CN104333424A - Frequency spectrum detection and unknown noise variance tracking estimation method and device thereof - Google Patents

Abstract

For a spectrum detection problem in the condition of dynamic unknown noise variance in the practical application, the invention discloses a dynamic state-space system model to reflect an intrinsic mechanism: an authorized user state and a time-varying noise variance are taken as two hidden states, and two-state one-order Markow and auto-regressive models are used to carry out modeling for the dynamic transfer characteristic. On this basis, a novel spectrum sensing method device is designed and provided. The method is rooted in the Bayes statistical inference theory, the marginalized particle filtering technique (in the figure) is employed, and the combined estimation of the authorized user and the noise variance can be realized. Especially, the invention discloses a marginalized particle filter two-level adaptive prediction coefficient adjustment method, the noise variance time-varying characteristic is fully utilized, and the accurate tracking of the noise variance can be realized. The obtained noise variance information is used, the algorithm device can be extended to (but not limited to) a single-node single-antenna sensing system, and good spectrum sensing performance is obtained.

Description

A kind of frequency spectrum detection and unknown noise variance follow the tracks of method of estimation and device

Technical field

Frequency spectrum detection problem under change unknown noise variance condition when the present invention is directed to, first design proposes a kind of dynamical state space system model (Dynamic State-space Model, DSM), respectively based on two state Markov state probability metastasis model and autoregression (Auto-regressive, AR) model, using authorized user state and time become noise variance as two hidden states (Hidden States); On this basis, a kind of frequency spectrum detecting method of novelty is proposed.The method is based on bayesian theory, utilize marginalisation particle filter (Marginal Partical Filtering, MPF) technology, the real-time tracking of accurate detection to authorized user state and unknown noise variance can be realized simultaneously, significantly improve the frequency spectrum perception performance under the dynamic unknown condition of noise variance.Belong to the communications field.

Background technology

Along with the fast development of Modern wireless communication technology, limited frequency spectrum resource and traditional static spectrum allocation schemes cannot meet the demand of high transfer rate communication equipment.Cognitive radio (Cognitive Radio, CR), as a kind of frequency spectrum smart share technologies, greatly can improve the availability of frequency spectrum, becomes one of hot technology being subject to extensive concern at present.

As the basic technology of cognitive radio, frequency spectrum perception detects by carrying out periodicity to mandate frequency range, finds frequency spectrum cavity-pocket, thus ensures that unauthorized user does not cause interference to authorized user while utilizing mandate frequency range.In recent years, multiple frequency spectrum perception mode is suggested, common method comprises energy measuring (Energy Detection, ED), matched filtering detects (Matched Filter Detection, MFD) and cyclostationary characteristic detection (Cyclostationary Detection).In actual applications, due to uncertainty and the time variation of noise, cognitive user is difficult to obtain real-time noise variance accurately, thus causes existing detection method detection perform significantly to decline.

In order to tackle the problems referred to above, the present invention proposes a kind of brand-new frequency spectrum sensing method.The method, based on bayesian theory, utilizes marginalisation particle filter technology, in single node single antenna sensory perceptual system, can realize the real-time tracking of accurate detection to authorized user dynamic behavior and unknown noise variance simultaneously, obtain good perceptual performance.

Summary of the invention

First the present invention proposes a kind of frequency spectrum detection dynamical state space model for becoming during noise variance under unknown condition.By authorized user operating state and time become noise variance and regard two hiding system modes as, adopt single order two state Markov model and AR Model Abstraction to become migrate attribute at that time respectively.Meanwhile, adopt non-coherent reception scheme to obtain measured value, in the time window of length-specific, carry out Cumulative sum by reception sampled signal energy and obtain observation signal.On this basis, according to edge particles filtering technique and conjugate gradient descent method concept, design proposes a kind of brand-new frequency spectrum sensing method.Especially, novelty of the present invention propose marginalisation particle filter two-stage adaptive prediction coefficient updating method, can more accurately pair time become noise variance and follow the tracks of.New departure, under the prerequisite ensureing perception algorithm requirement of real-time, greatly improves frequency spectrum perception performance, thus provides a kind of scheme having application potential for the design and implimentation of distributed cognition wireless network.

At receiving terminal, comprise authorized user state-detection and noise variance real-time tracking two modules.

