International journal of computer integrated manufacturing , tập 23, số 6, 2010

International Journal of Computer Integrated Manufacturing
Vol. 23, No. 6, June 2010, 487–499

Sources of variability in the set-up of an indoor GPS
Carlo Ferria*, Luca Mastrogiacomob and Julian Farawayc
a

Via XI Febbraio 40, 24060 Castelli Calepio, BG, Italy; bDISPEA, Politecnico di Torino, Corso Duca degli Abruzzi 24, Torino
10129, Italy; cDepartment of Mathematical Sciences, University of Bath, Bath BA2 7AY, UK
(Received 17 June 2009; final version received 5 January 2010)
An increasing demand for an extended flexibility to model types and production volumes in the manufacture of
large-size assemblies has generated a growing interest in the reduction of jigs and fixtures deployment during
assembly operations. A key factor enabling and sustaining this reduction is the constantly expanding availability of
instruments for dimensional measurement of large-size products. However, the increasing complexity of these
measurement systems and their set-up procedures may hinder the final users in their effort to assess whether the
performance of these instruments is adequate for pre-specified inspection tasks. In this paper, mixed-effects and
fixed-effects linear statistical models are proposed as a tool to assess quantitatively the effect of set-up procedures on
the uncertainty of measurement results. This approach is demonstrated on a Metris Indoor GPS system (iGPS). The
main conclusion is that more than 99% of the variability in the considered measurements is accounted for by the
number of points used in the bundle adjustment procedure during the set-up phase. Also, different regions of the
workspace have significantly different error standard deviations and a significant effect on the transient duration of
measurement. This is expected to affect adversely the precision and unbiasedness of measurements taken with Indoor
GPS when tracking moving objects.
Keywords: large scale metrology; large volume metrology; distributed coordinate measuring systems; Indoor GPS;
iGPS; uncertainty

1. Introduction
During the last decades research efforts in coordinatemeasuring systems for large-size objects have led to a
broadening of the range of instruments commercially
available (cf. Estler et al. 2002).

These coordinate measurement instruments can be
grouped into two categories: centralised and distributed systems (Maisano et al. 2008).
A centralised instrument is a measuring system constituted by a single hardware element that in performing a measurement may require one or more ancillary
devices such as, typically, a computer. An example of a
centralised instrument is a laser tracker that makes
use of a spherically-mounted reflector (SMR) to take
a measurement of point spatial coordinates and that
needs to be connected to a monitor of environmental
conditions and to a computer.
A distributed instrument is a collection of separate
independent elements whose separately gathered measurement information needs to be jointly processed in
order for the system to determine the coordinates of a
point. A single element of the system typically cannot
provide measurements of the coordinates of a point
when standing alone. Precursors of these apparatuses
can be identified in wireless indoor networks of sensors

*Corresponding author. Email:
ISSN 0951-192X print/ISSN 1362-3052 online
Ó 2010 Taylor & Francis
DOI: 10.1080/09511921003642147


for automatic detection of object location (cf. Liu
et al. 2007). These networks can be deployed for inspection tasks in manufacturing operations once their
trueness has been increased. The term trueness is
defined in BS ISO 5725-1:1994 as ‘the closeness of
agreement between the average value obtained from a
large series of test results and an accepted reference
value’ (Section 3.7).

When inspecting parts and assemblies having large
dimensions, it is often more practical or convenient
to bring the measuring system to the part rather than
vice versa, as is typically the case on a smaller scale.
Therefore, instruments for the inspection of large size
objects are usually portable. In performing a measurement task, a single centralised instrument, say a laser
tracker, can then be deployed in a number of different
postions which can also be referred to as stations. By
measuring some fixed points when changing station,
the work envelope of the instrument can be significantly enlarged enabling a single centralised instrument to be used for inspection of parts significantly
larger than its original work envelope. To illustrate this
concept, in Figure 1(a) the top view of three geometrical solids, a cylinder, a cube and an octahedron
(specifically a hexagonal prism) is displayed. These


488

Figure 1.

C. Ferri et al.

Centralised and distributed measurement systems.

solids are inspected by a single centralised instrument
such as a laser tracker, which is moved across different
positions (1, 2, . . . , 6 in the figure) from each of
which the coordinates of the points P1, P2 and P3 are
also measured. In this respect, a single centralised
system appears therefore comparable with a distributed system, whose inherent multi-element nature
enables work envelopes of any size to be covered,

provided that a sufficient number of elements are
chosen. This characteristic of a measuring system of
adapting itself to suit the scale of a measuring task is
often referred to as scalability (cf. Liu et al. 2007). The
concept above can therefore be synthesised by saying
that a centralised system is essentially scalable in virtue
of its portability, whereas a distributed system is such
due to its intrinsic modularity.
With a single centralised instrument, measurement
tasks within a working envelope, however extended,
cannot be performed concurrently but only serially.
Each measurement task to be performed at a certain
instant in time needs a dedicated centralised instrument. This is shown in Figure 1(a) where the cylinder is
measured at the current instant with the instrument in
position 2, whereas the hexagonal prism is going to be
measured in a future instant when the instrument will
be placed in position 3. With a distributed system this
limitation does not hold. With a distributed system,
concurrent measurement tasks can be performed provided that each of the concurrent tasks has a sensor or
subgroup of sensors dedicated to it at a specific instant
within the distributed instrument. In Figure 1(b), the
same three objects considered in the case of a centralised instrument are concurrently inspected using a
distributed system constituted by six signal transmitter
elements (1, 2, . . . , 6) and three probes, each carrying
two signal receiving elements whereby the coordinates
of the probe tips are calculated.

This characteristic of distributed systems is especially advantageous when concurrently tracking the
position of multiple large-size components during
assembly operations. The sole way of performing the

same concurrent operation with a centralised system
would require the availability and use of more than a
single centralised instrument (laser tracker, for instance), with potentially-detrimental economic consequences on the manufacturing organisation in terms of
increased fixed assets, maintenance costs and increased
complexity of the logistics.
A number of different distributed systems have
been developed recently, some as prototypes for research activities (cf., for instance, Priyantha et al. 2000;
Piontek et al. 2007), some others with a level of
maturity sufficient for them to be made commercially
available (cf., for instance, Welch et al. 2001; Maisano
et al. 2008). In this second case, the protection of
intellectual property (IP) rights prevents users’ transparent access to the details of the internal mechanisms
and of the software implemented in the systems. This
may constitute a barrier to a full characterisation of
the performance of the equipment. This investigation
endeavours to provide better insight into the performance of such systems by using widespread statistical
techniques. The main objective is therefore not to
criticise or evaluate the specific instrument considered
thereafter, but to demonstrate the use of techniques
that may be beneficially deployed also on other
distributed systems. In particular, the effect of discretionary set-up parameters on the variability and
stability of the measurement results has been analysed.
In the next section the main characteristics of
the Metris iGPS, which is the instrument considered,
are described. A cone-based mathematical model of
the system is then presented in Section 3. The experimental set-up is described in Section 4 and the results


International Journal of Computer Integrated Manufacturing
of the tests are analysed in Section 5. Conclusions are

drawn thereafter.
2. Physical description of the instrument
The instrument used in this study is the iGPS (alias
indoor GPS) manufactured by Metris. The description
of such a system provided in this section is derived
from publicly available information.
The elements constituting the system are a set of
two or more transmitters, a number of wireless sensors
(receivers) and an unit controlling the overall system
and processing the data (Hedges et al. 2003; Maisano
et al. 2008).
Transmitters are placed in fixed locations within
the volume where measurement tasks are performed.
Such a volume is also referred to as a workspace.
Each transmitter has a head rotating at a constant
angular velocity, which is different for each transmitter, and radiates three light signals: two infrared fanshaped laser beams generated by the rotating head,
and one infrared strobe signal generated by light
emitting diodes (LEDs). The LEDs flash at constant
time intervals ideally in all directions, but practically in
a multitude of directions. Each of these time intervals
is equal to the period of revolution of the rotating head
on which the LEDs are mounted. For any complete
revolution of the rotating head a single flash is emitted
virtually in all directions. In this way, the LED signals
received by a generic sensor from a transmitter
constitute a periodic train of pulses in the time domain
where each pulse is symmetric (cf. Hedges et al. 2003,
column 6).
The rotating fan-shaped laser beams are tilted by
two pre-specified opposite angles, f1 and f2 (e.g. 730

and 308, respectively) from the axis of rotation of the

489

head. These angles are also referred to as slant angles.
The fact that the angular velocity of the head is
different for different transmitters enables each transmitter to be distinguished (Sae-Hau 2003). A schematic
representation of a transmitter at the instant t1 when
the first fanned beam L1 intersects the sensor in
position P and at the instant t2 when the second fanned
beam L2 passes through P is shown in Figure 2, where
two values for the slant angles are also shown. Ideally,
the shape of each of the fanned beams should be
adjustable to adapt to the characteristics of the measurement tasks within a workspace. Although two
beams are usually mounted on a rotating head, configurations with four beams per head have also been
reported (Hedges et al. 2003, column 5). To differentiate between the two fanned beams on a transmitter, their time position relative to the strobe signal is
considered (see Figure 2).
The fanned beams are often reported as planar
(Liu et al. 2008; Maisano et al. 2008), as depicted in
Figure 2. Yet, the same beams when emitted from
the source typically have a conical shape that is first
deformed into a column via a collimating lens and
then into a fan-shape via a fanning lens (Hedges et al.
2003, column 6). It is believed that only an ideal chain
of deformations would transform completely and perfectly the initial conical shape into a plane. For these
reasons, the final shape of the beam is believed to
preserve traces of the initial shape and to be more
accurately modelled with a portion of a conical surface, rather than a plane. Each of the two conical
surfaces is then represented by a vector, called a cone
vector, that is directed from the apex to the centre of

the circular directrix of the cone. The angle between a
cone vector and any of the generatrices on the cone
surface is called the cone central angle. This angle is

Figure 2. Schema of a transmitter at the instants t1 and t2 when the first and second fanned beams, respectively, intersect the
position (P) of a sensor. 730 and þ308 are two arbitrary values of the slant angle.


