01m. Simulated imperfection – tank no. 1: (f) estimation of measured imperfections – sample no. 1; (g) simulated extreme imperfections – sample no. 234 Identification and simulation of initial geometrical... 419 Table 2. Statistical description of measured initial geometrical imperfections of tanks Tank d Mean value St. dev. |xmaxi | x/d σx/d |xmaxi |/dno. [m] x [m] σx [m] [m] 1 11.875 −0.763 ·10−2 0.0233 0.159 −0.643 ·10−3 0.00197 0.0134 2 11.875 −0.988 ·10−2 0.0240 0.163 −0.832 ·10−3 0.00202 0.0137 3 11.875 −0.485 ·10−2 0.0153 0.110 −0.408 ·10−3 0.00129 0.00926 4 12.250 −0.216 ·10−1 0.0213 0.075 −0.176 ·10−2 0.00173 0.00612 5 17.790 −0.994 ·10−3 0.0329 0.139 −0.559 ·10−4 0.00185 0.00781 6 24.750 −0.361 ·10−2 0.0302 0.167 −0.146 ·10−3 0.00122 0.00675 7 24.750 0.118 ·10−1 0.0252 0.097 0.476 ·10−3 0.00101 0.00392 8 27.005 0.297 ·10−2 0.0278 0.108 0.110 ·10−3 0.00103 0.00400 9 32.500 0.109 ·10−2 0.0491 0.226 0.334 ·10−3 0.00151 0.00696 Using the above assumption, the following nonhomogeneous correlation function is introduced K(y1,y2,z1,z2)=α z1z2 h2 cos(ω(y2−y1))exp(−β|y2−y1|−γ|z2−z1|) (2.1) where: y1, y2, z1, and z2 are the point coordinates (see Fig.1), α,ω,β, and γ are the correlation coefficients, and h denotes the tank height. The first term αz1z2h −2 determines a linear change of the standard deviation σx = √ αz/h along the vertical line with zero value at the tank bottom and maximum at the top. The element cos(ω(y2 − y1)) assumes that the random variables permute along the perimeter according to the cosine function. The last term, exp(−β|y2−y1|−γ|z2− z1|), describes how fast the correlation between the neighbouring points vanishes in the horizontal and vertical directions. The correlation function parameters α, ω, β, and γ are estimated on the basis of the measured data. Only the imperfection fields of tanks nos. 1-3 (V = 5000m3), no. 6 (V = 30000m3), and no. 9 (V = 50000m3) were me- asured using fine meshes (see Table 1). Thus, only five data sets were used to determine the constants of Eq. (2.1). The assumption that the random field of imperfections is ergodic along the horizontal lines makes it possible to analyse not a single (the measured) but hundreds of realizations. For exam- ple, in the first case (tank no. 1), as the imperfection values are given in 202 vertical lines, the same number of field realizations can be considered (NR=202). For the reason that the imperfectionsweremeasured at 16 points along thevertical line, thedimensionof the randomvector xi equals 16×202= 3232. The global experimental covariance matrix Ke (size 3232×3232) of the 420 J. Górski, T. Mikulski measured imperfection fieldwas obtained according to the following statistical formulas Ke = 1 NR−1 NR∑ i=1 (xi−x)(xi−x)⊤ (2.2) x= 1 NR NR∑ i=1 xi where xi (i = 1, . . . ,NR) is the measured imperfection vector, NR is the number of realizations, and x represents the mean value vector. Making use of the calculated matrix Ke (Eqs. (2.2)), the parameters of correlation function (2.1) are determined by a standard regression analysis. As any regression calculations concerning the global experimental covariance matrix size 3232×3232 are inefficient, only imperfectionsmeasured on limited areas of tank surfaces have been considered. In the cases of tanks nos. 1, 2 and 3, the random vectors xi of dimensions 16× 16 has been analysed. Similar reduction has been set up in the case of tanks nos. 6 and 9 (28×9 and 26× 10, respectively). In the calculation, all horizontal lines are taken into considerations and only a limited number of the vertical lines. Themagnitude of the reduced areas of the calculation are connectedwith the significant range of the correlation defined by (2.1). This simplification has reduced the time of regression analysis. The results – parameters of correlation function (2.1) – are presented in Table 3, in which the