CN115209418A

CN115209418A - Intelligent broadband spectrum sensing technology based on pre-training basic model

Info

Publication number: CN115209418A
Application number: CN202210662006.XA
Authority: CN
Inventors: 胡祝华; 李向辉
Original assignee: Hainan University
Current assignee: Hainan University
Priority date: 2022-06-13
Filing date: 2022-06-13
Publication date: 2022-10-18

Abstract

The invention provides an intelligent broadband spectrum sensing technology based on a pre-training basic model, which comprises the following steps: step 1, obtaining sampling data by using a modulation broadband converter technology; step 2, preprocessing the sampled data; and 3, inputting the data obtained in the step 2 into a time-frequency fusion adjustable depth convolution neural network (TFF _ aDCNN) model which is divided into a main network and an auxiliary network to obtain a support set perception result. The invention can obviously reduce the hardware cost and improve the reconstruction precision and reconstruction capability of the frequency spectrum support set in a complex environment with low signal-to-noise ratio; compared with the traditional deep learning method, the method can quickly obtain the migration spectrum sensing model, the model can learn the distribution rule of the spectrum more easily through the auxiliary network, and more complex electromagnetic spectrum characteristics can be learned through the fusion with the main network, so that the intelligent broadband spectrum sensing under various environmental conditions is realized.

Description

Intelligent broadband spectrum sensing technology based on pre-training basic model

Technical Field

The invention relates to the field of deep learning and spectrum sensing, in particular to an intelligent broadband spectrum sensing technology based on a pre-training basic model.

Background

Frequency spectrum is an important scarce strategic resource, and with the rapid development of the next generation wireless communication technology and the internet of things, the types and the number of devices accessed to a wireless network are increased explosively. In order to meet the requirements of various broadband devices on frequency spectrums, the 6G communication technology must advance into ultra-wide frequency spectrum ranges of millimeter waves and terahertz waves in the future. However, the current static spectrum allocation strategy makes the contradiction between low utilization rate of spectrum resources and shortage of spectrum resources increasingly sharp. Under a complex electromagnetic environment, a traditional broadband spectrum sensing method based on Cognitive Radio (CR) is difficult to meet future performance requirements, and along with the rapid development of an artificial intelligence technology, the realization of intelligent spectrum sensing in an extremely wide spectrum range by means of AI is a feasible solution.

Spectrum sensing allows Secondary Users (SUs) to sense the spectrum occupancy state of a licensed Primary User (PU) in a surrounding complex radio environment. The SU finds a spectrum cavity by using a spectrum sensing technology, and then accesses to the spectrum cavity to realize spectrum sharing, so that the spectrum utilization rate can be greatly improved, and the problem of scarcity of high-quality spectrum resources is solved. The conventional spectrum sensing method based on sub-Nyquist sampling mostly does not judge whether PU exists in a spectrum, and directly reconstructs a support set, so that the false alarm rate and the calculation cost are higher.

With the great improvement of computing power, deep learning exerts strong ability. The deep learning method can find out the mapping model of the signal and the support set in a complex radio environment without manually extracting features. Convolutional Neural Networks (CNNs) have become basic feature extraction networks in image processing due to their excellent feature extraction capabilities. There have been many studies attempting to perform spectrum sensing using the advantages of deep learning.

However, the current research (whether using the traditional or deep learning-based spectrum sensing method) still has the following three problems:

(1) The relationship between the PU's active habits and time periods is often ignored.

(2) Existing modulation-based Wideband Spectrum Sensing (WSS) algorithms for wideband converters (MWCs) require more parallel channels to guarantee the probability of support set reconstruction in low signal-to-noise ratio (SNR) environments.

(3) The reconstruction capability of the existing scheme is also greatly limited in low signal-to-noise ratio environments.

Disclosure of Invention

The invention aims to provide an intelligent broadband spectrum sensing technology based on a pre-training basic model, and provides a time-frequency fusion adjustable depth convolution neural network (TFF _ aDCNN) and a novel sensing framework based on the TFF _ aDCNN, aiming at the problem that the broadband spectrum sensing with high precision and high reconstruction capability is difficult to realize under an extremely low signal-to-noise ratio in an extremely wide frequency range at present.

The technical scheme of the invention provides an intelligent broadband spectrum sensing technology based on a pre-training basic model, which comprises the following steps:

step 1, obtaining sampling data by using a modulation broadband converter technology;

step 2, preprocessing the sampled data;

step 2.1, obtaining an estimated original signal with superposed noise and aliasing signals by using the sampling result and the observation matrix;

step 2.2, denoising and dimensionality reduction are carried out by using a principal component analysis method;

2.3, performing discrete Fourier transform on the data subjected to denoising and dimensionality reduction to obtain a frequency spectrum estimation signal;

and 2.4, performing real part and imaginary part separation on the frequency spectrum estimation signal as the input of the network.

And 3, inputting the data obtained in the step 2 into a time-frequency fusion adjustable depth convolution neural network (TFF _ aDCNN) model which is divided into a main network and an auxiliary network to obtain a support set perception result.

