LoRaWAN RSS Localization with LSTM and Transformer Models

The patent introduces neural network models (LSTM and transformer) as intermediary systems that process raw signal measurements and transform them into accurate location estimates. These models act as mediators between the noisy wireless signals affected by attenuation and multipath fading, and the final location determination, effectively filtering out the harmful effects through learned patterns from training data.

The system performs preliminary training of neural network models using labeled location data and corresponding signal measurements before actual localization. This preliminary action prepares the models to compensate for signal attenuation and multipath fading effects during operation, enabling accurate localization without requiring real-time correction for these harmful factors.

The patent uses copying by training neural network models offline using comprehensive training datasets that include various signal conditions and locations. The trained models are then copied and deployed to edge devices or servers, allowing complex computation to be performed during the training phase rather than during actual localization operations. This enables accurate localization with reduced runtime complexity.

The system implements dynamic model selection and configuration, where different neural network architectures (LSTM vs. transformer) can be chosen based on specific deployment requirements. The models adapt to different network conditions and can be retrained or updated as needed, providing flexibility that manages complexity while maintaining accuracy across varying operational contexts.

The patent implements partial action by providing uncertainty measures selectively based on application requirements. The system can output only the location estimate when basic tracking is needed, or add uncertainty information when higher reliability is required. This partial provision of information reduces computational overhead while maintaining the option to enhance reliability when necessary.

The uncertainty measures provide feedback to the localization system and end users about the quality of location estimates. This feedback mechanism allows the system to adjust its operation, such as requesting additional measurements or switching to alternative localization methods when uncertainty is high, thereby improving overall reliability without continuously consuming maximum computational resources.