You are on page 1of 20

IIE Transactions (2012) 44, 136154

Copyright 
C IIE

ISSN: 0740-817X print / 1545-8830 online


DOI: 10.1080/0740817X.2011.593609

A waste relationship model and center point tracking metric


for lean manufacturing systems
SAINATH GOPINATH and THEODOR I. FREIHEIT
University of Calgary, Mechanical & Manufacturing Engineering, 2500 University Dr. NW, Calgary, Alberta, T2N 1N4, Canada
E-mail: tfreihei@ucalgary.ca

Received December 2009 and accepted May 2011

Lean manufacturing is about eliminating waste, which requires the creation of waste metrics that are tracked in order to create
the conditions for its elimination. In this article, metrics used to monitor the seven traditional non-value adding wastes types of
overproduction, defects, transportation, waiting, inventory, motion, and processing are explored and a center point metric pair is
proposed that can give systematic insight into system waste performance and trade-offs. For example, lower work-in-process levels
(inventory waste) may require more replenishment (transportation waste) in order to maintain production. A waste relationship model
is proposed that can be used to derive the relationship between different wastes in a Pareto-optimal waste-dependent lean system.
The trade-off relationships are statistically verified using simulation experiments across different system configurations, complexities,
and planning scenarios.
Keywords: Lean manufacturing, waste relationships, performance metrics, trade-offs, decision making

1. Introduction Womack et al. (1991), Hines and Rich (1997), Russell and
Taylor (1999), Canel et al. (2000), Conner (2001), Svensson
Lean manufacturing emphasizes value creation by elimi- (2001), and Rawabdeh (2005) consider waste as an expen-
nating waste. Waste consists of non-value-adding activities diture of resources for any means other than the creation
that contribute to the product cost and for which the cus- of value for a customer and thus a target for elimination.
tomer is unwilling to pay. Eliminating waste can reduce Historically, little attention has been given to Non-Value-
product costs and improve quality, but it is not possible Adding (NVA) activities such as storage and transporta-
to completely eliminate waste even in an efficient system tion (Rawabdeh, 2005). The result is that only minimal
whose operations are waste dependent (that is, has waste differences have been realized in the reduction of overall
as a part of its functionality). It is therefore necessary to lead time, the improvement of quality, and the reduction
understand the waste relationships in order to minimize of cost. For instance, Conner (2001) reported that when
system waste to the lowest possible level. lead time was examined, value-adding activities only ac-
In lean manufacturing, seven types of waste have been counted for 5% of the lead time. Similarly, Suzaki (1987)
defined (Womack et al., 1991). claimed that only 5% of an operators time adds value and
the rest adds cost to the product. Todd (2000) concluded
1. Overproduction (production ahead of demand). that waste should be eliminated to reduce lead time and
2. Defects (any product/service that the customer is un- allow a manufacturer to respond quickly to customer re-
willing to accept). quirements. Samaddar and Heiko (1993, p. 19) summarized
3. Transportation (moving products when it is not actually that
required to perform a processing step).
4. Waiting (any resources/materials staying idle).
. . . the elimination of waste can be viewed as a common
5. Inventory (materials not being completely transformed). issue in any production/operations system. One should
6. Motion (resources moving more than is required to systematically identify and continuously work to eliminate
transform the material). such waste in order to achieve effectiveness and efficiency.
7. Processing (unnecessary or processing over the mini-
mum necessary for material transformation). The majority of the research literature focuses on the
complete elimination of waste in order to improve produc-
tivity and quality, respond quickly to customer require-

Corresponding author ments, and reduce manufacturing cost.


