Anda di halaman 1dari 10

Applied Ergonomics 35 (2004) 275284

Automobile seat comfort prediction: statistical model vs. articial


neural network
M. Kolich*, N. Seal, S. Taboun
Department of Industrial & Manufacturing Systems Engineering, University of Windsor, Windsor, Ont., Canada N9B-3P4
Received 20 January 2003; accepted 28 January 2004
Abstract
The current automobile seat comfort development process, which is executed in a trial and error fashion, is expensive and
outdated. The prevailing thought is that process improvements are contingent upon the implementation of empirical/prediction
models. In this context, seat-interface pressure measures, anthropometric characteristics, demographic information, and perceptions
of seat appearance were related to an overall comfort index (which was a single score derived from a previously published 10-item
survey with demonstrated levels of reliability and validity) using two distinct modeling approachesstepwise, linear regression and
articial neural network. The purpose of this paper was to compare and contrast the resulting models. While both models could be
used to adequately predict subjective perceptions of comfort, the neural network was deemed superior because it produced higher r
2
values (0.832 vs. 0.713) and lower average error values (1.192 vs. 1.779).
r 2004 Elsevier Ltd. All rights reserved.
Keywords: Automobile seat; Comfort; Neural network
1. Introduction
The typical approach to automobile seat comfort
development is to rst select a target from the
appropriate vehicle segment. The target, which is usually
a competitive vehicle, is selected through the joint efforts
of engineering, marketing, and program management.
One of the primary considerations is positive consumer
comfort ratings.
Next, the target seat is obtained and benchmarked. As
part of this exercise a subjective evaluation is performed.
This involves an extended duration ride and drive with a
highly structured survey, which directs occupants to
assign feelings of discomfort to specic regions of the
seat. The nature of the subjective evaluation methodol-
ogy makes it necessary to investigate the opinions of
large groups of occupants in order to determine the
impact of various design characteristics on perceived
seating comfort (Manenica and Corlett, 1973). Never-
theless, this feedback, in terms of likes and dislikes
attributable to the seat design, is used to drive comfort
development for the remainder of the program. That is,
prototypes are built and evaluated using the same
subjective evaluation approach. The subjective ratings
obtained from the prototype seat are compared to the
subjective ratings obtained from the target seat. Design
decisions are taken to address the shortcomings of the
prototype seat. This process continues until the proto-
type seat exceeds the comfort level offered by the target
seat. Any improvement, no matter how slight, is
considered success. The purported strength of this
process lies in the A to B comparison of seats. Since a
typical seat program takes 34 years to execute (it is not
uncommon for a program to require 15 separate
subjective evaluations), by the time the product is
launched, it is slightly more comfortable than the best
seat in the market 34 years ago (assuming that the
design team was successful).
From the perspective of providing design direction,
seat system design teams struggle with subjective
evaluations because, while offering credible evaluations
in terms of face validity, the output is poor in terms of
experimental rigor. This creates a scenario in which
prototypes built to address specic comfort issues (i.e.
those arising from one set of subjective evaluations) fail
ARTICLE IN PRESS
*Corresponding author. Johnson Controls, Inc., Automotive
Systems Group, 49200 Halyard Drive, Plymouth, MI 48170, USA.
Tel.: +1-734-254-5911; fax: +1-734-254-6277.
E-mail address: michael.kolich@jci.com (M. Kolich).
0003-6870/$ - see front matter r 2004 Elsevier Ltd. All rights reserved.
doi:10.1016/j.apergo.2004.01.007
to produce the expected results in future subjective
evaluations. That is, design decisions appear suspect/
faulty even though they were appropriate given the
available data.
The current trial and error approach to seat comfort
development is inefcient and outdated. It is extremely
time consuming (the excessively long development time
hinders advances in comfort), expensive, and prone to
measurement error (associated with reliability and
validity (described in the next paragraph). These
limitations could in some ways be justied if the process
could guarantee a comfortable seat. This is, unfortu-
nately, not the case. Since good seats are an exception
and not a rule, the seat comfort development process
needs to be augmented with more efcient and modern
evaluation techniques.
Measurement error associated with subjective evalua-
tions can, at least, partly be addressed through survey
design. Given the automotive seating industrys pro-
pensity to use subjective evaluations, it is surprising that
so little attention has been paid to the quantitative
aspects of survey design and analysis. This lack of
attention is best demonstrated through the shortage of
published literature pertaining to this topic (noteworthy
exceptions include Reed et al., 1991; Shen and Parsons,
1997; Kolich, 1999). A survey should not be used as the
basis for design decisions unless it: (1) provides
consistent ratings and is relatively free of random error
(reliable); and (2) reects exactly what the seat system
design team intends to measure (valid). To satisfy these
criteria, the survey must be designed with special
emphasis on the wording of the survey items, the type
and number of rating scale categories, the verbal tags
associated with the categories, the method of quantica-
tion, and the interest and motivation of the respondent
(as a function of survey length). Kolich (1999) has
published such a survey to be used for automobile seat
comfort development. It represents a signicant im-
provement over the surveys included as part of
traditional seat comfort development processes because
it has proven levels of test-retest reliability, internal
consistency, criterion-related validity, and construct-
related validity (refer to Kolich (1999) for details). The
survey can also be reduced to a single, overall comfort
index (OCI), which minimizes the bias (e.g. vehicle
nameplate, type of trim material (cloth vs. leather), etc.)
thought to plague seat comfort ratings (Kolich, 1999).
