Journal of Intelligent and Robotic Systems 17: 81-99, 1996.

82
@ 1996 Kluwer Academic Publishers. Printed in the Netherlands.

Dynamic Simulation and Neural Network

Compliance Control of an Intelligent Forging
Center

K. W. LILLY
The Pennsylvania State University, Department of Mechanical Engineering, University Park,
PA 16802, U.S.A.

A. S. M E L L I G E R I
StatDesign, Inc., 3319 Greenfield Rd. # 197, Dearborn, MI 48120, U.S.A.

(Received: 27 March 1995; in revised form: 31 October 1995)

Abstract. Automation of forging processes is important for both safety and efficiency in today's
advanced manufacturing operations. This work supports the development of an Intelligent Open Die
Forging System which will integrate state-of-the-art modelling techniques, automatic die selection
and sequencing, full system dynamic simulation, automatic machine programming and coordi-
nation, and sensor-based process control to enable the production of more general and complex
workpiece geometries than are achievable using current forging methods. Effective automation of
this open die forging system requires the coordination and control of the major system components:
press, robot, and furnace. In particular, forces exerted on the robot through its manipulation of the
workpiece during forging must be minimized to avoid damage to the manipulator mechanism. In
this paper, the application of neural networks for compliance control of the forging robot to min-
imize these forces is investigated. Effectiveness of the neural network-based compliance control
module is evaluated through a full dynamic system simulation, which will later form a central part
of the complete Intelligent Forging System. Dynamic simulation of the robot is achieved using an
efficient O(N) recursive algorithm, while material flow of the workpiece is modeled with a finite
element approach. Simulation and timing results for the complete processing system for a specific
open die forging example are presented.
Key words: forging automation, compliance control, neural networks, dynamic simulation, finite
element models, recursive algorithms, intelligent manufacturing.

1. I n t r o d u c t i o n

Forging is recognized as one of industry's most arduous tasks, combining inhos-

pitable working conditions with the potential danger involved in using heavy and
possibly fast-moving presses to deform sizable pieces of hot metal. Even with
intensive safety procedures and constant maintenance, forge operators are con-
tinually at risk, making the automation of forging processes an important goal.
In addition to improved safety, automation of the forging process will contribute
to the current m o v e m e n t towards environmentally conscious manufacturing in
today's world market. Intelligent integration of advanced technologies such as
82 K. W, LILLY AND A, S. MELLIGERI

3D modeling techniques, automatic die sequencing, robotic workpiece manipula-

tion, automatic machine programming and coordination, and sensor-based process
control in an agile forging system will enable the efficient production of complex
part geometries with minimal material waste [18]. This, in tum, will lead to the
increased use of 'as-forged' parts which require little or no additional machin-
ing or other processing to achieve final finished form, thereby conserving time,
money, and important natural resources.

1.1. BACKGROUND
This paper discusses the automation of open die or 'hand' forging, which makes
use of simple dies (typically flat or V-shaped) to progressively forge a specified
part geometry (see Fig. 1). Open die forging differs from impression or closed die
forging in that no patterns or impressions are 'sunk' (machined) into the top and
bottom dies. Like other types of forging processes, open die forging enables the
control of grain flow in the forged part to optimize its material property profiles
and increase its service life over a cast, welded, or machined part [9].
The advantages of open die forging over alternative forming processes are
evident: (1) for small quantities (reduced batch processing); (2) for large part
sizes; and (3) when desired property requirements can only be achieved by forg-
ing. Open die forging is ideal for producing small numbers of parts, often more
economically than casting, machining, or welding. Open-die-forged prototypes
may also be used to optimize the geometric, material property, or functional
characteristics of a part before large investments are made in costly closed die
tooling which will eventually be used to form the final product.
In other cases, a large part size dictates that only open die forging can be used
to manufacture a part with the desirable characteristics of forging. In general,
open die forgings can be produced much larger than closed die forgings, with
the largest open die-forged shafts weighting more than 200 tons [9]. Open die

5
For
Mm

Figure 1. Simple open die forging process example.

SIMULATION AND CONTROL FOR INTELLIGENT FORGING 83

forging is also typically chosen over alternative non-forging processes when the
desired part must withstand high repeated cyclical stresses. Other requirements
which dictate open die forging include the need for increased property values in
the transverse directions and/or the need for better torsional fatigue resistance.
Regardless of part size, open die forging is frequently the only process capable
of providing desired design properties and service life expectancies,
Although automation of the open die forging process is not unknown to date,
its scope has been limited thus far. The first use of manipulators in industrial
open die forging occured approximately 60 years ago in both the United States
and Europe. The development of manipulators for handling heavy workpieces
(e.g. 200+ tons) was initiated at that time and has continued to some extent
through today [33, 34]. Such manipulators are generally custom built for a specific
task, however, and are limited in their manipulative capacity. Typical forging
manipulators have three degrees of freedom or less, a fact which significantly
restricts their end-point motion, and hence, the complexity of the resulting forged
part geometries.
Effective use of more versatile and dexterous industrial robots in the open
die forging of more general and complex workpiece geometries has not been
fully investigated, although some preliminary results are presented by Appleton
[2, 3]. A study by Vitscheff [29] also discusses the potential application of indus-
trial robots in the forging process with proper end effector design. In addition,
Vitscheff demonstrates the need for compliance control when robots are used to
manipulate the workpiece during forging. Specifically, the external forces exert-
ed on the manipulator through its manipulation of the workpiece during forging
steps must be minimized to avoid damage to the robot mechanism. In Apple-
ton's later work [3], passive compliance is provided in the form of springs in the
end effector mechanism. Application of this passive approach is limited, how-
ever, based upon the type of workpiece being handled and the specific forming
operation under consideration.
In general, few results have been published regarding integrated systems for
automating the open die forging process until very recently. In [18], Lilly and
Jablokow present the conceptual design of a next generation Intelligent Open Die
Forging System which integrates state-of-the-art modelling techniques, automatic
die selection and sequencing, full dynamic system simulation, automatic machine
programming and coordination, as well as sensor-based process control. Earlier
supporting work in the area of dynamic simulation of a hardware-integrated Forg-
ing Center is discussed in [ 19], while relevant results in the areas of 3D modelling
and feature extraction are referenced in [18]. Preliminary control studies, includ-
ing the consideration of neural network-based compliance control in forging, are
discussed in [1, 22].
Effective automation of the open die forging process requires the close coordi-
nation and control of the major system components: press, robot, and furnace (for
hot forging). As with many advanced manufacturing applications, the need for
84 K. W. LILLY AND A. S. MELLIGERI

accurate simulation of this automated integrated processing system before imple-

mentation on the shop floor is apparent. Real-time simulation of a fully integrated
robot/forge processing system has not yet been accomplished, however, due in
part to the lack of appropriate 3D simulation tools for the forging process itself.
The computational complexity of most dynamic simulation algorithms for indus-
trial robots also makes real-time implementation difficult. Recent developments
in these two areas, along with the advent and use of neural networks in system
control, have raised expectations for more promising results in the near future.
New developments in the efficient simulation of automated forging systems are
a significant contribution of this work.

1.2. PREVIOUS AND SUPPORTING WORK

This paper presents results obtained as part of the on-going development and
testing of the Integrated Robot/Forge Processing Center in the Department of
Mechanical Engineering at The Pennsylvania State University. The fundamental
objective of this research effort is the realization of an automated Intelligent Open
Die Forging System, as described briefly above, which exploits the advantages
of modern industrial robots and current computer technology. In contrast to most
standard forging applications which restrict the use of robots to pick-and-place
operations, the Penn State Integrated Robot/Forge Processing Center is designed
to utilize an industrial robot to manipulate and directly support the workpiece
while it is being forged. This will enable the forming of more general and complex
workpiece geometries, which in turn will eliminate or reduce the labor of later
production steps and lead to more efficient use of limited resources.
To aid in the design, implementation, and testing of the advanced coordina-
tion and control algorithms required for the Integrated Robot/Forge Processing
Center, a general dynamic model of an industrial robot and its interaction with
a workpiece during the open die forging process was introduced in [19]. The
current paper presents a more detailed discussion of the full dynamic system
simulation scheme developed for task validation and forging cell management
and its application in the investigation of neural network-based compliance con-
trol for simple open die forging tasks. Relevant to the development of this full
system simulation is the consideration of available tools and results in the areas of
forging process modelling, dynamic simulation algorithms for industrial robots,
and compliance control.
The use of finite element analysis techniques for the modelling of metal
forming processes in recent years is well established [1 l, 14, 16]. Non-linear
finite element analysis tools have been successfully developed and applied with
the increased computational capability of modern computers. However, although
new remeshing schemes and 3-dimensional analyses are expanding finite ele-
ment applications, there is no previous record of researchers utilizing material
SIMULATION AND CONTROL FOR INTELLIGENT FORGING 85

flow simulation techniques in the design of better workpiece handling methods

as is proposed here (see also [18]).
Dynamic simulation of a robot manipulating a workpiece in an open die forg-
ing application requires both open and closed chain Direct Dynamics algorithms,
consistent with the robot's instantaneous kinematic configuration. In general, a
Direct Dynamics algorithm determines the motion of the robot joints, given the
driving actuator torques/forces and any external torques and forces exerted on the
robot. Effective open chain Direct Dynamics algorithms include those presented
by Walker and Orin [31], Featherstone [8], and Brandl et al. [5]. The general-
ized open chain algorithm given in [5] is the most efficient known at this time.
Closed chain Direct Dynamics algorithms include the work of Orin and McGhee
[23] and the linear operator methods of Rodriguez et al. [26]. The most efficient
Direct Dynamics algorithm for single closed chain manipulators currently known,
as presented by Lilly and Orin in [20], is used in conjunction with the algorithm
of Brandl et al. [5] for the dynamic simulation of the industrial forging robot in
this paper.
Compliance control has attracted considerable attention in the past decade,
with research efforts focussed on many different applications. While a compre-
hensive review of the literature is beyond the scope of this paper, some represen-
tative publications include [t2, 13, 21, 24, 30]. In recent years, the implementa-
tion of compliance control using dedicated neural networks has received special
attention. Asada has shown that a trained neural network can be used for both
linear and non-linear compliance control [4]. Other examples of neural networks
in compliance control include the work of Cha [6], Connolly [7], Kwan [15], and
Venkataraman et al. [28], however, none of these applications consider the use
of compliance control in automated forging processes. Evaluating the feasibility
and performance of a neural network-based compliance controller within an inte-
grated open die forging application is a primary objective of the work presented
here.

2. Dynamic Simulation Algorithm of the Industrial Robot

For the dynamic simulation of an industrial forging robot, both open chain and
closed chain Direct Dynamics solutions are required. When the simulated robot
end effector moves in an unconstrained manner in its workspace, an open chain
Direct Dynamics algorithm is sufficient. In this research effort, the efficient solu-
tion method of Brandl et al. [5] is used for this task. When the motion of the
simulated robot end effector is constrained through its interaction with the work-
piece as it is forged, a closed chain Direct Dynamics algorithm is required. The
efficient closed chain simulation algorithm of Lilly and Orin [20] is used to
model the constrained motion of the robot in this work.
This section briefly describes the efficient dynamic simulation of a single
closed chain robot with N degrees of freedom. The fundamental equations which
86 K.W. LILLY AND A. S. MELLIGERI

define the dynamic behavior of a robot may be expressed using joint space
variables as follows [17, 31]:

T = H(q). ~ + C(q,~) + G(q) + j(q)T, f (1)

where ~- is an N • 1 vector of joint actuator torques/forces; q, q and ~ are N • 1

vectors of joint positions, velocities, and accelerations; C(q, (t) is an N • 1 vector
of Coriolis and centripetal force terms; G(q) is an N • 1 vector of gravitational
forces; J(q) is the 6 • N Jacobian matrix; f is a 6 • 1 vector of external
forces/moments exerted on the robot gripper; and H(q) is the N • N robot joint
space inertia matrix. This vector equation may be used to describe the dynamics
of both open and closed chain configurations. In general, the vector of external
forces and moments ( f ) is unknown for a closed chain system, while it is zero or
known for an open chain system. This vector f must be determined along with/i
for a full closed chain dynamic system simulation. In the current application, the
robot will be constrained in a closed chain configuration whenever it supports
and/or manipulates the workpiece in contact with the forging press.
A brief outline of the efficient single closed chain dynamic simulation algo-
rithm of [20] is now provided. Using Equation (1), we may solve for the vector
of joint accelerations for the robot,/], as follows:
= H - l ( 7 - - C - G) - (H-1jT) 9f (2)

or:

qclosed = qopen -- ~Jconstrained (3)

where qopen is the vector of joint accelerations for the robot in an unconstrained,
open chain configuration with all external contact forces removed. The second
term on the right side of this equation, qconstrained, represents a correction term
which accounts for the closed chain contact. It is a function of the contact force
vector, f. By defining:
( H - I j T ) = f~ (4)

we may rewrite Equation (3) as follows:

qclosed = qopen -- ( ~ " f ) (5)

The N • 6 matrix, f~, is a function of the joint positions, q.
Similarly, by utilizing the operational space or Cartesian workspace formu-
lation of the robot dynamics equations, we may partition the closed chain accel-
eration vector of the robot end effector as follows [20]:

Xdosed = Xopen -- ( A - 1 . f ) (6)

where
( J H - 1jT) = A - l (7)
SIMULATION AND CONTROL FOR INTELLIGENT FORGING 87

The 6 x 6 matrix A is the robot operational space inertia matrix (OSIM). It

represents the combined inertia of the articulated manipulator links projected to
the robot end effector or tip. Efficient O(N) algorithms for the computation of
f~ and A are presented in [17] and [26].
To complete the closed chain simulation algorithm, the partitioned dynamic
equations [Eqs (5) and (6)] are combined with a constraint model of the contact
between the robot/workpiece and the environment to compute the contact force
vector, f . The contact model is based upon the general joint model presented
in [25]. Its application is further discussed in [17], [19], and [20]. For this open
die forging application, the contact model defines a rigid connection between the
robot end effector and the workpiece. Further, the contact between the workpiece
and the press dies is also modelled as a rigid connection during forming. The
contact force vector is determined by linear system solution of an equation of
the following form:

(A-I . f ) = Xopen -- ffc (8)

where 9 c represents the vector of end effector accelerations in the constrained

directions of the contact. Note that solution of Equation (8) requires an equivalent
matrix inverse which is only of dimension 6 x 6. Once the contact force vector
has been determined, it is used in Equation (5) along with the open chain joint
acceleration vector and ~ to calculate the closed chain joint accelerations, qdosed.
These closed chain joint accelerations are then integrated to find the next state
joint rates and positions for the robot.
Given the current state of the manipulator, q and 0, and the input joint actuator
torques/forces, T, the open chain acceleration t e r m s qopen and Xopen may be com-
puted. Any appropriate Direct Dynamic algorithm for open chain manipulators
may be employed for this computation [5, 8, 31]. The recursive algorithm of
Brandl et al. [5] may be used to compute these terms most efficiently, and it is
utilized in this work. A recently developed recursive O(N) algorithm, the Force
Propagation Method [17], is used for the efficient computation of ~2 and A -1
The complete algorithm for a serial N-link manipulator is presented in Table I
for illustration and reference. Note the following definitions for terms used in
this algorithm [8, 17, 25]:
q~i = (6 x 1) spatial axis of motion for joint i
iX~_l = (6 x 6) spatial coordinate transformation between frames i and i - 1
Ii = (6 x 6) spatial inertia matrix for link i
I* = (6 x 6) articulated-body inertia matrix for link i
K [ = (6 x 6) projected inverse inertia matrix for link i
L~ = (6 x 6) spatial acceleration propagator matrix for link i
D~ = (6 x 6) transformed acceleration propagator for link i
1 = (6 x 6) identity matrix
88 K.W. LILLY AND A. S. MELLIGERI

Table I. Force propagation algorithm for computation of A-~ and fl

Given: (Initialization)
d?i,Ii,iXi_l for i = 1,...,N;
I ~ = I N ; D*N = 1
K*N = r N ( r Nr N)-1r
L~v = 1 - (K~v" I~)
(,,&~)- 1 = 0

Backward Recursion
For i = ( N - 1 ) , . . . , 1:
I ; = Ii + i + ' X ~ r ( I g + t " L * + , ) i + ' X i
K~" = r ( cT I* r -t r
L; = 1 - ( K ; . I~, )
* i+lyTgr* XD*
Oi = -'~-i k.Ui+l] i+l

Forward Recursion
For i = 1,...,N:
('A*)-' = K , . + ( L : ) [ i X i - i (i-IA.
k i-l)
~-I i T
9 Xi-ll(Li
*)T
A-I - ( N A ~ ) - t
[ . ~ T I . , t ~'~-I~.Tf 1 t i * ~ i X ti-l^* ",-1 i x T (L*~T]D*
~r = k W l iY.'i] ~)i[ --k i ) " i--l'k lxi--1) " i--It i] J i
~-- [~'-~(l) ~(2) " ' ' •(N)] T

The structure of the algorithm is a double recursion. In the first recursion,

I/*, the articulated-body inertia of each link, is calculated starting at the robot
tip or last link and moving back to the first link (Backward Recursion). The
second recursion begins at link 1 (near the base) and moves out to the last link
(Forward Recursion), computing intermediate inverse operational space inertia
matrices (A*) -1 and components of f~ at each link. When this second recursion
reaches link N, A - l and f~ are known. For computation of the contact force
vector, f , nodal velocities from the finite element simulation of the forging pro-
cess are differentiated to obtain the end effector accelerations in the constrained
directions of the contact (9c). These are used in Equation (8), along with ~open
and A -1, and a solution for f is found using a linear system solver based on
Gaussian elimination. This solution for f is then used along with q~ and f~
in Equation (5) to complete computation of the closed chain joint accelerations,
/1closed. Finally, the closed chain joint accelerations are integrated to obtain the
next state joint rates and positions using a standard Euler integration routine.

3. Finite Element Modeling of the Forging Process

This section presents some basic concepts and features of the rigid viscoplastic
finite element analysis method used to model the open die forging process in this
work. The present discussion is based on the finite element method developed
SIMULATION AND CONTROL FOR INTELLIGENT FORGING 89

by Lee and Kobayashi [16]. This method has been used successfully to model
a variety of metal forming processes, including forging and extrusion [14]. The
formulation is briefly described below, with additional details found in [11].
The fundamental law for viscoplastic materials relates the strain rate and
deviatoric stress as follows [11]:

iij---~ ff aij (9)

where
~j = strain rate tensor
= effective strain-rate magnitude
= effective flow stress magnitude
I
cri~ = deviatoric stress tensor.
The integral form of the equilibrium equation during the deformation of a
rigid-plastic material may be written:

n=/vaadV-fsFiuidS I
(10)

where

st= surface on which material undergoes plastic deformation

under prescribed surface traction
= surface on which material undergoes plastic deformation
under prescribed velocities
S
H= functional of power
V= prescribed volume
traction component
Ui velocity component.
The first order variation of the functional H is:

5H=/v~(5~)dV-/s FiSuidS = O (11)

l

Metal deforms with constant volume. Therefore, the admissible velocity ui must
satisfy the incompressibility constraint. The constraints are imbedded in the vari-
ational form by introducing a penalty constant, Q. The modified first-order vari-
ation of the functional H is given as follows:

5 H = { , 6-(55)dV+ Q {, ~ , ( ~ , ) d V - {~ Fi'ui d S = O (12)

JV dV Y
90 K.W. LILLY AND A. S. MELLIGERI

where

gv = volumetric strain-rate magnitude.

At the interface of the die and the workpiece preform, the velocity boundary
condition is given by the die velocity in the direction normal to the interface.
The traction boundary condition, which is friction force, is represented using a
constant friction coefficient, slipping velocity, and yield stress. This relationship
is given as follows:

fs=mkl--mk( 2 ) tan-l(lusl]Ikuo/ (13)

where

fs = frictional stress
m = friction factor
k = shear yield stress magnitude
l = unit tangent vector
Us = relative sliding velocity
Uo = velocity constant (for numerical stability).

The matrix representation of the goveming equation for rigid viscoplasticity is:

KV = F (14)

where K is the material and process-dependent nonlinear stiffness matrix, V

is the preform nodal velocity vector, and F is the traction vector. The solution
to Equation (14) is obtained iteratively, with specific expressions for K and F
appearing as follows [11]:

OCCj dr) (15)

and

F = Ov,i list = ~ / s c tan-1 '

where

lip = penalty term in II

1-ID = deformation energy-rate term in II
1-isc = friction term in I-I
Pij = effective strain-rate tensor
SIMULATION AND CONTROL FOR INTELLIGENT FORGING 91

C~ = function of shape function derivative matrix

ui : nodal velocity component
Sc = tool-workpiece contact surface
qj = shape function.

A simulation approach based on the equations shown above was implemented

for a simple open die forging operation. Only fiat dies were used, and heat transfer
was not considered. Simulation inputs were die velocity, friction factor, flow
stress (as a function of strain rate), initial billet velocity, and total reduction in the
billet height. The outputs were nodal forces, displacements, and velocities. The
nodal velocities were differentiated to obtain the acceleration of the workpiece
at the contact points with the robot end effector. This information was then used
to solve for the contact force vector, as described in Section 2.

4. Neural Network Compliance Control

In recent years, active compliance control has received considerable attention
due to its flexibility of application for different processes ranging from assembly
operations to machining [12, 13, 21, 24, 30]. Active compliance control utilizes a
linear or nonlinear mapping from measured forces to corrected motion (typically
displacement and/or velocity) to control system behavior. For linear mappings, a
single matrix representation may be used to describe the force/motion relation-
ship. When a nonlinear mapping is required, multiple compliance matrices may
be specified along with a rule-based system to determine which matrix should be
used based on the process state. While such an approach is generally quite flex-
ible, it requires some kind of in-process monitoring to recognize the occurrence
of different process states.
For certain applications, the exact nonlinear mapping is known through mod-
eling of the process itself (e.g. a finite element model). However, it may not be
possible to directly incorporate such a model in the compliance control scheme
and continue to satisfy real-time computation constraints. In such a situation, a
metamodel which closely represents the exact nonlinear model and which can
be implemented in real time is required. Neural networks provide one means
to effectively realize such a metamodel. The representation and generation of
complex nonlinear compliance mappings using neural networks is investigated
by Asada in [4]. Through his and other related efforts, it has been proven that
multi-layer neural networks are capable of representing approximately any con-
tinuous nonlinear function on a compact set [4, 10]. The generalized delta rule
[27], which computes the weights for each neuron based on the difference in
the actual and desired output of training sets, can be used to train the network
to provide the desired output for a given input. Once the neural network has
been trained, it is used to replace the original analytical process model [10, 32].
92 K.W. LILLYAND A. S, MELLIGERI

Since all the elements in one layer of a neural network compute in parallel, the
total computing time is a function of the relatively small number of layers rather
than the large number of processors. This greatly reduces the time required to
compute the output of the model.
The neural network approach demonstrated in [4] is used here to generate
the complex nonlinear compliance mapping required to accurately represent the
open die forging process in a continuous manner. Details and results of the
neural network training procedure are now briefly discussed. Integration of the
neural network compliance control module into the overall forging application is
discussed in Section 5.

4.1. FORGING SIMULATION FOR NEURAL NETWORK TRAINING

In the current application, force and displacement data obtained from an off-
line simulation of the open die forging process were used as training sets for
the neural network. The simple upsetting of an aluminum billet was chosen
to represent a typical open die forging example. The finite element modelling
approach described in Section 3 was used to simulate the upsetting process.
Figure 2 illustrates the finite element mesh generated by the simulation for the
aluminum billet at the begining and end of the reduction process (XY-plane
only). The locations of nodes 1, 10, 91, and 100 are shown explicitly (before and
after reduction). The forging simulation required 60 steps; the workpiece was
compressed 0.02925 inches in each step.
The magnitude and direction of the contact forces and moments exerted on
the robot end effector during forging are of considerable interest in avoiding

Y
10
Initial form

J
1" I ii ~ ~ X ~ 0*

-- "1-t'-- V" r'x- - ~

I I LI t A I-/
91, Pl* 100 100"
X

Figure 2. Simple forging example.

SIMULATION AND CONTROL FOR INTELLIGENT FORGING 93

14 -0,006

12 Fz
9 -0.008

~2
q,) 9 -0.010

>4
< ' -0.012 >
6

, , , ,. , , .o.ol,
0.o 0.1 o 2 o 3 0., o.s ~

Time(Sec)

Figure 3. Contact forces at end effector.

damage to the robot mechanism. It is these forces and moments which the com-
pliance control scheme is designed to reduce and/or eliminate. Figure 3 shows
the computed contact forces at the end effector in the X and Y directions (Fz
and F v, respectively) throughout the upsetting process. As expected, the magni-
tude of the axial force, Fz, which results from the elongation of the workpiece
in the X direction, is considerably greater than the vertical force being applied
to the gripper, Fy. The neural network was trained to map given Cartesian forces
and moments exerted on the robot end effector to corresponding end effector
displacements using this force and displacement data.

4.2. NEURALNETWORK TRAINING RESULTS

The specific structure of the multi-layer neural network used here is as follows:
20-30-20-19. That is, it utilizes 20 inputs and 19 outputs, with 2 hidden layers of
30 and 20 neurons, respectively. Its 20 inputs are composed of the forces in the X
and Y directions at nodes 10, 2 0 , . . . , 100 of the workpiece model, all of which
are assumed to be in contact with the robot end effector. The network outputs
are the 19 displacements in the X and Y directions of these same nodes. (Note:
node 100 is constrained by the bottom die, and therefore, has no displacement
in the Y direction.)
The input/output data from every fifth step of the required 60 steps of the
finite element simulation formed a training set for the neural network. The back
propagation method [32] was used with global momentum and a learning rate
of 0.9 and 0.5, respectively. A total of 998 training cycles, corresponding to
94 K.W. LILLY AND A. S. MELLIGER1

0-8
0-7
o MAX Error
0.6
----o RMS Error
0.5
0-4
0.3
0.2
0-1
0
0 0 0 0 C) C) 0 0 0 0 0 0 0 0 0 0 0
r'.- r O) u') T- I'~ r O~ to ~ I'~ r (3) it) ~ D.,.
CO oo ,.. O)

Number of Cycles

Figure 4. Training errror plot.

570.7 sec on a VaxStation 3100 M38, were required for the neural network to
learn the compliance mapping. In the training error plot of Figure 4, both RMS
and maximum errors have dropped to less than 5% at the end of the training cycle,
demonstrating that the neural network was well-trained for the given input/output
pairs.

5. Integrated Simulation of the Automated Forging Process

This section briefly describes the integration and coordination of the computa-
tional components involved in the complete dynamic simulation of the Integrated
Robot/Forge Processing Center. Figure 5 presents a block diagram of the data
flow and module interconnections required for the simulated application of a
neural network-based active compliance control scheme in the automated open
die forging process. This simulation structure represents an important step in
the development of a real-time compliance control system for practical forging
applications.
As shown in Figure 5, T represents the input joint actuator torques/forces
applied to the industrial robot as calculated by the Joint Controller, based on the
desired task trajectory. This task trajectory defines the desired motion of the robot
end effector, and thus, the workpiece itself, as it interacts with the forging press.
Vectors q and q represent the actual positions and velocities of the robot joints,
respectively, while f represents the vector of spatial forces/moments acting on
the robot end effector (generally nonzero for closed chain configurations). These
three vectors are outputs of the appropriate open or closed chain Direct Dynamics
algorithm which forms the basis of the Robot Arm Dynamics module. Note that
SIMULATIONANDCONTROLFORINTELLIGENTFORGING 95

Osk[
Trajectory ~ Joint
"~[ C~176
I
I
x I
--I
Robot
Arm f
Dynamics (closedchain)

q,q
(openand closedchain)

qd Inverse 1 Xd NeuralNetwork ~1~

Kinematic _~ Compliance
Transform Control

Figure 5. Block diagram of the integrated simulationwith compliance control.

for open chain configurations of the robot, f is zero, and the Compliance Control
feedback loop is not activated.
The spatial forces generated by the closed chain dynamics algorithm during
constrained motion of the robot (i.e. during forging) are used as inputs to the
Neural Network Compliance Control module. The Compliance Control module
maps these forces to the corresponding displacements of the robot end effector,
Xd, which will reduce or eliminate the tip force loading on the robot. These
Cartesian displacements are then transformed to joint space using an Inverse
Kinematic Transformation. The resulting joint positions, qd, are used as supple-
mentary inputs to the Joint Controller (currently a simple PID control algorithm),
along with the actual joint positions and velocities obtained from the Robot Arm
Dynamics module (q, ~)). Specific simulation results using this integrated com-
putational structure are now discussed.

5.1. INTEGRATEDSYSTEMSIMULATIONRESULTS
In the simulation of all advanced robotic systems, computational complexity and
elapsed computation time are of considerable interest. Table II shows the time
required for the computation of specific quantities and/or the implementation of
individual components of the simulation (one iteration). All computations were
conducted on a VaxStation 3100 M38 with a clock resolution of 10 ms.
The Direct Dynamics computation of the closed chain joint accelerations for
constrained motion of the industrial robot results in the most significant computa-
tional burden (1 I0 ms). This fact further emphasizes the need for the development
of even more efficient closed chain robot dynamics algorithms. Integration of the
closed chain joint accelerations to obtain the joint velocities and positions, along
with the PID joint control scheme implemented by the Joint Controller, requires
96 K.W. LILLYAND A. S. MELLIGERI

Table II. Timing results

Computation Time (ms)

Closed chain joint accelerations ll0
Integration and PID control I0
Compliance control mapping 10
Inverse kinematic transform 5
Total 135

a total of only l0 ms of computation time. The mapping of end effector forces to

corresponding tip displacements in the Compliance Control module is also very
efficient (10 ms), which demonstrates the value of the neural network compliance
control approach. For comparison, note that the calculation of one step of the
finite element analysis of the forging process required 180 ms. The inverse kine-
matic transformation of desired end effector positions to desired joint positions
is the least computationally expensive task (5 ms).

5.2. EFFECTIVENESS OF THE COMPLIANCE CONTROL MODULE

The RMS step error plot shown in Figure 6 illustrates the performance of the
neural network at different steps of the forging process as it is used in the

0.14

0.12
,,r
ca 0.1
.=
9=- 0.08
t=

0.06

c~ 0-04

0.02

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Step #

Figure 6. RMS step error plot.

SIMULATION AND CONTROL FOR INTELLIGENT FORGING 97

0.09

o t°x. r
0.07

o.oo
I=

l- 0.04

003 I
0.01

0
0
T-
I
i
0
t"J
;
0
03
,
0
'~"
~
0
~
I
0
iO
I
0
D~
!
0
CO
',
0
(~)
0
0

Node #

Figure 7. Effectiveness of the compliance control module,

integrated simulation. Steps t, 5, 10, and 15 are the trained steps. From the
data of Figure 7, we see that the Neural Network Compliance Control module
provided a good estimation of the robot end effector position within an RMS
error of 0.025 inch for the trained points. From the RMS position error, we can
also observe how well the compliance control module performed for each nodal
displacement. Note that a good mapping was achieved in the X direction (as
compared to the Y direction) for nodes in the top half of the workpiece, while a
better mapping was achieved in the Y direction for nodes in the bottom half of
the workpiece.

6. Summary and Conclusions

This work investigated the application of neural networks for the compliance
control of an automated open die forging system through full dynamic system
simulation. Efficient algorithms and methods were utilized to simulate the motion
of the industrial robot and to model the material flow of the workpiece during
the forging process. A multi-layer neural network was trained to implement
the nonlinear force/displacement mapping required for the compliance control
scheme, which was designed to reduce or eliminate potentially damaging forces
transmitted to the robot during forging. These computational system components
were successfully coordinated in an integrated simulation environment.
A simple open die forging example (upsetting of an aluminum billet) was
used to test this neural network-based compliance control approach. The per-
98 K.W. LILLYAND A. S. MELLIGERI

formance and effectiveness of the Compliance Control module were examined

with good results. Timing issues for real-time implementation were also briefly
discussed. It is our conclusion that the use of neural networks for the compliance
control o f automated forging systems is a viable approach which deserves further
investigation.
Efficient dynamic simulation of the neural network-based compliance control
of an Integrated Robot/Forge System is a significant achievement. Further exten-
sions to the simulation are possible, however, and should be pursued. Extension
o f the finite element modelling analysis to a 3-dimensional form (with heat trans-
fer) should be implemented to enable the modelling of more complex workpiece
geometries. More advanced joint control schemes can be implemented within the
simulation model. Graphical visualization of the simulated forging task should
also be investigated. Finally, the simulation results generated here must be vali-
dated through implementation on forging/robot hardware.

