verification validation of integrated supply chain network ......develop credible simulations;...
TRANSCRIPT
Independent Verification & Validation of Integrated Supply‐Chain Network
Simulation and Optimization Models
Soroosh GholamiHessam S. Sarjoughian1
Arizona Center for Integrative Modeling & Simulation
Computer Science & Engineering Dept.
Arizona State University, Tempe, AZ, USA
1 Corresponding Author
Gary GoddingVictor ChangDaniel Peters
Supply Chain Intelligence and Analytics GroupIntel, Chandler, AZ, USA
1
Industrial Case StudiesPresentation without paper
WinterSim, December 8-11, 2013 – Washington, D.C., USA
Outline
• Project Description• Supply‐chain systems• Project objectives • Expected outcomes
• Supply‐chain system simulation specification• Multi‐chain component simulation• Strategic planner• Knowledge interchange broker
• Validation phases • Conclusion and future work
2
Project Overview
GOAL – Validation and verification of supply‐chain network simulation using a flexible, scalable simulation environment for interacting processes/logistics and decision plans that enables development and evaluation of existing/proposed supply‐chain systems
OUTCOMES – 1) V&V for Intel’s Discrete Event Simulation (DES) of production lines; 2) Impact of modeling theory on creditability of integrated DES, Linear Programming (LP) decision plans, and Knowledge Interchange Broker (KIB) simulations.
3
Outcomes and Data Analysis
• Comparison between commercial and DEVS simulation results
• This is reported in by‐product by‐site format
• Results may differ (up to a certain limit)
• Considering the variability and stocahsticity of models
• Different design perspectives• Various optimum solutions (integral
points) in the LP model
• Both outputs can be compared with the plan (strategic planner expectation)• DEVS can perfectly match the expectation (will be shown in the validation section)• Commercial simulation cannot yet match the plan (although the simulation models are validated separately before)
Qua
ntity
Date
DEVS
Commercial
4
Raw MaterialSupplier
Manufacturing
Warehouse (storage)Consumers
Supply Chain System
A system of inventories, processes, warehouses, and transportation mechanisms involved in producing and moving a product from raw material to finished goods and from production facilities to end customers
5
Supply Chain Simulation Issues & Considerations Stochasticity: production times and shipping times are inherently stochastic
Scale: the size of the optimization problem radically grows as the number of facilities, products, and transportation routes approach actual scales and complexity
Abstraction: accurate processes/logistics model specifications are necessary to develop credible simulations; interactions between simulation and optimization models must be separately modeled as in KIB to have flexible, integrated DES and LP simulation models
Impact: credibility of decision plans depends on credible simulation as well as credible interaction models.
6
A View of an Integrated Supply‐Chain Simulation System
ControlStrategy
(Linear Program)
Supply Chain Process (Discrete Event)
Knowledge Interchange
Broker
Fabrication
Assembly Test
Warehouse
CustomersFAB Supply
Customer Demand
Current Stock
Penalty Costs
7
The universal language
Language of system 1 (DEVS)
Language of the KIB
Language of system 2 (LP)
Poly‐Formalism
System Architecture
8
Network Generator
Supply‐Chain Network
Decision Connector
LP/CPLEX
Optim
izer
W1
BOM1
BOMn
DP1
BOM2
W2
BOM1
BOMn
BOM2
DP2
DP3
DP4
DP5
DP6
…
CTRL_INPUT_PORT
Structure Inform
ation
Inpu
t Data
DatasetDemand/Supply Information
Mod
el Con
struction
KnowledgeInterchange Broker
(KIB)
Adding Flow Information to the graph
Invoke Solver
Flow Information
Output Data
Release Commands
Partial View of the Dataset
Intransit_tptID
source
destination
default_tpt
distribution
Param 1
Param 2
Param 3
Param 4
Param 5
Site
site_name
site_type
longitude
latitude
Products
product_name
stage
Process_tptID
site
product
default_tpt
distribution
Param 1
Param 2
Param 3
Param 4
Param 5
BOM
bom_id
input
output
pct
ATM_Start_Schedule
bom_id
product
site
pln_date
pln_year
pln_month
pln_day
pln_quantity
CDP_Start_Schedule
bom_id
product
site
pln_date
pln_year
pln_month
pln_day
pln_quantity
9
Dataset Content ‐ DES
• Information coming from the dataset• Shipping times• Initial inventory• Demand at VMIs and CWs (for a full year; X)• Processing times (X different configurations)• Sites (X)• Shipping elements (X)• Products (X kinds) • Processing configurations (X individual routes)
• The model resulting from these record contains:• 620 atomic/coupled models• 1000+ couplings • …
10
LP Model
Process
W1
W2
Die Product1
Die Product 2
Die Product 3
Die Product 4
Die Product 5
Die Product 6
W1
BOM1
BOMn
DP1
BOM2
W2
BOM1
BOMn
BOM2
DP2
DP3
DP4
DP5
DP6
These connections have yields specified to represent the BOM splitThese connections have specified lead times (TPT)
Throughput time are supplied by the simulation model
We can hold inventory here, ahead of manufacturing process
We can hold inventory here, ahead of shipping
11
Loading Data from Simulation
• Coming from CTRL_OUTPUT_PORT (Separate initialization from run‐time simexec.)
• BoH information for inventories and processes (provides data for LP)
• Data generated inside Decision Connector• Flow information
• Demand: based on records in the dataset• Supply: FAB’s output
• Structure information• Shipping• BOM• …
In BOM
Out1
Out2
Out3
X1%
X2%
Xn%
100%
12
Commercial DES Simulator
• Systems are constructed using pre‐specified components:• Queues • Servers• Links• …
• No mathematical basis• Uses object‐oriented software concepts and methods • Java utility classes as data structures and Java methods for receiving/processing/sending actions
• Graphical visualization of the simulation
13
DEVS‐Suite Simulator
• Constructed using Parallel DEVS formalism
• Atomic and coupled models follow strict modularity; hierarchically models are closed under coupling
• Java methods for receiving/processing/sending messages:• External transition function• Output function• Internal transition function
• Graphical visualization of the simulation
14
Comparison of Commercial and DEVS‐Suite Simulators• DEVS‐Suite has a theoretical foundation• Both DEVS‐Suite and the commercial simulator follow object‐orientation design and implementation.
• There is one‐to‐one relationship between formal models and their implementation in DEVS‐Suite
• Both simulators are capable of other kinds of simulation (such as agent‐based) in addition to DES (however not used for this project)
• Fast and easy modeling of small systems in the commercial simulator• DEVS‐Suite is open‐source; commercial simulator is not
15
Validation (1)
• Phase 1: Single‐chain validation • A single chain model were designed to test simulation models • Simulation models for facilities and shipping components were tested with pulse and slope signals from a devised dataset
• We expected to see the same patterns in product actual output log (DEVS and commercial simulation)
• All processing and shipping times were set as deterministic
Transportation
16
FAB Process Assembly CW VMI
Phase 1 in DEVS‐Suite
17
Validation (2)
QUAN
TITY (O
F LO
TS)
DATE
DEVS Expected
QUAN
TITY (O
F LO
TS)
DATE
DEVS Expected
• Phase 2: Historic data validation • We expected to see a rough consistency between the actual output and historic delivery• It phases out LP and KIB validation• Plots are generated in a by‐product by‐site basis• Stochasticity is imported from the dataset for processing/shipping times
18
Comparison (released vs. expected) for FG10 Comparison (released vs. expected) for FG9Product A Product B
Validation (3)
• The complete model with KIB, LP, and simulation models• KIB is validated by comparing results between KIB‐based simulation and non‐KIB‐based ones• The totality of the simulation is validated (system‐wide validation) by comparing actual output log with the expected ones
