Citation: Gai, J.; Zhang, L.; Wei, Z.
Spectrum Sensing Based on
STFT-ImpResNet for Cognitive Radio.
Electronics 2022, 11, 2437. https://
doi.org/10.3390/electronics11152437
Academic Editors: Alexandros-
Apostolos Boulogeorgos,
Panagiotis Sarigiannidis,
Thomas Lagkas, Vasileios Argyriou
and Pantelis Angelidis
Received: 13 June 2022
Accepted: 1 August 2022
Published: 4 August 2022
Publisher’s Note: MDPI stays neutral
with regard to jurisdictional claims in
published maps and institutional affil-
iations.
Copyright: © 2022 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
Article
Spectrum Sensing Based on STFT-ImpResNet for Cognitive Radio
Jianxin Gai *, Linghui Zhang and Zihao Wei
The Higher Educational Key Laboratory for Measuring & Control Technology and Instrumentations of
Heilongjiang Province, Harbin University of Science and Technology, Harbin 150080, China
* Correspondence: jxgai@hrbust.edu.cn
Abstract:
Spectrum sensing is a crucial technology for cognitive radio. The existing spectrum sensing
methods generally suffer from certain problems, such as insufficient signal feature representation,
low sensing efficiency, high sensibility to noise uncertainty, and drastic degradation in deep networks.
In view of these challenges, we propose a spectrum sensing method based on short-time Fourier
transform and improved residual network (STFT-ImpResNet) in this work. Specifically, in STFT, the
received signal is transformed into a two-dimensional time-frequency matrix which is normalized
to a gray image as the input of the network. An improved residual network is designed to classify
the signal samples, and a dropout layer is added to the residual block to mitigate over-fitting effec-
tively. We conducted comprehensive evaluations on the proposed spectrum sensing method, which
demonstrate that—compared with other current spectrum sensing algorithms—STFT-ImpResNet
exhibits higher accuracy and lower computational complexity, as well as strong robustness to noise
uncertainty, and it can meet the needs of real-time detection.
Keywords: spectrum sensing; residual network; short-time Fourier transform; cognitive radio
1. Introduction
With the advent of the 5G era, the lack of spectrum resources has become a realistic
problem that is inevitable [
1
,
2
]. Spectrum sensing is of vital importance to the optimization
of the utilization rate of spectrum resources, and has become the key technology in cognitive
radio [
3
]. In cognitive radio, the secondary user (SU) is allowed to access the spectrum
dynamically and randomly without interfering with the primary user (PU) [
4
]. The main
task of spectrum sensing is to explore spectrum holes [
5
] in order to increase the usage of
spectrum resources.
The traditional spectrum sensing methods can be broadly categorized into energy
detection (ED) [
6
,
7
], matched filter detection [
8
], cyclostationary feature detection [
9
],
waveform-based sensing [
10
], and covariance-based detection [
11
], etc. However, the pre-
defined threshold set by the traditional method has a dramatic influence on the detection
probability. With the continuous development of machine learning techniques, the method
of realizing spectrum sensing is migrating gradually from conventional statistical methods
to machine learning ones. Nowadays, deep learning methods are becoming more and more
popular to train spectrum sensing models to classify signals, which improves the detection
probability of spectrum sensing, and the model is optimized to approach the pragmatic
application level.
At present, some commonly used machine learning methods such as support vector
machines (SVM), artificial neural networks (ANN), long-term and short-term memory
networks (LSTM), and convolutional neural networks (CNN) have achieved partial success
in spectrum sensing. Chen et al. [
12
] proposed a SVM-based spectrum sensing algorithm
to recognize the PU signal by training SVM classifiers based on the energy vectors sampled
from SU. Supervised learning and unsupervised learning algorithms such as the naive
Bayes classifier, SVM, and hidden Markov model are compared in terms of classification
accuracy in [
13
], in which the experimental results show that the performance of the
Electronics 2022, 11, 2437. https://doi.org/10.3390/electronics11152437 https://www.mdpi.com/journal/electronics
