CN109450573B - frequency spectrum sensing method based on deep neural network - Google Patents

Abstract

The invention belongs to the technical field of wireless communication, and relates to a frequency spectrum sensing method based on a deep neural network. The method mainly comprises the following steps: sampling: acquiring observation vectors at N moments through an M-element multi-antenna system, and recording the observation vectors as X; acquiring a detection statistic: constructing a detection statistic model based on the deep neural network, training by adopting the data obtained in the step S1 to obtain the trained deep neural network model, and obtaining a detection statistic T (X) according to the obtained observation vector; and (4) judging: comparing the detection statistic to a threshold γ: if T (X) is larger than gamma, judging that a main user signal exists; otherwise, judging that the main user signal does not exist. The scheme of the invention can achieve the performance of the optimal detector.

Description

frequency spectrum sensing method based on deep neural network

Technical Field

the invention belongs to the technical field of wireless communication, and relates to a frequency spectrum sensing method based on a deep neural network.

Background

the rapid development of wireless communication services has prompted people to have greater and greater demands for spectrum resources, which leads to the increasing scarcity of available spectrum resources, resulting in "spectrum crisis". According to the report of the federal communications commission in the united states, the conventional fixed spectrum allocation strategy causes the utilization rate of most of the existing frequency band spectrum to be low. In order to improve the utilization rate of frequency spectrum, a cognitive radio technology adopting a dynamic frequency spectrum access mechanism is developed, and the basic idea of the technology is frequency spectrum sharing or frequency spectrum multiplexing. Therefore, the spectrum sensing technology for detecting spectrum holes becomes an important precondition and a core link for realizing dynamic spectrum access of cognitive radio.

The design of the detection statistic directly influences the performance of spectrum sensing, and the existing spectrum sensing algorithm mainly utilizes the detection statistic based on a statistical model to carry out detection. One of the main disadvantages of these model-driven algorithms is that the performance of the algorithm is highly dependent on the accuracy of the test statistic model, and once the statistic model is uncertain or unavailable, the performance of the algorithm drops dramatically or fails.

disclosure of Invention

The present invention aims to solve the above problems, and provides a Deep Neural Network (DNN) based detection mechanism by utilizing a Deep learning technique to study the design of detection statistics.

the technical scheme adopted by the invention is as follows:

A spectrum sensing method based on a deep neural network is characterized by comprising the following steps:

s1, sampling: acquiring observation vectors at N moments through an M-element multi-antenna system;

s2, obtaining detection statistics: constructing a detection statistic model based on the deep neural network, training by adopting the data obtained in the step S1 to obtain a trained deep neural network model, and obtaining detection statistic according to the obtained observation vector;

S3, judging: comparing the detection statistic to a threshold: if the detection statistic is larger than the threshold value, judging that a main user signal exists; otherwise, judging that the main user signal does not exist.

the invention provides a detection framework based on a deep neural network, which mainly comprises offline training and online detection:

sampling: the multi-antenna system provides sampling data for the offline module and the online module, respectively. For the offline module: the multi-antenna system provides marked sampling data as a training set; for online modules: the multi-antenna system provides (unmarked) sampled data as test data.

Off-line training: the detection statistics are designed by off-line training. Considering spectrum sensing as a binary classification task, the output of DNN can be modeled as a posterior probability, so as to develop a cost function suitable for spectrum sensing (for example, the cost function based on the maximum posterior probability designed by the present invention); giving a training set, obtaining a trained DNN through off-line training, and designing detection statistics (for example, a DNN-based likelihood ratio LDNN designed by the invention) based on the DNN; and giving a training set under the condition that a master user does not exist, sequentially sending the training set into the trained DNN based on a Monte Carlo method to obtain the detection statistics corresponding to each sample, and searching for detection threshold values gamma corresponding to different false alarm probabilities according to a sorting method.

online detection: threshold-based online detection mechanisms. And sending the sampling point to the trained DNN to obtain detection statistic, and comparing the detection statistic with a detection threshold value to obtain a judgment result. Taking DNN-based Likelihood Ratio Test (DNN-based Likelihood Ratio Test, DNN-LRT) as an example: if the LDNN is more than gamma, judging that a main user signal exists; otherwise, judging that the main user signal does not exist.

the detection mechanism based on the deep neural network is a universal DNN detection framework, and the DNN can be any type of network, so that different networks can be generalized. The DNN detection framework of the present invention is not limited to the use of the spectrum sensing problem model, and any detection or estimation problems associated therewith (e.g., modulation identification, signal detection, and channel estimation problems in wireless communication systems) may be used.

