tutorial session 4: return on investment · discrete event simulation (des) • dynamic simulation...
TRANSCRIPT
1
University of MarylandCopyright © 2015 CALCE
1calceTM
1Prognostics and Health Management Consortium
Tutorial Session 4:Return on Investment
(Determining the Cost and Value of PHM)
Peter Sandborn(301) 405-3167
PrognosticsTM
PHM Conference 2015
University of MarylandCopyright © 2015 CALCE
2calceTM
2Prognostics and Health Management Consortium
Prognostics and Health Management
• Components• Interconnects• Boards• LRUs• Wiring• Systems
Prognostics Sensors
Analysis Methods
Predicted remaining life
Time
Acc
eler
atio
n F
acto
r (A
F)
Failure of prognostic sensors
Failure of actual circuit
Predicted remaining life
Time
Acc
eler
atio
n F
acto
r (A
F)
Failure of prognostic sensors
Failure of actual circuit
Time
Haz
ard
Rat
e
Infant Mortality
Useful life Wear-out
Prognostic cell
Actual circuitry
Time
Haz
ard
Rat
e
Infant Mortality
Useful life Wear-out
Prognostic cell
Actual circuitry
PrognosticsAnalysis
Accumulating damage
Determining remaining life
Data collection
and reduction
Deriving Value
Decision making
This Tutorial
2
University of MarylandCopyright © 2015 CALCE
3calceTM
3Prognostics and Health Management Consortium
Cost and Return on Investment (ROI) Analysis
Peter Sandborn(301) 405-3167
PrognosticsTM
PHM Conference 2015
University of MarylandCopyright © 2015 CALCE
4calceTM
4Prognostics and Health Management Consortium
The Fundamental Maintenance/Value Tradeoff
Goal (depending on application) is either to:
Perform maintenance so that the remaining useful life (RUL) and the remaining useful performance (RUP) are minimized (use up as much of the life as possible), while simultaneously avoiding failures (unscheduled maintenance)
or
Find the optimum mix of scheduled and unscheduled maintenance that minimizes the life-cycle cost
Note, we are not advocating the avoidance of all failures here, this is NOT a “safety” argument, it is an economic argument.
3
University of MarylandCopyright © 2015 CALCE
5calceTM
5Prognostics and Health Management Consortium
Deriving Value from PHM at the System and Enterprise Levels
– System-level PHM value means taking action based on prognostics to manage one specific instance of a system, e.g., one truck or one turbine. The actions tend to be “real-time” and consist of:
• Modify how you sustain the system (e.g., call ahead to arrange for a maintenance action)
• Modify the mission (e.g., reduce speed, take a different route)
• Modify the system (e.g., adaptive re-configuration)
– Enterprise-level PHM value means taking action based on prognostics to manage an enterprise, e.g., a whole fleet of trucks or a farm of turbines. The actions are longer-term strategic planning things (usually not real-time):
• Optimizing the logistics
• Management via availability and other outcome-based contracts
The optimum for an individual system instance ≠ the optimum for the system instance within a population if the population is managed via an outcome-based contract
University of MarylandCopyright © 2015 CALCE
6calceTM
6Prognostics and Health Management Consortium
Cost of PHM Implementation• Development cost
– Hardware and software design, development, testing and qualification– Integration costs
• Additional costs associated with product manufacturing– Recurring cost per product for additional hardware, additional processing, additional
recurring functional testing– Installation costs
• Cost of creating and maintaining the infrastructure to make effective use of the PHM data
– Cost of data archiving– Cost of maintaining the PHM structures (logistics footprint)– Cost of training personnel– Cost of creating and maintaining documentation– Cost of changing the logistics/maintenance culture
• Cost of performing the necessary analysis to make it work– Cost of data collection– Cost of data analysis– Cost of false positives
• Financial costs (cost of money)– $1 today (to implement PHM) costs more than $1 to repair tomorrow
4
University of MarylandCopyright © 2015 CALCE
7calceTM
7Prognostics and Health Management Consortium
Potential Cost Avoidance (Return)
Associated with PHM
• Failures avoided– Minimizing the cost of unscheduled maintenance– Increasing availability– Reducing risk of loss of system– Increased human safety
• Minimizing loss of remaining life– Minimizing the amount of remaining life thrown away by
scheduled maintenance actions
• Logistics (reduction in logistics footprint)– Better spares management (quantity, refreshment, locations)
o Lead time reductiono Better use of (control over) inventory
– Minimization of investment in external test equip– Optimization of resource usage
• Repair– Better diagnosis and fault isolation (decreased inspection time,
decreased trouble shooting time)– Reduction in collateral damage during repair– Reduction in post-repair testing
• Reduction in redundancy (long term)– Can redundancy be decreased for selected sub-systems?
