KR20210126912A - CNN Based Spectrum Sensing Technique for Cognitive Radio Communications - Google Patents

Abstract

The present invention relates to a method and a device for spectrum sensing in a cognitive radio system. According to the present invention, spectrum sensing can be performed using a convolutional neural network and without having to conduct noise power estimation. The present invention includes: a fast Fourier transform unit performing high-speed sampling of a reception signal of the entire band to be sensed and performing conversion into a frequency spectrum; an absolute value accumulation unit configuring a two-dimensional signal by accumulating a spectrum signal with respect to continuously received signals; an image data generation unit separating the two-dimensional signal by sensing channel bandwidth and combining a noise channel with each channel to generate convolutional neural network input data; and a spectrum sensing unit causing learning by inputting the signal generated at the image data generation unit to the convolutional neural network and determining the presence or absence of a main user signal with respect to the sensing channel based on the learning data.

Description

Spectral sensing method and apparatus based on convolutional neural network in cognitive radio system {CNN Based Spectrum Sensing Technique for Cognitive Radio Communications}

The present invention relates to a method and apparatus for spectrum sensing in a cognitive wireless system, and more particularly, to a method and apparatus capable of performing spectrum sensing without the need for estimating noise power using a convolutional neural network (CNN). will be.

The content described in this section merely provides background information for the present embodiment and does not constitute the prior art.

As the use of radio waves is rapidly increasing due to the development of wireless communication systems, the demand for efficient distribution and allocation of limited frequency resources is increasing. Cognitive radio (CR) is emerging as a technology to solve this problem.

The core technology of cognitive radio is spectrum sensing. Cognitive radio is a frequency sharing technology in which a secondary user (SU) finds an empty band and communicates without causing an interference signal to the primary user (PU) of the licensed band.

In general, cognitive radio technology uses threshold-based energy detection. The performance of individual spectral sensing, including energy detection, is strongly affected by the signal to noise ratio (SNR) of the main user's signal. In particular, accurate estimation of noise power for SNR estimation affects sensing performance.

As described above, in the conventional spectrum sensing using a threshold value, in a situation where the sub-user has no prior information on the main user signal, the power of the noise is first estimated and a threshold is set based on this, and then the power of the received signal is the threshold value. If it exceeds the threshold, it is determined that the main user signal exists, and if the power of the received signal does not exceed the threshold, it is determined that there is no signal. However, conventional spectral sensing exhibits excellent performance when the noise power is accurately estimated, but has a problem in that sensing performance is deteriorated when the estimated noise power is not accurate.

Therefore, in performing spectrum sensing, there is a need for a method that is not affected by spectrum sensing performance even in an environment with a bad SNR, and there is a need to propose a spectrum sensing method capable of performing spectrum sensing without estimating noise power.

Korean Patent No. 10-1298175 (Title of Invention: Spectrum Sensing Method and Apparatus)

Accordingly, it is an object of the present invention to provide a method and apparatus for spectrum sensing based on a convolutional neural network in a cognitive wireless system capable of performing spectrum sensing without the need for estimating noise power using a convolutional neural network.

Another object of the present invention is to accumulate a frequency spectrum of a signal and reconstruct it into a two-dimensional signal, cut it by channel bandwidth unit, image a two-dimensional matrix with a noise channel added to each channel, and input it into a convolutional neural network to perform spectrum sensing. An object of the present invention is to provide a method and apparatus for spectrum sensing based on a convolutional neural network in a wireless system.

Another object of the present invention is to provide a convolutional neural network-based spectrum sensing method and apparatus in a cognitive wireless system capable of determining the presence or absence of a main user signal based on a convolutional neural network in a situation where no prior information about the main user signal is known.

Another object of the present invention is to provide a convolutional neural network-based spectrum sensing method and apparatus in a cognitive wireless system capable of rapidly sampling the presence or absence of a main user signal based on a convolutional neural network in consideration of the entire band to be sensed.

In order to solve the above technical problem, the convolutional neural network-based spectrum sensing apparatus in the cognitive wireless system according to an embodiment of the present invention samples a received signal of the entire band to be sensed at high speed and converts it into a frequency spectrum at high speed. Fourier transform unit; an absolute value accumulator configured to form a two-dimensional signal by accumulating spectral signals for consecutively received signals; an image data generation unit that divides the two-dimensional signal into units of bandwidth of a sensing channel and generates input data of a convolutional neural network by combining a noise channel with each channel; and a spectrum sensing unit for inputting the signal generated by the image data generating unit into the convolutional neural network to learn it, and determining whether a main user signal exists with respect to the sensing channel based on the learned data.

