CN113315593A - Frequency spectrum sensing algorithm based on FLOM covariance matrix and LSTM neural network - Google Patents

Frequency spectrum sensing algorithm based on FLOM covariance matrix and LSTM neural network Download PDF

Info

Publication number
CN113315593A
CN113315593A CN202110553337.5A CN202110553337A CN113315593A CN 113315593 A CN113315593 A CN 113315593A CN 202110553337 A CN202110553337 A CN 202110553337A CN 113315593 A CN113315593 A CN 113315593A
Authority
CN
China
Prior art keywords
neural network
lstm neural
spectrum sensing
data
covariance matrix
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
Application number
CN202110553337.5A
Other languages
Chinese (zh)
Inventor
赵韵雪
朱晓梅
李义丰
朱艾春
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing Tech University
Original Assignee
Nanjing Tech University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Nanjing Tech University filed Critical Nanjing Tech University
Priority to CN202110553337.5A priority Critical patent/CN113315593A/en
Publication of CN113315593A publication Critical patent/CN113315593A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B17/00—Monitoring; Testing
    • H04B17/30—Monitoring; Testing of propagation channels
    • H04B17/382—Monitoring; Testing of propagation channels for resource allocation, admission control or handover
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • 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
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W16/00—Network planning, e.g. coverage or traffic planning tools; Network deployment, e.g. resource partitioning or cells structures
    • H04W16/14—Spectrum sharing arrangements between different networks
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W72/00—Local resource management
    • H04W72/04—Wireless resource allocation
    • H04W72/044—Wireless resource allocation based on the type of the allocated resource
    • H04W72/0453—Resources in frequency domain, e.g. a carrier in FDMA
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W72/00—Local resource management
    • H04W72/50—Allocation or scheduling criteria for wireless resources
    • H04W72/53—Allocation or scheduling criteria for wireless resources based on regulatory allocation policies
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D30/00—Reducing energy consumption in communication networks
    • Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Data Mining & Analysis (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Electromagnetism (AREA)
  • Monitoring And Testing Of Transmission In General (AREA)

Abstract

The invention discloses a frequency spectrum sensing algorithm based on a FLOM covariance matrix and an LSTM neural network, which belongs to the technical field of cognitive radio. The invention is scientific and reasonable, is safe and convenient to use, improves the detection performance by the strong capability of the FLOM in reducing the influence of non-Gaussian characteristics and the strong processing capability of the LSTM neural network in extracting data time sequence characteristics, effectively solves the problem of spectrum sensing under the condition that non-Gaussian noise has no energy and other second-order statistics used for detection, improves the spectrum sensing performance, and has better performance compared with other networks under the condition of low signal-to-noise ratio.

