Model-Based Calibration Toolbox

Gasoline Case Study

This example shows how to automatically generate an mbcmodel project for the gasoline case study using the command-line tools in Model-Based Calibration Toolbox™ .

Requires DIVCP_Main_DoE_Data.xls from mbctraining folder.

Create a New mbcmodel Project

project = mbcmodel.CreateProject;

Load Data into Project

  • Group data into tests and add some filters.

  • Add filters to remove bad data.

  • Remove tests which do not have sufficient data to fit local models.

datafile = fullfile( mbcpath, 'mbctraining', 'DIVCP_Main_DoE_Data.xls' );
data = CreateData( project, datafile );

BeginEdit( data );

% Group data by test number GDOECT.
DefineTestGroups( data, 'GDOECT', 0.5, 'GDOECT', false );
% Get rid of data which is probably unstable.
AddFilter( data, 'RESIDFRAC < 35' );
AddFilter( data, 'AFR > 14.25' );
% Get rid of the tests that are too small.
AddTestFilter( data, 'length(BTQ) > 4' );

% Commit the changes to the project.
CommitEdit( data );

Build Test Plan

Define Inputs for test plan.

LocalInputs = mbcmodel.modelinput('Symbol','S',...
    'Name','SPARK',...
    'Range',[0 50]);
GlobalInputs = mbcmodel.modelinput('Symbol',{'N','L','ICP','ECP'},...
    'Name',{'SPEED','LOAD','INT_ADV','EXH_RET'},...
    'Range',{[500 6000],[0.0679    0.9502],[-5 50],[-5 50]});
% Create test plan.
testplan = CreateTestplan( project, {LocalInputs,GlobalInputs} );
% Attach data to the test plan.
AttachData( testplan, data );

Build Boundary Models

Create a global boundary model. CreateBoundary does not add the boundary model to the tree.

B = CreateBoundary(testplan.Boundary.Global,'Star-shaped');
% Add boundary model to the test plan. The boundary model is fitted when it
% is added to the boundary model tree. The boundary model is included in
% the best boundary model for the tree by default.
% All inputs are used in the boundary model by default.
B = Add(testplan.Boundary.Global,B);

% Now make a boundary model using only speed and load and add to the
% boundary tree.
B.ActiveInputs = [true true false false];
B = Add(testplan.Boundary.Global,B);
% Look at the global boundary tree.
testplan.Boundary.Global
ans = 

  Tree with properties:

         Data: [189x4 double]
       Models: {[1x1 mbcboundary.Model]  [1x1 mbcboundary.Model]}
    BestModel: [1x1 mbcboundary.Boolean]
       InBest: [1 1]
     TestPlan: [1x1 mbcmodel.testplan]

Build Responses

Build response models for torque, exhaust temperature and residual fraction * Use a local polynomial spline model for torque. * Use a local polynomial model with datum for exhaust temperature and residual fraction

LocalTorque  = mbcmodel.CreateModel('Local Polynomial Spline',1);
LocalTorque.Properties.LowOrder = 2;
% Use the default global model.
GlobalModel = testplan.DefaultModels{2};
CreateResponse(testplan,'BTQ',LocalTorque,GlobalModel,'Maximum');
%make exhaust temperature and residual fraction models
LocalPoly  = mbcmodel.CreateModel('Local Polynomial with Datum',1);
CreateResponse(testplan,'EXTEMP',LocalPoly,GlobalModel,'Linked');
CreateResponse(testplan,'RESIDFRAC',LocalPoly,GlobalModel,'Linked');

Remove Local Outliers

Remove data if abs(studentized residuals) > 3. Note that a different process was used in the project Gasoline_project to decide which outliers to remove.

TQ_response = testplan.Responses(1);
numTests = TQ_response.NumberOfTests;
LocalBTQ = TQ_response.LocalResponses;
for tn = 1:numTests
    % Find observations with studentized residuals greater than 3
    studentRes = DiagnosticStatistics( LocalBTQ, tn, 'Studentized residuals' );
    potentialOut  = abs( studentRes )> 3;
    if any(potentialOut)
        % Don't update response feature models until end of loop
        RemoveOutliersForTest( LocalBTQ, tn, potentialOut , false);
    end
    % get local model for test and look at summary statistics
    mdl = ModelForTest(LocalBTQ,tn);
    if ~strcmp(mdl.Status,'Not fitted')
        LocalStats = SummaryStatistics(mdl);
    end
end

Update response features.

UpdateResponseFeatures(LocalBTQ);

Remove Points Where MBT<0 or MBT>60

knot = LocalBTQ.ResponseFeature(1);
PointsToRemove = knot.DoubleResponseData<0 | knot.DoubleResponseData>60;
knot.RemoveOutliers(PointsToRemove);

Create Alternative Response Feature Models

Make a list of candidate models and select the best based on AICc.