490

C. Ferri et al.

designated by a1 and a2 for the first and the second
beams, respectively. The apex of each cone lies on the
axis of rotation of the spinning head. In Figure 3, a
schema of the portion of the conical surface representing a rotating laser beam is displayed. In this figure,
two portions of conical surfaces are shown to illustrate a2 and f2 (f2 4 0, having established counterclockwise angle measurements around the x-axis as
positive).
The angular separation between the optical axes of
the two laser modules in the rotating head is denoted
by yoff, when observed from the direction of the rotational axis of the spinning head. The rotation of the
head causes each of the cone surfaces, and therefore
their cone vectors, to revolve around the same axis.
The angular position of the cone vector at a generic
instant is denoted by y1(t) and y2(t) for the first and
second fanned beams, respectively. These angles are
also referred to as scan angles and are defined relative
to the strobe LED synchronisation signal, as illustrated
below.
Wireless sensors are made of one or more photodetectors and a wireless connection to the controlling

unit for the transmission of the positional information
to the central controlling unit. The use of the photodetectors enables the conversion of a received signal
(stroboscopic LED, first fanned laser, second fanned
laser) into the instant of time of its arrival (t0, t1 and t2
in Figure 2). The time intervals between these instants
can then be converted into measurements of scan
angles from the knowledge of the angular velocity of
the head for each transmitter (o in Figure 2). It is
expected that y1 ¼ o 6 (t17t0) and that y2 ¼ o 6
(t27t0). At the instant t0 when the LED signal reaches
the generic position P, the same LED signal also
flashes in any direction. Therefore, at the very same

instant t0, the LED fires also in the reference direction
where the angles in the plane of rotation are measured
from (i.e. y1 ¼ y 2 ¼ 0).
In this study, any plane orthogonal to the axis of
rotation is referred to as a plane of rotation. For any
spherical coordinate system having the rotational axis
of the transmitter as the z-axis and the apex common
to the aforementioned conical surfaces as the origin,
the angle y1 swept by the cone vector of the first fanned
beam in the time interval t17t0 is connected with the
azimuth of P measured from any possible reference
direction x established in the xy-plane, which is the
plane of rotation passing through the common apex of
the conical surfaces.
From a qualitative point of view, the elevation
(or the zenith) of P can be related to the quantity
o 6 (t27t1). By analogy with Figure 2, it is argued

that, also in the case of conical fanned shaped beams,
when the elevation (or zenith) of P is increasing
(decreasing), the time interval t27t1 is also increasing.
Vice versa, the reason why a time interval t27t1 is
larger than another can only be found in the fact that
the position of the sensor in the first case has a higher
elevation than in the second.
In the most typical configuration, two receivers are
mounted on a wand or a bar in calibrated positions. A
tip of the wand constitutes the point for which the
location is calculated based on the signals received by
the two sensors. When the receivers are mounted on a
bar, the bar is then often referred to as vector bar. If
such a receivers-mounted bar is short, say with a length
between 100 and 200 mm, it is then called a mini vector
bar. These devices are equipped with firmware providing processing capabilities. The firmware enables
the computation of azimuth and elevation of the wand
or bar tip for each of the spherical reference systems
associated with each of the transmitters in the system.
This firmware is called a position computation engine
(PCE).
A vector bar therefore acts as a mobile instrument for probing points as shown in the schema of
Figure 1(b). More recently, receiving instruments with
four sensors have been developed, enabling the user to
identify both the position of the tip and the orientation
of the receiving instrument itself.
3. The role of the bundle adjustment algorithms in the
indoor GPS

Figure 3. Schema of a shaped laser beam with two portions

of conical surfaces to show the central angle a2 and the slant
angle f2.

The computation of the azimuth and elevation of
the generic position P in the spherical reference system
of a generic transmitter enables the direction of the
oriented straight line l from the origin (the apex of the
cones) to P to be identified. However, it is not possible
to determine the location of P on l. In other words, it is


International Journal of Computer Integrated Manufacturing
not possible to determine the distance of P from the
origin. Therefore, at least a second transmitter is necessary to estimate the position of P in a user arbitrarily
predefined reference system {Uref}. In fact, assuming
that the position and orientation of the ith and jth
transmitters in {Uref} are known, then the coordinates
of the generic point on li and on lj can be transformed
from the spherical reference system of the transmitters
to the common reference system {Uref} (cf. Section 2.3
in Craig 1986). Then, P can be estimated with some
nonlinear least squares procedure, which minimises the
sum of the squared distances between the estimates of
the coordinates of P in {Uref} and the generic point on
li and lj. Only in an ideal situation would li and lj
intersect. In any measurement result, the azimuth and
elevation are only known with uncertainty (cf. Sections
2.2 and 3.1 in JCGM 2008). Very little likelihood exists
that these measured values for li and lj coincide with
the ‘true’ unknown measurands. The same very little

likelihood applies therefore to the existence of an
intersection between li and lj. When adding a third kth
transmitter, qualitative geometrical intuition supports
the idea that the distances of the optimal P from each
of the lines li, lj and lk are likely to be less variable until
approaching and stabilising around a limit that can be
considered typical for the measurement technology
under investigation. Increasing the number of transmitters is therefore expected to reduce the variability of
the residuals. The estimation of the coordinates of P,
when the position of the transmitters is known, is often
referred to as a triangulation problem (Hartley and
Sturm 1997; Savvides et al. 2001).
If the position and orientation of the transmitters
in {Uref} are not known, then they need to be determined before the actual usage of the measurement
system. To identify the position and orientation of a
transmitter in {Uref}, six additional parameters need
to be estimated (cf. Section 2.2 in Craig 1986). This
more general engineering problem is often referred to
as three-dimensional (3D) reconstruction and occurs in
areas as diverse as surveying networks (Wolf and
Ghilani 1997), photogrammetry and computer vision
(Triggs et al. 2000; Lourakis and Argyros 2009). The
estimation of three-dimensional point coordinates together with transmitter positions and orientations to
obtain a reconstruction which is optimal under a prespecified objective function and an assumed errors
structure is called bundle adjustment (BA). The objective or cost function describes the fitting of a mathematical model for measurement procedure to the
experimental measurement data. Most often, but not
necessarily, this results in minimising the sum of the
squares of the deviations of the measurement data
from their values predicted with nonlinear functions of
the unknown parameters (Triggs et al. 2000; Lourakis

491

and Argyros 2009). A range of general purpose optimisation algorithms, such as for instance those of
Gauss–Netwon and Levenberg–Marquardt, can be
used to minimise the nonlinear objective function.
Alternatively, significantly increased efficiency can be
gained if these algorithms are adjusted to account for
the sparsity of the matrices arising in the mathematical
description of 3D reconstruction problems (Lourakis
and Argyros 2009).
In the measurement system investigated, a BA
algorithm is run in a set-up phase whereby the position and orientation of each transmitter in {Uref} are
determined. Therefore, during the subsequent deployment of the system (measuring phase), the coordinates
of a point are calculated using the triangulation
methods mentioned above.
However, as is typically encountered in commercial
measurement systems, the BA algorithms implemented
in the system are not disclosed completely to the users.
This makes it difficult for both users and researchers to
devise analytical methods to assess the effects of these
algorithms on the measuring system. In this investigation, consideration is given to experimental design and
statistical techniques to estimate the effect that decisions taken when running the built-in BA algorithm
exert on measurement results.
4. Experimental set-up
Four transmitters were mounted on tripods and placed
at a height of about two metres from floor level. The
direction of the rotational axis of each transmitter
spinning head was approximately vertical. Each of the
four transmitters was placed at the corners of an

approximate square of side about eight metres.
A series of six different targets fields labelled I, II,
III, IV, V and VI and respectively consisting of 8, 9, 10,
11, 12 and 13 targets was considered during the BA
procedure. Each of these fields was obtained by adding
one target to the previous field, so that the first eight
targets are common to all the fields, the first nine
targets are common to the last five fields and so on. A
schema of this experimental configuration is shown in
Figure 4.
All the fields were about 1.2 m above floor level.
The target positions were identified using an isostatic
support mounted on a tripod which was moved across
the workspace. A set of the same isostatic supports was
also available on a carbon-fibre bar that was used to
provide the BA algorithm built in the system with a
requested measurement of length (i.e. to scale the
system). A distance of 1750 mm between two isostatic
supports on the carbon-fibre bar was measured on a
coordinate-measuring machine (CMM). The carbonfibre bar was then placed in the central region of


