1. Introduction
Information has changed human life, from sharing bicycles to sharing everything, from the commercialization of 5G to the research and development of 6G. In today’s era, communication technology is vigorously developed, and the low-Earth orbit satellite Internet of things is used as the architecture to create a blue ocean of satellite Internet of things applications to realize the Internet of everything. With a large number of devices being connected to the wireless network, the scarcity of wireless spectrum resources is becoming increasingly serious, and the whole wireless communication environment has become extremely severe. The current spectrum resource allocation strategy is mainly a static spectrum allocation strategy, which licenses the spectrum to a limited number of users. The decision stipulates that even if the licensed band is idle, other users are still not able to access and use it, which results in a shortage of spectrum resources, but it is a lot of waste.
According to the research report of the Federal Communications Commission (FCC) of the United States, the spectrum occupancy of different licensed frequency bands varies due to the different number of services and requirements, ranging from 15% to 85% [
1]. Additionally, according to the survey of global spectrum utilization in recent years, the spectrum utilization rate is very low, and there is a large amount of spectrum waste worldwide. The American University of Sharjah measured the local UHF band (300 MHz to 3000 MHz) in 2017, and the study showed that the average occupancy rate was between 10% and 35% [
2]. Yu et al. used R&S EM100 digital compact receiver to measure the spectrum occupancy of China’s radio broadcast allocated frequency band from 87.5 MHz to 108 MHz, and the results showed that there were a large number of spectrum holes [
3]. In 2018, spectrum measurements of 6 GHz band models at three different locations in San Luis Potosi, Mexico, showed that the spectrum utilization rate was only 4.73% [
4]. In 2019, using RF Explorer 6G Combo to detect the spectrum occupancy of the GSM900 band in the Samsun area, and the calculation results showed that the occupancy rate was about 8.5% and 82% at −40 dBm and −75 dBm, respectively [
5]. The 24 h detection was carried out on the frequency band from 30 MHz to 1030 MHz in Pakistan, and the results showed that the average spectrum utilization rate was only 35.41%, among which the minimum occupancy rate of the 850 MHz to 1030 MHz band was only 25% [
6]. As can be seen in the above data, spectrums in large Numbers in the world cannot be effectively used, according to research, at any time and any place, the average spectrum utilization rate is not more than 5% [
7].
As early as 1999, in order to solve the problem of spectrum resources, Dr. Joseph Mitola proposed the concept of Cognitive Radio (CR) [
8]. After several years of radio technology development, the current CR technology mainly includes spectrum sensing, spectrum decision, spectrum shifting and spectrum sharing [
9,
10,
11]. Through the above four modules, CR technology has largely improved the spectrum resources problems. According to the data of “Cognitive Radio Market Report”, CR technology can effectively improve spectrum utilization, reaching 3–10% [
12]. Spectrum sensing technology is an indispensable part of the four modules, which realizes the dynamic access of the spectrum by sensing the holes in the spectrum environment. However, in the process of spectrum sensing, in order to obtain more spectrum resources, the range of spectrum sensing is often increased, but the corresponding perception time will be extended, and a lot of calculations and more and more complex algorithm processes will be generated. In order to solve such problems, Haykin first proposed spectrum prediction technology based on spectrum sensing technology in 2005, and then around 2012, researchers strengthened the concept of spectrum prediction technology. First of all, spectrum prediction technology can find the relationship between the spectrum through the study of the historical spectrum segment, predict the future spectrum state, find a range of possible free frequency bands, and then carry out spectrum perception of this range, which can directly find the free spectrum resources in a short time, so as to solve the problem of high spectrum perception consumption. Secondly, when the received signal is incomplete, we can also predict the amplitude and trend of the complete signal according to the spectrum prediction technology and carry out information demodulation. Finally, it allows users to access the underutilized spectrum opportunistically in the time domain or spectrum domain without significantly affecting other users, which provides the possibility for subsequent spectrum decision making, spectrum migration and spectrum sharing, thus assisting cognitive radio technology.
In the current research, the prediction results of spectrum prediction technology are relatively single, which can only complete the prediction in the ordinary time dimension but cannot reflect the characteristics of the frequency band at that time. The existing spectrum prediction processes can be roughly divided into two categories [
13], namely, by predicting the quality of the channel and predicting from the perspective of channel occupancy. The current common spectrum prediction technology methods are the regression analysis Model and Hidden Markov Model (HMM). The model based on HMM is simpler and widely used in CR. This kind of method has better accuracy in predicting spectrum occupancy even in complex industrial environments [
14]. However, the traditional HMM spectrum prediction algorithm has the problems of time extension, and the prediction accuracy is easily affected by the uncertainty of the matrix to be tested. For this reason, [
15] proposes a spectrum occupancy prediction method based on a segmented Markov model, which uses a density clustering density algorithm to predict channel states in clusters. This algorithm not only improves the prediction accuracy but also reduces the prediction uncertainty. The piecewise prediction method of the periodic channel can better describe the change characteristics of historical data in the periodic channel and improve the final prediction accuracy without changing the algorithm complexity [
16]. The statistical learning method is used to learn the historical data of spectrum perception. Under the condition of ensuring the prediction rate, based on HMM, the optimal frequency of shortwave cognition is found to accurately predict the channel state of shortwave users [
17]. In order to solve the shortcomings of the existing channel access technology in the shortwave ALE system, the three-state HMM is proposed to be used in spectrum prediction technology in [
18], which greatly improves the prediction accuracy and utilization rate in the shortwave ALE system. Due to the nature of these two algorithms in mathematical calculation, although it will be convenient to understand the prediction process, in general, in the spectrum occupancy prediction method based on HMM, the system consumption is relatively large. Furthermore, the regression method is highly complex, and is not suitable for continuous prediction. In addition, the idea of cooperative prediction can also be applied to prediction technology. For the energy-constrained cognitive radio network, the combination of the cooperative prediction model and the spectrum sensing framework can overcome the problems of the local prediction model, so as to improve the spectrum efficiency in the case of low energy consumption [
19].
At present, these prediction techniques solve the spectrum problem by developing complex algorithms, which often consume a lot of time and energy in the prediction, which hinders the success of dynamic spectrum access. Therefore, we urgently need flexible and intelligent spectrum prediction technology. The emergence of machine learning brings new hope to many fields, among which deep learning has shown good ability in the classification and prediction of some complex systems, giving researchers the idea of applying deep learning to spectrum prediction technology.
In [
20], a spectrum prediction technology based on Multilayer Perceptron (MLP) is proposed. The structural parameters of the artificial neural network are configured, and all parameters are randomly initialized and then iteratively trained. The gradient is constantly calculated, and the parameter values are updated until the iteration conditions are met. However, due to the structure of the MLP network itself, the learning speed is slow, and it is easy to fall into the situation of local extremum. The complex nonlinear processing ability of feedforward neural networks can greatly improve the learning speed and prediction accuracy in prediction, among which the BP network [
21] and its variants are common. Later, researchers proposed Recurrent Neural Networks (RNN) on the basis of BP, which were constructed by sliding windows to analyze spectrum data in a specified time. However, RNN is prone to the problem of gradient disappearance in spectrum prediction, so a spectrum prediction method based on the Deep Recurrent Neural Network (DRRN) [
22] and a series of improved network algorithms are proposed. In [
23], spectrum prediction techniques based on the Long Short-term Memory network (LSTM) and BP network were discussed. In [
24], a new spectrum prediction framework was developed for LSTM, which improved its processing of time series. In [
25], prediction errors of LSTM were compared with those of integrated moving average autoregression and delay neural networks, which proved that LSTM has certain advantages and good prediction performance in time series problems. In order to solve the long-term predictions of spectrum data, ref. [
26] proposes a convolutional LSTM model for spectrum prediction—ConvLSTM. In the 450–520 MHz frequency band, the long-term spatial-spectral-time joint prediction of the spectrum signal is carried out. The results show that the model exhibits a stable value of Root Mean Squared Error (RMSE) for different channels, but it rises with the increase in the time step. In [
27], STS-PredNet was proposed by combining ConvLSTM and the predictive recurrent neural network (PredRNN). This model shows stable RMSE performance in multi-time step spatial-temporal spectrum prediction, and effectively improves the prediction performance in different prediction ranges. For the ISM band with high burst characteristics, ref. [
28] proposed an algorithm named Classified-based Deep Reinforcement Learning (C-DRL) to quickly and efficiently complete the matching between the state and the prediction model, achieving the effect of fast convergence, simple operation and high prediction rate. In order to reduce the time cost of convolutional network prediction in multi-channel, Gao et al. use LSTM, Seq-to-Seq modeling, exploit the intrinsic correlation and time correlation between channels, and propose a multi-channel multi-step joint spectrum prediction algorithm, which can effectively improve the prediction accuracy of multi-channel [
29].
In addition, for different industrial frequency bands and different communication systems, researchers have also proposed other prediction algorithms based on LSTM networks, which contribute greatly to the research of spectrum prediction technology [
30,
31,
32,
33,
34]. Despite the high performance of LSTM, it may bring challenges in terms of computational burden and handling missing data, and since spectral data is equivalent to long time series and LSTM has reduced prediction accuracy for long sequences, we propose the B-LTF algorithm to improve the problem of long sequence prediction. It slows down the impact of sequence length change on the prediction performance of the network so that the overall spectrum prediction rate is improved. Secondly, in this paper, we not only predict the channel state but also predict the spectrum signal. The prediction of the channel state is first processed by the gate threshold and then predicted, so as to know the channel state of the node in the future time, and the signal spectrum prediction goes deeper into the deeper level, that is, the trend and amplitude change of the future signal can be known through the sparse spectrum signal data, which is of great help to the subsequent cognitive radio technology.
Since deep learning algorithms are typically exploratory, they do not have any preconditions or assumptions about the data. Therefore, in many cases, they provide higher accuracy than traditional probabilistic and statistical algorithms. Although deep learning has been very popular in recent years, it is still in the initial stage in the field of communication, especially in CR technology. Therefore, how to use deep learning algorithms to make spectrum prediction more efficient and effective is an emerging research direction. Our main contributions in this paper are summarized as follows:
We study the problem of spectrum availability prediction and discuss the time spectrum occupancy characteristics together.
We proposed the B-LTF algorithm, combined the BP network with LSTM, built a new network structure and realized the spectrum prediction from the neural network. We studied the influence of the sequence length of spectrum data and the model prediction rate in detail and effectively improved the accuracy of spectrum prediction.
We show that the analysis of simulation prediction values obtained by simulating the current channel state shows that a long short-term memory network and its improved model can effectively process the time series, the GRU model has a simpler structure and the training time of the GRU model will be significantly reduced compared with the LSTM network, and the improved B-LTF algorithm compared with the LSTM, BP and GRU has better predictive performance. In addition, when the sequence length of spectrum data increases, the model prediction rate tends to be saturated or reduced.
The rest of this paper is organized as follows. We first introduce the spectrum prediction models, deep learning models and related knowledge in
Section 2.
Section 3 briefly describes the designed B-LTF model for spectrum prediction, which is followed by simulation results and discussions in
Section 4. Finally, we conclude the paper in
Section 5.