Wireless Signal DNN Training Across Dynamic Channel Environments

The patent segments the training channel environment into multiple discrete environments (e.g., line-of-sight, non-line-of-sight, urban, suburban, rural) and trains separate DNN models for each environment. This segmentation allows the system to apply specific pre-trained models based on the current channel conditions, avoiding the need for a single complex model trained on all environments, thus reducing overall system complexity while maintaining high adaptability.

The patent performs preliminary training of DNN models in advance for each specific channel environment before actual communication occurs. These pre-trained models are stored and can be quickly selected and applied when needed, eliminating the need for real-time training during communication. This preliminary action reduces the computational burden and system complexity during operation while maintaining strong learning capability.

The patent changes the parameter of training channel environment from a single static environment to multiple discrete environments with different characteristics. By training separate DNN models for each channel environment type and selecting the appropriate model based on current conditions, the system achieves both high training precision for each specific environment and adaptability to dynamic channel changes.

The patent implements a feedback mechanism where the system monitors current channel conditions and selects the appropriate pre-trained DNN model based on the detected environment. This feedback loop ensures that the system continuously adapts to changing channel conditions by selecting the most suitable model, thereby achieving both precision and adaptability.

The patent segments the continuous channel environment into a finite number of discrete environment types (e.g., 4-10 representative environments). This segmentation allows the system to handle channel diversity with a manageable number of pre-trained models, reducing training overhead and processing time compared to handling all possible channel variations continuously.

The patent performs all necessary training in advance for each discrete channel environment before actual communication. During operation, the system only needs to detect the current environment type and select the corresponding pre-trained model, which is a quick operation. This preliminary action eliminates iterative training during communication, significantly reducing training overhead and processing time while maintaining support for multiple channel environments.

The patent segments the training process by environment type and trains separate DNN models for transmitter and receiver for each environment. This segmentation allows each model to be trained on environment-specific data, reducing the amount of training data needed per model compared to training a single universal model on all environments. The segmented approach maintains high communication performance while reducing overall training data requirements.

The patent performs preliminary training of Tx and Rx DNN models for each channel environment in advance. These pre-trained models are then stored and reused during communication without requiring additional training data. This preliminary action eliminates the need for continuous training during operation, significantly reducing the effective training data requirements while maintaining high communication performance through the use of pre-trained models.