• Reduced waste stream– Less to end-of-life (dispose of) – disposal avoidance– Reduction in take-back cost
• Reduction in no-fault-founds• Eases design and qualification
of future systems• Reduced liability• Warranty claim verification
(resulting in warranty reserve fund reduction)
Problem: Future cost avoidance is a hard sell
University of MarylandCopyright © 2015 CALCE
8calceTM
8Prognostics and Health Management Consortium
PHM Cost ModelDiscrete-event simulation that follows a population of
sockets through their lifetime from first LRU installation to retirement of the socket.– “Discrete-event simulator” refers to the simulation of a
timeline, where specific events are added to the timeline and the resulting event order and timing can be used to analyze throughput, cost, availability, etc.
– “Socket” refers to one instance of an installation location for an LRU.
– “Population” means that the simulator is stochastic (governed by the laws of probability) so that a statistically significant number of non-identical fielded systems can be assessed and the results are distributions rather than single values.
5
University of MarylandCopyright © 2015 CALCE
9calceTM
9Prognostics and Health Management Consortium
Discrete Event Simulation (DES)
• Dynamic simulation (models changes over time)
• State variables change only at a discrete set of points in time (i.e., at “events”)
• Event = something that happens to the system at an instant in time that may change the state of the system (by definition, nothing relevant to the model changes between events)
• Discrete (successive changes are separated by finite amounts of time)
• Timeline = the sequence of events and their calendar times
• Path = a particular sequence of events
University of MarylandCopyright © 2015 CALCE
10calceTM
10Prognostics and Health Management Consortium
Following Sockets vs. LRUs
The discrete-event simulation follows a population of sockets through their lifetime (socket = the installation location of an LRU); issues with modeling sockets:– Easy to calculate socket cost and availability
– Implicit assumption of a stable population of LRUs
– Not-good-as-new repair – easy to model if you assume the same LRU comes back to the socket after repair
Alternatively, a simulation could follow LRUs; issues with following LRUs:– Repaired LRUs don’t necessarily go back into the same socket, so
you must model the LRU supply chain
– How are socket failures accounted for?
– Difficult to calculate socket cost and availability
6
University of MarylandCopyright © 2015 CALCE
11calceTM
11Prognostics and Health Management Consortium
Discrete Event Simulation for PHM
Time
Repeat the process for many socketsGenerate a histogram of the computed quantities
1) Add time zero implementation costs:• Base LRU recurring cost• PHM LRU recurring cost• LRU/socket non-recurring costs• System recurring cost
For one socket:
2) Predict failure date of LRU in the socket
3) Determine PHM predicted removal date
• Incorporates prognostic distance or safety margin
4) Maintain system on either the actual failure date or the PHM predicted removal date (whichever comes first)
• LRU replacement/repair cost• Costs associated with operational
profile
5) Start over at step 2) with a new or repaired LRU in the socket and continue process from the socket maintenance date until the end of the field support life
Infrastructure cost6) Compute/accumulate:
• Life-cycle cost/socket• Cost/operating hour • Availability• Failures avoided• Number of LRUs/socket
Every value used is “sampled” from a probability distribution that represents the input parameter
University of MarylandCopyright © 2015 CALCE
12calceTM
12Prognostics and Health Management Consortium
Discrete Event Simulation – Data-Driven
• 1,000 sockets simulated
• Small steps in the graph correspond to annual accumulation of infrastructure costs
• Big jumps in cost correspond to replacement of the LRU (average of 5 LRUs used per socket over the support life)
Unscheduled maintenance events (PHM missed these)
Scheduled maintenance events (PHM caught these)
7
University of MarylandCopyright © 2015 CALCE
13calceTM
13Prognostics and Health Management Consortium
Very Simple Baseline Data Assumptions for the Example Cases
