CN115438700A - Spectrum sensing method and device, electronic equipment and medium - Google Patents
Spectrum sensing method and device, electronic equipment and medium Download PDFInfo
- Publication number
- CN115438700A CN115438700A CN202211083076.6A CN202211083076A CN115438700A CN 115438700 A CN115438700 A CN 115438700A CN 202211083076 A CN202211083076 A CN 202211083076A CN 115438700 A CN115438700 A CN 115438700A
- Authority
- CN
- China
- Prior art keywords
- alexnet
- detection
- alexnet model
- data
- polarization
- 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
- 238000001228 spectrum Methods 0.000 title claims abstract description 38
- 238000000034 method Methods 0.000 title claims description 34
- 230000010287 polarization Effects 0.000 claims abstract description 48
- 239000011159 matrix material Substances 0.000 claims abstract description 24
- 238000001514 detection method Methods 0.000 claims description 74
- 238000012360 testing method Methods 0.000 claims description 42
- 238000012549 training Methods 0.000 claims description 30
- 230000006870 function Effects 0.000 claims description 25
- 238000007781 pre-processing Methods 0.000 claims description 20
- 230000008569 process Effects 0.000 claims description 15
- 238000003657 Likelihood-ratio test Methods 0.000 claims description 11
- 230000004913 activation Effects 0.000 claims description 9
- 238000011176 pooling Methods 0.000 claims description 9
- 238000005070 sampling Methods 0.000 claims description 9
- 238000003860 storage Methods 0.000 claims description 7
- 238000012163 sequencing technique Methods 0.000 claims description 4
- 238000004590 computer program Methods 0.000 claims description 2
- 238000002372 labelling Methods 0.000 claims description 2
- 238000013507 mapping Methods 0.000 claims description 2
- 238000004422 calculation algorithm Methods 0.000 abstract description 42
- 239000013598 vector Substances 0.000 abstract description 27
- 238000013135 deep learning Methods 0.000 abstract description 7
- 230000008447 perception Effects 0.000 abstract description 2
- 238000010586 diagram Methods 0.000 description 9
- 238000004458 analytical method Methods 0.000 description 8
- 230000014509 gene expression Effects 0.000 description 8
- 230000005684 electric field Effects 0.000 description 6
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 5
- 230000008859 change Effects 0.000 description 5
- 238000013461 design Methods 0.000 description 5
- 230000000694 effects Effects 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 4
- 238000012545 processing Methods 0.000 description 4
- 238000004364 calculation method Methods 0.000 description 3
- 238000005562 fading Methods 0.000 description 3
- 238000005457 optimization Methods 0.000 description 3
- 238000004088 simulation Methods 0.000 description 3
- 239000000654 additive Substances 0.000 description 2
- 230000000996 additive effect Effects 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 2
- 238000009826 distribution Methods 0.000 description 2
- 230000010355 oscillation Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000012795 verification Methods 0.000 description 2
- 230000001149 cognitive effect Effects 0.000 description 1
- 238000005314 correlation function Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000009795 derivation Methods 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 230000009977 dual effect Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 238000011423 initialization method Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 230000000644 propagated effect Effects 0.000 description 1
- 230000007480 spreading Effects 0.000 description 1
- 238000003892 spreading Methods 0.000 description 1
- 238000010998 test method Methods 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Radar Systems Or Details Thereof (AREA)
Abstract
In order to further improve the performance of spectrum sensing under the condition of no signal and noise prior knowledge, the invention explores the spectrum sensing by combining the polarization domain characteristics of signals with a deep learning algorithm. The polarization information of the signal contains time domain and frequency domain features, and the polarization information is a vector and contains the direction information of the signal, so that more features can be utilized for spectrum sensing by using the polarization information of the signal. The invention uses the orthogonal dual-polarized antenna to receive signals, uses the Jones vector to represent the polarization state of the signals, calculates to obtain the covariance matrix of the signals, converts the covariance matrix into an image and uses the image as the input of the AlexNet model, so that the AlexNet model learns the polarization information characteristics of the signals, thereby obtaining better perception performance without the prior knowledge of the signals and noise.
Description
Technical Field
The present invention relates to the field of spectrum sensing technologies, and in particular, to a spectrum sensing method and apparatus, an electronic device, and a medium.
Background
The cognitive radio improves the spectrum utilization rate by utilizing spectrum holes in a wireless environment, one of the core technologies is spectrum sensing, and meanwhile, exploring a new Spectrum Sensing (SS) technology is an important research challenge. At present, the research on spectrum sensing mainly uses time-frequency information of received signals, such as amplitude, power, frequency and phase information of signals, to sense spectrum occupancy according to a difference of certain characteristics of the received signals and noise, but there is very little sensing using polarization information of the signals, where such information is a set of characteristics of the signals received in time domain and frequency domain, and the polarization information is vector information, and includes the incoming direction of the signals, so that the signals received by an orthogonal dual-polarization antenna are used for spectrum sensing, which can further improve detection performance.
In the prior art, the traditional spectrum sensing algorithm based on polarization information has the defects of higher complexity, need of priori knowledge or poorer sensing performance.
