CN109450573A - A kind of frequency spectrum sensing method based on deep neural network - Google Patents
A kind of frequency spectrum sensing method based on deep neural network Download PDFInfo
- Publication number
- CN109450573A CN109450573A CN201811540309.4A CN201811540309A CN109450573A CN 109450573 A CN109450573 A CN 109450573A CN 201811540309 A CN201811540309 A CN 201811540309A CN 109450573 A CN109450573 A CN 109450573A
- Authority
- CN
- China
- Prior art keywords
- dnn
- neural network
- indicate
- deep neural
- frequency spectrum
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04B—TRANSMISSION
- H04B17/00—Monitoring; Testing
- H04B17/30—Monitoring; Testing of propagation channels
- H04B17/382—Monitoring; Testing of propagation channels for resource allocation, admission control or handover
Abstract
The invention belongs to wireless communication technology fields, are related to a kind of frequency spectrum sensing method based on deep neural network.Method of the invention specifically includes that sampling: acquiring the observation vector at N number of moment by M member multiaerial system, is denoted as X；Obtain detection statistic: detection statistic model of the building based on deep neural network is trained using the data that step S1 is obtained, obtains trained deep neural network model, according to the observation vector of acquisition, is obtained detection statistic T (X)；Judgement: it will test statistic and be compared with threshold gamma: if T (X) > γ, being judged to the presence of primary user's signal；Otherwise, primary user's signal is judged to be not present.The present invention program can achieve the performance of optimum detector.
Description
Technical field
The invention belongs to wireless communication technology fields, are related to a kind of frequency spectrum sensing method based on deep neural network.
Background technique
The rapid development of radio communication service promotes people to have increasing demand to frequency spectrum resource, causes available
Frequency spectrum resource is more and more rare, results in " frequency spectrum crisis ".It reports and shows according to US Federal Communication Committee, traditional fixed frequency
Spectrum allocation strategy causes most of existing frequency range availability of frequency spectrum relatively low.In order to improve the availability of frequency spectrum, connect using dynamic spectrum
The cognitive radio technology for entering mechanism comes into being, its basic thought is frequency spectrum share or spectrum reuse.Therefore, frequency spectrum is detected
The frequency spectrum perception technology in hole becomes the important prerequisite and core link that cognitive radio realizes dynamic spectrum access.
The design of detection statistic directly affects the performance of frequency spectrum perception, and existing frequency spectrum perception algorithm mainly utilizes base
It is detected in the detection statistic of statistical model.One major defect of these model-driven algorithms is exactly algorithm performance height
Dependent on the accuracy of detection statistic model, once statistical model is there are uncertainty or cannot obtain, the performance of algorithm
Sharply decline or fails.
Summary of the invention
The purpose of the present invention is mentioned aiming at the above problem using the design of depth learning technology research detection statistic
A kind of testing mechanism being based on deep neural network (Deep Neural Network, DNN) out.
The technical solution adopted by the present invention are as follows:
A kind of frequency spectrum sensing method based on deep neural network, which comprises the following steps:
S1, sampling: the observation vector at N number of moment is acquired by M member multiaerial system；
S2, obtain detection statistic: detection statistic model of the building based on deep neural network is obtained using step S1
Data be trained, obtain trained deep neural network model, according to the observation vector of acquisition, obtain detection statistics
Amount；
S3, judgement: it will test statistic and be compared with threshold value: if detection statistic is greater than threshold value, being judged to primary user
Signal exists；Otherwise, primary user's signal is judged to be not present.
A kind of detection framework based on deep neural network proposed by the present invention mainly includes off-line training and online inspection
It surveys:
Sampling: multiaerial system is respectively off-line module and provides sampled data in wire module.For off-line module: more days
Linear system system provides markd sampled data, as training set；For in wire module: multiaerial system provides adopting for (unmarked)
Sample data, as test data.
Off-line training: detection statistic is designed by off-line training.Regard frequency spectrum perception as a binary classification task, that
The output of DNN can be modeled as posterior probability, so that exploitation is suitable for the cost function of frequency spectrum perception (for example, the present invention is set
The cost function based on maximum a posteriori probability of meter)；Given training set obtains trained DNN, in turn by off-line training
Based on DNN design detection statistic (for example, the likelihood ratio L based on DNN that the present invention designsDNN)；Feelings are not present in given primary user
Training set under condition is sequentially sent to trained DNN based on monte carlo method, obtains the corresponding detection system of each sample
Metering, according to the corresponding detection threshold value γ of the different false-alarm probabilities of sort method search.
