Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Davide Caprioli
Marco Pierini
e-mail: marco.pierini@unifi.it
Dipartimento di Meccanica
e Tecnologie Industriali,
Universita` di Firenze,
Via di Santa Marta, 3,
50139 Firenze, Italy
This work presents an innovative approach to dynamic design that has the significant
advantage of allowing the dynamic requirements to be specified from the earliest design
stage. The method applies genetic algorithms to optimize the dynamic behavior of the
engine-subframe system and its links to the chassis. The optimization minimizes the sum of
the amplitudes of the forces transmitted to the chassis from each mounting, while complying with the static and dynamic constraints. The genetic algorithm was applied to a
multibody system model of the engine-subframe system and its links to derive new, improved configurations. DOI: 10.1115/1.1383968
Introduction
Quality and competitiveness are the challenges confronting todays automotive industry. To better meet them, numerical methodologies capable of predicting the behavior of complex system
are gaining increasing importance in the process. Not long ago,
computer-aided engineering CAE and computer-aided design
CAD drastically reduced the number of trial and error loops
required in the costly validation process. The next step will be to
provide designers with simple and efficient models that allow the
evaluation of alternatives and design optimization to be moved up
to the earliest design stages. Our objective in this work was to
develop a method for predicting the dynamic behavior of an automotive subsystem as part of the optimization process. Since
structural problems involve numerous independent variables, we
turned to genetic algorithms, which are eminently suited to such
applications.
Genetic Algorithms
Genetic algorithms are adaptive algorithms that function according to the principles of natural selection and genetics. Their
outstanding feature is the ability to search through complex multidimensional spaces, which derives from their being able to simultaneously explore in several directions. Moreover, needing
only objective functions, they do not require restrictive assumptions about the search spaces regarding the behavior of the function in the domain i.e., continuity and the existence of derivatives. For all the above reasons, genetic algorithms are
particularly advantageous with respect to the conventional gradient, heuristic, and analytical methods.
These features make genetic algorithms especially suited to numerous applications, including self-repair and self-guidance of artificial systems, machine learning, optimization processes, and
wherever biological flexibility is a prime requisite. Genetic algorithms have been successfully used to optimize sensor positions
1 and random signals 2. They are being increasingly applied to
structural problemsalthough the maximum number of independent variables used so far twenty is still quite limited. Some of
the most interesting structural applications to date include optimizing a model to minimize the frequency response function
FRF in a specific frequency band 3, performing threedimensional pipe routing 4, identifying the mounting locations
of a hydraulic pipe circuit that minimize energy transfer to the
support 5, defining the best shape for an elastomeric vibration
Contributed by the Technical Committee on Vibration and Sound for publication
in the JOURNAL OF VIBRATION AND ACOUSTICS. Manuscript received Dec. 1999;
revised Mar. 2001. Associate Editor: J. Q. Sun.
isolator 6, and optimizing the thickness of a beam damping material to achieve the best performance with the least amount of
material 7.
In the preliminary stage of our research, a specific geneticbased optimization routine was developed to obtain an algorithm,
that would converge quickly and stably to the optimum solution
when applied to complex multidimensional problems. We also
attempted to verify the possibility of solving the problem with an
undersized set of possible solutions population.
Current literature abounds with general purpose criteria to determine the appropriate population size 8 10. Among the simplest is the Levin-Lievin criterion 10, which sets the dimension
of the population to four times the length of a binary coded solution. The problem under investigation comprised a maximum of
48 variables, each of which was binary coded by five digits. Consequently, the coded solution, which was 240 digits long, would
have produced a population of 960 possible solutions. To avoid
the cumbersome and time-consuming process of evolving such a
large population toward optimum, we created a efficient genetic
algorithm to speed up and improve the optimization problem.
We developed a genetic-based optimization routine starting
from a basic structure 11, which comprised the simple genetic
operators reproduction, single-point crossover, and mutation.
Each variable was binary coded with five digits. Reproduction
was performed with a roulette wheel that was linearly biased on
the basis of the objective function values fitness values.
The algorithm was applied to a De Jong test environment as
defined by Goldberg 11 to verify its capabilities and determine
the proper values for the main parameters of the genetic operators.
The fact that the test environments five functions were potentially continuous/discontinuous, convex/nonconvex, unimodal/
multimodal, etc. in mathematical terms notably complicated the
optimization. We rated the validation process as sufficient, since
structural problems may generally be represented by better conditioned functions. Prior to testing, the functions had been extended
to fit the dimensionality of the problem.
The optimization of each function was repeated three times
with a maximum of 100 generations iterations per run. The dimension of the population was also a variable of the turning process. Each modification to the algorithm was tested separately to
properly evaluate its influence, and performance was evaluated on
the basis of optimum attainment and convergence stability.
We analyzed the results to determine the best values for crossover and mutation probabilities (p cross80%, p mut0.7%. In
addition, we modified the algorithms basic structure by adding
new genetic operators and changing some of the existing ones:
f f AVG
a f f AVG f AVG
f f AVG
(1)
In Eq. 1, a, which must exceed unitary, is the slope of the function. By attributing a higher fitness value to the individuals
which are greater than the population averagewe indirectly contribute to rewarding the most promising schemata, with the result
that the convergence of the optimization occurs at a significantly
faster rate.
Specific limits had to be respected for the engine displacements and rotations during coded situations, e.g., breaking with a
deceleration of 0.5 g, turning left and right on a flat road with a
friction coefficient of 0.5, applying 80 percent of the maximum
torque along the front axle. The loads were generated by the inertia of the system, but they were time-independent and were thus
simulated as static ones. For this reason, the displacements were
called quasistatic displacements.
The engines rigid body frequencies had to range between 4
and 15 Hz to avoid excitation from vehicle vibration lower limit
and to prevent the excitation of rigid body frequencies at idling
speeds upper limit.
The subframes rigid body frequencies had to be in excess of
50 Hz to prevent the system from being excited at the wheels
rotating frequency.
System Model. A multibody model of the system was created. The engine and subframe were modeled as rigid bodies having a mass and moments of inertia. In the case of the engine, there
Journal of Vibration and Acoustics
System Optimization
Objective Functions. Objective functions are the most critical aspect of genetic algorithms, because they strongly affect the
entire optimization procedure. Since we were dealing with a multiobjective problem, we had to include several terms that would
simultaneously evaluate performance in terms of constraints and
objectives. This was not an easy task owing to the terms differences in meanings and units of measure. We chose a function
composed of many modules, each of which would account for a
different constraint or objective. Their formulation is independent
of the specific system being analyzed and of the units of measure,
thereby making it possible to compare the different terms.
The first term of the fitness function was related to the quasistatic engine displacements and rotations. A nondimensional variable was defined for each component of the displacements translations: j1, . . . ,3 and rotations: j4, . . . ,6 according to Eq.
2.
x j x jLIM
xj
x jLIM
x j x jLIM
x jMAX
x jLIM
x j x jMAX
and
x j x jMAX
(2)
B y B y MAX
B y MINB y MAX
(3)
where y MIN was obtained at the x jLIM and y MAX at the x jMAX .
Equations 2 and 3 reveal that no distinction was made among
the solutions within the limits. To all these solutions the maximum
value of the objective function was assigned. In addition, no distinction was made for the solutions that were very far from the
target and above a fixed threshold; these solutions were assigned a
value of zero. The global value of the module for the quasistatic
displacements was obtained by assuming all the contributions
from all the displacement components of all the coded situations.
The second term of the fitness function Eq. 4 was related to
the systems rigid body frequencies. Since the engine and subframe requirements differed, it was imperative that their vibration
modes be distinguished. A check of the eigenvectors revealed that
the engines rigid body frequencies were the six lowest frequencies, since at least one of its components in each of the first six
eigenvectors was bigger than the corresponding components at the
other frequenciesa predictable effect of the engines substantial
mass. As in the case of the previous module, a nondimensional
variable was defined Eq. 4 before computing the fitness function
f r INF f r i
f r INF
f r i f r INF
f r i f r SUP
f r SUP
f r i f r SUP
(4)
f V MAX
F CALC
F RIF
(5)
yy MAX
y MINy MAX
(6)
Table 6 Comparison
stiffnesses
of
the
engine-subframe
mounting
Table 8 Comparison
stiffness
of
the
engine-subframe
mounting
Global Optimization.
In the second optimization, which
was performed on the same model, the mounting positions were
also considered variables, so that a total of 42 variables was available for the process. The position variability ranges were determined to allow the mountings to move along the subframe over a
the design process. The method is thus applicable from the earliest
design stages to specify and optimize the dynamic performance of
a generic mechanical system. In a forthcoming work, we intend to
improve the model by adding subframe material elasticity. Improvements in subframe modeling will also allow the introduction
of other optimization variables.