Disclosure of Invention
In view of this, the present invention is directed to a spectrum sensing method, apparatus, electronic device and medium, which use a deep learning technique to learn polarization domain characteristics of signals and noise, so that better sensing performance can be obtained without prior knowledge of the signals and noise.
In order to achieve the purpose, the technical scheme of the invention is realized as follows:
a method of spectrum sensing, comprising the steps of:
step 1: performing polarization data sampling and data preprocessing, including:
receiving a signal using an orthogonal dual-polarized antenna;
the received polarization signals are converted into covariance matrixes through data preprocessing, and the covariance matrixes are further converted into images for storage;
step 2: carrying out online training based on an AlexNet model, comprising the following steps:
an image obtained through data preprocessing is used as input data and is submitted to an AlexNet model learning characteristic;
constructing test statistics according to the output result of the AlexNet model by utilizing a Neyman Pearson criterion;
and step 3: setting a detection threshold, comprising:
the noise image obtained through data preprocessing is delivered to a trained AlexNet model;
sequencing and constructing the obtained results, and setting the false alarm probability according to actual needs so as to set a detection threshold;
and 4, step 4: carrying out off-line detection based on an AlexNet model, comprising the following steps:
test data are sent to an AlexNet model for testing;
and (4) constructing corresponding test statistics according to the test result, and comparing the test statistics with the detection threshold set in the step (3) so as to judge whether the authorized user exists.
Further, the setting of the detection threshold is to send the noise sample gray level graph to a trained AlexNet model for detection, and set the false alarm probability as required, so as to obtain a suitable threshold according to the false alarm probability.
Further, the AlexNet model comprises 5 convolutional layers for extracting the features of the polarization covariance matrix and 3 fully-connected layers for mapping the learned features to the space of the sample labels, and a ReLu activation function or a max pooling layer is used after the convolutional layers and the fully-connected layers.
Further, in the step 4, during off-line detection, a sample gray scale image without label labeling is collected as test data, the test data is transmitted to a trained AlexNet model, then AlexNet-LRT detection is performed based on an unlabeled sample, the maximum posterior probability is converted into a conditional probability to deduce a Likelihood Ratio Test (LRT) of a detection process, and after a detection stage obtains a test statistic, the test statistic can be compared with a preset detection threshold value, so that whether an authorized user exists or not can be quickly judged.
The invention also provides a spectrum sensing device, which comprises
Polarization data sampling and preprocessing device for polarization data sampling and data preprocessing include: receiving a signal using an orthogonal dual-polarized antenna;
the received polarization signals are converted into covariance matrixes through data preprocessing, and the covariance matrixes are further converted into images for storage;
the online training device is used for performing online training based on the AlexNet model and comprises:
taking an image obtained through data preprocessing as input data and delivering the input data to an AlexNet model learning characteristic;
constructing test statistics according to the output result of the AlexNet model by utilizing a Neyman Pearson criterion;
threshold setting means for setting a detection threshold, comprising:
the noise image obtained through data preprocessing is delivered to a trained AlexNet model;
sequencing and constructing the obtained results, and setting the false alarm probability according to actual needs so as to set a detection threshold;
the off-line detection device is used for carrying out off-line detection based on an AlexNet model, and comprises:
test data are sent to an AlexNet model for testing;
and constructing corresponding test statistic according to the test result, and comparing the test statistic with the detection threshold set by the threshold setting device so as to judge whether the authorized user exists.
The invention also provides an electronic device comprising at least one processor, an
At least one memory communicatively coupled to the processor, wherein:
the memory stores program instructions executable by the processor, which when called by the processor are capable of performing the methods described above.
The invention further relates to a computer-readable storage medium, in which a computer program is stored which, when being executed by a processor, carries out the above-mentioned method steps.
Compared with the prior art, the spectrum sensing method, the spectrum sensing device, the electronic equipment and the spectrum sensing medium have the following advantages:
(1) The orthogonal dual-polarized antenna is used for receiving signals, the polarization state of the signals is represented by Jones vectors, the covariance matrix of the signals is obtained through calculation and is converted into an image to be used as the input of an AlexNet model, the AlexNet model is made to learn the polarization information characteristics of the signals, and therefore good perception performance is obtained under the condition that priori knowledge of the signals and noise is not needed;
(2) The invention designs the test statistic of the JCM-AlexNet algorithm by applying Neyman Pearson's theorem (NP), so that the threshold value can be set through the false alarm probability, and the setting of the threshold value is more convenient.
Drawings
The accompanying drawings, which are included to provide a further understanding of the invention, illustrate embodiments of the invention and together with the description serve to explain the invention and do not constitute a limitation of the invention. In the drawings:
FIG. 1 is a schematic diagram of a spectrum sensing algorithm framework according to the present invention;
fig. 2 is a gray scale diagram of a covariance matrix of a noise and a signal sample for 40 orthogonal dual-polarized antennas according to the present invention;
FIG. 3 is a diagram of the AlexNet model architecture of the present invention;
fig. 4 is a schematic diagram of detection probabilities of 5 algorithms with different signal-to-noise ratios when pf =0.01 in the present invention;
FIG. 5 is a schematic diagram of detection probabilities of 5 algorithms with different false alarm probabilities when SNR = -15db in accordance with the present invention;
FIG. 6 is a schematic diagram showing variation of AlexNet model loss values under different epochs according to the present invention;
fig. 7 is a diagram illustrating the variation of the loss value under different snr according to the present invention.
Detailed Description
It should be noted that the embodiments and features of the embodiments may be combined with each other without conflict.
In the description of the present invention, it is to be understood that the terms "central," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," and the like are used in the orientations and positional relationships indicated in the drawings, which are based on the orientations and positional relationships indicated in the drawings, and are used for convenience in describing the present invention and for simplicity in description, but do not indicate or imply that the device or element so referred to must have a particular orientation, be constructed in a particular orientation, and be operated, and thus should not be construed as limiting the present invention. Furthermore, the terms "first", "second", etc. are used for descriptive purposes only and are not to be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first," "second," etc. may explicitly or implicitly include one or more of that feature. In the description of the present invention, "a plurality" means two or more unless otherwise specified.
In the description of the present invention, it should be noted that, unless otherwise explicitly specified or limited, the terms "mounted," "connected," and "connected" are to be construed broadly, e.g., as meaning either a fixed connection, a removable connection, or an integral connection; can be mechanically or electrically connected; they may be connected directly or indirectly through intervening media, or they may be interconnected between two elements. The specific meaning of the above terms in the present invention can be understood by those of ordinary skill in the art through specific situations.
The present invention will be described in detail below with reference to the embodiments with reference to the attached drawings.
Fig. 1 is a schematic diagram of the spectrum sensing method of the present invention. Specifically comprises
Polarization domain signal model
The classical spectrum sensing algorithm only uses one-dimensional scalar characteristics of a signal, such as a time domain, a frequency domain and the like, for sensing, and actually, polarization domain two-dimensional vector characteristics of the signal are also available important information, wherein the two-dimensional vector characteristics comprise Jones vectors representing polarization information of the signal. We will next generate a Jones vector signal model.
According to the classical electromagnetic wave polarization theory, the complex vector expression of the electromagnetic wave is as follows:
wherein,is a complex electric field vector, ω is the signal frequency, t = nT s N is the number of sampling points of the signal, 1/T s Is the sampling rate, r is the coordinate vector of any point in space,λ is the propagation vector of the electromagnetic wave and λ is the wavelength. If the electromagnetic wave is propagated in the + z direction, considering only a cross section with z =0, the vertical and horizontal components of the complex electric field vector are:
wherein | E H I and I E V I represents the polarization components of the signal in the horizontal and vertical directions, phi H And phi V Representing the horizontal and vertical polarization phases.
Thus, the complex vector expression of an electromagnetic wave can be simplified as:
electromagnetic waves of any single frequency can be decomposed in any two mutually orthogonal polarization states (called polarization groups). The invention is expressed by the common Vertical (V) and Horizontal (H) orthogonal polarization groups:
thus, the polarization state of an electromagnetic wave can be expressed as a two-dimensional vector, the Jones vector
Amount:
wherein, E A Representing the amplitude of the electric field of the signal, E A =(|E H | 2 +|E V | 2 ) 1/2 ,|E H |=E A cosθ,|E V |=E A sin theta, theta represents an included angle between the amplitude of the electric field and the vertical component of the amplitude; γ represents the amplitude of the polarization, γ = arctan (E) V /E H ),Phi denotes the phase difference of the polarization, phi = phi H -φ V ,φ∈[-π , π)。
As can be seen from equation (5), only E needs to be used A The three parameters of gamma and phi can completely represent the track shape of the plane electromagnetic wave electric field vector, and if only the horizontal and vertical components of the electric field are considered, a normalized Jones vector can be adopted:
signal model
Spectrum sensing can be expressed as a binary hypothesis testing problem:
wherein H 1 Indicating that the channel to be detected is occupied by a user, H 0 It indicates that the channel is idle. Samples representing the received signal, R x For the variance, x, of the received signal samples H [n],x V [n]Representing samples of the received signal in the horizontal and vertical directions, respectively. T is s Is the sampling time; representing independent identically distributed two-channel additive white Gaussian noise, R u To receive the variance of the noise samples, u H [n],u V [n]Representing samples of received noise in the horizontal and vertical directions, respectively.Representing samples of the authorized user signal after passing through a fading channel,is the variance of the samples of the authorized user signal after passing through the fading channel. When the effects of channel multi-path delay are ignored,representing samples of the authorized user signal after passing through a fading channel in the horizontal and vertical directions, respectively.Can be expressed ass[n]For transmitting samples of a signal, h [ n ]]For dual polarization depolarized channels, it can be expressed as:
wherein h is xy Represents the complex gain of the channel between the transmit polarization component y and the receive polarization component x, phi i Representing its random phase; due to s [ n ]]For a transmitted signal sample containing a vector of polarization information Jones, it can be expressed as:
the signal model thus generated is the Jones vector signal model.
Data pre-processing
Assuming that the signal model has M orthogonal dual-polarized antennas to receive signals, the signals received by the antennas can be expressed as:
wherein,s i [n]=Ae jωt ,A=[E′ Jones1 ,E′ Jones2 ,…E′ JonesM ]the vector matrix is steered for the polarization domain.
Order: x = [ X ] 1 ,x 2 ,…,x M ] T (11)
u=[u 1 ,u 2 ,…,u M ] T (12)
u i For the noise signal received by the i-th antenna, x i For the sampled signal received by the ith antenna, X spreading is written in the form of a matrix as follows:
the sample covariance matrix of the sampled signal is:
wherein,as a variance of the noise, I M Representing an identity matrix of order M, R s (N) is a sample covariance matrix of the transmitted signal.
Suppose there are 40 orthogonal dipolesChemical antenna, H 0 And H 1 Assuming that the gray-scale images of the lower sample covariance matrix are respectively shown in fig. 2 (a) and (b), when no main user exists, the values except the diagonal are very small, and the brightness distribution of the gray-scale images is relatively uniform; when a main user exists, the numerical value except the diagonal line is large, and the brightness distribution of the gray-scale image is not uniform.
The elements in the covariance matrix represent the correlation between random variables, and the smaller the correlation, the smaller the element value, and the darker the grayscale brightness. The diagonal elements are autocorrelation of random variables, have strong correlation, have larger numerical value, and the diagonal brightness of the gray-scale image is brighter; off-diagonal elements represent cross-correlation between random variables, with correlation weaker than auto-correlation, and off-diagonal luminance darker than the diagonal luminance of the grayscale map. As shown in fig. 2 (a), the noise has randomness, the correlation of the noise received by different secondary users is weak, and the gray scale image outside the diagonal is darker in color; as shown in fig. 2 (b), different secondary users receive the same transmitted signal, and the useful signals received by the users have stronger cross-correlation, and the brightness of the gray-scale images outside the diagonal is brighter than that of fig. 2 (a).
The gray-scale graphs of the useful signals and the noise are obviously different, namely the covariance matrix characteristics of the useful signals and the noise are obviously different, and the AlexNet model is good at learning the matrix characteristics, so that the covariance matrix characteristics are extracted by using the AlexNet model, and the spectrum sensing is realized.
On-line training
The JCM-AlexNet algorithm is that a marked JCM sample data gray level graph is sent to an AlexNet model to be trained, so that the AlexNet model learns to obtain appropriate parameters, two maximum posterior probabilities related to signals and noise are obtained, a likelihood ratio test is constructed by utilizing the maximum posterior probabilities, and the likelihood ratio test is compared with a threshold value to realize the correct classification of the signals and the noise.
1. AlexNet model structure
The internal structure of the AlexNet model is shown in fig. 3, where a Jones vector covariance matrix gray-scale map is used as input data for the model, and conv represents the convolutional layer; the largest pooling layer was used after conv1, volume conv2, conv 5; FC represents a fully connected layer; kernel stands for convolution kernel, since the data set used by the present invention is small, to speed up training, the number of convolution kernels per convolution layer is halved, the first convolution layer uses 48, 11 × 11 convolution kernels, the second convolution layer uses 128, 5 × 5 convolution kernels, the third convolution layer uses 192, 3 × 3 convolution kernels, the fourth convolution layer uses 192, 3 × 3 convolution kernels, the fifth convolution layer uses 128, 3 × 3 convolution kernels, and finally 3 fully-connected layers are used for final classification, since the present invention is a binary hypothesis testing problem, the last fully-connected layer has a size of 2.
2. Training phase
In the training phase, the labeled sample gray map is sampled as a training set:
wherein,represents R x (N), Y represents a set of labels Y,the kth data (k =1,2,3.... K) representing the training set, and further, the labels of the training data and the validation data are encoded as one-hot encoded vectors:
H 1 indicating an authorized user. Accordingly, the last fully-connected layer output of the AlexNet model is represented as a 2 × 1 class fraction vector:
wherein f is θ (. Cndot.) is an expression containing AlexNet model parameters θ,is H i And (4) corresponding expressions. At this time, the process of the present invention,represents H i The classification score of (1).
Therefore, we have the following two hypothetical probability expressions:
Therefore, the goal of AlexNet model training is to maximize the conditional probability, i.e.:
for ease of calculation, we introduce a logarithmic function:
we expect that the larger l (θ) the better, i.e., the smaller l (θ) the better, the loss function can be derived:
therefore, the goal of AlexNet training is to obtain the optimal θ to maximize the a posteriori probabilityAt maximum, i.e.:
wherein, theta * Represents the best θ under maximum a posteriori probability (MAP) estimation.
Based on the loss function obtained by derivation in (21), the parameter θ of the AlexNet model is gradually updated by using an Adam optimization algorithm, so that the training process is converged, that is, the AlexNet model obtains the optimal parameter θ, and finally the trained AlexNet model is obtained, which can be expressed as:
wherein, f θ* (R x ) Represents the input R x Then through the expression of the trained AlexNet model,represents H i The classification score of (1).
3. Test statistic design
From the results of the previous training process, i.e., equations (22) and (23), two maximum a posteriori probabilities can be obtained:
according to bayes' theorem, we can get:
wherein P (R) x ∣H i ) Denotes a given H i Conditional probability of (2),P(R x ) As edge probability, P (H) i ) Represents H in the training process i A priori probability of. Known as P (R) x ∣H 1 ) And P (R) x ∣H 0 ) In the case of (2), the neman-pearson theorem (NP) demonstrates the superior performance of the Likelihood Ratio (LR) as a test statistic, which is defined in the neman-pearson theorem.
Neman pearson theorem (NP): at false alarm probability P f When known, in order to make the detection probability P d Maximization, we assume at H 1 In the case of (2):
wherein P is f =P(H 1 |H 0 ),P d =P(H 1 |H 1 ),L(R x ) For the likelihood ratio, γ is the detection threshold, and the test method of equation (27) is called the Likelihood Ratio Test (LRT).
By substituting equations (25) and (26) into equation (27) according to the NP theorem, alexNet-LRT can be obtained:
wherein, the threshold value gamma can be obtained from false alarm probability constraint, and for convenient analysis, the training data set H is used 1 And H 0 The number of samples is the same, i.e. P (H) 1 )=P(H 0 ) =0.5, further:
off-line detection
The JCM-AlexNet frequency spectrum sensing model detection is to transfer the data gray level diagram of the unlabeled JCM sample to a trained model (namely, the model obtains proper parameters through training) for detection, thereby realizing the correct classification of signals and noise. Firstly, a detection threshold is designed according to a trained model, and then a probability value obtained by likelihood ratio test in the detection process is compared with the threshold, so that whether an authorized user exists or not is determined.
1. Detection threshold design
The threshold design is to send a noise sample gray level image to a trained AlexNet model for detection, and set P according to the requirement f Thus according to P f A suitable threshold is obtained if:
wherein R is u (N) is a sample covariance matrix formed by N noise vectors, u (N) represents that the mean is 0 and the variance is sigma 2 Independent and identically distributed additive white Gaussian noise (T) can be obtained JCM-AlexNet At H 0 The expression of time:
due to P f =P(H 1 |H 0 ) That is, the probability that an unauthorized user is determined as an authorized user, the following can be rewritten:
P f =P{T JCM-AlexNet ∣H 0 >γ} (32)
order:represents H 0 And the lower data set consists of L data, and the L data are sent to the trained JCM-AlexNet model to obtain all T JCM-AlexNet ∣H 0 Values, sorted in descending order, T can be constructed JCM-AlexNet ∣H 0 Set, represented asThe detection threshold may be obtained by:
whereinIndicating rounding down to the nearest integer,representing the l-th elementThe value is obtained.
2. Determination of detection result
In the model detection stage, a sample gray level graph without label marking is acquired as test data and expressed asWill be provided withTransmitting the data to a trained AlexNet model, and then carrying out AlexNet-LRT detection based on an unlabeled sample, wherein the maximum posterior probability expression obtained in the training process is the unknown model parameter, and the detection probability is calculated for the known model parameter in the detection process, so that P (H) is obtained i ∣R x ) Not suitable for detecting instances, converting them into conditional probabilities P (R) x ∣H 1 ) To derive a Likelihood Ratio Test (LRT) for the detection process.
Wherein:
substituting equations (36) and (37) into equation (35) yields:
therefore, after the test statistic is obtained in the detection stage, the test statistic can be compared with a preset detection threshold value, and whether the authorized user exists or not can be judged quickly.
The effect of the method of the present invention is verified by experiments below.
Firstly, the network structure of the proposed JCM-AlexNet algorithm is analyzed, and the AlexNet model comprises 5 convolutional layers (Conv 2 d) and 3 fully-connected layers (Linear), wherein the convolutional layers are used for extracting the features of the polarization covariance matrix, and the fully-connected layers map the learned features to the space of sample marks to play the role of a classifier, as shown in the following table, since the spectrum sensing can be expressed as a binary hypothesis testing problem, the output of the last fully-connected layer is 2. A ReLu activation function or a maximum pooling layer is also used after the convolutional layer and the full-link layer, the ReLu activation function is a nonlinear combination for changing output into input, so that input data can approximate to any function through model training; the maximum pooling layer is used for reducing feature dimensionality and removing redundant information, so that the parameter number is reduced, the network complexity is simplified, the AlexNet model uses Overlapping pooling (overlaying) on the basis of the maximum pooling layer, the output of the pooling layer is expanded into multi-stage smaller features, sparse coding is adopted for fusion of the multi-stage features, and the feature dimensionality output by the pooling layer is reduced. In addition, a Dropout layer is also used, namely in the training process of the deep learning network, for a neural network unit, the neural network unit is temporarily discarded from the network according to a certain probability, dropout has a very good effect on preventing overfitting, parameters of the model can be reduced, and the generalization capability of the model can be enhanced.
