PaperlayerView exact revision on arXiv

Source-linked AI summary

Clearing the Underbrush: AI-Enhanced RF Interference Suppression

Rahul Jain, Pierre Trepagnier, Rick Gentile, Joey Botero, Alexia Schulz

arXiv:2608.24974v1cs.LGcs.AIeess.SP

TL;DR

The paper addresses RF interference rejection for tactical-edge communications, where structured interference and deployment constraints make robust, fast processing important. It adds an FSQ tokenizer to autoregressive transformer-based models and evaluates inference optimizations, reporting improved rejection and real-time-compatible latency in the studied experiment.

  • Problem

    RF interference rejection must handle structured interference and challenging wireless conditions while meeting tactical-edge latency and SWaP constraints.

  • Method

    The work adds an FSQ tokenizer to an autoregressive RF Transformer and experiments with inference optimizations for faster processing.

  • Results

    AI-based models outperformed traditional methods on audio-quality metrics, while the RF Transformer with tokenizer showed interference rejection for SINR down to -6 dB.

  • Takeaways & Limitations

    Optimized inference supported a real-time communications application, with an example overall latency of 92 msec and throughput of roughly 1.2 MHz.

Abstract

from arXiv · show

AI-based structured interference rejection has grown more popular because deep learning approaches can outperform traditional methods by jointly considering the signal of interest (SOI) and the signal mixture (SOI plus interference). This work builds on a previous AI-enabled approach utilizing autoregressive transformer-based models by adding a Finite Scalar Quantization (FSQ) tokenizer layer which aims to improve the interference rejection performance while keeping overall latency to a minimum. Additionally, we experiment with other inference optimization techniques with the goal of speeding up inference without much accuracy loss. We explore this space with an experiment where the SOI is a digitally modulated radio frequency (RF) signal and the structured interference is a digital television signal, an extremely common type of Orthogonal Frequency-Division Multiplexing (OFDM) transmission. Our results achieve low latency and increased interference rejection over traditional techniques and prior work with other AI-enabled methods. We demonstrate the benefits of the AI-enabled approaches via audio metrics such as Perceptual Evaluation of Speech Quality (PESQ). Additionally, we explore a variety of applications and detail how our interference rejection algorithm may be used in operationally-relevant scenarios.

I. INTRODUCTION

The paper operationalizes AI-enhanced RF interference rejection for tactical-edge software-defined radios, where real-time inference must function under severe SWaP constraints. It addresses difficult wireless conditions, structured interference, and single-antenna settings relevant to resilient communications.

  • The work applies AI to enhance RF interference rejection on software-defined radios at the tactical edge.
  • Tactical-edge deployment requires real-time or near-real-time inference despite severe size, weight, and power constraints.
  • Wireless systems must recover signals across varied SINR conditions while propagation effects and crowded spectrum can cause bit errors.
  • OFDM provides structured time-frequency patterns that AI-enabled approaches can learn from training data, including when the SOI is digitally modulated.
  • The approach is suited to SWaP-constrained systems with only a single antenna element, unlike spatial techniques relying on phased arrays.

III. APPROACH

The approach advances from traditional RF interference rejection toward AI models designed for structured interference and operational deployment. It adds compact RF representations through tokenization to support learning from high-rate signal data.

  • Traditional methods can be suboptimal for structured interference, computationally slow, or dependent on detailed channel or interferer knowledge.
  • RF WaveNet, RF Transformer, and RF Transformer Decoder are presented as AI techniques that may be fast enough for operational structured-interference rejection.
  • The paper explores transformer variations to improve signal quality and/or inference time.
  • The tokenizer converts continuous RF sequences into finite discrete values so models can learn compressed representations instead of raw continuous signals.
  • FSQ maps continuous data points into discrete tokens by partitioning the input space and assigning each region a representative value.

B. RF Transformer using the RF FSQ Tokenizer

The RF Transformer uses an FSQ tokenizer at its output to convert predicted SOI tokens back into a continuous signal sequence. The design uses autoregressive generation while avoiding tokenization of the input mixture because that representation may be harder to learn under strong interference.