Description

Frequency spectrum sensing algorithm based on FLOM covariance matrix and LSTM neural network
Technical Field
The invention relates to the technical field of cognitive radio, in particular to a frequency spectrum sensing algorithm based on a FLOM covariance matrix and an LSTM neural network.
Background
The cognitive radio can reasonably utilize idle wireless channel resources, and is one of important ways for solving the problem of lack of wireless spectrum, and spectrum sensing is one of key technologies of cognitive radio. The task of spectrum sensing is to identify the frequency bands that fit the sub-users. In order to find a more efficient cognitive radio resource allocation scheme, a machine learning method is also widely used for solving the spectrum sensing problem.
Currently, many model-driven spectrum sensing schemes have been proposed. However, most spectrum sensing algorithms are based on gaussian noise environment, but the performance is superior in gaussian noise environment, and the noise in the actual wireless communication channel is affected by natural factors and human factors, and the noise contains "impulse noise" which is not consistent with the characteristics of gaussian noise distribution, and the noise is non-gaussian noise. The data-driven spectrum sensing algorithms depend on the extracted second-order statistic information to a great extent, and non-Gaussian noise has no second-order statistic due to the distribution characteristics of the non-Gaussian noise and only has fractional low-order moment, so that when spectrum detection is carried out in the non-Gaussian noise, the second-order statistic for detection is not available. In this case, the conventional spectrum sensing algorithm has a performance degradation phenomenon or even an ineffective phenomenon. Therefore, a spectrum sensing algorithm based on the FLOM covariance matrix and the LSTM neural network is needed to solve the above problems.
Disclosure of Invention
The purpose of the invention is as follows: the invention aims to provide a frequency spectrum sensing algorithm based on a FLOM covariance matrix and an LSTM neural network to solve the problem that performance of the algorithm is reduced or even invalid when frequency spectrum monitoring is carried out in non-Gaussian noise.
The technical scheme is as follows: the invention discloses a frequency spectrum sensing algorithm based on a FLOM covariance matrix and an LSTM neural network, which is used for improving the frequency spectrum sensing performance under the non-Gaussian noise environment and comprises the following steps:
s10, collecting digital modulation signal samples under different signal-to-noise ratios and Alpha noise samples without user signals as original data for the equipment to be detected; dividing the acquisition result by adopting quintuple cross validation, dividing the original data into five groups, sequentially selecting one group as a test set, and taking the other four groups as training sets for training so as to obtain five models;
s20, preprocessing samples in the training set by using FLOM (fractional low-order moment) to obtain a fractional low-order moment covariance matrix, namely obtaining a sample set; marking the sample set to obtain a label set;
s30, constructing an LSTM neural network model, wherein the LSTM neural network model comprises an input layer, an LSTM layer, a full connection layer and a Softmax function module, and mapping the output of the LSTM neural network model to a spectrum sensing result through the full connection layer and the Softmax function module after extracting high-dimensional features by the LSTM layer; initializing structural parameters of the LSTM neural network, wherein the structural parameters comprise input layer node number, hidden element dimension, data vector dimension, iteration times and learning rate; in the construction process, the flow of control information of a forgetting gate, an input gate and an output gate is respectively constructed, PU weight matrixes and offsets of all the gates are different, and the PU weight matrixes and the offsets are used for controlling the information flow on the RNN; adding three information control units in a hidden layer unit of an LSTM neural network according to different places for controlling information circulation so as to reserve useful information and remove useless information;
s40, inputting the label set into an LSTM neural network for learning, and calculating the loss value of the label set through a trained loss function; back propagation is carried out by using a random gradient descent method until iteration is finished;
s50, drawing a loss curve according to the loss value, and drawing a test set accuracy curve according to the detection probability of the algorithm performance under different false alarm probabilities; observing a loss curve and a test set accuracy curve, increasing hidden element dimensionality and reducing learning rate under the non-convergence state of the curve, and executing S30, and executing the next step when the curve is in the convergence state;
s60, inputting the test set into the learned LSTM neural network, obtaining an output result by a softmax module, and comparing the output result with a threshold value; when the output result is smaller than the threshold value, the frequency band is not used by the master user and can be accessed by the master user for communication; when the output result is greater than the threshold, the primary user is using the frequency band.
Preferably, the S20 includes the following steps:
s2001, processing samples in the training set by using FLOM:
Figure RE-GDA0003140254390000021
wherein the content of the first and second substances,
Figure RE-GDA0003140254390000022
for processed data, zj(n) is a sample in the training set, p is the order of the corresponding fractional low-order moment, and Alpha is an Alpha stable distribution noise characteristic index;
s2002, obtaining a fraction low-order moment covariance matrix after data preprocessing according to the definition of covariance:
Figure RE-GDA0003140254390000023
Figure RE-GDA0003140254390000031
wherein the content of the first and second substances,
Figure RE-GDA0003140254390000032
is a calculation method of a score low-order moment covariance matrix,
Figure RE-GDA0003140254390000033
represents the fractional low order moment of the data, and H represents the conjugate transpose of the matrix;
thus, a sample set of
Figure RE-GDA0003140254390000034
S2003, labeling the obtained sample set, wherein labels of the digital modulation signal samples and Alpha noise samples are respectively [1, 0 ]]And [0, 1 ]]The labeled tag set is
Figure RE-GDA0003140254390000035
Wherein the content of the first and second substances,
Figure RE-GDA0003140254390000036
fractional low-order moment covariance matrix, y, representing datamIndicating label [ y1，y2，...ym...yM]。
Preferably, in S30, the number of nodes in the input layer is 200, the number of nodes in the hidden layer is 100, the hidden element dimension is 10, the data vector dimension is 20, the number of iterations is 600, and the learning rate is 5 × 10-6。
Preferably, the S40 includes the following steps:
s4001, concentrating the labels
Figure RE-GDA0003140254390000037
Inputting LSTM neural network, and outputting training through Softmax function module
