Neural Network Self-Interference Cancellation in Full-Duplex Systems

The patent applies dynamics by making the self-interference cancellation system adaptive to time-varying channel conditions. The neural network model is trained offline to learn the nonlinear characteristics of the self-interference channel, and the trained model is then applied online to cancel self-interference in real-time varying environments. This dynamic adaptation allows the system to maintain high cancellation effectiveness regardless of channel variations.

The patent employs preliminary action by performing offline training of the neural network model before actual full-duplex operation. During the offline training phase, the system learns the nonlinear characteristics of the self-interference channel using training data. This pre-learning enables the system to quickly and accurately cancel self-interference during online operation without requiring complex real-time adjustments.

The patent replaces complex real-time adaptive signal processing with a pre-trained neural network model. Instead of using traditional mechanical or algorithmic approaches that require continuous real-time computation and adjustment, the system uses a neural network that has already learned the self-interference characteristics during offline training. This substitution significantly reduces online computational complexity while maintaining high communication capacity.

The patent applies preliminary action by performing offline training of the neural network model before actual full-duplex operation. During the offline training phase, the system learns the nonlinear characteristics of the self-interference channel using training data. This pre-learning enables the system to quickly and accurately cancel self-interference during online operation without requiring complex real-time adjustments.

The patent applies segmentation by dividing the self-interference cancellation process into two distinct stages: linear component cancellation and nonlinear component cancellation. The linear component is removed using traditional linear cancellation techniques, while the nonlinear component is handled by a neural network model. This segmentation allows each stage to focus on specific types of interference, improving overall cancellation effectiveness while keeping the system manageable.

The patent uses a composite approach by combining traditional linear cancellation techniques with neural network-based nonlinear cancellation. The system employs both linear filtering and nonlinear neural network processing to comprehensively address different types of self-interference components. This composite technique achieves complete interference removal by leveraging the strengths of both linear and nonlinear methods.