Robotics
and
Computer Integrated
Manufacturing
PERGAMON
Robotics and Computer Integrated Manufacturing 18 (2002) 197-203
www.elsevier.com/locate/rci
m
Impact of human error on lumber yield in rough mills
Urs Buehlmann
a,
*, R. Edward Thomas
b
a
North Carolina State University, Department of Wood and Paper Science, 3036B Biltmore Hall, Raleigh, NC 27695-8003, USA
b
USDA Forest Service, Northeastern Research Station, 241 Mercer Springs Road, Princeton, WV 24740 USA
Abstract
Rough sawn, kiln-dried lumber contains characteristics such as knots and bark pockets that are considered by most people to be defects.
When using boards to produce furniture components, these defects are removed to produce clear, defect-free parts. Currently, human
operators identify and locate the unusable board areas containing defects. Errors in determining a defect and its location, known as
operator error, lead to lower lumber yield and increased product cost. Technology exists that would alleviate these problems and is a viable
option to avoid wasting lumber because' of human error. This study was performed in a rough mill collecting data on the errors made by
humans when marking defects. Computer-based simulation tools were used to assess the significance of these errors. It was found tha
t
three-quarters of the decisions made by human operators are erroneous in some way resulting in an absolute yield loss of approximately
16.1 %. Thus, automated defect detection systems that perform more accurately than do humans could have a payback
p
eriod of 1 year o
r
less. Published by Elsevier Science Ltd.
K
eywords: Human error; Material loss; Lumber yield; Wood processing
1. Introduction
Lumber costs are the single most important cost for
furniture manufacturers. Between 12% and 15% of the
furniture production cost, depending on style and quality of
the furniture, is due to lumber costs [4]. A popular rule of
thumb states that saving 1 % of the incoming raw material
(i.e. lumber) reduces total production costs by as much as
2% [5]. As a result, manufacturers are aggressively trying
to improve yield. Yield is defined as the ratio of aggregate
part surface area output in relationship to aggregate lumber
surface area input [6].
Wood is a non-homogeneous material with unusable
areas randomly dispersed throughout the board. Each board
is classified in appropriate quality classes based on defect
sizes, locations, frequencies, and other
geometric
characteristics [7]. Boards of the same quality class are
then processed in rough mills that either employ crosscut-
first or rip-first sawing technology. Ripfirst technology is
the most commonly used technology today [8]. Classifying
all the different natural characteristics of lumber is a
difficult task performed by human operators. Some board
characteristics are considered acceptable, others as a
defect. It is the responsibility
of the operators, or "marker"
as they are called in the wood industry, to determine which
Solid wood dimension parts for furniture and cabinets are
cut from rough sawn, kiln-dried, random length, and rando
m
width lumber. In a rough mill, boards are processed into
rough-sized furniture parts utilizing two processing
methods: rip-first or crosscut-first. Rip-first processing
b
egins by ripping the board into narrow strips, then
chopping the strips into part lengths. Crosscut-firs
t
p
rocessing chops the board to the part lengths, then rips the
b
oard segments to the correct widths. This paper is
concerned only with rip-first processing. For a thorough
discussion of rough mills see [1]. Furniture parts are mainly
produced using hardwood lumber, however, the price o
f
hardwood lumber has almost doubled during the last 20
years [2]. During this same period, domestic furniture
manufacturers were competing with imported, low price,
good quality, solid wood furniture in most retail marke
t
sectors. In fact, furniture imports have grown significantly
over the last decade, comprising one-third of all furniture
sales in the US [3]. Not surprisingly, US manufacturers are
striving to minimize their production costs to compete with
foreign producers.
*Corresponding author. Tel.: + 1-919-515-5580; fax: + 1-919-
515-8739.
E-mail address:urs
_
buehlmann
@
ncsu.edu
(
U. Buehlmann
)
.
0736-5845/02/$ - see front matter Published by Elsevier Science Ltd.