Recent advances in seat comfort evaluation technol-
ogies can reduce the time and expense associated with
subjective evaluations, which stems from both schedul-
ing/organizing the subjective evaluation and building
the required prototypes, if, and only if, the output from
the advance (i.e. objective measure) can be linked to
perceptions of comfort. Examples of objective measures
include electromyography (Bush et al., 1995; Lee and
Ferraiuolo, 1993; Sheridan et al., 1991), disc pressure
measurement (Andersson et al., 1974), vibration trans-
missibility (Ebe and Grifn, 2000), and microclimate at
the occupantseat interface (Diebschlag et al., 1988).
One of the better-developed approaches is based on
pressure measurement at the occupantseat interface
(Hertzberg, 1972; Kohara and Sugi, 1972; Chow and
Odell, 1978; Kamijo et al., 1982; Bader et al. 1986;
Diebschlag et al., 1988). Some basic research regarding
the interpretation of pressure distribution proles has
already been conducted (beginning with Akerblom
(1948) and extending to Reed et al. (1991) and Park
and Kim (1997)). The implication is that pressure
distribution characteristics correspond to perceptions
of comfort. The ideal solution, from an automobile seat
development viewpoint, is to formalize the implication
through a model that can be used to predict subjective
perceptions of comfort from quantitative measures. In
this regard, the previously described OCI makes for an
ideal model output.
There are several modeling approaches available. In
practice, stepwise, multiple, linear regression is the most
common. Articial neural networks are another option.
They have not been applied as extensively by ergono-
mists and human factors professionals. Neural networks
take previously solved examples and look for patterns,
learn these patterns, and develop the ability to correctly
classify new patterns (i.e. provide forecasts/predictions).
The basic building block of neural network technology
is the simulated neuron. Independent neurons are of
little use, however, unless they are interconnected in a
network of neurons. The network processes a number of
inputs from the outside world to produce an output (i.e.
the networks predictions). The neurons are connected
by weights and grouped into layers by their association
to the outside world. For example, if a neuron receives
data from outside of the network, it is considered to be
in the input layer. If a neuron contains the networks
predictions, it is in the output layer. Neurons in between
the input and output layers are in the hidden layer,
which serve to: (1) add non-linearity to the system; and
(2) address interactions between input variables. There
can be many hidden layers (i.e. many levels of non-
linearity and many interactions). It should, at this point,
be stated that the addition of hidden layers is only one
of many available neural network based approaches.
The interested reader is referred to Chen (1996), Gately
(1996), Caudill and Butler (1990), and Goldberg (1989)
for additional detail.
The current automobile seat comfort development
process is replete with shortcomings. The common belief
is that seat system design teams need more efcient and
modern evaluation techniques. One alternative may take
the form of prediction models relating comfort ratings
to seat-interface pressure measures, anthropometric
characteristics, demographic information, and percep-
tions of seat appearance (the belief is that occupants are
ARTICLE IN PRESS
M. Kolich et al. / Applied Ergonomics 35 (2004) 275284 276
more likely to provide a positive seat comfort rating if
the seat is aesthetically pleasing (Branton, 1969)). The
purpose of this paper is to compare and contrast two
distinct types of modeling approaches; namely stepwise,
multiple, linear regression and articial neural network.
2. Method
Five different front driver bucket seats, representing a
range of good and bad seats (based on seat comfort
ratings provided by J.D. Power & Associates (1997)),
were obtained from local rental agencies. Only seats
from the compact car segment were selected. This
decision was based on the assumption that seats from
the same segment have comparable H-Point to heel
point relationships (i.e. similar packages/environments).
The H-Point (a) establishes the intended driving/riding
position of each seat, (b) has X, Y, and Z coordinates
relative to the designed vehicle structure, and (c)
simulates the position of the pivot center of the human
torso and thigh, along with associated angles (for more
detail the interested reader is referred to the Society of
Automotive Engineers (1995)). The seats were base level
(i.e. cloth with manual track and recliner). Both the
subjective and objective evaluations were conducted in
vehicle (as opposed to in laboratory). The vehicles, each
designed by a different manufacturer, were white with
gray interior (1997 model year). This precaution was
taken to minimize the effect of color preferences.
The procedure began by obtaining anthropometric
measurements (i.e. standing height and body mass) in a
self-report fashion from 12 occupants. The occupants
were volunteers of working age with no health
problems. Demographics and anthropometry were held
constant by using the same 12 occupants for all ve seats
(repeated measures experimental design). The demo-
graphic and anthropometric details are included in
Table 1. For modeling purposes, females were assigned a
zero and males were assigned a one. Based on an
anthropometric data published by Gordon et al. (1989),
it can be said that the sample spanned a broad range of
the population (in terms of percentiles), which is
important for occupant accommodation. Gender
(SEX), standing height (SH), and body mass (BM) were
considered model inputs.
The seat-interface pressure technology included thin,
exible sensor arrays (manufactured and supplied by
Tekscan). The sensors featured a grid-work of 48
columns and 44 rows based on 10 mm centers. At each
of the 2112 intersection points on the grid, a sensing cell
is created. An electrical resistance inversely proportional
to the pressure applied relative to the cells surface
characterizes each sensing cell. By scanning the grid and
measuring the electrical resistance at each grid point, the
pressure distribution on the sensors surface can be
determined. The scanning electronics are packaged in a
handle assembly that clips onto the sensors interface
tab and provides the electrical connection to each
sensing cell. The sensor arrays (also known as mats)
were calibrated prior to each seat evaluation according
to the instructions outlined by Tekscan, Inc. (1996). The
seat cushion and seatback were tted with the calibrated
mats. These mats were securely attached to the seat
using strips of masking tape. Care was exercised to
ensure that the mats were placed in a consistent location
from occupant-to-occupant and seat-to-seat. Provisions
included: (1) lining up the center of the mat with the
mid-point of the head restraint rods; and (2) tucking the
mats into the biteline, which is dened as the region
where the cushion and seatback converge.