• 0 investment cost• 0 infrastructure cost• Spares assumed to be available and purchased as needed
Variable in the model Value used for example analysis
Production cost (per unit) $10,000 Time to failure 5000 operational hours most likely value (symmetric triangular
distribution with variable d istribution width) Operational hours per year 2500
Sustainment life 25 years Unscheduled Scheduled
Value of each hour out of service $10,000 $500 Time to repair 6 hours 4 hours
Time to replace 1 hour 0.7 hours Cost of repair (materials cost) $500 $350 Fraction of repairs requiring
replacement of the LRU (as opposed to repair o f the LRU)
1.0 0.7
Various values and distributions(TTF)
University of MarylandCopyright © 2015 CALCE
14calceTM
14Prognostics and Health Management Consortium
PHM Approaches• Reliability-Based – Historical reliability predictions used to produce
failure distributions– Average part, average environmental stress
• Model-Based (Physics of Failure) – Observing the environmental stress history that a system has been subjected to and deciding based on an understanding of a nominal system, that the system may be unhealthy (indirect method) – degradation inferred from environmental stress
– Average part, actual environmental stress
• Data-Driven – Directly observing the system and deciding that it looks unhealthy (e.g., precursor to failure, LRU-dependent canaries, anomaly detection) – measured properties are correlated to actual failure
– Actual part, actual environmental stress
• Fusion – Combines data-driven and physics of failure approaches together to enable calculation of remaining useful life (RUL)
8
University of MarylandCopyright © 2015 CALCE
15calceTM
15Prognostics and Health Management Consortium
Background – Time to Failure (TTF) Distributions
• Doesn’t have to be “time”, could be thermal cycles, miles, takeoffs and landings, etc.
• Triangular distributions shown for simplicity
Pro
babi
lity
LRU Time-to-Failure (TTF)
Nom
inal
LR
U
The LRU’s TTF distribution represents variations in manufacturing and materials
LRU Instance
Actual TTF (sample)Time to Failure (t), hours
Fra
ctio
n of
par
ts f
aili
ng, f
(t)
0 100 200 300 400 500 600 700 800 9000
0.30
0.20
0.10
If the test was run until all instances of the product failed, then the area under the curve is 1 and we have a PDF of failures
University of MarylandCopyright © 2015 CALCE
16calceTM
16Prognostics and Health Management Consortium
100000
150000
200000
250000
300000
0 2000 4000 6000 8000 10000 12000
Fixed Scheduled Maintenance Interval (operational hours)
Eff
ec
tiv
e L
ife
Cy
cle
Co
st
(pe
r s
oc
ke
t)
1000 hours
2000 hours
4000 hours
6000 hours
8000 hours
10000 hours
Fixed Interval Scheduled Maintenance
TTF Distribution is
narrow
TTF distribution is wide
(10,000 sockets followed)
Width of the time to failure distribution
~Unscheduled maintenance solution
Single Socket
9
University of MarylandCopyright © 2015 CALCE
17calceTM
17Prognostics and Health Management Consortium
Data-Driven Methodologies(Precursor to Failure, Health Monitoring, LRU Dependent Canary)
• Example: A canary or other monitored structure is manufactured with the LRUs, i.e., it is coupled to a particular LRU’s manufacturing or material variations
• Prognostic Distance = length of time (in operational hours) before system failure that the prognostic structures are designed to indicate failure
Pro
babi
lity
LRU Time-to-Failure (TTF)P
roba
bili
tyPHM structure Time-to-Failure
Prognostic Distance
Sam
pled
TT
F
fore
cast
and
m
aint
enan
ce in
terv
al
for
the
sam
ple
Nom
inal
LR
U
The LRU’s TTF distribution represents variations in manufacturing and materials
LRU Instance
Actual TTF (sample)
Indicated Replacement Time
LRU Instance Actual TTF
(TTF distribution of the monitored
structure)
University of MarylandCopyright © 2015 CALCE
18calceTM
18Prognostics and Health Management Consortium
Model-Based Methodologies (LRU Independent, Life Consumption Monitoring (LCM), LRU Independent Canary)
• The PHM structure (or sensors) are manufactured independent of the LRUs, i.e., it is not coupled to a particular LRU’s manufacturing or material variations
• Safety Margin (Designed Prognostic Distance) = length of time (in operational hours) before failure of the nominal LRU that the PHM approach/structure is design to indicate failure.
