Post on 16-Sep-2020
Supply Chain Planning in the Consumer Electronics Industry
Real AI 2012
David Lesaint david.lesaint@univ-‐angers.fr
LERIA -‐ Université d’Angers -‐ France
Contents
• TFT-‐LCD Panels – Manufacturing operaCons – Supply chain planning – SCP as discrete opCmisaCon
• Constraint Programming Approach – MulC-‐pass adaptaCve planning – A generic CP model – Experiments
• Summary and Outlook
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 2
Contents
• TFT-‐LCD Panels – Manufacturing operaCons – Supply chain planning – SCP as discrete opCmisaCon
• Constraint Programming Approach – MulC-‐pass adaptaCve planning – A generic CP model – Experiments
• Summary and Outlook
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 3
TFT-‐LCD Panel Manufacturing
• Panel FabricaCon – etch color filters (CF) and thin-‐film
transistors (TFT) onto glasses – match and assemble CF and TFT
panels (cell producCon)
• Module Assembly – assemble display drivers, back-‐light
units (LEDs), PCB, etc
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 4
Manufacturing Process
• Reentrant flows • MulC-‐funcConal equipments
• High uClisaCon 24/7 x 365 • High contenCon
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 5
LCD Panel Fab Layout
• Reentrant flows • MulC-‐funcConal equipments
• High uClisaCon 24/7 x 365 • High contenCon
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 6
(from Choi et al. 2010, KAIST)
LCD Panel Supply Chain
• Panels for TVs, PCs and mobiles – about 20 fab and assembly plants – hubs, vendor-‐managed
inventories and customer warehouses located worldwide
– air, sea, road or rail shipping
• Three successive stages mixing producCon and transportaCon acCviCes 1. Panel fabricaCon 2. Module assembly 3. Shipping
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 7
Lead Cmes • 3-‐5 days for fab • 1 day for assembly • 1-‐30 days for shipping
Variability • 1K items • 10K routes • 1K resources • 1K inventories
Volume • 1K products • 10K orders / month • 10M panels / month
Contents
• TFT-‐LCD Panels – Manufacturing operaCons – Supply chain planning – SCP as discrete opCmisaCon
• Constraint Programming Approach – MulC-‐pass adaptaCve planning – A generic CP model – Experiments
• Summary and Outlook
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 8
Scope of Supply Chain Planning
• Plan producCon and transportaCon orders – for the next 4 months [28 daily Cme buckets + 12 weekly buckets] – every week [rolling plan with 20% refreshment rate] – plan communicated to plants and suppliers for execuCon and procurement
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 9
(from Rohde et al 2000, PPS-‐M
)
Business ObjecCves of SCP
• Just-‐in-‐Cme and complete delivery – minimise lateness and earliness
• Maximise resource capacity uClisaCon – balance and smoothen uClisaCon
• Reduce transportaCon costs • PrioriCse customer volume allocaCon
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 10
Build a plan for each confirmed/forecast
sales order
Challenges of SCP
• Dynamic nature of parts supplies – eg semiconductor shortages
• Excessive air freight • Unbalanced resource uClisaCon • Inflated demand
– eg sales mistrust SCP, forecast inaccuracies
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 11
Build a plan for each confirmed/forecast
sales order
Challenges of SCP
• Dynamic nature of parts supplies – eg semiconductor shortages
• Excessive air freight • Unbalanced resource uClisaCon • Inflated demand
– eg sales mistrust SCP, forecast inaccuracies
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 12
Build a plan for each confirmed/forecast
sales order
Subop&mal Planning!
