Physiological State Recognition Model Training for Individual Signal Variation

The patent uses a generator to create synthetic physiological signals that copy the characteristics of real physiological signals. These synthetic signals are then used to train the recognition model alongside real signals, effectively increasing the quantity of training data while maintaining the essential characteristics needed for accurate recognition. The generator creates copies that capture the underlying patterns without being constrained by the limited number of actual collected signals.

The patent transforms real physiological signals into synthetic signals by changing parameters such as signal characteristics, noise patterns, and individual variations. The generator modifies these parameters to create varied synthetic signals that represent different physiological states, allowing the model to learn from a diverse dataset without requiring actual diverse biological data from multiple subjects.

Instead of creating complex models to account for individualized differences, the patent uses the generator to create synthetic signals that inherently include various individual characteristics. By training on these diverse synthetic copies, the model learns to recognize patterns that are robust across different individuals without requiring complex individualization algorithms. The diversity is embedded in the data generation process rather than in the model architecture.

The patent segments the training process into two parts: a generator that creates synthetic physiological signals with varied individual characteristics, and a recognition model that learns from these segmented examples. This segmentation allows the complexity of individualized differences to be handled in the data generation stage rather than in the recognition model, simplifying the overall system while maintaining high accuracy.