CN102665221A - Cognitive radio frequency spectrum perception method based on compressed sensing and BP (back-propagation) neural network - Google Patents
Cognitive radio frequency spectrum perception method based on compressed sensing and BP (back-propagation) neural network Download PDFInfo
- Publication number
- CN102665221A CN102665221A CN2012100808404A CN201210080840A CN102665221A CN 102665221 A CN102665221 A CN 102665221A CN 2012100808404 A CN2012100808404 A CN 2012100808404A CN 201210080840 A CN201210080840 A CN 201210080840A CN 102665221 A CN102665221 A CN 102665221A
- Authority
- CN
- China
- Prior art keywords
- neural net
- frequency spectrum
- cognitive radio
- compression sampling
- output
- 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.)
- Pending
Links
Images
Landscapes
- Radar Systems Or Details Thereof (AREA)
Abstract
The invention discloses a cognitive radio frequency spectrum perception method based on compressed sensing and a BP (back-propagation) neural network. The method of the invention comprises a training process and an actual detection process. The training process comprises: carrying out compression sampling whose rate is lower than a Nyquist sampling rate on an original signal of cognitive radio; carrying out 1-bit quantification processing on sampling data; taking the compression sampling data after the quantification processing as training input, taking an actual frequency band occupancy situation as training output, and training a BP neural network detector. The actual detection process comprises: carrying out the compression sampling on the original signal of the cognitive radio and carrying out the 1-bit quantification processing on the sampling data; inputting the compression sampling data after the quantification processing into the trained BP neural network detector so as to obtain the output, wherein the output is the frequency spectrum occupancy situation of the cognitive radio. The invention also discloses a cognitive radio frequency spectrum perception system which uses the above method. By using the method of the invention, algorithm complexity is low and real-time performance of the frequency spectrum detection can be greatly increased.
Description
Technical field
The invention belongs to wireless communication technology field, concrete, relate to a kind of cognitive radio frequency spectrum cognitive method and system.
Background technology
Along with the broad development of radio communication service, frequency spectrum resource is deficient day by day, and its main cause is the unreasonable of wireless frequency spectrum access technology---and the existing wireless communications system all adopts inefficient fixed channel assignment strategy.Cognitive radio is considered to address this problem one of the most effective new technology, and its allows cognitive user under the prerequisite that does not influence validated user proper communication in the frequency band, to share this frequency range.
The frequency spectrum perception technology is the important foundation link of cognition wireless network design; Have only and understand main user in real time quickly, could under the situation that it is not caused extra interference, realize second use, the raising wireless frequency spectrum utilance of wireless frequency spectrum the situation that frequency spectrum uses.
According to the nyquist sampling law: in the transfer process of carrying out analog/digital signal; When sample frequency
greater than signal in during 2 times (
) of highest frequency
, the digital signal after the sampling can complete reservation raw information.Will guarantee generally that in practical application sample frequency is 5~10 times of signal highest frequency, under broadband even ultra broadband environment, signal will realize that harmless sampling needs very high sample frequency especially, and this is a test to sampling hardware.
Because the existing wireless communications system all adopts inefficient fixed channel assignment strategy, wireless frequency spectrum inserts unreasonable, and main CU frequency spectrum situation has very big frequency-domain sparse property under the actual conditions.Traditional theory is to handle these sparse signals like this: improve transmission efficiency for saving memory space, can do processed compressed to sparse signal.Suppose that length is that the signal x of N is that K-is sparse; Sampler is deferred to the nyquist sampling law and signal
is projected to its sparse territory
i.e.
, wherein
and
after to signal sampling.These non-zero parameters and position are sent, and
the individual component that will receive at receiving terminal is put back in the relevant position and other positions and is filled zero and obtain the decompress(ion) signal.
This thinking has intrinsic shortcoming:
(1) because sampling rate requires height, signal length can be very long, so the direct transform process is consuming time longer;
The position of
the individual component that (2) keeps is " self adaptation ", needs the space to deposit the component position;
(3) interference free performance is poor; Because this
individual component is " most important ", can have serious consequences if lose some components in the transmission course.
Summary of the invention
Technical problem to be solved by this invention is that the existing in prior technology computation complexity is high, the deficiency of online detection real-time difference; A kind of cognitive radio frequency spectrum cognitive method and system based on compressed sensing and BP neural net is provided; The compressed sensing technology is applied to the frequency spectrum detection of cognitive radio; And with the complicated algorithm for reconstructing of BP (Back-Propagation) neural net replacement; Computation complexity is transferred to the network training of off-line from online detection, thereby can significantly improve the online detection real-time of detection system.