Preferably, an analog/digital converter for converting the received signal of the entire band into a digital signal by high-speed sampling; It further includes an FFT frame collector for collecting the digital converted signal in units of FFT size and outputting it to a fast Fourier transform unit.

Preferably, the FFT frame collector collects data in units of FFT size, and collects data by overlapping adjacent signals.

Preferably, the absolute value accumulator takes an absolute value of each fast Fourier-transformed signal for all signal blocks to be sensed and accumulates it.

Preferably, the image data generator always leaves the last channel empty so that only noise can be observed, and combines the last channel with each sensing channel to form a two-dimensional matrix.

Preferably, the two-dimensional matrix Xc configured in the image data generator is calculated by the following [Equation 5].

[Equation 5]

는 크기가 K_C(센싱채널폭) × B(전체 블록 개수) 인 X의 부분 행렬이고, cK_C 부터 (c+1)K_C -1까지의 행을 선택해서 만든 행렬이다.

는 마지막 센싱 채널에 해당하는 부분 행렬이며 항상 잡음만 존재하는 채널이다. 그리고 N_FFT/K_C는 총 센싱채널 개수에 해당한다.) (here

is a partial matrix of X whose size is K _C (sensing channel width) × B (total number of blocks), and is a matrix created by selecting rows _{from cK C} to (c+1)K _{C -1.}

is a submatrix corresponding to the last sensing channel and is a channel in which only noise is always present. And N _FFT /K _C corresponds to the total number of sensing channels.)

Preferably, the image data generator reconstructs the two-dimensional matrix into a black-and-white image, and in the black-and-white image, a larger value is displayed closer to white, and a smaller value is displayed closer to black.

Preferably, the spectrum sensing unit determines the presence or absence of the main user signal using a binary classification convolutional neural network based on the learned data.

Preferably, the binary classification convolutional neural network comprises: an input layer to which a training data image with a size of 32 × B is input; first and second convolutional layers including a convolutional layer, a batch normalization layer, and a pooling layer; a third convolutional layer including a convolutional layer and a batch normalization layer; fully connected layer with two outputs; It includes a classification layer that judges the presence or absence of the main user signal, and the filter size in all convolutional layers is 3 × 3 and the stride is 1, and the pooling layer applies 2 × 2 max pooling with a stride of 2, The number of filters in the first convolution layer is 8, the number of filters in the second convolution layer is 16, the number of filters in the third convolution layer is 32, and the activation function in each convolution layer is ReLU.

Preferably, the spectrum sensing unit updates the parameters of the convolutional neural network so that cross entropy is minimized.

Preferably, the parameters for spectrum sensing by the spectrum sensing unit include a sampling clock, an FFT size, a sensing channel bandwidth, the number of channels, a length of a sensing channel bandwidth of a main user, a time shift, and an observation signal block length.

A method for spectrum sensing based on a convolutional neural network in a cognitive wireless system according to an embodiment of the present invention comprises: a first step of converting a received signal of an entire band to be sensed into a frequency spectrum by fast Fourier transforming; a second step of constructing a two-dimensional signal by accumulating the converted spectral signals with respect to the continuously received signals; A third step of generating two-dimensional image data by dividing a two-dimensional signal into a unit of bandwidth of a sensing channel, and combining a noise channel with each channel; and a fourth step of inputting image data into the convolutional neural network to learn, and determining whether a main user signal exists with respect to a sensing channel based on the learned data.

Preferably, the method further includes a fifth step of high-speed sampling of the received signal of the entire band, converting it into a digital signal, and collecting the digital signal in units of FFT size.

Preferably, step 5 collects FFT signals overlapped by a time shift.

Preferably, in the second step, the absolute value of the fast Fourier-transformed signal is taken and accumulated for all signal blocks to be sensed.

Preferably, in step 3, the last channel is always left blank so that only noise can be observed, and the last channel is combined with each sensing channel to form a two-dimensional matrix.

Preferably, the two-dimensional matrix (Xc) is calculated by the following [Equation 5].

[Equation 5]

는 크기가 K_C ×B인 X의 부분 행렬이고, cK_C 부터 (c+1)K_C -1까지의 행을 선택해서 만든 행렬이다.

는 마지막 센싱 채널에 해당하며 항상 잡음만 존재하는 채널이다.) (here

is a submatrix of X of _{size K C} × B, and is a matrix created by selecting rows _{from cK C} to (c+1)K _{C -1.}

corresponds to the last sensing channel and is a channel where only noise is always present.)

Preferably, in step 3, the two-dimensional matrix is reconstructed into a black-and-white image, and in the black-and-white image, a larger value is displayed closer to white, and a smaller value is displayed closer to black.