  • Quadratic

  • Cubic

  • RBF with a range of centers

  • Polynomial-RBF with a range of centers

Get the base model. You use this to create the other models.

rf = LocalBTQ.ResponseFeature(1);
BaseModel = rf(1).Model;

Make a quadratic model that uses Minimize PRESS to fit, and add it to the list.

m = BaseModel.CreateModel('Polynomial');
m.Properties.Order = [2 2 2 2];
m.FitAlgorithm = 'Minimize PRESS';
mlist = {m};

Make a cubic model and add it to the list.

m.Properties.Order = [3 3 3 3];
m.Properties.InteractionOrder = 2;
mlist{2} = m;

Make RBF models with a range of centers. The maximum number of centers is set in the center selection algorithm.

m = BaseModel.CreateModel('RBF');
Centers = [50 80];
Start = length(mlist);
mlist = [mlist cell(size(Centers))];
for i = 1:length(Centers)
    m.FitAlgorithm.WidthAlgorithm.NestedFitAlgorithm.CenterSelectionAlg.MaxCenters = Centers(i);
    mlist{Start+i} = m;
end

Make Polynomial-RBF models with a range of centers.

m = BaseModel.CreateModel('Polynomial-RBF');
m.Properties.Order = [2 2 2 2];
Start = length(mlist);
mlist = [mlist cell(size(Centers))];
for i = 1:length(Centers)
    % Maximum number of centers is set in the nested fit algorithm
    m.FitAlgorithm.WidthAlgorithm.NestedFitAlgorithm.MaxCenters = Centers(i);
    mlist{Start+i} = m;
end

Make alternative models for each response feature and select the best model based on AICc.

criteria = 'AICc';
CreateAlternativeModels( LocalBTQ, mlist, criteria );

Alter Response Feature Models

Get the alternative responses for knot and alter models using stepwise regression.

knot = LocalBTQ.ResponseFeature(1);
AltResponse = knot.AlternativeResponses(1);

Get the stepwise statistics.

knot_model = AltResponse.Model;
[stepwise_stats,knot_model] = StepwiseRegression(knot_model);

Use PRESS to find the best term to change, and toggle the stepwise status of the term with this index.

[bestPRESS, ind] = min(stepwise_stats(:,4));
[stepwise_stats,knot_model] = StepwiseRegression(knot_model, ind);

Get a VIF statistic.

VIF = MultipleVIF(knot_model)
VIF =

    3.1290
    1.2918
    1.6841
    1.1832
    1.3230
    2.6617
    1.6603
    1.3306
    1.2856
    1.4317
    2.6550

Get the RMSE.

RMSE = SummaryStatistics(knot_model, 'RMSE')
RMSE =

    5.1578

Change the model to a Polynomial-RBF with a maximum of 10 centers.

new_knot_model = knot_model.CreateModel('Polynomial-RBF');
new_knot_model.Properties.Order = [1 1 1 1];
new_knot_model.FitAlgorithm.WidthAlgorithm.NestedFitAlgorithm.MaxCenters = 10;
% Fit the model with current data.
[S,new_knot_model] = new_knot_model.Fit;

If there were no problems with the changes then update the response, otherwise you will continue to use the original model.

if strcmp(new_knot_model.Status,'Fitted')
    new_RMSE = SummaryStatistics(new_knot_model,'RMSE')
    % Update the response with the new model.
    UpdateResponse(new_knot_model);
end
new_RMSE =

    3.5712

Make the Two-stage Model for Torque

doMLE = true;
MakeHierarchicalResponse( LocalBTQ, doMLE );
% Look at the Local and Two-Stage RMSE.
BTQ_RMSE = SummaryStatistics( LocalBTQ, {'Local RMSE', 'Two-Stage RMSE'} )
BTQ_RMSE =

    0.8319    4.6626

Plot the Two-stage Model of Torque Against Spark

Plot the 5th test

testToPlot = 5;
BTQInputData = TQ_response.DoubleInputData(testToPlot);
BTQResponseData = TQ_response.DoubleResponseData(testToPlot);
BTQPredictedValue = TQ_response.PredictedValue( BTQInputData );
fig = figure;
plot( BTQInputData(:,1), BTQResponseData, 'o', BTQInputData(:,1), BTQPredictedValue, '-' );
xlabel( 'spark' );
ylabel( 'torque' );
title( 'Test 5' );
grid on

Build Other Responses

Build models for exhaust temperature and residual fraction.

  • Copy outliers from torque model

  • Make alternative models for each response feature

  • Make two-stage model without MLE

EXTEMP Response

EXTEMP = testplan.Responses(2).LocalResponses(1);
EXTEMP.RemoveOutliers(OutlierIndices(LocalBTQ));
CreateAlternativeModels( EXTEMP,mlist, criteria );
MakeHierarchicalResponse( EXTEMP, false );
EXTEMP_RMSE = SummaryStatistics( EXTEMP, {'Local RMSE', 'Two-Stage RMSE'} )
EXTEMP_RMSE =

   10.5648   29.2186

RESIDFRAC Response

RESIDFRAC = testplan.Responses(3).LocalResponses(1);
RESIDFRAC.RemoveOutliers(OutlierIndices(LocalBTQ));
CreateAlternativeModels( RESIDFRAC,mlist, criteria );
ok = MakeHierarchicalResponse( RESIDFRAC, false );
RESIDFRAC_RMSE = SummaryStatistics( RESIDFRAC, {'Local RMSE', 'Two-Stage RMSE'} )

if isgraphics(fig)
    % delete figure made during example
    delete(fig)
end
RESIDFRAC_RMSE =

    0.0824    0.5694