CN112968740B - Satellite spectrum sensing method based on machine learning - Google Patents

Satellite spectrum sensing method based on machine learning Download PDF

Info

Publication number
CN112968740B
CN112968740B CN202110136183.XA CN202110136183A CN112968740B CN 112968740 B CN112968740 B CN 112968740B CN 202110136183 A CN202110136183 A CN 202110136183A CN 112968740 B CN112968740 B CN 112968740B
Authority
CN
China
Prior art keywords
neural network
training
time slot
cnn
lstm
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
CN202110136183.XA
Other languages
Chinese (zh)
Other versions
CN112968740A (en
Inventor
丁晓进
倪韬
朱剑
张更新
吴尘
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 University of Posts and Telecommunications
Original Assignee
Nanjing University of Posts and Telecommunications
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 University of Posts and Telecommunications filed Critical Nanjing University of Posts and Telecommunications
Priority to CN202110136183.XA priority Critical patent/CN112968740B/en
Publication of CN112968740A publication Critical patent/CN112968740A/en
Application granted granted Critical
Publication of CN112968740B publication Critical patent/CN112968740B/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
    • G06N3/044—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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic 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
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B7/00—Radio transmission systems, i.e. using radiation field
    • H04B7/14—Relay systems
    • H04B7/15—Active relay systems
    • H04B7/185—Space-based or airborne stations; Stations for satellite systems
    • H04B7/1851—Systems using a satellite or space-based relay

Abstract

The invention discloses a satellite spectrum sensing method based on machine learning, which comprises three parts of data preprocessing, model offline training and online detection. According to the method, historical received data of different channels are processed, a corresponding training set training neural network model is constructed, the model is optimized by combining grid search, cross validation and early-stop correction, and the hiding rule of the data can be effectively extracted. In addition, the strong classification capability of the CNN neural network is combined with the advantage of the LSTM neural network processing sequence problem, a proper fusion mode is selected, a reasonable combined detection value calculation method is designed, and the false alarm probability is taken into consideration to obtain a proper detection threshold. And finally, inputting the real-time detection sample into the trained neural network model to realize the real-time judgment of the spectrum occupation state, thereby improving the spectrum sensing accuracy and further improving the spectrum resource utilization rate.

