Vehicle Transmission State Prediction With Physics-Constrained AI

The patent combines a data-based generative model (first model) with a physics-based model (second model) into a hybrid system. The generative model learns patterns from training data to predict intermediate transmission states, while the physics-based model independently calculates physically viable states based on vehicle dynamics equations. These two models are merged through a combination layer that integrates their outputs, ensuring predictions are both data-driven and physically valid. This resolves the contradiction by maintaining adaptability from the generative model while guaranteeing reliability through physics constraints.

The patent transforms the output of the generative model by applying physics-based parameter constraints. The physics-based model calculates transmission states using fundamental physical parameters (vehicle mass, gravitational acceleration, slope, resistance forces) and uses these to adjust and validate the generative model's predictions. This parameter transformation ensures that even when extrapolating to unseen scenarios, the predicted transmission states remain physically viable, thus resolving the reliability issue while preserving the generative model's adaptability.

The patent merges the strengths of physics-based modeling with data-based learning by creating a hybrid architecture. The physics-based model (second model) provides reliable physical constraints and deterministic calculations, while the generative model (first model) contributes adaptability by learning patterns from training data including unobservable driver behavior and road conditions. The combination layer integrates both approaches, allowing the system to satisfy physical constraints while adapting to new scenarios through data-driven insights.

The patent introduces an intermediate state representation that acts as a mediator between the generative model and physics-based model. The generative model predicts intermediate transmission states that capture learned patterns from data, while the physics-based model validates and adjusts these intermediate states to ensure physical viability. This intermediary layer allows the system to benefit from both data adaptability and physics reliability without requiring the physics model to directly learn from data or the generative model to guarantee physical constraints.