How can AI estimate physical quantities that are difficult, costly or impractical to measure directly?

This session explores how Transformer architectures and attention mechanisms can transform multivariate time-series data from vehicle CAN signals into virtual sensor estimates. By capturing temporal dependencies and relationships between signals, attention-based models can identify which measurements, and which moments in their history, are most relevant to estimating a target physical quantity.

We will explore the journey from raw CAN data to virtual sensors: signal synchronization and preprocessing, temporal feature learning, model training, and validation under different operating conditions.

Beyond model accuracy, the session will highlight the importance of physical understanding, robustness, generalization and real-world constraints when applying advanced AI architectures to vehicle dynamics and industrial applications.