492

C. Ferri et al.

the workspace. The coordinates of the two targets
1750 mm apart were measured with iGPS and their
1750 mm distance was used to scale the system in all
the target fields considered. In this way, the scaling

procedure is not expected to contribute to the variability of the measurement results even when different
target fields are used in the BA procedure. Figure 5
shows an end of the vector bar used in this set-up (the

large sphere in the figure), while coupled with an isostatic support (the three small spheres) during the
measurement of a target position on the carbon-fibre
bar.
The BA algorithm was run on each of these six
targets fields so that six different numerical descriptions of the same physical positions and orientations of
the transmitters were obtained.
Six new targets locations were then identified using
the isostatic supports on the carbon-fibre bar mentioned above. Using the output of the BA executions,
the spatial coordinates of these new target locations
were measured. The approximate position of the six
targets relative to the transmitters is shown in the
schema of Figure 6.
Each target measurement consisted in placing the
vector bar in the corresponding isostatic support and
holding it for about 30 s. This enabled the measurement system to collect and store about 1200 records of
target coordinates in {Uref} for each of the six targets.
In this way, however, the number of records for each
target is different, owing to the human impossibility of
manually performing the measurement procedure with
a degree of time control sufficient to prevent this
situation occurring.

Figure 4.

Target fields I, II, III, IV, V and VI.

5. Results
Each of the six target positions displayed in Figure 6
and labelled 1, 2, 3, 4, 5, 6 was measured using each of
the six BA set-ups I, II, . . . , VI, giving rise to a

Figure 5.

Isostatic support identifying a target.

Figure 6.

Target field when running the instrument.


International Journal of Computer Integrated Manufacturing
grouping structure of 36 measurement conditions
(cells).
When measuring a target location, its three Cartesian coordinates in {Uref} are obtained. To reduce the
complexity of the analysis from three-dimensional to
mono-dimensional, instead of these coordinates the
distance of the targets from the origin of {Uref} is
considered. Central to this investigation is the estimation of the effect on the target–origin distance due to
the choice of a different number of target points when
running the BA algorithm. The target locations 1,
2, . . . , 6 do not identify points on a spherical surface,
so they are at different distances from the origin of
{Uref}, regardless of any possible choice of such a
reference system. These target locations therefore
contribute to the variability of the measurements of
the target–origin distance whereby the detection of a

potential contribution of the BA set-ups to the same
variability can be hindered. To counteract this masking
effect, the experiment was carried out by selecting first
a target location and then randomly assigning all the
BA set-ups for that location to the sequence of tests.
This was repeated for all the six target positions. Such
an experimental strategy introduces a constraint to a
completely random assignment of the 36 measurement
conditions to the the run order. In the literature (cf.
Chapters 27, 16 and 8 in Neter et al. 1996, Faraway
2005 and Faraway 2006, respectively), this strategy is
referred to as randomised complete block design
(RCBD). The positions of the targets 1, 2, . . . , 6
constitutes a blocking factor identifying an experimental unit or block, within which the BA set-ups are
tested. The BA set-ups I, II, . . . , VI constitute a
random sample of all the possible set-ups that differ
only in the choice of the location and number of points
selected when running the BA algorithm during the
system set-up phase. On the other hand, the analysis
of the obvious contribution to the variability of the
origin–target distance when changing the location of
the targets would not add any interesting information
to this investigation. These considerations lead to describing the experimental data of the RCBD with a
linear mixed-effects statistical model, which is first
defined and then fitted to the experimental data.
5.1.

Mixed-effects models

The distance dij of the i th (i ¼ 1, . . . , 6) target from

the origin measured when using the jth (j ¼ I, . . . , VI)
BA procedure is modelled as the sum of four
contributions: a general mean m, a fixed effect ti due
to the selection of the i th target point, a random effect
bj due to the assignment of the jth BA set-up and a
random error eij due to all those sources of variability
inherent in any experimental investigation that is not

493

possible or convenient to control. This is described by
the equation
dij ¼ m þ ti þ bj þ eij :

ð1Þ

In Equation (1) and hereafter, the Greek symbols are
parameters to be estimated and the Latin symbols are
random variables. In particular, the bj’s have zero
mean and standard deviation sb; the eij’s have zero
mean and standard deviation s. The eij’s are assumed
to be made of independent random variables normally
distributed, i.e. eij * N(0,s2). The same applies to the
bj’s, namely bj * N(0,s2b ). The eij’s and the bj’s are also
assumed to be independent of each other. Under these
assumptions, the variance of dij, namely s2d , is given by
the equation
s2d ¼ s2b þ s2 :

ð2Þ

Using the terminology of the ‘Guide to the expression
of uncertainty in measurement’ (cf. Definition 2.3 in
JCGM 2008), sd is the standard uncertainty of the
result of the measurement of the origin–target
distance.
As pointed out in the previous section, the number
of the determinations of the target–origin distance that
have been recorded is different for each of the 36
measurement conditions. For simplicity of the analysis, the number of samples gathered in each of these
conditions has been made equal by neglecting the
samples in excess of the original minimum sample size
over all the cells. This resulted in considering 970
observations in each cell. The measurement result
provided by the instrument in each of these conditions
and used as a realisation of the response variable dij in
Equation (1) is then defined as the sample mean of
these 970 observations. There is a single measurement
result in each of the 36 cells. The parameters of the
model, i.e. m, ti, sb and s, have been estimated by the
restricted maximum likelihood (REML) method as
implemented in the lme() function of the package
nlme of the free software environment for statistical
computing and graphics called R (cf. R Development
Core Team 2009). More details about the REML
method and the package nlme are presented in
Pinheiro and Bates (2000). The RCBD assumes that
there is no interaction between the block factor (target
locations) and the treatment (BA set-up). This
hypothesis is necessary so that the variability within

a cell represented by the variance s2 of the random
errors can be estimated when only one experimental
result is present in one cell. In principle, such an
estimation is enabled by considering the variation of
the deviations of the data from their predicted values
across all the cells. This would estimate the variability


494

C. Ferri et al.

of an interaction effect, if it were present. If an
interaction between target locations and BA set-ups
^ of s provided in this
actually exists, the estimate s
study would account for both interaction and error
variability in a joint way and it would not be possible
to separate the two components. Therefore, from a
practical point of view, the more the hypothesis of no
^ overestimates s.
interaction is violated, the more s
After fitting the model, an assessment of the
assumptions on the errors has been performed on the
realised residuals, i.e. the deviation of the experimental
results from the results predicted by the fitted model
for corresponding cells (eˆij ¼ dij7dˆij). The realised
residuals plotted against the positions of the targets do
not appear consistent with the hypothesis of constant
variance of the errors. In fact, as shown in Figure 7(a),

the variability of the realised residuals standardised by
^, namely e^ij ¼ ðdij À d^ij Þ=^
s
s seems different in different
target locations.
For this reason, an alternative model of the data
has been considered which accounts for the variance
structure of the errors. This alternative model is defined as the initial model (see Equation 1), bar the
variance of the errors which is modelled as different in
different target locations, namely:

si ¼ snew  di ;

d1 ¼ 1:

ð3Þ

From Equation (3) it follows that snew is the unknown
parameter describing the error standard deviation in
the target position 1, whereas the di’s (i ¼ 2, . . . , 6) are
the ratios of the error standard deviation in the ith
target position and the first.
The alternative model has been fitted using one of
the class variance functions provided in the package
nlme and the function lme() so that also snew and the
di’s are optimised jointly with the other model
parameters (m, ti and sb) by the application of the
REML method (Section 5.2 in Pinheiro and Bates
2000).
For the alternative model, diagnostic analyses of

the realised residuals were not in denial of its underlying assumptions. The standardised realisations of the
residuals, i.e. e^ij ¼ ðdij À d^ij Þ=^
si , when plotted against
the target locations (Figure 7(b)) do not appear any
longer to exhibit different variances in different
target locations, as was the case in the initial model
(Figure 7(a)). The same standardised realisations were
also found not to exhibit any significant departure
from normality.
The fact that all the target fields have more than
50% of the targets in common together with the fact
that each field has been obtained by recursively adding
a single target to the current field may cause the
experimenters to expect that the measurement results
obtained when different target fields have been used in

Figure 7. Realisations of the standardised residuals (dimensionless) grouped by target positions for the initial and the
alternative mixed-effects models.