PII: S 0 7 3 6 - 5 8 4 5 ( 0 2 ) 0 0 0 1 0 - 8
198
U Buehlmann, RE. Thomas / Robotics and Computer Integrated Manufacturing 18 (2002) 197-203
Fig. 1. Two strips with defects that are perfectly marked.
characteristics are acceptable and which are not. Operators
are constantly fed narrow board strips from the ripsaw ove
r
roller conveyors. Both sides of each strip are inspected an
d
defects are marked with a scanner-readable chalk. A
computerized saw with a camera reads these marks and
optimizes the fit of parts into the clear areas, then cuts the
strips into segments of clear wood with no defects and
unusable segments containing defects. Fig. 1 shows two
strips as an operator could encounter them. The black areas
symbolize defects and cannot be used for the finished
product. The operator has to distinguish between clear areas
of the strip and the areas containing defects. The marks
encompassing the defects in Fig. 1 show the two strips
marked perfectly, i.e marked such that no clear wood is
wasted.
However, due to the difficulties in distinguishing some
defects from sound wood, the high speed at which operators
have to work, and the long hours on the job, human errors
occur. Operators theoretically can make four decisions when
marking strips. These decisions are shown in the decision
matrix in Fig. 2.
As shown in the decision matrix, an operator has a choice
of four decisions when marking strips. Two are correct and
two incorrect (Fig. 2). Positive outcomes result under one o
f
two conditions: a defect is identified and removed (Yes-
Yes) and defect free wood is not marked for removal (No-
N
o). Costly errors can result from the other two conditions.
When a defect-free area is marked as a defect, then usable
wood is wasted (Yes-
N
o, Type II error). When a defect is
not detected (No-Yes, Type I error), then this defective area
will be used to produce a part that will either require costly
remanufacturing or the part will be discarded later in the
p
roduction process. Rejects are especially costly as they
result not only in a loss of wood; but labor as well. In the
last case, when the operator marks an area that contains no
defect, (Yes-
N
o, Type II error), then a negative outcome is
achieved in that clear, usable wood is cut out and wasted.
Two other errors can occur: (1) incorrectly marking the
end of a defect, leaving a small portion of the defect in the
b
oard to be used for a part (Fig. 3, A), and (2) marking the
end of a defect beyond its true end resulting in wasted
usable material (Fig. 3, B).
Both errors A and B are mistakes, which can lead to
waste material and increased production costs. How
ever, error A is of more concern since it may lead to a part
containing a defect that will be rejected later in the
p
roduction process. The loss of usable wood is an error,
which is present regardless of human error in defect
Fig. 3. Two special cases of operator error: (A) when mark is within
defect, and (B) when mark is beyond defect.
identification since the optimization of a strip's clear areas
is rarely able to find chop solutions that use 100% of the
clear area. Since a minimal amount of usable wood is los
t
when optimizing the clear areas of a strip, error B is
increasing costs only marginally. Therefore, we did no
t
further investigate this error. In summary, there are three
operator errors that are of significance and will lead to
either a loss of usable, clear wood or to rejects because
defects will be contained in the part:
(1) Type II error, detection of a defect where there is
none.
(2) Type I error, when a defect goes undetected.
(3) Incorrect marking of a defect within its boundaries
(Fig. 3, A). We will refer to this type of error as
"partial Type I error."
Little knowledge exists concerning the amount of lumbe
r
or money lost due to operator error. However, Huber et al.
published a paper in 1985 concerning the ability o
f
operators to correctly detect and mark defects in boards [9].
This study assessed the performance of six experience
d
operators from three different plants by asking them to
assess a board in 1 min and to memorize the location an
d
type of defects. The operators then used an eight (length o
f
the board) by two (width of the board) matrix to indicate
the location and type of defects found in the board. This
test was performed twice for each employee on 30-2A
common southern red oak boards. The authors conclude
d
that operators need to be able to (a) "see and recognize
defects", (b) "have the mental aptitude to properly locate
the cuts", (c) "possess the physical strength to position the
board manually" , (d) "resist boredom and maintain an aler
t
mental attitude", and (e) "be able to remember defect
199
U. Buehlmann, RE. Thomas / Robotics and Computer Integrated Manufacturing 18 (2002) 197-203
West Virginia. Roughly 200 boards were collected and 158
were found to be usable for this study. The remaining 42
b
oards were discarded because they were either below
grade or had excessive crook. Crook is when a board has a
substantial arc from end to end across the grain. Thus, i
f
you would lay a board flat and put one side against a wall,
the middle of the board would not be touching the wall. The
b
oards were digitized according to Anderson et al. [12].