On-line checking: the on-line checking mechanism based on threshold value.Sampled point is sent to trained DNN, obtains detection statistics
Amount, by comparing detection statistic and detection threshold value, obtains court verdict.With the Likelihood ration test (DNN-based based on DNN
Likelihood Ratio Test, DNN-LRT) for: if LDNN> γ is judged to the presence of primary user's signal；Otherwise, it is judged to lead
Subscriber signal is not present.
It is the general DNN detection framework of one kind that the present invention, which mentions the testing mechanism based on deep neural network, and DNN can be with
For any type of network, therefore can be extensive to heterogeneous networks.DNN detection framework of the invention is not limited to utilize frequency spectrum sense
Know problem model, any associated detection or estimation problem (such as wireless communication system Modulation Identification, signal detection and channel
The problems such as estimation) it can use.
The invention has the benefit that (1) present invention develops a kind of general DNN detection framework, DNN can be to appoint
The network for type of anticipating, therefore can be extensive to heterogeneous networks.(2) the invention proposes the Likelihood ration test scheme based on DNN, roots
According to Nai Man-Pearson came theorem, in the case where training set is sufficiently large, the present invention program can achieve the performance of optimum detector.
(3) the already present spectrum sensing scheme based on deep learning directly replaces entire detection system with DNN, and threshold value cannot be arranged and change
Become false-alarm probability.Different, DNN detection scheme proposed by the present invention develops a kind of testing mechanism based on threshold value, it can
Easily to obtain desired false-alarm probability by changing threshold value.
Detailed description of the invention
Fig. 1 shows the frequency spectrum perception frame based on model-driven；
Fig. 2 shows the present invention is based on the spectrum sensing schemes of DNN；
Fig. 3 shows that the present invention is based on the derivations of the likelihood ratio of DNN；
Fig. 4 shows the ROC performance curve of each algorithm under independent signal model；
Fig. 5 shows the ROC performance curve of each algorithm under coherent signal model.
Specific embodiment
The present invention will be described in detail with simulated example with reference to the accompanying drawing, so that those skilled in the art can be more
Understand the present invention well.
Consider a multiple antennas cognitive radio scene.As shown in Figure 1, a cognition wireless electric terminals pass through M member antenna
The N number of observation vector of system acquisition carries out frequency spectrum perception.Enable x (n)=[x1(n),x2(n),…,xM(n)]TIndicate to observe for n-th to
It measures (n=0,1 ..., N-1), wherein xm(n) it is n-th of discrete time sampling (m=1,2 ..., M) in m-th antenna.Cause
This, the frequency spectrum perception problem under multiple antennas can be expressed as a binary hypothesis test problem:
Wherein, H1And H0Respectively represent that primary user exists and primary user is there is no two kinds of hypothesis testings, s (n) indicate signal to
Amount, u (n) indicate that mean value is 0, and covariance isIndependent same distribution circle symmetric complex
(CSCG) vector,Indicate noise variance.
The present invention utilizes the detection statistic of DNN development features design data driving, proposes a kind of detection based on DNN
Frame, the frame are made of two modules of off-line training and on-line checking.
(1) off-line training: the detection statistic design based on DNN
Training set can be expressed as (Y, Z)={ (y(1),z(1)),(y(2),z(2)),…,(y(K),z(K))} (2)
Wherein, Y indicates the set of input data y, the number that y can be original sampling data or be obtained based on crude sampling
According to；It is corresponding to it, Z represents the set of label.Therefore, (y(k),z(k)) indicate k-th of sample, for frequency spectrum perception model, z(k)
=1 and z(k)=0 respectively indicates H1And H0。
It enables
Indicate the output of DNN, it is the class scores vector of one 2 × 1 dimension.Wherein, hθ() indicates under parameter θ
DNN expression formula,Indicate that DNN corresponds to hypothesis testing HiExpression formula.In this way,It indicates to correspond to HiClass
Score.
Therefore, have
Wherein P () indicates probability.The target of so DNN training is exactly to maximize likelihood:
Or log-likelihood:
This is equivalent to minimize cost function:
Based on this, the target of DNN training is the optimal parameter θ of acquisition, so that posterior probability P (Z | Y) it is maximum, that is,
Wherein, θ*Represent the optimized parameter under maximum posteriori criterion.