Occupants were not permitted to sit in the seat (on top
of the mats) until they removed their wallets and belts.
This was done to avoid false seat-interface pressure
readings. Each occupant was allowed to adjust the track
position and the seatback angle. In this study, there were
no other seat features to adjust. The preferred setting
was called occupant selected seat position or com-
fort position. Recall that seats from the same market
segment were selected because they were thought to
have comparable H-Point to heel point relationships
(i.e. similar packages/environments). Coupled with the
fact that the same 12 participants were used for all ve
seats, occupant preferred seat position was expected to
be similar between seats. This would not be true if, for
example, van seats were compared to sport car seats
(these two market segments represent vastly different
packages/environments/H-Point to heel point relation-
ships (van seats are typically much higher with more
upright seatback angles)). Once set, the pressure
distribution was recorded. The system software then
ARTICLE IN PRESS
Table 1
Demographic and anthropometric characteristics of occupants evalu-
ating ve different front driver bucket seats using both subjective and
objective methods
Subject Gender Standing
height (cm)
Percentile Body
mass (kg)
Percentile
1 Female 176 90 55 20
2 Male 189 98 132 99
3 Male 198 99 105 98
4 Female 179 99 73 90
5 Male 189 98 82 65
6 Female 178 99 73 90
7 Female 153 5 61 50
8 Male 175 45 79 55
9 Female 154 10 64 60
10 Male 172 30 85 75
11 Female 152 5 73 90
12 Male 164 3 61 5
Mean 173 57 78 66
STD 15 44 21 30
M. Kolich et al. / Applied Ergonomics 35 (2004) 275284 277
produced the following objective measures, which
served as model inputs:
*
cushion contact area (cm
2
)CCA;
*
cushion total force (N)CTF;
*
cushion load at the center of force (N/cm
2
)CCF;
*
cushion peak pressure (N/cm
2
)CPP;
*
seatback contact area (cm
2
)BCA;
*
seatback total force (N)BTF;
*
seatback load at the center of force (N/cm
2
)BCF
*
seatback peak pressure (N/cm
2
)BPP.
The mats were removed and the occupant was asked
to re-enter the seat in order to complete the survey
without interference from the mats (the survey, which
was adopted from Kolich (1999), is shown in Table 2). It
should, at this point, be stated that some occupants
completed the appearance-rating item (a ve-point
scale5 is best) prior to sitting in the seat while others
completed the item after exiting the seat. There was no
standard procedure outlined for when occupants were to
complete the appearance-rating item. The reason should
be obviousit is difcult for occupants to rate the
appearance of the seat if they are sitting in it. The risk
associated with failing to control this experimental detail
was judged to be minimal. The remainder of the survey
was designed so that occupants who were satised with
the comfort or support being assessed by a particular
item would mark the just right box, which, in the
ensuing analysis, corresponded to a score of zero. Most
of the items could be rated from 3 to +3. To obtain a
single score from the survey, the absolute deviation of
each item from just right was summed. Therefore, the
closer the score was to zero, the more comfortable the
occupant. The worst-case score was 30. This score was
considered an OCI. In terms of model development, the
appearance rating (AR) was considered an input while
the OCI was considered an output.
The entire procedure took approximately 30 min to
complete (per occupant). Each seat evaluation was
completed within 1 day. There was a 1-month delay
between seat evaluations owing to the fact that the seat
evaluation was only one part of a much larger
investigative protocol applied to each vehicle. Although
the process was not truly randomized, occupants were
not tested in any particular order (i.e. the order was
denitely not consistent from seat-to-seat).
Descriptive statistics, together with a one-way ANO-
VA (degree of rigor was set to 0.05), were used to
determine which of the metrics (subjective and objective)
could be used to distinguish between seats. For the
one-way ANOVA, the independent variable was seat
(5 levels) and the dependent variables were (1) OCI, (2)
AR, and (3) eight seat-interface pressure measures. This
was necessary in order to demonstrate that the
differences between the selected seats could be quanti-
ed through the adopted protocol.
ARTICLE IN PRESS
T
a
b
l
e
2
R
e
l
i
a
b
l
e
a
n
d
v
a
l
i
d
s
u
r
v
e
y
u
s
e
d
t
o
c
o
l
l
e
c
t
s
u
b
j
e
c
t
i
v
e
d
a
t
a

a
d
o
p
t
e
d
f
r
o
m
K
o
l
i
c
h
(
1
9
9
9
)
S
t
o
p
!
s
t
a
r
t
o
v
e
r
P
o
o
r
,
m
a
j
o
r
i
m
p
r
o
v
e
m
e
n
t
s
n
e
e
d
e
d
F
a
i
r
,
m
i
n
o
r
i
m
p
r
o
v
e
m
e
n
t
s
n
e
e
d
e
d
G
o
o
d
,
s
l
i
g
h
t
i
m
p
r
o
v
e
m
e
n
t
s
n
e
e
d
e
d
W
o
r
l
d
c
l
a
s
s
s
e
a
t
O
v
e
r
a
l
l
s
e
a
t
a
p
p
e
a
r
a
n
c
e
1
2
3
4
5
&
&
&
&
&
I
t
e
m

1
J
u
s
t
r
i
g
h
t
1
2
3
S
e
a
t
b
a
c
k
A
.