Authorized user state-detection: according to current time receiving terminal observation signal and a upper moment Noise Variance Estimation value, utilize particle filter (Particle Filtering, PF) technology carries out sequential renewal to the posterior probability of authorized user operating state, thus realizes the detection to authorized user state.

Noise variance real-time tracking: export according to current time observation signal and authorized user state detection module, utilize conjugate prior concept and marginalisation thought, propose marginalisation particle filter two-stage adaptive prediction coefficient updating method, real-time tracking is realized to noise variance.

Advantage of the present invention is:

1) technical solution of the present invention is applicable to the single node frequency spectrum perception under the uncertain wireless transmission environments of noise, is cognitive radio technology application establish a firm foundation in wireless mobile communications environment;

2) dynamical state space model of design proposition, adopt HMM and the dynamic migration characteristic of autoregression model to authorized user state and noise variance to be described respectively, more effectively can reflect frequency spectrum perception inherent mechanism under wireless mobile communications environment; Adopt incoherent reception to obtain observation signal, reduce the complexity that scheme realizes.

3) new departure can realize the real-time tracking to noise variance while detecting authorized user state, thus the frequency spectrum perception performance under the uncertain wireless environment of noise is greatly improved, simultaneously for the real-time tracking of noise variance, promote unauthorized user to the cognitive ability of wireless environment, be conducive to realizing more effectively frequency spectrum access and sharing;

4) this invention makes full use of the priori transition probability information of authorized user operating state, adopt edge particles filtering technique, effectively overcome the non-stationary non-Gaussian feature that observation signal (cumlative energy) presents, and avoid the conventional particle filtering problem that the computation complexity that occurs is too high when tackling higher-dimension and detecting.

5) this invention becomes migrate attribute for during noise, proposes edge particles filtering two-stage adaptive prediction coefficient updating method, can more accurately pair time become noise variance and follow the tracks of.

6) along with incorrect noise increases, the joint detection algorithm that the present invention proposes still has excellent robustness, is thus had great advantage by tool in actual applications.

Accompanying drawing explanation

Fig. 1 is frequency spectrum perception receiving end signal processing unit block diagram.

Fig. 2 is actual noise variance and its estimated value comparison diagram

Fig. 3 is that new method frequency spectrum perception detects accuracy and traditional E D performance simulation comparison diagram.

Embodiment

The present invention sets up the frequency spectrum perception dynamical state space model under the dynamic unknown condition of noise variance, adopts edge particles filtering technique to carry out Combined estimator to change variance and authorized user state during noise simultaneously.Below dynamic system model and frequency spectrum perception process are set forth respectively.

1. the frequency spectrum perception dynamical state space model of the present invention's foundation is such as formula shown in (1) (2) (3).

$S_{x_n}=f\left(S_{x_{n-1}}\right)---\left(1\right)$

${\mathrm{\σ}}_n^2=g\left({\mathrm{\σ}}_{0:n-1}^2\right)---\left(2\right)$

y _n＝(x _n,z _n) (3)

In above formula, represent the authorized user state in n moment, shift according to specific state transition function f (.).X _nrepresent that authorized user transmits.There are two kinds of hypothesis testings in frequency spectrum perception, namely authorize frequency range idle and occupied, use H respectively ₀and H ₁represent.When authorization user signal do not exist namely authorize frequency range idle time, x _n=0; When there is authorization user signal, authorization user signal energy being normalized, both obtaining x _n=1.In the present invention, by abstract for the dynamic migration characteristic of authorized user be single order two state Markov model, namely authorized user two states shifts mutually with certain probability, and transition probability matrix (Transmission Probabilities Matrix, TPM) is such as formula shown in (4)

represent n moment noise variance.In the present invention, by noise time to become migrate attribute abstract be three rank AR models.That is:

${\mathrm{\σ}}_n^2=a_1{\mathrm{\σ}}_{n-1}^2+a_2{\mathrm{\σ}}_{n-2}^2+a_3{\mathrm{\σ}}_{n-3}^2+e_n---\left(5\right)$

E _nfor the white Gaussian noise signal that average is 0.While it is noted that, variable period T during noise variance _σbe greater than perception cycle T _s.For convenience of analyzing, further simplification and assumption T _σ=LT _s, wherein L be greater than 1 integer.

Y _nrepresent the observation signal that receives of cognitive user, for sampled signal in length-specific observation time window energy and, time window length is set to M, shown in (6).

