MBDyn - Multibody System Dynamics

LaMSID (EDF R&D - French CNRS Joint Laboratory)

July 24 - August 7, 2008

Pierangelo Masarati <[email protected]>

Politecnico di Milano
Dipartimento di Ingegneria Aerospaziale
 2
Preamble

This presentation is split in two parts:

• Multibody System Dynamics,

with specific reference to MBDyn,

the free general-purpose multibody solver
developed at “Dipartimento di Ingegneria Aerospaziale”
of “Politecnico di Milano”, Italy

• FEM interaction requirements of MBDyn,

with the outline of possible solutions for the interaction with Aster

LaMSID – July 24 – August 7, 2008
 3
Outline

• Overview
• Multibody dynamics
• Software architectures
• Problems
• Arbitrary motion description
• Deformable components
• Solving the problem
• Extracting useful information
• Examples of multibody modeling with MBDyn
• Future development
• Documentation and support
• Acknowledgments

LaMSID – July 24 – August 7, 2008
 4
Multibody dynamics: overview

• Multibody: a “buzz” word?

• Initial idea:
 automatically write equations of motion of arbitrary mechanisms

• Current status:
 efficiently and accurately integrate in time
 exact rigid-body kinematics, plus
 nonlinear finite elements, plus
 natural inclination towards multi-physics & system integration

• Future:
 scale to larger and larger problems, and
 higher performances when solving more complex problems

LaMSID – July 24 – August 7, 2008
 5
Multibody dynamics: overview

Multibody methods:

• Usually are general-purpose: can model a wide variety

of mechanical systems

• Must allow to automatically write motion equations

• Should support an arbitrary number of a variety of parts, forces,

geometries, constraints, etc.

• Most often use numerical methods to compute solutions.

• Often integrated in CAD tools, with Graphical User Interfaces (GUI)

LaMSID – July 24 – August 7, 2008
 6
Multibody dynamics: overview

Types of analysis:

• Kinematics: direct, inverse

• Statics (pseudo-time)

• Dynamics: direct IVP, BVP, inverse dynamics

• Miscellaneous:
 modal analysis
 sensitivity analysis
 parameter optimization
 topology optimization and mechanism synthesis

LaMSID – July 24 – August 7, 2008
 7
Multibody dynamics: overview

Applications:
• Robotics
• Automotive
• Crashworthiness
• Aerospace engineering
• Industrial automation
• Applied mechanics

• Videogames: revenue higher than engineering software!

• Anywhere realistic motion (or at least that looks realistic)
is required, multibody simulation is or can (and will) be used.

LaMSID – July 24 – August 7, 2008
 8
Multibody dynamics: overview

Classification of approaches:
• Based on coordinate set
classical approach
 Minimal coordinate set
 Redundant coordinate set modern approach

• Based on local solution methods

 Direct linear solvers classical approach
 Iterative solvers modern approach

• Based on problem formulation

 Explicit specialistic
 Implicit general-purpose

• Based on integration schemes

 Ordinary Differential Equations (ODE) w/ constraint stabilization
 Differential-Algebraic Equations (DAE)

LaMSID – July 24 – August 7, 2008
 9
Multibody dynamics

Basic equations:

M  x  ẍ=f x , ẋ , t
 mechanics of unconstrained system of bodies
 subjected to configuration-dependent loads

Can be obtained from many (equivalent!) approaches:

 Newton-Euler: linear/angular equilibrium of each body
 d'Alembert-Lagrange: virtual work of active forces/moments
 Gauss, Hertz, Hamilton, ...: variational principles

Note: non-smooth mechanics tend to resort to discrete equations

k
M  x ẋ k−1, k =∫k−1 f  x , ẋ , t d t

LaMSID – July 24 – August 7, 2008
 10
Multibody dynamics

Constrained system: kinematic constraints

 holonomic (algebraic)
 x , t=0 ẍ=aM −1 f c
 Non-holonomic
(differential, not integrable to holonomic)
x , ẋ , t=0
M−1 f c
usually
a=M −1 f
A  x , t ẋ=b x , t

