CN107801190B - HDP-NSHMM-based spectrum sensing method - Google Patents

Abstract

The invention relates to a frequency spectrum sensing method based on HDP-NSHMM, which adopts a layered Dirichlet process-non-stationary hidden Markov model to fuse and cluster historical sensing data, and sets a larger self-transition offset parameter when the retention time of the current state is shorter in a clustering cycle so as to ensure that the state of the sensing data does not change rapidly along with the time, and reduces the self-transition offset parameter along with the increase of the retention time of the current state, thereby reducing the self-transition probability of the state and making the sensing data more likely to select to transition to different states. Compared with the prior art, the method and the device have the advantages that the self-transition probability of the state is adjusted through the self-transition offset parameter related to the cluster category holding time, the change of the channel state can be judged more accurately, the occurrence of redundant states can be avoided through the fixed state categories, the clustering accuracy of historical sensing data is improved, the sensing performance is higher, and the spectrum judgment accuracy is improved.

Description

HDP-NSHMM-based spectrum sensing method

Technical Field

The invention relates to a cognitive radio spectrum sensing technology, in particular to a spectrum sensing method based on HDP-NSHMM.

Background

The spectrum sensing is a core technology in cognitive radio, and needs to monitor surrounding wireless environment in real time, provide available spectrum resources for unauthorized users, and simultaneously ensure that the occupation of the authorized users on the current frequency band is found in time, so as to avoid causing interference. Therefore, the accuracy of spectrum sensing plays a very critical role for cognitive radio networks. A great deal of research work on spectrum sensing technology provides a plurality of detection methods based on signal processing, and the detection methods can be mainly divided into two categories of uncooperative spectrum sensing and cooperative spectrum sensing.

The cognitive radio equipment using the uncooperative spectrum sensing can independently select different detection methods, process local sensing results respectively and make spectrum judgment. The spectrum sensing method does not need information interaction among different users, so that the spectrum sensing method is simpler, quicker and more convenient to realize. Common non-cooperative detection methods mainly include energy detection, matched filter detection, cyclostationary feature detection, and the like. Although the uncooperative sensing methods do not require information interaction and are easy to implement, these detection methods are limited because spatial diversity information is not considered. Therefore, many spectrum sensing algorithms select a cooperative detection method that fuses spectrum sensing results of multiple users. The cooperative sensing can reduce uncertainty caused by noise, channel fading and the like faced by the individual cognitive radio equipment in the detection process of spectrum sensing, thereby improving the spectrum detection performance and obtaining a more accurate judgment result.

According to the selection of the cognitive radio user data fusion strategy, cooperative sensing can be mainly divided into two categories, namely centralized cooperative sensing and distributed cooperative sensing. In centralized cooperative sensing, each cognitive radio user participating in cooperative sensing sends respective sensing conditions to a fusion center through multi-hop communication, and a central base station makes a uniform spectrum judgment based on sensing results of all users. However, each user needs to perform sensing data transmission and decision-making reception with the fusion center, which brings a large communication overhead, and at the same time, has high requirements on the processing capacity and processing speed of the data fusion center. In order to reduce the overhead and the requirement on the fusion center, each user can be selected to perform local information exchange with the adjacent users, and a frequency spectrum judgment result is independently obtained without establishing a fusion center method, namely a distributed cooperative sensing method. Such a distributed network is less costly and more flexible.

A non-stationary hidden markov model (NSHMM) is a special hidden markov model with a non-static system hidden state transition probability, which is different from a conventional HMM, and the hidden state transition probability of the NSHMM is changed according to a change in state retention time. Therefore, the layered Dirichlet process-non-stationary hidden Markov model (HDP-NSHMM) is more suitable for the application scene that the transition probability of the system hidden state changes along with the system state retention time at the current moment.

Disclosure of Invention

The invention aims to overcome the defects of the prior art and provide a spectrum sensing method based on HDP-NSHMM.