Description

Satellite spectrum sensing method based on machine learning
Technical Field
The invention relates to a satellite spectrum sensing method based on machine learning, and belongs to the cognitive radio spectrum sensing technology.
Background
With the development of wireless communication technology, the number of wireless frequency devices is rapidly increasing, and the spectrum demand is increasingly intensified. However, currently, spectrum resources are statically allocated, and this allocation strategy may cause a plurality of authorized spectrum to have insufficient utilization rate, and a spectrum hole exists, which results in a great waste of resources. The cognitive radio is a technology that an unauthorized user (a slave user) opportunistically accesses and uses idle spectrum resources of an authorized user under the permission of the authorized user (a master user) of a spectrum, and can improve the spectrum utilization rate to a certain extent and relieve the current situation of the shortage of the spectrum resources. Spectrum sensing plays an extremely important role in cognitive radio, and therefore, it is necessary to design a high-accuracy spectrum sensing method.
The most widely used method for sensing the traditional frequency spectrum is an energy detection method. Although the energy detection method is simple and low in implementation difficulty, the energy detection method as a semi-blind detection method needs to know prior information of noise and is lack of flexible adaptability to a measurement environment. Meanwhile, the energy detection method has poor robustness to noise under the condition of low signal-to-noise ratio, and misjudgment is easily caused when the noise fluctuation is large.
In recent years, with the improvement of computer computing capability and data quantization, a deep learning technology is rapidly developed, target features are intelligently extracted through various neural network models, a hiding rule is learned, the method plays a role in the fields of computer vision, natural language processing and the like, and at present, a spectrum sensing technology based on deep learning begins to completely reveal the head. Convolutional Neural Networks (CNNs) are widely used in classification problems, and Long-short-term memory (LSTM) is an improved Recurrent Neural Network (RNN), which is a Neural Network specially used to deal with sequence problems. Discrete received signals can be regarded as a kind of time series to some extent, so that it is feasible and reasonable to apply the CNN neural network and the LSTM neural network to perform predictive classification on the received signals representing the spectrum occupation state.
Disclosure of Invention
The purpose of the invention is as follows: in order to overcome the defects in the prior art, the invention provides a satellite spectrum sensing method based on machine learning, which is a joint spectrum sensing method based on CNN and LSTM neural networks. In addition, the strong classification capability of the CNN neural network is combined with the advantages of the LSTM neural network processing sequence problem, a proper fusion mode is selected, a reasonable combined detection value calculation method is designed, and the false alarm probability is taken into consideration to calculate a proper detection threshold; the samples corresponding to the real-time received data are input into the optimized neural network model to realize the real-time judgment of the spectrum occupation state, so that the utilization rate of spectrum resources can be improved.
The technical scheme is as follows: in order to achieve the purpose, the invention adopts the technical scheme that:
a satellite spectrum sensing method based on machine learning comprises three parts of data preprocessing, model offline training and online detection, and specifically comprises the following steps:
(1) data preprocessing: performing energy normalization on recent historical received data of a single user signal I channel by taking a time slot as a unit; the historical received data of I channels of the single-user signal are spliced and combined to obtain a CNN neural network training set (X, Y) CNN (ii) a The historical received data of I channels of the single-user signal are combined to obtain an LSTM neural network training set (X, Y) after energy normalization LSTM ；
(2) Model off-line training: using CNN neural network training set (X, Y) CNN Training the CNN neural network using the LSTM neural network training set (X, Y) LSTM Training the LSTM neural network;
(3) online detection: the method comprises three parts of joint detection value calculation, detection threshold calculation and online sample detection:
(3.1) calculating a joint detection value: fusing the output probability vector of the trained CNN neural network and the output probability vector of the LSTM neural network to obtain a combined detection value;
(3.2) calculating a detection threshold: CNN neural network training set (X, Y) corresponding to historical time slot CNN Noise data set in (1) and LSTM neural network training set (X, Y) LSTM Calculating a combined detection value set by the noise data set, and solving a detection threshold by combining the set false alarm probability;
(3.3) online sample detection: and comparing the joint detection value of the detection sample acquired in real time with a detection threshold, and judging the frequency spectrum occupation condition.
According to the method, historical received data of different channels are processed, a corresponding training set training neural network model is constructed, the model is optimized by combining grid search, cross validation and early-stop correction, and the hiding rule of the data can be effectively extracted. In addition, the strong classification capability of the CNN neural network is combined with the advantage of the LSTM neural network processing sequence problem, a proper fusion mode is selected, a reasonable combined detection value calculation method is designed, and the false alarm probability is taken into consideration to obtain a proper detection threshold. And finally, inputting the real-time detection sample into the trained neural network model to realize the real-time judgment of the spectrum occupation state, thereby improving the spectrum sensing accuracy and further improving the spectrum resource utilization rate.
Specifically, in the step (1), the historical received data of I channels of the single-user signal is subjected to energy normalization, and the received data of the ith channel of the kth time slot