A
m
o
u
n
t
o
f
l
u
m
b
a
r
s
u
p
p
o
r
t
T
o
o
l
i
t
t
l
e
&
&
&
&
&
&
&
T
o
o
m
u
c
h
B
.
L
u
m
b
a
r
c
o
m
f
o
r
t
U
n
c
o
m
f
o
r
t
a
b
l
e
&
&
&
&
C
.
A
m
o
u
n
t
o
f
m
i
d
-
b
a
c
k
s
u
p
p
o
r
t
T
o
o
l
i
t
t
l
e
&
&
&
&
&
&
&
T
o
o
m
u
c
h
D
.
M
i
d
-
b
a
c
k
c
o
m
f
o
r
t
U
n
c
o
m
f
o
r
t
a
b
l
e
&
&
&
&
E
.
A
m
o
u
n
t
o
f
b
a
c
k
l
a
t
e
r
a
l
s
u
p
p
o
r
t
T
o
o
l
i
t
t
l
e
&
&
&
&
&
&
&
T
o
o
m
u
c
h
F
.
B
a
c
k
l
a
t
e
r
a
l
s
u
p
p
o
r
t
U
n
c
o
m
f
o
r
t
a
b
l
e
&
&
&
&
G
.
S
e
a
t
b
a
c
k
f
e
e
l
/

r
m
n
e
s
s
T
o
o
s
o
f
t
&
&
&
&
&
&
&
T
o
o

r
m
C
u
s
h
i
o
n
H
.
I
s
c
h
i
a
l
/
b
u
t
t
o
c
k
s
c
o
m
f
o
r
t
U
n
c
o
m
f
o
r
t
a
b
l
e
&
&
&
&
T
o
o
m
u
c
h
I
.
T
h
i
g
h
c
o
m
f
o
r
t
U
n
c
o
m
f
o
r
t
a
b
l
e
&
&
&
&
J
.
C
u
s
h
i
o
n
l
a
t
e
r
a
l
c
o
m
f
o
r
t
U
n
c
o
m
f
o
r
t
a
b
l
e
&
&
&
&
M. Kolich et al. / Applied Ergonomics 35 (2004) 275284 278
Next, 45 of 60 (12 occupants 5 seats) examples were
randomly selected (i.e. 75% of the total data set). This
sample, referred to as the training set, was used to
develop two types of models. The rst was based on a
stepwise, multiple, linear regression approach. The
stepwise selection criteria used for the model develop-
ment were: (1) probability-of-F-to-enter=0.05 and (2)
probability-of-F-to-remove=0.10. These were default
criteria in SPSS, Inc.s (2000) statistical software
package.
The second was a neural network developed with the
help of NeuroShell
s
Predictor (commercial computer
software available through Ward Systems Group, Inc.
(1997)). As part of this process, input neurons were
created for every input variable and an output neuron
was created for the output variable (i.e. OCI). In the
context of traditional statistical modeling, this is
synonymous with identifying independent and depen-
dent variables. When the value of an input variable is
fed into an input neuron, the neuron is activated, along
with its links to other neurons. All of the input neurons
are rst directly linked to the output neuron (linear
relationship). Weight values are assigned to the links,
which indicate the strength of the connection. After the
preliminary relationships are found, neurons are added
to the hidden layer so that nonlinear relationships can
be found. Input values in the rst layer are multiplied by
the weights and passed to the second (hidden) layer.
Neurons in the hidden layer produce outputs that are
based upon the sum of weighted values passed to them.
The hidden layer passes values to the output layer in the
same fashion, and the output layer produces the desired
results (predictions). The neural network learns by
adjusting the interconnection weights between layers.
The neural networks predictions are repeatedly com-
pared with the correct answers, and each time the
connecting weights are adjusted slightly in the direction
of the correct answers. Additional hidden neurons are
added as necessary to capture features in the data set.
Eventually, if the problem can be learned, a stable set of
weights evolves and will produce good answers for all
of the sample decisions or predictions. The real power of
neural networks is evident when the trained network is
able to produce good results for data that the network
has not previously encountered. The specic architec-
ture of the neural network, in terms of the number of
hidden neurons, must be determined. Too many hidden
neurons can hinder the neural networks ability to
generalize to data not encountered during training,
which is referred to as over-tting. Too few can cripple
the neural networks ability to learn the relationships at
hand.
Both models were assessed using standard statistics.
The primary goal was to maximize the r
2
value and
minimize the average error for the training set, while
avoiding over-tting (which would compromise validity
as indicated by a poor cross-validated r). The remaining
25% of the total data, referred to as the test set, was
used for validation. For the sake of completeness,
average errors for the test set were also computed.
3. Results
Prior to modeling, it was necessary to determine
whether the adopted protocol could be used to
distinguish between seats. If, given the selected metrics,
all the seats were the same, then the creation of models
would be unnecessary. Along these lines, descriptive
statistics (Table 3), together with the one-way ANOVA
(Table 4), were used to demonstrate that a few of the
identied measures could be used to distinguish between
seats. Specically, OCI differences were statistically
signicant. The post hoc results (Table 5) revealed that
all comparison were different with the exception of Seat
B (mean=10.3) and Seat D (mean=8.6). Seat C
(mean=2.3) was the most comfortable. Similarly, AR
differences were statistically signicant (Table 4). The
post hoc results (Table 5) suggested that Seat B
(mean=2.8) was among the least aesthetically pleasing,
while Seat C (mean=4.4) was among the most. In terms
of seat-interface pressure measures, the results revealed
that only CPP and BCF could be used to quantitatively
distinguish between seats (i.e. the differences were
statistically signicantTable 4). Considering the post
hoc results for CPP (Table 5), Seat C (mean =1.5 N/
cm
2
) was different than Seat E (mean=0.8 N/cm
2
),
whereas, the post hoc results for BCF (Table 5)
suggested that Seat B (mean=0.2 N/cm
2
) and Seat C
(mean=0.2 N/cm
2
) were different than Seat E
(mean=0.4 N/cm
2
).