The invention has the beneficial effects that: (1) the invention develops a universal DNN detection framework, and the DNN can be any type of network, so that different networks can be generalized. (2) The invention provides a DNN-based likelihood ratio detection scheme, and the scheme of the invention can achieve the performance of an optimal detector under the condition that a training set is large enough according to the Nehmann-Pearson theorem. (3) Existing deep learning-based spectrum sensing schemes directly replace the entire detection system with DNN, and cannot set a threshold to change the false alarm probability. In contrast, the DNN detection scheme proposed by the present invention develops a threshold-based detection mechanism that can easily achieve the desired false alarm probability by changing the threshold.

drawings

FIG. 1 illustrates a model-driven based spectrum sensing framework;

FIG. 2 illustrates a DNN-based spectrum sensing scheme of the present invention;

FIG. 3 illustrates DNN-based likelihood ratio derivation of the present invention;

FIG. 4 shows ROC performance curves for each algorithm under an independent signal model;

FIG. 5 shows ROC performance curves for various algorithms under a correlation signal model.

Detailed Description

the invention is described in detail below with reference to the drawings and simulation examples so that those skilled in the art can better understand the invention.

consider a multi-antenna cognitive radio scenario. As shown in fig. 1, a cognitive radio terminal collects N observation vectors through an M-ary antenna system to perform spectrum sensing. Let x (N) ═ x1(N), x2(N), …, xm (N) ] T denote the nth observation vector (N ═ 0,1, …, N-1), where xm (N) is the nth discrete-time sample at the mth antenna (M ═ 1,2, …, M). Therefore, the spectrum sensing problem under multiple antennas can be expressed as a binary hypothesis testing problem:

H1 and H0 represent two hypothesis tests of existence of a main user and absence of the main user respectively, s (n) represents a signal vector, u (n) represents an independent homographic Circular Symmetry Complex Gaussian (CSCG) vector with a mean value of 0 and a covariance of 0, and represents a noise variance.

The invention designs data-driven detection statistics by utilizing DNN development characteristics, and provides a DNN-based detection framework which consists of an offline training module and an online detection module.

(1) off-line training: DNN-based detection statistic design

The training set may be expressed as (Y, Z) { (Y (1), Z (1)), (Y (2), Z (2)), …, (Y (k), Z (k)) } (2)

where Y represents a set of input data Y, which may be raw sampled data or data derived based on raw samples; in correspondence, Z represents a set of tags. Therefore, (y (k), z (k)) denotes the kth sample, and for the spectrum sensing model, z (k) 1 and z (k) 0 denote H1 and H0, respectively.

Order to

represents the output of DNN, which is a class score vector of dimension 2 x 1. Where h θ (·) represents the expression of DNN under the parameter θ, indicating that DNN corresponds to the expression of hypothesis test Hi. Thus, a class score corresponding to Hi is represented.

Thus, there are

Where P (-) represents the probability. Then the goal of DNN training is to maximize likelihood:

Or log-likelihood:

This is equivalent to minimizing the cost function:

based on this, the goal of DNN training is to obtain the optimal parameter θ such that the posterior probability P (Z | Y) is maximized, that is,

And theta represents the optimal parameter under the maximum posterior probability criterion.

based on the cost function, the DNN parameters can be updated step by step through a back propagation algorithm, and finally the trained DNN is obtained. As shown in fig. 2, the trained DNN may be modeled as

here, the trained DNN with y as an input is indicated, and the class score corresponding to Hi is indicated. Thus, the posterior probabilities for two hypotheses can be obtained:

then based on Bayes theorem, the conditional probability can be obtained:

And

where P (Hi) represents the prior probability of Hi. Given the conditional probabilities P (y | H1) and P (y | H0), the Neemann-Pearson theorem demonstrates that the optimal detection statistic is a likelihood ratio. Thus, a DNN-based likelihood ratio can be derived:

wherein the content of the first and second substances,

and also

As shown in particular in figure 3.

Next, a detection threshold needs to be set. Let a single sample input under label H0 be represented, the expression of LDNN in the case of H0 can be derived:

order to

representing the data set under H0. The DNN is sent to the trained DNN, and the value of LDNN | H0 corresponding to each sample can be obtained. The values are arranged in descending order to form a set, and the corresponding detection threshold value can be expressed as the value

Where the expression takes the nearest integer down, the l-th element of the set is expressed.

