MODEL UPDATING Endless Possibilities Peter Avitabile Modal Analysis and Controls Laboratory University of Massachusetts Lowell Lowell, Massachusetts USA ABSTRACT Finite element models are updated using test data to better represen the actual "as-built" characteristics ofa system or prototype. Using commercially available software, a detailed finite element model can be easily updated to reflect the measured characteristis - whether they be frequencies, mode shapes or response functions. However, the accuracy of the updated model is heavily dependent on the location and type of| ‘iserepancy tha exists and is strongly influenced by the analysts understanding ofthe model and selection of parameters that ae used forthe updating process. ‘This paper uses simulated test data to show the sensitivity ofthe updated model to parameters selected in the updating process ‘The models presented are not intended to creditor discredit the approach but rather to provide avery unbiased view ofthe capabilities ofthe approach with its strengths and limitations. Modes ar itentionally developed that will span the full range of acceptable updated models to unacceptable ones. Different ‘measured parameters are used (Frequency, mode shape and response function differences) to perform the updating procedure to illustrate the improvement (and degradation) ofthe finite clement model depending on the analysts proper (or impropes) selection of parameters for change. INTRODUCTION ‘Many analysts desire to obtain more accurate models of structures, for siractural dynamic applications. However, the generation of large finite element model isa very complicated process. While the analyst may be reluctant to admit thatthe models developed may contain inaccuracies, the reality is that these models are approximations of what the analyst believes the system to be. Generally, for important applications, prototypes are developed to confirm the adequacy of the model developed forthe application at hand ‘These large finite element models may yield firly good models that at times correlate reasonably wel for some sysiem modes but lack correlation with other system modes. If these modes are of 562 concer, then the analyst may wish to update or adjust the model using the measured modal data. Of course, we would all wish that ‘our models correlate perfectly to the measured data from the prototype. However, the reality is that our models are approximations that need to be adjusted, refined and tweaked to provide a bette representation af the actual system. Today, several commercial software codes exist which allow the updating ofa finite element mode! using measured data. Most of the available software uses a sensitivity approach forthe adjustment ofthe analytical model. While these sofware packages have brought forth the possiblity of updating large complicated finite element madels, there is general lack of, understanding ofthe implications of using these techniques by the general engineering population, ‘At the firs introduction of mode! updating software packages in the early 1990s, typically only frequency differences were. ‘addressed in the updating procedures (packages such as SYSTUNELI], LINKI2], COORDS|3} existed). While measured mode stapes could be included inthe updating process, mode shape differences were not considered as accurate as frequency differences and much more dificult to address, and generally not included mainly due to inaccuracies inthe expansion processes. Generally, models could be improved fairly well but researchers moved forward with techniques that addressed response function differences which was considered more realistic than just the frequency difference consideration (packages such as FEMtools{4], LINK{S] and SDRC FRFCORR(6) were developed). ‘While the measured response function (frequency response function) difference isan appropriate mechanism forthe updating, process, its often perceived as "more accurate” than frequency ifferences or mode shape differences. The response function is actually not any more accurate than other parameters, itis just perceived that way since i is a measured function (and how could it possibly be wrong!) Tn actuality, there are many reasons why the response difference also contains some possible areas for ‘concer when performing modeling updating, One key concem is the estimate of damping used for generating the FEM response functions; this may itself eause appreciable differences between the analytical and experimental response functions. In any event, the use of frequency, mode shape and response differences, llprovide mechanisms for the adjustment ofthe analytical model 10 better reflect the actual measured characteristics, The procedures incorporated in available software allow the analy to very easily ‘update his analytical mode (for a procedural standpoint) Of course, re afefne points of the different mode! updating techniques that are argued and presented in echnical papers discussing the merits and pitfalls of the different approaches However, there are several areas that ae citcal othe success of the improvement process that are often overlooked, These can be broken down as follows; ' There can be idealization errors inthe model assumplions. ‘These errors are associated with how joints are modeled, what boundary conditions were used in the model, ee, Incorrect ‘modeling assumptions here are sometimes difficult to address. For instance, a joint may be modcled as rigid whieh ‘ean bea good approximation for lower order modes but may be totally incorrect for higher order modes, Tere can also be errs associated with discrtiation in the ‘generation of the finite element model. These errors are ‘generally associated with mesh coarseness and lumped mass approximations, However, erors suchas element type ean also have effect in this category Te last error associated with model updating is related to the parameters selected for ihe updating process, he selection of elements and their characteristics whieh are allowed to change ean be one of the most rial parts ofthe updating process. Ifthe wrong parameters are selected for change then| ‘an incorrect description ofthe updated model may eos All ofthe types of errors above are important the development ‘of an improved model. For this paper, only errors associated with parameter selection are investigated, The other Wo areas are {equally important inthe generation of a good updated model for structural dynamic studies but are not addressed herein, ‘Commercially avilable software allows the fst category sted to bbe addressed. The tools available in all the sofware packages allows the analyst to easily update an existing finite clement model to minimize the difference benveen the analytical and test models ‘through the selection of parameters that ae alowed to change ring the adjustment process. However, the selection ofthese Parameters isthe eritcal part oF using dhe model updating software “The updating procedures are based on the sensitivity ofthe model for change. Unfortunatly, the acta errors that exist inthe mode! _may not be inthe most sensitive egions ofthe model. Therefore, using these sensitivity approaches to update the model must be