Transformer-Enhanced Localization via Adaptive PDP Representation Under Dynamic Bandwidths.

Area of Science:

• Wireless communication
• Signal processing
• Machine learning

Background:

• Accurate wireless positioning is crucial for Internet of Things (IoT) and autonomous aerial vehicles (AAVs).
• Existing methods struggle with dynamic bandwidths and multipath environments.
• Current learning-based localization methods lack adaptability across different signal bandwidths due to reliance on bandwidth-specific channel state information (CSI).

Purpose of the Study:

• To propose a unified, neural network-oriented framework for bandwidth-adaptive wireless positioning.
• To enhance the generalization and robustness of localization models across varying signal bandwidths.
• To overcome the limitations of conventional methods in dynamic bandwidth conditions.

Main Methods:

• A novel framework constructs bandwidth-adaptive power delay profile (PDP) representations.
• A PDP preprocessing scheme uses adaptive zero-padding and oversampled IFFT of heterogeneous CSI.
• A sub-band-sliced PDP representation is developed, processing PDPs as Transformer tokens.
• A dedicated Transformer model performs location estimation from multi-access point PDPs.

Main Results:

• The proposed framework demonstrates superior cross-bandwidth generalization capabilities.
• Achieved significantly higher localization accuracy compared to existing analytical and learning-based baselines.
• The PDP preprocessing and Transformer-based approach effectively handles dynamic bandwidth conditions.

Conclusions:

• The preprocessing-PDP-plus-Transformer framework offers a robust solution for wireless positioning under dynamic bandwidths.
• This approach enhances the adaptability and accuracy of localization for IoT and AAV applications.
• The method provides a significant advancement over traditional localization techniques.