Jtam.dvi JOURNAL OF THEORETICAL AND APPLIED MECHANICS 48, 3, pp. 623-643, Warsaw 2010 IDENTIFICATION OF OPTIMAL MULTIBODY VEHICLE MODELS FOR CRASH ANALYSIS1 Marta Carvalho Jorge Ambrósio Institute of Mechanical Engineering, Instituto Superior Técnico, Technical University of Lisbon, Lisboa, Portugal; e-mail: marta@dem.ist.utl.pt; jorge@dem.ist.utl.pt This work proposes an optimization methodology for the identification of vehicle multibody models for crashworthiness analysis based on the use of plastic hinge approach. The multiple objective functions for the optimal problem are built as the deviation of the model behavior from the required crash responses.The designvariables of the problemare the plastic hinge constitutive relations. The constraints are set not only as technological side constraints but also as some of the deviation functions of selected crash responses that would, otherwise, be used as objective functions.Thevehiclemodel identificationmethodology is demonstrated by its application to the construction of virtual vehiclemodels, designa- ted as the generic car model, for which the reference is available as a detailed finite element model. Key words: multibody vehicle models, optimization, crashworthiness, plastic hinge 1. Introduction The construction of computational models of vehicles for crashworthiness stu- dies requires the knowledge of most of its construction details being the out- come of the analysis, generally, very sensitive in the quality of the model. When a detailed geometric model of the vehicle is available, the software to- ols existing today allows for the automatic construction of very detailed finite element models suitable for crash analysis. But, even when such automatic model generation tools exist for the building of multibodymodels, the neces- sary information on the vehicle may not be disclosed to the analyst, even if 1The paper was presented at the ECCOMAS Thematic Conference on Multibody Dynamics which was held atWarsaw University of Technology on June 29–July2, 2009. 624 M. Carvalho, J. Ambrósio a commercial partner of the automaker. This is, generally, the case either be- cause such aspects may be confidential by the vehicle developer, representing technological advances or simply due to legal reasons that are associated to liability and, therefore, they cannot be disclosed even to commercial or de- velopment partners. A solution to this problem is the construction of virtual vehicle models that have the same crash response of the detailed models for selected crash scenarios studies, but that are not an exactmatch of the original vehicles (Sousa et al., 2008). Such a crash response can bemeasured in terms of accelerations of given points in the vehicle structure or the anthropometric testing devices, energy absorption characteristics of the vehicle subsystems, intrusion measures or by any other measurable characteristic of the vehicle behavior in the crash scenario. Therefore, it is convenient to develop vehic- le models that can be used at the conceptual design development stage, for quick analysis and redesign. These models may include, or not, all structural andmechanical features of the real vehicle but theymust provide high quality crash responses for the selected scenarios. As an example of application of this virtual model, a developer would be able to tackle the task of devising a subcomponent or protective system for a selected part of the vehicle being as- sured that the overall behavior of the vehicle model is validated against some reference performance. The genericmultibody vehiclemodel can also be used in the design of new vehicles from scratch. Since the vehiclemodel has all structural characteristics of a real vehicle, it can bemodified to present an improved crash performance. In the context of a new design, such a reference crash response can be defined as a functional objective. The generic model can be used to devise which parts of its structure should bemodified in order to match the targeted crash performance.The next stage on the use of the newvehiclemodel would be the identification of real structural components or mechanical systems for which themultibodymodel would be valid, using inverse engineeringmethodologies. The most important outcomes on the simulations of crash events that re- quire analysis are themechanisms of deformation of the structural components to identify intrusions, the amount of energy absorption due to the structural deformation of the vehicle and its subsystems and the acceleration level of vehicle structural components and occupants (Ambrósio and Pereira, 1997). With these crash responses, it is possible to evaluatemost of the structural in- tegrity and biomechanical injury indexes (Ambrósio, 2010). Therefore, rather detailed information is required on deformation mechanisms of the structural components and on surface contact forces that are associated to the impact of the substructures of the vehicle. The vehicle model developed must inclu- Identification of optimal multibody vehicle models... 