$y_n=\left\{\begin{array}{cc}\underset{m=1}{\overset{M}{\mathrm{\Σ}}}{z}_{n,m}^2& H_0\\ \underset{m=1}{\overset{M}{\mathrm{\Σ}}}{(x_n+z_{n,m})}^2& H_1\end{array}\right.---\left(6\right)$

Wherein, M=T _s× f _sprepresent perception cycle T _sinterior sampling number, f _spfor sample frequency.Z _n=[z _{n, 1}, z _{n, 2}..., z _n,M] representing the one dimension noise vector that the independent identically distributed Gauss's sample value of M is formed, interchannel noise is average, and to be 0 variance be additive white Gaussian noise (AWGN).Corresponding to authorized user idle/work two states, observation signal y obeys center/non-central card side distribution that the degree of freedom is M respectively.

2., based on above-mentioned dynamical state space model, the present invention carries out Combined estimator to authorized user state and noise variance further.From Bayes's angle, Combined estimator obtains, shown in (7) by maximizing posterior probability.

$({\widehat{\mathrm{\σ}}}_{0:n}^2,{\hat{x}}_{0:n})=\arg \underset{x_{0:n}\∈X}{\max }\left[p({\mathrm{\σ}}_{0:n}^2,x_{0:n}|y_{0:n})\right]=\arg \underset{x_{0:n}\∈X}{\max }\left[p\left({\mathrm{\σ}}_{0:n}^2\right|x_{0:n},y_{0:n})p\left(x_{0:n}\right|y_{0:n})\right]---\left(7\right)$

The combined estimation method being obtained the present invention's proposition by above formula comprises two important component parts: be first detect authorized user state based on PF, on this basis, according to conjugate prior concept and marginalisation thought, carry out real-time tracking to noise variance.Describe in detail for above-mentioned two parts below.

1) the authorized user state sequence based on particle filter is estimated

Can be obtained by formula (6), the acquisition of measured value is nonlinear transformation operation, and the Sequential detect for authorized user state is brought serious challenge by its nonlinear and non-Gaussian characteristic undoubtedly.Approach the particle filter technology of thought based on Monte Carlo (MC, Monte-Carlo) discrete digital, effectively can process nonlinear and non-Gaussian input problem.Particle filter mainly adopts one group with weight w ⁽ⁱ⁾discrete particle x ⁽ⁱ⁾approach complicated Posterior distrbutionp p (x), namely have p (x)=∑ _iw ⁽ⁱ⁾δ (x-x ⁽ⁱ⁾).On this basis, realize n moment authorized user state based on maximum a posteriori probability (Maximum a Posteriori, MAP) criterion to estimate in real time.In specific implementation, PF mainly comprises particle and generates and upgrade respective weights two important steps.

Generating on particle essential is that this distribution is different from Posterior distrbutionp probability, is called importance function for a specific distribution sampling process.Optimum importance function is adopted, shown in (8) in this programme.

$\mathrm{\π}\left(x_n\right|x_{0:n-1}^{\left(i\right)},y_{0:n},{\widehat{\mathrm{\σ}}}_{0:n-1}^2)=\frac{p\left(y_n\right|x_n,x_{0:n-1}^{\left(i\right)},y_{0:n-1},{\widehat{\mathrm{\σ}}}_{0:n-1}^2)p\left(x_n\right|x_{0:n-1}^{\left(i\right)})}{\underset{x_n=\mathrm{0,1}}{\mathrm{\Σ}}p\left(y_n\right|x_n,x_{0:n-1}^{\left(i\right)},y_{0:n-1},{\widehat{\mathrm{\σ}}}_{0:n-1}^2)}\∝p\left(y_n\right|x_n,{\widehat{\mathrm{\σ}}}_{n-1}^2)p\left(x_n\right|x_{n-1}^{\left(i\right)})---\left(8\right)$

Upgrade the particle weights produced, shown in (9).

$w_n^{\left(i\right)}=w_{n-1}^{\left(i\right)}\frac{p\left(y_n\right|x_n^{\left(i\right)},{\widehat{\mathrm{\σ}}}_{n-1}^2)p\left(x_n^{\left(i\right)}\right|x_{n-1}^{\left(i\right)})}{\mathrm{\π}\left(x_n^{\left(i\right)}\right|x_{n-1}^{\left(i\right)},y_n,{\widehat{\mathrm{\σ}}}_{n-1}^2)}---\left(9\right)$

According to MAP criterion, authorized user state is estimated.