The purpose of the invention can be realized by the following technical scheme:

a spectrum sensing method based on HDP-NSHMM comprises the following steps:

s1, collecting spectrum sensing observation data of all historical moments and carrying out initialization processing;

s2, fusing and clustering historical sensing data by adopting a layered Dirichlet process-non-stationary hidden Markov model, and in a clustering cycle, reducing self-transfer offset parameters when the channel state retention time is increased to be larger than a set time, so that the possibility of selectively transferring to different channel states is increased;

s3, constructing a Bayesian model by adopting gamma distribution as the conjugate prior of the index distribution parameters of the spectrum sensing data in different channel states, and updating the gamma distribution hyperparameters of the index distribution parameters of each class by utilizing all sensing data classified into the same channel state cluster;

and S4, estimating the power value of each type of channel state through the updated gamma distribution superparameter, and comparing the power estimation value with a preset threshold value to obtain a frequency spectrum judgment result.

Preferably, the initialization processing in step S1 specifically includes:

and allocating an initial channel state category to each historical moment, and counting the initial count, the length of the time sequence for keeping the clustering result as the initial channel state until each historical moment, and the start time and the end time of the keeping.

Preferably, in each clustering cycle in step S2, the following steps are performed for each time t:

s21, acquiring a channel hiding state clustering result and a fusion parameter, sampling the transition distribution beta, and acquiring statistical data of the clustering result before the current clustering cycle at the moment t;

s22, sampling to obtain a new channel state clustering result at the time t, if a new category is selected, updating the channel state clustering result at the time t and related statistical data thereof, updating transition distribution, and adding frequency spectrum sensing data corresponding to the time t into channel state fusion data of corresponding categories;

s23, comparing the new channel state clustering result with the old clustering result, if the new channel state clustering result is different from the old channel state clustering result, updating the state of the time sequence before and after t, and keeping the sequence length and the initial time sequence number;

s24, checking whether there is empty category, if there is empty category, removing empty category, reducing total category number by one, and beta_kAdding into

Wherein, beta_kRepresenting the probability of selecting a hidden state k, obeying beta to the parameters in the broken rod construction method obtained in the Dirichlet process_k|α,H～beta(1,α)，

Indicating selection of a new hidden state

The probability of (d);

and S25, sampling the auxiliary variable in the fusion parameter.

Preferably, the step S21 specifically includes:

s211, collecting channel hiding state clustering results and fusion parameters of each type of channel state at other moments except the moment t;

s212, sampling the transition distribution beta according to Dirichlet distribution;

s213, removing the clustering result before the current clustering cycle at the moment t from the related statistical data, and removing the frequency spectrum sensing data at the moment t from the fusion data of the clustering result.

Preferably, the step S22 specifically includes:

s221, respectively calculating the time sequence length of the continuous same clustering result until the time t when the existing category or the new category is selected at the time t;

s222, calculating the probability of transferring to each existing channel state category and the probability of selecting a new clustering category according to the prior distribution and the likelihood function selected by the time series clustering categories and the spectrum sensing data;

s223, sampling to obtain a new channel state clustering result at the moment t, if a new category is selected, updating the channel state clustering result at the moment t and related statistical data thereof, and updating transition distribution;

s224, adding the latest channel state clustering result at the moment t into related statistical data, and adding the spectrum sensing data corresponding to the moment t into corresponding channel state fusion data.

Preferably, the probability of transferring the existing channel state cluster category to each existing channel state category in step S222 is:

wherein, Y_tRepresenting a set of spectrally perceived observation data at time t, z_tIndicating the initial allocation of a channel state class, τ, at each time instant_tIndicates that the clustering result is kept as z until time t_tK represents the kth channel state class, K represents the total number of currently existing channel classes, K represents the self-transition offset parameter, δ (K, z)_t-1) And δ (k, z)_t+1) The impulse function is represented as a function of the impulse,

and n_k·Which respectively represent T-1, T +1 and the number of transitions of T out of state class k, alpha and beta, and all times T1, 2_kAlpha is the concentration parameter of Dirichlet in the second layer, beta is the parameter in the construction method of the broken rod obtained in the Dirichlet process_kDenotes the probability, β, of selecting the kth channel state class_k|α,H～beta(1,α)，p(Y_t；a_k,b_k) Representing a likelihood function, a_k、b_kTwo hyper-parameters representing a gamma distribution;

probability of selecting new cluster class K + 1:

wherein the content of the first and second substances,

the sequence number of the clustering result at the time of t +1 is z_t+1The probability of (a) of (b) being,

indicating selection of a new hidden state category

The probability of (c).

