AI Neural Network Positioning from Channel Responses in NLOS

The patent replaces traditional geometric positioning methods with AI/ML-based positioning. The neural network model processes channel response data to estimate UE position, substituting the mechanical/geometric calculation approach with an intelligent system that can handle NLOS conditions more effectively by learning from training data the relationship between channel characteristics and position.

The patent changes the input parameters for positioning from direct geometric measurements to channel response parameters. By using channel state information, reference signal received power, and other channel-related parameters as inputs to the neural network, the system can infer position indirectly through the radio channel characteristics, which remains informative even in NLOS conditions.

The patent performs preliminary training of the neural network model during off-line periods using extensive training data. The model learns the mapping between channel responses and positions in advance, so during actual positioning operations, only the processed channel response data needs to be transmitted, not the entire training dataset. This separates the data-intensive training phase from the lightweight inference phase.

The patent uses a trained neural network model that captures the essential relationships from extensive training data. Instead of transmitting or storing large amounts of raw training data, the system creates a compressed representation in the form of the trained model parameters, which can be deployed to UEs or LMF for efficient inference with minimal data overhead.

The patent segments the AI positioning system into different functional components that can be distributed between UE and network. The neural network model can be split such that feature extraction is performed at the UE using received channel data, while the inference or model updating can be performed at the network side. This allows accuracy enhancement while managing UE processing complexity through functional decomposition.

The patent introduces the neural network model as an intermediary between raw channel measurements and position estimation. Instead of requiring complex geometric calculations or signal processing at the UE, the trained model acts as a mediator that transforms channel responses into position estimates, simplifying the UE processing requirements while maintaining high accuracy.

The patent makes reference signals multi-functional by using them for both traditional purposes (channel quality, beam management) and positioning. The same downlink reference signals that are already transmitted for other radio access functions are also utilized as input for the AI-based positioning model, eliminating the need for separate dedicated positioning reference signals and reducing network overhead.

The patent changes the utilization approach of reference signals from traditional geometric measurement inputs to AI model inputs. By processing the same reference signal measurements through a neural network that has learned positioning-relevant features, the system extracts more positioning information from the same signal resources, effectively reducing the need for additional reference signals.