application of learning curves in the aerospace industry handout
TRANSCRIPT
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
Learning Curves – Some Alternative Approaches Alan R Jones, BAE Systems
“O! This Learning, what a thing it is.”
William Shakespeare (c.1594, The Taming of The Shrew)
The material presented here is based on a case study presented in the following publication: Jones, A.R. ‘Case Study - Applying Learning Curves in Aircraft Production - Procedures and Experiences’ in Zandin, K (editor) Maynards
Industrial Engineering Handbook, 5th Edition, McGraw-Hill, New York, 2001
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
Learning Curves – An Alternative Approach
Constituent Elements of Production Learning– Segmentation Theory
Applications– Effect of Output Rate Constraint– “End of Line” Effect– Assessing Loss of Learning– Multi-Ganging of Operations (Parallel Learning)
Cumulative and Cumulative Average Data– Formulae– Examples
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
Constituent Elements of Production Learning
Background
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
34%
6%
11%22%
23%
4%
Tooling Improvements
Manufacturing CostImprovementsQuality Control
Manufacturing Control
Operator Learning
Engineering Changes toAssist Production
Source: P Jefferson, ‘Productivity Comparisons with the USA – where do we differ?’ Aeronautical Journal, Vol 85 No844 May 1981, p.179
Constituent Elements of Production Learning
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
Segmenting the Learning Curve: Mathematical Model
Consider 4 cost driver components with values α, β, γ, and δwhere α + β + γ + δ = 1 (or 100%)
Equation of a Unit Learning Curve: TA = T1 Aε
where ε is the learning exponent: ε = log(p)/log(2)with p = the learning percentage expressed as a decimaland TA is the time at Unit A
Expand the exponent: TA = T1 A(α + β + γ + δ) ε
TA = T1 Aαε Aβε Aγε Aδε
In order to model data with breakpoints, re-define the variable A:TA = T1 A1
αε A2βε A3
γε A4δε
For the primary learning (where all cost drivers are “active”), the values of A1 A2 A3 and A4 are all equal
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
454510560
454510545
101010510
55555
44444
33333
22222
11111
A4A3A2A1ALogisticsToolingOperatorDesignBuild No
Segmenting the Learning Curve: Mathematical Model
Example based on a production run of 60 units All cost drivers active. Relative learning points are all equal
Impact of constant output rate truncates relative learning for this cost driver
Impact of design freeze truncates relative learning for this cost driver
“End of Line”truncates relative learning for these cost drivers
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
Learning Curve Segmentation: Points to Consider
Benefits of Approach:
• Allows discontinuities to be modelled easily (using an on/off switch approach)
• Allows scenarios to be modelled which assume learning rates greater than or less than the “norm” for a particular process or product type
• Allows multiple linear regression techniques to be applied in cost data analysis
Words of Caution:
• As with all modelling techniques, the approach requires calibration for the specific environment in which it is to be applied
• There should be a logical model or explanation of why particular cost drivers have been “switched in or out”
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
Effect of Output Rate Constraint on Learning
Application Example
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
AverageContents
Effect of Output Rate Constraint on Learning
Number of Operators Average Hours Workedin Time Period
Average Hours spentper Unit in Time Period
Number of Unitsproduced in Time Periodx x=
Constant(For Optimum
Learning)
“Constant”(Effective Upper& Lower Limits)
Reducing(Learning Curve)
Increasing(Rate Ramp-up)
Every operatorperforms same task
on every unit
Constrained by workinghour practices (basic
working week &sustainable overtime
The “Reduced Cost : Increased Output” is in part a natural response of increased product familiarity, and in part a response to market
expectations of affordability etc
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
AverageContents
Effect of Output Rate Constraint on Learning
Number of Operators Average Hours Workedin Time Period
Average Hours spentper Unit in Time Period
Number of Unitsproduced in Time Periodx x=
Constant(Fixed Output Rate)
Reducing “Constant”(Effective Upper& Lower Limits)
Reducing(Learning Curve)
Constrained by workinghour practices (basic
working week &sustainable overtime
Reducing the number of operators violates
the premise for optimum learning
A response to market expectations of
affordability etc to drive down costs
Customer contractual limitation or constraint
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
Example 1: Cumulative Deliveries of Product A
0
50
100
150
200
250
300
350
Years
Cumul
ative
Uni
tsConstant Rate
DeliveriesDelivery Rate
Build-up
117
9.75 per month
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
Example 2: Assembly Learning for Product A
1 10 100 1000Cumulative Units
Man
-hou
rs
Actual Regression 5% Confidence Level 95% Confidence Level
Constant RateDeliveries
Delivery RateBuild-up
Breakpoint@ 117
80.4% Learningafter the breakpoint
75.7% Learningup to the breakpoint
22%
78%
Swingometer
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
Example 2: Cumulative Deliveries of Product B
0
50
100
150
200
250
Years
Cumul
ative
Uni
tsConstant Rate
DeliveriesDelivery Rate
Build-up
60
4 per month
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
Example 2: Assembly Learning for Product B
1 10 100 1000Cumulative Units
Man
-hou
rs
Actual Regression 5% Confidence Level 95% Confidence Level
Constant RateDeliveries
Delivery RateBuild-up
Breakpoint@ 60
87.8% Learningafter the breakpoint
72.1% Learningup to the breakpoint
60%
40%
Swingometer
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
Effect of Output Rate Constraint on Learning
Other factors affecting the analysis:
• The examples emanate from different factories with differentmanagement styles and cultural heritage
• One product was essentially for a single customer variant/mark initially followed by small batch export orders
• The other product was a multiple variant/mark international collaboration• The level of continued investment was geared around the known and
perceived market opportunities• The level and timing of engineering change required to introduce export
variants and support customer modifications has to be considered• The underlying manufacturing technology used on the two products was
similar but not identical
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
“End of Line” Effect on Learning Curves
Application Example
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
30
40
50
60
70
80
90
100
1 2 3 4 5 6 7 8 9 10 11 12
“End of Line” Effect on Learning Curves
Premise:To enable ongoing learning curve reduction once a constant rate of output is achieved requires investment in new or improved technology, process or logistics etc
Cumulative Return on InvestmentReduced saving
per unit
Reduced quantity remaining over which investment can be recovered
Diminishing Return on Investment
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
“End of Line” Effect on Learning Curves
0.001
0.01
0.1
1
10
100
1000
1 10 100 1000Cumulative Units
Fact
or (C
umul
ativ
e R
etur
n on
Inve
stm
ent)
¾ Quantity
Learning Rate75%80%85%90%
75%80%85%90%
¾ Quantity
Diminishing Cumulative Return on Investment =(Unit Learning Curve Reduction) x (Units Remaining)
It would seem that there is a case that a learning curve will truncate
naturally somewhere between the 60% to 80% point of the total envisaged
production quantity, regardless of the learning curve rate?
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
“End of Line” Effect on Learning Curves
1000
10000
100000
1 10 100 1000Cumulative Units
Fact
or (C
umul
ativ
e Re
turn
on
Inve
stm
ent)
¾ Quantity
The empirical relationship of the “End of Line”Effect on a learning curve can be attributed to
the “Law of Diminishing Returns”.
It is not unreasonable to expect that a learning curve will truncate naturally
somewhere between the 60% to 80% point of the total envisaged production quantity.
Breakpoint@ Constant Rate
Example:• Constant rate of output at unit 50• 400 units planned in total• 75% Learning Curve
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
Assessing Loss of Learning
Application Example
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
Assessing Loss of Learning: Anderlohr Method
1 10 100 1000Cumulative Units
Man
-hou
rs
Basic AnderlohrSource: Anderlohr, G., ‘What production breaks cost’, Journal of Industrial Engineering,
September 1969, pp.34-36
Consider a Break in Production of 12 months after 50 units
1. Determine how many units have been produced in the previous 12 months
2. Back track up the learning curve by this quantity
3. This defines the re-start position for learning
4. Repeat the learning process (offset by the number of units lost)
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
Assessing Loss of Learning: Segmentation Method
1 10 100 1000Cumulative Units
Man
-hou
rs
Basic With Re-learning Continued Learning
Consider a Break in Production of 12 months after 50 units
1. Determine the proportion of learning that will continue by considering the cost drivers that might be affected
3. After the break the continued learning component still applies
4. Factor this by the re-learning component (offset by the number of units lost)
1. Determine the proportion of learning that will continue by considering the cost drivers that might be affected
2. This defines the re-start position for learning after the break
Component subject to Re-Learning
69%
31%
Example
Continued Component of Learning
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
Assessing Loss of Learning: Comparison of Methods
1 10 100 1000Cumulative Units
Man
-hou
rs
Basic Anderlohr Segmentation
Consider a Break in Production of 12 months after 50 units
Anderlohr method always lags the segmentation method for
the same re-start value
Anderlohr Method
Segmentation Method
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
Assessing Loss of Learning: Comparison of Methods
Practical Considerations:
• Small breaks in production will be more difficult to detect further down the curve due to potential “noise” in the actual data
• The Anderlohr Method assumes that the rate of learning loss is equivalent to the rate of learning gain. This is not necessarily the case, but a modified approach which “backtracks” only a proportion of the “lost”learning could be adopted
• What happens when the break in production occurs during the latter stages of the production run (often the case)? The learning curve may have “bottomed out” by this stage
• Either approach could be applied to other cases of learning loss other than time breaks; for example, a physical relocation or new start-up.
Consider the following example using the cost driver segmentation method
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
Example 3: Cumulative Deliveries of Product C
800
820
840
860
880
900
920
940
960
980
1000
Years
Cumul
ative
Uni
ts3-Year BreakIn Production
2 per month@ peak
2 per month@ end of line
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
Example 3: Assembly Learning for Product C
700 750 800 850 900 950 1000Cumulative Units
Man
-hou
rs
Actual Regression
Re-learningRate restricted learning
Break in Production
29%
71%
Swingometer
22%
78%
Swingometer
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
Multi-Ganging of Operations: Parallel Learning
Application Example
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
Multi-Ganging of Operations: Parallel Learning
1 10 100Cumulative Units
Man
-hou
rs
Common Learning Series Working 2-Gangs 4-Gangs 8-Gangs
Multi-Gang parallel working has the effect of deferring learning curve reduction by a proportion of the lost operator contribution
This has the apparent effect of reducing the observed learning and
increasing the theoretical First Unit Value
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
Cumulative and Cumulative Average Data
Alternative Approaches
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
Cumulative Average Data
Cumulative Average Model:
• The formula for the Cumulative Average version of a Learning Curve is the same as that for a Unit Learning Curve:
TA = T1 Aε
where ε is the learning exponent: ε = log(p)/log(2)with p = the learning percentage expressed as a decimaland TA is the Cumulative Average Time at Unit A
• The Cumulative Average version will be inherently “smoother” than its Unit counterpart, but the rate of learning indicated will be very similar for higher quantities (greater than 30 – depending on the accuracy required)
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
Cumulative Average Data
1 10 100 1000Cumulative Units
Man
-hou
rs
Unit Unit Cum Ave Unit Regression
Cumulative Average Curve runs parallel to the Unit
Curve for larger quantities
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
Cumulative Data Approximation Formulae
Cumulative Data Approximations for a Unit Learning Curve:
For a positive error1, CA ~ T1 [ (A + 0.5)ε+1 - 0.5ε+1 ]
(ε + 1)
For a negative error2, CA ~ T1 (Aε+1 - 1) + T1 (Aε + 1)
(ε + 1) 2
where ε is the learning exponent: ε = log(p)/log(2)with p = the learning percentage expressed as a decimal
Source: 1. Conway, R.W. and Schultz, A.Jr., ‘The Manufacturing Progress Function’, Journal of Industrial Engineering, Jan-Feb 1959, pp.39-542. Jones, A.R. ‘Case Study - Applying Learning Curves in Aircraft Production - Procedures and Experiences’ in Zandin, K (editor)
Maynards Industrial Engineering Handbook, 5th Edition, McGraw-Hill, New York, 2001
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
Cumulative Data Approximation Formulae Error
Source: Jones, A.R. ‘Case Study - Applying Learning Curves in Aircraft Production - Procedures and Experiences’ in Zandin, K (editor) Maynards Industrial Engineering Handbook, 5th Edition, McGraw-Hill, New York, 2001
-2.50%
-2.00%
-1.50%
-1.00%
-0.50%
0.00%
0.50%
1.00%
1.50%
2.00%
2.50%
0 10 20 30 40 50 60 70 80 90 100
Cumulative Units
% Error
Jones Approximation
Cauchy-Schwartz Approximation80% learning curve
75% learning curve
© 2005 BAE Systems
Cost Drivers Learning Event, 2nd November 2005
Cumulative Data Equivalent Unit Completion Method
0 . 1 1 1 0 1 0 0Cumulative Units
Man
-hou
rs
0
10
20
30
40
50
60
Cumulative Units
Calendar Time
Cumulative Average
Unit Learning Curve
Cumulative Average based onEquivalent Unit Completions