TABLE 1JCM-AlexNet model Structure and corresponding training parameters
Different super parameters are selected in deep learning, different loss functions and optimization functions can obtain different results, the super parameters are variables determined according to experience and the effect of a verification set on a model, and a super parameter combination with the highest detection probability is obtained through a simulation experiment, as shown in table 1. Different loss functions and optimization functions may also affect the final performance of the algorithm, but since binary _ cross-sensitivity loss functions and category _ cross-sensitivity loss functions can be used for processing the binary tasks, and different category _ cross-sensitivity loss functions can also be used for the multi-classification tasks, the selection of category _ cross-sensitivity loss functions or category _ cross-sensitivity loss functions herein is almost indistinguishable, but different optimizers have a greater impact on the detection probability because the optimizers are SGDs, where gradient updates are frequent, resulting in severe loss function oscillations, and eventually stay at a local minimum or saddle, resulting in a decrease in the detection probability; the RMSProp optimizer adds second-order momentum on the basis of the SGD and calculates the second-order momentum by using a window sliding weighted average value, so that the problems of local minimum and saddle points are solved; in addition, the Adam optimizer integrates the first-order momentum of SGD and the second-order momentum of RMSProp, the first-order momentum can reduce the parameter updating speed, so that oscillation is reduced, the parameter updating can be accelerated when the gradient direction is the same, so that convergence is accelerated, the second-order momentum can solve the problems of local minimum values and saddle points, the Adam optimizer integrates the advantages of the first-order momentum and the second-order momentum, and the Adam optimizer has strong robustness on selection of hyper-parameters. Thus, there is a higher accuracy with Adam optimizers.
In summary, a parameter combination which enables the detection probability of the four algorithms to be the highest is obtained through a simulation experiment, so that fairness of comparison of the detection probabilities of the four algorithms including MLP, LSTM, leNet5 and JCM-AlexNet in performance analysis is guaranteed.
TABLE 2 probability of detection of model parameters and correlation functions at SNR = -15dB
The detection probabilities of the JCM-AlexNet, leNet5, LSTM, MLP and PSD algorithms under different signal-to-noise ratios and different false alarm probabilities are compared through simulation, then the change conditions of the loss values of the JCM-AlexNet algorithms under different epochs and different signal-to-noise ratios are analyzed, and finally the JCM-AlexNet algorithm has better performance.
Detection probability performance analysis at different signal-to-noise ratios
To evaluate the performance of the proposed algorithm, we chose to compare the detection probabilities of the five algorithms PSD, MLP, LSTM, leNet5, JCM-AlexNet in the SNR range of [ -20dB,0dB ] at a false alarm probability of 0.01, as shown in FIG. 4. It can be seen from the figure that at high snr, the five algorithms all have higher detection probability, but at low snr, the proposed JCM-AlexNet algorithm has higher detection probability, followed by LeNet5, e.g. at snr of-15 dB, the detection probability of JCM-AlexNet and LeNet5 algorithms can reach 99.8% and 95.3%, respectively, while PSD is only 28%, MLP is 63.5%, and LSTM is 75.2%. The algorithm is used for learning the polarization domain characteristics of the signals, so that more information of the signals can be utilized, and AlexNet and LeNet5 algorithms are good at characteristic extraction of matrix data, so that the JCM-AlexNet and LeNet5 algorithms have higher detection probability compared with other algorithms.
Detection probability performance analysis under different false alarm probabilities
To further evaluate the performance of the proposed algorithm, we chose to compare the detection probabilities of five algorithms PSD, MLP, LSTM, leNet5, JCM-AlexNet in the false alarm probability range [0,0.1] when SNR = -15dB, as shown in fig. 5. It can be seen from the figure that both JCM-AlexNet and LeNet5 have higher detection probability at high false alarm probability, e.g. at 0.05, the detection probability can reach 100%, while PSD is only 75.3%, MLP is 92.5%, and LSTM is 96.8%. However, the detection probability of the proposed JCM-AlexNet algorithm is more obvious than that of other algorithms when the false alarm probability is low, for example, when the false alarm probability is 0.01, the JCM-AlexNet detection probability can reach 99.8%, the PSD is only 28.2%, the MLP is 63.5%, the LSTM is 75.2%, and the LeNet5 is 95.3%. As described above, the AlexNet and LeNet5 algorithms are good at extracting features of matrix data, and therefore have higher detection probability than other algorithms. However, compared with the LeNet5 algorithm, the AlexNet algorithm adds a dropout layer behind a full connection layer, so that the parameter quantity is reduced when a deeper network layer is provided, overfitting can be prevented, in addition, a ReLu activation function is used for replacing a Sigmod activation function to optimize input nonlinearity, the calculation speed of the model is accelerated, however, the ReLu activation function is asymmetric and nonlinear, so that a Kaiming initialization method specially used for processing the asymmetric and nonlinear activation function is used for initializing the weight in the model at random, and gradient explosion or disappearance can be effectively prevented when the gradient of the activation function is calculated, and therefore, the JCM-AlexNet algorithm has more excellent detection performance.
Loss value variation analysis
In deep learning, a loss function is used for evaluating the deviation degree of a predicted value and a true value of a model, and the smaller the loss value is, the better the model performance is. Therefore, the invention analyzes the change of the loss function value of the JCM-AlexNet algorithm, thereby verifying the performance of the JCM-AlexNet algorithm.