Compressed sensing is an emerging technology, provides a kind of estimation of directly non-self-adapting need not abandon the method for unnecessary component, and it has utilized the autocorrelation performance of sparse signal, uses a spot of projection to come recovering signal.The status of these projections is of equal value, thereby in transmission, loses and somely can not cause serious consequence.Compare with the high sampling rate of conventional thought; The only simple a part of data (but not total data) of gathering of compressed sensing; Partly give reduction of data with complex processing and bring in and do,, greatly reduce the sampling rate that needs in few extracting data information as much as possible of trying one's best.The compressed sensing technology is an extension of conventional information opinion; But for Signal Collection Technology has been brought revolutionary breakthrough; The prototype structure that its adopts non-self-adapting linear projection to come inhibit signal is with far below the nyquist frequency sampled signal, through the accurate reconstruct of numerical optimization problem.
Yet the compressed sensing technology exists a big inferior position: because algorithm for reconstructing is the NP-hard problem; Calculation of complex needs to consider that the associating of
sub spaces is exhaustive.Substitute
min algorithm with
min algorithm; Can reduce complexity to a certain extent; But its algorithm complex is (
expression signal length) also near
, is not suitable for energy consumption, the responsive perception scene of time consumption.Frequency spectrum perception is obtained too accurate frequency domain reconstruction signal and can be caused resource waste just in order to judge that roughly main user uses the frequency spectrum situation in addition.The present invention utilizes the powerful match function replacement compression process of reconstruction of BP neural net, is transferred on the prior network training during from the signal reconstruction process consuming time.Thereby both kept advantages such as compressed sensing low sampling rate, low time and space consumption, strong robustness, compression process are simple, and, reached real-time and detect requirement again from reducing the system complexity of decompressor end in essence.
The present invention is concrete to adopt following technical scheme to solve the problems of the technologies described above.
Based on the cognitive radio frequency spectrum cognitive method of compressed sensing and BP neural net, comprise training process and actual detected process; Wherein, training process may further comprise the steps:
Steps A 2, with the compression sampling data after the 1-bit quantification treatment as training input, take situation as training output with actual frequency range, BP neural net detector is trained;
The actual detected process may further comprise the steps:
Step B1, the primary signal of cognitive radio is lower than the compression sampling of Nyquist rate, and the compression sampling data that collect are carried out the 1-bit quantification treatment;
Step B2, with the BP neural net detector that the compression sampling data after 1-bit quantification treatment inputs trains, the frequency spectrum that the output that obtains is said cognitive radio takies situation.
Adopt the cognitive radio frequency spectrum sensory perceptual system based on compressed sensing and BP neural net of said method, this system comprises:
The compression sampling unit is used for the primary signal of cognitive radio is lower than the compression sampling of Nyquist rate, and exports the compression sampling data;
The 1-bit quantifying unit is used for the compression sampling data of compression sampling unit output are carried out the 1-bit quantification treatment, and the data after the output processing;
BP neural net detector is input with compression sampling unit dateout, and the frequency spectrum of exporting said cognitive radio takies situation.
As the present invention's one preferred version; Said BP neural net detector is one to comprise three layers of BP neural net of input layer, latent layer, output layer; The input layer number is the signal length of compression sampling signal, and output layer node number is the frequency spectrum segments of said cognitive radio.
When the frequency spectrum broad of cognitive radio, the output layer neuron number can be bigger, as adopt above-mentioned preferred version, may bring BP neural metwork training difficulty, for this reason, can adopt the strategy of hierarchical detection, and is specific as follows:
Said BP neural net detector forms for
individual BP neural net cascade; BP neural nets at different levels include input layer, latent layer, output layer; The band pass filter of connecting respectively between the upper and lower level BP neural net,
is the integer greater than 1; If said cognitive radio frequency spectrum is divided into
individual frequency range; Make
,
is the integer greater than 1; The input layer number of first order BP neural net is the signal length
of compression sampling signal, and output layer node number is
; The input layer number of second level BP neural net is
, and output layer node number is
; The input layer number of third level BP neural net is
, and the output layer neuron number is
; The rest may be inferred; The input layer number of
level BP neural net is
, and the output layer neuron number is
.
The hidden neuron of said BP neural net is several to be confirmed or confirms according to following formula according to test:
In the formula;
is the hidden neuron number;
is the input layer number;
is the output layer neuron number,
be constant and
.
For the frequency spectrum perception under the cognitive radio environment, the present invention introduces the compressed sensing technology first, thereby can adopt the sample rate that is lower than Nyquist rate to carry out compression sampling, has reduced the requirement to hardware; And utilize the BP neural net to replace compression algorithm for reconstructing process, computation complexity is taken the network training that situation is transferred to off-line from online detection; And, significantly reduce data complexity through 1-bit quantification treatment sampled data, and reduce the neural metwork training complexity, avoid " crossing training ", and can reduce The noise to a certain extent.
Description of drawings
Fig. 1 detects the scene sketch map for cognitive radio frequency spectrum;
Fig. 2 is a BP neural network structure sketch map;
Fig. 3 is the cognitive radio frequency spectrum cognitive method schematic flow sheet based on compressed sensing and BP neural net of the present invention;
Fig. 4 is the schematic flow sheet that substep detects;
Fig. 5 is the principle schematic that substep detects.