Preferably, the updating the transition distribution in step S223 specifically includes:

to distribution

Sampling is performed such that the probability of selecting a new cluster class K' is

Wherein

Indicating selection of a new hidden state category

The probability of (a) of (b) being,

β_kindicating the probability of selecting the kth channel state class, and K indicates the number of original hidden state classes.

Preferably, the step S4 specifically includes:

estimating the power value of each type of channel state through the updated gamma distribution hyper-parameters, comparing the power estimation value with a preset threshold value, if the power estimation value is smaller than the threshold value, the channel is considered to be available, and if the power estimation value is larger than the threshold value, the channel is considered to be occupied by the PU and cannot be used by the SU; if a plurality of available channels exist simultaneously, the channel with the smaller access power estimated value is selected.

Compared with the prior art, the method has the advantages that an access model associated with the state and duration of a channel occupied by the PU is established by utilizing the HDP-NSHMM, the state self-transition probability is adjusted through the self-transition offset parameter associated with the cluster category holding time, the change of the channel state can be judged more accurately, the occurrence of redundant states can be avoided through fixed state categories, the cluster accuracy of historical sensing data is improved, the sensing performance is higher, and the spectrum judgment accuracy is improved.

Drawings

FIG. 1 is a schematic diagram of a probability map model HDP-NSHMM;

FIG. 2 is a diagram illustrating the occupation situation of the primary user on the channel under the Markov model in the second embodiment;

FIG. 3 is a diagram illustrating the occupation of a primary user on a channel in a non-stationary Markov model according to a second embodiment;

fig. 4 is a result of channel hidden state allocation initialized when the category K is not fixed in the second embodiment;

fig. 5 shows the allocation result of the hidden state of the channel after 1 cycle when the category K is not fixed in the second embodiment;

fig. 6 shows the allocation result of the hidden states of the channels after 10 cycles when the class K is not fixed in the second embodiment;

fig. 7 shows the allocation result of the hidden states of the channel after 100 cycles when the class K is not fixed in the second embodiment;

fig. 8 shows the allocation result of the hidden channel state initialized when the category K is fixed in the second embodiment;

fig. 9 shows the allocation result of the hidden state of the channel after 1 cycle when the category K is fixed in the second embodiment;

fig. 10 shows the allocation result of the channel hidden state after 10 cycles when the category K is fixed in the second embodiment;

fig. 11 shows the allocation result of the hidden states of the channel after 100 cycles when the category K is fixed in the second embodiment;

fig. 12 is a diagram illustrating power estimation values corresponding to each type of channel state in the second embodiment;

FIG. 13 shows the actual channel availability in the second embodiment;

fig. 14 shows the result of channel state determination in the second embodiment;

FIG. 15 is a ROC contrast curve of the receiver SNR of 0dB according to the different spectrum sensing methods in the second embodiment;

FIG. 16 is a ROC comparison curve of the receiver SNR of 05dB according to different spectrum sensing methods in the second embodiment;

FIG. 17 is a ROC contrast curve of the receiver SNR of 10dB according to different spectrum sensing methods in the second embodiment;

fig. 18 shows the correct detection probability of the different spectrum sensing methods in the second embodiment when the false alarm probability is 0.2.

Detailed Description

The invention is described in detail below with reference to the figures and specific embodiments. The present embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation manner and a specific operation process are given, but the scope of the present invention is not limited to the following embodiments.

Example one

The method combines the layered Dirichlet process with the non-stationary hidden Markov model to realize the automatic clustering of the historical sensing data.

FIG. 1 illustrates a probability map model of the HDP-NSHMM method, in which the state transition probability is related to the state duration by changing the constant k to a time-varying function variable k (τ), which is a non-time stationary model. K in the rectangular box represents all τ in the box_t-1The class sequence numbers of the hidden states are all k, namely, the hidden states satisfy

The application provides a spectrum sensing method based on HDP-NSHMM, which comprises the following steps:

step 1, data initialization process

Step (11), collecting all frequency spectrum perception observation data sets { Y) of T historical moments_t|t＝1,2,...,T}；

Step (12), in a large-scale cognitive radio network spectrum sensing scene, firstly, users possibly having the same channel state distribution parameter at present are divided into the same group according to the spatial distribution of the users, and the specific method is to divide the users having the similar exponential distribution parameter lambda into the same group_kThe historical sensing signals are divided into a group, distributed spectrum sensing data fusion is carried out in the group, and a channel state class { z ] is initially allocated to each moment_t|t＝1,2,...,T}；

Step (13), counting the initial count n_ijI.e. the number of times T has transitioned from state class i to class j, at all times T1, 2_i·I.e. the number of times T has been transferred out of the state class i, i.e. all times T is 1, 2. Wherein i, j ═ 1, 2.., K denotes the number of initialized hidden state classes;

step (14) of counting τ₁,τ₂,…,τ_TWherein, τ_tIndicates that until time t, the clustering result remains z_tThe length of the time series of;

step (15), recording the starting time ts of the continuous state of each time₁,ts₂,…,ts_TAnd an end time te₁,te₂,…,te_TWherein, ts is_tAnd te_tRespectively representing the starting time and the ending time of the time sequence with continuous same clustering results where the time t is.

Step 2, fusion and clustering of historical perception data

When the retention time of the current state is short, a larger self-transition offset parameter is set to ensure that the state of the current state does not change rapidly along with the time to generate a redundancy category, and the self-transition offset parameter is reduced along with the increase of the retention time of the current state, so that the self-transition probability of the state is reduced, the state is more likely to be selected to be transitioned to different states, the accuracy of a frequency spectrum judgment result is finally improved, and in each clustering cycle, the following steps are respectively carried out at each moment t:

step (21), acquiring a channel hiding state clustering result and fusion parameters, sampling transition distribution, and acquiring statistical data of the clustering result before the current clustering cycle at the moment t;

step (211), collecting the channel hidden state clustering results { z) at other time except the time t _t1,2, T-1, T +1, T, and a fusion parameter { m } for each type of channel state_·k,a_k,b _k1, ·, K }; k denotes the kth channel state class.

Step (212), sampling the transition distribution beta according to Dirichlet distribution:

β～Dir(m_·1,m_·2,...,m_·K,γ)

step (213), assuming that the clustering result before the current clustering loop starts at time t is z'_tK 'is removed from the statistical data relating to class k', i.e. the count is to be counted

And n_k′·The number of the first and second electrodes is reduced by one,

and n_k′·Respectively representing T-1, T +1 and the times of transferring T from the state class k ', removing the spectrum sensing data at the time T from the fused data of the class k ', and simultaneously recording an old clustering result k ';

step (22), sampling to obtain a new channel state clustering result at the time t, if a new category is selected, updating the channel state clustering result at the time t and related statistical data thereof, updating transition distribution, and adding frequency spectrum sensing data corresponding to the time t into channel state fusion data of corresponding categories, specifically comprising:

step (221), respectively calculating that when K existing categories and new categories are selected at time t, the clustering results z which are continuously the same until time t are obtained_tLength of time series τ of_t(k)；

Step (222), for K current existing channel state cluster categories, the prior distribution formula selected by the time series cluster categories is as follows:

the likelihood function is formulated as:

wherein λ ═ { λ ═ λ_k|W_LT/π≤k≤W_HT/pi is the state parameter set of the channel, p (Y/lambda) represents the conditional probability of Y with known lambda, Lomax (Y (k); a)_k,b_k) Denotes Y (k) is in the hyperparameter a_kAnd b_kUnder the condition of Lomax distribution, CG (Y (k); l ═ 1, a_k,b_k) This represents a particular case where the composite gamma distribution is l 1, the component being the likelihood function of the observed data Y, W_LRepresenting the bandwidth of the low frequency signal, W_HRepresenting the bandwidth of the high frequency signal and l representing a specific parameter of the Gamma distribution.

And calculating the probability of transferring to each existing channel state class k by combining the spectrum sensing data:

wherein, Y_tRepresenting a set of spectrally perceived observation data at time t, z_tIndicating the initial allocation of a channel state class, τ, at each time instant_tIndicates that the clustering result is kept as z until time t_tK represents the kth channel state class, K represents the total number of currently existing channel classes, K represents the self-transition offset parameter, δ (K, z)_t-1) And δ (k, z)_t+1) The impulse function is represented as a function of the impulse,

and n_k·Respectively representing t-1, t +1 and all time instantsT1, 2, the number of transitions T has moved out of state class k, α and β_kTwo parameters in the construction method of the broken rod obtained in the Dirichlet process, alpha is the concentration parameter of the Dirichlet in the second layer, beta_k|α,H～beta(1,α)，p(Y_t；a_k,b_k) Representing a likelihood function, a_k、b_kTwo hyper-parameters representing a gamma distribution;

calculating the probability of selecting a new cluster class K + 1:

wherein the content of the first and second substances,

the sequence number of the clustering result at the time of t +1 is z_t+1The probability of (a) of (b) being,

representing a new hidden state class in addition to an existing hidden state

The probability of (d);

in particular, for the sensing data at the initial time, i.e. when t is 1, there are:

for the last perceptual data of the time series, i.e. when T equals T, there are:

step (223) of obtaining new channel state clustering result z at time t by sampling_t：

If a new category is selected, the value of K is increased, and the transition distribution beta is updated: to distribution

Sampling is performed such that the probability of selecting a new cluster class K' is

Wherein

Indicating selection of a new hidden state category

The probability of (a) of (b) being,

β_krepresenting the probability of selecting the kth channel state class;

step (224), adding the latest channel state clustering result at the time t into the related statistical data, namely counting

And n_k′·Adding one, and adding the frequency spectrum sensing data corresponding to the time t into the corresponding channel state fusion data;

step (23), the new channel state clustering result is compared with the old clustering result, if the new channel state clustering result is not the same as the old channel state clustering result, the secondary ts is used_t-1To te_t+1State of the time series, preserving sequence length and startA time sequence number;

and (24) checking whether an empty type exists at the moment, and if the empty type exists, checking whether n exists for all K which is 1,2_k·0 and n _·k0, if there is an empty class, remove empty class K, reduce the total number of classes K by one, and β_kAdding into

So as to improve the execution efficiency of the method;

and (25) sampling the auxiliary variable m. That is, for each channel state K1, 2,.. K in the step (12) packet q, m is initialized_qk0 denotes that the number of channel states k in the q group is 0, and the observed value Y of each channel state k is set for the group q_tSamples were taken from the following distribution:

wherein n represents the total number of channel states from 0 to n_jk-1 increasing n one by one, let x be from

If x is 1, m is the sample value taken out from the distribution_qkAnd adding 1.

Step 3, updating the distribution parameters of the historical perception data

Spectral sensing data exponential distribution parameter lambda adopting gamma distribution as different channel states_kThe bayesian model is constructed by conjugate prior:

Γ(a_k) Is represented by a_kA Gamma function as a parameter;

in the above formula, the first and second carbon atoms are,

denotes the average of the observed values of the n sampled frequency domain signals Y (k), p (λ)_k|{Y(k)}_n) Denotes λ after n observations of Y (k) are obtained_kThe posterior distribution of (1), Exponential ({ Y (k) }_n|λ_k) Denotes the exponential distribution, Gamma (lambda)_k；a_k,b_k) Denotes a gamma distribution, a'_k、b′_kDenotes a_kAnd b_kThe updated value.

Updating gamma distribution hyperparameters of exponential distribution parameters of each class with all perceptual data classified as same channel state cluster (a)_k,b_k)。

In the above formula, (1/lambda)_k) In order to be able to measure the channel conditions

Estimate of (a), E [ (1/lambda)_k)|{Y(k)}_n]Indicates that 1/lambda is obtained when n observed values Y (k) are obtained_kIs calculated from the expected value of (c).

Step 4, judging the frequency spectrum state

Step (41), after the channel state clustering and the spectrum sensing data clustering into the same class are fused, each class of channel state parameter lambda can be obtained_kUpdated distribution parameter (a)_k,b_k) The power value for each type of channel state can be estimated:

step (42), comparing the power estimation value with a preset threshold value, if the power estimation value is smaller than the threshold value, considering that the channel is available, and if the power estimation value is larger than the threshold value, considering that the channel is occupied by the PU and cannot be used by a Secondary User (SU); if a plurality of available channels exist simultaneously, the channel with the smaller access power estimated value is selected.

Example two

The embodiment verifies the clustering effect and the spectrum judgment performance of the spectrum sensing method provided by the application under the scene that the transition probability of a master user (PU) to the channel occupation and release conditions changes along with the state duration.

Considering the scene of the small cognitive radio network, the signals of the PU pass through Rayleigh channels and are attenuated according to a free space propagation model, except that the occupation and release conditions of the channels by the PU are subject to a non-stable Markov model, and the initial state transition probability is still set as p₀(0/0)＝0.975，p₀(1/0)＝0.025； p₀(0/1)＝0.05，p₀(1/1) is 0.95, and when the state retention time at the current moment is greater than the preset value, the self-transition probability of the state is reduced. In the simulation setup of this embodiment, the preset holding time is set to 10 time series lengths, and after the holding time is longer than the preset time, the state self-transition probability is reduced by the same proportion as the initial transition probability, i.e. let p be_t(0/0)＝p_t-1(0/0)·p₀(0/0) or p_t(1/1)＝p_t-1(1/1)·p₀(1/1)。

The occupation and release of the channel by the PU generated based on the non-stationary markov model with state transition probability varying with time duration and the markov model with fixed transition probability are shown in fig. 2 and 3, respectively. The abscissa is the time series sequence number, the total length is set to be T200, and the ordinate is the channel occupancy by the PU, "1" indicates occupied channel, and "0" indicates released channel. As can be seen from the figure, the non-stationary state transition probability makes the channel occupation and release duration of the PU not too long and more regular than the fixed state transition probability.

The clustering effect of the spectrum sensing data generated by the non-stationary markov model scene is verified below. Clustering is carried out on the spectrum sensing data with the time sequence length T being 200, and the signal-to-noise ratio of the SU receiver is set to be 10 dB. The parameters of the clustering algorithm are initialized as follows: α -2, γ -2, α and γ are two specific parameters of the Dirichlet distribution, the initial number of states K-4, a-l-2,

a. l, b are three specific parameters of the Gamma distribution,

the number of state classes K representing the average value of the spectral perception observations at all time instants is variable. Considering the perception scene of a single user to two channels, the occupation conditions of two PUs to the two channels are consistent with the same non-stationary Markov model. The hidden state clustering conditions of the spectrum sensing data are shown in fig. 4-7, and are channel hidden state distribution results after initialization, 1-time circulation, 10-time circulation and 100-time circulation in sequence. Since the state self-transition shift parameter κ (τ) will decrease with state duration, the probability of selecting a new class increases, thereby more easily creating redundant states and affecting the accuracy of the clustering. Therefore, on the premise of knowing the hidden channel state category, the clustering category number K can be fixed, the probability that the sensing data at each moment belongs to each existing category K is calculated, sampling is carried out from the K existing categories, and the clustering result is determined.

For the simulation scenarios of two PUs, the number of classes is fixed to K-4, and the channel hidden state allocation results after initialization, 1 cycle, 10 cycles, and 100 cycles are shown in fig. 8 to 11. After the cycle is completed, the calculated power estimation value of each class is shown in fig. 12, where the abscissa represents the cluster type and the ordinate represents the power estimation value S_k. In fig. 12, category 1 indicates that the first channel is available and the second channel is not available, denoted as (0, 1); class 2 indicates that both channels are available, denoted (0, 0); class 3 indicates that neither channel is available, denoted (1, 1); category 4 tableThe first channel is not available and the second channel is available, denoted (1, 0). Four available cases (0,0), (0,1), (1,0), (1,1) of two channels are respectively recorded as channel state 1, state 2, state 3, and state 4. The channel state decision result and the actual channel availability are shown in fig. 13 and 14, respectively.

The estimated value of the channel power is compared with different preset threshold values, and an ROC curve formed by the correct detection probability and the false alarm probability can be drawn. The fixed category clustering and spectrum judging method based on the HDP-NSHMM and the Sticky HDP-HMM method of fixed self-transfer offset parameters (the patent 'a cooperative spectrum sensing method based on multi-user historical sensing data mining', application publication No. CN 106972899A), the HDP-HMM method without the self-transfer offset parameters, and the performance of the methods of energy detection, matched filter detection and cyclostationary feature detection under the same receiver signal-to-noise ratio are compared. FIGS. 15-17 are ROC curve comparisons for receiver signal-to-noise ratios of 0dB, 5dB, and 10dB, respectively. Fig. 18 is a comparison graph of the correct detection probability of each spectrum sensing method under different signal-to-noise ratios with the false alarm probability of 0.2.

When the signal-to-noise ratio is 0dB, 5dB and 10dB, the correct detection probability of the method at the false alarm probability of 0.2 is respectively improved by 52 percent, 31 percent and 9 percent compared with the energy detection. Under the condition of higher signal-to-noise ratio, the correct detection probability when the false alarm probability is 0.2 is lower than that of matched filter detection and cyclostationary feature detection, the matched filter detection needs to know the priori knowledge of PU transmission signals in advance, the cyclostationary feature detection system is higher in complexity, needs longer detection time and has certain limitation in practical application. Under the condition of a low signal-to-noise ratio, the historical data is used for clustering and fusing the spectrum sensing data with high uncertainty, so that the method has the highest correct detection probability at the false alarm probability of 0.2. When the signal-to-noise ratio is 0dB, the correct detection probability of the method at the position where the false alarm probability is 0.2 is improved by 15% and 31% respectively compared with the matched filter detection and the cyclostationary feature detection. The static HDP-HMM method uses a fixed self-transition offset parameter in the clustering process, the HDP-HMM algorithm does not add the self-transition offset parameter, cannot distinguish the self-transition probability of the state, is not suitable for an application scene with the state transition probability changing along with the duration, and easily causes wrong clustering results under the condition of low signal-to-noise ratio, so that wrong data fusion is carried out, and the correct detection probability is lower under the condition of the same false alarm probability. The method utilizes the flexible self-transfer offset parameter, so that the change of the channel state is more accurately judged, the fixed state category can avoid the occurrence of redundant state, the clustering accuracy of historical sensing data is improved, and the sensing performance is higher. When the signal-to-noise ratio is 0dB, 5dB and 10dB, the correct detection probability of the method at the false alarm probability of 0.2 is respectively improved by 63%, 18% and 4% compared with that of a Sticky HDP-HMM method.

Claims (1)

1. A spectrum sensing method based on HDP-NSHMM is characterized by comprising the following steps:

s1, collecting the spectrum sensing observation data of all historical moments, carrying out initialization processing,

s2, using layered Dirichlet process-non-stationary hidden Markov model to fuse and cluster the historical sensing data, in the clustering cycle, when the channel state keeping time is increased to be larger than the set time, reducing the self-transfer offset parameter to increase the possibility of selectively transferring to different channel states,

s3, constructing a Bayesian model by adopting gamma distribution as the conjugate prior of the index distribution parameters of the spectrum sensing data of different channel states, updating the gamma distribution hyperparameters of the index distribution parameters of each class by utilizing all the sensing data classified into the same channel state cluster,

s4, estimating the power value of each type of channel state through the updated gamma distribution superparameter, and comparing the power estimation value with a preset threshold value to obtain a frequency spectrum judgment result;

the initialization processing of step S1 specifically includes:

allocating an initial channel state category to each historical moment, and counting the initial count, the length of a time sequence for keeping the clustering result as the initial channel state until each historical moment, and the start time and the end time of the keeping;

in step S2, in each clustering cycle, the following steps are performed for each time t:

s21, obtaining the channel hiding state clustering result and the fusion parameter, sampling the transition distribution beta, obtaining the statistical data of the clustering result before the current clustering cycle at the moment t,

s22, sampling to obtain new channel state clustering result at time t, if new category is selected, updating the channel state clustering result at time t and related statistical data, updating transition distribution, adding the spectrum sensing data corresponding to time t into the channel state fusion data of corresponding category,

s23, comparing the new channel state clustering result with the old clustering result, if not, updating the time sequence state before and after t, keeping the sequence length and the initial time sequence number,

s24, checking whether there is empty category, if there is empty category, removing empty category, reducing total category number by one, and beta_kAdding into

Wherein, beta_kRepresenting the probability of selecting a hidden state k, obeying beta to the parameters in the broken rod construction method obtained in the Dirichlet process_k|α,H～beta(1,α)，

Indicating selection of a new hidden state

The probability of (a) of (b) being,

s25, sampling auxiliary variables in the fusion parameters;

the step S21 specifically includes:

s211, collecting the channel hidden state clustering results of other time except the time t and the fusion parameters of each type of channel state,

s212, sampling the transition distribution beta according to Dirichlet distribution,

s213, removing the clustering result before the current clustering cycle at the moment t from the related statistical data, and removing the frequency spectrum sensing data at the moment t from the fusion data of the clustering result;

the step S22 specifically includes:

s221, respectively calculating the time sequence length of the continuous same clustering result until t time when the existing category or the new category is selected at the time t,

s222, calculating the probability of transferring to each existing channel state category and the probability of selecting a new clustering category according to the prior distribution and the likelihood function selected by the time series clustering categories and the spectrum sensing data,

s223, sampling to obtain new channel state clustering result at the moment t, if a new category is selected, updating the channel state clustering result at the moment t and related statistical data thereof, updating transition distribution,

s224, adding the latest channel state clustering result at the moment t into related statistical data, and adding the spectrum sensing data corresponding to the moment t into corresponding channel state fusion data;

the probability of transferring the existing channel state cluster category to each existing channel state category in step S222 is:

wherein, Y_tRepresenting a set of spectrally perceived observation data at time t, z_tIndicating the initial allocation of a channel state class, τ, at each time instant_tIndicates that the clustering result is kept as z until time t_tK represents the kth channel state class, K represents the total number of currently existing channel classes, K represents the self-transition offset parameter, δ (K, z)_t-1) And δ (k, z)_t+1) The impulse function is represented as a function of the impulse,

and n_k·Which respectively represent T-1, T +1 and the number of transitions of T out of state class k, alpha and beta, and all times T1, 2_kAlpha is a centralized parameter of the Dirichlet in the second layer, beta is a parameter in the construction method of the broken rod obtained in the Dirichlet process_kDenotes the probability, β, of selecting the kth channel state class_k|α,H～beta(1,α)，p(Y_t；a_k,b_k) Representing a likelihood function, a_k、b_kTwo hyper-parameters representing the gamma distribution,

probability of selecting new cluster class K + 1:

wherein the content of the first and second substances,

the sequence number of the clustering result at the time of t +1 is z_t+1The probability of (a) of (b) being,

indicating selection of a new hidden state category

The probability of (d);

the updating of the transition distribution in step S223 specifically includes:

to distribution

Sampling is performed such that the probability of selecting a new cluster class K' is

Wherein

Indicating selection of a new hidden state category

The probability of (a) of (b) being,

β_krepresenting the probability of selecting the kth channel state category, wherein K represents the number of the original hidden state categories;

the step S4 specifically includes:

estimating the power value of each type of channel state through the updated gamma distribution hyper-parameters, comparing the power estimation value with a preset threshold value, if the power estimation value is smaller than the threshold value, the channel is considered to be available, and if the power estimation value is larger than the threshold value, the channel is considered to be occupied by the PU and cannot be used by the SU; if a plurality of available channels exist simultaneously, the channel with the smaller access power estimated value is selected.

Images