A. Analysis of the variation of loss values at different epochs
In deep learning, one epoch represents that all data are input into the model, a forward propagation and backward propagation process is completed, the number of times of updating and iteration of the model weight is increased along with the increase of the number of epochs, a curve enters an optimized fitting state from an initial unfixed state, and finally an overfitting state is entered. Therefore, the AlexNet model is simulated when the epoch is 50, 100 and 200, and the optimal epoch is obtained by drawing a loss-epoch curve.
As can be seen from fig. 6, the variation of the loss value at a high snr is not significantly different, which enables a better fit of the AlexNet model, but at a low snr, the loss value is greater, i.e. the fit is worse, at epochs of 50 and 100 compared to epoch of 200. It can therefore be concluded that the AlexNet model can fit data better when the epoch is 200.
B. Loss value variation analysis under different signal-to-noise ratios
From the previous analysis, it can be seen that the AlexNet model can better fit the data when the epoch is 200, so we analyze the change of the proposed JCM-AlexNet algorithm at different signal-to-noise ratios when the epoch is 200. In addition, as the loss value is in a descending trend when the signal-to-noise ratio is between 0dB and 19dB, and the loss value is not obviously changed when the signal-to-noise ratio is 20 minus, in order to obviously distinguish the process that the loss value is reduced to the loss value and is not changed, the signal-to-noise ratio is selected to be between 0dB, 10dB, 18dB, 19dB and 20dB to represent the change condition of the loss value.
As shown in fig. 7, it can be clearly seen that when the signal-to-noise ratio is 0dB, the loss value is small, and the fluctuation change is small, which indicates that the model can correctly classify the signal and the noise at the initial stage of training; when the signal-to-noise ratio is-10 dB and-18 dB, the loss value is continuously reduced along with the training of the model, so that the model can not better classify the signals and the noise at the initial training stage, but parameters of the model are continuously optimized along with the training process of the model, and finally the signals and the noise can be correctly classified; when the signal-to-noise ratio is-19 dB, it can be seen that the loss value is large at the initial stage of model training, and the loss value is reduced and cannot reach the lowest value in the training process, so that signals and noise cannot be completely and correctly classified; in addition, the loss value hardly changes at a signal-to-noise ratio of-20 dB. It follows that the proposed algorithm also has good performance at signal-to-noise ratios of-18 dB.
In conclusion, the polarization information is combined with the deep learning technology, namely a Jones vector covariance matrix (JCM) which is subjected to normalization and other processing is converted into an image and then is used as the input of an AlexNet model, so that a JCM-AlexNet algorithm is provided for spectrum sensing; the invention uses the orthogonal dual-polarized antenna to receive signals, uses the Jones vector to represent the polarization state of the signals, and can use more signal characteristics to sense the frequency spectrum (such as the amplitude, the phase and the direction of the signals), thereby improving the performance of frequency spectrum sensing; the method comprises the steps of preprocessing polarization signals to convert the polarization signals into a polarization covariance matrix, and converting a spectrum sensing problem into an image processing problem according to the characteristics of the covariance matrix; the invention also designs the test statistic of the JCM-AlexNet algorithm according to Neyman Pearson's theorem (NP), so that the algorithm can set a threshold value through the false alarm probability, thereby realizing the verification of the performance of the spectrum sensing algorithm under the conditions of different false alarm probabilities.
The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like that fall within the spirit and principle of the present invention are intended to be included therein.
Claims (7)
1. A method of spectrum sensing, characterized by: the method comprises the following steps:
step 1: carrying out polarization data sampling and data preprocessing, comprising the following steps:
receiving a signal using an orthogonal dual-polarized antenna;
the received polarization signals are converted into covariance matrixes through data preprocessing, and further converted into images for storage;
step 2: carrying out online training based on an AlexNet model, comprising the following steps:
an image obtained through data preprocessing is used as input data and is submitted to an AlexNet model learning characteristic;
constructing test statistics according to the output result of the AlexNet model by utilizing a Neyman Pearson criterion;
and step 3: setting a detection threshold, comprising:
the noise image obtained through data preprocessing is delivered to a trained AlexNet model;
sequencing and constructing the obtained results, and setting false alarm probability according to actual needs so as to set a detection threshold;
and 4, step 4: carrying out off-line detection based on an AlexNet model, comprising the following steps:
test data are sent to an AlexNet model for testing;
and (4) constructing corresponding test statistics according to the test result, and comparing the test statistics with the detection threshold set in the step (3) so as to judge whether the authorized user exists.
2. The spectrum sensing method according to claim 1, wherein: and the step of setting the detection threshold value is to send the noise sample gray level image to a trained AlexNet model for detection, and to set the false alarm probability as required, so that a proper threshold value is obtained according to the false alarm probability.
3. The spectrum sensing method according to claim 1, wherein: the AlexNet model comprises 5 convolutional layers and 3 fully-connected layers, wherein the convolutional layers are used for extracting the characteristics of a polarization covariance matrix, the fully-connected layers are used for mapping the learned characteristics to a space of sample marks, and a ReLu activation function or a maximum pooling layer is used after the convolutional layers and the fully-connected layers.
4. The spectrum sensing method according to claim 1, wherein: in the step 4, during off-line detection, a sample gray scale image without label labeling is collected as test data, the test data is transmitted to a trained AlexNet model, then AlexNet-LRT detection is carried out based on an unlabeled sample, the maximum posterior probability is converted into a conditional probability to deduce a Likelihood Ratio Test (LRT) of a detection process, and after test statistics are obtained in a detection stage, the test statistics can be compared with a preset detection threshold value, so that whether an authorized user exists or not can be quickly judged.
5. A spectrum sensing apparatus, comprising: comprises that
Polarization data sampling and preprocessing device for polarization data sampling and data preprocessing include: receiving a signal using an orthogonal dual-polarized antenna;
the received polarization signals are converted into covariance matrixes through data preprocessing, and further converted into images for storage;
the online training device is used for performing online training based on an AlexNet model, and comprises:
taking an image obtained through data preprocessing as input data and delivering the input data to an AlexNet model learning characteristic;
constructing test statistics according to the output result of the AlexNet model by utilizing a Neyman Pearson criterion;
threshold setting means for setting a detection threshold, comprising:
the noise image obtained through data preprocessing is delivered to a trained AlexNet model;
sequencing and constructing the obtained results, and setting false alarm probability according to actual needs so as to set a detection threshold;
the off-line detection device is used for carrying out off-line detection based on an AlexNet model, and comprises:
test data are sent to an AlexNet model for testing;
and constructing corresponding test statistic according to the test result, and comparing the test statistic with the detection threshold set by the threshold setting device so as to judge whether the authorized user exists.
6. An electronic device, characterized in that: comprises at least one processor, an
At least one memory communicatively coupled to the processor, wherein:
the memory stores program instructions executable by the processor, the processor invoking the program instructions to perform the method of any of claims 1-4.
7. A computer-readable storage medium, characterized in that: the computer-readable storage medium has stored therein a computer program which, when being executed by a processor, carries out the method steps of any one of claims 1 to 4.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211083076.6A CN115438700A (en) | 2022-09-06 | 2022-09-06 | Spectrum sensing method and device, electronic equipment and medium |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211083076.6A CN115438700A (en) | 2022-09-06 | 2022-09-06 | Spectrum sensing method and device, electronic equipment and medium |
Publications (1)
Publication Number | Publication Date |
---|---|
CN115438700A true CN115438700A (en) | 2022-12-06 |
Family
ID=84247745
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202211083076.6A Pending CN115438700A (en) | 2022-09-06 | 2022-09-06 | Spectrum sensing method and device, electronic equipment and medium |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN115438700A (en) |
-
2022
- 2022-09-06 CN CN202211083076.6A patent/CN115438700A/en active Pending
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Zheng et al. | DL-PR: Generalized automatic modulation classification method based on deep learning with priori regularization | |
CN109379153B (en) | Spectrum sensing method | |
CN113050042A (en) | Radar signal modulation type identification method based on improved UNet3+ network | |
US11870619B2 (en) | Fast modulation recognition method for multilayer perceptron based on multimodally-distributed test data fusion | |
CN114268388B (en) | Channel estimation method based on improved GAN network in large-scale MIMO | |
Liu et al. | Deep learning-based spectrum sensing in space-air-ground integrated networks | |
CN113315593A (en) | Frequency spectrum sensing algorithm based on FLOM covariance matrix and LSTM neural network | |
CN109450573A (en) | A kind of frequency spectrum sensing method based on deep neural network | |
CN111817803A (en) | Frequency spectrum sensing method and system based on correlation coefficient and K-means clustering algorithm and computer readable storage medium | |
CN114943245A (en) | Automatic modulation recognition method and device based on data enhancement and feature embedding | |
CN113205140A (en) | Semi-supervised specific radiation source individual identification method based on generative countermeasure network | |
CN112232395A (en) | Semi-supervised image classification method for generating confrontation network based on joint training | |
CN116192307A (en) | Distributed cooperative multi-antenna cooperative spectrum intelligent sensing method, system, equipment and medium under non-Gaussian noise | |
CN114925720A (en) | Small sample modulation signal identification method based on space-time mixed feature extraction network | |
CN114757224A (en) | Specific radiation source identification method based on continuous learning and combined feature extraction | |
CN115276857B (en) | Full-blind spectrum detection method based on combination of Cholesky decomposition and convolutional neural network | |
CN115438700A (en) | Spectrum sensing method and device, electronic equipment and medium | |
CN115983390B (en) | Edge intelligent reasoning method and system based on multi-antenna aerial calculation | |
US6535641B1 (en) | Class specific classifier | |
CN115270891A (en) | Method, device, equipment and storage medium for generating signal countermeasure sample | |
CN113869238A (en) | Cognitive Internet of vehicles intelligent frequency spectrum sensing method and system | |
CN112968740B (en) | Satellite spectrum sensing method based on machine learning | |
Xiang et al. | Automatic modulation recognition based on SST and deep learning model | |
CN115378475A (en) | Method for estimating number of transmitting antennas based on convolutional neural network | |
CN115047422A (en) | Radar target identification method based on multi-scale mixed hole convolution |
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 |