FSA-Net: How Deep Learning Pinpoints Unknown Signal Sources with Sub-Degree Precision

Global, Source:

Imagine trying to locate multiple whispers in a noisy room without knowing how many people are speaking or where they stand. This is the daily reality for engineers working with wireless communications and radar systems. Traditional tools like MUSIC and ESPRIT have long been the go-to methods for determining the direction of arrival (DOA) of signals. Yet these classic techniques hit a wall when signal quality drops, sources cluster closely together, or the number of active transmitters remains unknown. A new deep learning framework called FSA-Net steps into this gap by estimating both the count of sources and their precise angles simultaneously. It does so without requiring any prior knowledge about the scene.

What makes FSA-Net particularly intriguing is its ability to handle three distinct challenges within a single unified model. Most existing solutions require separate configurations for one, two, or three sources. FSA-Net treats mixed scenarios as a natural part of its workflow. The network also tackles permutation ambiguity, a persistent headache in neural networks where ordered outputs fail to match unordered physical realities. By using a forward sort-and-gather mechanism, it aligns predicted angles with existence logits through synchronized permutations. This ensures that every angle prediction corresponds accurately to a specific source presence. Additionally, the inclusion of an adjacent-gap term in the objective function sharpens resolution for sources that sit very close to each other.

The input data for this system comes from a normalized sample covariance matrix generated by a uniform linear array. By stacking real, imaginary, and magnitude components, the model captures the full spatial correlation structure of the environment. The results speak for themselves. FSA-Net maintains sub-degree root mean square error across a broad spectrum of signal-to-noise ratios. Classical subspace methods often falter in low SNR conditions, but this deep learning approach holds steady. Source count accuracy reaches near-perfect levels, identifying the correct number of active sources with over 97% reliability even at -15 dB SNR.

Efficiency is another compelling aspect of this architecture. With just 4.439 million parameters and low floating-point operations per second, FSA-Net offers a compact solution ideal for resource-constrained devices. While the current version caps out at three simultaneous sources, it sets a high bar for joint estimation tasks. Researchers are already looking ahead to extending this capability to quantized measurements and coherent sources. These advancements promise more robust beamforming for next-generation reconfigurable intelligent surfaces and distributed time-modulated arrays.