Based on cost function, can be finally obtained trained by the parameter of back-propagation algorithm progressive updating DNN
DNN.As shown in Fig. 2, trained DNN can be modeled as
Wherein,It indicates to take y as the trained DNN inputted,It indicates to correspond to HiClass score.Therefore,
Available two kinds are assumed corresponding posterior probability:
It is then based on Bayes' theorem, so that it may obtain conditional probability:
With
Wherein, P (Hi) indicate HiPrior probability.Specified criteria probability P (y | H1) and P (y | H0), Nai Man-Pearson came is fixed
It is likelihood ratio that reason, which demonstrates optimal detection statistic,.Therefore, the likelihood ratio based on DNN can be derived:
Wherein,
And
It is specific as shown in Figure 3.
Next, needing that detection threshold value is arranged.It enablesIndicate label H0Under single sample input, available LDNN?
H0In the case of expression formula:
It enables
Indicate H0Under data set.It is sent to trained DNN, the corresponding L of available each sampleDNN|H0Value.
These values are arranged in descending order, a set is formed, is denoted asSo, correspond toDetection threshold value can be expressed as
Wherein,Expression takes downwards nearest integer,Indicate first of element of set.
(2) on-line checking: the testing mechanism based on threshold value
For in wire module, multiaerial system collects online data untagged as test data, it is denoted asIt is sent to
Trained DNN carries out the likelihood ratio test based on DNN:
As shown in Fig. 2, once obtaining detection statistic, so that it may by itself and threshold value comparison, and then make judgement.
Convolutional neural networks (Convolutional Neural Network, CNN) is applied to proposed DNN frequency spectrum perception
Mechanism, enabling covariance matrix is the input of CNN, and a kind of frequency spectrum perception algorithm based on CNN may be implemented.Fig. 4 and Fig. 5 give respectively
The ROC curve simulation result under independent CSCG signal model and correlation CSCG signal model is gone out.Wherein, CM-CNN, E-C,
MED, ED, BCED and CAV respectively indicate based on algorithm of the invention, estimator-correlator algorithm, maximum eigenvalue detection,
Energy measuring, blind combination energy measuring and covariance absolute value algorithm.From simulation result as can be seen that based on calculation of the invention
The setting detection threshold value that method can be convenient obtains desired false-alarm probability.In addition, no matter signal is independent or correlation, it is based on this hair
Receiver performance characteristics (Receiver Operating Characteristic, ROC) curve performance of bright algorithm all close to
The performance of optimal E-C algorithm.Particularly, when false-alarm probability is 0.001,96.7% inspection is achieved based on algorithm of the invention
Survey probability, about 4 times of the detection probability beyond traditional algorithm.
Claims (2)
1. a kind of frequency spectrum sensing method based on deep neural network, which comprises the following steps:
S1, sampling: the observation vector at N number of moment is acquired by M member multiaerial system；
S2, detection statistic is obtained: detection statistic model of the building based on deep neural network, the number obtained using step S1
According to being trained, trained deep neural network model is obtained, according to the observation vector of acquisition, obtains detection statistic；
S3, judgement: it will test statistic and be compared with threshold value: if detection statistic is greater than threshold value, being judged to primary user's signal
In the presence of；Otherwise, primary user's signal is judged to be not present.