Fig. 4 displays an example of a digitized board. The USDA
Forest Service's computerized ultimate grading and
remanufacturing system (UGRS) was employed to grade
the selected boards into appropriate quality classes [13].
The material used in this study consisted of 8.3% FIF, 6.8%
Selects, 51.6% 1 Common, 25.9% 2A Common, and 7.4%
3A Common. Appendix A shows the details of the lumbe
r
sample used.
Cutting bills, as shown by Buehlmann et al. [14], have
a
significant influence on yield. A cutting bill is the list o
f
p
ieces that need to be produced during a given production
run. To minimize the influence of cutting bill composition
on this study, we used a cutting bill that is considered easy
to complete [15]. Part quantities for this cutting bill were set
such that all pieces could be obtained from the 158 boards
available. Appendix B gives the details of the cutting bill
used. The prioritization algorithm we used was length-
square times width (L 2W) with the maximum part value
set at 1000. Prioritization algorithms are necessary to allow
the software to decide which parts to prefer over others in
situations where different part-choices exist. The par
t
values assigned help the prioritization algorithm determine
which part is more preferable than others for a given
situation. For a more complete discussion of par
t
prioritization, see [15].
2.2. Rough mill
The first processing step is to rip the boards into narrow
strips. An arbor, which is a steel rod holding the saw blades,
is used to rip the boards into strips. The part widths are the
distances between the saw blades. GRADS [16], the gang-
ripsaw arbor design program, was used to determine the
optimal arbor width spacing arrangement. The ripping
process optimizes the placement of the board with respec
t
to the saws such that the highest yielding strip combination
is obtained.
The strips resulting from gang-ripping were then
presented to the human operator. Since the relevance of ou
r
findings was directly connected with the abilities and
location on one side while marking on the other side". This
study highlights the difficulty of the tasks performed by
operators. Although all persons tested were experienced,
motivated individuals, the average mean composite score
(i.e. the combination of all individual parameters tested) was
68%, the minimum was 59%, the maximum 74%,
respectively. Highest mean scores were achieved for
location (75 percent), followed by number of defects (71
percent) with defect types being lowest (65%).
Although these tests demonstrated the magnitude of human
error in marking, they did not assess yield effects resulting
from operator errors. To fill this gap, the objective of this
study was to evaluate the frequency of operator errors (Type
I, Type II, and partial Type I errors) and the resulting yield
losses in a state-of-the-art rip-first rough mill. By
quantifying the potential yield increases, one can investigate
the economics of automated scanning systems for reducing
or eliminating yield losses due to operator error. Such
vision-
b
ased lumber defect detection systems are becoming
commercially available [10]. They recognize defective areas
of boards with a high accuracy and thus allow computer-
b
ased yield optimization and saw control systems to
efficiently use the clear areas of a board.
2. Methodology
Due to careful planning of an earlier study involving the
validation of a computer-based rough mill simulation model
called ROMI-RIP, previously ",ollected data could be used
[7,11]. These two publications also contain more details
about the rough mill and simulation setup used for this
study. This study used randomly sampled kiln-dried red oa
k
hardwood lumber from a sawmill in southeastern Wes
t
Virginia. After digitization of the boards, they were cut to
strips in a state-of-the-art rough mill and the strip solutions
were recorded. Next, an experienced employee marked the
strips. The results were compared to the digitized location o
f
the individual defects to establish the number of inaccurate
marks. Both datasets were then employed to perform
simulation to assess yield losses associated with operato
r
error.
2.1. Materials
Rough sawn, kiln-dried lumber was randomly sampled
fro the lumber grading conveyor of a large sawmill in m
200
U. Buehlmann, R.E. Thomas / Robotics and Computer Integrated Manufacturing 18 (2002) 197-203
Table 1
Results of the operator accuracy tests
simulated exactly as it occurred in the rough mill, the
average of four repetitions from the simulation was used.