global Ger and variance Ver errors of the calculations are also presented. They are obtained using the following formulas Ger(Ke,K̂e)= ∣∣∣‖Ke‖−‖K̂e‖ ∣∣∣ ‖Ke‖ ·100% (2.3) Ver(k e ii, k̂ e ii)= 1 MN MN∑ i=1 keii− k̂eii keii ·100% where ‖K‖ = √ trK2 is the matrix norm, kii denotes the diagonal element of thematrix, and MN describes the covariance matrix size. The elements of the covariance matrix estimator K̂e were obtained by inserting the calculated correlation parameters in (2.1). The graphical representation of the selected experimental covariance matrix values Ke and their estimator K̂e, calculated with the help of the vector xi defined on a 16×16 sizemesh, is given inFig.2. Identification and simulation of initial geometrical... 421 Table 3.Correlation parameters of covariance function (2.1) Tank Random vector Correlation parameters of Eq. (2.1) Errors no. xi dimension α ω β γ Ger Ver 1 16×16 0.000716 0.317389 0.175900 0.058001 4.49 26.84 2 16×16 0.000707 0.329261 0.163659 0.050385 3.04 29.00 3 16×16 0.000533 0.421175 0.113103 0.098789 8.25 15.75 6 28×9 0.001397 0.154658 0.061831 −0.00861 1.84 22.21 9 26×10 0.005234 0.134896 0.040879 0.001564 0.78 6.40 Fig. 2. Graphical representation of identified correlationmatrix of imperfection field (xi size equals 16×16); y=6.0m (constant, half the shell height), 0¬x¬ 37.3m (half the shell perimeter) Theerror values and theplots show that the scattered pattern of imperfections was modelled accurately. Analysing the results presented in Table 3, one can notice that only the first two sets of parameters describing the imperfections for tank no. 1 and no. 2 are similar. The results confirm the conclusions formulated above (see Table2). It is obvious that thecharacteristics of geometrical imperfectionsvary depending on different tank dimensions. Also the difference between the tanks of the same capacity (tanks no. 1, no. 2 and no. 3) can be easily predicted. Comparing Fig.1a,b, it is possible to notice that tank no. 3 was built much more precisely. Similar calculations of the correlation coefficients can bemade in the case of the limited measured data, i.e. for tanks nos. 4, 5, 7, and 8. For example, using the data of tank no. 4 (see Table 1), the following parameters have been obtained α=0.013612 ω=0.299437 β=0.062983 γ =−0.009868 (2.4) 422 J. Górski, T. Mikulski with the associated global and local errors: Ger = 16.51%, Ver = 4.48%. In this case, the calculations are performed with the help of all the available data (mesh 9× 15, 128 measured points). Despite the fact that tank no. 4 (V =5000m3) is almost identical to tanks 1-3, the results differ significantly (compare parameters (2.4) and those presented in Table 3). The differences result mainly from the limited data used in the calculations, i.e. 15 vertical lines in tank no. 4, and 202 when tank no. 1 is considered. The differences in the initial imperfection shapes and magnitudes (see Table 2) seem to have minor influence. To indicate the effect of the number of measured imperfections on the correlation parameters, an additional simple analysis was performed.Theme- asured data of tank no. 1 were reduced in such away that their size was close to the data of tank no. 4, i.e. 8×17=136 points. On that basis, the following correlation parameters were estimated α=0.009893 ω=0.335382 β=0.024198 γ =−0.013186 (2.5) As expected, the results differ significantly from those obtained for the 16×202 measured points of tank no. 1 (see Table 3). Thus, the imperfection data for tanks 4, 7, 8, and 9 seem to be insufficient to describe the random fields properly. The method of identification of the imperfection fields on the basis of the limited data remained an open problem. The analysis makes it possible to formulate the following conclusions: • the number of the measuring points of imperfections should depend on the quality of the