Model Monitoring (ASCMO-STATIC)

Data-based models can show poor extrapolation behavior. Predictions based on input data from regions that were not covered during training must therefore be treated with caution.

The model prediction itself does not indicate whether it is based on known or unknown input data. For dynamic models, not only the current inputs but also the current state of the dynamics must be considered.

Model Monitoring is used to detect whether the main model operates outside its valid operating domain. For this purpose, a second, small model is created. This Model Monitoring model monitors the main model during runtime.

The inputs of the Model Monitoring model can consist of the external inputs of the main model and the internal states of the main model. During runtime, the Model Monitoring model evaluates whether the current operating state is consistent with the data distribution seen during training.

To create the training data for the Model Monitoring model, the main model is executed on its training data. The generated data is used to train a probability density function over the extended feature space.

If the output of the Model Monitoring model falls below a defined threshold, an anomaly is detected. This indicates that the main model operates with high certainty outside its known data region.

The trained anomaly detection model can be exported to Embedded AI Coder together with the output models. In Embedded AI Coder, the code for the anomaly detection model and the selected output models can be generated for execution on the electronic control unit, see Model Export to Embedded AI Coder.

In ASCMO-STATIC, Model Monitoring can be used by adding an anomaly detection output and selecting the Gaussian Mixture Model for this output.

To add Model Monitoring

Select In/Outputs > Add Output for Anomaly Detection.
Open the model configuration for the anomaly detection channel.
Select Gaussian Mixture Model as model type.
Configure the model parameters.
Start the model training.
The anomaly detection channel is trained with a Gaussian Mixture Model.