Pro
babi
lity
LRU Time-to-Failure (TTF)
LRU Instance
Actual TTF (sample)
Pro
babi
lity
PHM structure Time-to-Failure
Safety Margin
Sam
pled
TT
F
fore
cast
and
m
aint
enan
ce in
terv
al
for
the
sam
ple
Nom
inal
LR
U
Nom
inal
LR
U
Indicated Replacement Time
The LRU’s TTF distribution represents variations in manufacturing and materials
LRU Instance Actual TTF
10
University of MarylandCopyright © 2015 CALCE
19calceTM
19Prognostics and Health Management Consortium
Comments
• The fundamental difference between the two models:• Data-Driven = the TTF distribution associated with the PHM structure
(or sensor/canary) is unique to each LRU instance• Model-Based = the TTF distribution associated with the PHM
structure (or sensor/canary) is tied to the nominal LRU and knows nothing about manufacturing/material variations between LRU instances
• Notes:• Failure does not have to be characterized by time – it could be cycles,
etc.• Triangular distributions are only used for simplicity
University of MarylandCopyright © 2015 CALCE
20calceTM
20Prognostics and Health Management Consortium
Data-Driven vs. Model-Based Methods(Varying TTF Distribution Width)
(10,000 sockets followed)
Single Socket
Data-DrivenModel-Based
100000
110000
120000
130000
140000
150000
160000
170000
180000
190000
200000
0 500 1000 1500
Safety Margin (operating hours)
Eff
ec
tiv
e L
ife
Cy
cle
Co
st
(pe
r s
oc
ke
t)
1000 hr TTF width, 1000 hr LCM width
2000 hr TTF width, 1000 hr LCM width
4000 hr TTF width, 1000 hr LCM width
100000
110000
120000
130000
140000
150000
160000
170000
180000
190000
200000
0 500 1000 1500
Prognostic Distance (operating hours)
Eff
ec
tiv
e L
ife
Cy
cle
Co
st
(pe
r s
oc
ke
t)
1000 hr TTF width, 1000 hr HM width
2000 hr TTF width, 1000 hr HM width
4000 hr TTF width, 1000 hr HM width
Pro
babi
lity
Pro
babi
lity
LRU Time-to-Failure (TTF)
TTF WidthVariations in TTF distribution width
DD width
DD width
DD width
MB width
MB width
MB width
11
University of MarylandCopyright © 2015 CALCE
21calceTM
21Prognostics and Health Management Consortium
(10,000 sockets followed)
Single Socket
Data-DrivenModel-Based
100000
110000
120000
130000
140000
150000
160000
170000
180000
190000
200000
0 500 1000 1500
Safety Margin (operating hours)
Eff
ec
tiv
e L
ife
Cy
cle
Co
st
(pe
r s
oc
ke
t)
2000 hr TTF width, 1000 hr LCM width
2000 hr TTF width, 2000 hr LCM width
2000 hr TTF width, 4000 hr LCM width
100000
110000
120000
130000
140000
150000
160000
170000
180000
190000
200000
0 500 1000 1500
Prognostic Distance (operating hours)
Eff
ec
tiv
e L
ife
Cy
cle
Co
st
(pe
r s
oc
ke
t)
2000 hr TTF width, 1000 hr HM width
2000 hr TTF width, 2000 hr HM width
2000 hr TTF width, 4000 hr HM width
PHM WidthVariations in PHM distribution width P
roba
bili
tyP
roba
bili
ty
PHM Approach Time-to-Failure
Data-Driven vs. Model-Based Methods(Varying PHM Distribution Width)
MB width
MB width
MB width
DD width
DD width
DD width
University of MarylandCopyright © 2015 CALCE
22calceTM
22Prognostics and Health Management Consortium
Data-Driven vs. Model-Based Method Observations
Single Socket
1) The model-based approach is highly dependent on the LRU’s TTF distribution
2) Data-driven methods are approximately independent of the LRU’s TTF distribution
3) All things equal,* optimum prognostic distances for data-driven methods are always smaller than optimum safety margins for model-based methods,** and therefore,
4) All things equal,* data-driven PHM methods will always result in lower life cycle cost solutions that model-based methods**
*All things equal = same LRUs, same shape and size distribution associated with the PHM approach
**Assumes that you have a choice, i.e., that there is a data-driven method that is applicable – there may not be
12
University of MarylandCopyright © 2015 CALCE
23calceTM
23Prognostics and Health Management Consortium
Multiple Socket Systems
Coincident time = time interval within which different sockets should be treated by the same maintenance action.