• By‐product by‐site plots are created as a result of this simulation• The complete model with
• 620 atomic/coupled models• 1,000+ couplings• X data record (demand, routes, …)• Stochastic processing/waiting times • X set of products (X kinds)• …
19
Validation Phases (summary)
20
Stochastic proc. Times
Multi‐chain model
Historicrelease
commandsDatabase LP KIB Input data
Phase 1 (single chain) Pulse and step functions
Phase 2(historic data)
Network topology, processing times, initialization, and release commands from the dataset
Phase 3‐1 (KIB validation)
1. Network topology, processing times, and initialization from the dataset.2. KIB transformations form configuration XML files
Phase 3‐2 (System‐wide) The output of the simulation against the
expected output
Quantity
(released lots)
Day Index
cumulative no KIB 1 cumulative no KIB 2
Results (1) – no KIB
• All processing/shipping times were changed to deterministic times
• Two experiments without the KIB
• Expected to see reasonably close results (not identical)
• Since LP is not under our control and multiple solutions may exist, two instances may have different results
Quantity
(released lots)
Day Index
no KIB 1 no KIB 2
Results (2) – with KIB
• All processing/shipping times were changed to deterministic times
• Expected to see reasonably close results (not identical)
• Since LP is not under our control and multiple solutions may exist, two instances may have different results
Quantity
(released lots)
Day Index
With KIB 1 with KIB 2
Quantity
(released lots)
Day Index
cumulative KIB 1 cumulative KIB 2
Results (3) – no KIB vs. with KIB
• One experiment with the KIB and the other without it
• Similar to the previous experiments, the results may have slight differences because of the unpredictability in the LP side
Quantity
(released lots)
Day Index
with KIB no KIB
Quantity
(released lots)
Day Index
cumulative KIB cumulative no KIB
Results (4)
• System‐wide validation: • The second plot is reporting the difference between the plan (LP results) and the actual outputs (reported in the first plot)
• The threshold of acceptable difference is set by domain experts and based on the current state of the system
Quantity
(of lots)
Day Index
Quantity
(of lots)
Day Index
Expected Actual
Simulation Platform Spec
• Software• Java 7 – 64bit • Windows 7 – 64bit• IBM ILOG CPLEX Optimization Studio 12.5• DEVS‐Suite 2.1.0• Microsoft SQL Server 2012 Developer version
• Hardware• Intel Core 2 Duo – 2.9GHz• 8GB of physical memory (DDR3)
Maximum memory usage: 5.8 GB
25
Performance Analysis
* Simulation time measures the duration of simulation for one week * Each solves takes 114 seconds on average
Minimum Simulation Time (for one week): 8.61 secMaximum Simulation Time (for one week): 62.29 secAVG Simulation Time (for one week): 28.30 secTotal simulation time (59 weeks) 131.8 minutes
26
Time Analysis
27
0.0010.0020.0030.0040.0050.0060.0070.0080.0090.00
Maximum Weekly Simulation Time AVG Weekly Simulation Time
Second
s
5 ‐RUN TIME ANALYSIS
High Low AVG
7.50
7.75
8.00
8.25
8.50
8.75
9.00
9.25
9.50
Minimum Weekly Simulation TimeSecond
s
5 ‐RUN TIME ANALYSIS
High Low AVG124
126
128
130
132
134
136
138
Total simulation time
Minutes
5 ‐RUN TIME ANALYSIS
High Low AVG
High Low AVG
Minimum Weekly Simulation Time 9.28(s) 8.13(s) 8.61(s)
Maximum Weekly Simulation Time 77.28(s) 50.26(s) 62.29(s)
AVG Weekly Simulation Time 33.13(s) 24.69(s) 28.30(s)
Total 59‐week simulation time 137(m) 129(m) 131.8(m)
Conclusion & Future Work
• Developed a simulation platform by integrating discrete event simulation, strategic control, and knowledge interchange broker
• Validated Intel’s commercial simulation model • Fed data for data analysis/mining team
Future Work• Simulating company‐level operation• Integration with tools where the simulation/optimization data can be mined for information aiding control and operation
28