The feeding sequence of strips influences yield, since the
p
lacement of individual parts within the available areas
depends on the sequence in which the strips are processed.
However, this effect, although discernible, is small and
does not alter the conclusions of this study.
4. Results
A total of 1303 defects were present in the 158 boards
used in the study (Table 1). Out of these boards, the ripsaw
p
roduced 404 strips: 59, 1.75in wide; 259,2.00 in wide; an
d
86 3.50 in wide. The operator made 1331 marks on the
strips or 3.45 marks per strip on average. No marks were
necessary for defects located < 1 in from either end of the
board since the system was set to make a I-in trim cut on
b
oth ends. The number of marks and the number of defects
are loosely correlated. Often several defects in close
p
roximity are marked together between two beginning an
d
ending marks.
Twenty-six of all marks set were Type II errors, i.e. the
operator marked areas where no defects were present. The
operator also made 578 Type I marking errors, i.e. no
t
marking a defect when there was one. Furthermore, 437
p
artial Type I errors, meaning that marks were place
d
inside the defective area, were made by the operator. Table
1 shows the test results obtained.
As these results show, 78.2% of the decisions made by
the human operator deviated from the optimum decision.
Type I error, not detecting a defect when one is present,
was the most common error constituting 43.4% of the total
number of errors committed (Table 1). Partial Ty
p
e I
errors, marking a defect inside the
,
motivation of the person doing the marking, a person
with more than 5 years of experience was employed. The
worker's ability to detect defects and react properly was
tested using the Wais-R digit symbol recognition test use
d
earlier in the Huber study [9]. For this test, a score of 40
p
oints was achieved by the operator, which indicated that
the worker's performance is comparable to the performance
of the average rough mill operator. The tests were
p
erformed at the worker's regular workstation with no
change to the workplace that could indicate testing was
b
eing done. The worker marked all areas of the strips.
containing defects with a scanner readable crayon. The
strips were directly fed to a Barr-Mullin Turbo Wondersaw
crosscut saw, with the sawblade removed, allowing for the
measurement and recording of the operator's marks. After
this step, there were two sets of strip-data for the
simulation: (a) the digitized board data, and (b) the defect-
marking solution as done by the human operator. Fig. 5
displays the marking solutions by the operator shown in
gray, with defects digitized before ripping shown in black.
3. Computer based rough mill simulation
The USDA Forest Service's ROugh Mill RIP (ROMI
RIP) first simulator, Version 2.0, was employed [11,17].An
ad itional computer program was written to assess the
accuracy of the operator. This program overlaid the
locations of the operator's grade marks with the actual
defect locations on the board. From this comparison several
important evaluation factors were determined including: (1)
how close the marks were to the defect(s), (2) defect sizes
and types not marked, (3) defects split, (i.e. defects not
entirely in a marked area), (4) marked areas with no
defects, and (5) defects properly marked and entirely
included in a marked area. This information described the
accuracy of the human operator.
A second test was designed to estimate yield losses due
to operator error. Here the boards were processed on the
ROMI RIP simulator using the marked area information
from the operator, cutting clear parts and parts tha
t
contained defects. Using a special software tool, these
computer generated parts were then checked for remaining
defects due to operator error. Parts containing defects were
rejected. By tracking the rejected parts, an accurate
assessment of the operator's performance in respect to yield
could be performed. Since the strip feeding sequence to the
chop saw could not be
d
Fig. 5. Sample board showing the actual defects and the operator solutions.
Total boards used
Total strips marked
Total no. of defects registered
Total marks placed
Type of error
Type II
Type I
Partial Type I
Total
Average error/strip
158
386
1303
1331
Missed (%)
20
43.4
32.8
Missed (no.)
26
578
437
1041
2.7
78.2
78.2
201
U. Buehlmann, RE. 1110mas / Robotics and Computer Integrated Manufacturing 18 (2002) 197-203
Table 2
Rejected parts due to human marking error
Input lumber Output parts Yield including Reject parts Rejects as percentage of Yield without
area area rejects area rejects
Run (m2) (m2) (%) (m2) Input (%) Output (%) (%)
A 88.0 55.6 63.2 14.0 15.9 25.1 47.3
B 87.9 55.9 63.6 14.2 16.1 25.3 47.5
C 85.3 54.4 63.8 13.5 15.8 24.8 48.0
D 87.1 55.2 63.4 14.4 16.6 26.2 46.8
Average 87.1 55.3 63.5 14.0 16.1 25.4 47.4
std. dev. 1.3 0.7 0.3 0.4 0.3 0.6 0.5
No. of boards No. of parts Rejected parts Rejected parts
Run (No.) (No.) (No.) (%)
A 155 1485 330 22.2
B 156 1494 323 21.6
C 149 1441 324 22.5
D 154 1463 316 21.6
Average 153.5 1470.8 323.3 22.0
std. dev. 3.1 23.7 5.7 0.4
defective area, accounted for 32.8% of the errors made and
Type II errors, identifying a defect when one is not present,
only constituted 2.0% of the total number of errors.
Four repetitions of the simulation of the strip cuts were
p
erformed to detect yield rates associated with human
operator error (Table 2). On average, 25.4%, or more than a
quarter of the parts produced (output), would be rejecte
d
b
ecause they contained a defect or a partial defect resulting
in a decrease of yield from 63.5% to 47.4% yield.
wood. If a mark was made one-sixteenth of an inch
(1.6mm) inside the recorded defect area, the verificatio
n
p
rogram registered that as an operator error. Therefore, the
error rate indicated in Table 1, although true, is
exceptionally rigorous. Since most furniture parts are cu
t
slightly overlength in the rough mill, some of these "errors"
would not be rejected.
Due to the traditionally high reject rates in furniture
p
lants, rough mill operators should be taught to mar
k
defects away from the a
p
parent end of the defects in the
clear wood to decrease the probability of part rejections.
This will help decrease rejection rates due to partial Type I
errors. It is hypothesized that the loss of 1-5 mm of clea
r
wood for each defect will not significantly reduce yield due
to the optimization schemes. Conversely, Type I errors will
require more effort, such as increased operator training,
b
etter lighting, and easier to understand defec
t
specifications, to be eliminated.
This phenomenon is also responsible for the significantl
y
lower part reject rate compared to the much higher operato
r
error rate. Although the operator error was 78.2%,
a
significantly fewer number of parts, 22% (25.4% on an are
a
basis), were rejected. The excess clear area relative to par
t
length required eliminates most, if not all, rejects fro
m
p
artial Type I operator errors. Additionally, Type II errors
(2.0%), although they lower yield, do not lead to rejecte
d
parts.
Type I errors (43.4%) are the major contributors o
f
rejected parts. The negative impact from these errors is
lessened by two factors: (a) overlooked defects tend to be
small, and (b) small defects are often clustered close
together. Therefore, many rejected parts contained more
than one Type I error. Also, some of these overlooke
d
defects may have been adjacent to a defect that the operato
r
marked correctly. Due to the excess clearance
5. Discussion
Marking lumber in a rough mill is a truly challenging
occupation. The average operator stands 8 hours a day in a
noisy, often poorly lighted workplace and is required to
mark hundreds of strips indicating defective areas. Also,
what constitutes a defect is not clearly defined and a wide
range of definitions exists within the industry an
d
sometimes even within the same operation [18]. In addition,
wood defects can have a wide variety of shapes and colors,
often making it challenging to: (a) recognize the defect, an
d
(b) determine the border between defect and clear wood.
These factors, when combined, lead to a high error rate tha
t
costs the industry millions of dollars every year. Kline et al.
[19] estimated that for an average rough mill, a 1 % yield
increase results in estimated savings of $150,000 to
$300,000 in lumber and operation costs per year.
This study's measurement quality standards were ver
y
demanding. The verification programs that compared a
mark and the beginning/end of a defect did not allow any
deviation from the optimum position, except if the mark was
made outside of the defective area in the clear