Embodiment
Below in conjunction with accompanying drawing technical scheme of the present invention is elaborated:
Fig. 1 has shown the cognitive radio scene in the typical fdma system (like GSM, radio data system, OFDM etc.); Wherein, Main CU spectral bandwidth is fixed, and inferior user and main user share frequency spectrum
.If the main number of users that is in activated state (active) is
; It is
that main user
sends bandwidth through radio-frequency unit; The modulation signal
that center spectrum is
,
.Because main user's frequency spectrum occupied bandwidth is fixed, signal has the piece sparse characteristic.The frequency spectrum of frequency spectrum perception specification requirement real-time perception channel takies situation, need not harmless reconstruction signal.According to this detection target, can introduce the compression sampling theoretical frame.The only simple a part of data (but not total data) of gathering of compressed sensing are partly given reduction of data with complex processing and are brought in and do, and in few extracting data information as much as possible of trying one's best, greatly reduce the sampling rate that needs.Yet the high shortcoming of algorithm for reconstructing computation complexity that it had also makes the real-time of frequency spectrum detection to meet the demands.
Thinking of the present invention is to replace complicated compression algorithm for reconstructing with the BP neural net, computation complexity is transferred to the network training of off-line from online detection, thereby can significantly improve the online detection real-time of detection system.
The structure of BP neural net is as shown in Figure 2, has three layers or above network configuration (being divided into input layer, latent layer and output layer), and realization is complete between the different layers neuron connects, then continuous mutually with the neuron of layer.Input layer obtains training sample, propagates through latent course output layer, and each neuron of output layer obtains the input response of network.Then, be that target begins successively to revise from output layer and is connected weights to reduce target and actual value error, until getting back to input layer.The BP neural net is the core of feedforward neural network, has drawn the elite of neural network theory.Verified, can approach with the BP network of single latent layer for a continuous function in any closed interval.That is to say that three layers of BP network can approach any continuous function with arbitrary accuracy.Therefore the BP network application is extensive, and the neural network model of 80%-90% has adopted BP network or its version.
As target function, the systematic learning process will be by making error function reduce fastest direction adjustment weight coefficient
till obtaining satisfied weight coefficient collection with formula 1 in system.
The broader frequency spectrum that the BP neural net is introduced based on compressed sensing detects, and makes that (compressed domain CD) or on the low dimension of the title territory realizes, all is favourable as far as compressed sensing and two angles of neural net in compression domain in the neural network learning training.As far as the compressed sensing frequency spectrum detection, reduced the complexity of compressed sensing signal reconstruction: like Fig. 2 scene, detecting requirement is the situation that takies of each frequency range, is similar to the classification problem of neural net, and need can't harm reconstruction signal itself.Introduce neural net,, not only reduced the scale that needs deal with data, also improved requirement real-time with this compressed sensing process of reconstruction that replaces NP-hard.Only need the good network of precondition in the reality, just can obtain the occupied information of each frequency range in the short time, help the real-time application at the utmost point.From the neural net angle; The signal length that the frequency spectrum detection problem need be handled is big, situation is more, is unfavorable for network training, and it is theoretical to introduce compressed sensing; Problem is transferred to lower dimensional space from higher dimensional space; The deal with data amount subtracts greatly, has reduced the time and the space cost of signal processing, has improved accuracy in detection.
Cognitive radio frequency spectrum compressed sensing system based on the BP neural net of the present invention, as shown in Figure 3, comprising:
The compression sampling unit is used for the primary signal of cognitive radio is lower than the compression sampling of Nyquist rate, and exports the compression sampling data;
The 1-bit quantifying unit is used for the compression sampling data of compression sampling unit output are carried out the 1-bit quantification treatment, and the data after the output processing;
BP neural net detector is input with compression sampling unit dateout, exports the frequency spectrum detecting result of said cognitive radio.
When adopting said system to carry out frequency spectrum perception, be divided into two processes: training process and actual detected process.Each process can be divided into three parts: compression sampling, quantification treatment, neural metwork training/neural net detect.
Adopt the BP neural net to carry out frequency spectrum detection, key is how to design input, output and network configuration, below this three aspect is described respectively.
