Deep Learning-Based Sensing With Deterministic Pilots and Random Data Payloads

doi:10.1109/lwc.2026.3725823

Deep Learning-Based Sensing With Deterministic Pilots and Random Data Payloads

2026-08-21

0

PRE

AI

DOI：10.1109/lwc.2026.3725823

Original

Original request for help

Save

Abstract

En 中文

Achieving high-precision wireless sensing using existing communication waveforms is a core challenge in integrated sensing and communication (ISAC) systems. Particularly for multi-source localization, traditional covariance-based subspace methods suffer from high computational complexity and fail to utilize higher-order statistics of transmitted signals. Meanwhile, existing approaches relying solely on unknown uplink data discard the valuable deterministic information contained in pilots. To overcome these limitations, we propose PiDaNet, an end-to-end transformer architecture that exploits both deterministic pilots and random data payloads. This method abandons explicit covariance matrix computation and extracts features directly from communication signals. We fully leverage the attention mechanism of transformers where a self-attention module is first employed to capture implicit higher-order statistics from the unknown data payloads. Subsequently, a cross-attention module utilizes pilot-derived channel features as query vectors to guide the effective fusion of pilot and data features. Experimental results confirm that PiDaNet significantly outperforms schemes relying solely on unknown data in terms of mean squared error and maintains robust direction-of-arrival (DOA) estimation accuracy across diverse sensing scenarios.

Keywords:

Deep learning

DOA estimation

ISAC

transformer

Deep Learning-Based Sensing With Deterministic Pilots and Random Data Payloads

Abstract

Journal

Organization

Cited Papers