Figure RE-GDA0003140254390000038
Wherein
Figure RE-GDA0003140254390000039
The detection probability of the algorithm under the two conditions of the existence of the used frequency band of the master user is represented;
s4002, conditional probability expressions assuming that the primary user is using the frequency band and the primary user does not use the frequency band, respectively:
Figure RE-GDA00031402543900000310
Figure RE-GDA00031402543900000311
Figure RE-GDA00031402543900000312
wherein H0Indicating that the primary user is not using the band, H1Indicating that the primary user is using the band,
Figure RE-GDA00031402543900000313
represents a probability in the case of Hi (i ═ 0, 1);
s4003, training the label set to improve the possibility to the maximum extent, wherein the training formula is as follows:
L(θ)＝P(ym＝1；θ)ymP(ym＝0；θ)1-ym
wherein, P (y)m1 is ═ 1; θ) represents data sample H1Probability of ymIs a data sample H1Label of (2), P (y)m0; θ) represents data sample H0Probability of (1-y)m) Is a data sample H0The label of (1);
s4004, calculating a loss value of training:
Figure RE-GDA0003140254390000041
wherein, ymIs a data sample H1Tag [1, 0 ] of]，
Figure RE-GDA0003140254390000042
Represents a prediction of a data sample as H1The probability of (d); (1-y)m) Is a data sample H0Tag [0, 1 ] of]，
Figure RE-GDA0003140254390000043
Represents a prediction of a data sample as H0The probability of (d);
s4005, adopting a random gradient descent method for back propagation, and optimizing theta to obtain a maximum P (y; theta) and a minimum J (theta) until the iteration is finished.
Preferably, in S60: the output result is a test
Figure RE-GDA0003140254390000044
The threshold value is 0.5.
Has the advantages that: preprocessing samples in a training set by using FLOM, building an LSTM neural network model, transmitting the training set for training to the LSTM neural network for learning, finally inputting a test set into the learned LSTM neural network, obtaining an output result by a softmax module, comparing the output result with a threshold value, and making a decision whether a master user exists or not; the detection performance is improved through the strong capability of the FLOM in the aspect of reducing the influence of non-Gaussian features and the strong processing capability of the LSTM neural network in the aspect of extracting data time sequence features, the spectrum sensing under the condition that non-Gaussian noise has no energy and other second-order statistics used for detection is effectively solved, the spectrum sensing performance is improved, and the spectrum sensing network has better performance compared with other networks under the condition of low signal-to-noise ratio. The invention trains the model by using BPSK signals and tests the network by using other signals such as 8PSK, QFSK and the like, so that the trained network model can adapt to different new signals, and the method has wide application range.
Drawings
FIG. 1 is a schematic flow chart of the algorithm of the present invention;
FIG. 2 is a ROC plot of the algorithm of the present invention under different activation functions;
FIG. 3 is an ROC curve of the algorithm of the present invention under different sampling point conditions;
FIG. 4 is a ROC curve for the algorithm of the present invention under different channel conditions;
FIG. 5 is a graph comparing the performance of different algorithms under Alpha noise in accordance with the present invention;
FIG. 6 is a ROC plot of the algorithm of the present invention under different signal-to-noise ratios;
FIG. 7 is a graph of model training performance for the present invention;
FIG. 8 is a ROC plot of the algorithm of the present invention under different test signals;
Detailed Description
The technical solution of the present invention is described in detail below with reference to the accompanying drawings, but the scope of the present invention is not limited to the embodiments.
A frequency spectrum sensing algorithm based on a FLOM covariance matrix and an LSTM neural network is used for improving the frequency spectrum sensing performance under the non-Gaussian noise environment, and comprises the following steps:
step 10, collecting digital modulation signal samples under different signal-to-noise ratios and Alpha noise samples without user signals as original data aiming at equipment to be detected; dividing the acquisition result by adopting quintuple cross validation, dividing the original data into five groups, sequentially selecting one group as a test set, and taking the other four groups as training sets for training so as to obtain five models;
step 20, preprocessing samples in the training set by using FLOM (fractional low-order moment) to obtain a fractional low-order moment covariance matrix, namely obtaining a sample set; marking the sample set to obtain a label set;
step 30, constructing an LSTM neural network model, wherein the LSTM neural network model comprises an input layer, an LSTM layer, a full connection layer and a Softmax function module, and after extracting high-dimensional features, the LSTM layer maps the output of the LSTM neural network model to a spectrum sensing result through the full connection layer and the Softmax function module; initializing structural parameters of the LSTM neural network, wherein the structural parameters comprise input layer node number, hidden element dimension, data vector dimension, iteration times and learning rate; in the construction process, the control information circulation of a forgetting gate, an input gate and an output gate is respectively constructed, PU weight matrixes and offsets of all the gates are different, and the control information circulation is controlled by the aid of the PU weight matrixes and the offsets which are arranged on an RNN (recurrent neural network); adding three information control units in a hidden layer unit of an LSTM neural network according to different places for controlling information circulation so as to reserve useful information and remove useless information;
step 40, inputting the label set into an LSTM neural network for learning, and calculating the loss value of the label set through a trained loss function; back propagation is carried out by using a random gradient descent method until iteration is finished so as to reduce the loss value of the loss function between the final test value and the actual value;
step 50, drawing a loss curve according to the loss value, and drawing a test set accuracy curve according to the detection probability of the algorithm performance under different false alarm probabilities; observing a loss curve and a test set accuracy curve, increasing hidden element dimensionality and reducing learning rate under the non-convergence state of the curve, and executing S30, and executing the next step when the curve is in the convergence state;
step 60, inputting the test set into the learned LSTM neural network, obtaining an output result by a softmax module, and comparing the output result with a threshold value; when the output result is smaller than the threshold value, the frequency band is not used by the master user and can be accessed by the master user for communication; when the output result is greater than the threshold, the primary user is using the frequency band.