Further, the method in step 2 further comprises: the TFF _ apdcnn network is a two-input neural network. The network is divided into two modules, namely a main network and an auxiliary network, wherein the input of the main network is preprocessed data, and the input of the auxiliary network is independent hot coding containing time information. The input of the auxiliary network is single hot coding, and the output of the auxiliary network is obtained through full-connection layer dimension increasing and a convolution module. The main network firstly carries out a layer of convolution to extract certain signal characteristics, and then the signal characteristics are fused with the output of the auxiliary network in the channel direction. After fusion, continuously extracting features through a convolution module (three-layer convolution), and finally outputting through a full connection layer

The invention has the beneficial effects that:

(1) The invention innovatively provides a time-frequency fused adjustable depth convolution neural network (TFF _ aDCNN). The network consists of a main network and a regulating network. The former is mainly used for learning complex and abstract signal characteristics, and the latter is mainly used for assisting the main network to learn different data distribution characteristics during training and is used for regulating and controlling the emphasis direction of the main network detection during detection. The network model can obviously reduce hardware cost, improve reconstruction accuracy and reconstruction capability, and can have good performance even under very low signal-to-noise ratio.

(2) The invention can directly obtain a new broadband spectrum sensing model (fine tuning model) under different sensing environments after obtaining a basic model through training under simple distribution and then carrying out transfer learning under a complex environment, compared with the traditional deep learning method, the invention can quickly obtain the spectrum sensing model and realize intelligent sensing under various environmental conditions.

Drawings

FIG. 1 is a drawing of an abstract of an intelligent broadband spectrum sensing technology based on a pre-training basic model according to the present invention;

FIG. 2 is a schematic diagram of a cognitive radio network model in which frequency bands occupied at different time periods vary with usage probability of a device according to an embodiment of the present invention;

FIG. 3 is a schematic diagram of a network architecture in an embodiment of the invention;

FIG. 4 is a graph illustrating a comparison of recovery probabilities for a support set under different channel conditions, in accordance with an embodiment of the present invention;

FIG. 5 is a schematic diagram of probability comparison of restoration of a support set of data (m = 20) with or without PCA denoising in the embodiment of the present invention;

fig. 6 is a schematic diagram of a comparison of recovery probabilities of support sets between DLWSS and TFF _ apdcnn using data without PCA processing (m =20, n = 6) in an embodiment of the present invention;

fig. 7 is a schematic diagram of reconstruction probability of TFF _ aDCNN support set under different channel numbers in the embodiment of the present invention;

FIG. 8 is a diagram illustrating the performance of a network according to an embodiment of the present invention at different signal-to-noise ratios and different signal frequency bands;

fig. 9 is a schematic diagram of the recall and accuracy performance of the TFF _ apdcnn model at m =20 and m =25 in an embodiment of the present invention;

FIG. 10 is a ROC curve (SNR = -10dB, sigma) reflecting model classification capability in an embodiment of the invention ₁ M =20 and N = 6);

FIG. 11 is a confusion matrix (SNR = -10dB, m = -20 and σ) of TFF _ aDCNN in the embodiment of the present invention ₁ )。

Detailed Description

The idea, specific steps and technical effects of the present invention will be clearly and completely described below in conjunction with the embodiments and fig. 1 to 11 to fully understand the objects, features and effects of the present invention. It should be understood that the described embodiments are only a few embodiments of the present disclosure, rather than all embodiments, and that functional, methodological, or structural equivalents thereof, or substitutions that may be made by those skilled in the art based on the described embodiments, are intended to be within the scope of the present disclosure.

In the description of the present invention, a plurality of means is one or more, a plurality of means is two or more, and greater than, less than, more than, etc. are understood as excluding the essential numbers, and greater than, less than, etc. are understood as including the essential numbers. If the first and second are described for the purpose of distinguishing technical features, they are not to be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated or implicitly indicating the precedence of the technical features indicated.

In the description of the present invention, unless otherwise specifically limited, terms such as set, installation, connection and the like should be understood in a broad sense, and those skilled in the art can reasonably determine the specific meanings of the above terms in the present invention by combining the specific contents of the technical solutions.

As shown in fig. 1, the example provides an intelligent broadband spectrum sensing technology based on a pre-training basic model, which includes the following features:

as shown in fig. 2, a broadband spectrum sensing link in a region is considered, where a plurality of primary users PU and a secondary user SU exist. According to the actual situation, the probability of the devices used by the master user at different time periods is different, and the frequency spectrum occupation situation in different time periods in the region has regularity.

Suppose the width is B _w Is divided into L consecutive narrowband bands of width B, denoted by the band index number {1,2,3.. L }, and all primary users in the system are allowed to use at most N consecutive non-overlapping narrowband bands of width B simultaneously. The invention divides a day into z time periods in a time period T _j ∈{T ₁ ,T ₂ ...T _z The state of the kth band in (S) } is recorded as S _k ∈{0,1}，S _k =1 denotes that the kth band is being used, S _k And =0 indicates that the k-th band is free. The primary user does not occupy all available frequency bands when using the secondary user, and the secondary user can sense B _w State B of each band in range _state ＝{S ₁ ,S ₂ ...S _L The invention will belong to T _j The set of occupied frequency band numbers M in a time period is called a support set of signals and is defined as:

Λ＝{M ₁ ,M ₂ ,......M _N } (1)

the sparse multiband signal is a signal type frequently encountered in cognitive radio communication, and in the embodiment of the invention, the received signal x (t) is assumed to be a sparse band-pass analog signal with a spectrum distributed in B _w ＝[-f _nyq /2,f _nyq /2]Inner, f _nyq Is the nyquist sampling rate of the signal. The received signals are as follows:

x(t)＝p(t)+w(t) (2)

w (t) is additive white noise obeying Gaussian distribution, p (t) is superposition of signals sent by active main users at t moment in the region, and the maximum bandwidth of the signals sent by the main users is B _max The signal energy of the used frequency band is E = [ E ] ₁ ,...E _N ]。

In consideration of the fact that the liveness of different primary users is closely related to the time periods of the primary users, the occupied frequency bands are distributed differently along with the observation of the different time periods, for example, the occupied navigation frequency band is increased sharply at about 8 am in the daytime, and the frequency bands used by operators at about 9 am are very active. After 12 pm, there are no particularly active bands, but there are still primary users randomly occupying the spectrum. In addition to certain specific active frequency bands in each time period in the daytime, other frequency bands are randomly occupied by the main users, and the characteristic is very similar to the characteristic of normal distribution. Thus, in the present embodiment, carrier frequency f in daytime signal p (t) is assumed _c (T _j ) Obeying a mean value of mu (T) _j )＝{μ(T ₁ ),...μ(T _z ) }, standard deviation σ (T) _j )＝{σ(T ₁ ),...σ(T _z ) The normal distribution of the } is:

f _c (T _j )～N(μ(T _j ),σ ² (T _j )) (3)

the spectrum occupancy situation after 12 nights is random, and the carrier frequency of p (t) is uniformly distributed:

f _c (T _j )～U(-f _nyq /2,f _nyq /2) (4)

the MWC algorithm is an under-Nyquist sampling method, and the under-sampling is realized by using low-speed sampling of parallel channels. Assuming that the number of parallel sampling channels is m, on a certain channel g, the signal x (T) is multiplied by a set of periods T _p ＝1/f _p 1 randomly alternating waveform C _g (T), g is belonged to {1,2.. M }, the frequency spectrum X (f) of the signal is shifted, and then the signal passes through a cutoff frequency of 1/2T _s To obtain a frequency range of F _s ＝[-f _s /2,f _s /2]Base band signal of (c), let f _p ＝f _s The sampled signal of the low-pass filter thus contains all the characteristics of the received signal x (t). The spectrum of the signal obtained on channel g is readily available from the fourier transform and knowledge of the signal:

c _gl is a signal C _g (t) the coefficient after Fourier expansion is C _g (t) spectrum. l is the coefficient of shifting the original spectrum X (f).

The spectrum of a broadband sparse signal is finite, so the number of X (f) shifts, l, is not infinite, and f is assumed to be _p ＝f _s Thus, the number of translations L can be determined ₀ :

Number of slices L and L in system model ₀ Is L =2L ₀ +1. (5) Formula (iv) can be rewritten as:

the spectrum of the compressed sampling results Y (n) of the m parallel channels, denoted Y (f), can be written in the form of a matrix:

Y(f)＝ΦZ(f) (8)

the observation matrix Φ consists of the row vectors c of the weighting coefficients of the m sampling channels, with a size of (m, L). Z (f) is a matrix of signals over L frequency bands with a size of (L, d), d being the number of samples under-sampled per channel.

Step 2, preprocessing the sampled data;

step 2.1, obtaining an estimation original signal with noise and aliasing signal superposition by using the sampling result and the observation matrix;

from equation (8), the compressed sampling result Y (n) can obtain the estimated original signal

The spectrum of (a) is represented as:

is a pseudo-inverse of the observation matrix, and an estimated original signal with superposition of noise and aliasing signals can be obtained through the formula (10).

2.2, carrying out denoising and dimensionality reduction treatment by using a principal component analysis method;

principal Component Analysis (PCA) is a commonly used data dimension reduction method. The PCA works by finding the direction with large variance according to the data, projecting the high-dimensional features to the direction with large variance, and reserving most information to finish dimensionality reduction. The noise with smaller variance will be filtered while retaining only the features of the principal component.

The MWC undersampling is performed twice consecutively, the estimated signal obtained from equation (10) is converted to a magnitude of (L × D, 1) and expanded in the column direction to a matrix D of (L × D, 2):

solving a covariance matrix C of D:

then, the eigenvalue and eigenvector of the covariance matrix C are obtained. And (4) arranging the eigenvalues from large to small, and taking a dimension which can represent the data characteristic most, namely the direction of the eigenvector corresponding to the maximum eigenvalue. And projecting the original signal to a new dimension to obtain PCA dimension-reduced data.

The horizontal axis of D represents the characteristics, the vertical axis represents the samples, the results of two times of sampling are regarded as two dimensions of one sample, the two dimensions represent estimation signals, but the two dimensions are influenced by white Gaussian noise. And reducing the two dimensions to one dimension by utilizing PCA to obtain a characteristic which most represents the original signal, wherein the direction of a new coordinate axis found by the PCA is the direction with the largest variance and can most represent the original signal, the other dimension discarded contains most of Gaussian white noise, and the PCA reduces the one dimension to remove part of the Gaussian white noise.

data denoised by PCA

The spectrum estimation signal is obtained by discrete Fourier transform

Will be provided with

After separation of the real and imaginary parts, they are stacked in the third dimension as two signal features, resulting in an input signal I of size (L, d, 2).

The traditional compressive sensing method is characterized in that a most relevant element is selected for a signal or a residual error through an algorithm in each iteration by utilizing an MWC (wrap-around coefficient) under-sampling result through a greedy algorithm, and a support set is restored through multiple iterations:

wherein j is _k Is the sum residual vector r in the dictionary matrix V _k-1 Most relevant column index, Λ _k Is a support set, updated by iteration。

The deep learning approach also uses the results of the MWC to reconstruct the support set. Considering that in the model proposed by the present invention, the frequency band occupied by the PU is related to the time period, the present invention increases the information of the time dimension. The present invention therefore proposes a system framework, as shown in fig. 3, to map signals to spectral states. One-hot encoding uses c-state representation bits to represent N time segments, each bit representing a time segment, the bit value being either 0 or 1, where 1 represents that the current state is valid. Preprocessed data I and R ^z Is obtained by artificially dividing time periods on the time dimension

Is sent to a TFF _ aDCNN network, and the estimated signal is mapped to a frequency band state B by using a TFF _ aDCNN model _state ：

Where θ is a parameter of the network. Meanwhile, the original support set reconstruction problem is converted into a multi-label classification problem through deep learning, and the final frequency band state can be obtained only by judging the states of all frequency bands. The multi-label problem continues to reduce to a multiple single-label binary problem. And finally, solving the two-classification problem of judging the L spectrum states by adopting a Sigmoid activation function.

The invention provides a double-input convolution neural network. The two inputs to the network represent the signal dimension and the time dimension, respectively. For the input of the time dimension, when a certain H is selected, a specific state bit is 1, other state bits are 0, the output of the full-connection layer in the adjusting network is only influenced by the current state bit being 1, and the output of the full-connection network cannot be influenced by the state bit being 0. In the detection process, the main network adjusts the frequency spectrum induction of specific data distribution by inputting a specific thermal code.

The training set of the network can then be defined as:

χ＝{[(I ⁽¹⁾ ,H ⁽¹⁾ ),y ⁽¹⁾ ],…,[(I ^(w) ,H ^(w) ),y ^(w) ]} (16)

where w represents the number of samples in the training set, the input of each sample being represented by the input signal I,

and the one-hot encoding H of the time period in which the input signal is located,

and (4) forming. Assuming that the trained network parameters are theta, labels are y,

the function of the web learning is ψ. The invention adopts supervised training, finally, the output value range is limited between (0, 1) through the Sigmoid function activation of the full connection layer, and the cross entropy loss function

Comprises the following steps:

wherein y is _i For the value of the ith spectral slice in the training label, the band occupied is 1, the band free is 0, ψ _i (I|H,θ) _n A map representing the i-th spectral slice state of the network for the N-th training sample represents the probability that the band is occupied in the form of 0 to 1, N being the number of training samples. The training of the TFF _ apdcnn network is mainly data driven. During training, the TFF _ aDCNN network learns by using an Adam (Adaptive motion estimation) optimizer according to the unique hot code representing the current time period input by the regulatory network and the data conforming to the current time period distribution input by the main network.

When the one-hot code representing a certain time period is input into the regulation network, the main network inputs data with the distribution characteristics of the time period. After a large amount of data with the distribution characteristics of the current time period are learned, the network reduces the loss of the current input combination (the one-hot code of the current time period and the data conforming to the distribution characteristics of the current time period), and under the input of the current one-hot code, the main network passively learns the distribution characteristics of the input data. Extensive training makes TFF _ aDCNN more sensitive to data distributed across the current time period when TFF _ aDCNN inputs unique hot codes for the current time period in the regulatory network.

After the training is finished, the independent hot codes representing different time periods are input into the regulation and control network, and the emphasis of TFF _ aDCNN main network detection is actively controlled, so that the network can still perform spectrum sensing emphasizing current data distribution through the independent hot codes input into the regulation and control network under the condition of low signal to noise ratio, and the performance is improved. The training of the network can be seen as an optimization problem:

when different conditions are selected to input H for learning, parameters of the control network are learned according to the currently input H, so that the control network can learn corresponding parameters for H representing each time period through back propagation, and the control network can assist the main network in learning the distribution characteristics of data input by different H.

As shown in fig. 2, the proposed network model consists of two parts, a primary network and a secondary network. The main network consists of a convolution part and a full-connection part, wherein the convolution part comprises four convolution units, and each convolution unit consists of a convolution layer with the convolution step length of 1, a batch normalization layer and an activation layer adopting a ReLU function. The full-connection part is composed of a layer of full-connection network with L neuron, and the output range is limited to (0, 1) by adopting a Sigmoid activation function to represent the band state.

The input data I of the main network is connected to the convolution section. Data passing volumeNucleus K ₁ After a first convolution unit with the channel number of 16 and the channel number of 1x21 is obtained, effective information is extracted to obtain a first characteristic diagram:

since the convolution kernel is set to 16 channels, the output is

To

16 matrices, each as follows:

the values in the matrix are obtained by convolving the input matrix with a convolution kernel, and the value at each position is given by:

where ξ () represents the ReLU activation function, a and b are the values of row a and column b in a certain channel of the feature map. Because the size of the input data I is (L, d, 2), and L is the number of the spectrum slices, the first dimension of the input data I has the structure and position information of the spectrum slices, in order to ensure that the information of the data is not lost, the first dimension of the convolution kernel is set to be 1, and in the output characteristic diagram of each subsequent convolution unit, the first dimension is L and is not changed.

The auxiliary network consists of a fully connected part and a convolutional part. Considering that the auxiliary network and the main network need 16-channel feature graphs with the same dimension and shape for feature splicing, for the number matching of parameters, the full-connection part consists of a full-connection layer with one layer of neurons L x (d-20), the setting of the parameters not only meets the subsequent feature splicing, but also maps the information of low-dimensional time to higher-dimensional timeIn addition, the learnable parameters of the network are increased, and the learning capability of the network is improved. The convolution part consists of two convolution units, wherein each convolution unit consists of a convolution layer with a convolution kernel of 1x21, a channel number of 16, padding of same, a batch normalization layer and an activation layer using a ReLU activation function. The one-hot code representing the time dimension information is connected with a full connection part in the auxiliary network, data after data dimensionality increasing cannot be directly input into a convolution part, and then output data reshape of full connection is formed into a two-dimensional matrix and then connected to a convolution layer. The convolutional layer has strong capability of mining information characteristics, can mine the characteristics of time information, and finally outputs a characteristic diagram R with time information ₂ The technology of the invention splices the characteristic diagram containing the input data information and the characteristic diagram containing the time information output by the auxiliary network to form a new characteristic diagram with the channel number of 32:

splicing feature map R containing time dimension features and signal features _concatenate And inputting a second convolution unit with convolution kernel of 1x21 and channel number of 32. And the second convolution unit fuses the information of the time dimension and the information of the input data, and further extracts the relation and the characteristics between the time and the data. The convolution kernels of the third convolution unit and the fourth convolution unit are set to be 1x31 in size, the number of channels is set to be 32, the fused features are further extracted, and finally the output of the convolution part is obtained. Since the present technique uses a fully connected network for final classification, the signature graph output by the convolution portion is connected to the fully connected portion after passing through a flat layer (scatter). The fully-connected part obtains the score conditions of L frequency spectrum slices through a sigmoid activation function, wherein 1 represents that the frequency band is occupied, and 0 represents that the frequency band is free. In order not to lose information, the embodiment of the invention does not use pooling layer in convolution unit, and in order to reduce parameters between the result of the flat layer and the full connection layer, the parameter quantity is reduced as much as possible through the setting of convolution kernel and channel number, and the convolution output size of the last layer is (L, 1, 32)And (5) feature diagrams.

<xnotran> (V aclav Valenta, roman Mar ˇ s alek, genevi ` eve Baudoin, martine Villegas, martha Suarez, and Fabien Robert.Survey on spectrum utilization in europe: measurements, analyses and observations.In 2010Proceedings ofthe fifth international conference on cognitive radio oriented wireless networks and communications,pages 1-5.IEEE,2010.) (Roger B Bacchus, antoni J Fertner, cynthia S Hood, and Dennis A Roberson.Long-term, wide-band spectral monitoring in support of dynamic spectrum access networks at the iit spectrum observatory.In 2008 3rd IEEE Symposium on New Frontiers in Dynamic Spectrum Access Networks, pages 1-10.IEEE,2008.) (Tanim M Taher, roger B Bacchus, kenneth J Zdunek, and Dennis A Roberson.Long-term spectral occupancy findings in chicago.In 2011 IEEE international symposium on dynamic spectrum access networks (DySPAN), pages 100-107.IEEE,2011.) (Adnan Ahmad Cheema and Sana Salous.Spectrum occupancy measurements and analysis in 2.4 ghz wlan.Electronics,8 (9): 1011,2019.) (Yunfei Chen and Hee-Seok Oh.A survey of measurement-based spectrum occupancy modeling for cognitive radios.IEEE Communications Surveys & Tutorials, 18 (1): 848-859,2014.) , . </xnotran> Therefore, it is difficult to have a general model to learn the regularity of spectrum usage in various regions.

The model obtained by training TFF _ apdcnn is to explore the effectiveness of the proposed spectrum sensing framework and to use it as a pre-trained basic model. Obviously, the basic model cannot be used as a general and specific deep learning model, and cannot cope with broadband spectrum sensing tasks in all electromagnetic environments.

In practical application, for a specific task, firstly, spectrum measurement is carried out for at least one day on site, and the distribution rule of the authorized spectrum is found according to the result of the spectrum measurement. The distribution is then modeled, and a significant amount of training set data is generated by the simulation. Next, the training set is used for transfer learning based on the pre-trained base model. In the process of transfer learning, the technology freezes the main network in the pre-trained basic model, and only activates the adjustable auxiliary sub-network, so that a practical new model suitable for a new scene can be quickly trained, namely, the method can quickly adapt to a local spectrum use mode.

Suppose the signal sent by the primary user is:

wherein B is _max =50MHz, nyquist sampling frequency of signal f _nyq If the frequency band is not greater than 10GHz, the ultra-wideband Bw is divided into 195 continuous narrowband bands, allowing the primary users to share at most N =6 bands including the symmetric band at the same time, ei = [1,2,3 ]]. To simplify the experiment, the day is divided into z =5 time periods in the present embodiment. The first four time periods are assumed to be the radio environment in the daytime, the comparison has regularity, and the carrier frequency obeys the mean value of [0.2,0.4,0.6,0.8 = [ ]],σ ₁ ＝[0.1,0.1,0.1,0.1]The carrier frequency is specified by a normal distribution random number. The last time period is assumed that the carrier frequency obeys uniform distribution at night, and the carrier frequency is specified by uniformly distributed random numbers. The MWC sampling parameter is set to be m =25, the dimension of Y (f) finally obtained by sampling 101 data of each channel is 25x101, the dimension of an observation matrix phi is 25x195, and a signal with the dimension of 195x101x2 is obtained through data preprocessing

The bandwidth of each divided narrow-band frequency band is f _nyq L =51.28mhz, w (t) represents white gaussian noise.

The software environment of the simulation experiment is as follows, 64-bit Win10 operating system, MATLAB 2018a, tensorFlow1.15.2 and Cuda10.0/cudnn7. The hardware environment is Inter (R) Core (TM) i5-8300H CPU @2.30GHz, RAM 169B, and the GPU is NVIDIA GeForce GTX 1080 Ti.

TABLE 1 setting of simulation experiment parameters

Experiment 1: support set reconstruction performance at different signal-to-noise ratios

The model performance for different signal-to-noise ratios was compared. Data is first denoised by PCA. The signal-to-noise ratio of the received signal { x (t) } is SNR = { -10, -8, -6, -4, -2,0} db, other conditions are the same as the simulation settings. Under each SNR condition, the training set has 8,500 data, and the test set has 1,500 data. Each set of data is divided into 5 groups, each set of data obeying a particular distribution. Experiments were carried out using the proposed TFF _ aDCNN network, the TFF _ aDCNN network without a tuning network, swSOMP (Zhuhua Hu, yongBai, yaochi Zha, and Yiran Zhang. Support retrieval for multiband spread sensing networks, pages 146-159.Springer, 2017.), SOMP (journal A trunk, anna C Giert, and mark J stratum for spatial networks, part i: grdy throughput, signal processing, 86-572, and vibration processing, 13. Sub.27, track J.S. transport processing, shift, and shift J.S.S. transport processing, 14. Sub.27, and shift J.S.sub.27. Sub.27. Sub.11. Sub.27. Sub.10. Sub.27. Sub.12. Sub.7. Sub.13. Sub.7. Sub.J.S.14. Sub.J.J..

The results of the ablation experiments are shown in fig. 4. When m =25,n =6,snr = { -10, -8, -6, -4, -2,0} db, the average reconstruction probability of the proposed model is 3.5% higher than the TFF _ apdcnn model without a tuned network. The performance of the DLWSS is similar to TFF _ apdcnn without a regulated network. The average improvement rates of the proposed model were 65% and 69% compared to the traditional methods like SOMP and SwSOMP, respectively. The results also demonstrate that TFF _ aDCNN can learn the data characteristics for each time segment, so that the model can also obtain a higher support set reconstruction probability with a single encoding of the input at low signal-to-noise ratio. Meanwhile, in the embodiment of the invention, the standard deviation condition of the data distribution is set as sigma ₂ ＝[0.04,0.04,0.04,0.04]Support set reconstruction probability experiments were performed with m =20 and N =6. From FIG. 4, when the standard deviation of data is set to σ ₂ The more obvious the usage pattern of the primary user in the current time period is, the more the network can discover the pattern and learn the data distribution characteristics.

The PCA denoising enables the performance of the TFF _ aDCNN network to be greatly improved. As shown in fig. 5, at m =20,n =6, σ ₁ In the case of (3), the network support set reconstruction probability performance with or without PCA processing data is compared. As can be seen from fig. 5, the results of the ablation experiment are: PCA does not remove noise effectively when the signal-to-noise ratio is low. However, as the signal-to-noise ratio increases, the TFF _ apdcnn network performs better on the denoised data of the PCA. Compared with the data processed by PCA, the de-noising can improve the performance of network support set reconstruction, and particularly when SNR is not equal to-2 dB, the support set reconstruction rate is improved by 24.66%.

Then, the estimation signals without PCA denoising are input into a TFF _ aDCNN network and a DLWSS network, and support set reconstruction performance comparison is carried out. Data without PCA denoising contains more white gaussian noise, which is more indicative of the performance of the model itself under low snr conditions. Since the DLWSS is also designed for specific MWC parameters, the convolution filter size in the DLWSS network is adapted to the data set in the embodiment of the present invention, and the network is reproduced as in the original paper. At N =6,m =20, σ ₁ And SNR = { -10, -8, -6, -4, -2,0} dB, the experimental results are shown in FIG. 6. Experiments prove that under the condition of no PCA denoising, the TFF _ aDCNN network support set reconstruction probability provided by the technology of the invention is averagely higher than DLWSS by 9.73% under the condition of SNR = { -10, -8, -6, -4, -2,0} dB, and only in terms of the performance of the network per se under the environment with low signal to noise ratio. The SNR is improved by the data subjected to PCA denoising, and the performance of the network can be better reflected by the data without PCA processing, which strongly proves that the TFF _ aDCNN network successfully fuses the input time. This also shows that TFF _ apdcnn can learn the data distribution characteristics of each H input, and use the learned data distribution to assist the network in spectrum sensing at low signal-to-noise ratio.

Experiment 2: performance at different number of parallel channels

The more channels of the MWC, the better the model behaves. In the embodiment of the invention, the data after PCA denoising is used for researching the condition that the number of channels is [15,35 ]]When the value is taken by 2 in the interval, N =6, sigma ₁ SNR = -6dB and SNR = -8dB performance of TFF _ apdcnn network.

Through comparison, the TFF _ aDCNN network can obtain higher reconstruction probability of the support set under the conditions of low channel number and low signal-to-noise ratio. And the reconstruction precision of the SwSOMP of the traditional method is lower than 20/%. As shown in fig. 7, under the condition that SNR = -6db, n = -35, the reconstruction probability of TFF _ aDCNN reaches 90% or more, which is a high probability reconstruction performance. The method reduces the hardware requirement on the WSS, so that the network is more suitable for deployment, the hardware consumption is less, and the method is more environment-friendly.

Importantly, as shown in fig. 7, TFF _ aDCNN network and DLWSS network are also compared in terms of reconstruction accuracy of the support set. When the number of parallel channels approaches the lower theoretical limit (i.e., m =15, lower theoretical limit of 2N +1=13 (Moshe Mishaali and Yonin C electric, from the theory to practice: sub-nyquist sampling of broad bandwidth and analog signals, IEEE Journal of selected protocols in signal processing,4 (2): 375-391, 2010.) (HJ Landa. Novel sampling conditions for sampling and interpolation of theoretical functions, acMathemia, 117-52, 1967.), the performance improvement of TFF _ aDCNN is not significant compared to that of WSS, however, as the number of channels m =15, the performance improvement of TFF _ aDCNN is very large, the performance improvement of TFF _ aDCNN is very gradual, when TFF _ nN + 1% is less than that of TFaDCN, the performance improvement of TFF _ aDCnN is 1% and the performance improvement of TFaDCaDCaDCN is moderate, the performance improvement of TFF _ DLNN is 1% as compared to that of TFaDCN, the performance improvement of TFaDCaDCnN is 1.6.

Therefore, the TFF _ aDCNN model has better reconstruction performance than the same type of model and a traditional optimization model under the conditions of low signal-to-noise ratio and low hardware complexity.

Experiment 3: performance under different signal frequency bands

Then, the embodiment of the invention researches the influence of different N on the TFF _ aDCN network support set reconstruction capability. Considering the symmetric frequency bands, the number of frequency bands must be an even number, so a new data set, N =2, 4, and 6, is created in the embodiment of the present invention, and the size of the data set is the same as that of the data set in the support set reconstruction performance experiment under different signal-to-noise ratios. The signal was denoised using PCA with values of signal-to-noise ratio of { -10, -8, -6, -4, -2,0} db, number of parallel MWC channels m =20, and other parameters identical to the simulation settings. As shown in fig. 8, as the number of signal bands increases, the network performance at low signal-to-noise ratio decreases significantly.

When N =2 and the number of channels m =20, the support set reconstruction rate is still close to 90/%, when SNR = -10 dB. When N =4 and the number of channels m =20, the performance degradation is worse than when N =2, but when SNR =0dB, the support set reconstruction rate can still reach 90/%. When N =6 and the number of channels m =20, the performance is further degraded and it is impossible to obtain a high reconstruction probability lower than SNR =0 dB. The number of signal bins has a significant impact on the performance of TFF _ apdcnn.

Experiment 4: TFF _ aDCNN Structure discussion

In order to discuss the rationality of the network structure, in the embodiment of the present invention, two different adjusting network structures are compared, and the performances of the two structures in the reconstruction capability are compared, the data after PCA denoising is used, the number of parallel channels is m =25, σ is ₁ ，N＝6。

The experimental results are shown in table 2, from which it can be found that when two layers of CNNs are used as a tuning network in the TFF _ apdcnn structure proposed in the present invention, it has advantages compared to a tuning network using only one layer of CNN. The TFF _ apdcnn network proposed by the present invention has a suitable number of parameters and advantageous performance in consideration of the number of parameters and the support set reconstruction probability of the frequency band.

TABLE 2 two tuning network architectures (σ) ₁ M =25 and N = 6).

Experiment 5: TFF _ aDCNN Classification capability

For the classification capability of TFF _ apdcnn, receiver Operating Characteristic (ROC) curves and confusion matrices are used in the embodiments of the present invention for evaluation. The confusion matrix is a matrix reflecting the classification capability of the model, and since the binary classification method is used to determine the spectrum state in the embodiment of the present invention, a matrix with a size of (2, 2) can be obtained, as shown in fig. 11. The vertical 0 and 1 represent true tags, and the horizontal 0 and 1 represent tags predicted by the model, then the importance of these four regions is defined as follows. True Positive (TP), true Negative (TN), false Positive (FP) and False Negative (FN).

Since the spectrum is sparse, the positive and negative samples are disproportionate. The judgment ability of the model on positive samples needs to be described by recall rate and accuracy.

Recall＝TP/(TP+FN) (24)

Precision＝TP/(TP+FP) (25)

Wherein emph in formula (24) reflects the sensitivity of the model to the positive cases, and emph in formula (25) reflects the judgment accuracy of the model to the positive cases. However, in spectrum sensing, accuracy is not as important as recall. When the false alarm rate of the network is high, the reconstructed support set is likely to contain the true support set, which also does not affect the use of the network by the PU. Conversely, when the recall rate is low, the reconstructed support set lacks a truly occupied frequency band, which would greatly affect the normal communications of the PU.

The data sets in the support set reconstruction performance experiment under different signal-to-noise ratios are used for calculating recall rate and precision, wherein N =6 and sigma ₁ . Looking at the recalled image in FIG. 9, it can be seen that when SNR is<At-4 dB, the recall rates of the two models are similar, and the TFF _ apdcnn network proposed by the present invention can show better performance at lower SNR. This indicates that the TFF _ apdcnn network is more sensitive to band occupation at low signal-to-noise ratios and can capture the occupied band with higher probability at low signal-to-noise ratios.

According to the precision ratio data, the state of the PU can be more accurately judged when the number of channels is m =25 by TFF _ apdcnn. In the case of m =20, the accuracy performance is worse than that of the TFF _ apdcnn without adjusting the network, however, the TFF _ apdcnn can obtain a higher support set reconstruction probability in the case of the worse accuracy performance, which indicates that after the number of channels is reduced, the TFF _ apdcnn can still assist the main network to correctly judge the spectrum state by using the learned relation H between the frequency band occupation characteristic and the input. Experiments have shown that it is necessary to adjust the presence of the network.

The ROC may reflect the classification capability of the model. In the embodiment of the present invention, the spectrum sensing problem is regarded as a binary classification problem for each frequency band, and thus, as shown in fig. 10, the classification capability of the model can be reflected by using an ROC curve. In the embodiment of the invention, data under SNR = -10dB, m = -20 and N =6 are used for classification test, the first four time periods obey normal distribution, and the standard deviation sigma is ₁ ＝[0.1,0.1,0.1,0.1]. The results show that the network of the present technology has better banding state classification capability than TFF _ aDCNN without adjusting the network.

Through the experiments of the confusion matrix in fig. 11, it can be clearly found that the TFF _ apdcnn network can correctly predict more positive classes under a low signal-to-noise ratio, and the recall rate and the accuracy are higher than those of the TFF _ apdcnn network without an adjustment network. The TFF _ apdcnn model correctly predicted 738 positive cases over the TFF _ apdcnn model without the adjusted network.

Experiment 6: model migratability

In order to verify the migratability of the pre-trained basic model proposed by the present invention, a new data set is created for a new electromagnetic environment in the embodiments of the present invention. The parameters are set to SNR = [ -10, -8, -6, -4, -2,0] = db, m =20, N =6.

The pre-trained base model assumes that the sub-band activity for each time segment exhibits a normal distribution. By changing the distribution of the frequency spectrum, the embodiment of the invention also simulates an unfamiliar electromagnetic environment, namely, the single normal distribution of each time period is changed into two normal distributions of each time period, after a pre-trained basic model is loaded, the trained main network is frozen, and only model parameters of an adjustable network are trained. Therefore, a broadband spectrum sensing fine tuning model can be obtained in the embodiment of the invention, and a new electromagnetic distribution environment can be adapted through transfer learning.

Table 3 shows that the pre-trained base model performs well in a single distributed electromagnetic environment. For sensing tasks in complex electromagnetic environments, embodiments of the invention use pre-trained base models for migration learning, during which only the adjustable sub-networks are retrained. It can be seen from table 3 that the fine tuning model also has good performance. Compared to a model that is completely trained from scratch, it can be concluded that the fine-tuned model can adapt to different electromagnetic environments. Meanwhile, the pre-trained basic model has the ability of transfer learning.

TABLE 3 Performance of the three models (. Sigma.) ₁ 、m＝20、N＝6)。

Experiment 7: time overhead

The overhead of the whole system framework is simply estimated. The whole system can be divided into a MWC sampling stage, a data preprocessing stage and a TFF _ aDCNN network stage. In the embodiment of the invention, the MWC is used for sampling a signal for 100 times, and the average time is calculated. Then, the time of data preprocessing is recorded. And finally, predicting by using the trained model, and recording the time. The tensor of the prediction result is obtained.

The time cost of MWC is found to be highest by the experimental results in table 4, followed by TFF _ apdcnn network and finally by the data pre-processing section. In order to denoise the signal during the data preprocessing, two MWC samples are used in the embodiment of the present invention, which takes some time, but these times are traded for higher probability of support set reconstruction. It is noted that the IEEE802.22 standard proposes two sensing methods for the service of PU, namely fast sensing and fine sensing. The spectrum uses a band space that varies over a large time scale, and the real-time requirements on the sensing period are less stringent.

As can be seen from table 5, a fast training speed can be obtained by the transfer learning method based on the pre-training basic model, which also verifies the correctness of the theoretical analysis of the TFF _ apdcnn model. Therefore, in practical application deployment, a grounded broadband spectrum sensing model can be obtained at a higher speed.

TABLE 4 time cost analysis using m =25 and 100 iterations

TABLE 5.50 time cost analysis of training

It will be evident to those skilled in the art that the invention is not limited to the details of the foregoing illustrative embodiments, and that the present invention may be embodied in other specific forms without departing from the spirit or essential attributes thereof. The present embodiments are therefore to be considered in all respects as illustrative and not restrictive, the scope of the invention being indicated by the appended claims rather than by the foregoing description, and all changes which come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein. The details in the examples are not to be construed as limitations on the scope of the invention, and any obvious modifications, equivalent alterations, simple substitutions, etc. based on the technical solution of the present invention are intended to fall within the scope of the present invention without departing from the spirit and scope of the present invention.

Claims

1. The utility model provides an intelligence broadband frequency spectrum perception technique based on base model trains in advance which characterized in that: the method comprises the following steps:

step 2, preprocessing the sampled data;

and 2.4, performing real part and imaginary part separation on the frequency spectrum estimation signal as input of the network.

2. The method of claim 1, wherein step 2 further comprises: the TFF _ aDCNN network is a two-input neural network. The whole network is divided into: the system comprises a main network and an auxiliary network, wherein the input of the main network is preprocessed data, and the input of the auxiliary network is a one-hot code containing time information. The input of the auxiliary network is single hot coding, and the output of the auxiliary network is obtained through full-connection layer dimension increasing and a convolution module. The main network firstly extracts certain signal characteristics through a layer of convolution, and then is fused with the output of the auxiliary network in the channel direction. And after fusion, continuously extracting features through a convolution module (three-layer convolution), and finally outputting through a full connection layer.