2. a kind of frequency spectrum sensing method based on deep neural network according to claim 1, which is characterized in that the step
Rapid S2's method particularly includes:
If x (n)=[x1(n),x2(n),…,xM(n)]TIndicate n-th of observation vector, n=0,1 ..., N-1, wherein xm(n) it is
In n-th of discrete time sampling of m-th of antenna, m=1,2 ..., M；Frequency spectrum perception problem under multiple antennas is set as one
Binary hypothesis test problem:
H1: x (n)=s (n)+u (n)
H0: x (n)=u (n)
Wherein, H1And H0It respectively represents primary user to exist with primary user there is no two kinds of hypothesis testings, s (n) indicates signal vector, u
(n) indicate that mean value is 0, covariance isIndependent same distribution circle symmetric complex vector,Indicate noise variance；
Form deep neural network DNN frame using two modules of off-line training and on-line checking, then:
(1) off-line training is for the detection statistic design based on DNN:
If training set are as follows:
(Y, Z)={ (y(1),z(1)),(y(2),z(2)),…,(y(K),z(K))}
Wherein, Y indicates the set of input data y, and y is original sampling data or the data that are obtained based on crude sampling；Therewith
Corresponding, Z represents the set of label, i.e. (y(k),z(k)) indicating k-th of sample, k=1,2 ..., K, K is total sample number, for
Frequency spectrum perception model, z(k)=1 and z(k)=0 respectively indicates H1And H0；
It enables
Indicate the output of DNN, it is the class scores vector of one 2 × 1 dimension；Wherein, hθ() indicates the DNN expression under parameter θ
Formula,Indicate that DNN corresponds to hypothesis testing HiExpression formula, i.e.,It indicates to correspond to HiClass score；
If the target of DNN training is to minimize cost function:
That is the target of DNN training is the optimal parameter θ of acquisition, so that posterior probability P (Z | Y) maximum:
Wherein, θ*Represent the optimized parameter under maximum posteriori criterion；
Trained DNN is obtained by the parameter of back-propagation algorithm progressive updating DNN based on cost function are as follows:
Wherein,It indicates to take y as the trained DNN inputted,It indicates to correspond to HiClass score；
It enablesIndicate label H0Under single sample input, obtain LDNNIn H0In the case of expression formula:
It enables
Indicate H0Under data set, be sent to trained DNN, obtain the corresponding L of each sampleDNN|H0Value, by these values
It arranges in descending order, forms a set, be denoted asCorrespond toDetection threshold value indicate are as follows:
Wherein,Expression takes downwards nearest integer,Indicate first of element of set；
(2) on-line checking:
Multiaerial system collection online data is denoted asIt is sent to trained DNN and carries out the likelihood ratio test based on DNN:
After obtaining detection statistic, S3 is entered step.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811540309.4A CN109450573B (en) | 2018-12-17 | 2018-12-17 | frequency spectrum sensing method based on deep neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811540309.4A CN109450573B (en) | 2018-12-17 | 2018-12-17 | frequency spectrum sensing method based on deep neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109450573A true CN109450573A (en) | 2019-03-08 |
| CN109450573B CN109450573B (en) | 2019-12-06 |
Family
ID=65559232
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811540309.4A Active CN109450573B (en) | 2018-12-17 | 2018-12-17 | frequency spectrum sensing method based on deep neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109450573B (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110213814A (en) * | 2019-07-04 | 2019-09-06 | 电子科技大学 | A kind of distributed power distributing method based on deep neural network |
| CN112101460A (en) * | 2020-09-16 | 2020-12-18 | 中国电子科技集团公司第三十六研究所 | Spectrum sensing method and device based on deep learning classification |
| CN114285701A (en) * | 2021-11-30 | 2022-04-05 | 西安电子科技大学重庆集成电路创新研究院 | Method, system, equipment and terminal for identifying transmitting power of master user |
| CN114710221A (en) * | 2022-03-21 | 2022-07-05 | 上海应用技术大学 | Frequency spectrum sensing method based on convolutional neural network |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104135327A (en) * | 2014-07-10 | 2014-11-05 | 上海大学 | Spectrum sensing method based on support vector machine |
| CN104270210A (en) * | 2014-10-14 | 2015-01-07 | 中国人民解放军总参谋部第六十三研究所 | Soft-decision spectrum sensing method based on compression non-reconstruction |
| CN107995628A (en) * | 2017-12-18 | 2018-05-04 | 北京工业大学 | A kind of cognition wireless network multi-user Cooperation frequency spectrum sensing method of deep learning |
| US20180324595A1 (en) * | 2017-05-05 | 2018-11-08 | Ball Aerospace & Technologies Corp. | Spectral sensing and allocation using deep machine learning |