625 de the possibility for the development of such mechanisms of deformation on the structural components and on the structural regions where deformation energy has to be controlled. Thevehiclemultibodymodel, suitable for crashworthiness application, uses the plastic hinge approach (Nikravesh et al., 1983; Ambrósio, 2001), in which localized areas of the structural component experience the plastic deforma- tions. The first step of the model construction methodology is to identify the potential location of the plastic hinges well in advance of the simulation. Usu- ally, the plastic hinges occur near the joints of themember, weak areas of the element and load application points. A plastic hinge is modeled by a kine- matic joint, describing the kinematics of the deformation and a generalized spring element which is used to represent the constitutive characteristics of deformation. Such a constitutive spring element represents the elastic-plastic stiffness of themember and the energy absorption characteristics when plastic deformations occur (Nikravesh et al., 1983). The spring that represents the plastic hinge constitutive relation can be applied with any common type of joint used in multibodymodeling, as those depicted in Fig.1a-d, for one axis bending, two axis bending; torsion and axial loading (Ambrósio, 2001). Then, the structural component is modeled as a collection of rigid bodies connected by plastic hinges, such as in Fig.1e. Depending on the particular location of the component and the joint, the choice of each joint type is required and relevant for the deformation mechanism. Fig. 1. Plastic hinge concepts (a) to (d); and application in a vehicle model (e) 626 M. Carvalho, J. Ambrósio After devising the topological structure of the multibody system that re- presents the structural vehicle components that describe the most relevant mechanisms of deformation, it is necessary to identify the constitutive beha- vior of the plastic hinges and adjust the vehicle response to that of a reference vehicle. This task is actually done by using a trial and error procedure. Howe- ver, the improvement of themodel predictability, or its validation, can bedone by using an optimization procedure based on the minimization of deviation between the observed response of the vehicle model and the reference respon- se. The reference data is obtained by simulating crash tests of a validated FE model of the vehicle or by using data directly obtained in an experimental EuroNCAP test. Such a procedure is outlined in Fig.2, deemed as themetho- dology used to validation of the MB models. The vehicle model obtained is said to be validated and constitutes a virtual model of the reference vehicle (Carvalho, 2009). Fig. 2. Scheme of the methodology for validation of MB vehicle models Different optimization procedures can beused to solve the problem. In this work, themultibodydynamic analysis code is linkedwith general optimization algorithms included in theMATLABOptimizationToolbox (MATLAB,2008). After an evaluation of themost suitable optimization algorithm to handle this type of problems, the choice is the Sequential Quadratic Programming. Note that theuse of genetic or swarmoptimization (Eberhard et al., 2003; Sedlaczek andEberhard, 2006;Ambrósio andEberhard, 2009) has notbeen investigated. The optimization methodology requires the evaluation of the sensitivities of the problem. The simplest procedure to calculate the sensitivity derivatives is the finite-difference approximation (Adelman and Haftka, 1986). However, Identification of optimal multibody vehicle models... 627 small perturbations may result in errors in the derivative due to the limited accuracy of the dynamic response variation, while large perturbations can lead to truncation errors (Greene andHaftka, 1989). Analytical sensitivities, obta- ined either by direct implementation (Neto et al., 2009) or by using automatic differentiation tools (Bischof et al., 1992) are preferred if access to the source analysis code is possible (Neto et al., 2009).However, whenusing a commercial code, their use in not an option and, consequently, numerical sensitivities are the only option that can be used, regardless of the drawback of this method related to the additional analysis required (Dias and Pereira, 1997). The identification of the design variables and the objective function and constraints are of extremely importance for the resolution of the optimization problem. In this work, the objective is to identify a vehicle multibody model that has prescribed crash responses, in which each response is associated to accelerations, velocity and intrusion histories of different pick-up points of the structure and, eventually, with energy absorption ability of several structural regions of the vehicle. The problem can be defined as beingmultiple objective optimization or by summing the different objectives in a single function (Am- brósio and Eberhard, 2009). In this work, the time interval of the analysis is discretized into time points and the objective functions are evaluated as the sum of the square deviations at such points. Some of the objective functions are used as constraints in the optimal problem instead of using a procedure similar to that described byGonçalves andAmbrósio (2005) for the optimiza- tion of road vehicle suspensions. The choice of the design variables to be used in the identification of the vehicle multibody model is the final step required by the procedure. The potential design variables include the location of the plastic hinges, as in reference (Pereira and Ambrósio, 1997), the shape of the curves that make the plastic hinges constitutive relations, or the scaling of selected plastic hinge relations. Due to the ability to change the model crash behavior demonstrated by the scaling of the plastic hinge constitutive relation (Sousa et al., 2008) and to the complexity of the vehicle multibody models (Ambrósio and Dias, 2005); the design variables used in this work are the scaling factors of selected plastic hinge relations. Themethodology here proposed is suitable to be usedwith anymultibody dynamics approach or code, but it is demonstrated by its application to the identification of the vehiclemultibodymodel of a large family car by using the MADYMO multibody code as solver (MADYMO, 2004). It is foreseen that the use of optimization procedures can lead tomore robustmodels obtained in a shorter development time. The selection of the optimization methodologies usedand the strategies used to avoid premature convergence or instability pro- 628 M. Carvalho, J. Ambrósio blems in theoptimal problemarediscussedherein and theprocedureenvisaged is demonstrated by an application to the identification of a vehicle multibody model of a large family car (GCM3 MB model) based on the reference crash response obtained from a detailed vehicle finite element model (GCM3 FE model). 2. Vehicle multibody models: matching reference models 2.1. Motivation The generic car model is to be used in crash test scenarios described by the regulations approved inEurope.Of particular relevance are theECE regu- lations for frontal and side impacts that must be fulfilled by any new vehicle that seeks approval for release in the EU countries. In the application descri- bed here, theGCM3MBmodel is tested according to theECER33European Regulation for full frontal rigid barrier impact (1995). The same frontal crash test is conducted with the GCM3 FE model. Due to the easy access to fini- te element models, these are used here as reference vehicles. This MB model has been validated for frontal impact according to the procedure reported by Puppini et al. (2005). However, note that real vehicle data can be used as re- ference data with the proposed procedure when available. In agreement with the regulation, the impact speed is 48.3Km/h. The crash response of the vehicle is measured by accelerations, velocities and intrusions in selected points of the structure, shown in Fig.3. It should be noticed that the acceleration signals are filtered with CFC 60, the velocities and displacements signals are filtered with CFC 180 in both FE models and MBmodels. Fig. 3. Accelerometers position in GCM3 model: (a) FE model; (b) MBmodel Frames showing the deformations of the simulations results of the initial MB and FE models, for the ECE R33 crash scenario, are presented in Fig.4. Identification of optimal multibody vehicle models... 629 With the dynamic response and deformations shown, these results suggest that the initial MB model is too stiff when compared with the reference mo- del. In particular, someplastic hinges have constitutive functions, which cause rotation in the longitudinal bars of the frame instead of allowing for the defor- mation of the crashbox.Theplastic hinge approach disregards the generalized elastic deformation of the structural members and considers the plastic defor- mation region as represented by a single point. Two corrective measures can be applied to the MBmodels, one is to increase the number of plastic hinges in each structural member, the second, which is used in what follows, is to decrease the stiffness of some plastic hinges constitutive functions. Fig. 4. Sequence of deformation of GCM3 MB and FEmodels in ECER33 test The objective is to correlate the displacement, velocity and acceleration in the five accelerometers displayed in Fig.3 of the MB model with the FE model, which is the reference. In the first approach considered, the usual trial and error procedure used in industrial applications, the correlation between theMB and FE models is attempted. The strategy pursued consists in scaling the stiffness of the constitutive functions of all plastic hinges in each subcomponent. Several simulations are performed attempting different scaling factors for the constitutive functions. The results that lead to the best correlation between theMBmodel response and the FE model are achieved by using different scaling factors for different subsystems. The scaling factors are the following (Sousa et al., 2008): • Forces on the plastic hinges of thebody structure are scaled by x1 =0.1. • Forces on the plastic hinges of the frame are scaled by x2 =0.2. • Forces on theplastic hinges of thebumper frameare scaledby x3 =0.35. 630 M. Carvalho, J. Ambrósio • Forces on the plastic hinges of the sub-frame are scaled by x4 =0.45. • Forces in the plastic hinges of the bumper sub-frame are scaled by x5 =0.65. The location of the kinematic joints with plastic hinges referred is repre- sented for the left side of the car in Fig.5, being symmetrically located for the right side of the vehicle. Fig. 5. Joints localization and respective design variables for the GCM3-MBmodel All responses measured in terms of accelerations, velocities and displace- ments of the improved model show a better correlation with the reference response than the original model. This improvement is measured with the er- ror function defined by Eq. (2.1), and in Table 1 the error for the initial and for the improvedMBmodel is listed fi(x)= N ∑ i=1 √ (Rref (i)−Rn(x,i)) 2 (2.1) Table 1. Measure of the error for the initial and improved MB model calcu- lated with Eq. (2.1) Accelerometer i fi,acc (×104) fi,vel (×102) fi,disp Initial Improved Initial Improved Initial Improved A-Pillar 1.64 0.72 2.57 0.36 5.45 0.33 FWCenter 1.49 1.40 2.47 0.48 5.14 0.21 Front Sill 1.59 0.77 2.56 0.41 5.23 0.54 B-Pillar (middle) 1.52 0.47 2.86 0.50 5.24 0.47 B-Pillar (bottom) 1.50 0.73 2.59 0.39 5.25 0.44 Identification of optimal multibody vehicle models... 631 The response of the initial and improved multibody and reference finite element models are presented in Fig.6 for accelerations, velocities and displa- cements measured in the sensor located in the bottom B-Pillar. All accelera- tion, velocity and displacement responses of the improved model show good correlations with the reference responses as depicted in Table 1. Fig. 6. Dynamic responses for the initial and improvedMB model versus the FE model in the front impact test measured in the sensor located in the left BP bottom 3. Identification of the multibody vehicle models The improvement of the vehicle multibody model is obtained at the cost of the constitutive equations and location of the plastic hinges and the potential mechanisms of deformation, which aremodifiedwithin a given range of varia- tion. This has been the procedure used by other researchers as well, seeMooi et al. (1999), Gielen et al. (2000), Zweep et al. (2005). The acceptable range of variation of themodel data in the validation step is related to the approxima- tions madewhen defining the discretization of themodel and the constitutive equations of theplastic hinges.These ranges are represented as corridors inside 632 M. Carvalho, J. Ambrósio which different force-displacement characteristics are accepted. The validation process includes: 1) Collect the dynamic responses of the reference vehicle as intrusions, ve- locities and accelerations of selected points of the body structure or occupant models. 2) For the MB model, define which parameters are allowed to vary and identify their variation range. The constitutive relation for a particular plastic hinge is shown in Fig.7. 3) For each of thedynamic responses considered in step1 calculate the error between the reference response and the actual response of the model at a finite number of instants. Fig. 7. Constitutive relation for a generic plastic hinge and variation corridor 3.1. Formulation of the optimization problem Although the trial and error procedure leads to a better correlated MB model, the outlined procedure constitutes in fact an optimization process to identify the MB model that best represents the reference vehicle. Sequential Quadratic Programming (SQP) methods are the recommended optimization methods for this type of problems (Schittkowski, 1985). The fmincon, which is the SQP method, implemented in MATLAB Optimization Toolbox (MA- TLAB, 2008) is used in this work. In this method, the function solves a qu- adratic programming (QP) subproblem at each iteration. First, an estimate of the Hessian of the Lagrangian is updated using the BFGS (Fletcher and Powell, 1963; Goldfarb, 1970), which is used to generate a QP subproblem that is solved using an active set strategy similar to that described in Gill et al. (1981). The solution of this problem is used to form a search direction for a line search procedure. The line search is performed using a merit function si- milar to that proposed byGill et al. (1981), Han (1977) andPowell (1978a,b). This gradient based optimization algorithm can achieve fast convergence re- quiring less function evaluations when compared with the genetic algorithms. Identification of optimal multibody vehicle models... 