  • The RF Transformer architecture is adapted to include the RF FSQ tokenizer.
  • The tokenizer’s output-side upsampler converts predicted SOI tokens back into a continuous signal sequence.
  • The model generates tokens autoregressively and converts them to the continuous signal one step at a time.
  • The authors hypothesize that tokenizing the input signal mixture is more difficult, especially when interference levels are high.

C. Training Methodology

The training pipeline preprocesses SOI and interference data to align sample rates, control magnitudes, and vary SINR before training the tokenizer and transformer. The experiment uses QPSK voice as the SOI and DVB-T2 television signals as structured interference.

  • Preprocessing brings SOI and interference datasets to a common sample rate and controls signal magnitudes for stable, efficient learning.
  • The processing pipeline adds bandpass filtering to interference data and calculates SINR across a defined range.
  • The RF FSQ tokenizer is trained on the SOI, after which the RF Transformer predicts tokenized representations that the tokenizer converts into continuous signals.
  • The experiment mixes digitally modulated QPSK voice as the SOI with synthetic DVB-T2 digital television signals as interference.

IV. RESULTS

The digital radio experiment evaluated AI-based models using processed SOI quality metrics and a fixed held-out test set. AI-based models outperformed traditional methods, with the RF Transformer plus tokenizer maintaining strong interference rejection down to -6 dB SINR.

  • SOI Quality: AI-based models were evaluated for processed SOI quality using audio-domain metrics including PESQ, SDR, LSD, and Mel-CD.The evaluation used a fixed held-out test set rather than k-fold cross-validation.
  • SOI Quality: AI-based models outperformed traditional methods across all reported audio-quality cases.The comparison used processed audio against ground-truth audio.
  • SOI Quality: -6 dB SINR: the RF Transformer with tokenizer maintained good interference rejection down to this level.This performance was reported as beyond the other evaluated methods and aligned with prior experiments.
  • SOI Quality: PESQ evaluation used audio downsampled from 44.1 kHz to 16 kHz, while BER analysis was omitted because audio metrics better represented communication quality for this dataset.BER validation was presented separately using a non-communications IQ dataset.
  • SOI Quality: The RF Transformer + Tokenizer’s performance advantage may partly reflect its larger parameter count, leaving the tokenizer’s independent contribution unresolved.Other trained architectures were omitted because they performed poorly or showed inconsistent interference rejection.

B. Overall Latency and Output Throughput

The study treated buffer latency, inference time, and output throughput as key deployment measures for real-time streaming. On the Jetson, the tokenizer model stayed within an approximately 5-second latency budget for batch sizes up to 64, but FP32 throughput required further optimization.

  • Latency Measures: Buffer latency and inference time were combined to estimate overall system latency, with output throughput treated as a real-time streaming requirement.Latency results used an NVIDIA Jetson AGX Orin, with an NVIDIA H100 NVL as a comparison platform.
  • Buffer Latency: 4 msec: buffer latency for L = 2560 and B = 1, compared with 929 msec at B = 256 in the digital radio experiment.The experiment used a 705.6 kHz sample rate; waiting to fill large buffers was considered unacceptable for real-time use.
  • Overall Latency: Around 5 sec: RF Transformer + Tokenizer latency on the Jetson for all B ≤64, primarily because of high inference time.RF WaveNet and the RF Transformer Decoder remained below 1 sec for B ≤64, while the tokenizer model’s latency was described as perceptible but tolerable.
  • Output Throughput: B = 1 minimizes buffer latency, but real-time tactical voice also requires sufficient throughput to avoid a backlog of samples.The FP32 baseline did not meet this throughput requirement on the Jetson, so additional optimizations were needed.

C. Optimizations for Latency and Output Throughput

The study evaluates mixed-precision and compilation techniques to reduce inference latency and increase output throughput. Torch-TensorRT provides the strongest reported acceleration, enabling a real-time communications configuration on the Jetson.