Preferably, step 4 determines the presence or absence of a main user signal using a binary classification convolutional neural network based on the learned data.

Preferably, the binary classification convolutional neural network comprises: an input layer to which a training data image with a size of 32 × B is input; first and second convolutional layers including a convolutional layer, a batch normalization layer, and a pooling layer; a third convolutional layer including a convolutional layer and a batch normalization layer; fully connected layer with two outputs; It includes a classification layer that judges the presence or absence of the main user signal, and the filter size in all convolutional layers is 3 × 3 and the stride is 1, and the pooling layer applies 2 × 2 max pooling with a stride of 2, The number of filters in the first convolution layer is 8, the number of filters in the second convolution layer is 16, the number of filters in the third convolution layer is 32, and the activation function in each convolution layer is ReLU.

Preferably, step 4 updates the parameters of the convolutional neural network so that cross entropy is minimized.

Preferably, the parameters for the four-step spectrum sensing include a sampling clock, an FFT size, a sensing channel bandwidth, the number of channels, a length of a sensing channel bandwidth of the main user, a time shift, and an observation signal block length.

A method and apparatus for spectrum sensing based on a convolutional neural network in a cognitive wireless system according to an embodiment of the present invention performs spectrum sensing based on a convolutional neural network. Therefore, according to the present invention, spectrum sensing is possible without the need for estimating noise power, so that the estimation accuracy of noise power is not affected and spectrum sensing performance is not affected even in an environment with a bad SNR.

In addition, the method and apparatus for spectrum sensing based on a convolutional neural network in a cognitive wireless system according to an embodiment of the present invention stacks a frequency spectrum of a signal into a two-dimensional signal without threshold-based energy detection, and divides each channel bandwidth into a channel Input the reconstructed image data including the noise channel to the convolutional neural network. Then, using the learning parameters learned from the convolutional neural network, spectral sensing of the received signal is performed. As described above, the present invention performs spectrum sensing without estimating SNR to achieve high spectral sensing performance.

In addition, the present invention has an advantage in that it is possible to determine the presence or absence of a main user signal based on a convolutional neural network in a situation where any prior information about the main user signal is not known, thereby increasing the efficiency of utilization of limited frequency resources.

In addition, the present invention can sample the presence or absence of a main user signal at high speed based on a convolutional neural network in consideration of the entire band to be sensed, thereby obtaining an effect of achieving high-speed distribution of frequency resources.

1 is a block diagram illustrating a signal reception path for performing spectral sensing based on a convolutional neural network in a cognitive wireless system according to an embodiment of the present invention.
2 illustrates an example of overlapping data when collecting FFT frames in an embodiment of the present invention.
3 is a diagram illustrating an example of learning data of a convolutional neural network-based spectrum sensing method in a cognitive wireless system according to an embodiment of the present invention.
4 illustrates a network structure of a convolutional neural network for spectrum sensing based on a convolutional neural network according to an embodiment of the present invention.
5 is a diagram showing the main characteristics used in the convolutional neural network-based spectrum sensing method according to an embodiment of the present invention.
6 is a graph illustrating a learning curve of a convolutional neural network-based spectrum sensing method according to an embodiment of the present invention to which the main characteristics shown in FIG. 5 are applied.
7 is a graph illustrating FDR and MDR error probabilities of observation time (signal block length B) according to the convolutional neural network-based spectrum sensing method according to an embodiment of the present invention.
8 is a graph comparing the performance of FDR and MDR of the convolutional neural network-based spectrum sensing method and the conventional threshold-based sensing method according to an embodiment of the present invention.
9 is a variable (α) determined as a function of the conventional SNR, and is an exemplary diagram of a positive real value experimentally found for each SNR section to provide optimal sensing performance.
10 is a flowchart illustrating an operation process of a spectrum sensing method based on a convolutional neural network according to an embodiment of the present invention.

Hereinafter, the embodiments disclosed in the present specification will be described in detail with reference to the accompanying drawings, but the same or similar components are assigned the same reference numerals regardless of reference numerals, and overlapping descriptions thereof will be omitted. The suffixes "part" and "group", "module" and "part", "unit" and "part", "device" and "system", "terminal" and "node" for components used in the description below. and "digital walkie-talkie" are given or mixed in consideration of the ease of writing the specification, and do not have distinct meanings or roles by themselves.

In addition, in describing the embodiments disclosed in the present specification, if it is determined that detailed descriptions of related known technologies may obscure the gist of the embodiments disclosed in the present specification, the detailed description thereof will be omitted. In addition, the accompanying drawings are only for easy understanding of the embodiments disclosed in the present specification, and the technical spirit disclosed herein is not limited by the accompanying drawings, and all changes included in the spirit and scope of the present invention , should be understood to include equivalents or substitutes.

Terms including an ordinal number, such as first, second, etc., may be used to describe various elements, but the elements are not limited by the terms. The above terms are used only for the purpose of distinguishing one component from another.

When a component is referred to as being “connected” or “connected” to another component, it is understood that the other component may be directly connected or connected to the other component, but other components may exist in between. it should be On the other hand, when it is said that a certain element is "directly connected" or "directly connected" to another element, it should be understood that no other element is present in the middle.

The singular expression includes the plural expression unless the context clearly dictates otherwise.

In the present application, terms such as "comprises" or "have" are intended to designate that a feature, number, step, operation, component, part, or combination thereof described in the specification exists, but one or more other features It should be understood that this does not preclude the existence or addition of numbers, steps, operations, components, parts, or combinations thereof.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. It is apparent to those skilled in the art that the present invention may be embodied in other specific forms without departing from the spirit and essential characteristics of the present invention.

1 is a block diagram illustrating a signal reception path for performing spectral sensing based on a convolutional neural network in a cognitive wireless system according to an embodiment of the present invention.

The cognitive radio system of the present invention receives a signal through the antenna (10). The received signal is subjected to frequency down-regulation of the received signal in the down conversion unit 20 , and is converted into a digital signal in the analog/digital conversion unit 30 . The output signal x(n) of the analog/digital converter 30 is input to the FFT frame collector 40 .

The FFT frame collector 40 reads the output signal of the analog/digital converter 30 in _{units of N FFTs} that are FFT (fast Fourier transform) sizes, and collects samples by overlapping adjacent N _{FFT signals.}

A process of taking a signal by overlapping data in the FFT frame collector 40 is shown in FIG. 2 .

In FIG. 2, N ₀ represents the time shift of the first sample when the next signal block is taken, and B represents the number of observed total signal blocks. At this time, the block of the m+1th collection signal (x _m ) is determined as in [Equation 1].

[Equation 1]

The sample collected by the FFT frame collector 40 is input to a fast Fourier transform unit 50 (FFT) to observe a frequency spectrum, and is converted into a spectrum. In the present invention, since it is necessary to determine the presence or absence of a signal of the main user for each frequency channel, it is determined by converting it into a frequency spectrum rather than a time domain.

라고 정의하면, [수학식 2]와 같다.The fast Fourier transform unit 50 inputs the sample collected by the FFT frame collector 40 (x _m ) and outputs a transformed signal (X _m ). N _FFT points FFT of the input signal (x _{m )}

If defined, it is the same as [Equation 2].

[Equation 2]

의 크기는 N_FFT ×1이며, 스펙트럼 센싱은 스펙트럼의 크기를 이용하여 판단해야 하므로

의 N_FFT개 원소에 대해 절대값 축적부(60)는 FFT 출력의 절대값을 취한다. 이 절대값을 취한 수신신호는 [수학식 3]과 같다.here

The size of is N _FFT × 1, and spectrum sensing should be determined using the size of the spectrum.

The absolute value accumulator 60 takes the absolute value of the FFT output for _{N FFT elements.} The received signal taking this absolute value is shown in [Equation 3].

[Equation 3]

The absolute value accumulator 60 performs FFT on a total of B signal blocks, respectively, and builds a _{matrix with a size of N FFT} × B by stacking blocks with absolute values to form a two-dimensional signal (X) and Equation 4].

[Equation 4]

On the other hand, the size of X, which is the output signal of the absolute value accumulator 60, is a _{two-dimensional matrix of N FFT} × B, and in order for this signal to be input to the convolutional neural network, data to be input to the convolutional neural network is separated by sensing channels. creation process is required.

If the sensing channel width is K _C , the total number of sensing channels is N _FFT / K _C . If the analog/digital conversion sampling clock is F _S , the bandwidth of each sensing channel becomes K _C F _S / N _FFT (Hz).

In the present invention, if there are a total of A frequency channels available to the main user, only A-1 can be allocated, and the remaining one channel is always set to be empty. This is to perform spectrum sensing using one empty channel as a noise channel. Therefore, the last sensing channel, N _FFT /K _C -1 th sensing channel, is always empty, and only noise is always observed.

The output signal X of the absolute value accumulation unit 60 is input to the image data generation unit 70 and is separated for each sensing channel. Then, a new two-dimensional matrix is created by including the noise channel, which is the last sensing channel, in the separated sensing channel, and this can be expressed by [Equation 5].

[Equation 5]

는 크기가 K_C(센싱채널폭) × B(전체 블록 개수) 인 X의 부분 행렬이고, cK_C 부터 (c+1)K_C -1까지의 행을 선택해서 만든 행렬이다.

는 마지막 센싱 채널에 해당하는 부분 행렬이며 항상 잡음만 존재하는 채널이다. 그리고 N_FFT/K_C는 총 센싱채널 개수에 해당한다. 따라서 본 발명은 이미지 데이터 생성부(70)에서 각 센싱 채널과 잡음 채널을 결합하여 새롭게 만든 2차원 행렬 X_C를 이용하여 주 사용자의 신호 존재 유무를 판단한다.here

is a partial matrix of X whose size is K _C (sensing channel width) × B (total number of blocks), and is a matrix created by selecting rows _{from cK C} to (c+1)K _{C -1.}

is a submatrix corresponding to the last sensing channel and is a channel in which only noise is always present. And N _FFT /K _C corresponds to the total number of sensing channels. Therefore, in the present invention, the presence or absence of a signal of the main user is determined using a _{two-dimensional matrix X C} newly created by combining each sensing channel and a noise channel in the image data generator 70 .

Accordingly, the output signal of the image data generating unit 70 is a two-dimensional matrix signal newly constructed by dividing each sensing channel in units of bandwidth and combining a noise channel with each of the separated sensing channels. Accordingly, the image data generator 70 generates data to be input to the convolutional neural network.

3 is a diagram illustrating an example of learning data of a convolutional neural network-based spectrum sensing method in a cognitive wireless system according to an embodiment of the present invention.

3 shows a two-dimensional signal X considering the entire sensing bandwidth and an example image X _C of a training data set.

The two-dimensional matrix may be represented as a black-and-white image as shown in FIG. 3 . In the example diagram shown, when N _FFT = 512, B = 64, K _C = 16, SNR = 20dB, X and X _C , the results obtained from [Equation 4], are white and black based on the presence or absence of the main user signal. can be represented as an image of

In the illustrated example, black represents a value close to 0, and white represents a large value close to 1. And the values in between can be displayed in gray.

Here, X _C is an example image of training data used in the spectrum sensing method of the present invention in which a noise channel is combined with each sensing channel, Busy indicates a case in which a main user signal is present, and Idle indicates a case in which a channel is empty.

That is, in the embodiment of the present invention, the spectrum sensing unit 80 determines the presence or absence of the main user signal through this image data. And, in the process of learning the convolutional neural network, learning is performed in a state where the result value (Busy or Idle) of the input learning data is known.

4 illustrates a network structure of a convolutional neural network for spectrum sensing based on a convolutional neural network according to an embodiment of the present invention.

The spectrum sensing unit 80 includes an input layer 703 to which a training data image is input, a first convolutional layer 704 , a second convolutional layer 705 , and a third convolutional layer ( Convolutional 706 ). , a fully connected layer having two outputs (707), and a classification layer (708) that finally determines whether the main user's signal is present.

The input layer 703 inputs learning data Xc corresponding to 2Kc×B, as shown in FIG. 3 . Since the width of each sensing channel was previously set to 16, the input image data has a size of 32 × B.

The first convolutional layer 704 and the second convolutional layer 705 include a convolutional layer, a batch normalization layer, and a pooling layer (Max pooling). The pooling layer applies 2x2 max pooling with a stride of 2. The number of filters in the first convolutional layer 704 is 8, and the number of filters in the second convolutional layer 705 is 16.

The third convolutional layer 706 includes a convolutional layer and a batch normalization layer, and the number of filters in the third convolutional layer 706 is 32.

The filter size of all convolutional layers included in the first, second, and third convolutional layers is 3 × 3, and the stride is unified as 1. The activation function in each convolutional layer uses a Rectified Linear Unit (ReLU).

The output of the third convolutional layer 706 goes through the fully connected layer 707 , has two outputs for binary classification, and finally determines the presence or absence of a signal of the main user through the classification layer 708 .

Next, in order to verify the performance of a convolutional neural network-based spectrum sensing method and apparatus in a cognitive wireless system according to an embodiment of the present invention, known MATLAB, which is engineering software supporting a matrix-based calculation function, is used. So, let's take a look at the simulation.

The main characteristics used in the spectrum sensing technique of the embodiment are shown in FIG. 5 .

The sampling frequency of the received signal is 16 MHz, and the bandwidth of the sensing channel is 0.5 MHz. Therefore, the main user signal has a channel bandwidth of 0.5 MHz, and therefore, there are a total of 32 channels in the received signal.

The number of collected signal blocks B is increased by 4 times from 4 to 64 to compare the performance, and as B is smaller, shorter signals are observed, and spectrum sensing is performed, so rapid sensing is possible. The main characteristic shown in Fig. 5 is a set value and can be updated in order to derive an optimal value.

The signals used for training are a total of 20,000 sets of main user signals. The bandwidth of the main user signal is 0.5 MHz, which is the same as the sensing bandwidth, and the main user signal is allocated to 31 channels except for the last one among 32 channels.

The probability of the main user signal in each channel is set to 0.5. Therefore, once a signal is generated, about half of the 31 allocable channels have the main user signal, and the other half is distributed without the signal. In the learning process, learning is performed in a state where the determination result of the channel in which the main user signal exists and the channel in which the main user signal does not exist is known.

Also, when generating 20,000 sets of signals, the SNR is randomly selected. Specifically, it is randomly selected in the range of -20dB to 50dB. In this range, the actual field environment is considered, and since 31 Xc (c=0,...,30) can be created with one set of signals, a total of 620,000 training data images are input to the convolutional neural network during the simulation process. do.

In this way, the two-dimensional data set is input through the input layer 703 of the convolutional neural network, training is performed so that cross entropy is minimized, and the mini-batch size is 620. At this time, the optimization technique for learning uses SGDM (Stochastic Gradient Descent with Momentum), and the learning rate is 0.001.

6 is a graph illustrating a learning curve of a convolutional neural network-based spectrum sensing method according to an embodiment of the present invention to which the main characteristics shown in FIG. 5 are applied.

6 shows the transition of the loss function during learning. For training, the entire training data image is repeatedly used, and as can be seen in the graph, the learning curve converges quickly, and it can be confirmed that the learning curve converges within about 100 updates.

Next, test data for performance verification is set as follows.

In the same environment as the training data, the SNR generates 4,000 sets of signals in the range of -20dB to 6dB at intervals of 2dB. The presence or absence of the main user signal for each channel is also randomly selected, and since the signal is composed of 31 channels, about 124,000 test data images are input to the convolutional neural network.

And in an embodiment of the present invention, two indicators are measured to verify the performance of the convolutional neural network-based spectrum sensing method.

The first is called missed detection when it is determined that the channel is being used even though it is empty, and the ratio is defined as MDR (Miss detection ratio). Second, a case where it is determined that the channel is in use and empty is called a detection error, and the ratio is defined as a false detection ratio (FDR). Here, since it is important to minimize the interference to the main user, it can be seen that FDR performance is more important than MDR.

7 is a graph illustrating FDR and MDR error probabilities of observation time (signal block length B) according to the convolutional neural network-based spectrum sensing method according to an embodiment of the present invention.

As shown, the FDR shows a performance gain of 3 to 4 dB whenever the observation time B increases by 4 times, and the performance is the best when B is 64. MDR also shows a similar gain with increasing observation time B. Therefore, it can be confirmed that there is a trade-off relationship between the detection speed and accuracy of the spectrum, and MDR and FDR show similar performance at the same B value.

8 is a graph comparing the performance of FDR and MDR of the convolutional neural network-based spectrum sensing method and the conventional threshold-based sensing method according to an embodiment of the present invention.

In the conventional method, the threshold λ is determined by the following [Equation 6] and [Equation 7].

[Equation 6]

[Equation 7]

In this process, since α is a positive real value and is a variable determined as a function of SNR, the experimentally found value as shown in FIG. 9 was used for each SNR section to provide optimal sensing performance.

In this case, B is 64, and as shown in FIG. 8 , it can be confirmed that the spectrum sensing method of the present invention has an overall performance gain of 2 dB compared to the conventional method.

Next, FIG. 10 is a flowchart illustrating an operation process of a spectrum sensing method based on a convolutional neural network according to an embodiment of the present invention.

In consideration of the entire band to be sensed by the simulator, a fast Fourier transform is performed on the received signal, and a process of converting the received signal into a frequency spectrum is performed (S110).

As shown in FIG. 3, the spectrum of the continuously received signal is stacked and reconstructed into a two-dimensional signal (S120). At this time, a one-dimensional vector of 512 × 1, which is the result of FFT of the signal, is converted into a two-dimensional matrix of 512 × 64. The two-dimensional matrix is reconstructed into a black-and-white image, and in the black-and-white image, a larger value is closer to white, a smaller value is closer to black, and an intermediate value is expressed in gray.

Then, the main user channel is divided into bandwidth units, and a noise channel is combined with each channel (S130). The training data for learning the convolutional neural network is generated in an additive white Gaussian noise (AWGN) environment.

