Joint Signal Detection and Automatic Modulation Classification via Deep Learning

Huijun Xing, Xuhui Zhang, Shuo Chang, Jinke Ren, Zixun Zhang, Jie Xu, Shuguang Cui

Abstract

This paper presents a novel joint deep learning framework, JDM, designed for the simultaneous detection and modulation classification of multiple coexisting wireless signals. Unlike traditional methods that treat signal detection and automatic modulation classification (AMC) as disjoint tasks, JDM integrates these processes into a unified pipeline inspired by object detection techniques in computer vision. To evaluate this framework in realistic scenarios, the authors introduce CRML23, a comprehensive dataset featuring multi-signal environments with various channel impairments. Experimental results demonstrate that JDM significantly outperforms conventional approaches, maintaining robustness across diverse signal parameters and noise levels.

Introduction

Cognitive radio systems rely heavily on the ability to sense the spectrum and identify active transmissions. Two critical components of this process are signal detection—locating signals in the frequency domain—and automatic modulation classification (AMC)—determining the encoding scheme used. Historically, these tasks have been researched in isolation, with AMC studies often assuming perfectly pre-processed and centered signals.

However, modern spectrums are increasingly crowded, with multiple signals overlapping in time and frequency. This paper addresses the limitations of decoupled systems by proposing a Joint framework for Detection and Modulation classification (JDM) and a corresponding realistic dataset, CRML23, which moves beyond idealized single-signal scenarios to reflect the multi-signal reality of modern spectrum environments.

The JDM Framework

The JDM framework employs a two-stage deep learning architecture designed to handle signals in a single pipeline.

Architecture of the Joint Detection and Modulation (JDM) framework.
  1. Radio Detection Module: Inspired by the YOLO (You Only Look Once) object detection paradigm, this module performs a Fast Fourier Transform (FFT) on raw I/Q data. A Convolutional Neural Network (CNN) then identifies signal "bounding boxes" in the frequency domain, defined by center frequency and bandwidth.
  2. The "Proposal" Link: The detection module outputs a "proposal" containing the specific frequency coordinates of detected signals, serving as the bridge to the next stage.
  3. Modulation Classification Module: This module receives the proposals, performs carrier frequency removal and low-pass filtering to isolate the signal of interest, and uses a second CNN backbone to predict the modulation scheme (e.g., BPSK, QPSK, 16QAM).

The CRML23 Dataset

A significant contribution of this work is the CRML23 (Coexisting RADIOML) dataset. Existing benchmarks like RADIOML.2016.10A typically feature only a single signal per sample, failing to capture the complexity of real-world interference.

Statistical distribution of the CRML23 dataset across bandwidth, frequency, signal count, and modulation types.

CRML23 provides a more rigorous testbed by including multiple signals per entry, randomly distributed across the spectrum. It incorporates realistic channel effects such as Rayleigh and Rician fading, Doppler shifts, and clock offsets. The dataset covers five primary modulation schemes: BPSK, QPSK, 8PSK, 16QAM, and 64QAM.

Results and Discussion

Evaluation of the JDM framework reveals several key insights into joint signal processing:

  • Superior Performance: The JDM framework significantly outperformed traditional combinations of Matched Filter detection and SVM or Decision Tree classifiers, which often struggled to accurately estimate bandwidth.
  • Detection Accuracy: The system achieved a mean Average Precision (mAP) of approximately 0.95 in ideal AWGN conditions. In complex simulated environments with interference, accuracy remained robust, dropping by only about 10%.
  • The Classification Bottleneck: System accuracy is highly sensitive to the classification module's performance at low Signal-to-Noise Ratios (SNR). While simple schemes like BPSK reach near-perfect accuracy quickly, higher-order modulations like 64QAM require significantly higher SNR for reliable classification.
  • Bandwidth and Channel Nuances: Narrower bandwidth signals were found to be easier to detect due to their concise spectral features. Interestingly, the model performed slightly better in Rayleigh channels than Rician channels in certain metrics, suggesting that Line-of-Sight components do not necessarily simplify the detection task for deep learning models.

Conclusion

The JDM framework represents a significant step toward more integrated and realistic cognitive radio systems. By bridging the gap between signal detection and classification, the authors provide a robust solution for managing crowded spectrum environments. Future work will focus on optimizing the interaction between the detection and classification modules and enhancing performance in extremely low SNR conditions.

Summary updated

BibTeX
@article{xing2024joint,
  author = {Huijun Xing and Xuhui Zhang and Shuo Chang and Jinke Ren and Zixun Zhang and Jie Xu and Shuguang Cui},
  title = {{Joint Signal Detection and Automatic Modulation Classification via Deep Learning}},
  journal = {IEEE Trans. Wireless Communications},
  year = {2024},
  doi = {10.1109/twc.2024.3450972},
}
Figure