  • Baseline evaluation: FP32 evaluation establishes latency and throughput baselines for RF WaveNet, RF Transformer Decoder, and RF Transformer + Tokenizer models.The models were evaluated on the NVIDIA Jetson AGX Orin and NVIDIA H100 NVL.
  • Automatic mixed precision: The transformer models receive limited low-batch-size AMP speedup because their operations are memory-bound and their sequential architecture limits GPU parallelism.Higher batch sizes begin to provide equal or lower latency with AMP.
  • Automatic mixed precision: AMP with FP16 provides acceptable latency and theoretical output throughput for RF Transformer Decoder on the Jetson when B ≥8.B = 16 or B = 32 may be more practical after accounting for miscellaneous system-processing overheads.
  • Torch-TensorRT optimization: 1.2-5.7x improvement in τ is achieved on the Jetson after compiling RF Transformer Decoder with Torch-TensorRT.FP16 with Torch-TensorRT performs best, exceeding the improvement from AMP; kernel fusion is among the applied optimizations.
  • Optimization boundary: The RF Transformer + Tokenizer could not be compiled with Torch-TensorRT because of issues likely related to dynamically fed-back decoded SOI tokens.This constrains the reported Torch-TensorRT optimization results to the RF Transformer Decoder.
  • Torch-TensorRT optimization: 92 msec overall latency and roughly 1.2 MHz theoretical output sample throughput are achieved with FP16 Torch-TensorRT at B = 16.The latency combines 58 msec buffer time and 34 msec inference time, while the throughput exceeds the 705.6 kHz input sample rate.

A. Use Cases

The paper targets operational use of AI interference rejection in tactical-edge communications, especially real-time voice. It also distinguishes voice requirements from more relaxed data-communication scenarios.

  • Tactical-edge deployment: The proposed receiver preprocessor could be added ahead of existing tactical radios without changing their hardware or software.Its output is converted back into an RF signal and fed into the existing radio connector where the antenna was installed.
  • Real-time voice: Real-time voice requires latency low enough not to distract operators at either end of the radio interface.RF Transformer Decoder can support this use case, whereas RF Transformer + Tokenizer has noticeable delay and low output throughput.
  • Data communications: Data communications may relax latency and output-throughput requirements depending on the application, such as file transmission.In this setting, RF Transformer + Tokenizer can also recover the signal when interference is present.

B. Future Work

Future work expands evaluation across unseen OFDM interferers, improves robustness and deployment realism, and broadens applications including spectrum sharing. The authors also plan additional inference optimization.

  • New interference sources: The authors plan to test transformer-based signal separation on new OFDM interference sources that were not used for training.Digital television signals are identified as convenient future test sources because their subcarrier modulation techniques are configurable.
  • Robustness and deployment: Future modeling work targets improved estimated signal quality, robustness, and generalization to over-the-air transmissions.The authors plan outdoor propagation testing at government-sponsored test ranges and representative dataset collection across variations.
  • Expanded applications: Spectrum sharing is proposed as an application in which multiple transmissions could operate in the same band despite potential interference.The paper identifies possible 5G use of S-band spectrum occupied by military radars as an example.
  • Expanded applications: The authors intend to extend these algorithms to multifunction RF systems beyond the tactical edge.This expansion is motivated by interference concerns arising from increasing contention between civil and military spectrum uses.

VI. CONCLUSION

The work operationalizes AI-enhanced structured-interference rejection for tactical radios using digitally modulated SOI and interferer signals. It reports tokenizer-based improvement, optimized support for higher sample rates, and broader potential applications.

  • Conclusion: The RF Transformer with a tokenizer improves upon previous AI-enabled and traditional interference suppression methods in experiments with digital SOI and interferer signals.The conclusion frames this result as progress toward operationalizing AI-enhanced interference rejection for tactical radios.
  • Conclusion: Optimized inference supports higher-sample-rate datasets for a real-time resilient communications application at the tactical edge.The conclusion presents this as a demonstrated capability of the work.
  • Conclusion: AI-enhanced interference rejection improves detection, demodulation, and decoding over a range of SINR levels without detailed design-level knowledge of the interferer or propagation conditions.The conclusion extends the stated application scope to military and commercial settings where interference rejection is crucial.
  • Conclusion: The paper discusses spectrum sharing and data transfer over communications links as future use cases for the interference-rejection algorithms.These use cases are presented alongside future directions for broader deployment.
Loading 2608.24974v1…