(1) detects input
Input obviously should be relevant with the result that node sample obtains.The most directly method is that information that node is obtained with the speed sampling of nyquist sampling law regulation is as input.But input signal too complex in this way is unfavorable for storage, calculation cost and network training.Adopt the compression sampling in the compressed sensing technology among the present invention, the primary signal of cognitive radio is lower than the compression sampling of Nyquist rate, thereby obtain comparatively simply input.Its principle is following: suppose
be the unknown vector (can refer to signal or image) on
territory according to actual conditions, need obtain in some way and reconstruction signal.To accomplish this purpose generally speaking and need
individual sampled point at least.But suppose to have such priori: under a known conversion; The transformation results of
is sparse or compressible, the data-signal that this moment, we can obtain
from
individual linear equation (rather than simple pixel samples).If
is individual
Linear equation is selected appropriate and is allowed reconstruction error to a certain degree, and
can be much smaller than the sampling number of common needs
Order
.If the resultant signal
of time user's perception is the one-dimensional signal of length
in the current frequency spectrum, degree of rarefication (promptly contains for
individual nonzero value is arranged), the two-dimensional observation matrix (
) that gaussian random matrix
is
.Itself is not or not signal
Sparse; But it is sparse or have a compressibility on frequency domain; Thereby change base
but following rarefaction representation Fourier, then be
The compressible sampled signal of uniting
.This compression sampling speed is to advise far below nyquist sampling theorem
Fixed minimum sampling rate.Can be referring to following document [1]-[3] about the more detailed content of compressed sensing:
[1]?Candès?E?J,?Romberg?J,?Tao?T.?Stable?signal?recovery?from?incomplete?and?inaccurate?measurements?[J].?Communications?on?Pure?and?Applied?Mathematics,?2006,?59(8):1207–1223.
[2]?Donoho?D?L.?Compressed?sensing?[J].?IEEE?Transactions?on?Information?Theory,?2006,?52(4):1289-1306.
[3]?Zhi?T,?Giannakis?G?B.?Compressed?sensing?for?wideband?cognitive?radio[C].?2007?IEEE?international?conference?on?acoustics,?speech,?and?signal?processing,?April?15-20,?2007:?IV-1357-?IV-1360.
According to the theoretical discussion of front, when satisfying certain condition, reconstruction signal as a result that can compression sampling, this this such input value of meaning is comprising and enough is being used to the spectrum information judged.Even if but it is theoretical to introduce compressed sensing, it is still comparatively complicated to detect input.Signal input value is continuous, and possibility is unfavorable for the convergence of learning process too much.
In addition; The transfer function that the node of neural net is taked has certain requirement to the span of characteristic; Like the sigmoid function value is [1; 1], though its domain of definition is
, the excessive too small function that all will cause of argument value is saturated; Make that the output valve discrimination is too little, convergence with learning process is unfavorable for classifying.
Comprehensive above-mentioned considering, the signal that the present invention obtains compression sampling is done the 1-bit quantification treatment, as the input of BP neural net.The complexity, the simplification that have reduced target function on the one hand detect input; Also avoided the generation of above-mentioned function saturated conditions on the other hand.
The principle that 1-bit quantizes is following: establish primary signal
length for
; Obtain the detection signal
of length through compression sampling for
; According to following formula observation signal
is done the 1-bit quantification treatment, obtains length and be the input of the detection behind the coding of
:
(2)
Emulation experiment proves: the signal after quantizing through 1-bit, and convergence rate is obviously accelerated, and the training error of network significantly reduces, and takies in various frequency ranges and almost can both obtain better to detect effect under the situation, and its faults rate is much smaller than quantification treatment situation not.
(2) detect output
The output of neural net should be a continuous variable, in order to realize classification with it, need do suitable coding to classification according to output.Considering that
individual segmentation spectrum region is equiprobable maybe be by main CU; Can regard each output node as one 0,1 two-valued variable,
individual output node codified obtains
type.
For example if can obtain 32 types of codings with 5 nodes:
00000,00001,00010,00011,00100,00101,00110,00111,
01000,01001,01010,01011,01100,01101,01110,01111,
10000,10001,10010,10011,10100,10101,10110,10111,
11000,11001,11010,11011,11100,?11101,11110,11111。
Wherein
expression judged result is a channel idle,
represent that judged result is that channel is occupied.Also can represent with the corresponding decimal system of coding.
(3) neural network structure
The concrete structure of neural net comprises several node layers (or claiming neuron), every layer of how much node of employing etc., need decide according to practical problem.Generally speaking, for most application scenarios, adopt the three-layer neural network of a latent layer just can perfectly accomplish artificial tasks.The input layer number is exactly the dimension of sample characteristics; Adopt the length
of coding back input signal in the present invention; The output layer interstitial content can be confirmed as frequency spectrum segmentation number
according to top discussion.Unique uncertain be in the middle of the number of latent layer.Generally speaking, the hidden node number will cause bigger training error very little, be referred to as " owing study " in the neural network theory.The big more neural network of hidden node number will have stronger learning ability, on training set, converge to separating of the littler target function of acquisition more easily.But relative, strong excessively learning ability can cause more weak popularization ability, though promptly can obtain very little training error, on new independent sample, but obtains very high test error, and in neural network theory, this is referred to as " crossing study ".To obtain balance " crossing study " and " owing study ", need to confirm good the number of hidden nodes.The number of middle latent layer does not have definite theoretical value, needs to obtain through test method.Tentative calculation is several times done by system; Select several different hidden node numbers; Respectively training sample set is made an experiment; Adopt the leaving-one method cross validation, select interstitial content
preferably according to error rate.Also can confirm according to following empirical equation:
In the formula;
is the hidden neuron number;
is the input layer number;
is the output layer neuron number,
be constant and
.
