Source-linked AI summary
Spectrum Sensing Based on Deep Learning Classification for Cognitive Radios
Shilian Zheng, Shichuan Chen, Peihan Qi, Huaji Zhou, Xiaoniu Yang
TL;DR
Spectrum sensing requires robust detection across diverse signals and noise conditions. The paper formulates it as deep-learning classification using normalized received power, diverse signal/noise training, and transfer learning for real-world signals. It reports better performance than two traditional methods, adaptation to untrained signals, improved ACARS detection after transfer learning, and stronger colored-noise performance.
Problem
General spectrum sensing must handle diverse signals, noise-power uncertainty, untrained signals, and non-ideal noise beyond methods designed for specific signals or AWGN.
Method
The method classifies signal versus noise with a CNN using power-spectrum inputs, diverse signal and noise training data, normalized received power, and transfer learning for real-world signals.
Results
The method outperforms maximum-minimum eigenvalue ratio-based and frequency domain entropy-based methods, adapts to untrained signals, improves ACARS detection after transfer learning, and performs better under colored noise.
Takeaways & Limitations
Deep-learning classification provides a spectrum-sensing approach that generalizes beyond trained signal types and remains effective under colored noise.
Takeaways & Limitations
Most experiments use simulation data; large-scale over-the-air real-world signal experiments remain future work.
Abstract
from arXiv · showhide
Spectrum sensing is a key technology for cognitive radios. We present spectrum sensing as a classification problem and propose a sensing method based on deep learning classification. We normalize the received signal power to overcome the effects of noise power uncertainty. We train the model with as many types of signals as possible as well as noise data to enable the trained network model to adapt to untrained new signals. We also use transfer learning strategies to improve the performance for real-world signals. Extensive experiments are conducted to evaluate the performance of this method. The simulation results show that the proposed method performs better than two traditional spectrum sensing methods, i.e., maximum-minimum eigenvalue ratio-based method and frequency domain entropy-based method. In addition, the experimental results of the new untrained signal types show that our method can adapt to the detection of these new signals. Furthermore, the real-world signal detection experiment results show that the detection performance can be further improved by transfer learning. Finally, experiments under colored noise show that our proposed method has superior detection performance under colored noise, while the traditional methods have a significant performance degradation, which further validate the superiority of our method.
I. INTRODUCTION
Wireless-spectrum demand is increasing, making spectrum sensing important for cognitive radio. The paper formulates sensing as binary classification and applies a deep CNN, comparing it with traditional methods.
- Spectrum sensing is a key prerequisite for dynamic spectrum access in cognitive radio as demand for wireless spectrum increases.
- The paper formulates spectrum sensing as a two-category classification problem and proposes a deep convolutional neural network solution.
- Extensive simulation experiments compare the proposed method with traditional spectrum sensing methods.
A. Related Work
Prior spectrum-sensing methods often require handcrafted statistics or features and may target cooperative sensing or specific signals. This paper instead addresses general single-node sensing with deep learning trained on diverse signals and noise.
- A. Related Work: Traditional methods rely on decision statistics whose design requires extensive analysis and domain-specific knowledge.
- A. Related Work: Earlier machine-learning approaches commonly use cooperative sensing or pre-extracted features, making performance depend on those feature choices.
- A. Related Work: Some deep-learning studies outperform traditional detectors but focus on specific signals rather than spectrum sensing in the general sense.
- A. Related Work: The proposed method models general spectrum sensing as binary classification, trains on diverse signals and noise, and normalizes received power for noise-power uncertainty.
- A. Related Work: The method uses power spectra as CNN inputs, learns signal and noise features automatically, and aims to detect varied—including untrained—signals without prior feature extraction.
- A. Related Work: Experiments compare the method with maximum-minimum eigenvalue ratio and frequency-domain entropy detectors and test untrained signal types.
II. PROBLEM FORMULATION
Spectrum sensing is expressed as a binary hypothesis test distinguishing received signal-plus-noise from noise alone. The paper recasts these hypotheses as signal-versus-noise classification.