633 However, this algorithm is heavily dependent on the initial design and requ- ires accurate gradient information for the iterative approximation of thedesign space. In this case, the gradient is obtained with the forward finite difference formula, which requires n+1 function evaluations, being n the number of design variables. If the updating of the Hessian matrix is deemed inefficient, n+1 extra function evaluations are needed. The gradient information is severely affected by numerical noise that is normally inherent to the nonlinear simulation of complex numerical models, such as the full vehicle models used in this application. The success of the implementation of the SQP algorithm to solve the present optimization pro- blem is affected not only by the initial design but is also conditioned to the constraints imposed, limiting the feasible region for search and avoiding the lack of precision of the gradient information. The objective of themodel validation is described by a scalar optimization problem defined by min x F(x)= n ∑ i=1 wifi,acc(x)= 5 ∑ i=1 N ∑ t=1 √ ( acci ref (t)−accin(x,t) )2 ·10−4 (3.1) subject to: 0.1¬xj ¬ 2, j=1, . . . ,5 and fi,vel(x)= nt ∑ t=1 √ ( veliref (t)−vel i n(x,t) )2 ¬ f∗i,vel f ∗ vel = [40,50,45,55,40] where the objective function is defined as the sum of the mean square error of the accelerations, represented in Fig.8, measured in different points of the vehicle, represented in Fig.5, and sampled for the duration of the crash event. Thedesignvariables xj, j=1, . . . ,5, are the samedefined for the improvement made by the trial and error process previously described, which are the scale factors for the plastic hinge forces of the constitutive relations for the body structure, frame, bumper, sub-frame and bumper sub-frame hinges. The error in the intrusion velocity profile of the vehicle is used as a con- straint rather than in the objective functions, being the total mean square de- viation between themodel and the goal velocities given by f∗vel. These values are chosen based in the values obtained for the improvedMBmodel (Table 1), in order to obtainmodels by an optimization algorithmwith better or similar correlations than this one. The maximum number of iterations allowed and tolerances used for the convergence criteria of the SQP algorithm used is listed in Table 2. 634 M. Carvalho, J. Ambrósio Fig. 8. Model and reference responses for a selected point of the vehicle Table 2. Stopping criteria used in fmincon Options Description Set values MaxFunEval Maximum number of function default evaluations allowed {100*numberOfVariables} MaxIter Maximum number of default {400} iterations allowed TolFun Termination tolerance on 0.01 the function value TolCon Termination tolerance on default {10−6} the constraint violation TolX Termination tolerance on x 0.1 3.2. Results of the application Starting from the raw model, x0j = 1, j = 1, . . . ,5, the optimum de- sign is achieved for the 14th iteration, with an optimum objective func- tion value of 3.72 and with the design variable xopt = [0.1,0.22771,0.32932, 0.11232,0.18595]. The iterations history is represented in Fig.9 and listed in Table 3. Fig. 9. History of the value of the objective function Identification of optimal multibody vehicle models... 635 Table 3.History of SQP iterations of the vehicle model optimization x1 x2 x3 x4 x5 f(x) 0 1.00000 1.00000 1.00000 1.00000 1.00000 7.73 1 0.80555 0.68728 1.90560 1.81640 0.14442 7.15 2 2.00000 0.10000 0.10000 1.68070 0.35768 6.92 3 0.15989 0.10000 0.64067 0.79964 0.10000 3.74 4 0.29380 0.10000 0.10000 0.10000 0.10000 4.07 5 0.33259 0.23192 0.10000 0.10000 0.25839 4.33 6 0.30370 0.19192 0.29173 0.10000 0.13880 4.23 7 0.30372 0.19184 0.29340 0.10000 0.13888 4.04 8 0.30371 0.19182 0.29340 0.10001 0.13888 3.99 9 0.33662 0.19616 0.29471 0.11049 0.13672 4.07 10 0.14522 0.22951 0.32976 0.10000 0.18649 3.79 11 0.10000 0.23157 0.32915 0.11539 0.19119 3.70 12 0.10000 0.22536 0.32988 0.12794 0.19057 3.86 13 0.10000 0.22741 0.32877 0.11518 0.18403 3.82 14 0.10000 0.22771 0.32932 0.11232 0.18595 3.72 Table 4 shows the objective function, i.e., the error measure between the reference andMB crash response for the initial and optimal MBmodels. The results show an improvement of all dynamic responses as shown in