Prediction of frequency-domain data for EMC of an electric vehicle powertrain with Gaussian processes and ANN-based embedding

For the electromagnetic compatibility (EMC) of an electric vehicle (EV), investigation of the frequency-domain behavior of the powertrain and its components is critical to ensure compliant, robust and cost-effective design. Time-consuming and expensive measurements and simulations are common to obtain the frequency-domain data, like EMC filter transfer functions and emissions. Machine learning (ML) techniques can be used to build surrogate models with very fast evaluation speed, though these models lose accuracy when training data is scarce regarding design-parameter variations. This issue is aggravated by the fact that the frequency axis is typically densely sampled which can quickly overload data-efficient ML-techniques. Here, an embedding of the frequency-domain data using artificial neural nets in an autoencoder configuration is used to reduce the number of data points that need to be handled. This allows for the usage of the data-efficient Gaussian process regression algorithm to perform the predictions over the EMC design space on the embedded data. The method is demonstrated with two EMC tasks showing high predictive accuracy and efficiency. First, the prediction of the insertion loss of an EMC filter for the high-voltage powertrain of an EV with data obtained from 3D simulation. Second, the prediction of conducted emissions of the high-voltage powertrain of an EV with data obtained from a transient simulation.

Content

Subjects

Electromagnetic Compatibility (EMC)

Machine Learning

Electric Vehicles

Conducted Emissions

Autoencoder

Gaussian Process Regression

Passive Filters

DDC Class

537: Electricity and Electronics

Funding References

Funding(s)

Thermo-Electrical EMC Filter Design for Electric Vehicles