(2) Online detection: threshold-based detection mechanism

for the online module, the multi-antenna system collects online label-free data as test data, and records that the test data is sent to a trained DNN for DNN-based likelihood ratio test:

once the detection statistic is obtained, it can be compared to a threshold to make a decision, as shown in fig. 2.

A Convolutional Neural Network (CNN) is applied to the proposed DNN spectrum sensing mechanism, and the covariance matrix is an input of the CNN, so that a spectrum sensing algorithm based on the CNN can be implemented. Fig. 4 and 5 show ROC curve simulation results under an independent CSCG signal model and a related CSCG signal model, respectively. Wherein, CM-CNN, E-C, MED, ED, BCED and CAV respectively represent an algorithm, an estimator-correlator algorithm, maximum eigenvalue detection, energy detection, blind combination energy detection and covariance absolute value algorithm based on the invention. As can be seen from the simulation results, the detection threshold can be conveniently set to obtain the expected false alarm probability based on the algorithm of the invention. In addition, whether the signals are independent or correlated, the Receiver Operating Characteristic (ROC) curve performance based on the algorithm of the invention is close to the performance of the optimal E-C algorithm. In particular, when the false alarm probability is 0.001, the algorithm based on the invention achieves the detection probability of 96.7 percent, which is about 4 times higher than that of the traditional algorithm.

Claims (1)

1. a spectrum sensing method based on a deep neural network is characterized by comprising the following steps:

s1, sampling: acquiring observation vectors at N moments through an M-element multi-antenna system, and dividing the observation vectors into marked sampling data and unmarked sampling data, wherein the marked sampling data is used as a training set, and the unmarked sampling data is used as test data;

s2, obtaining detection statistics: constructing a detection statistic model based on the deep neural network, training by adopting the training set obtained in the step S1 to obtain a trained deep neural network model, and obtaining detection statistic according to the obtained test data; the specific method comprises the following steps:

Let x (N) ([ x1(N), x2(N), …, xm (N), …, xm (N)) ] T denote the nth observation vector, N ═ 0,1, …, N-1, where xm (N) is the nth discrete time sample at the mth antenna, M ═ 1,2, …, M; the spectrum sensing problem under multiple antennas is set as a binary hypothesis testing problem:

H:x(n)＝s(n)+u(n)

H:x(n)＝u(n)

H1 and H0 represent that a main user exists and a main user does not exist in two hypothesis tests respectively, s (n) represents a signal vector, u (n) represents that the mean value is 0, and an independent homodistributed circular symmetric complex Gaussian vector with covariance represents noise variance;

Forming a deep neural network DNN framework by adopting two modules of offline training and online detection, and then:

(1) Offline training is used for DNN-based detection statistic design:

The training set is set as follows:

(Y,Z)＝{(y,z),(y,z),…,(y,z),…,(y,z)}

Wherein Y represents a set of input data Y, Y being either raw sample data or a training set derived based on raw samples; correspondingly, Z represents a set of labels, i.e. (y) (K), Z (K)) represents the kth sample, K is 1,2, …, K is the total number of samples, and for the spectrum sensing model, Z (K) 1 and Z (K) 0 represent H1 and H0, respectively;

Order to

represents the output of DNN, which is a class score vector of 2 x1 dimensions; wherein h θ (·) represents a DNN expression under the parameter θ, representing that DNN corresponds to an expression of hypothesis testing Hi, i.e., representing a class score corresponding to Hi;

let the goal of DNN training be to minimize the cost function:

That is, the goal of DNN training is to obtain the optimal parameter θ such that the posterior probability P (Z | Y) is maximized:

Wherein theta represents the optimal parameter under the maximum posterior probability criterion;

based on the cost function, gradually updating the DNN parameters through a back propagation algorithm to obtain the trained DNN as follows:

Wherein, the trained DNN with y as input is represented, and the class score corresponding to Hi is represented;

let a single sample input under the label H0 be represented, resulting in an expression for the DNN-based likelihood ratio LDNN in the case of H0:

Order to

and (3) representing a data set under H0, sending the data set to a trained DNN to obtain the corresponding LDNN | H0 values of each sample, arranging the values in a descending order to form a set, and representing the detection threshold value corresponding to the false alarm probability phi as:

wherein, the expression takes the nearest integer downwards and represents the ith element of the set;

(2) Online detection:

Test data collected by the multi-antenna system is recorded to be sent to a trained DNN for DNN-based likelihood ratio test:

After the detection statistic is obtained, the process proceeds to step S3;

s3, judging: comparing the detection statistic to a threshold: if the detection statistic is larger than the threshold value, judging that a main user signal exists; otherwise, judging that the main user signal does not exist.