The relationship between each of the 12 input
variables (i.e. SEX, SH, BM, AR, CCA, CTF, CCF,
CPP, BCA, BTF, BCF, and BPP) and the OCI (output
variable representing subjective perceptions of comfort)
was examined using Pearson product moment correla-
tion coefcients (Table 6). Only three of the variables
(AR, BCF, and CPP) were statistically related to the
OCI at the 0.05 level. The fact that three input variables
were linearly related to the OCI suggested that a viable
linear model could be developed.
Not only were the majority of relationships non-
linear, they were non-quadratic. This became apparent
when the input variables were plotted against the OCI
(scatter plots) in a separate exercise. This was a
surprising result, particularly with respect to the seat-
interface pressure measures. One would expect an
optimal amount of CCA, for example, with too much
or too little CCA being comparably detrimental. Since
the one-to-one relationships were not as straightforward
as one would hope, the data were assumed to be beset
with several important interactions and/or complicated
ARTICLE IN PRESS
M. Kolich et al. / Applied Ergonomics 35 (2004) 275284 279
levels of non-linearity. This triggered the application of
neural network technology.
As mentioned, 75% of the total sample was randomly
selected and used to develop the stepwise, multiple,
linear regression model and the neural network. The
nal linear model can be expressed as follows:
OCI 13:749 2:038 AR6:32314 BCF
0:99796 CPP 0:01046 BTF
0:0204 CTF 0:133 BM:
The model explained 71.3% of the variance in OCI
with an average error of 1.8 (Table 7). This can be
contrasted with the nal neural network, which con-
tained 31 hidden neurons. This number produced the
maximum r
2
values and the minimum average error
values for both the training and test sets. The neural
network explained 83.2% of the variance in OCI with an
average error of 1.2 (Table 7). The remainder of the total
sample (i.e. 25%) was used for validation purposes. In
this regard, both models resulted in a signicant
relationship between the actual and predicted OCI
ratings (as demonstrated through the high cross-
validated r-values and low average error valuesTable
7). Both models were considered valid.
In addition to predicting the output for the test set,
the models were used to determine which variables were
most effective at estimating the OCI. This information
provided a relative measure of the signicance of each
input variable (in terms of its ability to predict the
output). Weights could range from zero to one.
Higher values were associated with more important
variables (inputs). Since the sum of all importance
values was approximately one, the importance values
were thought of as the percent contribution to the
model (Table 8).
4. Discussion
As stated, the automobile seat design process pro-
ceeds by comparison with existing models (i.e. targets).
From a comfort perspective, seats evolve only because
the design objective is to exceed the performance of the
target. Even marginal improvement is considered
success. Rarely is statistical signicance considered. It
should also be noted that, often times, the process fails.
For these reasons, automobile seat comfort has not
improved/evolved substantially over the last decade.
The prevailing thought is that prediction models may
ARTICLE IN PRESS
Table 3
Descriptive statistics for overall comfort index, appearance rating, and seat interface pressure measures
Seat OCI AR CCA (cm
2
) CTF (N) CCF (N/cm
2
) CPP (N/cm
2
) BCA (cm
2
) BTF (N) BCF (N/cm
2
) BPP (N/cm
2
)
A Mean 6.0 3.8 1717 598 0.3 1.1 1318 273 0.3 0.7
STD 2.2 0.7 113 160 0.2 0.3 191 74 0.2 0.3
Min 2 2.5 1585 377 0.0 0.7 1086 192 0.0 0.5
Max 11 4.5 1967 1010 0.7 1.6 1653 422 0.5 1.3
B Mean 10.3 2.8 1699 588 0.3 1.2 1338 240 0.2 0.7
STD 1.9 0.6 122 194 0.1 0.5 248 74 0.1 0.2
Min 7 2.0 1494 367 0.0 0.6 990 137 0.0 0.4
Max 13 4.0 1964 1066 0.4 2.4 1896 363 0.4 1.0
C Mean 2.3 4.4 1746 697 0.2 1.5 1342 277 0.2 1.1
STD 1.1 0.6 112 172 0.2 0.7 281 108 0.1 0.9
Min 1 3.0 1623 537 0.0 0.6 850 140 0.0 0.4
Max 4 5.0 2002 1186 0.5 3.4 1908 518 0.4 2.9
D Mean 8.6 3.8 1630 564 0.3 1.1 1219 250 0.3 0.7
STD 1.3 1 119 155 0.2 0.3 183 88 0.1 0.2
Min 6 2.5 1494 359 0.1 0.6 1005 126 0.1 0.5
Max 10 5.0 1917 958 0.5 1.7 1711 451 0.5 1.0
E Mean 12.8 3.2 1725 579 0.2 0.9 1358 322 0.4 0.7
STD 1.4 0.7 117 149 0.1 0.3 254 116 0.1 0.2
Min 10 2.0 1494 424 0.0 0.5 953 154 0.1 0.5
Max 15 4.5 1948 970 0.4 1.6 1978 557 0.6 1.1
Total Mean 8 3.6 1703 605 0.3 1.2 1315 272 0.3 0.8
STD 4 0.9 120 168 0.1 0.5 232 95 0.1 0.5
Min 1 2.0 1494 359 0.0 0.5 850 126 0.0 0.4
Max 15 5.0 2002 1186 0.7 3.4 1978 557 0.6 2.9
M. Kolich et al. / Applied Ergonomics 35 (2004) 275284 280
ameliorate this situation by helping seat system design
teams understand, earlier in the development process,
how to exceed the comfort level offered by the target.
The new found efciency may (1) create shorter product
life cycles or (2) provide design teams with more time to
develop and integrate innovative ideas/solutions for
comfort enhancement. In either case, seat comfort will
advance at a faster rate.
This research found that both an articial neural
network and a stepwise, multiple linear regression model
could be used to adequately predict subjective percep-
tions of comfort. The decision regarding which model to
advocate was, rst and foremost, based on r
2
values and
error estimates derived from the training set. The
secondary consideration was to avoid over-tting, which
would compromise validity as indicated by a poor cross-
validated r (derived from the test set). The neural
network was deemed superior to the regression model
because, while still validating (i.e. generalizing well), it
explained more of the variance in OCI with lower
average error. The neural networks ability to deal with
interaction effects is offered as the principle reason for
its superior performance.
The neural network, which considers a larger number
of inputs, is also more useful to seat system design teams
because constraints (which are an inevitable part of the
design process) are less limiting. That is, the more
options design teams have for how to improve comfort,
the better. Note also that the regression model
categorizes occupants according to only body mass.
The neural network considers body mass, standing
height, and gender. For this reason the neural network is
capable of determining how a particular set of inputs
will affect a more focused subset of a target population.
ARTICLE IN PRESS
Table 4
One-way ANOVA for seat-interface pressure measure differences between seats
Sum of squares df Mean square F Sig.
Overall Between seats 780.267 4 19.067 74.677 0.000
Comfort Within seats 143.667 55 2.612
Index Total 923.933 59
Appearance Between seats 18.400 4 4.600 8.680 0.000
Rating Within seats 29.146 55 0.530
Total 47.546 59
CCA Between seats 94,346.798 4 23586.699 1.731 0.156
Within seats 749,382.925 55 13625.144
Total 843,729.723 59
CTF Between seats 133,067.292 4 33266.823 1.195 0.323
Within seats 1,531,350.961 55 27842.745
Total 1,664,418.253 59
CCF Between seats 554.233 4 138.558 0.604 0.661
Within seats 12,613.500 55 229.336
Total 13,167.733 59
CPP Between seats 30,635.136 4 7658.784 3.632 0.011
Within seats 115,966.372 55 2108.479
Total 146,601.507 59
BCA Between seats 147,976.804 4 36994.201 0.674 0.613
Within seats 3,020,261.389 55 54913.843
Total 3,168,238.193 59
BTF Between seats 47,893.236 4 11973.309 1.369 0.257
Within seats 480,967.674 55 8744.867
Total 528,860.910 59
BCF Between seats 3436.433 4 859.108 4.930 0.002
Within seats 9584.500 55 174.264
Total 13,020.933 59
BPP Between seats 18,360.978 4 4590.244 2.278 0.072
Within seats 110,814.024 55 2014.800
Total 129,175.002 59
M. Kolich et al. / Applied Ergonomics 35 (2004) 275284 281
Based on both models, CTF was the most important
predictor of OCI. This raises an interesting discussion
topic in that CTF was not one of the pressure measures
found to distinguish between seats (i.e. all ve seats had
similar CTFs). The explanation lies in the fact that
occupants, within specic seats, tended to be more
comfortable when there was high CTF. Design teams
interested in improving comfort should, therefore, focus
on this variable. It is somewhat intuitive to suggest that
CTF can be increased through the provision of a more
upright recline angle. As the recline angle becomes more
upright, more of the occupants mass is taken up by the
cushion and less by the seatback. In practice, this type of
design solution would require a change in H-Point
ARTICLE IN PRESS
Table 5
Post-hoc (Tukeys honestly signicant difference test) in support of one-way ANOVA
(I)Seat (J) Seat Mean difference (IJ) Sig. (I)Seat (J) Seat Mean difference (IJ) Sig.
Overall comfort index CPP
A B 4.25 0.00 A B 12.44 0.96
C 3.75 0.00 C 46.02 0.12
D 2.58 0.00 D 2.08 1.00
E 6.75 0.00 E 22.49 0.75
B A 4.25 0.00 B A 12.44 0.96
C 8.00 0.00 C 33.58 0.39
D 1.67 0.10 D 14.53 0.94
E 2.50 0.00 E 34.93 0.35
C A 3.75 0.00 C A 46.02 0.12
B 8.00 0.00 B 33.58 0.39
D 6.33 0.00 D 48.10 0.09
E 10.50 0.00 E 68.51 0.01
D A 2.58 0.00 D A 2.08 1.00
B 1.67 0.10 B 14.53 0.94
C 6.63 0.00 C 48.10 0.09
E 4.17 0.00 E 20.41 0.81
E A 6.75 0.00 E A 22.49 0.75
B 2.50 0.00 B 34.93 0.35
C 10.50 0.00 C 68.51 0.01
D 4.17 0.00 D 20.41 0.81
Appearance rating BCF
A B 0.96 0.02 A B 7.92 0.59
C 0.63 0.23 C 9.08 0.45
D 0.04 1.00 D 3.17 0.98
E 0.63 0.23 E 11.50 0.22
B A 0.96 0.02 B A 7.92 0.59
C 1.58 0.00 C 1.17 1.00
D 1.00 0.01 D 11.08 0.25
E 0.33 0.79 E 19.42 0.01
C A 0.63 0.23 C A 9.08 0.45
B 1.58 0.00 B 1.17 1.00
D 0.58 0.30 D 12.25 0.17
E 1.25 0.00 E 20.58 0.00
D A 0.04 1.00 D A 3.17 0.98
B 1.00 0.01 B 11.08 0.25
C 0.58 0.30 C 12.25 0.17
E 0.67 0.18 E 8.33 0.54
E A 0.63 0.23 E A 11.50 0.22
B 0.33 0.79 B 19.42 0.01
C 1.25 0.00 C 20.58 0.00
D 0.67 0.18 D 8.33 0.54
M. Kolich et al. / Applied Ergonomics 35 (2004) 275284 282
(recall that the H-Point establishes the intended driving/
riding position of each seat (i.e. design position)this
includes the torso angle).