one so with the urmost af eae, If not then updated ‘characteristics may resul which may not tuly eeflect the proper dynamics of the system, (One basic understanding ofthe sensitivity approach used is that ‘the mass and stiffness sensitive regions ofthe structure forthe sets ‘of modes identified for the updating process wil always doninate {the solution if they are ellowed to participate in the updating process. This is true whether or not the diseepancy exists in that region ofthe structure. Ifthe diserenancy does exist ip that sensitive region ofthe structure, then the updating process will have the ability t identify this region for updating. However, if the discrepancy does not exist in the mass o stiffness sensitive areas ofthe model and these sensitive areas are allowed 10 Participate in the updating process, then the solution will naturally tend to update the most sensitive regions of the model even though ‘he discrepancy does not exist in that rogion af the model. This above statements ate true for any sensitivity based approach, ‘Those statements hold tre for both modal based updating approaches and frequency response based updating approaches. ‘Both techniques are fundamentally restricted by the above statements, Unfortunately, commercializaton ofthese techniques has, et limes, clouded the basic restrietions ang limitations ofthese techniques, Oftentimes, there isan impression in the mind of Potential users of these techniques thatthe updating process Produces the "ime" or "correct" mass and stiffness matrices, While the updating process will detiitely indicate trends in the improvement ofthese matrices if care is enereised in the seletion ‘of updating parameters, usually the analyst isnot sute a 10 Whi parameters should be selected for optimization. Therefore, the model is updated and a solution is achieved which satisfies the minimization ofthe parameters selected atthe expense af possibly distorting the mass and stiffness matrices, This often occurs since the analysts not exactly sure where the actual diserepaney exists ‘The iment of this paper isto demonsirate some of the common misinterpretations that may result through the use oF several simple examples ‘THEORY The basic theory of model updating is presented in detail in the ‘theoretical manual of the various software packages and papers have addressed this. Gener the system response and design parameters to be ‘changed can be related through a sensitivity equation which is approximated by {ar} = {ae} +[S}{¢e} (eel {me [en Hace} ‘The sensitivity coefficients, [S], with respect to the design parameters p, ae used to minimize the difference between the current state, a. andthe target state q. This equation is used ina ‘minimization scheme and oftentimes weighting fonctions ae used {express the analysts’ confidence in the parameters and responses sed forthe process.CONSIDERATION FOR MINIMIZATION PROBLEMS, Before any detailed diseussion is presented on particular models and test eases studies, it is beneficial o briefly review just what i required ofthe analyst for the wiliztion ofthese updating tools. The problem can be simply explained through the use of the following simple example «a least squares ft of data fora force age calibration, for instance. Consider the set of data shown in Figure 1. A frst glance, this data would be expected to be best fitusing a staight line that passes through zero, So using all the data with Teast squares fit af the data would produce the fit shown, However, it appears that tne fit does not look very good: the ezor between the data and the Tine fitting the data seems alittle large. But based on the constraots placed nthe fit, dhe parameters that bes ithe data areas shown. So isthe ft wrong? Or is the selection of Parameters used to deseribe the data wrong? The fitis not weong = it deseribes the data "as best as it can” given the parameters and ‘constraints given, yemx x Figure | Now consider the same data but with an exe parameter included to allow for something other than a zo crassng 3s shown in Figure 2. Notice hat the fit looks much beter and the error between the data and the line fiting the data is much less. yemx+b Figure 2 Now in reviewing the dara, why were all she data points used for the least squates process. Cleerly, it could he argued that the lowest measured value may be contaminated by noise and may be 'bclaw the actual useful range of the measuring device, Likewise the largest data may be outside the useful range of the measuring, deviee and not measured accurately. This is shown in Figure 3 [Notice that these estimated parameters produce a fit which looks ‘yet beter with the error between the data and te line fitting the data yet even less. 564 y=mx+b x Figure3 Another possibility on the use of all the data is to provide ‘weighting factors (or confidence factors) to identify how accurate the analyst believes each dats point is, All of the measured data could be assigned individual weighting factors to determine how ‘much emphasis each data point has on the minimization ofthe ‘orto produce parameters to describe the phenomena. So instead of arbitrarily eliminating data from the set used 1 estimate parameters, the analyst could assign a very low confidence inthe firs and last data points so that these data points do not hold equal ‘weight withthe other datapoints. Now that we have revisited this least squares error process, We can relat all ofthe above items to the model updating process that we ‘must address. First, we must selet the data which we wish to use in the model updating process. This ineludes which madal Srequoncy differences are selected; nt all the modal frequencies need tobe selected in the process. This also includes which response funetion differences of mode shape differences are used Jor the updating process. Second, the parameters that are allowed to change must be identified. This involves a determination of ‘which elements or parameter characteristics are selected Parameters such as area, thickness, bending inenia, mass, modulus can all he selected as parameters for individual elements, groups of element or all elements. Last the confidence tha the analyst has in the parameters selected needs tobe i unknown or very difficule to determine and sievitar ot identical confidences are usually applied throwehout the mode! for lack of more appropriate diseibution Al of these items identified above are very important (ond difficult) to select and have a very critical part in the development of the updated model. If incorrect or inappropriate items are selected for the updating process, then the results obtained may adjust the model in an incorrect fashion, However, iF appropriate parameters are selected, then a very good updated model can result, Clearly, the burden is on the analyst to select the appropriate characteristics for change in the model PRACTICAL CONSIDERATIONS The intent ofthis paper isnot to deve into the formulation ofa the equations used forthe model updating process. Rather, the inten is to identify the use ofthis tool and to help better identify {he critical role thatthe analyst plays when selecting parameters that ae allowed to change inthe mode! updating process.