Model Based Calibration Optimization Using MachineLearning Algorithms

MathWorks Automotive Conference - 2018Plymouth, MIMay 02, 2018

Shehan Haputhanthri and Shuzhen Liu

May 1, 2018

Electrified Powertrain Engineering (EPE)Ford Motor Company

Outline

Part I : IntroductionPart II : Infrastructure and Process DevelopmentPart III : Example Study

Shehan Haputhanthri and Shuzhen Liu Model Based Calibration Optimization Using Machine Learning Algorithms

Outline

Part I : Introduction

Shehan Haputhanthri and Shuzhen Liu Model Based Calibration Optimization Using Machine Learning Algorithms

Calibration Optimization Efficiency

Manual powertrain calibration optimizationRequires hardware prototypesSubstantial costEngineering (design validation) time

Design Validation (DV) efficiency improvementUse high fidelity CAE powertrain models (Simulink) for initial calibrationOptimized combination of design variablesReduce prototype vehicles and engineering timeExplore designs which would otherwise not be possible

Shehan Haputhanthri and Shuzhen Liu Model Based Calibration Optimization Using Machine Learning Algorithms

Calibration Optimization of Electrified Powertrains

Electrified vehiclesComplex hardware architecture and software controls1000s of design (calibration) variables

Finding the optimum designMillions of design combinationsHigher computation time for high fidelity Simulink SIL modelsSubstantial amount of computing resourcesCluster/ parallel computing for simulationMachine learning (Genetic Algorithm) for optimization100s (even 1000s) of designs to run in parallel

Shehan Haputhanthri and Shuzhen Liu Model Based Calibration Optimization Using Machine Learning Algorithms

Parallalization, Licenses and Scalability

Parallel computingRunning 100s/1000s of designs in parallelNot cost effective with standalone licenses

Parallel/distributed computing licenses for MATLAB & SimulinkParallel Computing Toolbox (PCT)MATLAB Distributed Computing Server (MDCS)

Scalable parallel/distributed computingFinal solution test in workstations (before deployed to HPC)HPC utilization for computation

With varying number of workersScalable from sequential to 100s CPUs in parallel

Shehan Haputhanthri and Shuzhen Liu Model Based Calibration Optimization Using Machine Learning Algorithms

Outline

Part II : Infrastructure and Process Development

Shehan Haputhanthri and Shuzhen Liu Model Based Calibration Optimization Using Machine Learning Algorithms

Distributed Computing Capability

Distributed Computing

HPCFord Internal

Online ClustersPay per Use

Local Cluster (Unused CPUs)MATLAB DCS/Other SW

Shehan Haputhanthri and Shuzhen Liu Model Based Calibration Optimization Using Machine Learning Algorithms

Cost Effective Parallel/Distributed Computing Licenses

Model Based Calibration Optimization (MBCO) teamPurchased 1000 MDCS licensesAdded to existing PCT/MDCS license poolDedicated ”Project Group”For license management

Shehan Haputhanthri and Shuzhen Liu Model Based Calibration Optimization Using Machine Learning Algorithms

Machine Learning Algorithms for OptimizationNeed of machine learning algorithms

All design possibilities can’t be evaluatedWith in the simulation budget and available time

Change calibration variablesOptimize for multi-objectivesSubjected to multi-constraints

Alternatives evaluating (in-progress)No ”all-in-one” solution algorithmMATLAB Global Optimization ToolboxmodeFrontierSeveral other commercial optimization SW packages

Currently modeFrontier is used as the optimization toolCompatibility with non MathWorks SW toolsOff the shelf optimization algorithms

Optimization SW packages and Large Scale Parallelization

”Off the shelf” optimization SW packages - IssuesParallel jobs are launched using internal schedulersIndividual job submitted to clusterUses standalone licenseDo not use PCT and MDCS for MATLAB & SimulinkNot suitable for large scale parallel computing

Final Requirements and Solution

Combining modeFrontier and MATLABMATLAB - Capable of parallelization to any scalemodeFrontier - License issueSimulations to run in HPC

Must use distributed computing licensesPCT and MDCSJobs submitted to HPC from a remote MATLAB Client