0740-817X 
C 2012 IIE
Waste relationship model 137

It is not possible to completely eliminate all of the seven rior center point metric that is highly correlated with other
waste types even in an efficient system whose operations system waste types. They verified its efficacy using simple
are waste dependent. A system that has waste as a part of correlation analysis; however, this approach only provides
its functionality is referred to as a waste-dependent system the direction and degree of linear relationships.
in this article. In these systems, complete waste elimination In this article, the center point metric concept is more
is not possible. Instead, all wastes can only be reduced to rigorously examined using regression analysis, which pro-
the minimum level that exhibits Pareto-optimality. Pareto- vides additional insight into the system sensitivity. The first
optimality (Feldman, 1980) is a term from economics that metric of the pair is a metric that receives strong signals
is used to describe a set of solutions for a multiple objective from the system or magnifies the effect of changes or ab-
problem that exhibits the property that no single objective normalities in the system and is termed the Detection Cen-
criterion can be improved without a trade-off making some ter point Metric (DCM). The DCM is highly correlated to
other criterion worse (Feldman, 1980; Petrie et al., 1995). other waste types and it can be used to monitor the system
To improve upon the Pareto-optimal waste-dependent performance. The second metric of the pair sends strong
system, work tasks within the production system must signals to the system and is termed the Pivot Center point
be redesigned to achieve functionality without the waste Metric (PCM). Small changes in the PCM are magnified
(waste independent), which may require significant capital in other system waste types, and it can be used for system
investment. The fullest application of lean principles leads waste optimization or system design and decision mak-
naturally to transformation into Pareto-optimal waste- ing. The best center point metric pair will be determined
dependent lean systems if we consider the use of capi- through statistical evaluation of system waste responses.
tal as wasteful if the system can be improved (all wastes This article proposes a waste relationship model that
reduced; i.e., there is no Pareto trade-off between waste can be used for decision-making about trade-offs with the
types) without its expenditure. Until such time that capital objective to reduce all waste types to the minimum pos-
is available to redesign the system, it will operate in this sible level in a waste-dependent efficient system without
Pareto-optimal state. The proposed Pareto-optimal waste jeopardizing its intended functionality. Moreover, this ar-
relationship model is intended to aid in effective opera- ticle identifies, develops, and integrates a set of metrics;
tional decisions to cut costs and improve efficiency without determines the waste relationship; and statistically verifies
the requirement of immediate capital. the proposed waste relationship and the center point met-
The relationship between the waste types must be first ric across different production planning scenarios and dif-
understood in order to achieve a waste-dependent efficient ferent manufacturing system complexities. The following
system. Wastes must therefore be measured quantitatively sections of this article elaborate further on the research
in order to derive their relationships. Feld (2001) defines methodology, quantification of the wastes using metrics,
a manufacturing metric as a standard measure that de- derivation of the waste relationships, and statistical verifi-
scribes a performance criterion for a manufacturing process cation of the waste relationship model and the center point
so that everyone in the organization is working towards the metric using simulation experiments.
same goal. An attempt has been made to identify and inte-
grate a set of metrics that can quantitatively measure differ-
ent waste types. However, it was found that not all metrics in 2. Methodology
the existing literature fulfill the requirements of this article
in that they be simple and shop floor feasible. For example, A three-step methodology was followed in this research.
the build-to-schedule (Khadem et al., 2006) metric requires First, the literature on lean manufacturing was reviewed to
computer simulations and is not simple. Therefore, a set of define waste types and explore potential metrics. Second,
metrics is proposed to provide the information necessary the logical relationship between waste types was mapped
to understand the waste relationships. using concept mapping and a relationship model was de-
veloped. Finally, the model was statistically tested using
discrete-event simulation to determine the trends in the
1.1. The center point metric
magnitude and direction between the waste relationships.
It is desirable to have a simple, feasible metric that can pro- Note that an absolute measure of the relationship between
vide, at least to some degree, a measure of all waste types the waste types is not the intent of this analysis.
in the manufacturing system. An objective of this article is The review of current research in lean manufacturing
to develop a metric or set of metrics to understand the per- identified a few shortcomings in existing research such as
formance of the system in the least possible time and cost. the lack of simple, shop floorfeasible dedicated metrics
A single metric, highly correlated to other system metrics, to quantify waste. Next, concept mapping was selected to
can reveal critical information about the whole systems assimilate the relationships between waste types, which is
performance. Therefore, a center point metric pair is pro- a powerful technique for the graphical representation of
posed that can give systematic insight into the system and knowledge. Moreover, it is a technique that can aid in the
can be used for decision making. Gopinath and Freiheit understanding of relationship concepts (in this case, an
(2009) have proposed customer waiting waste as a supe- identified waste) with other concepts (other waste types).
138 Gopinath and Freiheit
Table 1. Waste definitions useful to decision-making processes in manufacturing in-
dustries, metrics must be feasible for collection on a real,
Waste Definition
dynamic manufacturing shop floor. In addition, the num-
Defects Any product that is unacceptable to the ber of metrics should be kept to a minimum in order to keep
customer. Handling and transformation the data collection costs as low as possible and to minimize
defects are considered the time necessary to understand what is happening in the
Overproduction Production ahead of demand, which is system.
captured by the finished inventory Many manufacturing performance metrics can be iden-
Inventory Raw materials and work-in-process not tified in the research literature, but few are both feasible
being processed and simple. A table of proposed waste metrics identified
Motion Operators movement between workstations
in the literature is summarized and reviewed in Gopinath
Processing Processing more than the minimum required
for material transformation
and Freiheit (2009). Performance metrics have been de-
Transportation Transporters movement between inventories veloped in different contexts such as lean manufacturing,
Waiting Any resource staying idle during work hours total productive maintenance, and theory of constraints.
Unfortunately, many of these metrics cannot be used di-
rectly to measure shop floor performance because they
This process involves creating a global map that shows are too general (provide global measures), require overly
the main topics and their relationships, and more detailed complicated calculations (e.g., dock-to-dock as defined by
map(s) showing specific details of a particular portion of Khadem et al. (2006)), or do not capture waste as defined
the map. by lean manufacturing, even though, as in machine re-

R
Simulation with Arena was used to test the waste re- liability, they are good performance indicators. In some
lationship model. Design of Experiment (DoE) techniques cases, metrics are better suited for computer simulation
were used to run the simulation experiments, and linear models than direct shop floor measurements; e.g., build-to-
regression and Analysis of Variance (ANOVA) statistical schedule (Khadem et al., 2006). Additionally, some shop
analysis were performed on the simulation data in order floor performance metrics provide superfluous or redun-
to understand the systems sensitivity to the center point dant information and make things look unnecessarily com-

R
metric using the statistics toolbox of MATLAB . plex.
Table 1 summarizes how waste types are defined in
this article, and Table 2 summarizes the proposed waste
3. Metrics metrics. The defects (A) waste metric should measure
anything that is unacceptable for the customer. Therefore,
Metrics for measuring waste in a manufacturing systems the metric proposed by Rother and Shook (1999) quantifies
should be easy to collect and simple to understand. To be the waste by providing the percentage of unacceptable

Table 2. Waste metrics table

Waste code Waste Metric Definition


 Si
A Defects P
Rother and Shook (1999)
all MC
1 T

B Overproduction T 0
F Idt Time-persistent measure of finished inventory (Kelton et al., 2007)
Tm
C Motion T
Percentage of time spent in motion
F Tt
D Transportation T
Percentage of time spent in transportation
Tw
E Waiting (customer) T
Percentage of time spent waiting
 1 N
F Waiting (material WIP) N i =1 WQi The time spent waiting (Kelton et al., 2007)
all WI P
all
G Waiting (machine) 1 MC TRi
nT
Percentage of time spent waiting
 Tw + Tm 
H Waiting (operator) 1 T
1% Operator saturation
T
I InventoryWarehouse 0 WHdt Time-persistent measure of raw material inventory (Kelton et al.,
2007)
 1
T
J InventoryWork-in-progress T 0 WI Pi dt Time-persistent measure of WIP inventory (Kelton et al., 2007)
all WI P
K Processing C p , C pk Chuan et al. (2001), NIST Handbook
Waste relationship model 139