References
1. Ang, C. S.: Implementation of the virtual environment force control concept on an industrial
robot, MS Thesis, Department of Mechanical Engineering, Penn State University, 1994.
2. Appleton, E., Higginbotham, W. B. and Law, D.: Open die forging with an industrial robot,
The Industrial Robot (December 1979), 191-194.
3. Appleton, E.: Open-die forging operations by industrial robot, Robotic Technology 15 (1983),
160-168.
4. Asada, H.: Teaching and learning of compliance using neural nets: Representation and gener-
ation of nonlinear compliance, in Proc. 1990 IEEE Int. Conf. Robotics and Automation, 1990,
pp. 1237-1244.
5. Brandl, H., Johanni, R. and Otter, M.: A very efficient algorithm for the simulation of robots
and similar multibody systems without inversion of the mass matrix, Proc. IFAC/IFIP/IMACS
Int. Symp. Theory of Robots, 1986.
6. Cha, D. H.: Compliance control of telerobot systems using a recurrent neural network, in Proc.
Japan-USA Symp. Flexible Automation, Part 1, 1992, pp. 271-274.
7. Connoly, T. H.: Neural network hybrid position/force control, in Proc. 1993 Int. Conf. Intelli-
gent Robots and Systems, 1993, pp. 240-244.
8. Featherstone, R.: Robot Dynamics Algorithms, Kluwer, Boston, 1987.
9. FIA, Open Die Forging Manual, 3rd edn, Forging Industry Association, Cleveland, OH, 1982.
10. Funahashi, K. I.: On the approximate realization of continuous mappings by neural networks,
Neural Networks 2 (1989), 183-192.
11. Han, C. S., Gandhi, R. V. and Srinivasan, R.: Optimum design of forging die shapes using
nonlinear finite element analysis, AIAA J. 31(4) (1993), 774--781.
12. Her, M. G.: Automated robotic deburring of parts using compliance control, ASME J. Dynamic
Systems, Measurement, and Control 113(1) (1991), 60-66.
13. Kazerooni, H.: Compliance control and stability analysis of cooperating robot manipulators,
Robotica 7(3) (1989), 191-198.
14. Kobayashi, S., Oh, S. and Altan. T.: Metal Forming and the Finite Element Method, Oxford
University Press, New York, 1989.
15. Kwan, C. M.: Robust force/motion control of constrained robots using neural networks, in
Proc. IEEE Conf. Decision and Control, Vol. 2, 1994, pp. 1862-1867.
16. Lee, C. H. and Kobayashi, S.: New solution to rigid plastic deformation problems using a
matrix method, ASME Trans. J. Engrg. Industry 95 (1973), 865-873.
17. Lilly, K. W.: Efficient Dynamic Simulation of Robotic Mechanisms, Kluwer, Boston, 1993.
SIMULATION AND CONTROL FOR INTELLIGENTFORGING 99

18. Lilly, K. W. and Jablokow, A. G.: Intelligent open die forging, Int. J. Industrial Engrg. 2(3)
(1995).
19. Lilly, K. W. and Melligeri, A. S.: Efficient dynamic simulation of an integrated robot/forge near
net-shape processing center, in R. P. Judd and N. A. Kheir (eds), Intelligent Manufacturing
Systems, Pergamon, New York, 1992, pp. 347-351.
20. Lilly, K. W. and Orin, D. E.: Efficient dynamic simulation of a single closed-chain manipulator,
in Proc. 1991 IEEE Int. Conf. Robotics and Automation, Vol. 1, 1991, pp. 210-215.
21. Maekawa, A.: Compliance control of an articulated manipulator, JSME Int. J. Ser. C: Dynamics,
Control Robotics, Design and Manufacturing 37(1) (1994), 155-161.
22. Melligeri, A. S. and Lilly, K. W.: Application of neural networks in compliance control of
an integrated robot/forge processing center, in R. Sodhi (ed.), Advances in Manufacturing
Systems: Design, Modeling and Analysis, Elsevier, New York, 1993, pp. 445--450.
23. Orin, D. E. and McGhee, R. B.: Dynamic computer simulation of robotic mechanisms, Int. J.
Robotics Res. 7(5) (1981), 32-47.
24. Raibert, M. and Craig, J.: Hybrid position/force control of manipulators, ASME J. Dynamic
Systems, Measurement, and Control 102(2) (1981), 126-133.
25. Roberson, R. E. and Schwertassek, R. E: Introduction to the Dynamics of Multibody Systems,
Springer, Berlin, 1987.
26. Rodriguez, G., Jain, A. and Kreutz-Delgado, K.: A spatial operator algebra for manipulator
modeling and control, Int. J. Robotics Res. 10 (1991), 371-381.
27. Rumelhart, D. E. and McClelland, J. L.: Parallel Distributed Processing, MIT Press, 1986.
28. Venkataraman, S. T., Gulati, S., Barhen, J. and Toomarian, N.: Parameter learning and compli-
ance control using neural networks, in G. A. Bekey and K. Y. Goldberg (eds), Neural Networks
in Robotics, Kluwer, Boston, 1993, pp. 193-216.
29. Vitscheff, V.: A programmable manipulator for closed die forging, Proc. 9th Int. Drop Forging
Convention, Kyoto, Japan, 1977.
30. Waibel, B. J.: Theory and experiments on the stability of robot compliance control, IEEE
Trans. Robotics and Automation 7(1) (1991), 95-104.
31. Walker, M. W. and Orin, D. E.: Efficient dynamic computer simulation of robotic mechanisms,
ASME J. Dynamic Systems, Measurement, and Control 104 (1982), 205-211.
32. Wasserman, P. D.: Neural Computing, Van Nostrand/Reinhold, New York, 1989.
33. Westmeyer, W. B.: Modernization speeds open die operation, Forging, Winter Issue (1991),
30-34.
34. Wright, J. R.: Investing in Excellence, Forging, Winter Issue (1991), 16--19.