Fig.10. Table 4.Measure of the criterion for the optimized MBmodel Accelerometer i fi,acc (×104) fi,vel (×102) fi,disp Initial Optim. Initial Optim. Initial Optim. A-Pillar 1.64 0.66 2.57 0.35 5.45 0.28 FWCenter 1.49 1.24 2.47 0.44 5.14 0.17 Front Sill 1.59 0.70 2.56 0.40 5.23 0.47 B-Pillar (middle) 1.52 0.44 2.86 0.54 5.24 0.54 B-Pillar (bottom) 1.50 0.68 2.59 0.38 5.25 0.38 Though the results showan improvement of the correlations, with gradient based algorithms, no guarantee exists that the global minimum is reached. With a grid of initial designs, it is possible to look for the lowest objective function value, but always without being able to ensure that it is the global minimum. The design obtained by the trial and error procedure, which shows also a satisfactory correlation with the reference response, is a good candidate for the initial design in this methodology. 636 M. Carvalho, J. Ambrósio Fig. 10. Dynamic response for the optimizedMB model versus the FEmodel in the front impact test Fig. 11. History of the value of the objective function Starting from the improvedmodel, the optimumdesign is achieved for the 5th iteration, with an optimal function value of 3.63. The iterations history is listed in Table 5 and represented in Fig.11. The optimum design is defi- ned by the plastic hinge scale factors shown in the vector xopt = [0.13199, 0.13782,0.22923,0.59378,0.42484]. Identification of optimal multibody vehicle models... 637 Table 5.History of SQP iterations x1 x2 x3 x4 x5 f(x) 0 1.00000 0.20000 0.35000 0.45000 0.65000 4.08 1 0.10587 0.20163 0.37578 0.47422 0.67109 4.02 2 0.10604 0.20124 0.37470 0.48018 0.67289 3.99 3 0.10648 0.20084 0.37363 0.48372 0.67354 3.97 4 0.14023 0.15042 0.23682 0.57870 0.38677 3.67 5 0.13199 0.13782 0.22923 0.59378 0.42484 3.63 Table 6 shows the error measure for the optimized and improved by trial and errorMBmodels. The results show an improvement in almost all dynamic responses considered, as shown in Fig.12. Table 6.Measure of the criterion for the optimized MBmodel Accelerometer i fi,acc (×104) fi,vel (×102) fi,disp Initial Optim. Initial Optim. Initial Optim. A-Pillar 0.72 0.59 0.36 0.20 0.33 0.07 FWCenter 1.40 1.38 0.48 0.49 0.21 0.21 Front Sill 0.77 0.58 0.41 0.24 0.54 0.21 B-Pillar (middle) 0.47 0.65 0.50 0.52 0.47 0.94 B-Pillar (bottom) 0.73 0.59 0.39 0.22 0.44 0.14 The optimization problem is formulated as the minimization of an objec- tive function, described by the sum of the mean square error of the vehicle accelerations measured in different points of the structure and sampled for the duration of the crash event. The error of the intrusion velocity profile of the vehicle is used as a non-linear constraint. Table 7 contains a summary of the results obtained in the application with the two initial designs: the first concerns the raw model and the second is the model improved by trial and error. Starting from the raw MB model, the SQP algorithm took 14 iterations to obtain an optimal design, being the minima basically reached in three ite- rations. However, there appear numerical problems with the gradients and, consequently, with the Hessian matrix update, which lead to oscillations of the iterative optimization procedure about the local minima until convergen- ce is obtained without constraint violations. This optimization problem takes about 5CPUhours to reach theminimumwhich improves all crash responses, when comparing with the initial design. 638 M. Carvalho, J. Ambrósio Fig. 12. Dynamic response for the optimizedMB model versus the FEmodel in the front impact test Based on the knowledge of a gooddesign, the secondoptimization runused the improvedMBmodel by trial and error. Because this design is close to the minimum, in this optimization the convergence was reached in five iterations only. Furthermore, the velocity non-linear constraints are the upper limits for this initial design. This design leads to the lowest value of the objective function. All crash responses of the optimal problem are better correlated with the reference response than themodel improved by trial and error, with the exception of the acceleration, velocity and displacement in the middle of B-Pillar. The crash responses of the validated and reference vehicles are shown in Fig.12 showing that a good predictability of the model has been achieved. Notice that the deformations of the optimized and FE model are very similar for the frontal crash test, as shown in Fig.13, which confirms the results obtained through the crash response. Identification of optimal multibody vehicle models... 