CN111431645B - Spectrum sensing method based on small sample training neural network - Google Patents

Spectrum sensing method based on small sample training neural network Download PDF

Info

Publication number
CN111431645B
CN111431645B CN202010237889.0A CN202010237889A CN111431645B CN 111431645 B CN111431645 B CN 111431645B CN 202010237889 A CN202010237889 A CN 202010237889A CN 111431645 B CN111431645 B CN 111431645B
Authority
CN
China
Prior art keywords
neural network
detector
training
network detector
gradient
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.)
Active
Application number
CN202010237889.0A
Other languages
Chinese (zh)
Other versions
CN111431645A (en
Inventor
赵海涛
魏急波
高士顺
熊俊
张晓瀛
周力
辜方林
唐麒
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.)
National University of Defense Technology
Original Assignee
National University of Defense Technology
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 National University of Defense Technology filed Critical National University of Defense Technology
Priority to CN202010237889.0A priority Critical patent/CN111431645B/en
Publication of CN111431645A publication Critical patent/CN111431645A/en
Application granted granted Critical
Publication of CN111431645B publication Critical patent/CN111431645B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

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
    • G06N3/045—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

Abstract

The invention provides a frequency spectrum sensing method based on a small sample training neural network, which comprises the following steps: pre-training a neural network detector, deploying the detector in an actual operation environment of a system in advance, iteratively calculating a fine adjustment value and a test loss value of the detector under different perception environments through a small amount of data and a small amount of gradient, then calculating the gradient of an initial parameter corresponding to the loss value, and performing gradient updating on the initial parameter, so that the initial parameter of the detector can be rapidly adapted to the change of the environment; performing online adjustment based on initial parameters of a pre-trained neural network detector; the sampled received signal is input into an adjusted neural network detector to predict the probability of whether the spectrum is occupied at that time. The spectrum sensing method can achieve detection performance similar to that of the existing neural network detector based on a large number of samples and gradient iteration, and can effectively reduce the calculated amount and the number of samples required by the adjustment of the detector.