It was also interesting to note that AR was related to
OCI. Although correlation does not imply causality,
automobile seat design studios would, almost denitely,
be interested in knowing that perceptions of seat
appearance are related to perceptions of seat comfort.
This nding substantiates the claim originally made by
Branton (1969).
Having established that a neural network can be used
to predict automobile seat comfort, future research
should focus on the derivation of an optimized set of
inputs (i.e. a set of inputs that will result in maximum
comfort). There are many optimization algorithms
available for this purpose. Also, future investigations
should broaden the scope of the models to include seats
from other market segments (i.e. in addition to compact
car) and other seating positions (i.e. in addition to front
driver). The assumption is that expectations of auto-
mobile seat comfort vary by market segment (compare
the compact car consumer to the luxury car consumer)
and seating position (compare passengers to drivers,
which are much more constrained). Once completed, the
ndings could be published in the form of seat comfort
design guidelines/standards.
Perceptions of comfort are constantly changing. A
comfortable automobile seat from 1970, for example,
would probably not be considered comfortable today.
This suggests that prediction models, in order to remain
useful, will need to be periodically updated. Consumer
researchers, at both the vehicle manufacturer and seat
supplier levels, will have a vital role to play in specifying
the frequency of these updates.
In all applications, comfort degrades with time on
task (Zhang et al., 1996). A recent article by Helander
(2003) found that all seats, even those that conform to
some basic set of ergonomic criteria, decrease comfort.
This was attributed to the physiological aspects of
sitting for extended durations rather than the seat
design. So while comfort will decrease over time, the
rank order of preference among a set of seats will not
change over time. Said another way, occupants can
reliably assess comfort immediately. It can be argued
that this immediate evaluation (sometimes referred to as
static or showroom comfort) is what sells automobiles.
The present investigation was focused on this type of
comfort. Having said this, it would still be valuable, as
part of future research, to understand the time
dependency associated with seat-interface pressure
measures. Unfortunately, with the current state of
technology, this will not be easy. Todays pressure
sensors are obtrusive (i.e. they may, themselves, cause
the occupant to modify his/her posture or to inuence
comfort directly), impractical to use when driving,
adversely affected by the seated occupant over a long
drive, and cumbersome from a data management
perspective (Gyi et al., 1998). There is, in short, a
technological divide, which represents a product oppor-
tunity. This technological divide may be why Porter et al.
(2003), who considered driving comfort a dynamic
phenomenon, were unsuccessful in relating subjective
perceptions of comfort to seat-interface pressure char-
acteristics.
ARTICLE IN PRESS
Table 8
Comparison of relative importance of input variables
Stepwise linear
regression
Articial neural
network
SEX 0.009
HT 0.203
BM 0.261 0.018
AR 0.161 0.024
CCA 0.002
CTF 0.335 0.658
CCF 0.011
CPP 0.050 0.012
BCA 0.034
BTF 0.100 0.010
BCF 0.093 0.008
BPP 0.011
Table 6
Relationship between predictor variables and overall comfort index
(n 60)
Predictor variable r p
GenderSEX 0.076 0.562
Standing height (cm)HT 0.163 0.213
Body mass (kg)BM 0.031 0.817
Appearance ratingAR 0.645 0.000
Cushion contact area (cm
2
)CCA 0.181 0.166
Cushion total force (N)CTF 0.219 0.092
Cushion load at the center of force (N/cm
2
)CCF 0.032 0.808
Cushion peak pressure (N/cm
2
)CPP 0.381 0.003
Seatback contact area (cm
2
)BCA 0.062 0.635
Seatback total force(N)-BTF 0.203 0.119
Seatback load at the center of force (N/cm
2
)BCF 0.505 0.000
Seatback peak pressure (N/cm
2
)BPP 0.201 0.124
Table 7
Comparison of model performance
Performance
statistics
Stepwise, linear
regression
Articial neural
network
Model development
r
2
0.713 0.832
Average error 1.8 1.2
Model validation
Cross-validated-r r (15)=0.952, p 0 r (15)=0.847, p 0
Average error 0.5 0.7
M. Kolich et al. / Applied Ergonomics 35 (2004) 275284 283
References
Akerblom, B., 1948. Standing and sitting posture with special reference
to the construction of chairs. Unpublished Doctoral Dissertation,
Karolinska Institute, A.B. Nordiska Bokhandeln, Stockholm,
Sweden.
Andersson, B.J.G., Ortengren, R., Nachemson, A., Elfstrom, G., 1974.
Lumbar disc pressure and myoelectric back muscle activity during
sitting. IV. Studies on a car drivers seat. Scand. J. Rehabil. Med. 6,
128133.