639 Table 7.Measures of the criteria for the optimized MBmodel Initial Opt. 1 Improved Opt. 2 model model model model Iterations – 14 – 5 f(x) 7.73 3.72 4.08 3.63 x(1) 1 0.10000 0.1 0.13199 x(2) 1 0.22771 0.2 0.13782 x x(3) 1 0.32932 0.35 0.22923 x(4) 1 0.11232 0.45 0.59378 x(5) 1 0.18595 0.65 0.42484 AP 1.64 0.66 0.72 0.59 FW Ctr 1.49 1.24 1.40 1.38 fi,acc (×104) F Sill 1.59 0.70 0.77 0.58 BP mid 1.52 0.44 0.47 0.65 BP bot 1.50 0.68 0.73 0.59 AP 2.57 0.35 0.36 0.20 FW Ctr 2.47 0.44 0.48 0.49 fi,vel (×102) F Sill 2.56 0.40 0.41 0.24 BP mid 2.86 0.54 0.50 0.52 BP bot 2.59 0.38 0.39 0.22 AP 5.45 0.28 0.33 0.07 FW Ctr 5.14 0.17 0.21 0.21 fi,disp F Sill 5.23 0.47 0.54 0.21 BP mid 5.24 0.54 0.47 0.94 BP bot 5.25 0.38 0.44 0.14 Fig. 13. Sequence of deformation of the improvedMBmodel versus the FEmodel in the front impact test 640 M. Carvalho, J. Ambrósio 4. Conclusions The methodology proposed here is demonstrated by its application to the identification of the vehiclemultibodymodel of a large family car forwhich the reference crash behavior is available. It is foreseen that the use of optimization procedures can lead tomore robustmodels than those commonly obtained by trial and error approaches. The identification of the design variables and the objective functionand constraints are of extreme importance for the resolution of the optimization problem. Themultiple objectives, represented by themean square errors of the diffe- rencebetween themultibodymodel and reference crash responses,measured in several points of thevehicle, are reduced to a single objective functionobtained as a weighted sum of the individual error accelerations. The main difficulty of this approach is the choice of the initial designs and of the values for the weights wi, which has influence on the optimization results. In this applica- tion, the optimization results for two initial designs, the raw model and the improved model, being assumed that all acceleration criteria in the objective function are equally important, are compared.The function that evaluates the velocity error, defined by the vector f∗vel is introduced in the optimal problem as constraints and kept below specified values. This approach allows reducing the design feasible region, leading to a better convergence. The disadvantage of this approach is that the solutions to the optimal problem depend on the values of the non-linear constraint vector. Overall, a large improvement of the vehicle multibodymodel is observed as the acceleration errors have been gre- atly minimized and the velocity and intrusion errors basically eliminated. In an industrial application, the methodology proposed in this work is suitable to lead to better vehicle designs, satisfying a large number of the requirements imposed by regulations restrictions, project goals, thus shortening the deve- lopment time and reducing the inherent costs. The designs obtained with the optimization methodology presented in this framework are suitable to serve as the basis for a detailed design of the vehicle with improved crashworthiness characteristics. Acknowledgements The support of EC through project APROSYS-SP7 with the contract number TIP3-CT-2004-506503, having as partners Mecalog (Fr), CRF (It), TNO (Nl), Poti- tecnico di Torino (It), CIDAUT (Sp) Technical University of Graz (At) is gratefully acknowledged. Inparticular, the supportofTNOformultibodymodelingandMecalog and CRF for finite element models is greatly appreciated. The financial support of the first author has been received fromFCT –Fundação para a Ciência e Tecnologia through grant SFRH/BD/23296/2005. Identification of optimal multibody vehicle models... 641 References 1. AdelmanH.M., HaftkaR.T., 1986, Sensitivity analysis of discrete structu- ral systems,AIAA Journal, 24, 5, 823-832 2. Ambrósio J., 2001,Crashworthiness: EnergyManagement andOccupant Pro- tection, Springer-Verlag,Wien 3. Ambrósio J., 2010,Automotive structural crashworthiness and occupant pro- tection, In:Road and off-road Vehicle SystemDynamics Handbook, G.Mastinu andM. Ploeck (Eds.), Taylor and Francis, London 4. Ambrósio J., Dias J., 2007, A road vehicle multibodymodel for crash simu- lation based on the plastic hinges approach to structural deformations, Inter- national Journal of Crashworthiness, 12, 1, 77-92 5. Ambrósio J., Eberhard P., 2009,Advanced Design of Mechanical Systems: From Analysis to Optimization, Springer-Verlag,Wien 6. Ambrósio J., Pereira M.S., 1997, Multibody dynamic tools for crashwor- thiness and impact, In:Crashworthiness Of Transportation Systems: Structural Impact and Occupant Protection, J.A.C. Ambrósio et al.. (Eds.), NATO ASI Series E. Vol. 332, 475-521, Kluwer Academic Publishers, Dordrecht 7. Bischof C., Carle A., Corliss G., Griewank A., Hovland P., 1992, ADIFOR: generating derivative codes from Fortran programs, Scientific Pro- gramming, 1, 1, 11-29 8. Carvalho M., 2009, Development of Optimal Multibody Vehicle Models for Crash Analysis, Ph.D. Dissertation, Technical University of Lisbon, Lisbon 9. Dias J.P.,PereiraM.S., 1997,Sensitivity analysis of rigid-flexiblemultibody