Deep Learning for Spectrum Sensing

Jiabao Gao; Xuemei Yi; Caijun Zhong; Xiaoming Chen; Zhaoyang Zhang

Source-linked AI summary

Deep Learning for Spectrum Sensing

Jiabao Gao, Xuemei Yi, Caijun Zhong, Xiaoming Chen, Zhaoyang Zhang

arXiv:1909.02730v1Sep 6, 2019cs.ITeess.SP

TL;DR

Reliable primary-signal detection is needed in cognitive radio, but energy detection suffers from an SNR-wall caused by noise uncertainty. The paper develops deep-learning detectors that exploit modulated-signal structure without requiring prior signal or noise information, and reports improved standalone and cooperative sensing performance.

Problem
Energy detection is simple and blind but is constrained by noise uncertainty, producing an SNR-wall that limits reliable detection at low SNR.
Source: §I · p. 1
Method
The paper develops a CLDNN-based DetectNet and a soft cooperative fusion network, using signal structure and distributed soft information for detection.
Source: §I · p. 1, A. Network Architecture Design · p. 2, C. Customized Two-stage Training · p. 2
Results
DetectNet lowers the energy detector's SNR-wall from −5.35 dB to −8.5 dB at Pf = 5.92% and Pd = 90%, while SoftCombinationNet achieves nearly the same Pd as Logical-OR with significantly lower Pf.
Source: 1) Comparison with Different Networks · p. 3, B. Simulation Results · p. 4
Takeaways & Limitations
Performance improves over conventional energy detection, remains relatively insensitive to modulation order for similar schemes, and cooperative soft combination achieves high Pd and low Pf simultaneously.
Source: 2) Impact of Modulation Scheme · p. 3, 3) Generalization Ability · p. 3, §V · p. 4

Abstract

from arXiv · show

In cognitive radio systems, the ability to accurately detect primary user's signal is essential to secondary user in order to utilize idle licensed spectrum. Conventional energy detector is a good choice for blind signal detection, while it suffers from the well-known SNR-wall due to noise uncertainty. In this letter, we firstly propose a deep learning based signal detector which exploits the underlying structural information of the modulated signals, and is shown to achieve the state of the art detection performance, requiring no prior knowledge about channel state information or background noise. In addition, the impacts of modulation scheme and sample length on performance are investigated. Finally, a deep learning based cooperative detection system is proposed, which is shown to provide substantial performance gain over conventional cooperative sensing methods.

I. INTRODUCTION

Cognitive radio requires reliable detection of weak primary signals, but conventional energy detection is limited by noise uncertainty and the SNR-wall. The paper proposes deep-learning detectors that exploit signal structure without requiring prior signal or noise information.

Reliable detection of weak primary signals is essential for protecting licensed users while enabling opportunistic spectrum use.Source: §I · p. 1
Energy detection depends heavily on noise-density knowledge and fails below the SNR-wall when noise uncertainty is present.The practical SNR-wall is about −6 dB, compared with the −15 dB limit required by IEEE 802.22.Source: §I · p. 1
Blind sensing is desirable because secondary users often lack prior knowledge of primary signals, motivating methods that identify their underlying structure.Source: §I · p. 1
The paper proposes a CLDNN-based detector applicable to arbitrary primary-signal types without requiring online information about the signal or noise density.Source: §I · p. 1
Signal detection is formulated as binary hypothesis testing for the absence or presence of a primary signal in a band.The received sample is modeled under H0 as noise and under H1 as a channel-scaled primary signal plus noise.Source: §II · p. 1
Energy detection evaluates normalized received-signal energy and uses false-alarm and missed-detection probabilities to assess performance.The noise density can be estimated from noise-only samples, and the resulting SNR-wall depends on the sensing and noise-sample settings.Source: §II · p. 1

III. DL BASED DETECTOR

The deep-learning detector maps received-sample vectors through a customized multilayer network to classify whether a primary signal is absent or present. Its final SOFTMAX layer produces hypothesis probabilities, and argmax selects the detected class.

The DL detector processes the received-sample vector through a customized neural network with L layers.Source: §III · p. 2
Intermediate layers apply learned weights and activation functions, while the final SOFTMAX layer outputs probabilities for the two hypotheses.Source: §III · p. 2
The detector returns the index of the hypothesis with the largest output probability.Source: §III · p. 2

A. Network Architecture Design

The proposed DetectNet uses a CLDNN architecture combining convolutional, recurrent, and fully connected layers. Its configuration was selected through cross-validation and includes dropout to reduce overfitting.

A. Network Architecture Design

The paper adopts a CLDNN architecture motivated by its performance in modulation-recognition tasks and validates it through numerical simulations.Source: A. Network Architecture Design · p. 2
DetectNet uses two convolutional layers, two LSTM layers, one intermediate fully connected layer, and two additional fully connected layers.Source: A. Network Architecture Design · p. 2
The final fully connected layer uses SOFTMAX, other layers use ReLU, and dropout follows every layer to prevent overfitting.Source: A. Network Architecture Design · p. 2
The model hyperparameters were determined through extensive cross-validation and are reported in Table I.Source: A. Network Architecture Design · p. 2, Table I · p. 2
DetectNet's network architecture is illustrated in Fig. 1.Source: Figure 1 · p. 2, A. Network Architecture Design · p. 2

B. Dataset Generation and Preprocessing

The dataset uses digitally modulated signals as positive samples and CSCG noise as negative samples, with a 3:1:1 training, validation, and testing split. Energy normalization is applied before training and inference to expose modulation structure and reduce dependence on signal energy.