• algebraic relationship between kinematic variables

• explicitly dependent on time: rheonomic
• scleronomous otherwise

LaMSID – July 24 – August 7, 2008
 11
Multibody dynamics

Minimal set:
x=xq , t
usually, this relationship:
 is not known in advance, or
 cannot be easily made explicit with respect
to Lagrangian coordinates q

Cordinate partitioning is required, e.g.:

 direct elimination from derivative of constraint equation
 QR, SVD or similar decomposition

Results in Maggi-Kane equations and similar approaches

Small system is obtained by expensive numerical reduction

(unless topology knowledge can be exploited)

LaMSID – July 24 – August 7, 2008
 12
Multibody dynamics

Redundant set:
 ⋅= ⋅ x⋅T/x 
T
 ⋅= ⋅ x⋅ / ẋ 
By Lagrange multipliers:
 dynamics of constrained system using physical coordinates
 constraint reactions applied to equations of motion
 algebraic constraints explicitly added to the system

• mechanism made of multiple bodies with multiple constraints

and few actual DoFs:
 system size nearly doubles

• mechanism made of deformable bodies with few constraints (e.g. FEM):

 system size not significantly altered

[ ]{ } { }
T
M  / x ẍ f
=
• sparsity is almost preserved  /x 0  b'

LaMSID – July 24 – August 7, 2008
 13
Multibody dynamics

Constraint equations written “as is”:

 x , t=0
problem becomes differential algebraic (DAE); issues:
• needs specific care to be solved: (nearly) L-stable integration, i.e.
 unconditionally stable (A-stable), and
 x  0 for  t ∞
k1

• the constraint equation implies the additional constraints

̇ x , t=0
̈ x , t=0
but they are not explicitly enforced:

may need constraint stabilization (Gear et al.)

LaMSID – July 24 – August 7, 2008
 14
Multibody dynamics

Constraint equations:
• x is correct
• derivatives may be inaccurate
• multipliers may be inaccurate

numerical solution x j , t j =0

k+2
numerical solution  x j , t j =0
with constraint k+2
stabilization ̇ x j , t j =0
k+1

k
 x , t=0 exact solution

LaMSID – July 24 – August 7, 2008
 15
Multibody dynamics

Alternative: constraint equations differentiated to second order:

 / x ẍ=b '
problem remains ordinary differential (ODE);
• can be solved by conditionally stable algorithms
• the constraint equation does not imply the original constraints

 x , t=0
̇ x , t=0

suffers from drift! definitely needs constraint stabilization!

common technique: Baumgarte

 / x ẍ=b ' −2  ̇−2 
(violation governed by asymptotically stable linear differential eq.)

LaMSID – July 24 – August 7, 2008
 16
Software architectures

• Monolithic:
 user prepares specific model using built-in library elements
 general-purpose solver swallows model and spits results

• Library:
 user writes specific solver using library elements
• usually needs programming skills; the solver must be compiled
 specific solver solves the problem and spits results

• Symbolic manipulators:
 user writes equations
 symbolic manipulation engine solves equations and spits results

• Modelica (and Modelica-like):

 user prepares model using a modeling language and libs
 general-purpose interpreter generates specific solver
 specific solver solves the problem and spits results

LaMSID – July 24 – August 7, 2008
 17
Software architectures

Free software examples (surely there are more):

• Monolithic:
 MBDyn
non-free counterparts omitted
• Library:
 DynaMechs (C++)
 Mbs3d (requires Matlab)
 Open Dynamics Engine (ODE) (C++)

• Symbolic manipulators:
 3D_MEC
 EasyDyn (MuPad)
 RoboTran (requires Matlab) frequently architectures overlap

• Modelica:
 OpenModelica?

LaMSID – July 24 – August 7, 2008
 18
Software architectures
• MBDyn is monolithic

• Input consists in a text file

• The input syntax allows some flexibility, e.g.:

 math expressions evaluation
 variables definition
 “rigorous” syntax checking, but free style, indentation, ...

• Relevant portions of the code are modular and can be extended by:
 writing run-time loadable modules
 hacking the code (it's free, all in all!)

• There is no built-in pre-post processing facility

• Help in this area would be warmly appreciated!

 MBDyn output can be translated into EasyAnim
 there is an independent, partial customization based on Blender

 what about a custom Salome toolbox for multibody?

LaMSID – July 24 – August 7, 2008
 19
Software architectures
MBDyn interacts with other software:
• Generic streamed input/output
 Input: “drive” (scalar); can be fed to several entities:
• imposed load
• imposed joint motion
• imposed deformable component prestrain
• Inputs can be aggregated and manipulated
 Output: “measurement” (scalar); can be any exportable entity:
• any model state
• internal parameters of elements and nodes
• outputs can be aggregated and manipulated
 Used to interact with Simulink, Scicos, ... also in hard real-time

• Generic motion output/force input

 Used for interaction with CFD solvers
 Could be (easily?) adapted for interaction with Aster?

• Other software needed to pre/post-process data (e.g. FEM)

LaMSID – July 24 – August 7, 2008
 20
Problems

Equations of motion: for each node (purely geometrical entity),

• Newton-Euler, written as first-order system of equations:
M ẋ= p (for dynamic nodes only)
ṗ= f x , ẋ , t  (=0 for static nodes)
• Momentum and momenta moment are used
instead of pseudo-velocities

• allows multiple contributions to inertia of a single node

(but constraint stabilization is no longer straightforward...)

Constrained equations are directly written

in differential-algebraic form:
M ẋ= p
T
ṗ / x = f x , ẋ , t
x , t= 0

LaMSID – July 24 – August 7, 2008
 21
Problems

• Fundamental problem:
 integration of Initial Value Problem (IVP) in time
 also in Hard Real-Time (by way of RTAI, RT-Net & Scicos)
for Hardware-In-the-Loop (HIL) simulations

• Static analysis as degeneration of IVP dynamic analysis:

 momentum and momenta moment definitions omitted
 only gravity is considered
 system determination must be provided by kinematic constraints
and deformable components

• Kinematic analysis as degeneration of IVP dynamic analysis:

 inertia elements omitted
 system determination must be provided by kinematic constraints
 deformable components can act as “regularization”

LaMSID – July 24 – August 7, 2008
 22
Problems
K '=I : Moore-Penrose pseudo-inverse

Experimental inverse dynamics problem

• inverse kinematics

[ ]{ } { }
T
(allows under- K ' / x  x 0
constrained = nonlinear
systems):  /x 0  0 −x , t

[ ]{ } { }
T
K '  / x  ẋ 0
same matrix! = linear
 /x 0 1 b x , t 

[ ]{ } { }
T
K '  / x  ẍ 0
= linear
 /x 0 2 b ' x , ẋ , t

• the RHS contains the desired motion and its derivatives

• the (regularized) static analysis provides the kinematic inversion
T/ x =f ' x , ẋ , ẍ , t  linear

LaMSID – July 24 – August 7, 2008
 23
Problems

• Direct eigenanalysis (eXperimental)

 issues with constraints formulation
(mainly with rotations)

 issues with equations implementation

(matrices not available for some elements; e.g. aerodynamics)

• Future development: relative coordinate frame dynamics

 imposed frame motion: modifications only to RHS inertia elems

 instrumental for many helicopter rotor/wind turbine

dynamics problems

LaMSID – July 24 – August 7, 2008
 24
Arbitrary motion description

• Mechanical degrees of freedom:

 structural node positions in the absolute reference frame
 structural node orientations with respect to the absolute frame
(but in updated form...)

• Kinematics is always written with respect to the absolute frame

• Newton-Euler equations are written in the absolute frame

 moment equilibrium (Euler) equations are written
with respect to the (moving) nodes

• Special elements may introduce further approximations

 e.g. Component Mode Synthesis (CMS) element

LaMSID – July 24 – August 7, 2008
 25
Arbitrary motion description

Orientation handling:
• orientation variables: Cayley-Gibbs-Rodrigues parameters
• orientation matrix:
R=R g
• orthonormality:
R T =R−1
• derivative:
Ṙ R T =×=G g ġ×

• incremental approach from step k to k+1 to eliminate the

singularity issue that arises from orientation parameters
(increments must be small anyway for accuracy):
R k =R g k R k−1

LaMSID – July 24 – August 7, 2008
 26
Arbitrary motion description

• Orientation handling:
 the actual orientation variables are the Cayley-Gibbs-Rodrigues
parameters relative to the correction phase of each time step
 k: time step counter
 i: correction iteration counter (0: predicted value)

“updated-updated” (0)
lagrangian k

k-1 k(i)
updated lagrangian
• Orientation matrix:
R ki=R g i

 R 0 
k

• Derivative:
Ṙ ki R ki T=ki×=R  gi

  0
k
×G g i

 ġ  i

×

LaMSID – July 24 – August 7, 2008
 27
Arbitrary motion description

• Incremental orientation from previous step:

 Orientation parameters order of magnitude:

g~O∥∥ t
• Incremental orientation from prediction:
 Orientation parameters order of magnitude:

g ~O  t n1 

where n is the min between the order of the predictor and that
of the integration method (MBDyn: 3 and 2, respectively)

 As a consequence: R g  ≈I
G g ≈ I
(only in Jacobian
matrix, of course!) Ġ  g ≈0

LaMSID – July 24 – August 7, 2008
 28
Deformable components

• Lumped deformable components

 rod (1D)
 linear, angular components (3D)
 linear & angular component (6D)

• Intrinsic, composite-ready Finite-Volume beam element

 arbitrary constitutive law
 piezoelectric constitutive law
 aerodynamic (strip-theory with inflow model) beam element

• Component Mode Synthesis (CMS)

 attached to a floating frame (a multibody node) for rigid body motion
 linear state-space representation of unsteady aerodynamics

ẋ= A xB q s
f a= qC xD0 q s b/V D1 q̇ s  b/V 2 D2 q̈ s 

LaMSID – July 24 – August 7, 2008
 29
Deformable components

Lumped deformable components (3D, 6D):

T
= ax  exp  R a R b 
−1

m= R   m
 
• Attached form:
 constitutive properties referred to either of the connected nodes

=0, =1

• Intrinsic form (invariant: =1/2 ):

 constitutive properties referred to a floating reference frame
 intrinsically handles geometrical nonlinearity related to rotations
 correctly captures bending-torsion buckling behavior
 essential for anisotropic deformable components

The intrinsic/invariant form is unique to MBDyn.

It resulted from a work for Hutchinson CdR

LaMSID – July 24 – August 7, 2008
 30
Deformable components

Intrinsic, composite-ready beam

• Topology:
 1D reference line p, 1D reference structure R
 2D section characterization

reference motion: p , R
reference line

reference
orientation
p R

x= pR t
t
x / = p /×R t

warping

LaMSID – July 24 – August 7, 2008
 31
Deformable components

Intrinsic, composite-ready beams

• strain measure:
T T
= R p /−R 0 p 0 /
= R T  −R T0  0

• equilibrium (from VWP):

f /= 
m / p /×f = 

• constitutive properties:

f = f , 
m= m  , 

LaMSID – July 24 – August 7, 2008
 32
Deformable components

Intrinsic, composite-ready beams: 3-node discretization

• Finite Volume approach: equilibrium of finite portions of beam
• internal forces function of node kinematics thru constitutive laws
• warping goes into constitutive properties computation
f II , m II f 3 , m3
f 2 , m2
f I , mI

f 1 , m1
point II
point I
node 3

node 2

node 1 constitutive properties from detailed 2D FEM

LaMSID – July 24 – August 7, 2008
 33
Solving the problem

Numerical integration
• implicit, (quasi-)L stable 2 step algorithm
y k =a 1 y k−1 a 2 y k−2  t b 0 ẏ k b 1 ẏ k−1 b2 ẏ k−2 

• tunable algorithmic dissipation: asymptotic spectral radius 1→0

 asymptotic spectral radius = 0: 2nd order BDF
 “optimal” dissipation: spectral radius ~ 0.6

• second-order accurate, with third-order accurate predictor

• variable time step
• not ideal for non-smooth problems (multi-step)
• different integrators can be used; new ones can be implemented

LaMSID – July 24 – August 7, 2008
 34
Solving the problem

• Prediction:
ẏ0k =m 1 y k−1m 2 y k−2 / tn 1 ẏ k−1n 2 ẏ k−2
y0k =a1 y k−1a2 y k−2  t b 0 ẏk0b 1 ẏ k−1 b 2 ẏ k−2 

• Correction iteration:
f / ẏ  ẏ if /y  yi =−f  ẏki−1 , yki−1 , t k 
but
 y i = t b 0  ẏ  i

the problem becomes algebraic

 f / ẏ  t b 0 f / y  ẏ i=−f  ẏi−1
k
, y i−1
k
, tk
ẏki= ẏki−1 ẏi
 i  i−1  i
y k =y k  t b0  ẏ

LaMSID – July 24 – August 7, 2008
 35
Solving the problem

Model assembly
• model could be input incorrectly
• initial values of the state (position, velocity, reactions) are needed
• this task might not be trivial
• initial state values must comply with constraints:
 x 0, t 0 =0
̇ x 0, t 0 =0
• a dummy static nonlinear problem is solved (regularization):
T
K ' x−x 0  /x  '= f '
C ' ẋ− ẋ0  T/x '= ḟ '
x , t 0 = 0
̇x , t 0 = 0

LaMSID – July 24 – August 7, 2008
 36
Solving the problem

Solution initialization (so-called “derivatives”)

• explicit problem:
ẏ=f y , t 

• implicit problem:
0=f  y , ẏ , t 

• modified correction phase to initialize solution:

f / ẏ c f /y   ẏ i=−f  ẏ 0i−1 , y 0 , t 0 
ẏ0i= ẏ0i−1 ẏi
y 0 =y 0

• convergence no longer quadratic, but saves lots of code duplication

• Setting c=0 might not work (problem can be structurally singular)

LaMSID – July 24 – August 7, 2008
 37
Extracting useful information

• Detailed analysis requires detailed models, but...

• excessive details endanger the chance to extract useful information

• Proper Orthogonal Decomposition allows to extract information

from redundant measures

• Consider a set of N measurements X for n time steps; their SVD:

XT  ∈ℝ n× N =U  V T

• The singular values allow to determine the m most relevant signals

XT1: m, n=U n , 1: m  1 : m, 1: m V TN , 1 : m

• Note that T 2 T T 2 T
X X= U  U XX = V V
U T XT =  V T XT V= U 
• This allows to efficiently compute the singular values and the POMs

LaMSID – July 24 – August 7, 2008
 38
Extracting useful information

• The POMs can be used to identify a transition matrix

X  k1= X k
• If X contains the free response of the system, the transition matrix
allows to estimate the relevant eigenvalues (AR model)

• More sophisticated system identification techniques can be used

• A technique based on covariance estimates from time histories has

been recently proposed; works for:
 free response
 forced response
 unmeasured forced response

[1] G. Quaranta, P. Masarati, and P. Mantegazza: “Continuous-Time Covariance Approaches for

Modal Analysis”, Journal of Sound and Vibration, Vol.310/1-2 pp. 287-312, 5 February 2008.
[2] G. Quaranta, P. Masarati, and P. Mantegazza: “Assessing the Local Stability of Periodic
Motions for Large Multibody Nonlinear Systems Using POD”, Journal of Sound and
Vibration, Vol 271/3-5, pp. 1015-1038, 2004.

LaMSID – July 24 – August 7, 2008
 39
Examples of multibody modeling with MBDyn

Robotics:
delta robot

inverse dynamics
for computed
torque control

LaMSID – July 24 – August 7, 2008
 40
Examples of multibody modeling with MBDyn

Robotics: PA-10
inverse kinematics
with path optimization
of cooperating robots

LaMSID – July 24 – August 7, 2008
 41
Examples of multibody modeling with MBDyn

Robotics:
biomimetic robot
real-time motion planning
by inverse kinematics
with fault detection

LaMSID – July 24 – August 7, 2008
 42
Examples of multibody modeling with MBDyn

Industrial processes:

• simulation of automotive components assembly (car brake pipe) to:

 check stresses introduced during assembly
 check loads on supports introduced during assembly
 check interference with other parts during assembly
 check interference with other parts during operation

• the model has been developed by a rubber manufacturer

(Hutchinson CdR)