Description

Spectrum sensing method based on small sample training neural network
Technical Field
The invention belongs to the technical field of wireless communication networks, and relates to a frequency spectrum sensing method based on a small sample training neural network.
Background
Spectrum sensing, i.e., receiving a signal to determine whether a current spectrum is occupied, has received increasing attention. Due to the wide application range of the spectrum sensing technology, the spectrum sensing technology is widely applied in many communication scenes. Such as cognitive radio technology and interference-free communication technology.
Research on spectrum sensing techniques generally assumes that certain a priori information, such as the communication characteristics or channel characteristics at the transmitting end, are known. However, the prior information is often difficult to obtain in practical application, so that the practicability of the method is poor. With the rapid development of machine learning techniques, existing research gradually focuses on combining machine learning techniques with blind spectrum sensing techniques. Under the condition that prior information is not needed, a machine learning method can obtain a more accurate detection result, but the method consumes a large amount of computing resources of a sensing end and needs a large amount of training samples.
Therefore, there is a need for a new technique for a spectrum sensing method based on a small sample training neural network.
Disclosure of Invention
The invention aims to provide a spectrum sensing method based on a small sample training neural network, and aims to solve the technical problems that a large number of samples and a large number of calculations are needed in the spectrum sensing method based on the neural network in the prior art.
In order to achieve the above object, the present invention provides a spectrum sensing method based on a small sample training neural network, comprising the following steps:
step S1, pre-training stage: pre-training a neural network detector, deploying the neural network detector in a system actual operation environment in advance, randomly initializing initial parameters of the neural network detector, and circularly processing according to the following process until pre-training is performed for specified times, wherein the specific process is as follows:
s11, extracting K samples in the current sensing environment, calculating a fine adjustment value of the neural network detector through J gradient iterations, and storing the fine adjustment value; wherein K and J are integers greater than 0;
s12, extracting K new samples under the current sensing environment, calculating the fine adjustment value and the loss value of the new samples, and calculating the gradient of the initial parameter corresponding to the loss value of the new samples;
s13, updating the initial parameters according to the gradient of the initial parameters corresponding to the loss value of the new sample;
s14, checking the pre-training times, and executing a step S2 if the pre-training times reach the preset times; if the pre-training times are not reached, checking the detection performance of the fine tuning value of the neural network detector at the moment, and if the detection performance deterioration of the fine tuning value of the neural network detector exceeds a threshold value, returning to the step S11 for the next pre-training;
step S2, on-line adjustment stage: extracting at least K samples under the current perception environment based on the initial parameters of the neural network detector pre-trained in the step S1, and adjusting the initial parameters of the neural network detector through at least J gradient iterations;
step S3, on-line detection stage: the current received signal is input to the neural network detector adjusted in step S2, and the neural network detector outputs whether the current spectrum is occupied.
Further, in the step S3, it is checked whether the detection performance degradation of the neural network detector exceeds a threshold value during the operation process, and if the detection performance degradation exceeds the threshold value, the operation returns to the step S2 to readjust the neural network detector.
Further, the detection performance deterioration means that the difference of the missed detection probabilities at different time points exceeds a threshold value.
Further, K in step S1 is an integer of 10 or more and 100 or less.
Further, J in step S1 is an integer of 1 or more and 10 or less.
Further, the initial parameters of the neural network detector in step S1 include an initial weight and an initial bias of each neuron in the neural network.
Further, the samples in step S1 and step S2 both refer to the received signal and its corresponding set of spectrum states.
Further, the loss value in step S12 refers to a value calculated by a loss function in machine learning.
Further, the loss function is cross entropy.
The invention has the following beneficial effects:
1. the invention provides a frequency spectrum sensing method based on a small sample training neural network, which is characterized in that a set of initial parameters of the neural network detector are obtained by deploying the neural network detector in advance in an actual environment and performing pre-training. On one hand, the initial parameters of the detector can effectively learn similar parts of the parameters of the neural network detector under different perception environments, so that the debugging process of the neural network detector is simplified. On the other hand, the initial parameters are very sensitive, and a small change can cause a large positive change in the detector performance. Whether the current sensing environment is changed or not is checked by checking whether the detection performance deterioration of the fine adjustment value of the neural network detector exceeds a threshold value or not, and the initial parameters are continuously optimized and adjusted. However, the pre-training aims to make only the detector with the fine-tuned initial parameter of the neural network detector have good detection performance, but the detection performance cannot be guaranteed, so that after the pre-training, the initial parameter of the detector needs to be fine-tuned on line according to the current sensing environment. Through combining the pre-training and the online fine-tuning, whether the current sensing environment changes or not is detected in real time in the pre-training stage and the online detection stage, and the current sensing environment is continuously optimized and adjusted, so that the interference of environment change can be effectively reduced, the leak detection probability is reduced, and the detection performance is improved.