Acknowledgments
The authors are grateful to Giovanni Boreanaz, Davide Vige`,
and Marco Danti of the FIAT Research Center for their invaluable
support
Nomenclature
Fig. 5 Forces of the configuration with the optimized stiffnesses and positions solid line: sum of all forces; line with
circles: second-order forces; line with xs: second-and-halforder forces
References
Fig. 6 Top and front view of the subframe and the mounting
positions for the optimized configuration square: engine center of gravity; circle: engine-subframe mounting: x: subframechassis mounting
Conclusions
In this work a novel approach for the design of the dynamic
behavior of an engine suspension system was presented. In the
first part a genetic algorithm for high dimensional problems was
developed and illustrated. A new kind of fitness function was defined to improve performance so that the attainment of each objective was evaluated by a function module. The algorithm was
then applied to the optimization of an actual automotive system
composed of an engine, its subframe, and elastic mountings. The
system was optimized with two different approaches: In the first,
the mounting stiffness was modified; in the second, the mounting
stiffness and positions were modified. In both cases, the results
showed that the optimization improves the configuration, thereby
providing the designer with the information needed to facilitate
Journal of Vibration and Acoustics
1 Concilio, A., De Vivo, L., and Sollo, A., 1996, A Comparison of Different
Strategies for Interior Active Noise Control by Piezoactuators, Proceedings,
ISMA21 Conference, Leuven, Belgium, pp. 247258.
2 Cooper, J. E., De Fonseca, P., and Swivers, J., 1996, Optimal Random Excitation, Proceedings ISMA21 Conference, Leuven, Belgium, pp. 903913.
3 Heylen, W., and Mas, P., 1996, Optimization of Structural Modifications
Based on Frequency Response Functions, Proceedings, ISMA21 Conference,
Leuven, Belgium, pp. 18431850.
4 Chen, W., Fadel, G. M., and Sandurkar, S., 1997, GAPRUS: ThreeDimensional Pipe Routing Using Genetic Algorithms and Tessellated Objects, Proceedings, 1997 Design Engineering Technical Conferences, Sacramento, CA.
5 Edge, K. A., and Kwong, A. H. M., 1996, Minimization of Noise Transmission through Optimizing Hydraulic Pipe Mounting Locations, Proceedings,
ISMA21 Conference, Leuven, Belgium, pp. 549558.
6 Muscia, R., 1997, Development of a Methodology for Vibration Elastomeric
Isolators Design, Proceedings, 10th International ADM Conference, Florence, Italy, pp. 255264.
7 Bregant, L. and Pediroda, V., 1998, Optimal Unconstrained Damping Distributions for Vibrating Elements Using Algorithms, Proceedings, ISMA23
Conference, Leuven, Belgium, pp. 10471054.
8 Cant-Paz, E., Goldberg, D. E., Harik, G. R., and Miller, B. L., 1997, The
Gamblers Ruin Problem, Genetic Algorithms, and the Sizing of Populations,
Proceedings, 1997 IEEE Conference on Evolutionary Computation, IEEE
Press, New York, NY.
9 Giguere, P., and Goldberg, D. E., 1998, Population Sizing for Optimum Sampling with Genetic Algorithms: A Case Study of the Onemax Problems, Proceedings, 3rd Annual Conference on Genetic Programming, San Francisco,
CA.
10 Levin, R. I., and Lievin, N. A. J., 1997, Dynamic Finite Element Model
Updating Using Simulated Annealing and Genetic Algorithms, Proceedings,
15th International Modal Analysis Conference, Orlando, FL, pp. 11951201.
11 Goldberg, D. E., 1989, Genetic Algorithms in Search Optimizations and Machine Learning, Addison-Wesley, New York, NY.
12 Rivin, E. I., 1985, Passive Engine MountsSome Directions for Further
Developments, SAE Paper #850481.
13 Schmitt, R. V., and Leingang, C. J., 1976, Design of Elastomeric Vibration
Isolation Mounting Systems for Internal Combustion Engines, SAE paper
#760431.
14 Swanson, D. A., Miller, L. R., and Norris, M. A., 1992, Multi-dimensional
Mount Effectiveness for Vibration Isolation, AIAA/ASME/AHS/ASC, Proceedings, 33rd Conference on Structures, Structural Dynamics, and Materials,
Dallas, TX.
15 Zhu, J., Haghighi, K., and Krutz, G. W., 1988, Dynamic Analysis of an
Engine Chassis Mount Bracket Using the Finite Element Method, SAE paper
#881854.
16 Heisler, H., 1999, Vehicle and Engine Technology, SAE Edition, Warrendale, PA.