Eight kinds of digitally modulated signals at different SNRs are generated as positive samples, while CSCG noises serve as negative samples.The signal generation follows RadioML2016.10a.Source: B. Dataset Generation and Preprocessing · p. 2
The dataset is partitioned into training, validation, and testing sets using a 3:1:1 split ratio.Source: B. Dataset Generation and Preprocessing · p. 2
Energy normalization is performed before training or inference instead of directly using received time-domain complex signals.The stated motivations are to minimize signal-energy effects, expose modulation structure, and improve robustness when background noise changes.Source: B. Dataset Generation and Preprocessing · p. 2

C. Customized Two-stage Training

A customized two-stage training strategy is designed for CFAR detection. It first trains the model to convergence, then stops within a preset false-alarm-probability interval to control the performance–training-time trade-off.

The first training stage uses early stopping with six epochs of patience to train the model to convergence.Source: C. Customized Two-stage Training · p. 2
The second stage resumes from the first stage's best model and stops when false alarm probability falls within a preset interval.This provides some control over detection performance despite DL methods' lack of precise performance control.Source: C. Customized Two-stage Training · p. 2
Smaller false-alarm intervals provide more precise performance control but require longer training.Source: C. Customized Two-stage Training · p. 2

D. Simulation Results

Simulations evaluate model comparisons, modulation effects, and reproducibility settings. DetectNet lowers the SNR-wall relative to the energy detector, while its performance varies across modulation families and is relatively insensitive to PSK modulation order.

Comparison with Different Networks

DetectNet and CNN outperform DNN and LSTM on QAM16 signals with sample length 128, while DetectNet achieves similar Pd to CNN with lower Pf.The comparison uses cross-validated hyperparameters and 0.2 dropout for all models.Source: 1) Comparison with Different Networks · pp. 2–3, Figure 2 · p. 3
DetectNet's SNR-wall is −8.5 dB, 3.15 dB lower than the energy detector's −5.35 dB at Pf = 5.92% and Pd = 90%.Source: 1) Comparison with Different Networks · p. 3

Impact of Modulation Scheme

FSK signals, especially GFSK, achieve better DetectNet performance than PSK and QAM signals at sample length 128.Source: 2) Impact of Modulation Scheme · p. 3, Figure 3 · p. 3
Detection-performance differences among BPSK, QPSK, and 8PSK are insignificant, indicating insensitivity to modulation order within these PSK signals.Source: 2) Impact of Modulation Scheme · p. 3

3) Generalization Ability:

DetectNet generalizes well across signals sharing a modulation type, but performance deteriorates when modulation types differ. Across sample lengths, it maintains an approximately 5 dB SNR-wall advantage over the energy detector.

Same-type modulation pairs, including QAM16/QAM64 and BPSK/QPSK, show decent DetectNet generalization.Source: 3) Generalization Ability · p. 3
Different-type pairs, such as BPSK and GFSK, produce significant performance deterioration.Source: 3) Generalization Ability · p. 3
DetectNet generalization reflects shared structural information among signals with the same modulation type and substantial structural differences across types.Source: 3) Generalization Ability · p. 3
Around 5 dB SNR-wall improvement over the energy detector persists across different sample lengths.The comparison uses operating points with Pd = 90% and a Pf stop interval of [7%, 9%] on GFSK signals.Source: 4) Impact of Sample Length · p. 3

IV. DL BASED COOPERATIVE DETECTION

The cooperative system combines local DetectNet outputs at a fusion center using a learned soft-information fusion rule. This design uses confidence information and node priorities that conventional hard-decision fusion does not exploit.

System Design

The proposed soft combination strategy implicitly exploits decision confidence and differing node priorities.Source: §IV · p. 3
Each sensing node uses DetectNet to produce a probability vector for the two primary-signal hypotheses.Source: A. System Design · p. 4
A three-layer fully connected neural network learns the fusion rule directly from the distributed node outputs.Source: A. System Design · p. 4
The cooperative model is called SoftCombinationNet, with a network architecture illustrated in Fig. 5.Source: A. System Design · p. 4, Figure 5 · p. 4

B. Simulation Results

Simulations evaluate cooperative sensing with QAM16 signals and 2, 4, or 8 sensing nodes. SoftCombinationNet achieves nearly the same detection probability as Logical-OR while substantially reducing false alarms in the practical operating regime.

Simulation Setup

QAM16 experiments use sample length 128 and independently and identically distributed channel gains across sensing nodes.Source: B. Simulation Results · p. 4
Two-, four-, and eight-node cooperative systems are evaluated.Source: B. Simulation Results · p. 4

Results

For Pd larger than 90%, SoftCombinationNet achieves almost the same Pd as Logical-OR with a significant reduction in Pf across all three systems.Source: B. Simulation Results · p. 4
The conclusion reports that SoftCombinationNet achieves high Pd and low Pf simultaneously.Source: §V · p. 4