Step 20 comprises the steps of:
step 2001, processing the samples in the training set by using FLOM:
Figure RE-GDA0003140254390000061
wherein the content of the first and second substances,
Figure RE-GDA0003140254390000062
for processed data, zj(n) is a sample in the training set, p is the order of the corresponding fractional low-order moment, and Alpha is an Alpha stable distribution noise characteristic index;
step 2002, obtaining a score low-order moment covariance matrix after data preprocessing according to the definition of covariance:
Figure RE-GDA0003140254390000063
Figure RE-GDA0003140254390000064
wherein the content of the first and second substances,
Figure RE-GDA0003140254390000065
is a calculation method of a score low-order moment covariance matrix,
Figure RE-GDA0003140254390000066
represents the fractional low order moment of the data, and H represents the conjugate transpose of the matrix;
thus, a sample set of
Figure RE-GDA0003140254390000067
Step S2003, labeling the obtained sample set, wherein labels of the digital modulation signal sample and the Alpha noise sample are respectively [1, 0 ]]And [0, 1 ]]The labeled tag set is
Figure RE-GDA0003140254390000068
Wherein the content of the first and second substances,
Figure RE-GDA0003140254390000069
fractional low-order moment covariance matrix, y, representing datamIndicating label [ y1，y2，...ym...yM]。
In step 30, the number of nodes of the input layer is set to be 200, the number of nodes of the hidden layer is set to be 100, the hidden element dimension is 10, the data vector dimension is 20, the iteration frequency is 600, and the learning rate is 5 x 10-6。
Step 40 comprises the steps of:
step 4001, centralizing the labels
Figure RE-GDA00031402543900000610
Inputting LSTM neural network, and outputting training through Softmax function module
Figure RE-GDA0003140254390000071
Wherein
Figure RE-GDA0003140254390000072
The detection probability of the algorithm under the two conditions of the existence of the used frequency band of the master user is represented;
step 4002, assuming conditional probability expressions that the primary user is using the frequency band and the primary user does not use the frequency band, respectively:
Figure RE-GDA0003140254390000073
Figure RE-GDA0003140254390000074
Figure RE-GDA0003140254390000075
wherein H0Indicating that the primary user is not using the band, H1Indicating that the primary user is using the band,
Figure RE-GDA0003140254390000076
represents a probability in the case of Hi (i ═ 0, 1);
step 4003, training the label set to maximize the probability, the training formula is as follows:
L(θ)＝P(ym＝1；θ)ymP(ym＝0；θ)1-ym
wherein, P (y)m1 is ═ 1; θ) represents data sample H1Probability of ymIs a data sample H1Label of (2), P (y)m0; θ) represents data sample H0Probability of (1-y)m) Is a data sample H0The label of (1);
step 4004, calculating a loss value of training:
Figure RE-GDA0003140254390000077
wherein, ymIs a data sample H1Tag [1, 0 ] of]，
Figure RE-GDA0003140254390000078
Represents a prediction of a data sample as H1The probability of (d); (1-y)m) Is a data sample H0Tag [0, 1 ] of]，
Figure RE-GDA0003140254390000079
Represents a prediction of a data sample as H0The probability of (c).
Step 4005, using a stochastic gradient descent method to back-propagate, and optimizing θ to obtain a maximum P (y; θ) and a minimum J (θ), until the iteration is over.
In step 60: the output result is a test
Figure RE-GDA0003140254390000081
The threshold value is 0.5.
Example 1: as shown in fig. 2-4, the detection probability (P)d) And false alarm probability (P)f) Is a measure of the spectrum sensing algorithm. On the premise that the false alarm probability is fixed, if the detection probability is higher than 90%, the performance of the perception algorithm is considered to be good. The experimental simulation was based on MATLAB and Pytorch. The training and test sets were generated by MATLAB, using a Pytrch, scikitleren library for programming the algorithm under the ubuntu16.04 operating system. The simulation data set parameters are shown in table 1:
TABLE 1 simulation dataset parameters
Figure RE-GDA0003140254390000082
The FLOM pre-processed data was first input into a modified LSTM neural network. When algorithm performance comparison is carried out, under the condition that other conditions are the same, data which are not subjected to fractional low-order preprocessing and data which are subjected to fractional low-order preprocessing are input into an improved LSTM neural network, a DNN (deep neural network) spectrum sensing algorithm and a CNN (convolutional neural network) spectrum sensing algorithm, and the influence of network parameters on detection performance is analyzed.
The activation function can increase the nonlinearity of the model, improve the modeling precision of the complex function and reduce the weightAnd (5) overfitting. Sigmoid (S-shaped growth curve function), tanh (hyperbolic tangent function) and ReLU (linear rectification function) are used as activation functions of the hidden layer and are respectively used for training the network, and finally, an ROC (receiver operating characteristic) curve of spectrum sensing is obtained, as shown in fig. 2, a model obtained by using ReLU as the activation function has high convergence speed and highest detection probability. Thus, ReLU is chosen as the activation function of the model hidden layer. In particular, when GSNR is 0dB, P (fractional lower moment order) is 0.7, Pf(false alarm probability) 0.5, improved LSTM algorithm Pd(detection probability) is 87%, and P with sigmoid and tanh as activation functionsd75% and 78%, respectively. Improved P of CNN algorithm with activation function of ReLUdUp to 95%, while the activation functions are about 85% and 87% for sigmoid and tanh, respectively.
As can be seen from fig. 3, increasing the sampling points can improve the detection performance of the model. This is because, as the number of sampling points increases, the data features become more and more obvious, and the neural network will be more accurate and efficient in feature extraction, classification and decision making. When GSNR is 0dB, P is 0.7, PfP of modified LSTM algorithm for 1000 samples at 0.1dP of up to 78% and 100 sample pointsdIt is only 65%. P of improved DNN algorithm of 1000 sampling pointsdUp to 73%, and P of 100 samplesdIt is only 70%.
As can be seen from fig. 4, channel fading has a certain impact on the detection performance. When the GSNR is high, the fading of the channel adds noise and the user signal is still much larger than pure noise data, which is not characterized by the overall signal characteristics. However, in the case of a low signal-to-noise ratio, the overall characteristics of the signal are substantially similar to the overall characteristics of noise, and channel fading exists, so that the data characteristics of the user signal occupy a small proportion. The accuracy can reach 70% under the condition of high signal-to-noise ratio no matter whether the user is faded or not. Under the condition of low signal-to-noise ratio, the detection probability of the fading of the channel plus the fading of non-Gaussian noise is only 55%. Under the condition of low signal-to-noise ratio, when GSNR is-10 dB, P is 0.7, PfImproved LSTM algorithm P without channel fading at 0.01dUp to88% with improved LSTM algorithm P with channel fadingdIt is only 60%. P of improved CNN algorithm without channel fading under same conditiond80% P of the modified CNN algorithm with channel fadingdIt is only 60%. In summary, when the signal data set is subjected to channel fading, the detection performance of the three neural network models is reduced, and it can be concluded from the above graph that the improved LSTM algorithm based on the FLOM perception has better detection performance than the detection method based on DNN and CNN when the signal-to-noise ratio is low.