International Journal of Computer Integrated Manufacturing
the BA procedure have some degree of correlation. If
that were the case, then the experimental results
should be in denial of the assumed independence of
the random effects bj’s. The random effects, like the
errors, are unobservable random variables. Yet, algorithms have been developed to predict the realisations
of these unobservable random effects on the basis of
the experimental results and their assumed model
(Equations 1, 2 and 3 with the pertinent description
above). The predictor used in this investigation is referred to as the best linear unbiased predictor (BLUP).

It has been implemented in nlme and it is described,
for instance, in Pinheiro and Bates (2000). The
predicted random effects bˆj’s for the model and the
measurement results under investigation are displayed
in Figure 8(a). To highlight a potential correlation
between predicted random effects relative to target
fields that differ by only one target, the bˆjþ1’s have been
plotted against the bˆj in Figure 8(b) (j ¼ 1, . . . , 5).
From a graphical examination of the diagrams of
Figure 8 it can be concluded that, in contrast with what
the procedure for establishing the targets fields may
lead the experimenter to expect, the measurement
results do not appear to support a violation of the
hypothesis of independence of the random effects.
Similar values for the BLUPs and therefore similar
conclusions can be drawn also for the initial mixedeffect model (the BLUPs for the initial model have not
been reported for brevity).
As suggested in Pinheiro and Bates (2000) (Section
5.2, in particular), to support the selection between

495

the initial and the alternative model, a likelihood ratio
test (LRT) has been run using the generic function
anova() implemented in R. A p-value of 0.84% led to
the rejection of the simpler initial model (8 parameters
to be estimated) when compared with the more complex alternative model (8 þ 5 parameters to be estimated). The same conclusion would hold if the
selection decision is made on the basis of the Akaike
information criterion (AIC) also provided in the output of anova() (read more about AIC in Chapters 1
and 2 of Pinheiro and Bates 2000).

This model selection bears significant practical
implications. From a practitioner’s point of view, in
fact, selection of the alternative model means that the
random errors have significantly different variances
when measuring targets in different locations of the
workspace. The workspace is not homogeneous: there
are regions where the variability of the random errors
is significantly lower than in others. This also means
that a measurement task can therefore be potentially
designed so that this measuring system can perform it
satisfactorily in some regions of its workspace but not
in others.
REML estimates of the parameters that have
practical implications are as follows:
^b ¼ 160:7 mm;
s

ð4Þ

^ ¼ 14:28 mm; ^d2 ¼ 0:2625; ^d3 ¼ 0:8599;
s
^d4 ¼ 0:3706; ^d5 ¼ 0:1260; ^d6 ¼ 0:5446:
ð5Þ

Figure 8. BLUPs of the random effects for each targets field and the graphically insignificant autocorrelation between BLUPs
of random effects associated with consecutive targets fields.


496

C. Ferri et al.

Estimates ^ti confirm the tautological significance of
^,
the location of the targets or block factor, whereas m
depending on the the parametrisation of the model,
can for instance be the centre of mass of the point
locations or can also be associated with a particular
target location (cf. Chapters 13 and 14 in Faraway
2005). All these estimates do not convey any practical
information. They are therefore not reported.
The significance of the random effect associated
with the BA set-up procedure has been tested using a
likelihood ratio approach, where the alternative model
has been compared with a null model characterised
by an identical variance structure of the errors but
without any random effect (i.e. sb ¼ 0). The p-value
was less than 10732 under the assumption of a chisquared distributed likelihood ratio. In reality, as
explained in Section 8.2 of Faraway (2006), such an
approach is quite conservative, i.e. it tends not to reject
the null hypothesis by overestimating the p-value.
However, given the extremely low p-value (510732),
there is strong evidence supporting the rejection of the
null hypothesis of an insignificant random effect (H0:
sb ¼ 0).
From a practical point of view, this indicates that
caution should be exerted when selecting the target
locations for running the BA algorithm during the setup phase: when repeating the BA procedure during the
set-up with identical positions of the transmitters, the
consideration of a different number of targets significantly inflates the variability of the final measurement results.

Substituting the estimates of Equations (4) and (5)
in the adaptation of Equation (2) to the alternative
model, it is derived after a few passages that the choice
of a different number of targets when running the BA
algorithm during the set-up phase accounts for 99.22,
99.94, 99.42, 99.89, 99.99 and 99.77% of the variance
of the measured origin–target distance when the target
is in locations 1, 2, 3, 4, 5 and 6, respectively. If there
was no discretion left to the operator when selecting
the number of targets and their locations during the
BA procedure, then the overall variability of the final
results in each of the location tested could have been
reduced by the large percentages reported above.
It may be worth pointing out that the designed
experiment considered in this investigation could be
replicated K times, on the same or in different days.
The obtained measuring results could then be modelled
with the following equation:
dijk ¼ m þ ti þ bj þ ck þ eijk

ð6Þ

À
Á
with ck $ N 0; s2c ; k ¼ 1; 2; . . . ; K, being the random effect associated with the kth repetition of the
experiment. The significance of the random effects ck’s

could then be tested in a similar way as the significance
of the bj’s has been tested above. The practical use of the
model of Equation (6) is twofold. First, it enables the

experimenter to detect if a significant source of varibility
can be associated with the replication of the whole
experiment. For instance, if each replication takes place
in slightly different natural and/or artificial light
conditions, then testing the significance of the ck would
tell if these enviromental conditions had significant
^c
effects on the measurement results (dijk). The estimate s
would quantify the increased variability of the response
varible attributable to them. Second, the increased
number of measurements taken would raise the
confidence of the experimenter in the estimates of
^b ; s
^c and s
^. For instance, it would dissipate (or
s
confirm) the suspicion that the experimenter may have
that the random effects attributed in Equation (1) to the
different setups, namely the bj’s, may be contributed to
by the natural variability due to repetition which was
estimated in Equations (4) and (5). This further study
can be considered as future work.
5.2.

Transient definition and analysis

In the above analysis, the average of all the 970
experimental data in a cell has been considered. The
variability of each of these 970 determinations of
distance, say st, is significantly larger than that of their

average (sd). If these determinationspwere
mutually
ffiffi
independent, then it would be sd ¼ st = ð970Þ. But the
determinations are instead highly correlated, owing
to the fact that they are taken at varying sampling
intervals of the order of milliseconds. Identifying the
correlation structure of these determinations is beyond
the scope of this investigation. In this study, when the
instrument is measuring the tth determination, say
dt,ij, a running average of all the determinations measured until that instant, say dt,ij, is considered. An
interesting question that arises is: ‘How many determinations are sufficient for the instrument to provide a
measurement dt,ij that does not differ much from the
measurement result dij?’. A 2mm maximum deviation
from dij has been considered for differentiating the
steady and the transient states of dt,ij. The value t?
has been used to identify the end of the transient. In
other words, for any index t 4 t? it holds that
jdt,ij7dijj 5 1mm.
In Figure 9, for each of the 36 experimental conditions, two continuous horizontal lines 1 mm apart
from the measurement result dij delimit the steady-state
region, whereas a single vertical dashed line indicates
the transition index t? from the transient to the steady
state as defined above.
From Figure 9, it is observed that for the same
target location (panels in the same column) the


International Journal of Computer Integrated Manufacturing

Figure 9.

497

Transition from the transient and the steady state.

transition from transient to steady state may occur at
different t?’s for different BA set-ups (different vertical
dashed lines in each panel). This suspicion is even
stronger when considering t? for the same BA set-up
but for different target locations (panels on a row in
Figure 9).
To ascertain whether the variation of t? with the
BA set-ups and with the target locations examined is
significant or is only the result of uncontrolled or uncontrollable random causes, the experimental values of
t? calculated starting from the RCBD already discussed have been analysed with a fixed-effects ANOVA
model (cf. Section 16.1 in Faraway 2005). The values
of t? have been computed by an ad hoc function
implemented in R by one the authors. The t?’s are
assumed as though they have been generated by the
following equation:
t?ij ¼ m þ bi þ gj þ eij ;

ð7Þ

where the bi’s and the gj’s are the effects of the blocking
factor (the target locations) and of the BA set-ups,

respectively, whereas the eij’s are the random errors,
assumed independent, normally distributed with constant variance and zero mean. The parameters have

been estimated using the ordinary least squares method
as implemented in the function lm() in R (cf. R Development Core Team 2009). The assumptions underlying the models have been checked on the realised
residuals and nothing amiss was found. To test the
potential presence of interaction between the two
factors in the form of the product of their two effects,
a Tukey test for additivity was also performed (cf.
Section 27.4 in Neter et al. 1996). This test returned a
p-value of 30.43%. It is therefore concluded that the
experimental data do not support the rejection of
the hypothesis of an additive model in favour of this
particular type of interaction effect of target locations
and BA set-ups on tij?.
The effect of the target positions on t? was
significant, i.e. H0 : bi ¼ 0(i ¼ 1, 2, . . . , 6) gives rise
to a p-value of 3.88% (under the hypotheses of the
model). However, the effect of the BA set-ups did not
appear to be significant, i.e. H0 : gi ¼ 0(i ¼ 1, 2, . . . , 6)