production. Overproduction (B) is production ahead of not considered in this model as it is specific to particular
demand and it is captured by the finished inventory. The manufacturing processes. Rawabdeh (2005) also has simi-
finished inventory should be measured based not only on lar views on processing waste. Whether a waste type has a
the inventory content but also the duration of time that the direct or inverse relationship with another waste type can
parts stay in the inventory. Therefore, the time-persistent be obtained by multiplying the signs on the path between
metric suggested by Kelton et al. (2007) is used, where the them. For example, the relationship between overproduc-
inventory is time-weighted based on part content duration. tion (B) and WIP (J) is obtained by multiplying the 1
Similarly, work-in-process (J) and warehouse inventories +1 +1 +1 = 1, showing that it has an inverse
(I) are quantified using time-persistent measures. Motion relationship.
(C) waste occurs when an operator walks between work-
stations and is measured as a percentage of time spent in
motion by the operator. Transporting material between 5. Testing the relationship model using simulation
processes, which adds no value to the product, leads to
transportation (D) waste. This metric should measure The waste relationship model is tested using discrete-event
both transportation duration and frequency to provide a simulation in three similar serial manufacturing system
transportation percentage. Waiting material (F) waste is models of increasing complexity and different production
quantified by measuring the average time waiting in inven- planning time horizons. The results of the simulation runs
tory (Kelton et al., 2007). Waiting resource (G, H) waste were also used to identify and screen center point metric
is quantified by the time that a resource is idle, expressed candidates. Each complexity model was run at three lev-
as a percentage of NVA activity by the resource. Similarly, els of demand rate where no resource was deliberately set
waiting customer (E) waste is quantified by the percentage as a bottleneck. Rather, the resources become system bot-
of time that customers wait for product availability. Pro- tlenecks automatically by the randomness induced by the
cessing (K) waste occurs whenever the material is processed simulation. The demand rate was varied by 10% with re-
inappropriately during transformation. Process capability spect to the system throughput time. The system constants
indices like C p and C pk (Chuan et al., 2001) are appropriate are summarized in Table 3. The system responses are over-
metrics to use to indicate the statistical potential of the production (finished inventory (B)), inventory (WIP (J),
process toward exceeding customer requirements. warehouse (I)), waiting (material (F), machine (G), cus-
tomer (E), and operator (H)), motion (C), transportation
(D), and defects (A).
4. The relationships between waste types

The concept map illustrated in Fig. 1 was developed to un- Table 3. Simulation model constants
derstand the logical relationships between the waste types.
The starting point for this map was Rawabdeh (2005), who Variables Values
examined similar waste relationships but missed a few criti- Batch size B1 30 Units
cal relationship scenarios such as transportation and inven- Batch size B2 4 Units
tory (Work-In-Process (WIP)). Building on his model, an Lead time LT 90 Min
understanding of the waste relationships was developed, Distance D1 1900 Ft
giving the trade-offs between different waste types in an Velocity V1 265 Ft/min
efficient waste-dependent system. Scrap S1 0.4 %
The concept map shows how different waste types are Scrap S2 0.2 %
conceptually linked together in a system. Darker shaded Scrap S3 0.0 %
nodes are the wastes and lighter shaded nodes are con- Scrap S4 0.0 %
Scrap S5 0.1 %
necting concepts. An example of the interpretation of the
Scrap S6 0.0 %
connections is that as a transportation resource replen- Cycle time CT1 4 Min
ishes inventory, larger batch sizes will lead to higher mate- Cycle time CT2 4.3 Min
rial storage resulting in high inventory waste and lower Cycle time CT3 4 Min
replenishment frequencies (lower transportation waste). Cycle time CT4 3.9 Min
Therefore, an inverse relationship can be derived between Cycle time CT5 4.2 Min
transportation (D) and inventory waste (J). Reorder point ROP1 30 Units
From the concept map, a trade-off model of relation- Reorder point ROP2 15 Units
ships was developed and is summarized in Fig. 2. These Availability A1 90 %
relationships can be used for multi-level decision making Availability A2 95 %
by selecting the appropriate metrics from Table 2 and deter- Availability A3 97.5 %
Availability A4 90 %
mining the relative impact between the waste types result-
Availability A5 95 %
ing from waste reduction program. Processing waste (K) is
140
Fig. 1. A concept map of logical waste relationships. (Color figure available online.)
Waste relationship model 141

Fig. 2. The waste relationship model.

The intermediate complexity model, illustrated in Fig. 3, goods supplier replenishes the warehouse inventory with
has one raw goods warehouse (WH), a raw materials trans- large batch sizes. Production control is a pull system with
porter (TR), five serial work stations (WS) with five inter- the transporter and other system resources triggered by
spersed WIP buffers (W), a finished goods inventory (FI), the finished goods inventory level. The transporter moves
two machine operators (O), and a customer (CU). The raw small batches from the warehouse to the WIP buffer. Then,

Fig. 3. Schematic of intermediate complexity manufacturing system model.


142 Gopinath and Freiheit

the workstations pick the parts from their respective WIP ship verification models were chosen to be in between the
buffers and process them. Operator 1 loads, unloads, and Pareto-optimal bounds obtained from this optimization.
transfers parts between workstations 1, 2, and 3 and oper-
ator 2 controls workstations 4 and 5. Finally, the customer
consumes from the finished goods inventory. 5.1. Preliminary data visualization
The simple model has the same features as the above- The raw simulation waste measures from the various com-
described model except for the number of resources. It has plexity and demand LTSS runs were aggregated and then
two machines and one operator to load, unload, and trans- plotted pair-wise in a matrix scatterplot to visualize the re-
fer between them. In contrast, the complex model has eight lationships between the wastes; see Fig. 4. Note that there
machines and three operators. The increase in the resources is considerable dispersion in the data and in many cases
complicates the material flow and increases the overall sys- multiple distinct lines can be seen. In general, wastes with
tem complexity. All models are considered to be efficient highly consistent data spread, defined as a tendency to be
because the resources were saturated to the maximum pos- coherent yet have a distributed data frequency (short bars
sible level, line balanced, made highly reliable, use pull in the histogram), have higher sensitivity to other waste
signals to control the excess inventory build-up, and the types. For example, consider waiting customer and waiting
system was operated close to the takt time. At the same material wastes. The histogram for waiting customer waste
time, stochasticity was introduced into the system to repli- (E) is more consistently distributed (more toward a uni-
cate the real, dynamic environment, such as resource failure form distribution) than waiting material (F), which results
patterns following an exponential distribution, demand fol- in a slope tending to the extremes of either zero or large
lowing a Poisson distribution, and process times following when predicting one waste from the other. Motion (C),
a triangular distribution. While these models are appropri- transportation (D), waiting machine (G), waiting operator
ate to explore the relationship between the waste types, this (H), and warehouse (I) wastes have similar histograms to
simulation model is limited to efficient waste-dependent waiting customer (E) waste and tend to show coherent dis-
lean systems. tributions and are expected to have higher sensitivity levels
Each complexity simulation run was performed across to the other wastes. Overproduction (B) waste is incoher-
three different production planning time horizons, namely, ent and has a concentrated interval frequency that results
Long-Term Steady State (LTSS), Short-Term (ST), and in lower system sensitivity. Defects (A) and WIP (J) waste
Production Ramp-up (PR). The LTSS (truncated replica- show random scatter, which is also expected to show lower
tion steady state) was run for 249 600 min of production system sensitivity. The diverging multi-streams that lead to
time (representing a year) for 30 replications using Arena. the incoherence in these plots is due to noise factors such as
System performance statistics were collected after a 4800- demand and complexity, which can be filtered out to give
min warm-up to avoid the initialization biases. The ST sim- better insight into the system.
ulation was run for 4800 min after a 4800-min warm-up for The raw waste data were filtered of the obscuring effects
30 replications, whereas the PR simulation used terminat- of the known noise factors, namely, demand and complex-
ing replications and was run for a total time of 4800 min ity, by regressing each waste against another waste type but
without clearing the initial statistics. The waste responses blocking for the noise, which is possible in a controlled ex-
from the simulation model were normalized in order to periment. Blocking is a statistical technique used to remove
address slope magnitude variation because waste response the obscuring effect of factors and their sources of variabil-
magnitudes can vary significantly. ity (Montgomery, 2008). Equation (1) is the pair-wise waste
A test for Pareto optimality was conducted on the sim- relationship model used in the regression analysis:
ple model to ensure that the selected system parameters
showed appropriate trade-off relationships. The waste types Wi = 0 + 11 C1 + 12 C2 + 2 D + j Wj + ,
were minimized by integrating the Arena simulation soft- i, j = , . . . , 10, i = j. (1)
ware with the optimization toolbox of MATLAB. Ten sys- The system complexity has two regression coefficients for
tem parameters, consisting of the independent variables blocking. Demand, a continuous variable, was scaled in
operator transfer time, machine cycle times, finished inven- the regression equation and took the values 0, 0.5, and 1.0.
tory re-order point, supplier lead time, operator cycle time, Unlike demand, complexity is a categorical variable and a
warehouse re-order point, internal batch size, warehouse DoE-based coding convention was adapted for complexity
order batch size, and transporter scrap rate, were numer- blocking, where (C1 , C2 ) are respectively (1, 0), (0, 1), and
ically optimized using the MATLAB fmincon function to (1, 1) for low, intermediate and high complexity (Mont-
minimize the objective function of the weighted sum of gomery, 2008). This regression allows the response waste
the simulation response waste metrics. The variation of the from every simulation sample k to be adjusted to account
simulation model was controlled by using random num- for the effect of complexity and demand, giving a noise
ber seeds, steady-state truncated replication simulation filtering equation:
strategy, and averaging the waste function value from five
replications. The system parameters for the waste relation- i k = Wi k 0 11 C1 12 C2 2 D.
W (2)
Waste relationship model 143

Fig. 4. Pair-wise waste comparisonpreliminary data visualization.

In other words, the effect of the predicted variation of the waste variation. Table 4 summarizes the effect of noise
the complexity and demand resulting from the regression factors for the LTSS simulation waste measures. It shows
model is removed from the simulation waste response data, that defects (A), motion (C), waiting machine (G), waiting
and the resultant pair-wise relationships can be examined operator (H), and WIP (J) waste variation is influenced by
with minimal influence of noise; see Fig. 5. As can be seen, complexity. This follows because these factors are scaled
the data are much more coherent and the histograms are directly by the amount of processing required and this in-
more symmetric. Consider again the example of waiting creases with system complexity. Example patterns in the
customer (E) and waiting material (F), the diverging multi- magnitude of the complexity coefficients are illustrated in
streams of waiting customer waste have for the most part Fig. 6(a). Low complexity tends to decrease the overpro-
disappeared, and the data are generally more coherent for duction (B) waste response, whereas intermediate complex-
both waiting customer and waiting material wastes. The ity is more neutral, and high complexity tends to increase
filtered data make response prediction much clearer; for ex-
ample, when customer waiting waste is predicted by waiting
material waste, the slope is generally more distinct, and vice Table 4. Contribution of noise (in %) to waste variation
versa. Complexity and demand parameters were included
in all subsequent regression analyses of the data. Waste (see Table 2 for code)
Response A B C D E F G H I J
5.2. The influence of complexity and demand on waste Complexity 54 17 65 19 21 13 64 73 14 47
Demand 4 64 3 8 0 0 3 2 10 9
An ANOVA test was conducted on the complexity and
Other 43 19 32 73 79 87 33 25 76 44
demand noise factors to determine their contribution to
144 Gopinath and Freiheit

Fig. 5. Noise-filtered pair-wise waste comparisonpreliminary data visualization.