The detection times for the different algorithms are shown in table 2. For the spectrum sensing technique mentioned in table 2, the sensing time of the proposed sensing scheme was evaluated. Model-driven spectrum sensing requires about 7.8 × 103 seconds, while the LSTM neural network herein operates about 10 epochs, requiring only 48 seconds to obtain the final detection result. The result shows that the scheme is superior to the spectrum sensing performance based on the model driving method in detection time and detection complexity.
TABLE 2 time comparison of the two algorithms
Figure RE-GDA0003140254390000091
Example 2: as shown in fig. 5-6, comparing the frequency spectrum sensing algorithm based on the flo m matrix sensing LSTM neural network with the DNN and CNN frequency spectrum sensing algorithms based on the flo m matrix sensing, and the LSTM frequency spectrum sensing algorithm, the DNN frequency spectrum sensing algorithm, the CNN frequency spectrum sensing algorithm, and the energy detection algorithm:
in the simulation, the detection threshold is initially set to η 0.5, and then a decision is made by comparison with the threshold. Finally, based on the monte carlo implementation, the probability detection values can be obtained under different conditions.
FIG. 5 is a comparison of performance of different algorithms under Alpha noise. It can be seen from fig. 8 that in the background of Alpha noise, the detection performance of the energy detection algorithm, DNN algorithm, CNN algorithm and LSTM algorithm is almost ineffective, and when the low-order processing is performed, the spectrum sensing performance of the algorithm proposed by the present invention is significantly better than that of the energy detection algorithm and the original DNN algorithm,CNN, LSTM algorithm. In other words, the spectrum sensing performance of the spectrum sensing method combined with the low-order processing is significantly better than that of the algorithm proposed in the early reference, and further results show that the improved LSTM algorithm based on the flo perception has better detection performance than the detection method based on DNN, CNN under the condition of low signal-to-noise ratio. For example, P of the original LSTM algorithm under Alpha noise when GSNR is-10 dBdOnly 45%, and the P of the LSTM algorithm presented hereindUp to 85 percent; p of original DNN algorithmdOnly 40%, and P for improving DNN algorithmdCan be 65 percent. In conclusion, after the algorithm provided by the invention is subjected to low-order processing, the spectrum sensing performance is obviously superior to that of an energy detection algorithm and the original DNN, CNN and LSTM algorithms.
Fig. 6 shows ROC curves for improving the LSTM algorithm after fractional low order processing under different signal-to-noise ratios in a non-gaussian environment and when the channel is a non-fading channel.
Under the condition of low signal-to-noise ratio, the detection probability of the proposed algorithm subjected to the fractional low-order processing is 0.3 higher than that of the original DNN, CNN and LSTM detection methods. The reason for this is that the noise component is significantly larger than the signal component in case of a low signal-to-noise ratio. Thus, the dominant features extracted from the data are noise features, whereas non-gaussian noise without fractional low-order moment processing is featureless. This results in poor detection performance.
It can be seen from fig. 6 that the detection probability of the LSTM neural network proposed by the present invention is optimal under low signal-to-noise ratio conditions when other conditions are the same. For example, when GSNR is-20 dB, PfP of DNN at 0.1dOnly 45% and P of LSTMdUp to 60%. When GSNR is-20 dB, PfP of CNN at 0.5dOnly 68% of the total mass of LSTMdUp to 80%. Other conditions being equal, the detection probability of CNN is optimal at high signal-to-noise ratio. For example, when GSNR is 20dB, PfP of DNN at 0.6dIs 90%, and P of CNNdUp to 96%. In conclusion, the FLOM perception-based LSTM detection system has better detectability than the other two networks under the condition of low signal-to-noise ratioCan be used.
The training performance of the proposed LSTM algorithm, when processed through fractional low order processing, is shown in fig. 7. As the number of training steps increases, the updated weights improve the performance of the model. As the number of training steps increases, the updated weights will improve the performance of the model. As shown in fig. 7, the loss of the LSTM neural network proposed herein decreased from 0.7 to 0.35 after 10 cycles, and almost to 0.09 after 50 cycles, showing the best detection effect.
Example 3: as shown in fig. 8, in order to verify the capability of the neural network model trained by several signal types to detect other types of untrained signals, signal data of 8PSK and QFSK modulation modes are simulated and input to the neural network as a test sample of an unknown signal for performance analysis in consideration of the influence of the input signal on the performance of the algorithm. The results are shown in FIG. 8. From fig. 8, the following two conclusions can be drawn here:
firstly, the method comprises the following steps: under the condition that the sample size and the network model parameters are the same, the detection performance of the 8PSK modulation signal is best, and when the GSNR is-15 dB, the P is the maximumdStill up to over 0.8. The detection performance of the QFSK modulation signal is slightly poor, and when the GSNR is 15dB, P isdOnly 0.75. For example, when GSNR is-15 dB and P is 0.7, the LSTM at 8PSK has Pd80% and P of LSTM of QPSKdIt is only 65%. When GSNR is 0dB and P is 0.7, the P of LSTM of 8PSKd90% while the P of LSTM for QPSKdThe content was 78%. When GSNR is 15dB, p is 0.7. P of LSTM for 8PSKd95% P of LSTM in QPSKdThe content was 80%. It can be concluded from this that these signals can still be detected with a high probability, which further suggests that the improved methods herein can accommodate the detection of a variety of unknown signals.
II, secondly: under the condition of low signal-to-noise ratio, the FLOM perception-based LSTM detection system has better detection performance than other two networks. For example, the detection performance of the LSTM neural network can reach 90% when the GSNR is between-20 dB to 0dB, while the detection performance of the DNN and CNN algorithms is only 70%.
As noted above, while the present invention has been shown and described with reference to certain preferred embodiments, it is not to be construed as limited thereto. Various changes in form and detail may be made therein without departing from the spirit and scope of the invention as defined by the appended claims.

Claims (5)

1. A frequency spectrum sensing algorithm based on a FLOM covariance matrix and an LSTM neural network is used for improving the frequency spectrum sensing performance under the non-Gaussian noise environment, and is characterized by comprising the following steps:
s10, collecting digital modulation signal samples under different signal-to-noise ratios and Alpha noise samples without user signals as original data for the equipment to be detected; dividing the acquisition result by adopting quintuple cross validation, dividing the original data into five groups, sequentially selecting one group as a test set, and taking the other four groups as training sets for training so as to obtain five models;
s20, preprocessing samples in the training set by using FLOM to obtain a score low-order moment covariance matrix, namely obtaining a sample set; marking the sample set to obtain a label set;
s30, constructing an LSTM neural network model which comprises an input layer, an LSTM layer, a full connection layer and a Softmax function module, and mapping the output of the LSTM neural network to a spectrum sensing result through the full connection layer and the Softmax function module after extracting high-dimensional features by the LSTM layer; initializing structural parameters of the LSTM neural network, wherein the structural parameters comprise input layer node number, hidden element dimension, data vector dimension, iteration times and learning rate;
s40, inputting the label set into an LSTM neural network for learning, and calculating the loss value of the label set through a trained loss function; back propagation is carried out by using a random gradient descent method until iteration is finished;
s50, drawing a loss curve according to the loss value, and drawing a test set accuracy curve according to the detection probability of the algorithm performance under different false alarm probabilities; observing a loss curve and a test set accuracy curve, increasing hidden element dimensionality and reducing learning rate under the non-convergence state of the curve, and executing S30, and executing the next step when the curve is in the convergence state;
s60, inputting the test set into the learned LSTM neural network, obtaining an output result by a softmax module, and comparing the output result with a threshold value; when the output result is smaller than the threshold value, the frequency band is not used by the master user and can be accessed by the master user for communication; when the output result is greater than the threshold, the primary user is using the frequency band.
2. The FLOM covariance matrix and LSTM neural network-based spectrum sensing algorithm of claim 1, wherein S20 comprises the steps of:
s2001, processing samples in the training set by using FLOM:
Figure RE-FDA0003140254380000011
wherein the content of the first and second substances,
Figure RE-FDA0003140254380000012
for processed data, zj(n) is a sample in the training set, p is the order of the corresponding fractional low-order moment, and Alpha is an Alpha stable distribution noise characteristic index;
s2002, obtaining a fraction low-order moment covariance matrix after data preprocessing according to the definition of covariance:
Figure RE-FDA0003140254380000021
Figure RE-FDA0003140254380000022
wherein the content of the first and second substances,
Figure RE-FDA0003140254380000023
is a calculation method of a score low-order moment covariance matrix,
Figure RE-FDA0003140254380000024
represents the fractional low order moment of the data, and H represents the conjugate transpose of the matrix;
thus, a sample set of
Figure RE-FDA0003140254380000025
S2003, labeling the obtained sample set, wherein labels of the digital modulation signal samples and Alpha noise samples are respectively [1, 0 ]]And [0, 1 ]]The labeled tag set is
Figure RE-FDA0003140254380000026
Wherein the content of the first and second substances,
Figure RE-FDA0003140254380000027
fractional low-order moment covariance matrix, y, representing datamIndicating label [ y1，y2，...ym...yM]。
3. The flo m covariance matrix and LSTM neural network-based spectrum sensing algorithm of claim 1, wherein: in S30, the number of nodes in the input layer is set to 200, the number of nodes in the hidden layer is set to 100, the hidden element dimension is set to 10, the data vector dimension is set to 20, the iteration number is set to 600, and the learning rate is set to 5 × 10-6。
4. The FLOM covariance matrix and LSTM neural network-based spectrum sensing algorithm of claim 2, wherein S40 comprises the steps of:
s4001, concentrating the labels
Figure RE-FDA0003140254380000028
Inputting into LSTM neural network, and outputting as training through Softmax function module
Figure RE-FDA0003140254380000029
Wherein
Figure RE-FDA00031402543800000210
The detection probability of the algorithm under the two conditions of the existence of the used frequency band of the master user is represented;
s4002, conditional probability expressions assuming that the primary user is using the frequency band and the primary user does not use the frequency band, respectively:
H1:
Figure RE-FDA00031402543800000211
H0:
Figure RE-FDA00031402543800000212
Figure RE-FDA00031402543800000213
wherein H0Indicating that the primary user is not using the band, H1Indicating that the primary user is using the band,
Figure RE-FDA0003140254380000031
represents a probability in the case of Hi (i ═ 0, 1);
s4003, training the label set to improve the possibility to the maximum extent, wherein the training formula is as follows:
L(θ)＝P(ym＝1；θ)ymP(ym＝0；θ)1-ym
wherein, P (y)m1 is ═ 1; θ) represents data sample H1Probability of ymIs a data sample H1Label of (2), P (y)m0; θ) represents data sample H0Probability of (1-y)m) Is a data sample H0The label of (1);
s4004, calculating a loss value of training:
Figure RE-FDA0003140254380000032
wherein, ymIs a data sample H1Tag [1, 0 ] of]，
Figure RE-FDA0003140254380000033
Represents a prediction of a data sample as H1The probability of (d); (1-y)m) Is a data sample H0Tag [0, 1 ] of]，
Figure RE-FDA0003140254380000034
Represents a prediction of a data sample as H0The probability of (d);
s4005, adopting a random gradient descent method for back propagation, and optimizing theta to obtain a maximum P (y; theta) and a minimum J (theta) until the iteration is finished.
5. The FLOM covariance matrix and LSTM neural network-based spectrum sensing algorithm of claim 1, wherein in S60: the output result is a test
Figure RE-FDA0003140254380000035
The threshold value is 0.5.
CN202110553337.5A 2021-05-20 2021-05-20 Frequency spectrum sensing algorithm based on FLOM covariance matrix and LSTM neural network Pending CN113315593A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110553337.5A CN113315593A (en) 2021-05-20 2021-05-20 Frequency spectrum sensing algorithm based on FLOM covariance matrix and LSTM neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110553337.5A CN113315593A (en) 2021-05-20 2021-05-20 Frequency spectrum sensing algorithm based on FLOM covariance matrix and LSTM neural network

Publications (1)

Publication Number Publication Date
CN113315593A true CN113315593A (en) 2021-08-27

Family

ID=77373911

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110553337.5A Pending CN113315593A (en) 2021-05-20 2021-05-20 Frequency spectrum sensing algorithm based on FLOM covariance matrix and LSTM neural network

Country Status (1)

Country Link
CN (1) CN113315593A (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114337880A (en) * 2021-11-15 2022-04-12 宁波大学 Spectrum sensing method considering random arrival and departure of main user signal
CN114337883A (en) * 2021-12-28 2022-04-12 杭州电子科技大学 CNN cooperative spectrum sensing method and system based on covariance matrix Cholesky decomposition
CN114826500A (en) * 2022-04-21 2022-07-29 中国人民解放军国防科技大学 Constant envelope modulation burst signal detection method and system based on deep learning
CN115021845A (en) * 2022-06-20 2022-09-06 西安航空学院 Spectrum sensing method, system, medium, device and terminal

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200153535A1 (en) * 2018-11-09 2020-05-14 Bluecom Systems and Consulting LLC Reinforcement learning based cognitive anti-jamming communications system and method
CN111600667A (en) * 2020-05-25 2020-08-28 电子科技大学 CNN-LSTM-based spectrum sensing method
CN111865863A (en) * 2020-07-20 2020-10-30 山东大学 RNN neural network-based OFDM signal detection method
CN112003663A (en) * 2020-07-22 2020-11-27 哈尔滨工业大学 Cognitive radio frequency spectrum sensing method based on LSTM neural network
CN112787736A (en) * 2020-12-30 2021-05-11 杭州电子科技大学 Long-short term memory cooperative spectrum sensing method based on covariance matrix

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200153535A1 (en) * 2018-11-09 2020-05-14 Bluecom Systems and Consulting LLC Reinforcement learning based cognitive anti-jamming communications system and method
CN111600667A (en) * 2020-05-25 2020-08-28 电子科技大学 CNN-LSTM-based spectrum sensing method
CN111865863A (en) * 2020-07-20 2020-10-30 山东大学 RNN neural network-based OFDM signal detection method
CN112003663A (en) * 2020-07-22 2020-11-27 哈尔滨工业大学 Cognitive radio frequency spectrum sensing method based on LSTM neural network
CN112787736A (en) * 2020-12-30 2021-05-11 杭州电子科技大学 Long-short term memory cooperative spectrum sensing method based on covariance matrix

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
YUNXUE ZHAO等: "Spectrum Sensing for Cognitive Radio Using FLOM and CNN in Alpha Noise", 《2021 IEEE WIRELESS COMMUNICATIONS AND NETWORKING CONFERENCE (WCNC)》 *
刘祥龙: "《飞桨PaddlePaddle深度学习实战》", 31 August 2020 *
宋永健等: "非高斯噪声中基于分数低阶矩协方差MME检测的频谱感知算法", 《信号处理》 *

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114337880A (en) * 2021-11-15 2022-04-12 宁波大学 Spectrum sensing method considering random arrival and departure of main user signal
CN114337880B (en) * 2021-11-15 2023-12-05 中电中南(武汉)计量检测有限公司 Spectrum sensing method considering random arrival and departure of main user signals
CN114337883A (en) * 2021-12-28 2022-04-12 杭州电子科技大学 CNN cooperative spectrum sensing method and system based on covariance matrix Cholesky decomposition
CN114337883B (en) * 2021-12-28 2024-03-01 杭州电子科技大学 CNN collaborative spectrum sensing method and system for covariance matrix Cholesky decomposition
CN114826500A (en) * 2022-04-21 2022-07-29 中国人民解放军国防科技大学 Constant envelope modulation burst signal detection method and system based on deep learning
CN115021845A (en) * 2022-06-20 2022-09-06 西安航空学院 Spectrum sensing method, system, medium, device and terminal
CN115021845B (en) * 2022-06-20 2023-04-18 西安航空学院 Spectrum sensing method, system, medium, device and terminal

Similar Documents

Publication Publication Date Title
CN113315593A (en) Frequency spectrum sensing algorithm based on FLOM covariance matrix and LSTM neural network
CN108696331B (en) Signal reconstruction method based on generation countermeasure network
CN110110862A (en) A kind of hyperparameter optimization method based on adaptability model
CN111600667B (en) CNN-LSTM-based spectrum sensing method
CN109495214B (en) Channel coding type identification method based on one-dimensional inclusion structure
CN109886464B (en) Low-information-loss short-term wind speed prediction method based on optimized singular value decomposition generated feature set
CN112364729A (en) Modulation identification method based on characteristic parameters and BP neural network
CN111199127B (en) Radar interference decision method based on deep reinforcement learning
CN112949387B (en) Intelligent anti-interference target detection method based on transfer learning
CN110929842B (en) Accurate intelligent detection method for non-cooperative radio signal burst time region
CN112949820B (en) Cognitive anti-interference target detection method based on generation of countermeasure network
CN114337880B (en) Spectrum sensing method considering random arrival and departure of main user signals
CN112787736B (en) Long-short term memory cooperative spectrum sensing method based on covariance matrix
CN113205140B (en) Semi-supervised specific radiation source individual identification method based on generative countermeasure network
Xu et al. Spectrum sensing based on parallel CNN-LSTM network
CN112307927A (en) BP network-based identification research for MPSK signals in non-cooperative communication
Xie et al. Virtual Adversarial Training-Based Semisupervised Specific Emitter Identification.
CN106529604A (en) Adaptive image tag robust prediction method and system
CN115659254A (en) Power quality disturbance analysis method for power distribution network with bimodal feature fusion
Tang et al. Specific emitter identification for IoT devices based on deep residual shrinkage networks
CN112836788B (en) Low-power-consumption deep learning network method for interference type identification
CN108631817B (en) Method for predicting frequency hopping signal frequency band based on time-frequency analysis and radial neural network
CN113109782A (en) Novel classification method directly applied to radar radiation source amplitude sequence
CN114337883B (en) CNN collaborative spectrum sensing method and system for covariance matrix Cholesky decomposition
CN116386847A (en) DB-LSTM neural network-based intelligent prediction algorithm for mobile medical Internet of things spectrum

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
RJ01 Rejection of invention patent application after publication
RJ01 Rejection of invention patent application after publication

Application publication date: 20210827