Figure BDA0002926739280000031
Expressed as:
Figure BDA0002926739280000032
wherein: complex vector
Figure BDA0002926739280000033
A discrete complex signal of length N is received for the ith channel of the kth slot,
Figure BDA0002926739280000034
is a complex vector
Figure BDA0002926739280000035
K is the total number of time slots, I is the total number of channels contained in the single-user signal, and N is the signal length;
reception data of ith channel of kth time slot
Figure BDA0002926739280000036
Energy of
Figure BDA0002926739280000037
Expressed as:
Figure BDA0002926739280000038
wherein: energy (·) represents the energy addition to · s;
after the historical received data of the single user signal I channel in the kth time slot is subjected to energy normalization, the CNN neural network training set (X, Y) is obtained by splicing and combining CNN Training sample of the k-th time slot
Figure BDA0002926739280000039
The corresponding label is marked as
Figure BDA00029267392800000310
Figure BDA00029267392800000311
Figure BDA00029267392800000312
Figure BDA00029267392800000313
Wherein: training sample
Figure BDA00029267392800000314
Is a matrix of size nxi x 2,
Figure BDA00029267392800000315
is composed of
Figure BDA00029267392800000316
Of the first matrix of channels of (a),
Figure BDA00029267392800000317
is composed of
Figure BDA00029267392800000318
Of the second channel matrix of (a) is,
Figure BDA00029267392800000319
indicating the state of spectral occupancy, H, at the k-th time slot 0 Indicating that the spectrum is unoccupied, H 1 Indicating that the spectrum is occupied;
after the historical received data of the single user signal I channels in the kth time slot are subjected to energy normalization, an LSTM neural network training set (X, Y) is obtained by combination LSTM Training sample of the k-th time slot
Figure BDA0002926739280000041
The corresponding label is marked as
Figure BDA0002926739280000042
The spectrum occupation state of the k time slot is obtained by predicting the energy value sequence of the k-1, …, k-s time slot by an LSTM neural network:
Figure BDA0002926739280000043
Figure BDA0002926739280000044
wherein: training sample
Figure BDA0002926739280000045
Is composed of the energy values of the first s time slots, s is the length of a backtracking window,
Figure BDA0002926739280000046
indicating the state of spectral occupancy, H, at the k-th time slot 0 Indicating that the spectrum is unoccupied, H 1 Indicating that the spectrum is occupied.
Specifically, in the step (2), a CNN neural network training set (X, Y) is used CNN And LSTM neural network training set (X, Y) LSTM The CNN neural network and the LSTM neural network are trained respectively based on the followingPrinciple:
(2.1) respectively determining the optimization directions and the hyper-parameters of the two neural networks by using a grid search method;
(2.2) reasonably dividing the two training sets by using a K-fold cross verification method respectively, and reducing the fluctuation of the training precision of the two neural networks;
and (2.3) monitoring the training process of the two neural networks by using an early-stopping method, and reducing the negative influence brought by the training process.
Specifically, the step (3) specifically includes the following steps:
(3.1) the joint detection value calculation includes the steps of:
(3.1.1) training sample of CNN neural network corresponding to k time slot
Figure BDA0002926739280000047
The output probability vector of the trained CNN neural network is represented as:
Figure BDA0002926739280000048
Wherein:
Figure BDA0002926739280000049
representing the CNN neural network after the training is completed,
Figure BDA00029267392800000410
the CNN neural network which represents the completion of the training judges that the frequency spectrum occupation state of the k time slot is H 0 The probability of (a) of (b) being,
Figure BDA00029267392800000411
the CNN neural network which represents the completion of the training judges that the frequency spectrum occupation state of the k time slot is H 1 The probability of (d);
(3.1.2) training samples of the LSTM neural network corresponding to the kth time slot
Figure BDA0002926739280000051
Output profile of trained LSTM neural networkThe rate vector is represented as:
Figure BDA0002926739280000052
wherein:
Figure BDA0002926739280000053
representing the LSTM neural network after training is complete,
Figure BDA0002926739280000054
the LSTM neural network which represents the completion of training judges that the frequency spectrum occupation state of the k time slot is H 0 The probability of (a) of (b) being,
Figure BDA0002926739280000055
the LSTM neural network which represents the completion of training judges that the frequency spectrum occupation state of the k time slot is H 1 The probability of (d);
(3.1.3) pairs
Figure BDA0002926739280000056
And
Figure BDA0002926739280000057
the fusion probability vector obtained by fusing the output probability vectors is represented as:
Figure BDA0002926739280000058
Figure BDA0002926739280000059
wherein:
Figure BDA00029267392800000510
a joint neural network fusing the CNN neural network and the LSTM neural network is shown,
Figure BDA00029267392800000511
representing that the joint neural network judges the frequency spectrum occupation state of the kth time slot to be H 0 The probability of (a) of (b) being,
Figure BDA00029267392800000512
representing that the joint neural network judges the frequency spectrum occupation state of the kth time slot to be H 1 The probability of (d);
(3.1.4) the joint detection value for determining the spectrum occupation state of the k-th slot is expressed as:
Figure BDA00029267392800000513
Wherein: t (-) represents the joint detection value;
(3.2) the detection threshold calculation comprises the following steps:
(3.2.1) training set (X, Y) from CNN neural network CNN Separating L frequency spectrum occupation states as H 0 Forming a noisy data set of training samples
Figure BDA00029267392800000514
From LSTM neural network training set (X, Y) LSTM Separating L frequency spectrum occupation states as H 0 Forming a noisy data set of training samples
Figure BDA00029267392800000515
Figure BDA00029267392800000516
And
Figure BDA00029267392800000517
the corresponding time slots are the same;
(3.2.2) calculating the joint detection value of the frequency spectrum occupation state of each time slot in the noise data set by using the method in the step (3.1), and performing descending order arrangement on the joint detection values to obtain the joint detection value set of the noise data set
Figure BDA0002926739280000061
T(W (l) ) Is shown as being as large asThe small joint detection value arranged at the l-th position;
(3.2.3) setting the detection threshold
Figure BDA0002926739280000062
Wherein:
Figure BDA0002926739280000063
representing sets of joint detection values
Figure BDA0002926739280000064
The first element T (W) in (1) (l) )，
Figure BDA0002926739280000065
In order to set the false alarm probability,
Figure BDA0002926739280000066
represents rounding-down;
(3.3) when in-line sample detection, firstly calculating the joint detection value of the detection samples of a certain time slot acquired in real time
Figure BDA0002926739280000067
Re-contrast joint detection value
Figure BDA0002926739280000068
And obtaining the spectrum occupation condition State of the detection sample with the detection threshold value gamma:
Figure BDA0002926739280000069
wherein: state indicates the spectrum occupancy of the detection sample.
Has the advantages that: the satellite spectrum sensing method based on machine learning combines the CNN neural network and the LSTM neural network to sense the spectrum, combines grid search, cross validation and early-stop correction methods to optimize the model, and can effectively extract the hidden rule of data. In addition, the strong classification capability of the CNN neural network is combined with the advantages of the LSTM neural network processing sequence problem, a proper fusion mode is selected, a reasonable combined detection value calculation method is designed, the false alarm probability is taken into consideration to calculate a proper detection threshold, the frequency spectrum occupation state of a detection sample can be well judged, and the frequency spectrum resource utilization rate can be improved.
Drawings
FIG. 1 is a schematic flow chart of the present invention;
FIG. 2 is a graph comparing the detection probability of the method of the present invention with that of the energy detection method under the conditions of PFA 0.1 and N64;
FIG. 3 is a graph showing the comparison of the detection probability under the conditions of PFA 0.01 and N64 in the method of the present invention and the energy detection method.
Detailed Description
The present invention will be further described with reference to the accompanying drawings.
As shown in fig. 1, a joint spectrum sensing method based on CNN and LSTM neural networks, which constructs training sets for the CNN neural network and the LSTM neural network respectively by processing recent historical received data; training the CNN neural network and the LSTM neural network through a training set, and optimizing the model by combining a grid search method, a cross validation method and an early-stop correction method; obtaining a combined detection value for judging the spectrum occupation state by fusing the output probability vectors of the trained CNN neural network and the trained LSTM neural network; the noise data set separated from the training set is sent to a CNN neural network and an LSTM neural network, a corresponding joint detection value set is calculated, and then a detection threshold can be obtained by using the false alarm probability and the set length of the noise data set; and calculating the joint detection value of the real-time test sample, and comparing with a threshold to obtain the spectrum occupation state. The method comprises three parts of data preprocessing, model offline training and online detection, and each part is specifically explained below.
First, data preprocessing
(1.1) energy normalization of historical received data
The historical received data of I channels of the single user signal is subjected to energy normalization, and the received data of the ith channel of the kth time slot
Figure BDA0002926739280000071
Expressed as:
Figure BDA0002926739280000072
wherein: complex vector
Figure BDA0002926739280000073
A discrete complex signal of length N is received for the ith channel of the kth slot,
Figure BDA0002926739280000074
is a complex vector
Figure BDA0002926739280000075
K is the total number of time slots, I is the total number of channels contained in the single-user signal, and N is the signal length.
Reception data of ith channel of kth time slot
Figure BDA0002926739280000076
Energy of
Figure BDA0002926739280000077
Expressed as:
Figure BDA0002926739280000078
wherein: energy () represents the energy of.
(1.2) construction of CNN neural network training set (X, Y) CNN
After the historical received data of the single user signal I channel in the kth time slot is subjected to energy normalization, the CNN neural network training set (X, Y) is obtained by splicing and combining CNN Training sample of the k-th time slot
Figure BDA0002926739280000079
The corresponding label is marked as
Figure BDA00029267392800000710
Figure BDA00029267392800000711
Figure BDA0002926739280000081
Figure BDA0002926739280000082
Wherein: training sample
Figure BDA0002926739280000083
Is a matrix of size nxi x 2,
Figure BDA0002926739280000084
is composed of
Figure BDA0002926739280000085
Of the first matrix of channels of (a),
Figure BDA0002926739280000086
is composed of
Figure BDA0002926739280000087
Of the second channel matrix of (a) is,
Figure BDA0002926739280000088
indicating the state of spectral occupancy, H, at the k-th time slot 0 Indicating that the spectrum is unoccupied, H 1 Indicating that the spectrum is occupied.
(1.3) constructing LSTM neural network training set (X, Y) LSTM
After the historical received data of the single user signal I channels in the kth time slot are subjected to energy normalization, an LSTM neural network training set (X, Y) is obtained by combination LSTM Of the k-th slotTraining sample
Figure BDA0002926739280000089
The corresponding label is marked as
Figure BDA00029267392800000810
The spectrum occupation state of the k time slot is obtained by predicting the energy value sequence of the k-1, …, k-s time slot by an LSTM neural network:
Figure BDA00029267392800000811
Figure BDA00029267392800000812
wherein: training sample
Figure BDA00029267392800000813
Is composed of the energy values of the first s time slots, s is the length of a backtracking window,
Figure BDA00029267392800000814
indicating the state of spectral occupancy, H, at the k-th time slot 0 Indicating that the spectrum is unoccupied, H 1 Indicating that the spectrum is occupied.
Second, off-line training of model
Using CNN neural network training set (X, Y) CNN Training the CNN neural network using the LSTM neural network training set (X, Y) LSTM Training the LSTM neural network based on the following principle:
(2.1) respectively determining the optimization directions and the hyper-parameters of the two neural networks by using a grid search method;
(2.2) reasonably dividing the two training sets by using a K-fold cross verification method respectively, and reducing the fluctuation of the training precision of the two neural networks;
and (2.3) monitoring the training process of the two neural networks by using an early-stopping method, and reducing the negative influence brought by the training process.
Third, on-line detection
The online detection comprises three parts of joint detection value calculation, detection threshold calculation and online sample detection.
(3.1) calculating a joint detection value: the method for obtaining the combined detection value by fusing the output probability vector of the trained CNN neural network and the output probability vector of the LSTM neural network comprises the following steps:
(3.1.1) training sample of CNN neural network corresponding to k time slot
Figure BDA0002926739280000091
The output probability vector of the trained CNN neural network is represented as:
Figure BDA0002926739280000092
wherein:
Figure BDA0002926739280000093
representing the CNN neural network after the training is completed,
Figure BDA0002926739280000094
the CNN neural network which represents the completion of the training judges that the frequency spectrum occupation state of the k time slot is H 0 The probability of (a) of (b) being,
Figure BDA0002926739280000095
the CNN neural network which represents the completion of the training judges that the frequency spectrum occupation state of the k time slot is H 1 The probability of (c).
(3.1.2) training samples of the LSTM neural network corresponding to the kth time slot
Figure BDA0002926739280000096
The output probability vector of the trained LSTM neural network is represented as:
Figure BDA0002926739280000097
wherein:
Figure BDA0002926739280000098
representing the LSTM neural network after training is complete,
Figure BDA0002926739280000099
the LSTM neural network which represents the completion of training judges that the frequency spectrum occupation state of the k time slot is H 0 The probability of (a) of (b) being,
Figure BDA00029267392800000910
the LSTM neural network which represents the completion of training judges that the frequency spectrum occupation state of the k time slot is H 1 The probability of (c).
(3.1.3) pairs
Figure BDA00029267392800000911
And
Figure BDA00029267392800000912
the fusion probability vector obtained by fusing the output probability vectors is represented as:
Figure BDA00029267392800000913
Figure BDA0002926739280000101
wherein:
Figure BDA0002926739280000102
a joint neural network fusing the CNN neural network and the LSTM neural network is shown,
Figure BDA0002926739280000103
representing that the joint neural network judges the frequency spectrum occupation state of the kth time slot to be H 0 The probability of (a) of (b) being,
Figure BDA0002926739280000104
representing a joint neural networkJudging the frequency spectrum occupation state of the k time slot to be H 1 The probability of (c).
(3.1.4) the joint detection value for determining the spectrum occupation state of the k-th slot is expressed as:
Figure BDA0002926739280000105
wherein: t (-) denotes the joint detection value.
(3.2) calculating a detection threshold: CNN neural network training set (X, Y) corresponding to historical time slot CNN Noise data set in (1) and LSTM neural network training set (X, Y) LSTM Calculating a combined detection value set by the noise data set, and solving a detection threshold by combining the set false alarm probability; the method comprises the following steps:
(3.2.1) training set (X, Y) from CNN neural network CNN Separating L frequency spectrum occupation states as H 0 Forming a noisy data set of training samples
Figure BDA0002926739280000106
From LSTM neural network training set (X, Y) LSTM Separating L frequency spectrum occupation states as H 0 Forming a noisy data set of training samples
Figure BDA0002926739280000107
Figure BDA0002926739280000108
And
Figure BDA0002926739280000109
the corresponding time slots are the same.
(3.2.2) calculating the joint detection value of the frequency spectrum occupation state of each time slot in the noise data set by using the method in the step (3.1), and performing descending order arrangement on the joint detection values to obtain the joint detection value set of the noise data set
Figure BDA00029267392800001010
T(W (l) ) The representations are arranged from large to small at the lThe joint detection value of the bits.
(3.2.3) setting the detection threshold
Figure BDA00029267392800001011
Wherein:
Figure BDA00029267392800001012
representing sets of joint detection values
Figure BDA00029267392800001013
The first element T (W) in (1) (l) )，
Figure BDA00029267392800001014
In order to set the false alarm probability,
Figure BDA00029267392800001015
indicating a rounding of the · down.
(3.3) online sample detection: calculating a joint detection value of detection samples of a certain time slot acquired in real time
Figure BDA0002926739280000111
Re-contrast joint detection value
Figure BDA0002926739280000112
And obtaining the spectrum occupation condition State of the detection sample with the detection threshold value gamma:
Figure BDA0002926739280000113
wherein: state indicates the spectrum occupancy of the detection sample.
FIG. 2 is a graph showing the comparison of the detection probability under the condition of the false alarm probability PFA being 0.1 and the detection probability under the condition of N being 64 by using the method of the present invention and an energy detection method; FIG. 3 is a graph showing the comparison of the detection probability under the condition of the false alarm probability PFA being 0.01 and the detection probability under the condition of N being 64 by using the method of the present invention and an energy detection method; obviously, the detection probability is better by adopting the method of the invention.
The above description is only of the preferred embodiments of the present invention, and it should be noted that: it will be apparent to those skilled in the art that various modifications and adaptations can be made without departing from the principles of the invention and these are intended to be within the scope of the invention.

Claims (3)

1. A satellite spectrum sensing method based on machine learning is characterized in that: the method comprises three parts of data preprocessing, model offline training and online detection, and specifically comprises the following steps:
(1) data preprocessing: carrying out energy normalization on historical received data of I channels of a single user signal by taking a time slot as a unit; the historical received data of I channels of the single-user signal are spliced and combined to obtain a CNN neural network training set (X, Y) CNN (ii) a The historical received data of I channels of the single-user signal are combined to obtain an LSTM neural network training set (X, Y) after energy normalization LSTM ；
(2) Model off-line training: using CNN neural network training set (X, Y) CNN Training the CNN neural network using the LSTM neural network training set (X, Y) LSTM Training the LSTM neural network: based on the following principles:
(2.1) respectively determining the optimization directions and the hyper-parameters of the two neural networks by using a grid search method;
(2.2) reasonably dividing the two training sets by using a K-fold cross verification method;
(2.3) monitoring the training process of the two neural networks by using an early-stopping method;
(3) online detection: the method comprises three parts of joint detection value calculation, detection threshold calculation and online sample detection:
(3.1) calculating a joint detection value: fusing the output probability vector of the trained CNN neural network and the output probability vector of the LSTM neural network to obtain a combined detection value; the method comprises the following steps:
(3.1.1) training sample of CNN neural network corresponding to k time slot
Figure FDA0003674085400000011
The output probability vector of the trained CNN neural network is represented as:
Figure FDA0003674085400000012
wherein:
Figure FDA0003674085400000013
representing the CNN neural network after the training is completed,
Figure FDA0003674085400000014
the CNN neural network which represents the completion of the training judges that the frequency spectrum occupation state of the k time slot is H 0 The probability of (a) of (b) being,
Figure FDA0003674085400000015
the CNN neural network which represents the completion of the training judges that the frequency spectrum occupation state of the k time slot is H 1 The probability of (d);
(3.1.2) training samples of the LSTM neural network corresponding to the kth time slot
Figure FDA0003674085400000016
The output probability vector of the trained LSTM neural network is represented as:
Figure FDA0003674085400000017
wherein:
Figure FDA0003674085400000018
representing the LSTM neural network after training is complete,
Figure FDA0003674085400000019
the LSTM neural network which represents the completion of training judges that the frequency spectrum occupation state of the k time slot is H 0 The probability of (a) of (b) being,
Figure FDA0003674085400000021
the LSTM neural network which represents the completion of training judges that the frequency spectrum occupation state of the k time slot is H 1 The probability of (d);
(3.1.3) pairs
Figure FDA0003674085400000022
And
Figure FDA0003674085400000023
the fusion probability vector obtained by fusing the output probability vectors is represented as:
Figure FDA0003674085400000024
Figure FDA0003674085400000025
wherein:
Figure FDA0003674085400000026
a joint neural network fusing the CNN neural network and the LSTM neural network is shown,
Figure FDA0003674085400000027
representing that the joint neural network judges the frequency spectrum occupation state of the kth time slot to be H 0 The probability of (a) of (b) being,
Figure FDA0003674085400000028
representing that the joint neural network judges the frequency spectrum occupation state of the kth time slot to be H 1 The probability of (d);
(3.1.4) the joint detection value for determining the spectrum occupation state of the k-th slot is expressed as:
Figure FDA0003674085400000029
wherein: t (-) represents the joint detection value;
(3.2) calculating a detection threshold: CNN neural network training set (X, Y) corresponding to historical time slot CNN Noise data set in (1) and LSTM neural network training set (X, Y) LSTM Calculating a joint detection value set by the noise data set, and solving a detection threshold by combining the set false alarm probability; the method comprises the following steps:
(3.2.1) training set (X, Y) from CNN neural network CNN Separating L frequency spectrum occupation states as H 0 Forming a noisy data set of training samples
Figure FDA00036740854000000210
From LSTM neural network training set (X, Y) LSTM Separating L frequency spectrum occupation states as H 0 Forming a noisy data set of training samples
Figure FDA00036740854000000211
Figure FDA00036740854000000212
And
Figure FDA00036740854000000213
the corresponding time slots are the same;
(3.2.2) calculating the joint detection value of the frequency spectrum occupation state of each time slot in the noise data set by using the method in the step (3.1), and performing descending order arrangement on the joint detection values to obtain the joint detection value set of the noise data set
Figure FDA0003674085400000031
T(W (l) ) Representing the combined detection value arranged at the l-th position from large to small;
(3.2.3) setting the detection threshold
Figure FDA0003674085400000032
Wherein:
Figure FDA0003674085400000033
representing sets of joint detection values
Figure FDA0003674085400000034
The first element T (W) in (1) (l) )，
Figure FDA0003674085400000035
In order to set the false alarm probability,
Figure FDA0003674085400000036
represents rounding-down;
(3.3) online sample detection: and comparing the joint detection value of the detection sample acquired in real time with a detection threshold, and judging the frequency spectrum occupation condition.
2. The machine learning-based satellite spectrum sensing method according to claim 1, wherein: in the step (1), the historical received data of I channels of the single-user signal is subjected to energy normalization, and the received data of the ith channel of the kth time slot
Figure FDA0003674085400000037
Expressed as:
Figure FDA0003674085400000038
wherein: complex vector
Figure FDA0003674085400000039
A discrete complex signal of length N is received for the ith channel of the kth slot,
Figure FDA00036740854000000310
is a complex vector
Figure FDA00036740854000000311
K is the total number of time slots, I is the total number of channels contained in the single-user signal, and N is the signal length;
reception data of ith channel of kth time slot
Figure FDA00036740854000000312
Energy of
Figure FDA00036740854000000313
Expressed as:
Figure FDA00036740854000000314
wherein: energy (·) represents the energy addition to · s;
after the historical received data of the single user signal I channel in the kth time slot is subjected to energy normalization, the CNN neural network training set (X, Y) is obtained by splicing and combining CNN Training sample of the k-th time slot
Figure FDA00036740854000000315
The corresponding label is marked as
Figure FDA00036740854000000316
Figure FDA00036740854000000317
Figure FDA0003674085400000041
Figure FDA0003674085400000042
Wherein: training sample
Figure FDA0003674085400000043
Is a matrix of size nxi x 2,
Figure FDA0003674085400000044
is composed of
Figure FDA0003674085400000045
Of the first matrix of channels of (a),
Figure FDA0003674085400000046
is composed of
Figure FDA0003674085400000047
Of the second channel matrix of (a) is,
Figure FDA0003674085400000048
indicating the state of spectral occupancy, H, at the k-th time slot 0 Indicating that the spectrum is unoccupied, H 1 Indicating that the spectrum is occupied;
after the historical received data of the single user signal I channels in the kth time slot are subjected to energy normalization, an LSTM neural network training set (X, Y) is obtained by combination LSTM Training sample of the k-th time slot
Figure FDA0003674085400000049
The corresponding label is marked as
Figure FDA00036740854000000410
The spectrum occupation state of the k time slot is obtained by predicting the energy value sequence of the k-1, …, k-s time slot by an LSTM neural network:
Figure FDA00036740854000000411
Figure FDA00036740854000000412
Wherein: training sample
Figure FDA00036740854000000413
Is composed of the energy values of the first s time slots, s is the length of a backtracking window,
Figure FDA00036740854000000414
indicating the state of spectral occupancy, H, at the k-th time slot 0 Indicating that the spectrum is unoccupied, H 1 Indicating that the spectrum is occupied.
3. The machine learning-based satellite spectrum sensing method according to claim 1, wherein: in the step (3.3), when detecting the online sample, the joint detection value of the detection sample of a certain time slot acquired in real time is calculated first
Figure FDA00036740854000000417
Re-contrast joint detection value
Figure FDA00036740854000000416
And obtaining the spectrum occupation condition State of the detection sample with the detection threshold value gamma:
Figure FDA0003674085400000051
wherein: state indicates the spectrum occupancy of the detection sample.
CN202110136183.XA 2021-02-01 2021-02-01 Satellite spectrum sensing method based on machine learning Active CN112968740B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110136183.XA CN112968740B (en) 2021-02-01 2021-02-01 Satellite spectrum sensing method based on machine learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110136183.XA CN112968740B (en) 2021-02-01 2021-02-01 Satellite spectrum sensing method based on machine learning