-
2018
- 2018-12-17 CN CN201811540309.4A patent/CN109450573B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104135327A (en) * | 2014-07-10 | 2014-11-05 | 上海大学 | Spectrum sensing method based on support vector machine |
| CN104270210A (en) * | 2014-10-14 | 2015-01-07 | 中国人民解放军总参谋部第六十三研究所 | Soft-decision spectrum sensing method based on compression non-reconstruction |
| US20180324595A1 (en) * | 2017-05-05 | 2018-11-08 | Ball Aerospace & Technologies Corp. | Spectral sensing and allocation using deep machine learning |
| CN107995628A (en) * | 2017-12-18 | 2018-05-04 | 北京工业大学 | A kind of cognition wireless network multi-user Cooperation frequency spectrum sensing method of deep learning |
Non-Patent Citations (2)
| Title |
|---|
| MAUNIL R. VYAS等: "Artificial Neural Network Based Hybrid Spectrum Sensing Scheme for Cognitive Radio", 《2017 IEEE 28TH ANNUAL INTERNATIONAL SYMPOSIUM ON PERSONAL, INDOOR, AND MOBILE RADIOCOMMUNICATIONS (PIMRC)》 * |
| 刘畅: "认知无线电中基于空间谱的频谱感知技术研究", 《万方博硕论文》 * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110213814A (en) * | 2019-07-04 | 2019-09-06 | 电子科技大学 | A kind of distributed power distributing method based on deep neural network |
| CN110213814B (en) * | 2019-07-04 | 2021-06-08 | 电子科技大学 | Distributed power distribution method based on deep neural network |
| CN112101460A (en) * | 2020-09-16 | 2020-12-18 | 中国电子科技集团公司第三十六研究所 | Spectrum sensing method and device based on deep learning classification |
| CN114285701A (en) * | 2021-11-30 | 2022-04-05 | 西安电子科技大学重庆集成电路创新研究院 | Method, system, equipment and terminal for identifying transmitting power of master user |
| CN114285701B (en) * | 2021-11-30 | 2024-03-29 | 西安电子科技大学重庆集成电路创新研究院 | Method, system, equipment and terminal for identifying transmitting power of main user |
| CN114710221A (en) * | 2022-03-21 | 2022-07-05 | 上海应用技术大学 | Frequency spectrum sensing method based on convolutional neural network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109450573B (en) | 2019-12-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109450573A (en) | A kind of frequency spectrum sensing method based on deep neural network | |
| CN109379153A (en) | A kind of frequency spectrum sensing method | |
| CN111126471A (en) | Microseism event detection method and system | |
| CN109190537A (en) | A kind of more personage's Attitude estimation methods based on mask perceived depth intensified learning | |
| CN111562108A (en) | Rolling bearing intelligent fault diagnosis method based on CNN and FCMC | |
| CN106792465A (en) | A kind of indoor fingerprint map constructing method based on mass-rent fingerprint | |
| CN112039687A (en) | Small sample feature-oriented fault diagnosis method based on improved generation countermeasure network | |
| CN108960404A (en) | A kind of people counting method and equipment based on image | |
| CN111103325A (en) | Electronic nose signal drift compensation method based on integrated neural network learning | |
| CN104252625A (en) | Sample adaptive multi-feature weighted remote sensing image method | |
| CN110929842B (en) | Accurate intelligent detection method for non-cooperative radio signal burst time region | |
| CN112749633B (en) | Separate and reconstructed individual radiation source identification method | |
| CN113126038A (en) | High-frequency ground wave radar working frequency optimization method, system, storage medium and application | |
| Butter et al. | Classification of Fermi-LAT blazars with Bayesian neural networks | |
| CN113109782B (en) | Classification method directly applied to radar radiation source amplitude sequence | |
| CN105743756A (en) | Frame detection method based on Adaboost algorithm in Wi-Fi system | |
| CN114584230B (en) | Predictive channel modeling method based on countermeasure network and long-term and short-term memory network | |
| CN112348700B (en) | Line capacity prediction method combining SOM clustering and IFOU equation | |
| CN112616160B (en) | Intelligent short-wave frequency cross-frequency-band real-time prediction method and system | |
| CN110784887B (en) | Method for detecting number of abnormal signal sources in gridding radio signal monitoring system | |
| Foldager et al. | On the role of model uncertainties in Bayesian optimisation | |
| Cominelli et al. | Accurate Passive Radar via an Uncertainty-Aware Fusion of Wi-Fi Sensing Data | |
| Eldeeb et al. | LoRaWAN-enabled Smart Campus: The Dataset and a People Counter Use Case | |
| CN112085079B (en) | Rolling bearing fault diagnosis method based on multi-scale and multi-task learning | |
| CN116561685A (en) | Sequential anomaly detection method based on differential component |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |