Selecting Local Models :: Gasoline Engine Calibration Case Study (Model-Based Calibration Toolbox™)

Learn more about Model-Based Calibration

Selecting Local Models

You should inspect the local and global models in turn, removing outliers if appropriate and trying different model types, before creating a two stage model. If the fit is good at the local and global levels you have the best chance of creating a two-stage model that accurately predicts the engine behavior. First, inspect the local models.

When you create the models the view switches automatically to the new local model node (PS22) in the model tree, in the All Models pane, and the local model view appears on the right. Make sure the local model node (PS22) is selected before following these steps.
Look through the tests to inspect the fits. Use the Test controls.
To quickly identify problem tests, click RMSE Plots in the toolbar (or View menu).
Inspect tests with high error values in the local model view. You can navigate to a test of interest from the RMSE Explorer by double-clicking a point in the plot to select the test in the Model Browser local model view. Plot all response features in the RMSE Explorer and investigate tests with extreme values.
Consider removing outliers to improve fits if some points are badly distorting the torque spark curve (use right-click or Outliers menu). Pay particular attention to end points. For example, for tests where the majority of points are at higher spark angles than the maximum (at MBT), it can improve the fit to remove some of these long "tails". It can be useful to remove outliers in this region, because there is likely to be knock at spark values much higher than MBT where the engine is less stable. Similarly, as there is no knock in simulation data, points can be collected far in advance of MBT, and it can improve the fit to remove these.
If removing some outliers does not bring MBT within the range of the data points, consider removing the whole test (use the Outliers menu).
After making changes, check the RMSE Explorer plots again for problem tests.
Check all tests before inspecting the global models. Try looking at the Local diagnostics (select from the list in the Diagnostic Statistics pane). Check for high values of Cond(J) (e.g., > 10⁸). High values of this condition indicator can be a sign of numerical instability.