Publications (2)

Publication Number Publication Date
CN112968740A CN112968740A (en) 2021-06-15
CN112968740B true CN112968740B (en) 2022-07-29

Family

ID=76272284

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110136183.XA Active CN112968740B (en) 2021-02-01 2021-02-01 Satellite spectrum sensing method based on machine learning

Country Status (1)

Country Link
CN (1) CN112968740B (en)

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106557809A (en) * 2015-09-30 2017-04-05 富士通株式会社 Nerve network system and the method is trained by the nerve network system
CN107995628A (en) * 2017-12-18 2018-05-04 北京工业大学 A kind of cognition wireless network multi-user Cooperation frequency spectrum sensing method of deep learning
WO2019010861A1 (en) * 2017-07-11 2019-01-17 北京邮电大学 Frequency spectrum prediction method and apparatus for cognitive wireless network
CN110309797A (en) * 2019-07-05 2019-10-08 齐鲁工业大学 Merge the Mental imagery recognition methods and system of CNN-BiLSTM model and probability cooperation
CN111600667A (en) * 2020-05-25 2020-08-28 电子科技大学 CNN-LSTM-based spectrum sensing method

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106557809A (en) * 2015-09-30 2017-04-05 富士通株式会社 Nerve network system and the method is trained by the nerve network system
WO2019010861A1 (en) * 2017-07-11 2019-01-17 北京邮电大学 Frequency spectrum prediction method and apparatus for cognitive wireless network
CN107995628A (en) * 2017-12-18 2018-05-04 北京工业大学 A kind of cognition wireless network multi-user Cooperation frequency spectrum sensing method of deep learning
CN110309797A (en) * 2019-07-05 2019-10-08 齐鲁工业大学 Merge the Mental imagery recognition methods and system of CNN-BiLSTM model and probability cooperation
CN111600667A (en) * 2020-05-25 2020-08-28 电子科技大学 CNN-LSTM-based spectrum sensing method

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
Deep Learning Aided Spectrum Prediction for Satellite Communication Systems;Xiaojin Ding等;《IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY》;20201231;全文 *

Also Published As

Publication number Publication date
CN112968740A (en) 2021-06-15

Similar Documents

Publication Publication Date Title
CN111126482B (en) Remote sensing image automatic classification method based on multi-classifier cascade model
CN111832225A (en) Method for constructing driving condition of automobile
CN111562108A (en) Rolling bearing intelligent fault diagnosis method based on CNN and FCMC
CN105303179A (en) Fingerprint identification method and fingerprint identification device
CN113852432B (en) Spectrum Prediction Sensing Method Based on RCS-GRU Model
CN110224771B (en) Spectrum sensing method and device based on BP neural network and information geometry
CN111597991A (en) Rehabilitation detection method based on channel state information and BilSTM-Attention
CN112787736B (en) Long-short term memory cooperative spectrum sensing method based on covariance matrix
CN112312541A (en) Wireless positioning method and system
CN113554156B (en) Multitask image processing method based on attention mechanism and deformable convolution
CN113315593A (en) Frequency spectrum sensing algorithm based on FLOM covariance matrix and LSTM neural network
CN112305441A (en) Power battery health state assessment method under integrated clustering
CN109450573A (en) A kind of frequency spectrum sensing method based on deep neural network
CN111144462A (en) Unknown individual identification method and device for radar signals
CN113343123B (en) Training method and detection method for generating confrontation multiple relation graph network
CN110717602A (en) Machine learning model robustness assessment method based on noise data
CN110072205A (en) A kind of layering aggregation method for wireless sense network anomaly data detection
CN112968740B (en) Satellite spectrum sensing method based on machine learning
CN116777183B (en) Unmanned ship cluster intelligent scheduling method and system
CN113379059A (en) Model training method for quantum data classification and quantum data classification method
CN112560981A (en) Training method, apparatus, device, program and storage medium for generating countermeasure model
CN115511012B (en) Class soft label identification training method with maximum entropy constraint
WO2023231374A1 (en) Semi-supervised fault detection and analysis method and apparatus for mechanical device, terminal, and medium
CN113705715B (en) Time sequence classification method based on LSTM and multi-scale FCN
CN112003662B (en) Cooperative spectrum sensing method and device based on dimensionality reduction and clustering in cognitive network

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