498

C. Ferri et al.

gives rise to a p-value of 84.96% (under the hypotheses
of the model).
From a practical point of view, there are two main
implications of these findings. First, the selection of a
different number of targets when running BA algorithms during the set-up phase does not appear to
have significant consequences on the duration of the
transient for obtaining a measurement. Second, the

duration of the transient appears to be significantly
different for different target locations within the workspace. This may be stated as follows: there are regions
of the workspace that require longer transient periods
than others before a measurement result stabilises, and
this is expected to have consequences for the accuracy
and precision of the determination of the position of
moving objects (tracking).
In fact, if a target point is in motion at a speed
sufficient for a number of determinations greater than
t? to be recorded in each measured point of its trajectory, then all the measurements results will be
representative of a steady state. But this may hold for
only some portions of the target trajectory; for others,
characterised by a larger t?, such a condition may not
be satisfied with a consequent inflation of the variability of those estimated positions, which may also be
biased.
6. Conclusions
The main characteristics of the Metris Indoor GPS
system have been reviewed on the basis of information
in the public domain. In particular, the working
principles of the system have been presented in terms
of a cone-based mathematical model.
The overall description of the system has been
instrumental in highlighting the key role of bundle
adjustment procedures during the set-up of the system. The selection of the number and location of target
points that are used when running the bundle adjustment procedure during the set-up phase can be affected
by discretionary judgements exerted by the operators.
To investigate the statistical significance of the
effects of this selection, a randomised complete block
design has been run on the distance between the origin
of the reference system and the measured positions of

target locations different from those used during the
bundle adjustment in the set-up phase. This design
enhances the possibility for the potential effects of
different set-ups on the origin–target distances to be
detected by discriminating them from the obvious
effects of the target positions. The set-ups considered
were different only in the number of the targets used
when executing the bundle adjustment procedure.
A mixed-effects and a fixed-effects linear statistical
model were fitted to the measurements results using the

restricted maximum likelihood method and the ordinary least squares technique, respectively.
The measurement results defined as the sample
average of the 970 determinations of distance recorded
in each target location for each set-up have been analysed with the mixed-effects model. By analysing the
realisations of the residuals, statistically different standard deviations of the random errors were identified
for different target positions. The work envelope of
the instrument do not therefore appear homogeneous:
in some areas the variability of the random error is
greater than in others, when performing measurements
of the distance of a target from the origin. Owing to
this heterogeneity, the punctual estimates of the standard uncertainty of the measured distances (sd) were
different for different target position and lay in a range
between 160.8 and 161.4 mm. The different set-ups,
tested to be statistically significant, always accounted
for more than 99.2% of the estimated standard uncertainty (the percentage varies for different target
positions). This quantitative evidence suggests that the
selection of points when running the bundle adjustment algorithms in the set-up phase should not be
overlooked. Performing this selection in a consistent
way according to some rule that ideally leads to

chosing the same points when the transmitters are in
the same positions may be a course of action worth
considering. Also, for replication and comparison
purposes, it may be advisable to quote the locations
of the targets used in setting up the system when
reporting the results of a measurement task.
The duration of the transient, i.e. the number of
determinations of distance needed for their current
average to be within +1 mm from the measurement
result (the average of the 970 determinations), has been
analysed with the fixed-effects model. The different
set-up configurations considered did not have any
significant effect on the duration of the transient.
However, this duration was significantly different in
different target locations. It can therefore be concluded
that the working space of the instrument is heterogeneous also for the characteristics of the transient of
measurement. It is expected that this conclusion has
negative implications on the precision and unbiasedness of the measurements obtained when using the
instrument for tracking moving points or moving
objects that the target points (or vector bars) are
attached to. Given a pre-specified configuration of
iGPS transmitters without any zone partitions having
been pre-established among them, if an object is
moving within an area of the working space of such
an iGPS, say area A, its position may be tracked
correctly, because the transient is sufficiently short
there. But if the same movement of the same object is
tracked by the same iGPS in another area of the same

International Journal of Computer Integrated Manufacturing
iGPS working area, say area B, the system may not be
able to track its location correctly because the transient
may not yet have finished.
Acknowledgements
This study is part of the research initiatives of The Bath
Innovative Design and Manufacturing Research Centre
(IdMRC), which is based in the Department of Mechanical
Engineering of the University of Bath and which is supported by
the United Kingdom Engineering and Physical Sciences
Research Council (EPSRC). In particular, this investigation
was carried out within the scope of the IdMRC research theme
‘Metrology, Assembly Systems and Technologies’ (MAST),
which is coordinated by Professor Paul Maropoulos.

References
BS ISO 5725-1, 1994. Accuracy (trueness and precision) of
measurement methods and results – Part 1: General
principles and definitions [online]. London: British Standards Institution. Available from: group.
com/ [Accessed 14 March 2010].
Craig, J.J., 1986. Introduction to robotics mechanics & control.
Reading, MA, USA: Addison-Wesley.
Estler, W.T., et al., 2002. Large-scale metrology – an update.
Annals of the CIRP, 51 (2), 587–609.
Faraway, J., 2005. Linear models with R. Boca Raton, FL,
USA: Chapman & Hall/CRC.
Faraway, J., 2006. Extending the linear model with R:
generalized linear, mixed effects and nonparametric
regression models. Boca Raton, FL, USA: Chapman &
Hall/CRC.

Hartley, R.I. and Sturm, P., 1997. Triangulation. Computer
Vision and Image Understanding, 68 (2), 146–157.
Hedges, T.M., et al., 2003. Position measurement system and
method using cone math calibration. United States
Patent US6535282B2, March.
JCGM (Joint Committee for Guides in Metrology), 2008.
JCGM:100:2008 (GUM:1995 with minor corrections)
Evaluation of measurement data – guide to the expression
of uncertainty in measurement [online]. Se`vres, France:
BIPM – Bureau International des Poids et Mesures.
Available from: [Accessed 7 June
2009].
Liu, H., et al., 2007. Survey of wireless indoor positioning
techniques and systems. IEEE Transactions on Systems,
Man, and Cybernetics – Part C: Applications and Reviews,
37 (6), 1067–1080.

499

Liu, Z., Liu, Z., and Lu, B., 2008. Error compensation of
indoor GPS measurement. In: C. Xiong et al., eds.,
Proceedings of the first international conference on
intelligent robotics and applications (ICIRA2008), Part
II, 15–23 October 2008. Wuhan, People’s Republic of
China. Lecture Notes in Computer Science 5315. Berlin:
Springer-Verlag, 612–619.
Lourakis, M.I.A. and Argyros, A.A., 2009. SBA: a software
package for generic sparse bundle adjustment. ACM
Transactions on Mathematical Software, 36 (1), 1–30.
Maisano, D., et al., 2008. Indoor GPS: system functionality

and initial performance evaluation. International Journal
of Manufacturing Research, 3 (3), 335–349.
Neter, J., et al., 1996. Applied linear statistical models. 4th ed.
New York: Irwin.
Pinheiro, J.C. and Bates, D.M., 2000. Mixed-effects models in
S and S-plus. New York: Springer-Verlag.
Piontek, H., Seyffer, M., and Kaiser, J., 2007. Improving the
accuracy of ultrasound-based localisation systems. Personal and Ubiquitous Computing, 11 (6), 439–449.
Priyantha, N.B., Chakraborty, A., and Balakrishnan, H.,
2000. The Cricket location-support system. In: Proceedings of the 6th annual international conference on mobile
computing and networking (MobiCom’00), 6–11 August.
Boston, MA. New York: ACM, 32–43.
R Development Core Team, 2009. R: a language and
environment for statistical computing [online]. Vienna: R
Foundation for Statistical Computing. Available from:
. ISBN 3-900051-07-0.
Sae-Hau, C., 2003. Multi-vehicle rover testbed using a new
indoor positioning sensor. Thesis (Master’s). Massachusetts Institute of Technology.
Savvides, A., Han, C.C., and Strivastava, M.B., 2001.
Dynamic fine-grained localization in ad-hoc networks
of sensors. In: Proceedings of the 7th annual international
conference on mobile computing and networking (MobiCom’01), 16–21 July. Rome. New York: ACM, 166–179.
Triggs, B., et al., 2000. Bundle adjustment – a modern
synthesis. In: B. Triggs, A. Zisserman, and R. Szeliski,
eds. Proceedings of the international workshop on vision
algorithms (in conjunction with ICCV’99), 21–22 September 1999, Kerkyra, Corfu, Greece. Lecture Notes on
Computer Science 1883. Berlin: Springer-Verlag, 298–
372.
Welch, G., et al., 2001. High-performance wide-area optical
tracking – the HiBall tracking system. Presence: Teleoperators and Virtual Environments, 10 (1), 1–21.

Wolf, P. and Ghilani, C., 1997. Adjustment computations:
statistics and least squares in surveying and GIS. New
York: Wiley.


International Journal of Computer Integrated Manufacturing
Vol. 23, No. 6, June 2010, 500–514

A real-time simulation grid for collaborative virtual assembly of complex products
X.-J. Zhen, D.-L. Wu*, Y. Hu and X.-M. Fan
CIM Institute, Shanghai Jiao Tong University, Shanghai, China
(Received 9 April 2009; final version received 9 February 2010)
Simulation of collaborative virtual assembly (CVA) processes is a helpful tool for product development. However,
existing collaborative virtual assembly environments (CVAE) have many disadvantages with regard to computing
capability, data security, stability, and scalability, and moreover it is difficult to create enterprise applications in
these environments. To support large-scale CVAEs offering high fidelity and satisfactory interactive performance
among various distributed clients, highly effective system architectures are needed. In this paper, a collaborative
virtual assembly scheme based on grid technology is proposed. This scheme consists of two parts: one is a grid-based
virtual assembly server (GVAS) which can support parallel computing, the other a set of light clients which can
support real-time interaction. The complex and demanding computations required for simulation of virtual
assembly (VA) operations, such as model rendering, image processing (fusion), and collision detection, are handled
by the GVAS using network resources. Users at the light clients input operation commands that are transferred to
the GVAS and receive the results of these operations (images or video streams) from the GVAS. Product data are
managed independently by the GVAS using the concept of RBAC (role-based access control), which is secure
enough for this application. The key related technologies are discussed, and a prototype system is developed based
on the web services and VA components identified in the paper. A case study involving a car-assembly workstation
simulation has been used to verify the scheme.
Keywords: grid; collaborative virtual assembly; complex product; real-time collaborative simulation

Notation

CDM
CVA
CVAE
DM
GVAS
VA
VE
VAGrid

1.

collision detection model
collaborative virtual assembly
collaborative virtual assembly
environment
data management
grid-based virtual assembly server
virtual assembly
virtual environment
virtual assembly based on grid

Introduction

CVA technology is used to develop complex products
such as automobiles and ships. It provides an effective
experimental assembly environment for designers
working at different locations (Lu et al. 2006), who
can exchange product data and discuss and verify the
assembly scheme to improve the previous design
scenario. Many CVA systems or prototypes have

been built for product development (Bidarra et al.
2002, Shyamsundar and Gadh 2002, Chen et al. 2004,
Chryssolouris et al. 2009). However, existing CVA
systems still have many disadvantages. In general,

*Corresponding author. Email:
ISSN 0951-192X print/ISSN 1362-3052 online
Ó 2010 Taylor & Francis
DOI: 10.1080/09511921003690054


virtual environments (VEs) have no modelling function, which means that products must generally be
modelled in a CAD environment, creating product
models that cannot be imported into VE directly; as a
result, much preparatory work such as model transformation must be done before the VA task can be
performed. Moreover, in the context of expanding
requirements for assembly simulation of complex
products, current CVA systems lack adequate computing power. Most CVA systems support only single-PC,
not parallel, computing, which is seriously insufficient
to meet requirements. For example, the frame rate for
rendering a model of a whole car is about 2*6F/S
(frames/second), which is not compatible with interactive operation. In addition, the computing resources
of all user nodes are allocated statically before the task
is begun, which limits the stability and extensibility of
the system.
Grid technology provides a new way to promote
collaborative virtual assembly using the concept of
sharing distinct resources and services. This approach
has been applied successfully in areas of computer
science requiring massive computing, such as parallel

computing and massive data processing, but less so in
the areas of design and manufacturing, especially when


International Journal of Computer Integrated Manufacturing
real-time computing is required. However, the characteristics of grid technology are a perfect fit for the
requirements of a CVA system, and CVA based on grid
technology offers many advantages with regard to
computing power, data security, stability, and scalability, as well as ease of constructing enterprise
applications.
In this paper, a collaborative virtual assembly
scheme based on grid technology is proposed, and a
prototype system called VAGrid (Collaborative Virtual Assembly-based Grid) is developed based on web
service and VA components. This system consists of
two parts: a grid-based virtual assembly server
(GVAS) and a set of light clients. Computing tasks
are handled by the GVAS using resources available
over the internet, with users performing only simple
interactive operations using a graphical interface. The
key related technologies are discussed in detail.
The rest of this paper is organised as follows. In
Section 2, related research on CVA and grid technology is reviewed. Section 3 describes the structure and
workflow of the system. The representative features
and capabilities of VAGrid are described in Section 4.
Section 5 provides a case study, and Section 6 states
conclusions and directions for future work.
2.

Related research

2.1. CVA
Many researchers have already conducted extensive
research into CVA, and significant results have been
achieved. An internet-based collaborative product
assembly design tool has already been developed
(Shyamsundar and Gadh 2002). In this system, a new
assembly representation scheme was introduced to
improve assembly-modelling efficiency. Liang has also
presented a collaborative 3D modelling system using
the web (Liang 2007). Lu et al. developed a collaborative assembly design environment which enabled
multiple geographically dispersed designers to design
and assemble parts collaboratively and synchronously
using the internet (Lu et al. 2006). Web-based virtual
technologies have also been applied to the automotive
development process (Noh et al. 2005, Dai et al. 2006,
Pappas et al. 2006).
These researchers have proposed various approaches to enable collaboration among multiple
designers, but the only interactive modes supported
by the CVA environment, such as chat channels, were
not found to be effective or intuitive. Moreover, the
performance of these systems, especially when supporting real-time assembly activity for complex products, was considered inadequate. In an effort to solve
these problems, the relative performance of various
distribution strategies which support collaborative

501

virtual reality environments, such as client/server
mode, peer-to-peer mode, and several hybrid modes,
has been discussed (Marsh et al. 2006). These
researchers proposed a hybrid architecture which

successfully supported real-time collaboration for
assembly. For supporting the interactive visualisation
of complex dynamic virtual environments for industrial assemblies, a dynamic data model has been
presented, which integrates a spatial data set of
hierarchical model representations and a dynamic
transformation mechanism for runtime adaptation
(Wang and Li 2006). Based on this model, complexity
reduction was accomplished through view frustum
culling, non-conservative occlusion culling, and geometry simplification.
2.2.

Grid computing

Grid computing was first proposed by Ian Foster in the
1990s (Foster and Kesselman, 1999). It aims to share
all the resources available on the internet to form a
large, high-performance computing network. An important characteristic of a grid-computing environment is that a user may connect to the grid-computing
system through the internet, and the grid-computing
system can provide all kinds of services for the user.
Some grid toolkits, such as Globus (Foster 2006),
Legion (Grimshaw and Natrajan 2005), and Simgrid
(Emad et al. 2006), which provide basic capabilities
and interfaces for communication, resource location,
scheduling, and security, primarily use a client-server
architecture based on centralised coordination. These
grid-computing applications use client-server architectures, in which a central scheduler generally manages
the distribution of processing to network resources and
aggregates the processing results. These applications
assume a tightly coupled network topology, ignoring
the changing topology of the network, and are suitable

for large enterprises and long-time collaboration.
However, in the area of production design,
especially for real-time design simulation, few relevant
studies have been reported. Li et al. (2007a, 2007b)
presented the concept of a collaborative design grid
(CDG) for product design and simulation, and a
corresponding architecture was set up based on the
Globus toolkit, version 3.0. Fan et al. (2008) presented
a distributed collaborative design framework using a
hybrid of grid and peer-to-peer technologies. In this
framework, a meta-scheduler is designed to access
computational resources for design, analysis, and
process simulation, which can help in resource
discovery and optimal use of resources. To meet
industrial demands for dynamic sharing of various
resources, Liu and Shi (2008) proposed the concept of
grid manufacturing. According to the characteristics of


502

X.-J. Zhen et al.

the resources to be shared and the technologies to be
used, grid manufacturing distinguishes itself from webbased manufacturing by providing transient services
and achieving true interoperability. To support realtime design simulation, a hybrid of HLA (high-level
architecture) and grid middleware was used (Rycerz
et al. 2007) These systems can solve some problems
from a special viewpoint, but to support collaborative
simulation design for complex products, further

research is needed.
3.

. Multi-user scene synchronisation. Consistency of
the scene across all the system nodes must be
maintained. This means that when the scene at a
particular node changes because of user manipulation, the information must transfer to all other
nodes and update their scenes synchronously.
. Multi-modal interaction. Users interact with VE
through multiple modalities for different hardware, such as the data glove and FOB (flock of
birds), as well as the common keyboard and
mouse as in CAD.

Structure and workflow of VAGrid

3.1. Functions and performance requirements
Complex products such as automobiles and ships have
similar features: (1) numerous components, (2) complex
structure, (3) high research and development costs, and
(4) requirement for a large number of designers. Some
form of collaborative design has been used by most
companies making complex products. However, to date
only some indirect applications, such as physically colocated meetings and CAD-based conferencing, have
been attempted. Unfortunately, these approaches have
many deficiencies with regard to service efficiency,
application effectiveness, and convenience (Trappey
and Hsiao 2008). In contrast, this research targets the
entire distributed team-design scenario, involving all
the participating designers and supervisors. A direct
way is needed to enable geographically distributed

designers to assemble their individual designs together
in real time. To make this possible, the following set of
system requirements should be incorporated into the
development process:
. Ease of use. System configuration, including
allocation of computing resources, is done
automatically, and the user can obtain the
desired results by means of simple interactions.
. Convenient data conversion. Product data requirements in the virtual environment are different from those in the CAD environment, so data
conversion is required. This process should be
simple or automatic and should support common
CAD software such as UG, CATIA, and PRO/E.
. Good real-time scene rendering performance.
The virtual scene at each user station should be
responsive enough to meet the needs of interactive operation, which requires powerful computing resources to perform model rendering,
collision detection, and similar tasks.
. Strong data security. These systems have many
users, including product designers, assembly
technologists, and even component suppliers.
User authorisation or similar measures must be
taken to maintain data security.

3.2. Structure of a grid-based virtual assembly
simulation for complex products
The basic idea of collaborative assembly is shown in
Figure 1. Geometric modelling and assembly design of
products are carried out in a CAD system by designers
at different locations. Then geometric and assembly
information are transferred to the collaborative assembly environment by means of a special data interface.
All designers can share the same virtual assembly

verification environment collaboratively to perform
assembly analyses and assembly process planning for
products, as well as assembly operation training. The
system can run over the internet or on a LAN (Local
Area Network), or particularly on a company intranet
for reasons of performance and stability.
In the context of the functional requirements and
the basic concept described above, several structures
can be used. A distributed parallel architecture (Zhen
et al. 2009) based on HLA and MPI (Message Passing
Interface), with many client nodes and one master
node, has been used, in which each client node was
supported by PCs in a LAN. The advantage of this
approach was fast execution speed, especially at
initialisation because data were saved at each local
position. However, this approach also had many
disadvantages: (1) data protection is difficult to achieve
for distributed data; (2) manual configuration of the
supporting resource nodes and all the resource nodes is
static and unreliable; (3) the system can support only
one task at a time and is therefore not suitable for
general use. For these reasons, an implementation
scheme based on grid technology has been proposed,
as shown in Figure 2. The computing resources needed
during the assembly process for tasks such as rendering, collision detection, and image processing (these
may be any idle resource available on the internet or a
corporate intranet) are managed dynamically by the
grid. Multi-tasking can be supported, but there is only
one virtual assembly scene for each task, maintained
and updated by the grid. Product data and evaluation

results are stored at an independent location. The
configuration of each user is simple, involving only an


International Journal of Computer Integrated Manufacturing

Figure 1.

The basic idea of collaborative virtual assembly.

Figure 2.

CVA solution based on grid.

I/O device or a multi-channel stereo system if using an
immersive virtual environment. The scene at each user
client station is a sequence of continuous images or a
video stream that a user can ‘see’ at his location. Users
are classified as an assembly task manager or a
participant.

503

3.3. Architecture and workflow of VAGrid
The system architecture of VAGrid is illustrated in
Figure 3. All users first register with the system using
the grid portal. Once a user has finished designing
assemblies or subassemblies using a CAD system, the
product models will be submitted to a given location in

504

X.-J. Zhen et al.
assembly path and sequence tracking, dynamic
collision detection, and assembly process evaluation report generation, which can be interactively used by each user.

4. Representative features and capabilities
4.1. Grid management
The grid platform is the main framework of the
system, which handles the management of users
(registration, logon, and user status monitoring), tasks
(startup, maintenance), and resources (monitoring,
dynamic configuration).
Figure 3.

System architecture of VA-Grid.

the grid database. The assembly-task manager accesses
the database with a valid authorisation, prepares the
corresponding data in the VE, and then sets up and
launches a new task. The system will search automatically for resources to support this task. If
insufficient resources are available, a rejection message
will be generated; otherwise, a collaborative assembly
environment will be established. The related users then
can join the task and enter the CVE to perform
assembly verification together.
The system consists of five key parts:
(1) Grid platform management: the basis of the
system. This part provides management of

communication status and computing resources for the CVE. Management of users,
tasks, and resources is also performed by this
part.
(2) Product management: convenient data transformation is an essential requirement for
complex products, and management of these
data in a distributed grid environment is a
complex task.
(3) Remote real-time collaboration: collaboration
based on virtual user models in a virtual
environment and remote real-time interactive
assembly operations are handled by this
module.
(4) Virtual scene graphics management: this module dynamically maintains the virtual scene
displays, including assembly based on solution
of constraints and constraint navigation. The
design and implementation of this module have
already been described in the paper on IVAE
(Yang et al. 2007).
(5) Tools for evaluation and analysis: a set of tools
for distance and dynamic gap measurement,

. User management
Users can be classified into two types: the task
manager and ordinary users. The former is in
charge of the whole task and holds the highest
authority; the latter is the operator in the virtual
environment. Although there is only one virtual
environment, with no scene-synchronisation problems among client nodes, the operating authority of each user is different, and therefore
problems with operating collisions do arise and
must be solved. In this case, the manager defines

roles related to the task and sets levels of
authority according to these roles. When a user
joins a task, he enters into a role and operates
with its corresponding authority.
. Task management
Unlike existing systems, the VAGrid system can
support multiple tasks running in parallel. Each
task has its own group of users, simulation data,
and supporting resources, all managed uniformly
by the task scheduling module. The workflow is
shown in Figure 4. A task manager sets up a new
task and submits it to the task queue. The
resource scheduling module queries the resource
and starts the task by calling CVA services, which
are the basic grid services deployed among the
grid nodes, such as virtual scene graphics management, model rendering, and collision detection.
. Resource management
This is the most complex part of the system.
Large amounts of real-time data must be
processed, and large amounts of many kinds of
computing resources, for example for rendering,
are required. These resources are distributed
among the computing nodes on the internet. A
resource cannot be used before a plug-in is
installed, which is a small program that encapsulates all the resources with the same kind of
access interface, as shown in Figure 5. Each node
registers with the registration centre with an


International Journal of Computer Integrated Manufacturing

Figure 4.

Figure 5.

505

The workflow of tasks management.

Resource encapsulated with the same interface.

attributes message. When a node fails, its work
will be delivered to new resource nodes
dynamically.

4.2. Data management
Data management (DM) is one of the key factors
which provide a flexible and extensible mechanism for
data manipulation in the grid and data delivery to grid
sites based on the participating entities’ interests.
Depending on the system architecture and operation,
users need only to upload CAD models related to the
task to grid data-storage nodes. Once in the virtual
environment, CAD models cannot be loaded directly,
because some preparatory work, including data
storage, processing, and access, is required when
moving from CAD to the virtual environment.
A virtual assembly task requires three main types
of data: product data, including a CAD model and
data in the virtual environment; information on

assembly tools and processes (saved as a file); and
simulation results, as shown in Figure 6. The first type,
product data in the virtual environment, includes the
display model, the part information model, assembly

hierarchy information (saved as a file), and the
collision detection model (CDM). The second data
type consists of important auxiliary information for
evaluation. Simulation results include elements such as
an assembly process video, an assembly sequence, a
path information file, and an evaluation report.
Among these data, product data is managed and
maintained by, and only by, the task manager, because
security requirements dictate that he has the sole
authority to write to the database; others can only
upload data. The second data type is shared by all
users of a particular task. Assembly results are saved in
a folder accessible to users and can be downloaded by
them consistently with their authorisations.
Some preparatory work must also be done for data
transformation. Assembly hierarchy information and
constraint information can be obtained from the CAD
environment after a static interference check. The
display model is the ‘visible’ part, and the CDM is used
to perform a dynamic interference check; these models
must be transformed from CAD models. A special
interface based on the ACIS solid modelling kernel has
been developed for this complex task. Each CAD
model is transformed into a .sat file, from which the
display model and the collision detection model can be

obtained.
. Display model: several types of model can be
used such as .step, .flt (Open Flight), etc. All of
these are polygon models with colour, texture,
and other appearance properties. Several types
are supported by the system, and all the models
will be transformed automatically before the
system starts.
. CDM: This is an internally defined type which
can be transformed by a polygon model as
described in the last step and then simplified into
a hierarchical bounding box depending on the
collision-detection precision.


506

X.-J. Zhen et al.

Figure 6.

Data management in VA-Grid.

Figure 7 shows a display model of a car chassis part
and its collision models at various precisions. All these
data are saved into grid data-storage nodes by the
MYSQL system. When the system starts, all the
computing resource nodes will load their corresponding data according to task requirements.
4.3.

Multi-user collaborative interactive operations

Unlike most existing CVA systems, users at each client
send operation commands to a single ‘remote’ virtual
scene and then receive the results of these operations,
which are scene fragments as seen from the users’
viewpoints. It frequently occurs that many users stay at
their assembly workstations, observing with two eyes,
operating with two hands, where each of them can feel
each other’s presence, but not see each other. This is
real intuitive and effective collaboration.
4.3.1. Co-operation based on virtual user models
According to the system architecture, users at each
client only send operating commands and receive
simulation results over the network; they do not save
product data nor perform computing tasks. The
management of users and of users’ operating commands is very heavy work. Therefore, a virtual-user
model is used, and each real user has a corresponding
virtual user in the virtual environment. The virtual user
processes, not only the current commands from the
user, but the user’s attribute information such as
location, viewpoint parameters, and so forth. The
architecture of the virtual user model is shown in
Figure 8.
Here, in the virtual user attributes information,
‘Type’ can be task manager or general collaborative

user. ‘RoleID’ is the ID number of the role which the
user takes on in the task; for example, if the ID of door
designer is ‘6’, then the ‘RoleID’ of a virtual user acting

as a door designer is ‘6’. Viewpoints are the ‘eyes’ of a
virtual user in the virtual assembly environment; sound
channels are his ‘ears’, operations are his two hands,
and the supporting resource records the list of
computing nodes which provide services to this user.
The system set up a new virtual user object automatically when a client joins in.
In VAGrid, there is no synchronisation problem
among the clients because there is one single virtual
scene, but manipulation conflicts still exist, for
example when:
. An object is grasped by two or more users
simultaneously; or
. Two objects that have an assembly constraint
relation are manipulated by two users
simultaneously.
In the former case, if the object has been grasped,
refuse all grasp requests; otherwise continue. Then the
level of authority and the role of the user are checked;
if they are inappropriate, refuse him; otherwise, accept
his application. Finally record the exact time, so as to
give precedence to the earlier user. In the latter case,
when a user tries to assemble one part with another,
first determine whether it is a free part; if yes, then
carry out the assembly operation, otherwise continue
checking. Then determine whether the operator is the
same person as the user. If yes, this means that the user
is grasping two objects simultaneously, and assembly
can be continued using a two-handed assembly
process; if no, a prompt message that the object is
already being operated on will be generated.

International Journal of Computer Integrated Manufacturing

Figure 8.

507

The architecture of virtual user model.

in real time. An image-based remote interactive
scheme is provided; the basic workflow and hardware
configuration of the client are shown in Figure 9. The
manipulator command is first coded and sent to the
grid over the network, where it is then decoded by the
system and sent to the assembly simulation scene
nodes. Then the parameters of related virtual user
changes and all rendering nodes within ‘RNodeList’
will be updated in response. In the same way, other
computing nodes will also be updated, and a new
video image will arrive at the client workstation. In
addition, users can communicate with each other by
sending text messages.
Using this scheme, users can manipulate virtual
objects conveniently according to their hardware
status. Because the inputs and outputs are separate,
multi-channel immersive stereo can be easily achieved,
as shown in Figure 9.
4.3.3. Interactive operation using virtual hands or by
assembly using tools and equipment in the virtual

environment
Figure 7.

Display model of part and its CDM models.

4.3.2. Remote real-time interactive assembly operation
Many kinds of equipment can be used for interactive
operation, such as FOB (Flock of Birds, a kind of
position-tracking device), data glove, mouse/keyboard, and other I/O equipment. While in the
VAGrid system, the data and its processing program
all reside at the grid site and the result at the client
site is the rendered scene image. To achieve interaction, the user’s command must be sent to the grid site

4.3.3.1. Interactive operation using virtual hands.
There are two virtual hand models in the virtual
scene, corresponding to the real hands of the user.
Virtual hands are driven synchronously by real hands
with the same position and orientation. The basic
manipulation of an object (part or assembly) in the
virtual scene includes grasping, moving, constraint confirmation, motion navigation, and object
release.
Grasping: a user who wants to pick up an object
sends an application request over the grid to the virtual
scene and waits for feedback information. If the object
has been grasped, either a message, ‘cannot be


508

Figure 9.

X.-J. Zhen et al.

The basic workflow and hardware configuration of client.

grasped now,’ will be generated or the grasp will be
successful.
Moving: after a successful grasp, the object is
affixed to the virtual hand and can be moved to
wherever the user wants. The whole scene will be
updated in real time.
Constraint confirmation and motion navigation:
when the object has been moved near to the desired
position, a marker appears (the location label of the
object), and the user can perform a gesture to confirm
the precise mating location based on assembly
constraint recognition.
Object Release: the client (user) sends a command
to release the object, and the relation between the
object and the virtual hand is broken.
4.3.3.2. Assembly using tools and equipment. Taking
the real assembly environment into account, including
assembly tools, fixtures, and assembly equipment, is an
important aspect of virtual assembly simulation.
Interactive assembly operations using assembly tools
should be provided. In the assembly process for a
complex product, equipment can be classified into
three types: automatic, semi-automatic, and manual
tools, as shown in Figure 10. Assembly tools act as a
special ‘assembly’, inheriting all the attributes,

including the collaborative attributes, of the assembly
object. Similarly, a part of an assembly tool inherits all
the attributes of a part object. In addition, the
assembly tool has particular attributes that make it
able to manipulate other objects.
The working process of a virtual tool is similar to
that in the real world. Most tools select their target
object dynamically by colliding with it, except for some
special tools. When a tool operates, for example a
screw tool, it first selects an object by colliding with a
bolt, then creates the axis constraint between the tool

and the target, and finally the tool can be navigated
using the axis constraint to the ending facet which is
the final position.
4.4. Tools for evaluation and analysis in assembly
simulation
To evaluate the assembly operation effectively, several
auxiliary tools are provided, such as distance and
dynamic gap measurement, assembly path record and
replay, trajectory display, and an assisted evaluation
report.
. Distance and dynamic gap measurement
The distance between two points, a point and a
line (a line segment), a point and a facet, or two
facets can be measured interactively. Dynamic
gap computation for two models is also provided. The user can select the measured object,
and the gap value will be shown in the virtual
scene in real time.
. Assembly path and sequence

In the virtual environment, the assembly sequence and the paths followed by the parts are
freely and arbitrarily selected using the data
glove, but there must be an optimal sequence and
set of paths. Recording and replaying the
sequence of motion and the paths followed can
be used to optimise the assembly process.
Trajectory display is used as an intuitive way to
show the path of motion of a component. In the
VAGrid system, replaying the assembly record
must be agreed upon by all the users because
there is only one virtual scene.
. Part interference and collision
The basic function of virtual assembly is to verify
the design intent. But here the interferences


International Journal of Computer Integrated Manufacturing

509

Figure 10. Assembly operation process with screw tool. (a) Select target object. (b) Create axis constraint. (c) Create face
constraint. (d) Run navigated by the axis.

among tools and fixtures can also be calculated,
except for those among parts.
. Assembly process evaluation report
An evaluation report will be created automatically when an assembly task is finished. The
information in the report is divided into three
parts. The first contains general statistical information about the whole assembly task. The
second contains status information for all the

parts. The last one contains interference information, which shows the interference of parts in the
assembly process using a special symbol.

5. Case studies
With the aim of verifying the feasibility of the
system while taking into account the assembly

space, an assembly simulation was performed using
the VAGrid system on a classical workstation to
model a rear suspension and front suspension. The
content of the simulation includes the layout of
fixtures and tools, the operating space, and the
dynamic interferences during the assembly process.
An assembly technologist acted as task manager
and was in charge of setting up the task and
coordinating all the participants. The participants
included designers of the rear suspension and front
suspension, designers of tools and fixtures, and
assembly operators.
To use resources over the internet, plug-ins must be
set up at each resource node. In this case, computing
resources were used over an enterprise intranet, and
product data were saved in the internal database of the
enterprise. The detailed steps followed using VAGrid
can be described as follows.


510
5.1.

X.-J. Zhen et al.
RA750 car assembly workstation simulation

5.1.1. Registration and logon for users
Each user first registers and obtains an account. The
task manager sets up a simulation task team and
defines roles and authority related to the present task.
Figure 11 shows the interface for a user registering at a
portal.
5.1.2. Preparation for simulation
Simulation initialisation consists of three steps: (1) all
CAD models are uploaded to the grid database by
designers from their dispersed locations; (2) the task
manager accesses the database, transforms the models
for simulation, and saves them to a given location; (3)
the folder path together with the car assembly
information are written into a simulation file (.xml as
the default format). A segment of the simulation file is
shown in Figure 12.

resources, rendering resources, and so forth. Once all
related resources are running, the status of this task is
set to be a collaborative task, and users can join in. The
whole assembly scene (at a default viewpoint) can be
browsed by opening an interactive interface, shown in
Figure 13, as part of the assembly scene at the task
manager workstation. Other users can apply for a role
and join the task.
5.1.4. Multi-user collaborative assembly operation
Once all the related users have joined the task, the

collaborative assembly operation will be performed
under the coordination of the task manger. Several
interactive modes can be supported by the system,
among them the ordinary keyboard and mouse as in
CAD, the 5DT Cyberglove and FOB (flock of birds),
and the Cyber Glove/Touch glove with haptic sensing
and FOB. Figure 14 illustrates two interaction modes:
(a) a user operating with a keyboard and mouse; (b) a
user operating with the Cyber Glove/Touch glove and
FOB.

5.1.3. Simulation task initialisation
The task manager starts up the task using the file
described above, and all related service nodes run
automatically to support this task: assembly scene

Figure 11.

User registers by portal.

5.1.5.

Aids to evaluation and analysis

Assembly evaluation tools can be used at any point
during the operation process as needed. To access these