${\hat{x}}_n^{\left(\mathrm{MAP}\right)}=\arg \underset{x_n\∈X}{\max }\left[\underset{i=1}{\overset{P}{\mathrm{\Σ}}}{w}_n^{\left(i\right)}\mathrm{\δ}(x_n-x_n^{\left(i\right)})\right]---\left(10\right)$

2) iteration of noise variance upgrades

For the white Gaussian noise expecting known Unknown Variance, adopt inverse its conjugate gradient descent method of gamma distribution definition.Its bivariate Bayesian hierarchical approach can be write:

$z_n|{\mathrm{\σ}}_n^2~N(0,{\mathrm{\σ}}_n^2)---\left(11a\right)$

${\mathrm{\σ}}_n^2~\mathrm{iG}({\mathrm{\ϵ}}_{n|n},{\mathrm{\φ}}_{n|n})=\frac{{\mathrm{\φ}}_{n|n}^{{\mathrm{\ϵ}}_{\mathrm{nn}}}}{\mathrm{\Γ}\left({\mathrm{\ϵ}}_{n|n}\right)}{\left(\frac{1}{{\mathrm{\σ}}_n^2}\right)}^{{\mathrm{\ϵ}}_{\mathrm{nn}}+1}\exp (-\frac{{\mathrm{\φ}}_{n|n}}{{\mathrm{\σ}}_n^2})---\left(11b\right)$

Then noise variance posterior probability meets inverse gamma distribution equally:

${\mathrm{\σ}}_n^2|x_{0:n}^{\left(i\right)},y_{0:n}~\mathrm{iG}({\mathrm{\ϵ}}_{n|n}^{\left(i\right)},{\mathrm{\φ}}_{n|n}^{\left(i\right)})---\left(12\right)$

Distributed constant upgrades as follows:

${\mathrm{\ϵ}}_{n|n}^{\left(i\right)}=\left\{\begin{array}{cc}{\mathrm{\ϵ}}_{n|n-1}^{\left(i\right)}+M& x_n^{\left(i\right)}=1\\ {\mathrm{\ϵ}}_{n|n-1}^{\left(i\right)}& x_n^{\left(i\right)}=0\end{array}\right.---\left(13a\right)$

${\mathrm{\φ}}_{n|n}^{\left(i\right)}=\left\{\begin{array}{cc}{\mathrm{\φ}}_{n|n-1}^{\left(i\right)}+y_n& x_n^{\left(i\right)}=1\\ {\mathrm{\φ}}_{n|n-1}^{\left(i\right)}& x_n^{\left(i\right)}=0\end{array}\right.---\left(13b\right)$

Wherein, Prediction Parameters computational methods are as follows:

${\mathrm{\ϵ}}_{n|n-1}^{\left(i\right)}=\mathrm{\λ}{\mathrm{\ϵ}}_{n-1|n-1}^{\left(i\right)}---\left(14a\right)$

${\mathrm{\φ}}_{n|n-1}^{\left(i\right)}=\mathrm{\λ}{\mathrm{\φ}}_{n-1|n-1}^{\left(i\right)}---\left(14b\right)$

λ represents predictive coefficient.Due to variable period T during noise variance _σbe greater than perception cycle T _s, T _σ=LT _s.Therefore comprise L perception cycle in a noise variance period of change, i.e. every L perception cycle, noise variance can produce migration according to AR model in formula (5).For this system model characteristic, the present invention proposes a kind of two-stage adaptive prediction coefficient updating method of novelty.The method can according to the perception cycle at noise variance period position, adjustment predictive coefficient.Particularly, be positioned at the moment in end of term noise variance week when the perception cycle, when namely next perception periodic noise variance is about to produce migration, λ=λ ₁=0.995; Other perception cycles, λ=λ ₂=0.960.

The more new estimation that finally can obtain noise variance calculates such as formula shown in (15).