- II. PROBLEM FORMULATION: The binary hypotheses distinguish received signal-plus-noise from noise-only observations.
- II. PROBLEM FORMULATION: The received signal r(n), transmitted signal s(n), noise w(n), and channel gain h define the sensing model.
- II. PROBLEM FORMULATION: From a classification perspective, the two categories are signal and noise.
A. The CNN Model
The supplied passages identify cross-entropy training and a residual-network architecture for the CNN model, with the network structure specified in Table I.
- A. The CNN Model: The CNN uses a residual network whose layered structure is specified in Table I.
- A. The CNN Model: Training uses cross-entropy loss with N training samples and per-sample output confidence for the true label.
C. The Decision Method
The method uses a residual CNN to classify power-spectrum inputs as signal or noise. Its architecture combines convolutional, pooling, residual, dropout, and fully connected layers before producing class confidences.
- C. The Decision Method: A residual CNN processes the signal power spectrum and outputs confidence scores for signal and noise classes.The network uses two convolutional layers, max-pooling, six residual modules, average pooling, a 32-neuron fully connected layer, dropout, and a final classification layer.
B. The Training Method
The training method broadens generalization by combining diverse signal types with matched noise data and using a thresholded softmax confidence for detection.
- B. The Training Method: The training set includes many wireless signal types and equally sized pure-noise data to support detection of untrained signals.The stated goal is to generalize from trained signal types to other signal types not seen during training.
- B. The Training Method: Detection uses the softmax confidence for the signal class rather than simply selecting the class above 0.5.A threshold γ between 0 and 1 is selected according to the required false-alarm probability.
- B. The Training Method: The hypotheses classify a sample as signal when 1 − t_noise exceeds γ and as noise otherwise.The threshold satisfies 0 < γ < 1 and is tied to the required probability of false alarm.
A. Data Generation
The experiments generate diverse simulated signals, compare the method with traditional detectors, test untrained signals and noise uncertainty, and evaluate transfer learning on real-world ACARS signals.
- B. Data Generation: Training data cover eight modulation types across SNRs from −20 dB to 20 dB, with 512-sample signals and matched AWGN or colored-noise data.Each modulation uses 1,000 training samples and 500 test samples at each SNR; three additional modulations are reserved for unknown-signal testing.
- B. Data Generation: The proposed method outperforms maximum-minimum eigenvalue ratio and frequency-domain entropy methods under various false-alarm probabilities, especially at low SNR.The comparison uses simulated AWGN, and Fig. 3 reports results for P_f = 0.1, 0.01, and 0.001.
- B. Data Generation: The trained model detects untrained 8PSK, 8FSK, and 64QAM signals with high probability and outperforms traditional methods at P_f = 0.01.These modulations are evaluated as unknown signal types rather than training classes.
- B. Data Generation: Noise-power uncertainty has little effect on the proposed method’s detection performance while maintaining the target false-alarm probability of 0.01.The uncertainty factor scales the real noise power, and the simulations use AWGN.
- B. Data Generation: After transfer learning, ACARS detection improves beyond the traditional methods, whereas the untuned model is similar to frequency-domain entropy detection.The ACARS evaluation uses over-the-air signal samples, with simulated noise injected for the plotted SNR values.
V. CONCLUSIONS
The deep-learning spectrum-sensing method outperforms two traditional methods, generalizes to untrained signals, and benefits from transfer learning for real-world ACARS signals. Under pink noise, it maintains performance close to AWGN while traditional methods degrade significantly.
- The proposed method outperforms maximum-minimum eigenvalue ratio-based and frequency domain entropy-based sensing methods in simulation.
- The method adapts to detecting a variety of untrained signals.
- Transfer learning further improves performance when detecting real-world ACARS signals.
- Under pink noise, the method maintains performance close to AWGN, while traditional methods degrade significantly.The experiment uses a false alarm probability of 0.01.