Driver State Classification Using Calibration-Free Bio-Signals

The patent uses Generative Adversarial Networks (GANs) to generate synthetic bio-signal data that copies the statistical characteristics and variability patterns of real driver bio-signals. This synthetic data serves as a substitute for actual calibration data, allowing the system to learn driver state classification without requiring real-time calibration from each driver. The generated data preserves the inter-variability and intra-variability properties of genuine bio-signals while eliminating the need for complex calibration procedures.

The system performs preliminary learning using synthetic bio-signal data generated before deployment. By pre-training the classification model with artificially generated but realistic driver bio-signal patterns, the system prepares itself in advance to handle various driver states without requiring on-site calibration. This preliminary action with synthetic data eliminates the time-consuming calibration step that would otherwise be required before actual use.

The system generates synthetic calibration data that replicates the temporal and statistical characteristics of real driver bio-signals. By using this copied synthetic data for training, the system achieves calibration-like performance without requiring actual drivers to undergo time-consuming calibration sessions. The synthetic data captures the essence of driver variability patterns without the time cost.

The system performs self-calibration by automatically adapting to new drivers without requiring manual calibration sessions. The classification model uses the generated synthetic data to self-adjust and accommodate individual driver characteristics, eliminating the need for drivers to spend 10-30 minutes in calibration sessions. The system serves itself by generating its own training data needs.

The patent employs Generative Adversarial Networks to create synthetic copies of real driver bio-signal data that preserve the statistical properties, variability patterns, and temporal dynamics of genuine measurements. This synthetic data copying approach provides sufficient training material for accurate model learning without requiring the complex logistics of collecting, storing, and managing large volumes of real driver data from multiple sources.

The GAN-based synthetic data generation acts as an intermediary between the need for accurate training data and the complexity of real data collection. Instead of directly collecting and processing real driver bio-signals, the system uses the synthetic data generator as a mediator that translates simple input parameters into realistic training datasets, simplifying the overall data acquisition process while maintaining learning accuracy.