${\widehat{\mathrm{\σ}}}_n^2\≈E\left({\mathrm{\σ}}_{n|n}^2\right|x_{0:n},y_{0:n})\≈\underset{i=1}{\overset{P}{\mathrm{\Σ}}}E\left({\mathrm{\σ}}_{n|n}^2\right|x_{0:n}^{\left(i\right)},y_{0:n})w_n^{\left(i\right)}=\underset{i=1}{\overset{P}{\mathrm{\Σ}}}\frac{{\mathrm{\φ}}_{n|n}^{\left(i\right)}}{{\mathrm{\ϵ}}_{n|n}^{\left(i\right)}-1}{w}_n^{\left(i\right)}---\left(15\right)$

In sum, the present invention designs cognitive method flow chart as shown in Figure 1.

Emulate above-mentioned frequency spectrum sensing method, obtain the contrast of Noise Variance Estimation value and actual value as shown in Figure 2, frequency spectrum perception performance curve as shown in Figure 3.

Become actual time noise variance into correspondingly, actual signal to noise ratio is SNR _n, for the unknown variable to be estimated.Then initial actual signal to noise ratio can be denoted as SNR ₀, corresponding noise variance is original hypothesis signal to noise ratio is denoted as correspondingly, noise variance is denoted as snr is that [-ε, ε] obeys equally distributed random number within the specific limits.

In Fig. 2, solid line represents noise variance actual value, and dotted line represents its estimated value.Abscissa represents detecting period.

In Fig. 3, solid line and dotted line represent the detection perform of associated detecting method and traditional E D-algorithm respectively.Abscissa is actual initial signal to noise ratio snr ₀, ordinate is detection probability P _d.The initial signal to noise ratio parameter ε that floats is set to 3,5 and 10 respectively.Can obviously be found out by figure, the joint detection algorithm that the present invention proposes comparatively ED algorithm detection perform is significantly improved.Notice, along with incorrect noise increases, ED detection perform can have degeneration, and the joint detection algorithm that the present invention proposes has excellent robustness, thus also will have more advantage in actual applications simultaneously.

Claims (6)

1. the implementation method of frequency spectrum perception and a device, becomes the high-performance frequency spectrum perception in unknown noise variance situation when can realize; It is characterized in that: based on proposed dynamical state space model, effectively achieve the Combined estimator to authorized user operating state and unknown noise variance.

2. the frequency spectrum perception implementation method according to claim 1 under the dynamic unknown situation of noise variance, it is characterized in that: the system model of frequency spectrum perception mechanism under the dynamic unknown situation of a kind of deeply reflection noise variance, using authorized user state and noise variance as system hidden state, two state single order Markov models and autoregression model is adopted to describe its dynamic migration characteristic respectively.Adopt incoherent reception, by the energy of sampled signal in special time window with as systematic perspective measured value.

3. the implementation method of frequency spectrum perception under the dynamic unknown situation of noise variance according to claim 1, it is characterized in that: utilize observation signal to carry out Combined estimator to authorized user state and unknown noise variance: first to detect authorized user state based on bayesian statistics inference framework, based on obtaining frequency spectrum detecting result and observation signal, variance Posterior distrbutionp parameter is upgraded, then obtains Noise Variance Estimation value.

4. according to claim 3 detection implementation method is carried out to authorized user state, it is characterized in that: adopt particle filter technology, the particle of one group of sequential renewal and respective weights value thereof is utilized to approach authorization user signal posterior probability, and the real-time Sequential Estimation of User Status of obtaining the authorization according to maximum posteriori criterion, overcome the non-gaussian nonlinear characteristic that observation signal presents, improve frequency spectrum detection accuracy.

5. implementation method of noise variance being carried out to real-time tracking according to claim 3, it is characterized in that: adopt inverse its conjugate gradient descent method of gamma distribution definition, according to conjugate gradient descent method characteristic, noise variance posterior probability meets inverse gamma distribution equally, and then can based on the estimated value of observation signal and authorized user operating state, the Posterior distrbutionp parameter of unknown noise variance is upgraded, and using the noise variance renewal estimated value of its statistical expection as current time.

6. according to the implementation method upgraded noise variance Posterior distrbutionp parameter according to claim 5, it is characterized in that: propose two-stage adaptive prediction coefficient updating method.The method at noise variance period position, adjusts predictive coefficient λ according to the perception cycle.Particularly, be positioned at the moment in end of term noise variance week when the perception cycle, when namely next perception periodic noise variance is about to produce migration, λ=λ ₁; Other perception cycles, λ=λ ₂.λ ₁<λ ₂。