Bader, D.L., Barnhill, R.L., Ryan, T.J., 1986. Effect of externally
applied skin surface forces on tissue vascularization. Arch. Phys.
Med. Rehabil. 67 (11), 807811.
Branton, P., 1969. Behavior, body mechanics and discomfort.
Ergonomics 12, 316327.
Bush, T.R., Mills, F.T., Thakurta, K., Hubbard, R.P., Vorro, J., 1995.
The use of electromyography for seat assessment and comfort
evaluation. Technical Paper No. 950143, Society of Automotive
Engineers, Inc., Warrendale, PA, USA.
Caudill, M., Butler, C., 1990. Naturally Intelligent Systems. The MIT
Press, Cambridge, MA, USA.
Chen, C.H., 1996. Fuzzy Logic and Neural Network Handbook.
McGraw-Hill, New York.
Chow, W.W., Odell, E.I., 1978. Deformations and stresses in soft body
tissues of a sitting person. J. Biomech. Eng. 100, 7987.
Diebschlag, W., Heidinger, F., Kuurz, B., Heiberger, R., 1988.
Recommendation for ergonomic and climatic physiological vehicle
seat design. Technical Paper No. 880055, Society of Automotive
Engineers, Inc., Warrendale, PA, USA.
Ebe, K., Grifn, M.J., 2000. Qualitative models of seat discomfort
including static and dynamic factors. Ergonomics 43 (6), 771790.
Gately, E., 1996. Neural Networks for Financial Forecasting. Wiley,
New York.
Goldberg, D., 1989. Genetic Algorithms in Search, Optimization &
Machine Learning. Addison-Wesley, Reading, MA, USA.
Gordon, C.C., Churchill, T., Clauser, C.E., Bradtmiller, B., McCon-
ville, J.T., Tebbetts, I., Walker, R.A., 1989. 1988 Anthropometric
survey of US army personnel: methods and summary statistics.
Technical Report NATICK/TR-89/044, Anthropology Research
Project, Yellow Springs, OH, USA.
Gyi, D.E., Porter, J.M., Robertson, N.K.B., 1998. Seat pressure
measurement technologies: consideration for their evaluation.
Appl. Ergon. 27 (2), 8591.
Helander, M.G., 2003. Forget about ergonomics in chair design?
Focus on aesthetics and comfort!. Ergonomics 46, 13061319.
Hertzberg, H.T.E., 1972. The human buttocks in sitting: pressures,
patterns, and palliatives. Technical Paper No. 72005, Society of
Automotive Engineers, Inc., New York.
J.D. Power & Associates, 1997. 1997 Seat Quality Report. J.D. Power
& Associates, Agoura Hills, CA.
Kamijo, K., Tsujimura, H., Obara, H., Katsumat, M., 1982.
Evaluation of seating comfort. Technical Paper No. 820761,
Society of Automotive Engineers, Inc., Warrendale, PA.
Kohara, J., Sugi, T., 1972. Development of biomechanical manikins
for measuring seat comfort. Technical Paper No. 720006, Society
of Automotive Engineers Inc., New York, NY.
Kolich, M., 1999. Reliability and validity of an automobile seat
comfort survey. Technical Paper No. 990132, Society of Auto-
motive Engineers Inc., Warrendale, PA, USA.
Lee, J., Ferraiuolo, P., 1993. Seat comfort. Technical Paper No.
931005, Society of Automotive Engineers Inc., Warrendale, PA,
USA.
Manenica, I., Corlett, E.N., 1973. Model of vehicle comfort and a
model for its assessment. Ergonomics 11 (6), 849854.
Park, S.J., Kim, C.B., 1997. The evaluation of seating comfort by the
objective measures. Technical Paper No. 970595, Society of
Automotive Engineers, Inc., Warrendale, PA, USA.
Porter, J.M., Gyi, D.E., Tait, H.A., 2003. Interface pressure data and
the prediction of driver discomfort in road trials. Appl. Ergon. 34,
207214.
Reed, M.P., Saito, M., Kakishima, Y., Lee, N.S., Schneider, L.W.,
1991. An investigation of driver discomfort and related seat design
factors in extended-duration driving. SAE Technical Paper No.
910117, Society of Automotive Engineers, Inc., Warrendale, PA,
USA.
Shen, W., Parsons, K.C., 1997. Validity and reliability of rating
scales for seated pressure distribution. Int. J. Ind. Ergon. 20,
441461.
Sheridan, T.B., Meyer, J.E., Roy, S.H., Decker, K.S., Yanagishima,
T., Kishi, Y., 1991. Physiological and psychological evaluations of
driver fatigue during long term driving. SAE Technical Paper No.
910116, Society of Automotive Engineers, Inc., Warrendale, PA,
USA.
Society of Automotive Engineers, 1995. Surface vehicle standard.
Devices for use in dening and measuring vehicle seating
accommodation, SAE J826. Issued 1962-11 Revised 1995-07.
Society of Automotive Engineers, Inc., Warrendale, Palestinians,
USA.
SPSS, Inc., 2000. SPSS for Windows (release 10.1.0). Computer
software.
Tekscan, Inc., 1996. Tekscan body pressure distribution measurement
system (Version 3.843). Computer software.
Ward Systems Group, Inc., 1997. NeuroShell
s
Predictor (Release 2.0).
Computer software.
Zhang, L., Helander, M.G., Drury, C.G., 1996. Identifying factors of
comfort and discomfort in sitting. Hum. Factors 38 (3), 377389.
ARTICLE IN PRESS
M. Kolich et al. / Applied Ergonomics 35 (2004) 275284 284