systems,Multibody System Dynamics, 1, 303-322 10. Eberhard P., Dignath F., Kübler L., 2003, Parallel evolutionary optimi- zation of multibody systems with application to railway dynamics, Multibody System Dynamics, 9, 2, 143-164 11. European Council for Europe, 1995,Regulation 33 – Uniform Provisions Con- cerning the Approval of VehiclesWith Regard to the Behaviour of the Structure of the Impacted Vehicle in a Head-on Collision 12. Fletcher R., Powell M.J.D., 1963, A rapidly convergent descent method for minimization,Computer Journal, 6, 163-168 13. Gielen A.W.J., Mooi H.G., Huibers J.H.A.M., 2000, An Optimization Methodology for Improving Car-to-Car Compatibility, IMechE Transactions C567/047/2000 14. Gill P.E., Murray W., Wright M.H., 1981,Practical Optimization, Aca- demic Press, London 642 M. Carvalho, J. Ambrósio 15. GoldfarbD., 1970,A family of variablemetric updates derivedbyvariational means,Mathematics of Computing, 24, 23-26 16. Gonçalves J., Ambrósio J., 2005, Road vehicle multiple objective optimi- zation for ride and stability,Multibody Systems Dynamics, 13, 1, 3-23 17. Greene W., Haftka R., 1989, Computational aspects of sensitivity calcula- tions in transient structural analysis,Computers and Structures,32, 2, 433-443 18. Han S.P., 1977, A globally convergent method for nonlinear programming, Journal of Optimization Theory and Applicatiobs, 22, 3 297-309 19. MADYMO Manuals, 2004, version 6.2. TNOMADYMO BV, Delft 20. MATLAB Release Notes, 2008, TheMathworks, Inc., Natick 21. Mooi H.G., Nastic T., Huibers J.H.A.M., 1999,Modeling and Optimiza- tion of Car-to-Car Compability, VDI Berichte NR 1471, 239-255, Delft 22. Neto M.A., Ambrósio J., Leal R., 2009, Sensivity analysis of flexible mul- tibody systems using composite materials components, International Journal of Numerical Methods in Engineering, 77, 3, 386-413 23. NikraveshP.E., Chung I.S., Benedict R.L., 1983, Plastic hinge approach to vehicle crash simulation,Computers and Structures, 16, 1/4, 395-400 24. PereiraM.S.,Ambrósio J.,Dias J., 1997,Crashworthinessanalysis andde- sign using rigid-flexible multibody dynamics with application to train vehicles, International Journal of Numerical Methods in Engineering, 40, 655-687 25. Powell M.J.D., 1978a, A fast algorithm for nonlinearly constrained opti- mization calculations. Numerical analysis, In: Lecture Notes in Mathematics, G.A.Watson (Ed.), 630, 144-175, Springer-Verlag, Berlin 26. PowellM.J.D., 1978b,The convergence of variablemetricmethods for nonli- nearly constrained optimization calculations, In:Nonlinear Programming, O.L. Mangasarian,R.R.Meyer andS.M.Robinson (Eds.),3, 27-63,AcademicPress, NewYork 27. Puppini R., Diez M., Ibba A., Avalle M., Ciglaric I., Feist F., 2005, GenericCar (FE)Models forCategories SuperMinis, Small FamilyCars,Large Family ExecutiveCars,MPVandHeavyVehicle, Technical Report APROSYS SP7-Virtual Testing, AP-SP7-0029-A 28. Schittkowski K., 1985, NLQPL: A FORTRAN-subroutine solving constra- ined nonlinear programming problems, Annals of Operations Research, 5, 485-500 29. Sedlaczek K., Eberhard P., 2006, Using augmented Lagrangian particle swarm optimization for unconstrained problems in engineering, Structural and Multidisciplinary Optimization, 32, 4, 277-286 Identification of optimal multibody vehicle models... 643 30. Sousa L., Verssimo P., Ambrósio J., 2008, Development of generic road vehiclemodels for crashworthiness,Multibody SystemsDynamics,19, 1, 135-158 31. Zweep C.D. van der, Kellendonk G., Lemmen P., 2005, Evaluation of fleet systems model for vehicle compatibility, International Journal of Crash- worthiness, 10, 5, 483-494 Identyfikacja optymalnych modeli wielobryłowych pojazdów samochodowych dla celów analizy testów zderzeniowych Streszczenie Praca prezentuje metodologię optymalizacji identyfikacji wielobryłowych modeli pojazdów samochodowych pod kątem ich przydatności do symulacyjnych badań zde- rzeniowych opartych na koncepcji plastycznych połączeń międzybryłowych. Funkcje celu w omawianymzagadnieniu polioptymalizacji sformułowano na podstawie odchy- leńodpowiedzi dynamicznychwstosunkudowymaganychprzebiegówobserwowanych podczas testów zderzeniowych. Dobieranymi funkcjami modelu są równania konsty- tutywne plastycznych połączeń międzybryłowych. Przyjęto, że więzy wynikają nie tylko ze względów technologicznych, ale także są funkcjami odchyleń w rejestrowa- nych odpowiedziach dynamicznych układu przy zderzeniu, które w przeciwnym razie same byłyby użyte jako funkcje celu.Metodologia identyfikacji została zaprezentowa- na na przykładzie jej aplikacji do budowywirtualnychmodeli samochodów, zwanych modelami generycznymi, dla których układami odniesienia są modele szczegółowo przeanalizowane za pomocąMetody Elementów Skończonych. Manuscript received November 27, 2009; accepted for print December 17, 2009