• it is used for product design and certification

• it required the development of specific features for solution

partitioning, which are now part of the standard MBDyn (the “hints”)

LaMSID – July 24 – August 7, 2008
 43
Examples of multibody modeling with MBDyn

Automotive: mechanical modeling of suspensions

purpose: determine loads in rubber bushings and other components

LaMSID – July 24 – August 7, 2008
 44
Examples of multibody modeling with MBDyn

Rotorcraft dynamics and aeroservoelasticity:

• WRATS (NASA/Army) tiltrotor aeromechanics

LaMSID – July 24 – August 7, 2008
 45
Examples of multibody modeling with MBDyn

Rotorcraft dynamics
and aeroservoelasticity:
• ERICA (AgustaWestland)
tiltrotor aeromechanics
(ADYN, NICETRIP EU 6FP,
in cooperation with
Eurocopter, DLR, ONERA
and more...)

LaMSID – July 24 – August 7, 2008
 46
Examples of multibody modeling with MBDyn

Wind-turbine dynamics simulation,

also in Real-Time
blades

gearbox
nacelle
tower
yaw teeter

pitch

LaMSID – July 24 – August 7, 2008
 47
Future development

• Multiscale handling of submodels with different dynamics;

for example, in the flying helicopter case:
 aircraft flight mechanics (~0 to 5 Hz: very slow)
 airframe & main rotor dynamics (~5 to 40 Hz: intermediate)
 tail rotor dynamics (~25 to >100 Hz: fast)

• Interfacing with different domains

 fluid-structure (Lagrangian/Eulerian modeling of work flows)
 structure-structure (dynamic loads to detailed static analysis)
 active control of large deformable systems

• Better abstraction/modularization of components/solution phases

 more freedom in model customization
 tight integration into nonlinear structural analysis
(structural damage growth under dynamic loads?)

• More...

LaMSID – July 24 – August 7, 2008
 48
Future development

• Development has always been problem (and customer!) driven

• Development model:
 research (possibly new) solutions in (possibly new) areas
 procure research grants and contracts in those fields
 never ask funds just for software development
 sell expertise, not just software; as a consequence...
 ...software is not the result of a contract: it's a pre-existing tool
 software remains free even when customers are narrow-minded
 when possible, work with open-minded customers :-)

• Drawback: cannot stick with tight development plans :-(

We are a University: we should not sacrifice our research freedom

to industry schedules

LaMSID – July 24 – August 7, 2008
 49
Documentation and support

• Freedom and features are not enough: software needs to be usable

• Users need:
 documentation
 forums to discuss issues
 issue tracking provisions
 reasonable guarantee of maintenance
 reasonable guarantee of stability across releases

• The main guarantee is the freedom of the software:

 in the worst case, one can always fork and go on its own; but...
 ... first, better talk to the developers: cooperation saves efforts

LaMSID – July 24 – August 7, 2008
 50
Documentation and support

• Theory manual:
 Incomplete; needs lots of work

• User manual
 available and complete, but probably hard to read; needs improvements

• Tutorials
 available, but reportedly too simple; need work

• Applications manual
 just started

• Installation manual
 available, but incomplete and outdated

• Mailing lists
 available: announce, users, devel
 right now, the “users” list also serves as issue tracking provision

LaMSID – July 24 – August 7, 2008
 51
Documentation and support

• Another important item that is missing is an automated test suite, that

 can be run automatically after building the software
 allows to check build errors
 allows to check regressions in new releases
 serves as example of modeling functionalities

• The rest is underway (always a work in progress)

Given the nature of the project, contributions are always welcome!

LaMSID – July 24 – August 7, 2008
 52
Acknowledgments

• Paolo Mantegazza (“grandfather” of MBDyn)

• Marco Morandini (lots of optimizations, friction, ...)
• Giuseppe Quaranta (schur decomposition, FSI, ...)
• Alessandro Fumagalli (inverse dynamics, ...)

• many Graduate and PhD students

• few contributions from outside as well (mostly bug fixes)

LaMSID – July 24 – August 7, 2008
 53

Questions?

LaMSID – July 24 – August 7, 2008