New in-house tool and a process developedMATLAB GUI - ModefrontierSCVSPCopyrightedCombines modeFrontier and MATLAB & SimulinkMeets above requirements

Shehan Haputhanthri and Shuzhen Liu Model Based Calibration Optimization Using Machine Learning Algorithms

Challenge - 4 Unconnected Dots

Shehan Haputhanthri and Shuzhen Liu Model Based Calibration Optimization Using Machine Learning Algorithms

Connecting Dots via a MATLAB GUI - ModefrontierSCVSP

1. Genetic algorithm2. Model + cal file2.5 License mgt.3. Job submission4. Results collection5. Results reporting

ModefrontierSCVSP - Standardized Modular Solution

MATLAB GUI - Glue CodeCombines HW, SW and Licenses seamlesslyAcross both Windows and Linux platformsUses PCT/MDCSManages the large scale Simulink simulations

Modular (runs simulations in)In workstation with PCT (limited by physical cores)HPC with MDCS (limited by MDCS licenses)

User friendly GUINo coding/programming requiredEasy to train, learn and apply

Shehan Haputhanthri and Shuzhen Liu Model Based Calibration Optimization Using Machine Learning Algorithms

ModefrontierSCVSP - Standardized Modular SolutionSelect to run in local workstation or HPC

Select the number of cores to run

Shehan Haputhanthri and Shuzhen Liu Model Based Calibration Optimization Using Machine Learning Algorithms

Possibilities With the New Tool

Scalabale parallel/distributed computingSequential computing(1 CPU) to parallel in HPC (1000 CPUs)Click of a button12000 Simulink simulations (4000-6000 hours)Completed in less than 15 hours (approx.)Using 512 simulations in parallel

modeFrontier combined with MATLAB & SimulinkMachine learning algorithmsDesign of Experiments (DoE)Scheduling (Genetic Algorithms)Multi-objectives (+ multi-constraints) optimizationSelected 12000 out of 1000000+ design combinations

Deployed and used by the MBCO teamShehan Haputhanthri and Shuzhen Liu Model Based Calibration Optimization Using Machine Learning Algorithms

Outline

Part III : Example Study

Shehan Haputhanthri and Shuzhen Liu Model Based Calibration Optimization Using Machine Learning Algorithms

Optimizing Vehicle Attributes with Battery Power TableBattery power tableOptimum power - different conditionsf(Driver Pwr, Bat. SOC)Linked attributesFuel Economy/EmissionsPerformance/Regeneration/NVH etc.modeFrontier - machine learningOptimize battery power table*Using Genetic Algorithms (MOGA-II)Evolutionary (generation) algorithmSimulation in HPC512 in parallelPopulation size

*Indirect method used - Non Uniform Rational Basis Spline (NURBS) based*Developed by Ford Model Based Systems Engineering (MBSE) team (@ RIC)

Shehan Haputhanthri and Shuzhen Liu Model Based Calibration Optimization Using Machine Learning Algorithms

Optimizing Vehicle Attributes with Battery Power Table*

Cost function - maximize vehicle attribute(s) mentioned aboveSingle objective optimization**

Design Iteration0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000

% In

crea

se o

f Cos

t Fun

ctio

n V

alue

-20

-15

-10

-5

0

5 "Evolution" of the Cost Function Value

*Exact details not shared due to confidentiality*Not validated with prototypes yet**Unfeasible/error designs excluded

Shehan Haputhanthri and Shuzhen Liu Model Based Calibration Optimization Using Machine Learning Algorithms

Acknowledgments

FordPoyu TsouTengko LienDoug BellCiro SotoGary SkiminAlex AkkermanBahij El FadlDennis Resso

EstecoZhendan XueDaniel SchmidtApurva Ghokal

MathWorksBill MaherNathan CrostyWill Wilson

Shehan Haputhanthri and Shuzhen Liu Model Based Calibration Optimization Using Machine Learning Algorithms

Thank You!!!

Shehan Haputhanthri and Shuzhen Liu Model Based Calibration Optimization Using Machine Learning Algorithms