Radio Environment Estimation Using Phase-Aware Indirect Wave Synthesis

The patent extracts and separately processes direct waves and indirect waves. Direct waves are calculated using simple distance-based path loss models, while indirect waves (reflected, diffracted, transmitted) are processed through machine learning models. This separation avoids the exponential complexity of traditional ray tracing that calculates all propagation paths simultaneously, thereby reducing calculation time while maintaining estimation accuracy.

The patent performs preliminary classification of propagation paths into direct and indirect categories before detailed calculation. By pre-identifying which paths are direct (line-of-sight) and which are indirect (obstructed), the system can apply appropriate simplified models to direct paths and ML models to indirect paths, avoiding unnecessary complex calculations for all paths and thus reducing overall computation time.

The patent creates a virtual copy of the radio wave environment through machine learning models trained on measured data. Instead of performing physical measurements at multiple candidate positions, the system uses a trained ML model to predict reception states at any candidate position based on the building layout and environmental features. This virtual copying approach maintains measurement accuracy while eliminating the need for repeated physical measurements, thereby significantly improving installation evaluation efficiency.

The patent changes the approach from direct physical measurement to parameter-based prediction. By training ML models on measured data to learn the relationship between environmental parameters (layout, obstructions, antenna positions) and reception characteristics, the system can predict reception states by inputting parameter values rather than performing actual measurements. This parameter change enables rapid evaluation of multiple installation candidates without additional field work.

The patent segments the radio wave propagation into distinct components: direct waves and indirect waves. Each segment is processed by specialized models - direct waves use geometric path loss calculations that inherently preserve phase information, while indirect waves use ML models trained to predict amplitude and phase separately. This segmentation allows the system to maintain phase accuracy for direct paths while using efficient ML processing for indirect paths, resolving the contradiction between speed and phase accuracy.

The patent extends the neural network model to handle phase information by adding a temporal or frequency dimension to the prediction output. Instead of predicting only amplitude, the model predicts both amplitude and phase components of the electric field. This dimensional extension allows the ML model to capture phase effects while maintaining the computational efficiency of machine learning approaches.