Main Content

Generate a Deep Learning SI Engine Model

If you have the Deep Learning Toolbox™ and Statistics and Machine Learning Toolbox™, you can generate a dynamic deep learning spark-ignition (SI) engine model to use for hardware-in-the-loop (HIL) testing, powertrain control, diagnostic, and estimator algorithm design. For example, fit a deep learning model to measured engine-out transient emissions data and use it for aftertreatment control and diagnostic algorithm development. The deep learning SI engine models the dynamic engine behavior from measured laboratory data or a high-fidelity engine model.

To train the deep learning SI engine model, Powertrain Blockset™ uses this SI engine data.

Input DataOutput Data

  • Engine speed

  • Intake manifold gas pressure

  • Wastegate area percent

  • Intake cam phaser angle

  • Exhaust cam phaser angle

  • Spark retard from nominal

  • Lambda

  • Torque

  • Airflow - Intake air mass flow

  • Exhaust gas temperature

  • Throttle inlet pressure

To generate the deep learning engine model, follow these steps.

  1. If it is not already opened, open the reference application.

  2. Double-click Generate Deep Learning Engine Model. Generating the model can take several hours.

    By default, to train the deep learning engine model, the reference application generates design of experiment (DoE) response data from the SI Core Engine block. Alternatively, you can use engine data generated by Powertrain Blockset from Gamma Technologies LLC engine models or other high-fidelity engine models.

    • View the training progress window to see the iteration or stop the training.

      Powertrain Blockset uses half the data to train the model and half to test the model.

  3. After you generate the deep learning SI model, view the results.

    • Review the pairwise overlay of test versus training dataset engine steady-state targets.

    • For each engine input, a plot displays the input signals that the deep learning model uses to train itself to match the output responses. The transient inputs stabilize to match the steady-state targets shown in the overlay plot. Instead of using throttle position as an input to the deep learning model, the model uses the measured intake manifold pressure response. The software uses a physical model to compute the intake manifold pressure and provide it to the deep learning model.

    • For the four engine outputs, a plot displays the SI engine deep learning model (predicted – red) and the test data (test – blue).

    • For the four engine outputs, a histogram displays the SI engine deep learning model error distribution between the responses predicted by the deep learning model and the measured test responses of the engine.

    • The Simulation Data Inspector displays the results of an engine performance test of the trained SI engine deep learning model over a grid of commanded engine speed and engine torque operating points. Use the commanded versus measured torque response comparisons to assess the deep learning model suitability for a vehicle model.

    • A mesh plot displays the quasi-steady torque response of the deep learning SI engine model. Use this plot for a qualitative visual understanding of whether the engine behavior is consistent with the real physical engine in steady-state operation.

  4. You can use the deep learning SI model, SiDLEngine, as an engine plant model variant in the conventional vehicle and hybrid electric vehicle (HEV) reference applications. For example, in the conventional vehicle reference application, on the Modeling tab, in the Design section, open the Variant Manager. Navigate to Passenger Car > Engine. Right-click to set SiDLEngine as the active choice.

  5. To fit your own deep learning SI engine model or adjust the deep learning training settings, use the FitSiEngineLSTM.m script in the reference application project folder.

See Also


Related Topics