tank assembly, • the data necessary for the acceptable definition of the random field of imperfections shouldbemeasuredprecisely, usingafinemesh, evenwhen only a limited part of the tank (for example one quarter) is taken into consideration; measuring the entire tank side by a rough mesh gives inadequate results. It is an important problem if a general rule can be established to describe the relation between the tank dimensions and the correlation coefficients. To this end, the parameter values (Table 3) are presented versus the tank dia- meters (see Fig. 3). Analysing the graph and taking into consideration the statistic data presented in Table 2, one can notice that the available data are too scattered to build a reliable general expression describing the randomfield of the tank imperfections.Only a draft formula can be estimated. For this pur- pose, the standard regression analysis is applied. The first coefficient α (Eq. Identification and simulation of initial geometrical... 423 Fig. 3. Correlation coefficients (Eq. (2.1)) vs. tank diameter (see Table 4) Table 4.Extreme values of geometrical imperfections Tank Extreme values Tank edge no. z/h xextr/σx xextr/σx 1 0.0625 4.5511 3.6533 2 0.1250 4.5764 3.5357 3 0.7333 3.8638 3.3294 4 0.3750 2.3853 1.6780 5 0.8571 2.5232 2.3843 6 1.0000 3.6732 3.6732 7 1.0000 2.6984 2.6984 8 0.0625 3.3931 2.4055 9 0.1000 4.7730 2.8431 Themeans: 3.6042 2.9112 (2.1)) is describedwith thehelp of an exponential function, and the coefficients ω, β and γ are defined by straight lines (see Fig. 3) α=0.006733exp(0.091963d) ω=−0.005894d+0.491321 β=−0.002876d+0.218356 γ=−0.001931d+0.112063 (2.6) where d is the tank diameter. Formulas (2.6) can be improved when new data are included in the cal- culations (Kowalski, 2004). Further improvements can be obtained describing the correlation coefficients as functions of standard deviations of imperfections and their extreme values. 424 J. Górski, T. Mikulski 2.2. Envelope of tank geometrical imperfections It is possible to describe the correlation function as an envelope of the extreme imperfections. For this purpose, the standard deviations of tank im- perfections for each horizontal level are calculated. The obtained normalized standard deviations for all tank data are plotted in Fig. 4. Examining these data, it is easy to notice that the tank side imperfections start from the bot- tom, and their values indicate the same order as the imperfections measured at higher levels. Therefore, the simplest solution is the approximation of the standard deviation to a parabolic function σ= √ αzh (2.7) Then, the correlation function of the imperfection field takes the form K(y1,y2,z1,z2)=α √ z1z2 h cos(ω(y2−y1))exp(−β|y2−y1|−γ|z2−z1|) (2.8) Fig. 4. Normalized standard deviations of geometric imperfections calculated separately for each level The coefficient α is evaluated in such a way that the standard deviation value at the top of the tank is increased by 10%. For example, in the case of tank no. 1 √ α= 0.0364. The other correlation parameters ω, β, and γ (see Eq. (2.8)) remain unchanged (Table 3). The normalized envelope for all tank data is presented in Fig. 4. The results have proved that the adopted envelope in the parabolic shape describes precisely possible extreme imperfections. Proposed correlation functions (2.1) and (2.8), and the estimated values of theparameters α,ω,β, and γmake it possible to simulate vertical geometrical imperfections of the tank. Identification and simulation of initial geometrical... 425 3. Simulation of geometrical imperfections The simulationprocess is presentedusing thedata of tankno. 1 (Table 1).Two cases are analysed, i.e. simulation of the measured imperfections as precisely as possible, and simulation of the maximal imperfection values. The field of imperfections is numerically simulated taking advantage of correlation function (2.1) and the estimated constants presented inTable 3. In the process, the conditional-rejection simulationmethod is used (seeBielewicz et al., 1994; Walukiewicz et al., 1997; Górski, 2006). This technique has been successfully applied to spatial descriptions of soil random properties (Prze- włócki and Górski, 2001; Tejchman and Górski, 2006) and to environmental problems (Jankowski and Walukiewicz, 1997). Here, only a brief description of this method is presented. 3.1. Conditional-rejection simulation method Adiscrete random field can be described bymultidimensional random va- riables defined at mesh nodes. The field is represented by the random vector x(m×1) and itsmean value x(m×1). The correlation function is replaced by the symmetric and positively defined covariance matrix K(m×m). The randomva- riable vector x(m×1) is divided into blocks consisting of the unknown xu(n×1) and the known xk(p×1) elements (n+p=m). The covariancematrix K(m×m) and the expected values vector x(m×1) are also appropriately cut x= { xu xk } (n+p) K= [ K11 K12 K21 K22 ] (n+p)×(n+p) x= { xu xk } (n+p) (3.1) The unknown vector xu is estimated from the following conditional trun- cated distribution (Jankowski andWalukiewicz, 1997) ft ( xu xk ) = 1 [(1− t)2π]mdetKc exp ( − 1 2(1− t) (xu−xc)⊤K−1c (xu−xc) ) (3.2) where Kc and xc are described as the conditional covariance matrix and con- ditional expected value vector Kc =K11−K12K−122K21 (3.3) xc =xu+K12K −1 22 (xk−xk) 426 J. Górski, T. Mikulski The constant t is the truncation parameter t= sexp(−s2/2)√ 2πerf (s) (3.4) The constant s (Eq. (3.4)) is usually related to the standard deviation of the random field points. The random vector of the unknown variables xu is simulated using the rejection method (Devroye, 1986). The single random points of the vector xu are generated according to the following formula xi = ai+(bi−ai)ui i=1, . . . ,m (3.5) where ui denotes the random variables uniformly distributed in the interval [0,1], and (ai,bi), i= 1,2, . . . ,m are the intervals of the reals. The value xi is generated according to (3.5) to fulfil condition (3.2) (the trial and error method). The intervals ai and bi are defined for all points of the mesh and establish an envelope of the randomfield. The envelope specifies the characte- ristic features of the field under consideration, for example, the maximal and minimal values and the field boundary conditions. It can be given in the form of a function or by experimental discrete data. Fig. 5. Simulation process – coverage of field points with moving propagation scheme; ◦ – known values, • – simulated value, × – unknown values simulated in one base scheme step, 2 – known values not included in base scheme calculations The direct application of the rejection method is inefficient. To improve it, a base scheme of random values is defined (Bielewicz et al., 1994). The scheme is placed at the nodal points of the mesh in such a way that it covers all the nodes i, 1 ¬ i ¬ m, m = M ×N (Fig.5). The simulation process is divided into three stages. First, four-corner random values are simulated. Next, the propagation scheme with a growing number of points covers the Identification and simulation of initial geometrical... 427 defined base scheme of the field mesh. In the example shown in Fig.5a the dimension of the base scheme equals c×d=5×5=20 random values (mesh points). In the third stage, the base scheme is appropriately shifted, and the next group of unknown random values is simulated (see Fig.5b). The base scheme is translated so as to cover all the field nodes. As the simulated field of tank imperfections is a formof circular data, the algorithmhasbeen improved. When the calculations along the circle (the horizontal lines) are being closed, the base scheme is modified accordingly. The base scheme and the random field envelope are the characteristic fe- atures of the conditional-rejection simulationmethod.Theymake it possible to simulate the fields describing any two-dimensional shapes (three dimensional problems can also be analysed). The size of the field is practically unlimited (a thousand of mesh points). 3.2. Simulation of measured and extreme tank imperfections According to the method of simulation, the random field envelope – i.e. the lower ai and the upper boundary bi (see Eq. (3.5)) – have to be defined for each mesh point i. The parameters determine the theoretical maximal and minimal values of imperfections and should be related to their standard deviations. Since the imperfection field is assumed to be ergodic, the same boundary values are assigned to each horizontal line. To estimate them, the imperfectionmeanvalues andthe standarddeviations calculated separately for all horizontal lines are analysed oncemore. The extreme imperfections related to their standarddeviations are presented inTable 4.They appear on different tank levels (z/h in Table 4). Only in two cases the extreme values occurred on the tank edges (tank no. 6 and 7). The maximal value xextr/σx = 4.7730 appeared in tank no. 9 (V = 50000m3). Analysing the data, the following values describing the random field boundary have been assumed ai,bi =±sσi =±5σi (3.6) where σi is the standard deviation at the point i (see Eq. (2.1)) σ2i =K(yi,yi,zi,zi)=α z2i h2 (3.7) Thenext simulation parameter, i.e. the simulation base scheme (seeFig.5) is definedas a rectangle of dimension 16×16points.Thisbase scheme ismoved horizontally step by step to cover all mesh points. The procedure allows for simulating the tank imperfections, i.e. along the cylindrical surface. As many 428 J. Górski, T. Mikulski as 2000 realizations have been simulated. Using formulae (2.3), the following errors in the field simulation have been found Ger =3.30% Ver =3.48% (3.8) Values (3.8) indicate excellent convergence of the field estimators. Additio- nally, samples no. 1 of the simulated field of imperfections is presented in Fig. 1f. It can be noticed that the essential features of the measured imperfection map have been numerically reproduced. The same calculation can be performed in order to obtain the extreme geometrical imperfection field. For this purpose, Eq. (2.8) with the following correlation coefficients √ α=0.0364 ω=0.317389 β=0.175900 γ=0.058001 (3.9) is implemented. The same boundary values (3.6) are used. After simulating 2000 realiza- tions, the following errors occurred Ger =0.15% Ver =0.41% (3.10) Asampleof the simulated initial imperfectionvector is presented inFig.1g. Themean value of the vector characterizes the extreme value chosen from all 2000 samples in the set. Thus, the sample is an extreme one (in the defined sense). 4. Numerical calculation of vertical petrol tank The preliminary numerical calculations include three cases of tank no. 1 (5000m3). Thefirst case refers to an ideal cylindrical shellwhosedata arepresented in Fig.1a (Orlik, 1976). The tank ismade of 6 strakes of plates, each ca. 2000mm inheight, andof thicknesses t=11, 10, 8, and7mm, starting fromthebottom. Steel St3VY is used for the lower side plates and the first row of the bottom plates, andsteel St3SYfor otherplates.The steel yield stresshasbeenassumed to be R = 255MPa. The petrol level, of specific gravity ρ = 50MN/m3, is raised in every incremental step (itsmaximal value hp =11.25m).Other data of the tank, for example, the soil (foundation) parameters k = 50MN/m3, Identification and simulation of initial geometrical... 429 are taken from Ziółko (1986). 15646 finite elements were used for the tank model mesh (Górski and Mikulski, 2005). Some of the results, obtained with the help ofMSCNASTRAN(MSCNastran forWindows, 2001), are presented in Table 5. Table 5. Extreme stresses, bending moments and axial forces in tanks with and without initial geometrical imperfections Tank surface Internal forces Ideal Measured Simulated imperfections imperfections Maximal stresses σR [MPa] 92.24 255 255 Bendingmomentsmz [kNm/m] −0.67-0.98 −2.51-4.19 −1.38-3.76 Bendingmomentsmϕ [kNm/m] −0.20-0.29 −4.11-2.68 −3.42-2.40 Axial force nz [kN/m] −7.97-20.56 −516.1-335.7 −712.7-448.2 Axial force nϕ [kN/m] −1.23-1025.0 −433.6-1319.0 −945.2-1227.0 The tank data for the second case include measured initial geometrical imperfections presented in Fig.1b (Orlik, 1976). As the imperfections refer only to 16 horizontal and 202 vertical lines (3232 points), the missing mesh values are determinedby linear interpolation using theneighbouringdata.The basis of the numerical calculation is the deformed shape of the cylindrical tank assumed to be stress free. The results are presented in Table 5. The third calculation is performed for the simulated extreme imperfections shown in Fig.1g. The results of nonlinear calculations (see Table 5) indicate that the tank initial imperfections can cause significant variations in stress fields in compa- rison with the solution related to an ideal surface. The steel of the tank with imperfections yields at a point connecting the bottomwith the side plates and in areas of extreme imperfections. It should be noted that the yielding process occurrs despite the fact that the initial field of imperfections is rather a typical one (see Fig.1b). 5. Conclusions Onthebasis of thepresented analysis, the following comments and suggestions having reference to future improvements of the identification andmodelling of the random fields of imperfections in tanks can bemade: • A method of measuring geometrical imperfections in tanks should be formulated. It is essential to define the minimal tank area and the size 430 J. Górski, T. Mikulski of themesh of measuring points. It has been shown that the parameters depend on the tank diameter as well as characteristics and magnitude of the imperfections. • A generalized version of the correlation function should be formulated for steel cylindrical tanks of various dimensions. The presented primary relations can help in future improvementswhenmoremeasureddata can be incorporated in the estimation. • The procedure of identification of the correlation function from limited data measured in situ needs further investigation. • The achieved results, aided by the statistical analysis, have proved the capability of the conditional-rejection method to precisely simulate geo- metrical imperfections in tanks when the correlation function is known. • The randomnumericalmodel of tanks can be easily extended, for exam- ple, by introducing randomvariability of soil foundationswhich canhave a degrading effect on the tank loading capacity. • The presented results of numerical calculations have revealed that the initial geometrical imperfections have influence on themechanical beha- viour of steel cylindrical tanks. Formulation of amethodology of identifi- cation, classification and description of the tank imperfections can lower the laborious andhigh cost of experiments and ensure a better andmuch safer design. Designing a new tank, the correlation function and the appropriate stan- dard deviations of the geometric imperfections can be described according to themeasureddataof existing tanksof similardimensions.Then, the simulation procedure can be applied and, from the simulated set, some extreme realiza- tions can be selected. The selection criteria of characteristic samples are, for example: the extreme value of imperfections, norms of the imperfection vec- tors, the number of sign changes (number of waves) along the horizontal lines. On the basis of the selected extreme samples, it is possible to calculate stress fields of the structure. In this way, the most dangerous stress state related to all simulated but realistic data can be obtained. A definition of the critical state of a steel petrol tank should also be established. Usually, plastic defor- mation of the tank side is considereddangerous as steel becomes brittle.Cyclic loading conditions of the tank also play an important role in the definition of the critical state. Identification and simulation of initial geometrical... 431 References 1. AndersM.,HoriM., 1999,Stochasticfinite elementmethod for elasto-plastic body, International Journal for Numerical Methods in Engineering, 46, 1897- 1916 2. Arbocz J., 1998,On accuracy of numerical buckling load predictions,Proc. of the 6thConference ”Shell Structures.Theory andApplicatios”, Gdańsk-Poland, 19-23 3. Arbocz J., Starnes J.H., 2002, Future directions and challenges in shell stability analysis,Thin-Walled Structures, 40, 729-754 4. AugustiG.,BarattaA.,Casciati F., 1984,ProbabilisticMethods in Struc- tural Engineering, London, NewYork: Chapman andHall 5. Bendat J.S., Piersol A.G., 1993, Engineering Applications of Correlation and Spectral Analysis, New York: JohnWiley & Sons 6. Bielewicz E., Górski J., 2002, Shell with random geometric imperfections. Simulation-basedapproach, International Journal ofNon-linearMechanics,37, 4/5, 777-784 7. Bielewicz E., Górski J., Schmidt R., Walukiewicz H., 1994, Random fields in the limit analysis of elastic-plastic shell structures, Computers and Structures, 51, 3, 267-275 8. Brandt S., 1997, Statistical and Computational Methods in Data Analysis, New York: Springer Verlag 9. Devroye L., 1986, Non-Uniform Random Variate Generation, New York: Springer-Verlag 10. Ferson S., GinzburgL.R., 1996,Differentmethods are needed to propagate ignorance and variability, Reliability Engineering and System Safety, 54, 133- 144 11. Fisher N.I., 1993, Statistical Analysis of Circular Data, Great Britain, Cam- bridge University Press 12. Górski, J., 2006, Non-linearmodels of structures with random geometric and material imperfections simulation-based approach,Gdansk University of Tech- nology, Monograph, 68 13. Górski J., Mikulski T., 2005, Statistical description and numerical calcu- lations of cylindrical vertical tanks with initial geometric imperfections, Proc. of the 7th International Conference Shell Structures. Theory and Applications, Gdańsk Jurata 2005; London: Balkema, Taylor & Francis, 547-551 14. HurtadoJ.E.,BarbatA.H., 1998,MonteCarlo techniques in computational stochastic mechanics,Archives of Computational Method in Engineering, 5, 1, 3-30 432 J. Górski, T. Mikulski 15. JankowskiR.,WalukiewiczH., 1997,Modeling of two-dimensional random fields,Probabilistic Engineering Mechanics, 12, 2, 115-121 16. Khamlichi A., Bezzazi M., Limam A., 2004, Buckling of elastic cylindri- cal shells considering the effect of localized axisymmetric imperfections, Thin- Walled Structures, 42, 1035-1047 17. Kowalski D., 2004, Influence of initial imperfections on the vertical tank side stresses, Ph.D. Thesis. Gdańsk University of Technology, Gdańsk [in Polish] 18. Li C.C., Der Kiureghian A., 1993,Optimal discretization of randomfields, Journal of Engineering Mechanics, 119, 6, 1136-1154 19. Marek P., Guštar M., Anagnos T., 1996, Simulation-Based Reliability Assessment for Structural Engineers, New York, London, Tokyo: CRC Press, Boca Raton 20. MatthiesH.G.,BrennerC.E., BucherC.G., SoaresC.G., 1997,Uncer- tainties in probabilistic numerical analysis of structures and solids – stochastic finite elements, Structural Safety, 19, 3, 283-336 21. Melchers R.E., 1999, Structural Reliability Analysis and Prediction, Chiche- ster: JohnWiley & Sons 22. Mignolet M.P., Spanos P.D., 1992, Simulation of homogeneous two- dimensional randomfields: Part I –ARandARMAmodels, Journal of Applied Mechanics, ASME, 59, 2, 260-269 23. MSCNastran forWindows, 2001, Los Angeles: MSC Software Corporation 24. Orlik G., 1976, Deformation of shapes of cylindrical steel shells, statisti- cal analysis and numerical simulations, Ph.D. Thesis. Technical University of Gdańsk, Gdańsk [in Polish] 25. Papadimitriou C., Beck J.L., Katafygiotis L.S., 2001, Updating robust reliability using structural test data, Probabilistic Engineering Mechanics, 16, 103-113 26. Papadopoulos V., Papadrakakis M., 2004, Finite-element analysis of cy- lindrical panels with random initial imperfections, Journal of Engineering Me- chanics, 130, 8, 867-876 27. Przewłócki J., Górski J., 2001, Strip foundation on 2-D and 3-D random subsoil,Probabilistic Engineering Mechanics, 16, 2, 121-136 28. RaizerV., 2004,Theory of reliability in structural design,AppliedMechanical Review, 57, 1, 1-21 29. Schenk C.A., Schuëller G.I., 2003, Buckling analysis of cylindrical shells with randomgeometric imperfection, International Journal of Non-LinearMe- chanics, 38, 1119-1132 Identification and simulation of initial geometrical... 433 30. Schuëller G., 2001, Computational stochastic mechanics – recent advances, Computers and Structures, 79, 2225-2234 31. Spanos P.D., Mignolet M.P., 1992, Simulation of homogeneous two- dimensional randomfields:Part II –MAandARMAmodels,Journal ofApplied Mechanics, ASME, 59, 2, 270-277 32. Stefanou G., Papadrakakis M., 2004, Stochastic finite element analysis of shells with combined random material and geometric properties, Computer Methods in Applied Mechanics and Engineering, 193, 139-160 33. Surdet B., Der Kiureghian A., 2002, Comparison of finite element reliabi- lity methods,Probabilistic Engineering Mechanics, 17, 337-348 34. Tejchman J., Górski J., 2006, Stochastic FE-analysis of shear localization in 2Dgranularmaterialwithin amicro-polar hypoplasticity,Archives of Hydro- Engineering and Environmental Mechanics, 53, 4, 1231-3726 35. Vanmarcke E., 1983, Random Fields: Analysis and Synthesis, Cambridge: MIT Press 36. Vanmarcke E., ShinozukaM., Nakagiri S., SchuëllerG.I., Grigoriu M., 1986, Random fields and stochastic finite elements, Structural Safety, 3, 143-166 37. Walukiewicz H., Bielewicz E., Górski, J., 1997, Simulation of nonhomo- geneous random fields for structural applications, Computers and Structures, 64, 1/4, 491-498 38. Wilde P., 1981, Random Fields Discretization in Engineering Calculations, Warsaw: PWN, [in Polish] 39. Zhang J., Ellinwood B., 1995, Error measure for reliability studies using reducedvariable set,Journal of EngineeringMechanics, ASCE,121, 8, 935-937 40. Ziółko J., 1986, Metal Tanks for Liquids and Gases, Warsaw: Arkady [in Polish] Identyfikacja i symulacja wstępnych imperfekcji geometrycznych w stalowych cylindrycznych zbiornikach Streszczenie Wyznaczenie obciążenia granicznego konstrukcji powłokowychwymaga uwzględ- nienia w obliczeniach odchyłek geometrycznych.Wykonanie pomiarów tego typu od- chyłek rzeczywistychobiektów jest trudne i kosztowne.Wyniki uzyskanenapodstawie analizy modeli laboratoryjnych nie są wiarygodne.W pracy zaproponowanometodę alternatywną. Wykazano, że na podstawie dostępnych danych można sformułować 434 J. Górski, T. Mikulski metody identyfikacji i symulacji wstępnych odchyłek geometrycznych. Zaprezento- wane procedury umożliwiają odwzorowanie pomierzonych pól wstępnych odchyłek stalowych cylindrycznych zbiornikówo osi pionowej na paliwa płynne. Symulacje nie- jednorodnych pól losowych imperfekcji wykonano za pomocą oryginalnej warunkowej metody akceptacji i odrzucania. Pomierzone dane pozwoliły także na wyznaczenie obwiedni odchyłek losowych i na tej podstawie wykonano symulację ekstremalnych, realistycznych pól imperfekcji. Dodatkowo przeprowadzono nieliniowe obliczenia nu- meryczne zbiorników idealnych oraz z imperfekcjami. Porównaniewynikówpozwoliło na określenie wpływu odchyłek na mechaniczną odpowiedź powłoki. Manuscript received May 21, 2007; accepted for print November 14, 2007