AI-Assisted Wireless Positioning With Adaptive Model Retraining

The patent introduces an AI/ML model as an intermediary component between the measurement data collection and the positioning result generation. This model processes raw measurement data from multiple sources (signal strength, time of arrival, angle of arrival) and transforms them into accurate positioning information, thereby achieving high positioning accuracy without directly increasing the complexity of the underlying positioning system architecture

The patent replaces traditional mechanical/mathematical positioning calculation methods with an AI/ML-based computational system. Instead of using conventional geometric or algebraic methods to calculate position from measurement data, the system employs trained neural networks or machine learning models that have learned optimal positioning patterns from training data, achieving superior accuracy while managing system complexity through software-based intelligence

The patent implements a dynamic positioning system where the AI/ML model can be retrained and updated based on changing environmental conditions. The system collects measurement data from the actual environment, retrains the model with this new data, and deploys updated models to maintain positioning accuracy as the environment evolves, thereby achieving high adaptability through controlled model complexity

The patent incorporates feedback mechanisms where positioning results and measurement data are continuously monitored and used to evaluate model performance. When performance degradation is detected or environmental changes occur, the system triggers model retraining using collected data, creating a closed-loop system that adapts to changing conditions while managing complexity through conditional updates rather than continuous changes

The patent applies preliminary action by pre-training the AI/ML model offline using comprehensive, high-quality training datasets that would be too large and complex to process in real-time. The trained model is then deployed for real-time positioning inference, separating the computationally intensive training phase from the time-critical application phase, thereby achieving both high accuracy and fast processing speed

The patent segments the data processing workflow into distinct phases: offline model training using comprehensive datasets, and online inference using the trained model. This segmentation allows the system to utilize large, complex datasets for training without compromising real-time performance, as the heavy computational burden is confined to the offline training phase while the online phase operates efficiently with the pre-trained model