Then, the convolutional neural network performs learning on the training data, which is the reconstructed image (S140). At this time, the parameters of the convolutional neural network are updated so that the replacement entropy is minimized.

Thereafter, when a channel signal for detecting the presence or absence of the main user signal is input, the presence or absence of the main user signal is determined through binary classification based on the learning data that is the result of learning in the convolutional neural network in the previous process ( S150 ).

As described above, the method and apparatus for spectrum sensing based on a convolutional neural network of a cognitive wireless system according to an embodiment of the present invention determines the presence or absence of a main user signal through spectrum sensing without knowing any prior information about the main user signal. To this end, the present invention samples the received signal at high speed in consideration of the entire band, and then converts it into a frequency spectrum through FFT of the signal. A two-dimensional signal is generated by continuously stacking the frequency spectrum converted in this way. Then, the generated two-dimensional signal is divided into units of a channel bandwidth to be detected, and the reconstructed image data is generated by combining each channel with a noise channel that does not include a main user signal. The image data generated in this way is input to the convolutional neural network, and it becomes possible to determine whether there is a main user signal in the input data for channel search based on the training data on which a series of learning is performed.

The above detailed description should not be construed as restrictive in all respects and should be considered as exemplary. The scope of the present invention should be determined by a reasonable interpretation of the appended claims, and all modifications within the equivalent scope of the present invention are included in the scope of the present invention.

10: antenna
20: down conversion unit
30: analog/digital conversion unit
40: FFT frame collector
50: fast Fourier transform unit
60: absolute value accumulation part
70: channel separation unit
80: spectrum sensing unit

Claims (22)

a fast Fourier transform unit for high-speed sampling of a received signal of the entire band to be sensed and converting it into a frequency spectrum;
an absolute value accumulator configured to form a two-dimensional signal by accumulating spectral signals for consecutively received signals;
an image data generation unit that divides the two-dimensional signal into units of bandwidth of a sensing channel and generates input data of a convolutional neural network by combining a noise channel with each channel;
Convolutional neural network-based spectrum in a cognitive wireless system including a spectrum sensing unit that inputs the signal generated by the image data generator to the convolutional neural network to learn, and determines whether a main user signal exists for a sensing channel based on the learned data sensing device.
The method according to claim 1,
an analog/digital converter that samples the entire band of the received signal at high speed and converts it into a digital signal;
A convolutional neural network-based spectrum sensing device in a cognitive wireless system further comprising an FFT frame collector that collects digital transformed signals in units of FFT size and outputs them to a fast Fourier transform unit.
3. The method according to claim 2,
The FFT frame collector is a spectral sensing device based on a convolutional neural network in a cognitive wireless system that collects data in units of FFT size and collects data by overlapping adjacent signals.
The method according to claim 1,
The absolute value accumulator is a convolutional neural network-based spectrum sensing device in a cognitive wireless system that takes and accumulates an absolute value of each fast Fourier-transformed signal for all signal blocks to be sensed.
The method according to claim 1,
The image data generator always leaves the last channel empty so that only noise can be observed, and a convolutional neural network-based spectrum sensing device in a cognitive wireless system that forms a two-dimensional matrix by combining the last channel with each sensing channel.
6. The method of claim 5,
A two-dimensional matrix (Xc) configured in the image data generator is a convolutional neural network-based spectrum sensing device in a cognitive wireless system calculated by the following [Equation 5].
[Equation 5]

(here

is a partial matrix of X whose size is K _C (sensing channel width) × B (total number of blocks), and is a matrix created by selecting rows _{from cK C} to (c+1)K _{C -1.}

is a submatrix corresponding to the last sensing channel and is a channel in which only noise is always present. And N _FFT /K _C corresponds to the total number of sensing channels.)
7. The method of claim 6,
The image data generator reconstructs a two-dimensional matrix into a black-and-white image, and a convolutional neural network-based spectrum sensing device in a cognitive wireless system in which a larger value is displayed closer to white in a black-and-white image, and a smaller value is displayed closer to black.
The method according to claim 1,
The spectrum sensing unit is a convolutional neural network-based spectrum sensing device in a cognitive wireless system that determines the presence or absence of a main user signal using a binary classification convolutional neural network based on the learned data.
9. The method of claim 8,
The binary classification convolutional neural network includes: an input layer to which a training data image with a size of 32 × B is input;
first and second convolutional layers including a convolutional layer, a batch normalization layer, and a pooling layer;
a third convolutional layer including a convolutional layer and a batch normalization layer;
fully connected layer with two outputs;
Includes a classification layer that determines the presence or absence of the main user signal,
In all convolutional layers, the filter size is 3 × 3 and the stride is 1, and the pooling layer applies 2 × 2 max pooling with a stride of 2, the number of filters in the first convolution layer is 8, and the number of filters in the second convolution layer is 8. A convolutional neural network-based spectrum sensing device in a cognitive wireless system in which the number of filters is 16, the number of filters in the third convolutional layer is 32, and the activation function in each convolutional layer is ReLU.
10. The method of claim 9,
The spectrum sensing unit is a convolutional neural network-based spectrum sensing device in a cognitive wireless system that updates parameters of a convolutional neural network so that cross entropy is minimized.
11. The method of claim 10,
The main characteristic values for spectrum sensing of the spectrum sensing unit include sampling clock, FFT size, sensing channel bandwidth, number of channels, length of sensing channel bandwidth of main user, time shift, and convolutional neural network in cognitive wireless system including observation signal block length. based spectrum sensing device.
A first step of converting the received signal of the entire band to be sensed into a frequency spectrum by fast Fourier transforming;
a second step of constructing a two-dimensional signal by accumulating the converted spectral signals with respect to the continuously received signals;
A third step of generating two-dimensional image data by dividing a two-dimensional signal into a unit of bandwidth of a sensing channel, and combining a noise channel with each channel;
A method for spectral sensing based on a convolutional neural network in a cognitive wireless system, comprising: inputting image data into a convolutional neural network to learn, and determining whether a main user signal exists for a sensing channel based on the learned data.
13. The method of claim 12,
Convolutional neural network-based spectrum sensing method in a cognitive wireless system, further comprising 5 steps of high-speed sampling of the entire band received signal, converting it to a digital signal, and collecting the digital signal in units of FFT size.
14. The method of claim 13,
Step 5 is a convolutional neural network-based spectrum sensing method in a cognitive wireless system that collects FFT signals overlapped by time shift.
13. The method of claim 12,
Step 2 is a spectrum sensing method based on a convolutional neural network in a cognitive wireless system that takes and accumulates absolute values of fast Fourier-transformed signals for all signal blocks to be sensed.
13. The method of claim 12,
Step 3 is a convolutional neural network-based spectrum sensing method in a cognitive wireless system that always leaves the last channel empty so that only noise can be observed, and forms a two-dimensional matrix by combining the last channel with each sensing channel.
17. The method of claim 16,
The two-dimensional matrix (Xc) is a spectrum sensing method based on a convolutional neural network in a cognitive wireless system calculated by the following [Equation 5].
[Equation 5]

(here

is a partial matrix of X whose size is K _C (sensing channel width) × B (total number of blocks), and is a matrix created by selecting rows _{from cK C} to (c+1)K _{C -1.}

is a submatrix corresponding to the last sensing channel and is a channel in which only noise is always present. And N _FFT /K _C corresponds to the total number of sensing channels.)
13. The method of claim 12,
Step 3 is a convolutional neural network-based spectral sensing method in a cognitive wireless system that reconstructs a two-dimensional matrix into a black-and-white image, and in a black-and-white image, larger values are displayed closer to white, and smaller values are displayed closer to black.
13. The method of claim 12,
Step 4 is a spectral sensing method based on a convolutional neural network in a cognitive wireless system that determines the presence or absence of a main user signal using a binary classification convolutional neural network based on the learned data.
20. The method of claim 19,
The binary classification convolutional neural network includes: an input layer to which a training data image with a size of 32 × B is input;
first and second convolutional layers including a convolutional layer, a batch normalization layer, and a pooling layer;
a third convolutional layer including a convolutional layer and a batch normalization layer;
fully connected layer with two outputs;
Includes a classification layer that determines the presence or absence of the main user signal,
In all convolutional layers, the filter size is 3 × 3 and the stride is 1, and the pooling layer applies 2 × 2 max pooling with a stride of 2, the number of filters in the first convolution layer is 8, and the number of filters in the second convolution layer is 8. The number of filters of is 16, the number of filters in the third convolutional layer is 32, and the activation function in each convolutional layer is ReLU. A convolutional neural network-based spectrum sensing method in a cognitive wireless system.
21. The method of claim 20,
Step 4 is a convolutional neural network-based spectrum sensing method in a cognitive wireless system that updates the parameters of the convolutional neural network so that cross entropy is minimized.
22. The method of claim 21,
The main characteristic values for the four-step spectrum sensing are the sampling clock, the FFT size, the sensing channel bandwidth, the number of channels, the length of the main user's sensing channel bandwidth, the time shift, and the convolutional neural network in the cognitive wireless system including the observation signal block length. based spectrum sensing method.

Images