If │Timecurrent maintenance action – Timerequired maintenance action on LRU i │ < Timecoincident
then LRU i is addressed at the current maintenance action
Coincident time = 0 means that each socket is treated independentlyCoincident time = infinite means that any time any LRU in the
system demands to be fixed, all sockets are fixed no matter what life expectancy they have
Next predicted maintenance action for socket i (from PHM approach)
Current maintenance action for the system
University of MarylandCopyright © 2015 CALCE
24calceTM
24Prognostics and Health Management Consortium
LRU2 Timeline
CumulativeTimeline L
RU
1 and LR
U2
Replaced
< Coincident time
LRU instance-specific TTF samples
LR
U1 and L
RU
2 R
eplaced
< Coincident time
Multi-Unit Timelines
LRU1 Timeline
> Coincident time
LR
U2 R
eplaced
Etc…end of support
Two different LRUs that are part of the same system
13
University of MarylandCopyright © 2015 CALCE
25calceTM
25Prognostics and Health Management Consortium
500000
550000
600000
650000
700000
750000
800000
850000
900000
1 10 100 1000 10000 100000
Coincident Time (operational hours)
Mea
n C
ost
(p
er
sys
tem
)
Many-LRU System
(Dissimilar LRUs)
6 different LRUs
LRU #1, γ* = 19900 hr (data-driven)LRU #2, γ* = 9900 hr (unscheduled maintenance)LRU #3, TTF** = 10,000 hr (fixed interval sched maintenance)LRU #4, TTF** = 20,000 hr (data-driven)LRU #5, TTF** = 20,000 hr (data-driven)LRU #6, TTF** = 5,000 hr (fixed interval sched maintenance)
*Weibull distributions**Symmetric triangular distributions
2 LRU System
6 LRU System
Minimum lifecycle cost
Multiple Sockets
Co
st p
er
Sys
tem
All LRUs maintained separately
All LRUs maintained at the same time
University of MarylandCopyright © 2015 CALCE
26calceTM
26Prognostics and Health Management Consortium
Evaluating the Return on Investment (ROI) Associated with PHM
What is ROI?
Investment
Investment-ReturnROI
Why evaluate the ROI?– To build a business case for implementation– To perform cost/benefit analysis on different prognostic approaches– Evaluate when PHM may not be warranted
Interpreting ROI:0 = breakeven (no cost impact)
> 0 there is a direct cost benefit
< 0 there is no direct cost benefit
Investment
Investment-AvoidanceCost
14
University of MarylandCopyright © 2015 CALCE
27calceTM
27Prognostics and Health Management Consortium
Formulating an ROI for PHM
• ROI relative to unscheduled maintenance gives
Cus = total life cycle cost using unscheduled maintenanceCPHM = total life cycle cost using the selected PHM approachIPHM = PHM investment cost
PHM
PHMus
I
CC ROI
• Investment cost
INFRECNREPHM CCCI
CNRE = PHM non-recurring costsCREC = PHM recurring costsCINF = PHM infrastructure costs
University of MarylandCopyright © 2015 CALCE
28calceTM
28Prognostics and Health Management Consortium
ROI for PHM (continued)
• Not so fast! Is IPHM complete? Are there other investment costs too?
• Example: Employing PHM will result in as many or maybe more maintenance events as unscheduled maintenance. If PHM results in the need for more spare replacement units, is the cost of these units an investment cost?
• The costs of: false alarm resolution, procurement of a different quantities of LRUs, and variations in maintenance costs are not included in the investment cost because they are the result of the investment and are reflected in CPHM
• CPHM must also include the cost of money differences associated with purchasing LRUs at differently timed maintenance events
This cost is a result of the investment, not part of the investment
15
University of MarylandCopyright © 2015 CALCE
29calceTM
29Prognostics and Health Management Consortium
Sandel ST3400 TAWS/RMI Electronic Display Unit
Example: PHM Return on Investment
• 502 Aircraft in fleet (Southwest Airlines)
• 2 sockets per aircraft• Support life: 20 years• Negligible false alarms assumed• 7% discount rate
Feldman, Jazouli, and Sandborn, “A Methodology for Determining the Return on Investment Associated with Prognostics and Health Management,” IEEE Trans. On Reliability, June 2009.
University of MarylandCopyright © 2015 CALCE
30calceTM
30Prognostics and Health Management Consortium
Precursor to Failure
$76,000
$78,000
$80,000
$82,000
$84,000
$86,000
$88,000
0 200 400 600 800 1000 1200
Prognostic Distance (hours)
Lif
eCyc
le C
ost
per
So
cket
Prognostic Distance
• For a data-driven PHM approach, an analysis is performed to find the prognostic distance that yields the lowest cost
• The prognostic distance that produces the lowest costs is a function of the inputs and is application-specific
• For the data-driven approach, used here, the prognostic distance was chosen as 475 hours
All failures avoided, but lots of remaining life thrown away
Few failures avoided, approaching unscheduled maintenance
Data-Driven
16
University of MarylandCopyright © 2015 CALCE
31calceTM
31Prognostics and Health Management Consortium
Time (years)
Cum
ulat
ive
Cos
t per
Soc
ket
Data-Driven PHM Approach
Unscheduled Maintenance
Cumulative Life-Cycle Costs
University of MarylandCopyright © 2015 CALCE
32calceTM
32Prognostics and Health Management Consortium
-5.0
0.0
5.0
10.0
15.0
20.0
25.0
30.0
35.0
0 500 1000 1500 2000 2500 3000 3500 4000 4500
Annual Infrastructure Cost per Socket ($)
Ret
urn
On
In
vest
men
t
ROI Analysis: Data-DrivenThe evaluation of ROI (relative to unscheduled maintenance) as a function of various implementation costs
Breakeven Point
> 0, Cost benefit
< 0, No cost benefit
Mea
n R
etu
rn o
n In
vest
men
t
17
University of MarylandCopyright © 2015 CALCE
33calceTM
33Prognostics and Health Management Consortium
ROI as a Function of Time
First maintenance event
Small ROI increase (at 3rd maintenance event), which coincides with inventory replenishment. PHM saves money on maintenance events, BUT buys more expensive spares (that include PHM recurring costs)
PHM missed this failure, therefore, it was resolved as unscheduled maintenance event
Annual inventory cost is greater for PHM, since the spares are more expensive
ROI of data-driven PHM relative to unscheduled maintenance
University of MarylandCopyright © 2015 CALCE
34calceTM
34Prognostics and Health Management Consortium
ROI for the Application of PHM to Wind Turbines (EADS)
TRIADE includes temperature, pressure, vibration, strain and acoustic sensors that can be used to monitor the health of turbine blades in a wind turbine
Time history of 1000 wind turbines:
Fixed interval maintenance
Using TRIADE (data-driven PHM)
Haddad, Sandborn, Jazouli, Foucher, and Rouet, “Guaranteeing High Availability of Wind Turbines,” Proc. ESREL, 2011.
18
University of MarylandCopyright © 2015 CALCE
35calceTM
35Prognostics and Health Management Consortium
Other Things to Consider …
• Redundancy• Not “as good as new” repair• Socket failures• Multiple failure mechanisms• Simple canaries are modeled as LRU independent fuses,
but may actually be mixtures of fuses and LRU-independent methods
• Second order uncertainty (uncertainty about uncertainty) may be a very real thing for this analysis
• Determining the right shape and size of distributions associated with various PHM approaches
University of MarylandCopyright © 2015 CALCE
36calceTM
36Prognostics and Health Management Consortium
Making Business Cases
• Cost avoidance is a hard sell
• Program/project management hears that the “sky is falling” every day from engineers
• Resource allocation is based on:1) The occurrence of serious
events2) Projected future impact (cost,
availability)
19
University of MarylandCopyright © 2015 CALCE
37calceTM
37Prognostics and Health Management Consortium
ResourcesGeneral Cost/Maintenance Model Description:
P.A. Sandborn and C. Wilkinson, “A Maintenance Planning and Business Case Development Model for the Application of Prognostics and Health Management (PHM) to Electronic Systems,” Microelectronics Reliability, Vol. 47, No. 12, pp. 1889-1901, December 2007.
http://www.enme.umd.edu/ESCML/Papers/MR_8772-SandbornComplete.pdf
Return on Investment Modeling:
K. Feldman, T. Jazouli, and P. Sandborn, “A Methodology for Determining the Return on Investment Associated with Prognostics and Health Management,” IEEE Trans. on Reliability, pp. 305-316, June 2009.
http://www.enme.umd.edu/ESCML/Papers/Feldman_et_al_IEEE_Trans_Rel.pdf
PHM Applied to Electronics:
M. Pecht editor, Prognostics and Health Management of Electronics, ed. M. Pecht, John Wiley & Sons, Inc., Hoboken, NJ, 2008.