CN102130732A - Cooperative spectrum detection method for cognitive radio based on neural network - Google Patents

Abstract

The invention discloses a cooperative spectrum detection method for cognitive radio based on a neural network, which comprises the following steps that: a cognitive user in a network unit has two states, namely a training period and a working period, and the cognitive user comprises a back propagation (BP) neural network module; the cognitive user in the training period carries out local spectrum detection, trains an own BP network by comparing detection results of a central node, predicts own correct detection probability, and does not participate in fusion; after the training period is over, the cognitive user enters the working period, uses the BP network to obtain a prediction value of the correct detection probability, carries out the local spectrum detection and participates in the fusion; the central node takes the prediction value of the correct detection probability of each cognitive user as reliability, and fuses each local detection result to obtain a final detection result; and the cognitive user in the working period judges whether the BP network is required to be corrected according to a particular case. In the method, factors of hidden terminal and instability of an electromagnetic environment are taken into full consideration in combination with the neural network and cooperative spectrum detection, so that the spectrum detection performance of a cognitive radio system is improved.

Description

Cognitive radio cooperation spectrum detection method based on neural net

Technical field

The invention belongs to wireless communication technology field, relate to the cooperation spectrum detection method in a kind of cognitive radio networks.Specifically, the cooperation spectrum detection technique is a kind of key technology of cognitive radio, cognitive user will be finished preliminary separate spectrum on the one hand and detect, on the one hand will with other user collaborative cooperations, the present invention is in conjunction with frequency spectrum detection, neural net and data fusion technology, under the prerequisite that guarantees feasibility, improve the reliability that cooperation spectrum detects.

Background technology

(Cognitive Radio, notion CR) originates from Joseph doctor's Mitola of Sweden KTH in 1999 the sex work of laying a foundation to cognitive radio, provides a kind of reliable solution thinking for solving the low problem of the availability of frequency spectrum.Its core concept be cognitive user promptly time user ((Primary User, appearance PU) can insert instant available spectrum adaptively to survey the promptly main user of authorized user for Secondary User, SU) perceived spectral environment reliably on broadband.Its key technology comprises frequency spectrum detection, dynamic frequency spectrum deployment and power control.

In the frequency spectrum detection problem, detect the unstable properties problem for solving " concealed terminal ", the electromagnetic environment single cognitive user that causes such as abominable, need a plurality of cognitive user carry out cooperation spectrum and detect.The thought that cooperation spectrum detects is the environment around cognitive user is constantly surveyed, and utilizes local detection algorithm to draw testing result and is sent to Centroid, draws the judgement whether main user exists by Centroid according to suitable blending algorithm.

At present, local detection method has energy measuring, matched filtering detection, the detection of circulation spectrum signature etc., and wherein energy measuring does not need main user's prior information because of it and realizes simply being widely adopted.

The blending algorithm that cooperation spectrum detects generally include with merge or merge, most fusions etc.Cognitive user in the certain limit sends local testing result to Centroid, and Centroid judges according to above-mentioned several method whether main user exists.Wherein, can reduce false alarm probability, but require too high the detection probability of local testing result with blending algorithm; Or blending algorithm then can bring higher false alarm probability, the reduction availability of frequency spectrum; Most fusions can be regarded the compromise of preceding two kinds of methods as, but its uncertainty is too high; Thereby said method is not the most feasible cooperation spectrum detection method.

Artificial neural net (Artificial Neural Network, ANN), being called for short neural net, is the network that is formed by a large amount of processing units (neuron Neurons) interconnection, be a kind of imitation animal nerve network behavior feature, the Mathematical Modeling of carrying out the distributed parallel information processing.Neural net has very strong robustness and fault-tolerance, is good at association, summary, analogy and popularization; Have very strong self-learning capability, can produce reasonably output for new input; Be an extensive self-adaptation nonlinear dynamical system, have the ability of collective's computing.

Neuron models generally should possess three key elements: 1) have one group of cynapse or connection; 2) has the input signal accumulator that reacts biological neuron space-time integration function; 3) have an excitation function, be used to limit neuron output.A large amount of neuron models are connected to constitute artificial neural net.According to network configuration and learning algorithm, artificial neural net is divided into individual layer feedforward network (as the LMS network), multilayer feedforward network (as the BP net), feedback network (as the Hopfield net), stochastic neural net (as the Boltzmann machine), competition neural net (as the Hamming net) etc.Wherein, famous error reflections propagate algorithm (Back-Propagation algorithm, the BP algorithm) be the core of feedforward network, BP network and version thereof are because of realizing simple and functional 80%～90% the application that occupied in artificial nerve network model.

Multilayer feedforward neural network is made of input layer, hidden layer (one or more layers) and output layer.If input layer is L, L input signal arranged promptly; Output layer is Q, and Q output neuron promptly arranged; Input layer and import out two hidden layers are arranged between the layer is respectively I and J, and I and J neuron are promptly arranged respectively.Cynapse weights w between input layer and the hidden layer I _LiExpression; Cynapse weights w between hidden layer I and the hidden layer J _IjExpression; Cynapse weights w between hidden layer J and the output layer _JqExpression.

The learning process of BP algorithm can be described below:

The first step is provided with variable and parameter:

P _k=[x _K1, x _K2..., x _KL], (k=1,2 ..., N), be input vector, or claim training sample that N is the number of training sample.

Weight vector when being the n time iteration between input layer and the hidden layer I.

Weight vector when being the n time iteration between hidden layer I and the hidden layer J.

Weight vector when being the n time iteration between hidden layer J and the output layer.

Q _k(n)=[q _K1(n), q _K2(n) ..., q _KQ(n)], (k=1,2 ..., N), the actual output of network when being the n time iteration.

d _k=[d _K1, d _K2..., d _KQ], (k=1,2 .. N), is desired output.

η is a learning rate.

The second step initialization is composed and is given W _LI(0), W _IJ(0), W _JQ(0) each small random nonzero value, n=0.

The 3rd step was imported sample P at random _k

The 4th step is to input sample P _k, every layer of neuronic input signal u of forward calculation BP network and output signal v.Wherein

$v_q^Q\left(n\right)=q_{k_Q}\left(n\right),$ p＝1，2，..，P

The 5th step is by desired output d _kThe actual output Q that tries to achieve with previous step _k(n) error of calculation E (n) judges whether it meets the demands, and goes to for the 8th step if satisfy; Do not go to for the 6th step if do not satisfy.

Whether the 6th step judged n+1 greater than maximum iteration time, if greater than going to for the 8th step, otherwise, to importing sample X _kEvery layer of neuronic partial gradient δ of backwards calculation.Wherein

${\mathrm{\δ}}_q^Q\left(n\right)=Q_q\left(n\right)(1-Q_q\left(n\right))(d_q\left(n\right)-Q_q\left(n\right)),$ p＝1，2，...，P

${\mathrm{\δ}}_j^J\left(n\right)=f^{\′}\left(u_j^J\left(n\right)\right)\underset{q=1}{\overset{Q}{\mathrm{\Σ}}}{\mathrm{\δ}}_q^Q\left(n\right)w_{\mathrm{jq}}\left(n\right),$ j＝1，2，...，J

${\mathrm{\δ}}_i^I\left(n\right)=f^{\′}\left(u_i^I\left(n\right)\right)\underset{j=1}{\overset{J}{\mathrm{\Σ}}}{\mathrm{\δ}}_j^J\left(n\right)w_{\mathrm{ij}}\left(n\right),$ i＝1，2，...，I

The 7th step was calculated as follows weights correction amount w, and revised weights; N=n+1 went to for the 4th step.Wherein

${\mathrm{\Δw}}_{\mathrm{jq}}\left(n\right)={\mathrm{\η\δ}}_q^Q\left(n\right)v_j^J\left(n\right),$ w _jq(n+1)＝w _jq(n)+Δw _jq(n)

${\mathrm{\Δw}}_{\mathrm{ij}}\left(n\right)={\mathrm{\η\δ}}_j^J\left(n\right)v_i^I\left(n\right),$ w _ij(n+1)＝w _ij(n)+Δw _ij(n)

${\mathrm{\Δw}}_{\mathrm{li}}\left(n\right)={\mathrm{\η\δ}}_i^I\left(n\right)x_{\mathrm{kl}}\left(n\right),$ w _li(n+1)＝w _li(n)+Δw _li(n)

The 8th step judged whether to finish all training samples, was then to finish, otherwise went to for the 3rd step.

The present invention combines artificial neural net with data fusion, propose a kind of method that the cognitive radio cooperation spectrum detects that solves on the basis of advantage separately making full use of.

Summary of the invention

The purpose of this invention is to provide a kind of cognitive radio cooperation spectrum detection method, be used for judging at cognitive radio system whether main user exists based on neural net.The present invention is by introducing the BP neural net at the cognitive user place, the correct probability that detects of this cognitive user of next sense cycle is made reasonable prediction, thereby take into full account the unsteadiness that each cognitive user detects performance, effectively reduce the influence that single cognitive user causes because of reasons such as " concealed terminal " and electromagnetic environment are changeable.

In order to achieve the above object, the invention provides a kind of cognitive radio cooperation spectrum detection method, it is characterized in that, may further comprise the steps based on neural net:

Step 1: after the cognitive user start, enter training period, the duration is the longest to be T, and in training period, cognitive user is operated as follows:

(1A) cognitive user is used energy detection algorithm in sense cycle, obtains local testing result, and receives the testing result from Centroid;

(1B) constantly more local testing result of cognitive user and corresponding Centroid testing result, the statistical value of calculating current detection week this cognitive user correct detection probability of after date;

(1C) these probable values in chronological sequence form sequence in proper order, with length is that the sliding window of L intercepts this sequence, obtain detecting in the nearest L sense cycle correct probability, the effect of this sliding window is to store these probable values in the mode of first-in first-out, and along with the continuity of time is brought in constant renewal in;

(1D) cognitive user is training sample with the correct detection probability sequence of storage as the input vector of cognitive user BP neural net, predicted value with current detection correct detection probability during the cycle is actual output, finishing correct detection probability afterwards with the current detection cycle is desired output, the BP network is trained, error between desired output and actual output meets the demands, and training period finishes; If the error of BP network backlog demand still still finished training period when the training time surpassed T;

Step 2: after training period finished, cognitive user entered duty cycle, and in the duty cycle, cognitive user is operated as follows:

(2A) cognitive user utilizes the BP network to obtain the predicted value of correct detection probability in the current detection cycle, uses energy detection algorithm simultaneously in sense cycle, obtains local testing result;

(2B) cognitive user sends to Centroid with the predicted value of local testing result and correct detection probability;

Step 3: Centroid merges according to the detection information of following step to each cognitive user:

(3A) the predicted value normalization of Centroid each correct detection probability that will receive;

(3B) value of Centroid after with normalization merges each the local testing result that receives as the reliability of cognitive user current detection performance;

(3C) Centroid sends to all cognitive user in the network element with final detection result;

Step 4: after the current detection cycle finished, more local testing result of cognitive user and final detection result from Centroid were upgraded the value of correct detection probability, judge whether and need make correction to the BP network, and go to step 2 in next sense cycle.

Advantage of the present invention is go out the correct probability that detects of cognitive user by introducing neural network prediction on the one hand, thereby the detection performance of judgement current time cognitive user to improve the accuracy of Centroid when merging each local testing result; Training to neural net makes cognitive user have good adaptivity to environment of living on the other hand, effectively improves the detection performance of cognitive radio system.In addition, when calculating the probability of the correct detection of cognitive user, introduced forgetting factor, effectively reduced the influence of historical data, made the result more realistic.

Description of drawings

Fig. 1 is that schematic diagram is formed in the cognitive radio networks unit of using cooperation spectrum detection model of the present invention.

Fig. 2 is the operating procedure schematic flow sheet of cooperation spectrum detection method of the present invention.

Fig. 3 is a cognitive user BP neural network structure schematic diagram.

Fig. 4 is the sliding window method of work schematic diagram of constraint BP network input vector.

Fig. 5 is the emulation matched curve of actual output of BP network and desired output.

Embodiment

For making technical scheme of the present invention and advantage clearer, the present invention is described in further detail below in conjunction with drawings and the specific embodiments.

Referring to shown in Figure 1, the cognitive radio networks unit that the present invention relates to comprises a main user (PU), a Centroid and several cognitive user (SU).Based on the notion of cognitive radio, following points: 1) main user's operate as normal, under existing system, can communicate by letter without restriction; 2) cognitive user need only be allowed to access idle frequency range according to sense cycle and the continuous channel perception of detection algorithm; Whether 3) Centroid is responsible for the detection information of each cognitive user in the UNE unit, exist to judge main user; 4) cognitive radio system is transparent to main user, and main user need not make any change; 5) be two-way link between cognitive user and the Centroid; 6) based on the present invention, each cognitive user all includes the BP neural network module.

Referring to shown in Figure 2, the cognitive radio cooperation spectrum detection method that the present invention is based on neural net has following operating procedure:

Step 1: establish the cognitive user that is total to N start in the network element, these cognitive user are in two kinds of different states respectively: training period and duty cycle.If i cognitive user just started shooting, enter training period, the longest sustainable time of training period is T.In training period, cognitive user is operated as follows:

(1A) cognitive user i adopts energy measuring, in sense cycle main subscriber signal is made energy accumulation, and with the energy of accumulation and set threshold ratio, makes local judgement: if cumlative energy surpasses threshold value, then judge main user's existence, use u _i=+1 expression; Otherwise, judge that then main user does not exist, and uses u _i=-1 expression.Because Centroid also merges the detection information of other cognitive user simultaneously, so cognitive user can be received testing result u from Centroid behind the current detection end cycle ₀, with local testing result representation class seemingly, u ₀=+1 expression Centroid judges that there is u in main user ₀=-1 expression Centroid judges that main user does not exist;

After (1B) sense cycle finishes, the more local testing result u of cognitive user _iCentroid testing result u with correspondence ₀The similarities and differences, comparative result is expressed as follows with m and n: for j sense cycle, if u _i=u ₀, note m _i(j)=1, n _i(j)=0; If u _i≠ u ₀, note m _i(j)=0, n _i(j)=1.By these data, cognitive user can obtain self correct probability statistics value that detects: this cognitive user detects correct probability after j the sense cycle

Wherein r is a forgetting factor, is used for reducing gradually the influence of historical data;

(1C) probable value that obtains in the step (1B) in chronological sequence forms sequence P in proper order _i=[p _i(1), p _i(2) ... p _i(j) ...].Referring to shown in Figure 4, be that the sliding window of L intercepts this sequence and gets P with length _i'=[p _i(k-L), p _i(k-L+1) ..., p _i(k-1)], wherein k represents the current detection cycle, and the effect of this sliding window is to store the correct probability that detects in the nearest L sense cycle in the mode of first-in first-out, and brings in constant renewal in along with the variation of sense cycle;

(1D) referring to shown in Figure 3, the probability sequence P that cognitive user obtains step (1C) _i' be training sample as the input vector of cognitive user BP neural net, with the predicted value q of current detection correct detection probability during the cycle _i(k) be actual output, the correct detection probability p after finishing with the current detection cycle _i(k) be desired output, the BP network is trained.The Rule of judgment that training finishes is as follows: (a) error between desired output and the actual output meets the demands; (b) training time surpasses T.Wherein the priority of condition (b) is higher than condition (a), the error of BP network backlog demand still when even the training time surpasses T, and training period, forced to finish;

Step 2: after training period finished, cognitive user i entered duty cycle.In the duty cycle, cognitive user is operated as follows:

(2A) method that provides according to step (1C) of cognitive user obtains up-to-date probability sequence P _i', as input vector, utilize the BP network that trains to obtain output valve it, i.e. the predicted value q of correct detection probability in the current detection cycle _iCognitive user is used energy detection algorithm in sense cycle simultaneously, obtains local testing result u _i

After (2B) sense cycle finished, cognitive user was with local testing result u _iPredicted value q with correct detection probability _iSend to Centroid;

Step 3: establishing at this moment has M to be in duty cycle in N cognitive user, and Centroid merges according to the detection information of following step to this M cognitive user:

(3A) Centroid is the predicted value normalization of M cognitive user correct detection probability receiving,

(3B) the value q of Centroid after with normalization _i' as the reliability of cognitive user current detection performance, each the local testing result that receives is merged according to following method: calculate

Value, if sum＞0, Centroid judges that main user exists, i.e. u ₀=+1; If sum＜0 judges that then main user does not exist, i.e. u ₀=-1;

(3C) Centroid is with final detection result u ₀Send to all N cognitive user in the network element;

Step 4: after the current detection cycle finished, whether cognitive user needed the BP network is revised by following condition judgment: (1) more local testing result u _iWith the final detection result u that obtains by step (3C) ₀, calculate the correct detection probability value of this sense cycle correspondence with the described method of step (1B), and with the predicted value that obtains by step (2A) relatively, if error does not meet the demands, then need to revise; (2) if surpass a specified time service time of existing BP network, then no matter whether error meets the demands, and all needs to revise.BP network modification method is: movable length is that the sliding window of L is to upgrade correct detection probability sequence P in time _i' (referring to shown in Figure 4) is with P _i' be training sample as the input vector of cognitive user BP neural net, with the predicted value q of current detection correct detection probability during the cycle _i(k) be actual output, the correct detection probability p after finishing with the current detection cycle _i(k) be desired output, the BP network is trained, and go to step 2 in next sense cycle.

Through above step, the cognitive radio cooperation spectrum detection method that the present invention is based on neural net is achieved.Referring to shown in Figure 5, abscissa is the iterations of BP network, and ordinate is the correct probability that detects of cognitive user, and the predicted value of this probability is the actual output of BP network, and statistical value is the desired output of BP network.Program has been observed 10000 times iteration altogether, has simulated the course of work over a long time, for guaranteeing the clear and legible of analogous diagram, has evenly chosen 100 sampled points analogous diagram of drawing.As seen from the figure, the BP network can be predicted the correct probability that detects of cognitive user well, makes it to tally with the actual situation, thereby for Centroid provides reliable fusion tolerance, improves and detect performance.

Claims (4)

1. cognitive radio cooperation spectrum detection method based on neural net is characterized in that this method comprises following operating procedure:

Step 1: the cognitive user start, enter training period, in training period, cognitive user is operated according to following steps:

(1A) cognitive user is used energy measuring in sense cycle, obtains local testing result, and receives the testing result from Centroid;

(1B) constantly more local testing result of cognitive user and corresponding Centroid testing result obtain the statistical value of correct detection probability;

(1C) cognitive user is stored the corresponding correct detection probability of some sense cycle in the mode of first-in first-out;

(1D) cognitive user with the probable value in the step (1C) as training sample, predicted value with current detection correct detection probability in the cycle is actual output, the statistical value of the correct detection probability that obtains with the method for step (1B) after finishing with the current detection cycle is a desired output, to self with the BP neural net train, meet the demands or surpass training period in the longest time limit until the error of BP net;

Step 2: after training period finished, cognitive user entered duty cycle, and in the duty cycle, cognitive user is operated according to following steps:

(2A) the BP network that trains of cognitive user utilization obtains the predicted value of correct detection probability in the current detection cycle, and adopts energy detection algorithm, obtains local testing result;

(2B) cognitive user sends to Centroid with the predicted value of local testing result and correct detection probability;

Step 3: the predicted value normalization of each correct detection probability that Centroid will receive, reliability as the cognitive user current detection performance that is in duty cycle, each the local testing result that receives is merged, obtain final detection result, and this result is sent to all cognitive user in the network element;

Step 4: the more local testing result of cognitive user with from the final detection result of Centroid, upgrade the value of correct detection probability, judge whether and need make correction, and go to step 2 in next sense cycle to the BP network.

2. method according to claim 1 is characterized in that, step 1 further comprises following content of operation:

Use u _iRepresent local testing result, u ₀Expression is from the testing result of Centroid, wherein u _i=+1 expression is local judges that there is u in main user _i=-1 expression is local judges that there is not u in main user ₀=+1 expression Centroid judges that there is u in main user ₀=-1 expression Centroid judges that main user does not exist;

The method that cognitive user obtains the statistical value of correct detection probability is: the constantly more local testing result u of cognitive user _iCentroid testing result u with correspondence ₀, comparative result is represented with m and n, for j sense cycle, if u _i=u ₀, note m _i(j)=1, n _i(j)=0; If u _i≠ u ₀, note m _i(j)=0, n _i(j)=1, calculate that this cognitive user detects correct probability after j the sense cycle:

$p_i\left(j\right)=\frac{\underset{d=1}{\overset{j}{\mathrm{\Σ}}}{r}^{j-d}{m}_i\left(d\right)}{\underset{d=1}{\overset{j}{\mathrm{\Σ}}}{r}^{j-d}[m_i\left(d\right)+n_i\left(d\right)]}$

Wherein r is a forgetting factor, is used for reducing gradually the influence of historical data;

The method of cognitive user storage probable value is: be the sliding window intercepting P of L with length _i=[p _i(1), p _i(2) ... p _i(j) ...] must the correct probable value P that detects of the interior cognitive user of a nearest L sense cycle _i'=[p _i(k-L), p _i(k-L+1) ..., p _i(k-1)], wherein k represents the current detection cycle, and keeps upgrading.

3. method according to claim 1 is characterized in that, step 3 further comprises following content of operation:

The method for normalizing of each node correct detection probability predicted value is:

${q_i}^{\′}=q_i/\underset{j=1}{\overset{M}{\mathrm{\Σ}}}{q}_j$

Q wherein _iBe the correct detection probability predicted value of the cognitive user i that obtains of step 2, i=1,2 ..., M, M cognitive user that is in duty cycle altogether;

The method that Centroid merges each local testing result is: the value q of Centroid after with normalization _i' as the reliability of cognitive user current detection performance, calculate

Value, sum＞0 o'clock judges that main user exists, i.e. u ₀=+1; Sum＜0 o'clock judges that main user does not exist, i.e. u ₀=-1.

4. according to the method for claim 1, it is characterized in that step 4 further comprises following content of operation:

Cognitive user judges that the condition whether current BP network needs to revise is: (1) more local testing result u _iWith the final detection result u that obtains by step (3C) ₀, calculate the correct detection probability value of this sense cycle correspondence with the method for step (1B) and claim 2, and with the predicted value that obtains by step (2A) relatively, if error does not meet the demands, then need to revise; (2) if surpass a specified time service time of existing BP network, then no matter whether error meets the demands, and all needs to revise;

The method of BP network correction is: movable length is that the sliding window of L is to upgrade correct detection probability sequence P in time _i', with P _i' be training sample as the input vector of cognitive user BP neural net, serve as actual output with the predicted value of current detection correct detection probability during the cycle, the correct detection probability after finishing with the current detection cycle is a desired output, and the BP network is trained.

Images