Adopt the BP neural net detector of said structure; When frequency spectrum segmentation number
is big; The form that corresponding number of categories
can increase with power series increases fast, is unfavorable for network training and accurately classification.For this reason; As
when big; Can adopt the strategy of hierarchical detection, specific as follows:
Said BP neural net detector forms for
individual BP neural net cascade; BP neural nets at different levels include input layer, latent layer, output layer; The band pass filter of connecting respectively between the upper and lower level BP neural net,
is the integer greater than 1; If said cognitive radio frequency spectrum is divided into
individual frequency range; Make
,
is the integer greater than 1; The input layer number of first order BP neural net is the signal length
of compression sampling signal, and output layer node number is
; The input layer number of second level BP neural net is
, and output layer node number is
; The input layer number of third level BP neural net is
, and the output layer neuron number is
; The rest may be inferred; The input layer number of
level BP neural net is
, and the output layer neuron number is
; The hidden neuron of said BP neural net is several to be confirmed or confirms according to following formula according to test:
In the formula;
is the hidden neuron number;
is the input layer number;
is the output layer neuron number,
be constant and
.
The flow process of hierarchical detection is as shown in Figure 4.The broader frequency spectrum segments
that order detects,
is the integer greater than 1.With
as the output node number of first order BP neural net; Be that first order detection is divided into entire spectrum
section
; Corresponding one-level band pass filter is
; Signal can extract the information of signal on frequency domain
through
filter; When the one-level neural net
of signal
through the training completion; Obtain the binary system output encoder
of length,
for
.Extracting one-level detects and takies frequency spectrum segment number set
in the coding; Show
the carrier frequency section of attaching most importance to district, meter
; If be eager to distribute the frequency spectrum cavity-pocket of bulk; Only need directly from set
, to extract to get final product
.
Element is done further detection in the set to
; Obtain the frequency spectrum cavity-pocket of smaller piece; Be about to signal
through band pass filter collection
; Obtain set of signals
; This
individual signal is passed through the secondary neural net that training is accomplished respectively; Obtain the binary system output encoder
of length for
;
; Extract and to take frequency spectrum segment number set
in the secondary detection coding; Show
the carrier frequency section of attaching most importance to district, meter
.If this moment, the bandwidth of frequency spectrum cavity-pocket was consistent with the required bandwidth of cognitive system transmission information; Then from set
, extract,
.
If element is done further detection in will gathering
; The band pass filter collection
that then signal
is finished coefficient correlation through adjusted carries out next step detection; So substep detects; A various challenge of classification is reduced to the simple problem of a plurality of less number of categories, and its principle is as shown in Figure 5.
This improvement strategy of hierarchical detection is through increasing the band pass filter of some; Relevant parameter through the adjustment band pass filter (because segmentation parameter
is fixed, can be controlled its parameter binary system output encoder that detection obtains according to upper level through simple software and upgrade the logical parameter of band automatically.), obtain interested signal flexibly, the progressively refinement testing result of classification.This improvement strategy can requirement of real time, can obtain the spectrum distribution situation fast roughly, also can satisfy the accurate needs that detect, and reaches the testing result under the arbitrary accuracy through refinement progressively.
The input layer of neural net has been simplified in this corrective measure, and complicated neural net is divided into a plurality of simple neural metwork training processes, has simplified problem; Because the relevant parameter of band pass filter can carry out the self adaptation adjustment through software, system is increased in the controlled range the needs of hardware, can not cause too big burden; The result that scheme obtains also can satisfy the requirement of real-time and accuracy simultaneously, and can adjust flexibly according to actual conditions, and this scheme has enlarged the scope of application based on the compression frequency spectrum perception of neural net greatly.
Claims (7)
1. the cognitive radio frequency spectrum cognitive method based on compressed sensing and BP neural net is characterized in that, comprises training process and actual detected process; Wherein, training process may further comprise the steps:
Steps A 1, the primary signal of cognitive radio is lower than the compression sampling of Nyquist rate, and the compression sampling data that collect are carried out the 1-bit quantification treatment;
Steps A 2, with the compression sampling data after the 1-bit quantification treatment as training input, take situation as training output with actual frequency range, BP neural net detector is trained;
The actual detected process may further comprise the steps:
Step B1, the primary signal of cognitive radio is lower than the compression sampling of Nyquist rate, and the compression sampling data that collect are carried out the 1-bit quantification treatment;
Step B2, with the BP neural net detector that the compression sampling data after 1-bit quantification treatment inputs trains, the frequency spectrum that the output that obtains is said cognitive radio takies situation.
2. according to claim 1 based on the cognitive radio frequency spectrum cognitive method of compressed sensing and BP neural net; It is characterized in that; Said BP neural net detector is one to comprise three layers of BP neural net of input layer, latent layer, output layer; The input layer number is the signal length of compression sampling signal, and output layer node number is the frequency spectrum segments of said cognitive radio.
3. according to claim 1 based on the cognitive radio frequency spectrum cognitive method of compressed sensing and BP neural net; It is characterized in that; Said BP neural net detector forms for
individual BP neural net cascade; BP neural nets at different levels include input layer, latent layer, output layer; The band pass filter of connecting respectively between the upper and lower level BP neural net,
is the integer greater than 1; If said cognitive radio frequency spectrum is divided into
individual frequency range; Make
,
is the integer greater than 1; The input layer number of first order BP neural net is the signal length
of compression sampling signal, and output layer node number is
; The input layer number of second level BP neural net is
, and output layer node number is
; The input layer number of third level BP neural net is
, and the output layer neuron number is
; The rest may be inferred; The input layer number of
level BP neural net is
, and the output layer neuron number is
.
4. like claim 2 or 3 said cognitive radio frequency spectrum cognitive methods, it is characterized in that the hidden neuron of said BP neural net is several to be confirmed or confirm according to following formula according to test based on compressed sensing and BP neural net:
5. cognitive radio frequency spectrum compressed sensing system based on the BP neural net that adopts the said method of claim 1 is characterized in that this system comprises:
The compression sampling unit is used for the primary signal of cognitive radio is lower than the compression sampling of Nyquist rate, and exports the compression sampling data;
The 1-bit quantifying unit is used for the compression sampling data of compression sampling unit output are carried out the 1-bit quantification treatment, and the data after the output processing;
BP neural net detector is input with compression sampling unit dateout, and the frequency spectrum of exporting said cognitive radio takies situation.
6. like the said cognitive radio frequency spectrum compressed sensing system of claim 5 based on the BP neural net; It is characterized in that; Said BP neural net detector is one to comprise three layers of BP neural net of input layer, latent layer, output layer; The input layer number is the signal length of compression sampling signal, and output layer node number is the frequency spectrum segments of said cognitive radio, and hidden neuron is several to be confirmed or confirm according to following formula according to test:
7. like the said cognitive radio frequency spectrum compressed sensing system of claim 5 based on the BP neural net; It is characterized in that; Said BP neural net detector forms for
individual BP neural net cascade; BP neural nets at different levels include input layer, latent layer, output layer; The band pass filter of connecting respectively between the upper and lower level BP neural net,
is the integer greater than 1; If said cognitive radio frequency spectrum is divided into
individual frequency range; Make
,
is the integer greater than 1; The input layer number of first order BP neural net is the signal length
of compression sampling signal, and output layer node number is
; The input layer number of second level BP neural net is
, and output layer node number is
; The input layer number of third level BP neural net is
, and the output layer neuron number is
; The rest may be inferred; The input layer number of
level BP neural net is
, and the output layer neuron number is
; The hidden neuron of BP neural nets at different levels is several to be confirmed or confirms according to following formula according to test:
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN2012100808404A CN102665221A (en) | 2012-03-26 | 2012-03-26 | Cognitive radio frequency spectrum perception method based on compressed sensing and BP (back-propagation) neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN2012100808404A CN102665221A (en) | 2012-03-26 | 2012-03-26 | Cognitive radio frequency spectrum perception method based on compressed sensing and BP (back-propagation) neural network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN102665221A true CN102665221A (en) | 2012-09-12 |
Family
ID=46774610
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN2012100808404A Pending CN102665221A (en) | 2012-03-26 | 2012-03-26 | Cognitive radio frequency spectrum perception method based on compressed sensing and BP (back-propagation) neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN102665221A (en) |
Cited By (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103346799A (en) * | 2013-06-05 | 2013-10-09 | 中国科学院微电子研究所 | Gas identification method based on compressed sensing theory |
CN104092503A (en) * | 2014-07-15 | 2014-10-08 | 哈尔滨工程大学 | Artificial neural network spectrum sensing method based on wolf pack optimization |
WO2014194482A1 (en) * | 2013-06-05 | 2014-12-11 | 中国科学院微电子研究所 | Gas recognition method based on compressive perception theory |
CN104540147A (en) * | 2015-01-23 | 2015-04-22 | 天津大学 | Collaborative compression perception quantifying and forwarding method |
CN105306779A (en) * | 2015-10-27 | 2016-02-03 | 西安电子科技大学 | Image encryption method based on compressive sensing and index scrambling |
CN105676178A (en) * | 2016-01-13 | 2016-06-15 | 南京邮电大学 | Wireless sensor network positioning method based on compressed sensing and BP neural networks |
CN106485316A (en) * | 2016-10-31 | 2017-03-08 | 北京百度网讯科技有限公司 | Neural network model compression method and device |
CN107480691A (en) * | 2017-07-04 | 2017-12-15 | 中国人民解放军总医院 | A kind of Thyroid Structure feature extracting method and system based on ultrasound data dimensionality reduction |
CN108055096A (en) * | 2018-02-13 | 2018-05-18 | 南通大学 | The frequency spectrum sensing method detected based on signal and noise characteristic |
WO2018107414A1 (en) * | 2016-12-15 | 2018-06-21 | 上海寒武纪信息科技有限公司 | Apparatus, equipment and method for compressing/decompressing neural network model |
CN108924847A (en) * | 2018-06-04 | 2018-11-30 | 东南大学 | A kind of cognitive radio frequency spectrum prediction technique and equipment based on ANN |
CN109379153A (en) * | 2018-12-17 | 2019-02-22 | 电子科技大学 | A kind of frequency spectrum sensing method |
CN109547961A (en) * | 2018-11-29 | 2019-03-29 | 北京理工大学 | Big data quantity compressed sensing decoding method in a kind of wireless sensor network |
CN112381462A (en) * | 2020-12-07 | 2021-02-19 | 军事科学院***工程研究院网络信息研究所 | Data processing method of intelligent network system similar to human nervous system |
CN112671486A (en) * | 2020-12-28 | 2021-04-16 | 电子科技大学 | Neural network-based combined spectrum sensing method and system |
CN115209508A (en) * | 2021-04-13 | 2022-10-18 | ***通信集团设计院有限公司 | Spectrum access method, device, equipment and storage medium |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20110241917A1 (en) * | 2010-02-23 | 2011-10-06 | Baraniuk Richard G | Method And Apparatus For Signal Reconstruction From Saturated Measurements |
CN102253327A (en) * | 2011-06-16 | 2011-11-23 | 长沙河野电气科技有限公司 | Diagnostic method for failure of switch current circuit |
CN102255675A (en) * | 2010-05-19 | 2011-11-23 | 索尼公司 | Spectrum sensing device, method and program based on cognitive radio |
-
2012
- 2012-03-26 CN CN2012100808404A patent/CN102665221A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20110241917A1 (en) * | 2010-02-23 | 2011-10-06 | Baraniuk Richard G | Method And Apparatus For Signal Reconstruction From Saturated Measurements |
CN102255675A (en) * | 2010-05-19 | 2011-11-23 | 索尼公司 | Spectrum sensing device, method and program based on cognitive radio |
CN102253327A (en) * | 2011-06-16 | 2011-11-23 | 长沙河野电气科技有限公司 | Diagnostic method for failure of switch current circuit |
Non-Patent Citations (2)
Title |
---|
P. BOUFOUNOS等: "1-Bit Compressive Sensing", 《PROC.CONF. INF. SCI. SYST. (CISS)》 * |
ROBERT CALDERBANK等: "Compressed Learning:Universal Sparse Dimensionality Reduction and Learning in the Measurement Domain", 《HTTP://CITESEERX.IST.PSU.EDU/VIEWDOC/SUMMARY?DOI=10.1.1.154.7564》 * |
Cited By (25)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2014194482A1 (en) * | 2013-06-05 | 2014-12-11 | 中国科学院微电子研究所 | Gas recognition method based on compressive perception theory |
CN103346799A (en) * | 2013-06-05 | 2013-10-09 | 中国科学院微电子研究所 | Gas identification method based on compressed sensing theory |
CN104092503A (en) * | 2014-07-15 | 2014-10-08 | 哈尔滨工程大学 | Artificial neural network spectrum sensing method based on wolf pack optimization |
CN104540147A (en) * | 2015-01-23 | 2015-04-22 | 天津大学 | Collaborative compression perception quantifying and forwarding method |
CN104540147B (en) * | 2015-01-23 | 2018-01-09 | 天津大学 | One kind cooperation compressed sensing quantifies retransmission method |
CN105306779A (en) * | 2015-10-27 | 2016-02-03 | 西安电子科技大学 | Image encryption method based on compressive sensing and index scrambling |
CN105676178A (en) * | 2016-01-13 | 2016-06-15 | 南京邮电大学 | Wireless sensor network positioning method based on compressed sensing and BP neural networks |
CN105676178B (en) * | 2016-01-13 | 2019-05-24 | 南京邮电大学 | Wireless sensor network locating method based on compressed sensing and BP neural network |
CN106485316B (en) * | 2016-10-31 | 2019-04-02 | 北京百度网讯科技有限公司 | Neural network model compression method and device |
CN106485316A (en) * | 2016-10-31 | 2017-03-08 | 北京百度网讯科技有限公司 | Neural network model compression method and device |
WO2018107414A1 (en) * | 2016-12-15 | 2018-06-21 | 上海寒武纪信息科技有限公司 | Apparatus, equipment and method for compressing/decompressing neural network model |
CN107480691B (en) * | 2017-07-04 | 2020-04-03 | 中国人民解放军总医院 | Thyroid structure feature extraction method and system based on ultrasonic data dimension reduction |
CN107480691A (en) * | 2017-07-04 | 2017-12-15 | 中国人民解放军总医院 | A kind of Thyroid Structure feature extracting method and system based on ultrasound data dimensionality reduction |
CN108055096A (en) * | 2018-02-13 | 2018-05-18 | 南通大学 | The frequency spectrum sensing method detected based on signal and noise characteristic |
CN108055096B (en) * | 2018-02-13 | 2019-08-09 | 南通大学 | The frequency spectrum sensing method detected based on signal and noise characteristic |
CN108924847A (en) * | 2018-06-04 | 2018-11-30 | 东南大学 | A kind of cognitive radio frequency spectrum prediction technique and equipment based on ANN |
CN109547961A (en) * | 2018-11-29 | 2019-03-29 | 北京理工大学 | Big data quantity compressed sensing decoding method in a kind of wireless sensor network |
CN109547961B (en) * | 2018-11-29 | 2020-06-09 | 北京理工大学 | Large data volume compressed sensing coding and decoding method in wireless sensor network |
CN109379153A (en) * | 2018-12-17 | 2019-02-22 | 电子科技大学 | A kind of frequency spectrum sensing method |
CN109379153B (en) * | 2018-12-17 | 2020-01-17 | 电子科技大学 | Spectrum sensing method |
CN112381462A (en) * | 2020-12-07 | 2021-02-19 | 军事科学院***工程研究院网络信息研究所 | Data processing method of intelligent network system similar to human nervous system |
CN112671486A (en) * | 2020-12-28 | 2021-04-16 | 电子科技大学 | Neural network-based combined spectrum sensing method and system |
CN112671486B (en) * | 2020-12-28 | 2022-06-24 | 电子科技大学 | Combined spectrum sensing method and system based on neural network |
CN115209508A (en) * | 2021-04-13 | 2022-10-18 | ***通信集团设计院有限公司 | Spectrum access method, device, equipment and storage medium |
CN115209508B (en) * | 2021-04-13 | 2023-09-19 | ***通信集团设计院有限公司 | Spectrum access method, device, equipment and storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN102665221A (en) | Cognitive radio frequency spectrum perception method based on compressed sensing and BP (back-propagation) neural network | |
CN107393526B (en) | Voice silence detection method, device, computer equipment and storage medium | |
CN105610525B (en) | A kind of broadband compression frequency spectrum sensing method based on sub-band match selection | |
CN102291341A (en) | Broadband frequency spectrum perceiving method based on compression perception | |
CN104780008B (en) | A kind of broader frequency spectrum cognitive method perceived based on self-adapting compressing | |
CN101630982A (en) | Broadband cognitive radio frequency spectrum detection method based on parallelly compressed sensing | |
CN103117818A (en) | Broadband spectrum sensing method based on space-frequency joint compressed sensing | |
CN112491442B (en) | Self-interference elimination method and device | |
CN103247028A (en) | Multi-hypothesis prediction block compressed sensing image processing method | |
CN108924148A (en) | A kind of source signal collaborative compression perception data restoration methods | |
CN111954250A (en) | Lightweight Wi-Fi behavior sensing method and system | |
Chen et al. | A spatiotemporal data compression approach with low transmission cost and high data fidelity for an air quality monitoring system | |
Feng et al. | A new technology of remote sensing image fusion | |
CN104123740A (en) | Image reconstruction method based on compressive sensing | |
Yoo et al. | On the role of ViT and CNN in semantic communications: Analysis and prototype validation | |
CN113837122B (en) | Wi-Fi channel state information-based contactless human body behavior recognition method and system | |
Wu et al. | A holistic approach to reconstruct data in ocean sensor network using compression sensing | |
CN103139828B (en) | Broadband spectrum sensing device and method | |
Raj et al. | Multilayered convolutional neural network-based auto-CODEC for audio signal denoising using mel-frequency cepstral coefficients | |
CN106878997A (en) | A kind of Bayes compression frequency spectrum sensing method in distributed network | |
Sivaroopan et al. | SyNIG: synthetic network traffic generation through time series imaging | |
Zamani et al. | On the use of compressive sensing (CS) exploiting block sparsity for neural spike recording | |
CN106656373B (en) | Broadband distributed cooperative compressed spectrum sensing method based on weighted consistent optimization | |
CN112671486B (en) | Combined spectrum sensing method and system based on neural network | |
CN115908969A (en) | Method and apparatus for image processing and model training |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
C02 | Deemed withdrawal of patent application after publication (patent law 2001) | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20120912 |