overproduction (B) response waste. The WIP (J) and mo- and transportation (D) were predictor wastes, demand in-
tion (C) response waste exhibits an opposite behavior, with creased over-production (B) and WIP (J) response waste.
low complexity increasing WIP and motion waste response This is because these predictor wastes are tightly coupled
and high complexity decreasing them. The waste response with higher production volumes. The demand coefficients
shows a trade-off between internal (WIP) (J) and finished for the other response wastes were of similar magnitude for
goods inventory (B) when going from low to high complex- each predictor waste.
ity. The pattern of motion (C) response waste follows di-
rectly from the simulation models, where for low complexity
there is one operator for two stations, for intermediate there 5.3. Pair-wise linear regression of waste relationships
are two operators for five stations complexity, and for high A pair-wise linear regression analysis was performed where
complexity there are three operators for eight stations. each waste response from the simulation experiments was
Table 4 also shows that, as expected, overproduction used to predict every other waste response to understand
(B) is most influenced by demand because finished inven- their relationship. The regression analysis provides a slope
tory, its metric, acts as a cushion for demand fluctu- whose direction determines whether the relationship is di-
ation. Likewise, other inventory waste measures such as rect or inverse, and a magnitude which determines the
warehouse (I) and WIP (J) are also influenced by demand. strength of the relationship. The regression also provides
The demand coefficients, Fig. 6(b), are much higher than an adjusted R2 , which, as a measure of the models fit, pro-
the complexity coefficients, indicating a strong coupling vides an estimate of how effective a given waste is in predict-
of demand to inventory waste. The demand coefficients ing changes in another waste. Equation (1) is the pair-wise
for the response wastes of both overproduction (B) and waste relationship model used for this regression analysis.
WIP (J) were both large and had mixed signs depend- Tables 5 and 6 summarize the slope and adjusted R2 val-
ing on the predictor waste. When defects (A), motion (C), ues derived for the time horizon of the LTSS. As can be
Waste relationship model 145

Fig. 6. Effect of system complexity and demand on waste relationship (LTSS).


146 Gopinath and Freiheit
Table 5. LTSS predictor waste regression coefficients (slope)

Predictor waste
Response waste A B C D E F G H I J

Defects (A) 1 0.69 0.86 0.90 0.83 1.08 0.91 0.86 0.86 0.81
Overproduction (B) 0.58 1 0.55 0.57 0.57 0.79 0.58 0.54 0.56 0.71
Motion (C) 1.08 0.82 1 1.05 0.97 1.21 1.06 1.00 1.00 0.95
Transportation (D) 1.02 0.78 0.95 1 0.92 1.13 1.01 0.95 0.95 0.90
Waiting customer (E) 1.10 0.90 1.02 1.08 1 1.26 1.09 1.02 1.03 0.99
Waiting material (F) 0.53 0.47 0.48 0.50 0.47 1 0.50 0.47 0.50 0.58
Waiting machine (G) 1.00 0.77 0.94 0.99 0.91 1.12 1 0.94 0.94 0.89
Waiting operator (H) 1.08 0.82 1.00 1.06 0.97 1.20 1.07 1 1.01 0.95
Warehouse inventory (I) 0.95 0.74 0.88 0.93 0.86 1.11 0.94 0.88 1 0.90
WIP inventory (J) 0.75 0.79 0.71 0.75 0.70 1.10 0.76 0.71 0.76 1

seen, the magnitude of the slopes and the fit of the rela- 5.4. Center point metric analysis
tionships between the wastes are not symmetric, whereas
Three broad criteria were established to test for the center
the directions are perfectly symmetric between the wastes.
point metrics. The first is the ability of the metric to
Consider waiting customer (E) and waiting material (F)
predict the system waste performance, as measured by
wastes. While their direction relationship is direct, posi-
the Pearson r correlation coefficient. The second is the
tive, and symmetric, their magnitude relationship is non-
overall sensitivity of the metric to other waste metrics, as
symmetric because the two wastes change in a dissimilar
measured by its regression coefficient slope, ignoring its
rate with respect to each other; i.e., one or more of the
direction. A third customized measure, referred to as a
wastes are non-linear. In this case, waiting customer waste
predictive slope, is the product of the adjusted R2 and the
changes at a rate 1.26 times that of the waiting material
slope and represents the sensitivity of the metric weighted
change, whereas the waiting material waste only changes
by its ability to predict the waste relationship.
by 0.47 times. In other words, waiting material waste is less
A four-step methodology was adopted to determine the
sensitive to change than waiting customer waste is.
center point metric pair. First, a correlation analysis was
The slope magnitudes and directions were found to be
conducted in order to identify the dependence relationship
uniformly consistent across the RU, ST, and LTSS produc-
between the wastes (Gopinath and Freiheit, 2009). Then, a
tion planning time horizons. Figure 7 shows an example of
statistical comparison was conducted between the correla-
the variation in predictor waste coefficients and adjusted
tion coefficients of all the system wastes to identify the top
R2 for customer waiting (E) and WIP (J) response wastes.
four center point metric candidates that have correlations
As can be seen, the variation in slopes between the time
significantly larger than the others. Second, each candidate
frames is generally small, less than 1020%, with only a
waste was compared using a two sample, one-tailed t-test
few predictor wastes such as defects (A) and material wait-
for statistical difference to determine the number of wastes
ing (F) having larger variation. These two predictor wastes
that have statistically higher mean waste sensitivity and pre-
are relatively rare events in the simulation model, as can be
dictive slope in order to identify the metrics that send or
seen by their poorer model fit at shorter time frames.
receive significantly higher signals. Third, the magnitude of

Table 6. LTSS waste relationship model fit (adjusted R2 )

Predictor waste
Response waste A B C D E F G H I J

Defects (A) 0.48 0.94 0.93 0.92 0.63 0.92 0.93 0.84 0.66
Over production (B) 0.85 0.86 0.86 0.88 0.84 0.86 0.86 0.85 0.89
Motion (C) 0.93 0.51 1.00 0.99 0.63 1.00 1.00 0.90 0.71
Transportation (D) 0.94 0.58 1.00 0.99 0.67 1.00 1.00 0.92 0.76
Waiting customer (E) 0.93 0.61 0.99 0.99 0.67 0.99 0.99 0.90 0.76
Waiting material (F) 0.62 0.44 0.62 0.61 0.64 0.61 0.62 0.61 0.68
Waiting machine (G) 0.94 0.59 1.00 1.00 0.99 0.68 1.00 0.92 0.76
Waiting operator (H) 0.93 0.54 1.00 1.00 0.99 0.65 1.00 0.91 0.73
Warehouse inventory (I) 0.86 0.54 0.91 0.91 0.91 0.65 0.91 0.91 0.76
WIP inventory (J) 0.73 0.70 0.78 0.78 0.79 0.76 0.78 0.78 0.79
Waste relationship model 147

Fig. 7. Effect of time frame on waste relationship.


148 Gopinath and Freiheit

the difference between the waste slopes was compared in or- The slopes and predictive slopes of the wastes were next
der to identify the wastes that send and receive the strongest compared to identify center point metric candidates that
signals to and from the system. Finally, the magnitude of had the strongest response to and from the system and
the predictive sensitivity difference was also compared. were most predictive of the system (Wuensch et al., 2002).
Pearson product-moment correlation coefficients, r , The t-statistic was calculated by first determining the stan-
were determined and compared for all waste pairs in or- dard error of each predictor waste slope (regression coeffi-
der to obtain the center point metric candidates. A Pearson cient) from the mean standard error (MSE) of the regres-
coefficient ranges from +1 to 0 to 1, indicating relation- sion model:
ships that are perfectly linear and direct to unrelated to 
perfectly linear and inverse, respectively. Correlation co- MSE j
s j = . (5)
efficients are not directly comparable because they ignore n
the inherent variance of the sample, so they must be trans- Then, determining the standard error of the difference
formed into a normally distributed variable using a Fisher between two predictor slopes:
z-transformation (Shen and Lu, 2006): 
 
1 1 + r  si j = s2i + s2 j , (6)

r = loge   . (3)
2 1r
and then the two sample t-test formula at a 95% confidence
The statistical comparison methodology suggested by level was used to determine whether a predictor waste slope
Wuensch et al. (2002) was used to compare the correla- was significantly greater than another candidates slope:
tion coefficients as well as the slope and predictive slope
magnitudes of different wastes. The transformed correla- i j
t= . (7)
tion coefficients, r  , can now be directly compared using a si j
statistical z-test:
For both the slope and predictive slope, the DCM candidate
ri r j wastes ranked as having the highest number of significant
z= . (4)
(1/(n i 3)) + (1/(n j 3)) better differences are (i) waiting customer (E) (100% and
97% of the slope and predictive slope, respectively, were
Table 7 summarizes the number of times a predictor higher in their paired comparison); (ii) waiting operator
waste had a significantly higher correlation coefficient to (H) (86% and 76%); (iii) motion (C) (76% and 69%); and
a response waste than the other predictor wastes. The z (iv) transportation (D) (60% and 58%). The PCM candi-
statistic is used as the t in a one-tailed t-test conducted at date wastes ranked as having the highest number of signifi-
a 95% confidence level for the LTSS production planning cant differences are (i) waiting material (F) (100% and 42%
scenario. The top five candidates with the highest count of for slope and predictive slope, respectively); (ii) defects (A)
correlations that are statistically significantly higher than (74% and 56%); (iii) waiting machine (G) (65% and 81%);
other wastes are motion (C), transportation (D), waiting and (iv) transportation (D) (56% and 72%). As can be seen,
customer (E), waiting machine (G), and waiting operator weighting the slope for its model fit reduces the attractive-
(H), which were significantly higher approximately 57% of ness of some candidate wastes. In fact, motion performed
the time. The number of waste types that tested significantly better than material waiting with a count of significantly
higher was found to be similar for all three production plan- better predictive slope differences of 43%.
ning scenario time frames. Taking into account the correlation, slope, and predictive
slope significant difference counts for the LTSS planning
Table 7. Correlation analysisdetection center point metric period, the top three DCM candidates are waiting customer
(LTSS) (E), waiting operator (H), and motion (C), whereas the top
Predictors
three PCM candidates are waiting machine (G), defects
(A), and transportation (D). While waiting material (F)
Responses A B C D E F G H I J Sum was a promising PCM candidate because of its high slopes,
as can be seen by inspecting Table 5, its model fit is poor
Defects (A) 0 7 0 0 0 1 0 0 0 2 10 and was eliminated from the finalists.
Overproduction (B) 1 0 0 0 0 0 0 0 0 0 1
The difference in the magnitude of the slope and predic-
Motion (C) 8 7 0 1 5 6 2 0 4 6 39
Transportation (D) 8 7 2 0 4 6 1 1 4 6 39 tive slope for the DCM was then calculated using Equation
Waiting customer (E) 8 8 2 2 0 8 2 2 2 6 40 (8), where Ri2 is set to one for the slope difference and set
Waiting material (F) 3 7 2 1 0 0 1 2 2 3 21 to the adjusted R2 for the predictive slope difference. In
j
Waiting machine (G) 8 7 2 0 4 6 0 1 4 6 38 this equation, k is the coefficient for predictive waste j for
Waiting operator (H) 8 7 3 0 5 6 1 0 5 6 43 j
candidate response k, whereas i is the coefficient for the
Warehouse inventory (I) 8 7 0 0 1 6 0 0 0 6 29
same predictive waste j for all other response wastes i . This
WIP inventory (J) 3 8 1 1 1 2 1 1 1 0 19
difference represents how much stronger a candidate waste
Waste relationship model 149

Fig. 8. DCM candidate slope and predictive slope comparison (LTSS).


150 Gopinath and Freiheit
Table 8. Contribution (in %) of factors with an inverse relationship to DCM slope variation

Predictor waste j
A B C D E F G H I J

Demand 11.2 5.7 5.0 4.8 12.7 9.4 10.2 10.5 4.7 33.0
Complexity 14.1 9.1 1.6 1.4 10.3 33.3 11.0 11.7 61.4 42.9
Response waste
A 0.0 0.1 0.0 0.0 0.1 0.0
B 0.6 0.0 0.4 0.4 0.0 0.0
C 0.0 0.0 0.0 0.0 0.0 0.0
D 0.0 0.0 0.0 0.0 0.0 0.0
E 3.4 4.3 3.9 3.9
F 0.7 1.3 1.5 1.5
G 6.4 8.0 7.7 7.7
H 12.8 16.1 15.6 15.6
I 5.8 6.0 5.6 5.6
J 0.0 0.0 0.0 0.0
Other 45.6 49.6 59.4 59.4 76.4 57.1 78.4 77.4 33.9 24.1

receives a signal from the predictive waste. Figure 8 illus- waste i . Figure 9 illustrates this difference for the three final-
trates this difference for the three finalist candidate wastes ist PCM candidates. While defects (A) sends the strongest
and indicates that for both slope and predictive slope, cus- signal to the response wastes, second only to material wait-
tomer waiting (E) receives the strongest signal from the ing (F), when weighing for the regression model fit, waiting
predictive wastes. machine (G) waste is a more favorable PCM.
 j  j
This analysis indicates that the best DCM is customer
DCM =  Rk2 k   Ri2 i , i, k = j. (8) waiting (E) and the best PCM is waiting machine (G). The
ST and RU planning scenarios generally yield the same
Similar to the DCM, the predictor waste j that sends the
conclusion as the LTSS. The PCM predictive slope analysis
strongest signal to response wastes can be calculated for
conducted for the ST planning scenario found transporta-
the PCM, Equation (9)
    tion (D) to be the best PCM, although it was only a little
PCM =  Rk2 ki   R2j ij  , j, k = i, (9) better than waiting machine (G) in terms of sending strong
signals to the system. However, this result may be due to
where in this equation, ki is the coefficient for candidate k unknown outliers or noise in the model. Therefore, it is sug-
predictive waste for response waste i and ij is the coeffi- gested that transportation waste should not be discounted
cient for all other predictive waste j for the same response as a PCM for short-term planning.

Table 9. Contribution (in %) of factors with a direct relationship to DCM slope variation

Predictor waste j

A B C D E F G H I J

Demand 13.9 15.2 12.6


Complexity 13.6 11.7 6.9 28.8 16.7
Response waste
A 1.0
B 1.7
C 7.1 3.7
D 20.5 10.9
E 3.3 3.6
F
G 5.0 1.8
H 8.3 3.5
I 0.0 17.3 0.0 0.0 12.6
J 21.8 29.2 21.2 21.2 10.4
Other 43.3 57.5 100.0 100.0 78.3 29.9 78.9 78.9 60.8 49.2
Waste relationship model 151

Fig. 9. PCM candidate slope and predictive slope comparison (LTSS).


152 Gopinath and Freiheit

6. Discussion A sense of how important a given variance factor is to


the overall strength of a signal can be derived by examining
This article assumes that the only noise factors in a sys- the contribution to variation, F (%), as weighted by a ratio
tem are demand and complexity. However, real, dynamic of the range of variation to its average value, Equation
manufacturing systems have many potential noise factors. (11):
The sources of variance outlined in Table 4 support that
only demand and complexity are noise factors in this arti- max( j ) min( j )
F = F. (11)
cles simulation models. Wastes such as defects (A), motion E( j )
(C), waiting machine (G), waiting operator (H), and WIP
(J) are significantly affected by complexity because it is as- For all sources of variance, this weighted variation around
sumed that increased system complexity can be modeled as the waiting customer DCM is only 15.6%, the lowest of
increases in the number of machines, interspersed buffers, the predictive wastes. Furthermore, Table 8 indicates that
scrap accumulation locations, and operators. However, this 12.7% of the DCM slope variation is due to the noise fac-
assumption may not hold in all cases. For instance, system tor demand, but this contribution to the slope magnitude
complexity may increase as a result of complex material is small when applying Equation (11), where its weighted
flow, resource movement, or batch size. Moreover, all waste contribution is only 2.0%.
types in this article are equally weighted, which may not re-
veal accumulation of the most critical costs in the system
that could be understood if the wastes were weighted by us- 7. Conclusions and recommendations
ing financial metrics. The PCM, which is only identified in
this article, has the potential for effective system waste cost The relationship between the waste types in waste-
minimization and prevention when used with the financial dependent Pareto-optimal lean manufacturing systems has
metrics. been examined in this article. Metrics that are simple and
The objective of this article was to find the waste types feasible to measure on the shop floor have been proposed.
that are most coupled to other wastes in the system. The A trade-off relationship between the waste types in efficient
waste types are not only sensitive to system noise such as waste-dependent systems was demonstrated using discrete-
complexity and demand but are also sensitive to variation event simulation and examined across different system con-
in other wastes. Two tests for robustness of the DCM were figurations, complexities, and planning scenarios. A center
conducted. First, the contribution of noise to waste re- point metric pair has been identified by examining the pair-
sponse variation was evaluated and, second, ANOVA tests wise slopes between the wastes. Customer waiting time (E)
were conducted on the predictive waste slopes. Equation has been identified as a DCM, which is an important per-
(10) is the regression model used to calculate the response formance measure for responsive production systems and
variable, j , for the ANOVA factors of noise (complexity can be used for system monitoring and abnormality diag-
and demand) and the other predictor wastes. The response nosis. Waiting machine time (G) has been identified as a
waste Wi was kept constant for the ANOVA tests. The re- PCM, which can be readily obtained from the common
sponses were separated out for the direct relationships (pos- shop floor performance measure machine saturation and
itive slopes) and the inverse relationships (negative slopes) can be used for system optimization, waste prevention, and
to prevent the misleading conclusion that results from ag- decision making. In addition, the center point metric pair
gregating all of the waste slopes. has been tested for robustness.
The DCM is expected to reflect and magnify system
Wi = 0 + j Wj + . (10) changes. Most systems are designed to have waiting cus-
tomer waste, which is a measure of the inverse of the sys-
Table 4 shows the effect of noise factors on waste re- tems service level, to be as low as possible because low
sponses. It can be observed that most of the variation in service levels are viewed to be detrimental to the business.
the DCM (waiting customer (E)) is due to factors other However, this may not always be true in waste-dependent
than noise. This is desirable because the DCM should be Pareto-optimal lean systems, as it depends on the nature,
as independent as possible from sources of noise. Second, location of waste cost accumulation, and efficiency priority
Tables 8 and 9 show the percentage contribution of differ- of a particular system. Therefore, it is suggested that cus-
ent system factors and noise to the waste slope variations. tomer waiting waste be used to monitor the system changes
Robust wastes, i.e., those that have lower sensitivity to vari- and machine waiting waste be used to aid in making effi-
ations in the system, have a high percentage in the other cient system changes.
field, as higher unexplained variation indicates that waste is In future work, the above assumptions and limitations
insensitive to other factors and most of its slope variation is should to be addressed. It is suggested that financial waste
due to inherent random variation. By this measure, waiting metrics be developed and their relationships verified in or-
customer waste is a robust metric, performing as well or der to determine the critical waste cost contributors of the
better as any other waste. system. Moreover, it is suggested that the potential of the
Waste relationship model 153

PCM be further explored by combining it with the finan- Svensson, G. (2001) Just-in-time: the reincarnation of past theory and
cial metrics for product waste cost minimization. Finally, practice. Management Decision, 39(10), 86679.
Todd, P. (2000) Lean manufacturing: building the lean ma-
the waste relationships should be verified in a real, dynamic
chine, available from: http://www.advancedmanufacturing.com/
manufacturing environment. leanmanufacturing/part1.htm, accessed February 2010.
Womack, J.P., Jones, D.T. and Roos, D. (1991) The Machine That Changed
the World, Lean Enterprise Institute, Cambridge, MA.
Wuensch, K.L., Jenkins, L.W. and Poteat, G.M. (2002) Misanthropy,
Acknowledgement idealism, and attitudes towards animals. Anthrozoos, 15, 139149.

The authors thank the support of the Canadian Auto21 Appendix


Network Centres of Excellence and the University of
Calgary. Notation
WS Workstation
PI Inventory from previous station
References TR Transportation resource
WIP Work-in-process inventory
Canel, C., Rosen, D. and Anderson, E.A. (2000) Just-in-time is not just I Inspection
for manufacturing: a service perspective. Industrial Management & FI Finished inventory
Data Systems, 100(2), 160. S Ship
Chuan, C.W., Chen, K.S. and Kuo, H.L. (2001) Decision making in
assessing process capability indices. International Journal of Infor-
WH Warehouse inventory
mation and Management Sciences, 12(3), 2942. O Operator
Conner, G.B. (2001) Lean Manufacturing for the Small Shop, Society of CU Customer
Manufacturing Engineers, Dearborn, MI. MC Machine
Feld, W.M. (2001) Lean Manufacturing: Tools, Techniques, and How to Util Utilization
Use Them, St. Lucie Press, Boca Raton, FL.
Feldman, A.M. (1980) Welfare Economics and Social Choice Theory,
Sat Saturation
Kluwer, Boston, MA. SU Supplier
Gopinath, S. and Freiheit, T. (2009) Development of a waste relation- F Transportation frequency
ship model for better decision making in lean manufacturing, Pro- n Number of machines, buffers, or workers
ceedings of the IIE Annual Conference and Expo, Miami, FL, May MCi i th Machining center or workstation
30June 3, 2009.
Hines, P. and Rich, N. (1997) The seven value stream mapping tools. In-
P Total units produced
ternational Journal of Operations & Production Management, 17(1), Si Scrap from the i th machining center
4664. T Total horizon time
Kelton, W.D., Sadowski, R.P. and Sturrock, D.T. (2007) Simulation with Tm Time spent in motion
Arena, fourth edition, McGraw Hill, New York, NY. TRi i th Machine operation time
Khadem, M., Ali, S.A. and Seifoddini, H. (2006) Efficacy of lean matrices
in evaluating the performance of manufacturing systems. Interna-
Tt Transportation time
tional Journal of Industrial Engineering, 15(2), 176184. Tw Idle or waiting time
Montgomery, D.C. (2008) Design and Analysis of Experiments, seventh i, j Wastes
edition, John Wiley & Sons, New York, NY. n i /n j Number of samples with respect to the waste i /j
National Institute of Standards and Technology. NIST/SEMATECH Si /j Standard error with respect to the waste i /j
handbook on engineering statistics, available from: http://www.itl.
nist.gov/div898/handbook/, accessed February 2010.
b Slope with respect to the waste
Petrie, C.J., Webster, T.A. and Cutkosky, M.R. (1995) Using Pareto op- ri Transformed correlation coefficient
timality to coordinate distributed agents. Artificial Intelligence for MSEi /j Mean squared error
Engineering Design, Analysis, and Manufacturing, 9(4), 269281. 0 Regression constant
Rawabdeh, I.A. (2005) A model for the assessment of waste in job j Wj Slope of the predictor waste Predictor waste
shop environments. International Journal of Operations & Produc-
tion Management, 25(8), 800822.
Wi Response waste in the simple linear regression
Rother, M. and Shook, J. (1999) Learning to See, Lean Enterprise Error term in the regression equation
Institute, Cambridge, MA. 1 Regression coefficient of complexity
Russell, R.S. and Taylor, B.W. (1999) Operations Management, second 2 Regression coefficient of demand
edition, Prentice-Hall, Englewood Cliffs, NJ.
Samaddar, S. and Heiko, L. (1993) Waste elimination: the common de-
nominator for improving operations. Industrial Management & Data
Systems, 93(8), 1319.
Shen, D. and Lu, Z. (2006) Computation of correlation coefficient and its Biographies

R
confidence interval in SAS , SUGI 31 Proceedings, San Francisco,
CA, March 2629. Sainath Gopinath received an M.Sc. degree in Mechanical Engineering
Suzaki, S. (1987) The New Manufacturing Challenge-Techniques for Con- from the University of Calgary, Canada, in 2010; an M.S. degree in
tinuous Improvement, The Free Press, New York, NY. Industrial Management from the Royal Institute of Technology, Sweden,
154 Gopinath and Freiheit
in 2008; and a B.Eng. degree in Production Engineering from Anna Calgary. His expertise is in both product and manufacturing sys-
University, India, in 2005. He has been a Manufacturing Engineer at Lean tem design and testing. He has a Ph.D. (2003) in Mechanical En-
Manufacturing Engineer at Magna International since 2010. His research gineering an MBA (1995) from the University of Michigan, and
interests include manufacturing systems, competitive manufacturing, and an MSE (1988) in Design Optimization from Purdue University.
statistical engineering. At the University of Calgary, he runs the senior capstone design
course. He has an active research program in the psychology of de-
Theodor Freiheit is an Associate Professor in the Department of sign innovation, design and analysis of manufacturing systems, and
Mechanical and Manufacturing Engineering at the University of micro-engineering.
Copyright of IIE Transactions is the property of Taylor & Francis Ltd and its content may not be copied or
emailed to multiple sites or posted to a listserv without the copyright holder's express written permission.
However, users may print, download, or email articles for individual use.

You might also like