Efficient Vehicle State Estimation on CAN-Bus/IMU Time-Series: Leveraging Convolutional Methodologies for Real-Time Maneuver Recognition

This thesis presents an industrial data-driven AI solution for real-time vehicle maneuver recognition and state estimation using continuous CAN-bus and IMU telemetry. Traditional rule-based systems are limited by high-frequency sensor noise, causing severe label jitter and unstable state transitions. To resolve this, this research classifies telemetry into six driving phases, strictly prioritizing temporal consistency and physical continuity. The pipeline features three primary stages. To overcome the scarcity of manually annotated data, a data augmentation ensemble pairs a MultiRocket classifier with smoothing techniques (majority voting and Hidden Markov Models (HMM)) to generate robust pseudo-labels for unlabelled telemetry. Optimal sequences are selected using the Silhouette Index with Transitions Penalty, a custom metric ensuring feature-space cohesion and physically feasible temporal continuity. During validation against the reference set of supervised data, the MultiRocket baseline shows competitive accuracy but severe vulnerability to classification jitter. Consequently, an InceptionTime Convolutional Neural Network is employed in its standard and "informed" variant, that injects domain-specific physical priors to mathematically mask dynamically impossible state transitions, vastly improving structural cohesion and suppressing erratic output jitter. Finally, the testing phase evaluates the model's generalization capabilities on completely unseen tracks, such as the Nürburgring, and novel vehicular configurations. The informed InceptionTime model successfully mitigates contradictory predictions in these unencountered environments. Furthermore, moving toward the real-time application, a structurally reduced network version maintains exceptional predictive clustering quality and resists overfitting, proving the framework's viability for integration into resource-constrained, embedded Electronic Control Units.