Contents
• TFT-‐LCD Panels – Manufacturing operaCons – Supply chain planning – SCP as discrete opCmisaCon
• Constraint Programming Approach – MulC-‐pass adaptaCve planning – A generic CP model – Experiments
• Summary and Outlook
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 13
Key Planning Decisions
• Route(s) selecCon – mulC-‐level BOMs for products – alternate routes for each item
• Resource(s) allocaCon – single or simultaneous resources – atomic or aggregate resources
• Lot-‐sizing – how much to consume, produce
and store • Scheduling
– when to consume, produce and store
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 14
100
100
100 100
100
100
17”DF-X
Route1
1
1 use
100 100 init 1/2 1/1
supply 300 0 demand 1/2 1/1
ITC
supply demand
1/2 1/1
use 150 150 init 1/2 1/1
WC1 WC2
2/1 2/2 2/3 2/4 2/5 2/6 2/7 2/8
24
16
8
0
Efficiency
FOR EACH ORDER:
Core Combinatorial Model
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 15
Resource
Route Inventory Inventory
Route Inventory
Route
... Resource Resource
... Resource Resource
Inventory
alternaCve resources
aggregate resources
Material
BOR BOR BOR
BOR-‐set
BOR
p-‐BOM
p-‐BOM
p-‐BOM
c-‐BOM c-‐BOM
c-‐BOM
c-‐BOM
c-‐BOM
c-‐BOM
c-‐BOM c-‐BOM c-‐BOM
c-‐BOM c-‐BOM
Core Constraints
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 16
Resource
Route Inventory Inventory
Route Inventory
Route
... Resource Resource
... Resource Resource
Inventory
alternaCve resources
aggregate resources
Material
BOR BOR BOR
BOR-‐set
BOR
p-‐BOM
p-‐BOM
p-‐BOM
c-‐BOM c-‐BOM
c-‐BOM
c-‐BOM
c-‐BOM
c-‐BOM
c-‐BOM c-‐BOM c-‐BOM
c-‐BOM c-‐BOM
• Cme-‐phased yield • priority, cost, proporCon
Core Constraints
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 17
Resource
Route Inventory Inventory
Route Inventory
Route
... Resource Resource
... Resource Resource
Inventory
alternaCve resources
aggregate resources
Material
BOR BOR BOR
BOR-‐set
BOR
p-‐BOM
p-‐BOM
p-‐BOM
c-‐BOM c-‐BOM
c-‐BOM
c-‐BOM
c-‐BOM
c-‐BOM
c-‐BOM c-‐BOM c-‐BOM
c-‐BOM c-‐BOM
• Cme-‐phased yield • priority, cost, proporCon
• consumpCon rate • priority, cost, proporCon
Core Constraints
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 18
Resource
Route Inventory Inventory
Route Inventory
Route
... Resource Resource
... Resource Resource
Inventory
alternaCve resources
aggregate resources
Material
BOR BOR BOR
BOR-‐set
BOR
p-‐BOM
p-‐BOM
p-‐BOM
c-‐BOM c-‐BOM
c-‐BOM
c-‐BOM
c-‐BOM
c-‐BOM
c-‐BOM c-‐BOM c-‐BOM
c-‐BOM c-‐BOM
• Cme-‐phased yield • priority, cost, proporCon
• capacity consumpCon rate • cycle-‐Cmes • Cme-‐phased efficiency • lot size min & increments
• consumpCon rate • priority, cost, proporCon
Required Features
5 Oct. 2012 Supply Chain Planning in Consumer Electronics
Id Feature 1 Due date fence 2 Material constraints 3 Route details & calendars 4 Shipping lead Cmes 5 Time-‐phased yield 6 Time-‐phased qty consumpCon 7 Efficiency rates 8 Resource capacity calendar 9 Resource down-‐Cmes 10 Over-‐prod. resource capacity 11 Resource capacity uClizaCon 12 ObjecCve-‐based simulaCon 13 Disconnected inventories 14 Resource-‐less routes 15 Incomplete BOMs 16 MulCple due dates (RTF) 17 Build-‐ahead days constraints 18 Demand/trans. mode pegging 19 EffecCve dates 20 Product life-‐cycles 21 Planned orders creaCon rule
Id Feature 22 Resource pooling 23 MulC-‐resource sales orders 24 MulC-‐bucket sales orders 25 MulC-‐route sales orders 26 Time-‐phased safety stock 27 Stock keeping-‐Cme 28 Policies -‐ delivery, demand priority 29 Policies -‐ cost-‐driven rouCng 30 Policies -‐ resource uClizaCon
maximizaCon & smoothing 31 Alternate materials 32 WIP projecCon 33 WIP pegging (Nenng) 34 Stock pegging (Nenng) 35 Time-‐phased line/plant assignmt. 36 Shipping schedule constraints 37 Time-‐phased efficiency 38 Time-‐phased takt Cme 39 Lot size 40 Net change producCon planning 41 Frozen plan, NOH handling
IN SCO
PE
OUT-‐O
F-‐SCOPE
19
RepresentaCve Instance Size
• Data Model – 10s of DB tables and columns – 100K records – 1mn to load an instance
• RepresentaCve Instance Size
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 20
• 20,000 BOMs • 10,000 c-‐BOMs • 10,000 p-‐BOMs
• 6,000 inventories • 10,000 routes • 1,500 resources
• 16-‐weeks long horizon • 28 daily buckets + 12 weekly buckets
• 1500 items (30 materials) • 10 item groups
• 10,000 sales orders • 12,000 planned orders
• 4,500 stock deliveries • 7,500 WIPs
• 12,000 BORs • 20 BOR-‐sets
Other Requirements
• Acceptable response Cmes
• Adaptable to problem instance characterisCcs – eg, seasonal demand, new product launch
• Decision-‐support – support what-‐if analysis – diagnose failure and subopCmality – repair or improve parts of a plan
• Plan stability
• System integraCon with MRP, ERP, etc
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 21
Contents
• TFT-‐LCD Panels – Manufacturing operaCons – Supply chain planning – SCP as discrete opCmisaCon
• Constraint Programming Approach – MulC-‐pass adaptaCve planning – A generic CP model – Experiments
• Summary and Outlook
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 22
ExisCng SCP System
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 23
Rule-‐based order-‐by-‐order planning
SequenCal process 1. WIP projecCon 2. stock order planning 3. inventory nenng 4. planned order planning
Incremental strategy • 1 stock/sales order at a Cme
Greedy algorithm • pre-‐ranking orders, routes, resources • backward-‐planning for planned orders • forward-‐planning for stock orders
ExisCng SCP System
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 24
Rule-‐based order-‐by-‐order planning
No search • alternaCves are not explored • may miss soluCons • may miss improvements
Same model for all orders • manually reconfigured aqer each run
No inference • does not prune search space using
constraints before and during search • may explore useless parts of search
space
Analysis
• MathemaCcal (Linear and Mixed Integer) Programming unsuitable – low-‐fidelity models – prohibiCve response Cmes
• Problem decomposiCon is a must – instance size – data-‐rich – orthogonal decisions – heterogeneous requirements ...
• Constraint Programming ... – ensures high-‐fidelity models – proven track-‐record on real-‐life
planning/scheduling problems
• ... But no one-‐size-‐fits-‐all soluCon! – use different CP models for
different subproblems – leverage COMET to this effect – develop a run-‐Cme configurable
CP engine
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 25
CP versus LP/MIP
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 26
FEATURE Mathema&cal Programming Constraint Programming
Relaxa&on Yes No
GAP measure Yes No
Op&mality Proof Yes Yes
Modeling Limita&ons QuadraCc problems limited to PosiCve Semi Definite problems and Second Order Cone Programming problems
Discrete problems
Specialised Constraints No Yes
Logical Constraints Yes Yes
Theore&cal Grounds Algebra AI, Graph Theory, Algorithms
(from ILOG 2009, OPL manual)
MulC-‐Paradigm OpCmisaCon Languages
• COMETTM
– mulC-‐paradigm opCmizaCon DSL – with Java-‐like programming layer – interpreted language
• Benefits – high-‐fidelity system engineering – eases validaCon – rapid prototyping – strong performances
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 27
IMPLEMENTATION PARADIGM
Embedded (eg. C++ library) Domain-‐Specific Language
MP ECLIPSE | OSCAR | ... OPL | COMET | MINIZINC ...
CP ECLIPSE | OSCAR | ILOG SOLVER |CHOCO | GECODE | ...
OPL | COMET | MINIZINC ...
Meta-‐Heuris&cs ECLIPSE | OSCAR | LOCALIZER | ... OPL | COMET ...
P. Van Hentenryck Brown univ. -‐ Dynadec
Problem DecomposiCon Approach
• MulC-‐pass planning – each pass iterates over a class of orders – one or more orders planned at each iteraCon – failed orders revisited in subsequent passes
• CP model – created at each iteraCon and customised for each (set of) order(s) – uses COMET global constraints and – porwolio of saCsfacCon/opCmizaCon algorithms
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 28
Problem too large for global approach Too much variability
for a single CP model
configure CP model
plan orders with CP model
select set of orders
next pass
MulC-‐Pass Planning Procedure
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 29
soluCon?
build CP model
solve
select orders
save soluCon
record failure
next itera*on
pass [CP solver]
yes
mulC-‐pass [Meta-‐Solver]
no
Contents
• TFT-‐LCD Panels – Manufacturing operaCons – Supply chain planning – SCP as discrete opCmisaCon
• Constraint Programming Approach – MulC-‐pass adaptaCve planning – A generic CP model – Experiments
• Summary and Outlook
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 30
Search Space
• RouCng search space is an OR-‐AND-‐tree – inventories = OR-‐nodes
• choose a single upstream route – routes = AND-‐nodes
• follow all upstream inventories – a plan is a tree of acCviCes
• Other decision points at each node – quanCCes and Cming of
producCon/consumpCon – resource allocaCon – stock consumpCon/creaCon – order shortage, etc
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 31
... ...
...
Materials
Inventory
Resource
Route
OR-‐node AND-‐node
Pazern
Search Space
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 32
...
...
... ...
...
...
...
...
... ...
... ...
...
Pazern
Search Space
Materials
Inventory
Resource
Route
OR-‐node AND-‐node
SoluCon Plans
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 33
...
...
... ...
...
...
...
...
... ...
Search Space
SoluCon Plan
CP Model Build & Solve
• Declare finite domains • Create decision variables • Post constraints • Search for a soluCon • Save soluCon or return failure
• All steps traverse the OR-‐AND-‐tree – either visit each node once – or visit each node mulCple Cmes
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 34
...
...
... ...
...
...
...
...
... ...
Search space
Variable CreaCon
• At each node by traversing supply net top-‐down
• Variables indexed by order ids, node ids and/or object ids
• Each variable models a specific aspect – rouCng decision, quanCty to
produce, stock to consume, start date of acCvity, etc
• Finite domain variables – symbolic – integers – booleans
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 35
...
...
... ...
...
...
...
...
... ...
... ...
...
Variable CreaCon Example
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 36
// order o!// inventory i!// RTE(i) = 1..#upstream_routes_of_i!!xRoute{<o,i>} = new var<CP>{int} = var<CP>{int}(_cp,RTE(i));!
... ...
...
Variable CreaCon Example
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 37
// order o!// inventory i!// POR(o) = set of planned orders for o!// RTE(i) = 1..#upstream_routes_of_i!// HOZ = 1..#buckets_in_planning_horizon!!xAddBck{<o,i>} = new var<CP>{int}[p in POR(o), r in RTE(i)]! = var<CP>{int}(_cp,HOZ);!
... ...
...
Variable CreaCon Example
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 38
//order o!// route r!// POR(o) = set of planned orders for o!// RSC(r) = 1..#resources_of_r!// CPY(r) = 0..#max_capacity_consumable_on_r!!xCsdCpy{<o,r>} = new var<CP>{int}[p in POR(o),w in RSC(r)]! = var<CP>{int}(_cp,CPY(r));!
Constraint PosCng
• At each node (visited once) • Each constraint addresses a
specific requirement or rule – match stock ins and outs – RTF/DDF – safety stocks – route possible Cmes – item SOL/EOL – capacity constraints – yield rate – etc
• Finite domain constraints – arithmeCc, logical – indexing, etc
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 39
...
...
... ...
...
...
...
...
... ...
... ...
...
Constraint PosCng Example
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 40
// order o!// RTF of o rtf!!if (_cfg.solver().enableOrderRTF()) {! if (! _cfg.solver().enableJustInTimeDelivery()) {! _cp.post(xSrlsBck{o} <= rtf,! onBounds);! } else {! _cp.post(xSrlsBck{o} == rtf,! onBounds);! }!}!
... ...
...
Constraint PosCng Example
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 41
// order o!// inventory i!// si = <o,i>!// POR(o) = set of planned orders for o!// RTE(i) = 1..#upstream_routes_of_i!!forall(p in POR(o)) {! _cp.post(xRaddStk{si}[p,route{<o,i>}] == xIaddStk{si}[p],!
! onBounds);! _cp.post((sum(r in xRaddStk{si}.getRange(1)) xAddStk{si}[p,r]) == xIaddStk{si}[p],!
! onBounds);!}!
Search Algorithm
• Different labelling strategy per class of variables – rouCng, allocaCon, stock, quanCty, Cme
• Each class has – a priority
• 0..5 (0 ó disabled)
– a traversal direcCon • backward or forward
– value selecCon heurisCcs • domain-‐specific or not • greedy or exhausCve • randomized or systemaCc • single value or domain splinng • min first or max first
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 42
...
...
...
...
... ...
Search Strategy Example
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 43
...
...
...
...
... ... 1. Rou
Cng + AllocaCo
n
Search Strategy Example
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 44
...
...
...
...
... ...
2. Q
uant
ities
Search Strategy Example
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 45
...
...
...
...
... ...
3. T
ime
Problem DecomposiCon Strategy (Meta-‐Solver)
• Cluster orders using any criteria combinaCon – unsolved orders – same customer – same module-‐out inventory – similar due dates – contenCon on resources – contenCon on materials – etc
• Use best-‐fit model for each class of orders
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 46
Model and Solver ConfiguraCon
• ReconfiguraCon aqer each pass – switch on or off problem features/constraints
• allow lateness, disable aggregate resources, do not fragment sales orders, shorten sales orders, etc
– switch on or off algorithmic opCons • aspect prioriCes, heurisCcs, search type, etc
• Other algorithmic opCons – maximum number of failures allowed – don’t explore non-‐contribuCng subtrees – saCsfy or opCmize – restarts (eg, Large-‐Neighbourhood Search)
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 47
Example of Strategy
• Quickly plan the bulk of the demand in first passes – using efficient but constraining CP model
• eg, no order fragmentaCon, no lateness, no shortening
• Tackle hard orders and perform fine-‐grained planning in subsequent passes – using more flexible model with all features turned on
• MulC-‐criteria opCmizaCon may be carried out – at iteraCon-‐level
• eg, using weighted objecCve funcCon inside CP model – at pass-‐level
• eg, different passes focus on different KPIs
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 48
Contents
• TFT-‐LCD Panels – Manufacturing operaCons – Supply chain planning – SCP as discrete opCmisaCon
• Constraint Programming Approach – MulC-‐pass adaptaCve planning – A generic CP model – Experiments
• Summary and Outlook
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 49
Experiments
• Tests on representaCve dataset – similar structure and size to customer’s instances
• Measures – on-‐Cme delivery (OTD) and fulfillment for orders and quanCCes – assembly resource capacity uClizaCon – run-‐Cme
• Comparison of COMET with exisCng rules system – lazer fragments a sales order's plan, former does not
• over different routes and dates and resources on a route
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 50
Search Strategy
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 51
CLASS PRIORITY
rou&ng 1
resource alloca&on 1
stock 0
quan&ty 2
&me 3
HEURISTIC dom-‐spec all-‐random
greedy-‐min
greedy-‐max
all-‐min all-‐max dicho-‐tomic
rou&ng x
resourcing x
stock x
quan&ty x
&me x
CLASS TRAVERSAL
rou&ng backward
resource alloca&on backward
stock none
quan&ty backward
&me forward
COMET vs. Rule-‐based Planner
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 52
total orders to plan #orders fulfilled #orders on-‐Cme COMET 7102 6658 5588
total quanCCes #quanCCes delivered #quanCCes on Cme COMET 21,909,458 18,550,314 16,411,630
run-‐Cme COMET 764 sec
#orders delivered #orders on Cme Improvement + 9.7% + 6.7%
#quanCCes delivered #quanCCes on Cme Improvement + 2.0% + 3.1%
Contents
• TFT-‐LCD Panels – Manufacturing operaCons – Supply chain planning – SCP as discrete opCmisaCon
• Constraint Programming Approach – MulC-‐pass adaptaCve planning – A generic CP model – Experiments
• Summary and Outlook
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 53
Summary
• DecomposiCon approach based on CP – high-‐fidelity model – solver leverages COMET propagaCon and search capabiliCes – "best-‐fit model" approach through run-‐Cme configurable model
• Results very encouraging – much bezer soluCons than Rules system – acceptable perfs: 10mns / run (30 orders/sec)
• Project – implementaCon enCrely in COMET – 10 man-‐month effort end-‐to-‐end – 30K lines of code – core model 3K lines of code
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 54
Outlook
• Missing features should further improve OTD, fulfillment and resource uClisaCon by tapping in remaining pockets of capacity and exploiCng inventories
5 Oct. 2012 Supply Chain Planning in Consumer Electronics 55
Id Feature 22 Resource pooling 23 MulC-‐resource sales orders 24 MulC-‐bucket sales orders 25 MulC-‐route sales orders 26 Time-‐phased safety stock 27 Stock keeping-‐Cme 28 Policies -‐ delivery, demand priority 29 Policies -‐ cost-‐driven rouCng 30 Policies -‐ resource uClizaCon maximizaCon & smoothing 31 Alternate materials
should improve OTD, fulfillment and resource uClizaCon