2. According to the spectrum sensing method based on the small sample training neural network, the neural network detector is deployed in the actual environment in advance for pre-training, so that the neural network detector can obtain high detection performance only by a small number of samples and a small number of gradient iterative adjustments when in use, the detection performance requirements are met, and the calculation cost and the data volume requirements of a detection end are reduced. Moreover, under the condition that the neural network detector adopts the same neural network structure, the method can achieve the performance similar to the neural network detector based on a large number of samples and a large number of gradient iterations under the condition of being based on a small number of samples and a small number of gradient iterations, and can effectively reduce the required training calculation amount and the required data amount. The amount of training computation required by the present invention can be reduced by a factor 1382 at the maximum relative to neural network detectors based on a large number of samples and a large number of gradient iterations.
In addition to the objects, features and advantages described above, other objects, features and advantages of the present invention are also provided. The present invention will be described in further detail below with reference to the drawings.
Drawings
The accompanying drawings, which are incorporated in and constitute a part of this application, illustrate embodiments of the invention and, together with the description, serve to explain the invention and not to limit the invention. In the drawings:
FIG. 1 is a schematic flow chart of a spectrum sensing method based on a small sample training neural network according to the present invention;
FIG. 2 is a diagram of a neural network architecture employed in simulation verification;
FIG. 3 is a graph comparing the performance of different spectrum sensing methods at a false alarm probability of 0.05, without noise uncertainty and with delay;
FIG. 4 is a graph of performance versus signal-to-noise ratio of-20 dB using different spectrum sensing methods with no noise uncertainty and with delay;
FIG. 5 is a graph comparing the performance of different spectrum sensing methods with a false alarm probability of 0.05 under uncertain noise;
FIG. 6 is a graph comparing the performance of different spectrum sensing methods at a false alarm probability of 0.05 at any delay;
the device comprises a first winding layer, an input layer, a second winding layer, a first pooling layer, a second pooling layer, a third pooling layer, a fourth pooling layer and a full connection layer.
Detailed Description
Embodiments of the invention will be described in detail below with reference to the drawings, but the invention can be implemented in many different ways, which are defined and covered by the claims.
As shown in fig. 1, the invention provides a spectrum sensing method based on a small sample training neural network, comprising the following steps: firstly, pre-training a neural network detector, and deploying the detector in a system actual operation environment in advance; then, initializing initial parameters of a network in the neural network detector at random; then under the current perception environment, the detector extracts K samples and performs J times of gradient iteration to calculate the fine adjustment value of the neural network detector; then, extracting the same number of samples under the environment and calculating the loss value of the detector relative to the samples and the gradient of the initial parameter corresponding to the loss value; then, updating the initial parameters, checking whether the pre-training times meet a stopping condition, if not, when the performance deterioration of the fine tuning value of the neural network detector exceeds a threshold value, namely when the difference value of the missed detection probabilities at different moments exceeds the threshold value, indicating that the environment has a certain degree of change, and carrying out the next pre-training in the changed new environment by the neural network detector; and if the stopping condition is met, the neural network detector enters an online adjustment stage. Based on the initial parameters of the neural network detector obtained by pre-training, the sensing end extracts at least K samples under the current sensing environment and performs at least J times of gradient iteration to fine tune the initial parameters of the neural network detector. The neural network detector then enters an online detection phase. At this time, the received signal may be input into the fine-tuned neural network detector, the detector will output the related probability of whether the frequency spectrum is occupied at this time, and if the detection performance degradation of the detector exceeds the threshold value, that is, the difference between the missed detection probabilities at different times exceeds the threshold value, it indicates that the environment has changed to a certain extent at this time, and the detector needs to be adjusted again in a new environment.
Wherein, K and J are integers more than 0 in the pre-training, preferably, K is more than or equal to 10 and less than or equal to 100, and J is more than or equal to 1 and less than or equal to 10.
The initial parameters of the neural network detector include an initial weight and an initial bias for each neuron in the neural network. Samples each refer to a set of received signals and their corresponding spectral states. The loss value refers to a value calculated by a loss function in machine learning, which is generally cross entropy in frequency perception.
Example 1:
for a further understanding of the present disclosure, reference will now be made to the following examples. The present invention is based on the following common and practical assumptions: the fading a, the noise epsilon and the time delay tau in the sensing environment h where the detector is positioned are distributed according to a certain rule in a certain range, namely h-rho (a, epsilon, tau); the following parameters are defined: the number of sensing end antennas M and the number of sensing time signal samples N, so that the sampled received signal can be expressed as XM×N(ii) a Considering the relevant spectral state of the received signal, the training samples of length K in the perceptual environment h may be represented as
Figure BDA0002431619090000041
y(k)∈{0,1}，y (k)1 indicates that the spectrum is occupied, whereas y(k)0; theta denotes a parameter of the neural network in the neural network detector, theta0Which is indicative of an initial parameter of the detector,
Figure BDA0002431619090000042
denotes the fine-tuning value, f, of the detector after J gradient iterations in environment hθRepresenting a neural network detector.
The method comprises the following specific steps:
the first step, pre-training stage: neural network detector fθThe method comprises the steps of deploying in an actual operation environment of a system in advance, and then initializing initial parameters theta of a network in a neural network detector randomly0Then based on pre-training targets
Figure BDA0002431619090000051
I.e. find a set of initial parameters theta that perform best after K samples and J iterations of gradient under any environment in the distribution p (a, epsilon, tau)0Go to loop iterationThe method specifically comprises the following substeps:
substep 1: extracting K samples under the current sensing environment h-rho (a, epsilon, tau)
Figure BDA0002431619090000052
The loss values L of these samples are then calculatedh(fθ) For example, cross entropy, which is expressed as:
Figure BDA0002431619090000053
then, the gradient of the current parameter corresponding to the loss value is calculated
Figure BDA0002431619090000054
Then according to
Figure BDA0002431619090000055
Executing J times in a loop to obtain a fine tuning value
Figure BDA0002431619090000056
Where α represents the learning rate.
Substep 2: under the current sensing environment h-rho (a, epsilon, tau), K new samples are extracted
Figure BDA0002431619090000057
Then calculating the test error according to the loss function expression
Figure BDA0002431619090000058
Then according to
Figure BDA0002431619090000059
Calculating the test error
Figure BDA00024316190900000510
Corresponding initial parameter theta0Of the gradient of (c).
Substep 3: based on the gradient of the initial parameter corresponding to the loss value of the neural network fine tuning value test under the environment calculated by the substep 2, according to
Figure BDA00024316190900000511
To update the initial parameters, where β is the learning rate. Then, checking whether the pre-training meets the stop condition or not, namely whether the pre-training times meet the stop condition or not, and if so, jumping to the step 2; if not, checking whether the fine tuning value detection performance deterioration of the neural network detector at the moment exceeds a threshold value gamma, namely whether the difference value of the missed detection probabilities at different moments exceeds the threshold value, if so, indicating that the environment is changed at the moment, and skipping to the substep 1 by needing to perform next pre-training.
Step two, an online fine tuning stage: extracting K samples under the current sensing environment h-rho (a, epsilon, tau)
Figure BDA00024316190900000512
And then based on the initial parameter theta of the pre-training0Calculating the loss values L of the samplesh(fθ). Then, the gradient of the current parameter corresponding to the loss value is calculated
Figure BDA00024316190900000513
Then according to
Figure BDA00024316190900000514
And performing loop execution for J times, and performing fine adjustment on the neural network detector under the current environment.
Step three, an online detection stage: the currently sampled signal X isM×NNeural network detector after input fine adjustment
Figure BDA0002431619090000061
To predict the probability that the current spectrum is occupied; then, checking whether the performance deterioration of the trimmed neural network detector exceeds a threshold value gamma, namely whether the difference value of the missed detection probabilities at different moments exceeds the threshold value gamma, if so, indicating that the environment is changed, and jumping to the second step for readjustment.
Comparative example 1: (without on-line trimming stage)
The difference from example 1 is that: after the pre-training stage, the online detection stage is directly entered without the online fine-tuning stage, and the currently sampled signal is input to the pre-trained neural network detector to predict the probability of whether the current frequency spectrum is occupied. The other steps and parameters were the same as in example 1.
Compared with a deep convolutional neural network architecture spectrum sensing method (CM-CNN for short), an energy detection method (ED for short), a maximum eigenvalue detection method (MED for short) based on covariance sensing, which is not subjected to pre-training, adopts a large amount of training data and gradient iteration and is based on prior art, and a spectrum sensing method (MAML-SS-Init for short) of proportion 1, simulation tests are carried out.
The neural network detector used in the simulation verification is a deep convolutional neural network architecture based on covariance perception, and as shown in fig. 2, the neural network architecture is composed of an input layer 1 (with a scale of 28 × 28 × 2), a convolutional layer one 2 (with a scale of 5 × 5 × 32), a pooling layer one 3 (with a scale of 2 × 2), a convolutional layer two 4 (with a scale of 5 × 5 × 64), a pooling layer two 5 (with a scale of 2 × 2), and a full-link layer 6 (with a scale of 3136 × 512) in this order. The environmental parameters of the simulation system are shown in table 1, wherein the transmitting end adopts orthogonal frequency division multiplexing (OFDM for short), and the signal-to-noise ratio and the delay in the channel are uniformly changed within a given range. In addition, the hyper-parameters of the neural network detector in the method of embodiment 1 of the present invention are shown in table 2, and the hyper-parameters of the neural network detector in the CM-CNN method are shown in table 3.
TABLE 1 simulation Environment parameters of System at verification time
Block Size Nd CP Length Nc Channel with a plurality of channels Modulation system
64 8 Rayleigh BPSK
Center frequency Bandwidth of Range of variation of signal to noise ratio Range of delay
2.4GHz 5MHz [-20dB,0dB] [0,Nd+Nc-1]
Table 2 neural network hyperparameters used in example 1
Figure BDA0002431619090000062
Figure BDA0002431619090000071
TABLE 3 neural network hyper-parameters adopted by CM-CNN method
TrainingCollective scale EarlyStopping Learning rate Batch-Size
10000 50 0.001 128
Test set Scale Epochs
2000 500
Fig. 3 and 4 are graphs comparing the performance of the method of example 1 (MAML-SS) of the present invention with the performance of the CM-CNN method, the energy detection method (ED), the maximum eigenvalue detection Method (MED), and the spectrum sensing method of comparative example 1 (MAML-SS-Init) without noise uncertainty and delay under various signal-to-noise ratios, which are not pre-trained and employ a large amount of training data and gradient iteration in the prior art. Fig. 3 is a performance comparison graph of five different spectrum sensing methods when the false alarm probability is 0.05, and fig. 4 is a performance comparison graph of five different spectrum sensing methods when the signal-to-noise ratio is-20 dB. As can be seen from fig. 3, the CM-CNN method, which has undergone a large amount of data and multiple gradient iterations, has the best performance among the five different spectrum sensing methods. And for the spectrum sensing method (MAML-SS-Init) of the comparative example 1, namely the pre-training stage is carried out like the method of the invention, but the method directly enters the on-line detection stage without the on-line fine-tuning stage, the performance is poor, and the missed detection probability is always about 97% under each signal-to-noise ratio. However, compared with the comparative example 1, after the neural network detector is subjected to fine adjustment through 5 gradient iterations under the condition of 10 samples, the neural network detector in the embodiment 1 can be quickly adapted to the current sensing environment, the performance of the neural network detector is greatly improved, the performance of the neural network detector is very close to that of CM-CNN, the missed detection probability of the neural network detector is only 0% -3% different from that of CM-CNN under each signal-to-noise ratio, and the method is far better than an energy detection method (ED) and a maximum eigenvalue detection Method (MED). Because the pre-training in the method aims to ensure that the detector with the initial parameters of the neural network detector subjected to fine tuning has better detection performance, the initial parameters only become sensitive and learn similar parts of the parameters of the neural network detector under different perception environments, but the detection performance cannot be guaranteed, the initial parameters of the detector also need to be subjected to online fine tuning after the pre-training. Through the combination of pre-training and online fine adjustment, the interference of environmental change can be effectively reduced, the leak detection probability is reduced, and the detection performance is improved. In addition, as can be seen from fig. 4, the detection performance of the method (MAML-SS) of the present invention is very close to the detection performance of the CM-CNN method based on a large amount of data and a large number of gradient iterations at each false alarm probability.
Fig. 5 is a performance comparison test result of different spectrum sensing methods under various signal-to-noise ratios under the condition of noise uncertainty, wherein η in fig. 5 represents noise uncertainty. As can be seen from FIG. 5, when the noise uncertainty is increased from 0.5dB to 1dB at a signal-to-noise ratio of-14 dB, the missed detection probability by the method of the present invention is increased from 2.03% to 2.72%, and the missed detection probability by the CM-CNN method is increased from 0.48% to 0.71%. It follows that the method of the present invention is more affected by noise uncertainty than CM-CNN. However, compared with the energy detection method (ED) and the maximum eigenvalue detection Method (MED), the invention is less affected, and the false detection probability of the two methods is increased from 73.38% and 63.28% to 79.69% and 68.53%, respectively. Fig. 6 shows the performance comparison test results of different spectrum sensing methods under various signal-to-noise ratios under any delay condition. Fig. 6 also shows a similar conclusion as fig. 5, so that it can be demonstrated that the method of the present invention has a certain robustness to noise uncertainty and delay.
During verification, the same neural network framework is adopted in the CM-CNN method which is not pre-trained and adopts a large amount of training data and gradient iteration in the prior art, so that when the training calculated amount is analyzed, the calculated amount C required by one-time previous item propagation (FP) of the adopted neural network framework is usedFPAs a unit of computation, the details are shown in Table 4, where FLOPs denote floating point number operations. Considering that the computation amount required for Back Propagation (BP) is twice as much as that of the previous propagation, the training computation amount of the neural network detector can be expressed as
Figure BDA0002431619090000081
Therefore, by performing 1000 Monte Carlo simulations without noise uncertainty and delay, the amount of computation required by the present invention and the CM-CNN method based on a large amount of data and a large number of gradient iterations is shown in Table 5, and the amount of training computation required by the present invention is reduced 1382 times at the maximum. It can be concluded from fig. 3 and table 5 that the present invention can effectively reduce the required training calculation amount and the required data amount when the same neural network structure is adopted for the neural network detector and similar detection performance is achieved.
TABLE 4 details of computational losses of neural network architecture forward propagation employed for simulation experiments
Convolutional layer Pooling layer Full connectionLayer(s) Total calculated amount
11296064FlOPs 37632FlOPs 3213826FlOPs 14547522FlOPs
TABLE 5 comparison of the calculated amounts of the inventive method and the CM-CNN method
Signal to noise ratio -12 -14 -16 -18 -20
CM-CNN/CFP 54080 82940 57200 47840 40560
MAML-SS/CFP 60 60 60 60 60
In summary, according to the spectrum sensing method based on the small sample training neural network provided by the invention, the neural network detector is deployed in the actual environment in advance for pre-training, so that the neural network detector can obtain higher detection performance only by a small number of samples and a small number of gradient iterative adjustments when in use, the detection performance requirements are met, and the calculation cost and the data volume requirements of the detection end are reduced. Moreover, under the condition that the neural network detector adopts the same neural network structure, the method can achieve the performance similar to the neural network detector based on a large number of samples and a large number of gradient iterations under the condition of being based on a small number of samples and a small number of gradient iterations, and can effectively reduce the required training calculation amount and the required data amount.
The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention, and various modifications and changes may be made by those skilled in the art. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims (5)

1. A spectrum sensing method based on a small sample training neural network is characterized by comprising the following steps:
step S1, pre-training stage: pre-training a neural network detector, deploying the neural network detector in a system actual operation environment in advance, randomly initializing initial parameters of the neural network detector, and circularly processing according to the following process until pre-training is performed for specified times, wherein the specific process is as follows:
s11, extracting K samples in the current sensing environment, calculating a fine adjustment value of the neural network detector through J gradient iterations, and storing the fine adjustment value; wherein K and J are integers greater than 0;
s12, extracting K new samples under the current sensing environment, calculating the fine adjustment value and the loss value of the new samples, and calculating the gradient of the initial parameter corresponding to the loss value of the new samples;
s13, updating the initial parameters according to the gradient of the initial parameters corresponding to the loss value of the new sample;
s14, checking the pre-training times, and executing a step S2 if the pre-training times reach the preset times; if the pre-training times are not reached, checking the detection performance of the fine tuning value of the neural network detector at the moment, and if the detection performance deterioration of the fine tuning value of the neural network detector exceeds a threshold value, returning to the step S11 for the next pre-training;
step S2, on-line adjustment stage: extracting at least K samples under the current perception environment based on the initial parameters of the neural network detector pre-trained in the step S1, and adjusting the initial parameters of the neural network detector through at least J gradient iterations;
step S3, on-line detection stage: inputting the current received signal to the neural network detector adjusted in step S2, and outputting whether the current spectrum is occupied by the neural network detector;
k in the step S1 is an integer which is more than or equal to 10 and less than or equal to 100;
j in the step S1 is an integer of more than or equal to 1 and less than or equal to 10;
the initial parameters of the neural network detector in the step S1 include an initial weight and an initial bias of each neuron in the neural network;
the samples in step S1 and step S2 both refer to the received signal and its corresponding set of spectral states.
2. The method for sensing frequency spectrum based on training neural network of claim 1, wherein the step S3 is executed to check whether the detection performance degradation of the neural network detector exceeds a threshold value, and if the detection performance degradation exceeds the threshold value, the step S2 is executed to readjust the neural network detector.
3. The method as claimed in claim 2, wherein the performance degradation is that the difference between the missed detection probabilities at different time instances exceeds a threshold.
4. The method for sensing frequency spectrum based on training neural network of claim 1, wherein the loss value in step S12 refers to a value calculated by a loss function in machine learning.
5. The method for spectrum sensing based on small sample trained neural network of claim 4, wherein the loss function is cross entropy.
CN202010237889.0A 2020-03-30 2020-03-30 Spectrum sensing method based on small sample training neural network Active CN111431645B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010237889.0A CN111431645B (en) 2020-03-30 2020-03-30 Spectrum sensing method based on small sample training neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010237889.0A CN111431645B (en) 2020-03-30 2020-03-30 Spectrum sensing method based on small sample training neural network

Publications (2)

Publication Number Publication Date
CN111431645A CN111431645A (en) 2020-07-17
CN111431645B true CN111431645B (en) 2022-02-08

Family

ID=71549297

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010237889.0A Active CN111431645B (en) 2020-03-30 2020-03-30 Spectrum sensing method based on small sample training neural network

Country Status (1)

Country Link
CN (1) CN111431645B (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111814978A (en) * 2020-09-11 2020-10-23 鹏城实验室 Method, apparatus and medium for calculating training computation of neural network model

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108777872A (en) * 2018-05-22 2018-11-09 中国人民解放军陆军工程大学 A kind of anti-interference model of depth Q neural networks and intelligent Anti-interference algorithm
CN109245840A (en) * 2018-10-15 2019-01-18 哈尔滨工业大学 Spectrum prediction method in cognitive radio system based on convolutional neural networks

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8521092B2 (en) * 2009-05-27 2013-08-27 Echo Ridge Llc Wireless transceiver test bed system and method
US20180075347A1 (en) * 2016-09-15 2018-03-15 Microsoft Technology Licensing, Llc Efficient training of neural networks
CN110830124A (en) * 2019-11-21 2020-02-21 长春理工大学 Spectrum sensing method based on quantum particle swarm optimization extreme learning machine

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108777872A (en) * 2018-05-22 2018-11-09 中国人民解放军陆军工程大学 A kind of anti-interference model of depth Q neural networks and intelligent Anti-interference algorithm
CN109245840A (en) * 2018-10-15 2019-01-18 哈尔滨工业大学 Spectrum prediction method in cognitive radio system based on convolutional neural networks

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
Sparse hybrid precoding and combining in millimeter wave MIMO systems;Aryan Kaushik等;《Radio Propagation and Technologies for 5G (2016)》;20170116;全文 *
一种基于深度学习的辐射源信号调制识别新算法;陶冠宏，周林;《科学技术与工程》;20200131;全文 *

Also Published As

Publication number Publication date
CN111431645A (en) 2020-07-17

Similar Documents

Publication Publication Date Title
CN111600667B (en) CNN-LSTM-based spectrum sensing method
CN110213003B (en) Wireless channel large-scale fading modeling method and device
CN104703216A (en) Energy-efficient-based detecting method of multi-task Bayes compressed sensing broadband spectrum
EP2743924A1 (en) Method and apparatus for adaptively detecting a voice activity in an input audio signal
Manesh et al. An optimized SNR estimation technique using particle swarm optimization algorithm
CN111800795B (en) Spectrum sensing method under non-Gaussian noise in cognitive unmanned aerial vehicle network
CN101507151A (en) Method, device and computer program for classifying a received signal
CN112364729A (en) Modulation identification method based on characteristic parameters and BP neural network
CN111431645B (en) Spectrum sensing method based on small sample training neural network
CN113315593A (en) Frequency spectrum sensing algorithm based on FLOM covariance matrix and LSTM neural network
Kim et al. Massive data generation for deep learning-aided wireless systems using meta learning and generative adversarial network
CN113033457B (en) Soft fault recognition model training and soft fault recognition method and device
CN105656577B (en) Towards the cluster-dividing method and device of channel impulse response
CN112101487B (en) Compression method and device for fine-grained recognition model
CN109600181B (en) Spectrum sensing method for multiple antennas
Youssef et al. Efficient cooperative spectrum detection in cognitive radio systems using wavelet fusion
Salehi et al. Kernelized-likelihood ratio tests for binary phase-shift keying signal detection
CN108599880B (en) Civil aviation ground-air intercom system interference early warning method based on convolutional neural network
Zhang et al. Machine learning based protocol classification in unlicensed 5 GHz bands
Lee et al. Device Free Localization Using Prior Correction-Based Logistic Regression
Cho et al. Low computational complexity spectrum sensing based on cyclostationarity for multiple receive antennas
CN114785433A (en) Channel scene recognition method, network device and storage medium
Arts et al. Quickest eigenvalue-based spectrum sensing using random matrix theory
Maras Adaptive nonparametric locally optimum Bayes detection in additive non-Gaussian noise
US9692465B1 (en) Aggregate interference model and use thereof to evaluate performance of a receiver

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant