Cyclic self-correlation spectrum sensing method and device based on cognitive radio

Abstract

The invention relates to the technical field of cognitive radio, in particular to a cyclic autocorrelation spectrum sensing method and device based on cognitive radio, and the method comprises the steps: inputting an OFDM (Orthogonal Frequency Division Multiplexing) signal, carrying out windowing processing and convolution solving, averaging the convolution, analyzing the cyclic autocorrelation of the OFDM signal, and obtaining a cyclic spectrum through an FAM (Fast Fourier Transform Accumulation) algorithm. The method comprises the following steps of: constructing a cyclic spectrum, normalizing the cyclic spectrum to form a cyclic autocorrelation gray level image, extracting depth features layer by layer through an improved AlexNet model, training a CNN (Convolutional Neural Network) by using a back propagation (BP) algorithm to perform a spectrum sensing training process and a spectrum sensing test process, and finally inputting a test set to verify the trained CNN model. According to the method, an AlexNet model is adopted to train a CNN, and the model uses more convolutional layers and a larger parameter space to fit a large-scale data set ImageNet. According to the FAM algorithm, the advantages of a cyclic spectrum are utilized, noise and a master user are distinguished, existence of the master user is detected, and better OFDM signal spectrum sensing performance is obtained under the condition of a low signal-to-noise ratio.

Description

Technical Field

[0001] This invention relates to the field of cognitive radio technology, specifically to a method and apparatus for cyclic autocorrelation spectrum sensing based on cognitive radio. Background Technology

[0002] Cognitive radio (CR) has garnered increasing attention as a potential solution to the current scarcity of spectrum. Spectrum sensing is a key technology in cognitive radio, aiming to identify available spectrum gaps that cognitive users (SUs) can exploit. Cyclic stationary feature detection (CPS) is a crucial technique in CPS, utilizing the cyclostationary characteristics of the received signal to detect primary users (PUs). However, CPS methods have particularly high computational complexity. In CR communication systems, PUs are expected to use Orthogonal Frequency Division Multiplexing (OFDM) waveforms. Therefore, OFDM-specific characteristics are crucial for spectrum sensing. Cyclic Prefix Autocorrelation (CP-AC) is an effective spectrum sensing method that overcomes the noise uncertainty problem in energy detection. Wideband spectrum sensing, effectively operating on multiple PU channels, can increase the probability of identifying unoccupied frequency bands.

[0003] Prior art document 1 (application number CN201810502819.6) discloses a spectrum sensing method based on the eigenvalues ​​of a cyclic matrix. This method aims to improve the spectrum sensing performance of unlicensed users with the same number of sampling points. The method includes: multiple unlicensed users collaboratively sensing signals from licensed users, who may sense the signals transmitted by the licensed users through their own devices; dividing the sensed signals into multiple received vectors according to sampling time, and determining the cyclic matrix for each received vector, which is obtained by cyclically shifting the received vectors in a certain order; calculating the average covariance matrix of the cyclic matrices of each received vector, and obtaining the eigenvalues ​​of the average covariance matrix, which is obtained by averaging the cyclic matrices of all received vectors; constructing a detection statistic for binary hypothesis detection, where the two hypotheses are: the existence of a licensed user and the non-existence of a licensed user, and the construction of the detection statistic depends on the eigenvalues ​​of the average covariance matrix; if the detection statistic is greater than a preset threshold, the licensed user is determined to exist; otherwise, the licensed user is determined to not exist. This method is applicable to scenarios in cognitive radio networks where the existence of a licensed user is determined. By combining signals sensed collaboratively by multiple unauthorized users with the calculation of cyclic matrices and average covariance matrices, the spectral sensing performance of unauthorized users can be improved. Furthermore, the presence or absence of authorized users can be more accurately determined through detection statistics. However, the eigenvalue detection method suffers from poor monitoring performance under low sampling point numbers and low signal-to-noise ratios.

[0004] Prior art document 2 (application number CN202111263098.6) discloses a cyclostationary feature spectrum sensing method based on pilot interpolation in the OFDM frequency domain. To achieve the goal of spectrum sensing, this invention employs a time-domain smoothing periodogram method to obtain the spectral correlation function of the OFDM signal. Then, feature points with pilot correlation data are selected in the spectral correlation function according to the rules of this invention. Using these feature points, a test statistic is constructed according to the method proposed in this invention, and its statistical characteristics are analyzed. Finally, spectrum sensing is achieved through statistical decision-making. The method includes: calculating the spectral correlation function of the OFDM signal using the time-domain smoothing periodogram method, which can obtain the spectral correlation function by periodically smoothing the signal; selecting feature points with pilot correlation data in the spectral correlation function according to the rules of this invention, where these feature points represent data associated with the pilot in the spectral correlation function; and constructing a test statistic using the selected feature points according to the method proposed in this invention. The test statistic is obtained by calculating and combining selected feature points. Analyzing the statistical characteristics of the test statistic, specifically its statistical distribution, reveals its performance in spectrum sensing. Statistical decision-making is then performed; based on set thresholds and statistical characteristics, a spectrum sensing decision can be made according to the magnitude of the test statistic to determine the presence of licensed users in the channel. By employing a time-domain smoothing periodogram method to obtain the spectral correlation function and selecting specific feature points within this function, the proposed method constructs a test statistic to achieve spectrum sensing and makes decisions based on its statistical characteristics. These steps help improve the spectrum sensing performance of unlicensed users with the same number of sampling points. However, noise and interference levels during the detection process may change with time and location, leading to uncertainty in noise power.

[0005] Prior art document 3 (application number CN201910659454.2) discloses a spectrum sensing method based on symmetrical peaks of the cyclic autocorrelation function of a modulated signal. The method of this invention is based on the calculation of the cyclic autocorrelation function and constructs a two-dimensional signal detection domain. Symmetrical peak points are then searched within this domain to determine the presence of the primary user signal. If a symmetrical peak point exists, the primary user signal is determined to exist; otherwise, it is determined not to exist. To ensure the false alarm probability of spectrum sensing, this invention introduces a symmetric peak point significance level factor, which is used during the search for symmetrical peak points. This method, by judging the symmetry of the peaks of the cyclic autocorrelation function of the modulated signal, requires no prior knowledge of the primary user signal or the channel. This method eliminates the impact of channel noise fluctuations on spectrum sensing performance and solves the signal spectrum sensing problem in environments with low signal-to-noise ratios and fluctuating channel noise. The improved algorithm can reliably determine the presence or absence of the primary user signal, thereby achieving efficient spectrum sensing. However, how to fully utilize the characteristics of the cyclic autocorrelation function of the master user modulation signal in the cognitive network for detection, and further improve the accuracy of spectrum sensing in low signal-to-noise ratio and noise fluctuation environments, is a problem that has not yet been completely solved.

[0006] The technical problem to be solved by this invention is that traditional spectrum sensing algorithms have poor spectrum sensing performance in complex and ever-changing wireless environments.

[0007] To address this, a cyclic autocorrelation spectrum sensing method and apparatus based on cognitive radio is proposed. Summary of the Invention

[0008] The purpose of this invention is to provide a method and apparatus for cyclic autocorrelation spectrum sensing based on cognitive radio. By using deep learning and cyclic spectrum, OFDM cyclic spectrum features are used to improve the dataset of CNN models, turning the spectrum sensing problem into an image processing problem and better leveraging the powerful learning capabilities of CNN.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] A cognitive radio-based cyclic autocorrelation spectrum sensing method includes the following steps:

[0011] S10, Input OFDM signal, the OFDM signal is windowed and then subjected to Fast Fourier Transform (FFT);

[0012] S20, the OFDM signal transformed by the FFT is respectively compared with... and Perform convolution to obtain the first convolution result and the second convolution result;

[0013] S30, convolve the first convolution result and the second convolution result again and calculate the average to output the OFDM signal cyclic spectrum;

[0014] S40, the cyclic spectrum of the OFDM signal is normalized to form a cyclic autocorrelation grayscale image, and then the depth features are extracted layer by layer through the improved AlexNet model. The CNN is trained using the backpropagation BP algorithm to carry out the spectrum perception training process.

[0015] The training process is divided into forward computation of data, backpropagation of error, and weight update; in the forward computation of data, the hidden layer output is defined as:

[0016]

[0017] in, X represents the weights of layer H. i f(·) represents the input of the current node, and f(·) represents the activation function of the current layer.

[0018] The output value of the output layer is defined as:

[0019]

[0020] in, This represents the output value of the hidden layer, and W represents the weight.

[0021] S50: After the spectrum sensing training process is completed, the spectrum sensing test process is carried out, and the test set is input to verify the trained CNN model.

[0022] S60, statistical test results, determine whether the main user exists.

[0023] Preferably, the spectrum sensing training process in S40 includes the following steps:

[0024] S401, acquire OFDM signal;

[0025] S402, Extract the cyclic peak characteristics of the OFDM signal;

[0026] S403, map the cyclic peak features;

[0027] S404, Train the set of mapped signals;

[0028] S405, the set obtained by training the AlexNet model in S404.

[0029] Preferably, the spectrum sensing test process in S50 includes the following steps:

[0030] S501, real-time acquisition of OFDM signals;

[0031] S502, Extract the cyclic peak characteristics of the OFDM signal;

[0032] S503, map the cyclic peak features;

[0033] S504, Test the set of mapped signals;

[0034] S505, training the spectrum sensing model;

[0035] S506, obtain test results.

[0036] Preferably, the OFDM baseband signal in S10 can be represented in the following form:

[0037]

[0038] Where t∈[0,T];

[0039] The OFDM signal model can be obtained through formula (4):

[0040]

[0041] Where L represents the total length of the OFDM signal, T c T represents chip time. s Let q(t) represent the effective period, and m represent the rectangular pulse. k，l This represents the l-th sampling point when the k-th OFDM symbol is inserted with the cyclic prefix CP.

[0042] Preferably, the m k，l The time-domain representation is:

[0043]

[0044] Where N represents the effective length, D represents the length of insertion of CPl = 0, 1, ..., L-1, and α k，n This represents the modulation data of the k-th OFDM signal on the n-th carrier, where n = 0, 1, ..., N-1.

[0045] Preferably, in the k-th OFDM symbol, there are n delayed sampling points. τ Time-domain data m k，l The autocorrelation function can be represented as

[0046] for:

[0047]

[0048] Where, 0≤n τ ≤N.

[0049] Preferably, the autocorrelation function of formula (4) can be expressed as:

[0050]

[0051] Where, τ N =|τ|-NT c OFDM has 2 cycles T c and T s R x (t, τ) has periodicity.

[0052] Preferably, the R x The expression for (t, τ) using the Fast Fourier Transform (FFT) can be as follows:

[0053]

[0054] Where α represents the cycle frequency; the Fourier transform of t will show discrete spectral lines. and Place.

[0055] Preferably, in S506, the accuracy of the test set can be expressed as:

[0056]

[0057] Where v represents when λ < λ th The number of tests at time U is such that there are U pairs of test sets {(x1, y1), (x2, y2), ..., (x... i y i )},λ th This indicates the test error threshold; the higher the accuracy of the test set, the better the spectrum sensing performance.

[0058] The difference between the predicted value and the true value can be expressed as:

[0059] λ=||Y W，b (x i )-y i || (10)

[0060] Among them, Y W，b (x i ) represents the output value of the last layer of the CNN model, W represents the weight, and b represents the offset.

[0061] Preferably, a cognitive radio-based cyclic autocorrelation spectrum sensing device includes:

[0062] The FFT transformation module is used to input OFDM signals, which are windowed and then subjected to Fast Fourier Transform (FFT).

[0063] The convolution module is used to convolve the OFDM signal transformed by the FFT with... and Perform convolution to obtain the first convolution result and the second convolution result; then convolve the first convolution result and the second convolution result again and calculate the average to output the OFDM signal cyclic spectrum;

[0064] The spectrum sensing training module is used to normalize the cyclic spectrum of the OFDM signal to form a cyclic autocorrelation grayscale image, and then extract deep features layer by layer through the improved AlexNet model, and use the backpropagation BP algorithm to train the CNN for spectrum sensing training process.

[0065] The spectrum sensing test module is used for the spectrum sensing test process, inputting a test set to verify the trained CNN model;

[0066] The decision module is used to statistically analyze test results and determine whether the main user exists.

[0067] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0068] 1. This invention uses the AlexNet model to train a CNN, which employs more convolutional layers and a larger parameter space to fit the large-scale ImageNet dataset. The ReLU nonlinear function is used as the activation function, significantly accelerating training time. Overlapping sampling is used in the pooling layers to effectively prevent overfitting. A novel Dropout method is introduced, preventing Dropout neurons from participating in forward propagation or backward propagation, thus reducing network training costs. It features good cross-GPU parallelism based on GPUs, and training is conducted using multiple GPUs.

[0069] 2. This invention analyzes the cyclic autocorrelation of OFDM signals and obtains the cyclic spectrum using the FFT accumulation algorithm (FAM). The cyclic spectrum is normalized to form a grayscale image of the cyclic autocorrelation, and then a modified AlexNet model is used to extract deep features layer by layer. Finally, the results are input into a test set to validate the trained CNN model. This FAM algorithm utilizes the advantages of the cyclic spectrum to achieve better OFDM signal spectral sensing performance under low signal-to-noise ratio conditions.

[0070] 3. This invention uses deep learning and cyclic spectrum, OFDM cyclic spectrum features to improve the dataset of CNN models, turning the spectrum perception problem into an image processing problem, and better leveraging the powerful learning capabilities of CNN. Attached Figure Description

[0071] Figure 1 This is a flowchart of the cyclic autocorrelation spectrum sensing process of the present invention;

[0072] Figure 2 This is a schematic diagram of the AlexNet structure of the present invention;

[0073] Figure 3 This is a graph showing the relationship between the amplitude of the cyclic autocorrelation function and the cyclic frequency in this invention.

[0074] Figure 4 This is a detection probability diagram for the cyclic autocorrelation contrast energy detection of the present invention;

[0075] Figure 5 The graph shows the detection performance of the present invention at different FFT lengths;

[0076] Figure 6 This is a graph showing the relationship between accuracy and training steps under five different network structures of the present invention. Detailed Implementation

[0077] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0078] Please see Figures 1 to 6 This invention provides a method and apparatus for cyclic autocorrelation spectrum sensing based on cognitive radio, the technical solution of which is as follows:

[0079] A cognitive radio-based cyclic autocorrelation spectrum sensing method includes the following steps:

[0080] S10, Input OFDM signal, the OFDM signal is windowed and then subjected to Fast Fourier Transform (FFT);

[0081] The OFDM baseband signal in S10 can be represented in the following form:

[0082]

[0083] Where t∈[0,T];

[0084] The OFDM signal model can be obtained through formula (2):

[0085]

[0086] Where L represents the total length of the OFDM signal, T c T represents chip time. s Let q(t) represent the effective period, and m represent the rectangular pulse. k，l This represents the l-th sampling point when the k-th OFDM symbol is inserted with the cyclic prefix CP.

[0087] The m k，l The time-domain representation is:

[0088]

[0089] Where N represents the effective length, D represents the length of insertion of CPl = 0, 1, ..., L-1, and α k，n This represents the modulation data of the k-th OFDM signal on the n-th carrier, where n = 0, 1, ..., N-1.

[0090] In the k-th OFDM symbol, there are n delayed sampling points τ Time-domain data m k，l The autocorrelation function can be expressed as:

[0091]

[0092] Where, 0≤n τ ≤N.

[0093] The autocorrelation function of formula (2) can be expressed as:

[0094]

[0095] Where, τ N =|τ|-NT c OFDM has 2 cycles T c and T s R x (t, τ) has periodicity.

[0096] The R x The expression for (t, τ) using the Fast Fourier Transform (FFT) can be as follows:

[0097]

[0098] Where α represents the cycle frequency; the Fourier transform of t will show discrete spectral lines. and Place.

[0099] S20, the OFDM signal transformed by the FFT is respectively compared with... and Perform convolution to obtain the first convolution result and the second convolution result;

[0100] S30, convolve the first convolution result and the second convolution result again and calculate the average to output the OFDM signal cyclic spectrum;

[0101] S40, the cyclic spectrum of the OFDM signal is normalized to form a cyclic autocorrelation grayscale image, and then the depth features are extracted layer by layer through the improved AlexNet model. The CNN is trained using the backpropagation BP algorithm to carry out the spectrum perception training process.

[0102] The training process is divided into forward computation of data, backpropagation of error, and weight update; in the forward computation of data, the hidden layer output is defined as:

[0103]

[0104] in, X represents the weights of layer H. i f(·) represents the input of the current node, and f(·) represents the activation function of the current layer.

[0105] The output value of the output layer is defined as:

[0106]

[0107] in, This represents the output value of the hidden layer, and W represents the weight.

[0108] The spectrum sensing training process in S40 includes the following steps:

[0109] S401, acquire OFDM signal;

[0110] S402, Extract the cyclic peak characteristics of the OFDM signal;

[0111] S403, map the cyclic peak features;

[0112] S404, Train the set of mapped signals;

[0113] S405, the set obtained by training the AlexNet model in S404.

[0114] S50: After the spectrum sensing training process is completed, the spectrum sensing test process is carried out, and the test set is input to verify the trained CNN model.

[0115] The spectrum sensing test process in S50 includes the following steps:

[0116] S501, real-time acquisition of OFDM signals;

[0117] S502, Extract the cyclic peak characteristics of the OFDM signal;

[0118] S503, map the cyclic peak features;

[0119] S504, Test the set of mapped signals;

[0120] S505, training the spectrum sensing model;

[0121] S506, obtain test results.

[0122] S506, the accuracy of the test set can be expressed as:

[0123]

[0124] Where v represents when λ < λ th The number of tests at time U is such that there are U pairs of test sets {(x1, y1), (x2, y2), ..., (x... i y i )},λ th This indicates the test error threshold; the higher the accuracy of the test set, the better the spectrum sensing performance.

[0125] The difference between the predicted value and the true value can be expressed as:

[0126] λ=||Y W，b (x i )-y i || (10)

[0127] Among them, Y W，b (x i ) represents the output value of the last layer of the CNN model, W represents the weight, and b represents the offset.

[0128] S60, statistical test results, determine whether the main user exists.

[0129] A cognitive radio-based cyclic autocorrelation spectrum sensing device, comprising:

[0130] The FFT transformation module is used to input OFDM signals, which are windowed and then subjected to Fast Fourier Transform (FFT).

[0131] The convolution module is used to convolve the OFDM signal transformed by the FFT with... and Perform convolution to obtain the first convolution result and the second convolution result; then convolve the first convolution result and the second convolution result again and calculate the average to output the OFDM signal cyclic spectrum;

[0132] The spectrum sensing training module is used to normalize the cyclic spectrum of the OFDM signal to form a cyclic autocorrelation grayscale image, and then extract deep features layer by layer through the improved AlexNet model, and use the backpropagation BP algorithm to train the CNN for spectrum sensing training process.

[0133] The spectrum sensing test module is used for the spectrum sensing test process, inputting a test set to verify the trained CNN model;

[0134] The decision module is used to statistically analyze test results and determine whether the main user exists.

[0135] As one embodiment of the present invention, refer to Figure 1 Flowchart of cyclic autocorrelation spectrum sensing.

[0136] Based on the above cyclic autocorrelation analysis, the FAM algorithm is used in simulation software, and the simulation parameters are shown in Table 1.

[0137] In Table 2, labels 0 and 1 represent the channel state as empty or busy, respectively. The positive and negative sample ratio is balanced at 1:1. Data is sampled and preprocessed at the receiver via Rayleigh fading and Gaussian white noise channels, with the data passing through these channels. The model is trained and tested using the TensorFlow deep learning framework. Spectrum sensing performance is determined by two factors: firstly, by P... f and P m Evaluation, by P m Derivation of the detection probability P d (P d =1-P m On the other hand, it is evaluated by the training and testing time of the CNN, which is the perceptual efficiency. The CNN parameter settings are shown in Table 3.

[0138] Table 1 Simulation Experiment Parameters

[0139]

[0140] Table 2 Dataset

[0141]

[0142] Table 3 CNN parameter settings

[0143]

[0144] Table 4 shows the design of CNNs using five different network architectures.

[0145]

[0146] As one embodiment of the present invention, refer to Figure 2The diagram shows the AlexNet architecture. The AlexNet network consists of eight layers, including five convolutional layers and three fully connected layers, using Softmax to achieve 1000-class classification. The input to AlexNet is a 224*224*3 image, which is preprocessed to a size of 227*227*3 before being used as the network's input. Using the ReLU nonlinear function as the activation function significantly speeds up training time. ReLU helps neural network models learn nonlinear relationships. During neural network training, ReLU can better capture nonlinear relationships in the data, thereby improving model performance. The output value of ReLU is fixed when the input is greater than 0, thus filtering out some unnecessary information and increasing the sparsity of the neural network. This reduces the complexity of the neural network, improves training speed, and avoids overfitting. ReLU is computationally very fast, only requiring a check if the input is greater than 0, making it faster than other commonly used activation functions (such as sigmoid and tanh). This is crucial in deep neural networks because deep neural networks are computationally very expensive; ReLU can effectively reduce computation time and cost. The derivative of ReLU is a constant 1 when the input is greater than 0, unlike sigmoid and tanh which gradually approach 0 as the input value increases or decreases. Therefore, it effectively avoids the vanishing gradient problem. The vanishing gradient problem is common in deep neural networks. It prevents gradients from being propagated to shallower layers during training, causing the weights of shallower layers to fail to update and ultimately preventing the model from training. Dropout is a method that randomly removes neurons during the learning process. During training, hidden layer neurons are randomly selected and then deleted. Each time data is passed, neurons to be deleted are randomly selected. During testing, the output of each neuron is multiplied by the deletion ratio during training.

[0147] As one embodiment of the present invention, refer to Figure 3 Graph showing the relationship between the magnitude of the cyclic autocorrelation function and the cyclic frequency. Figure 3 This describes the OFDM cyclic spectrum under condition H1 (H1: indicating the presence of an OFDM signal, i.e., the presence of the primary user). Baud rate f s =f m / 10, where f m It is the sampling frequency, when the cycle frequency α = 5f s At this point, the amplitude of the cyclic autocorrelation function reaches its maximum value.

[0148] As one embodiment of the present invention, refer to Figure 4The figure shows the detection probability diagram of cyclic autocorrelation (CAC) versus energy detection. As shown, the false alarm probability is approximately 0.0078 for both algorithms. When the signal-to-noise ratio (SNR) is less than -12 dB, the CAC algorithm outperforms the energy detection algorithm. The CAC algorithm uses fewer samples than the energy detection algorithm. While the energy detection algorithm uses 16,000 samples, the CAC algorithm only uses 768 samples.

[0149] As one embodiment of the present invention, refer to Figure 5 The detection performance graphs for different FFT lengths are shown. Figure 5 As shown, for different FFT lengths N fft =1024,N fft The detection probability at N=2048 increases with different signal-to-noise ratios. The figure shows that the detection probability increases with increasing FFT length. Increasing N... fft The symmetry of the estimated cyclic autocorrelation vector can be examined to improve its accuracy. Therefore, by more accurately estimating the cyclic autocorrelation function, the symmetry of the improved cyclic autocorrelation function is detected.

[0150] As one embodiment of the present invention, refer to Figure 6 The graph shows the relationship between accuracy and training steps for five different network architectures. The input dataset remained constant while the convolutional kernel size was modified for each of the five different CNN network architectures. The different layer structures are shown in Table 4. The parameter settings for training the five CNN models with different architectures under the same training and test sets are shown in Table 3. The accuracy of the training process for the five CNN models with different architectures was analyzed under the same training time conditions. Performance curves are described in... Figure 6 From Figure 6 It can be seen that when the number of training iterations is 10001 (the higher the number of training iterations, the better the optimization effect, but this invention only requires 10001 iterations and has no impact on the research problem), under the max pooling aggregation mode, the robustness of the five different CNN structures continuously increases with the increase of the number of training iterations. Compared with the other three structures, (3) and (4) are the best, with an actual training accuracy of 0.96.

[0151] In summary, a cyclic autocorrelation spectrum sensing method based on cognitive radio includes the following steps:

[0152] S10, Input OFDM signal, the OFDM signal is windowed and then subjected to Fast Fourier Transform (FFT);

[0153] The OFDM baseband signal in S10 can be represented in the following form:

[0154]

[0155] Where t∈[0,T];

[0156] The OFDM signal model can be obtained through formula (2):

[0157]

[0158] Where L represents the total length of the OFDM signal, T c T represents chip time. s Let q(t) represent the effective period, and m represent the rectangular pulse. k，l This represents the l-th sampling point when the k-th OFDM symbol is inserted with the cyclic prefix CP.

[0159] The m k，l The time-domain representation is:

[0160]

[0161] Where N represents the effective length, D represents the length of insertion of CPl = 0, 1, ..., L-1, and α k，n This represents the modulation data of the k-th OFDM signal on the n-th carrier, where n = 0, 1, ..., N-1.

[0162] In the k-th OFDM symbol, there are n delayed sampling points τ Time-domain data m k，l The autocorrelation function can be expressed as:

[0163]

[0164] Where, 0≤n τ ≤N.

[0165] The autocorrelation function of formula (2) can be expressed as:

[0166]

[0167] Where, τ N =|τ|-NT c OFDM has 2 cycles T c and T s R x (t, τ) has periodicity.

[0168] The R x The expression for (t, τ) using the Fast Fourier Transform (FFT) can be as follows:

[0169]

[0170] Where α represents the cycle frequency; the Fourier transform of t will show discrete spectral lines. and Place.

[0171] This invention employs FFT transform to calculate the cyclic autocorrelation function, providing a more accurate estimate and improving the symmetry of the cyclic autocorrelation function. The FAM algorithm leverages the advantages of the cyclic spectrum to achieve better OFDM signal spectral sensing performance under low signal-to-noise ratio conditions.

[0172] S20, the OFDM signal transformed by the FFT is respectively compared with... and Perform convolution to obtain the first convolution result and the second convolution result;

[0173] S30, convolve the first convolution result and the second convolution result again and calculate the average to output the OFDM signal cyclic spectrum;

[0174] S40, the cyclic spectrum of the OFDM signal is normalized to form a cyclic autocorrelation grayscale image, and then the depth features are extracted layer by layer through the improved AlexNet model. The CNN is trained using the backpropagation BP algorithm to carry out the spectrum perception training process.

[0175] The training process is divided into forward computation of data, backpropagation of error, and weight update; in the forward computation of data, the hidden layer output is defined as:

[0176]

[0177] in, X represents the weights of layer H. i f(·) represents the input of the current node, and f(·) represents the activation function of the current layer.

[0178] The output value of the output layer is defined as:

[0179]

[0180] in, This represents the output value of the hidden layer, and W represents the weight.

[0181] S50: After the spectrum sensing training process is completed, the spectrum sensing test process is carried out, and the test set is input to verify the trained CNN model.

[0182] S60, statistical test results, determine whether the main user exists.

[0183] The spectrum sensing training process in S40 includes the following steps:

[0184] S401, acquire OFDM signal;

[0185] S402, Extract the cyclic peak characteristics of the OFDM signal;

[0186] S403, map the cyclic peak features;

[0187] S404, Train the set of mapped signals;

[0188] S405, the set obtained by training the AlexNet model in S404.

[0189] This invention trains a CNN using the AlexNet model and employs the backpropagation (BP) algorithm. This model utilizes more convolutional layers and a larger parameter space to fit the large-scale ImageNet dataset. The ReLU nonlinear function is used as the activation function, significantly accelerating training time. Overlapping sampling is employed in the pooling layers to effectively prevent overfitting. A novel Dropout method is introduced, preventing Dropout neurons from participating in forward and backward propagation, thus reducing network training costs. It features good cross-GPU parallelism based on GPUs, and training is conducted using multiple GPUs.

[0190] The spectrum sensing test process in S50 includes the following steps:

[0191] S501, real-time acquisition of OFDM signals;

[0192] S502, Extract the cyclic peak characteristics of the OFDM signal;

[0193] S503, map the cyclic peak features;

[0194] S504, Test the set of mapped signals;

[0195] S505, training the spectrum sensing model;

[0196] S506, obtain test results.

[0197] S506, the accuracy of the test set can be expressed as:

[0198]

[0199] Where v represents when λ < λ th The number of tests at time U is such that there are U pairs of test sets {(x1, y1), (x2, y2), ..., (x... i y i )},λ th This indicates the test error threshold; the higher the accuracy of the test set, the better the spectrum sensing performance.

[0200] The difference between the predicted value and the true value can be expressed as:

[0201] λ=||YW，b (x i )-y i || (10)

[0202] Among them, Y W，b (x i ) represents the output value of the last layer of the CNN model, W represents the weight, and b represents the offset.

[0203] This invention employs a novel Dropout method, which essentially ensures the best accuracy on the test set, surpassing the accuracy on the training set. A single threshold is used to evaluate the test results; this method is low in complexity, simple to implement, and produces good detection results.

[0204] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A cyclic autocorrelation spectrum sensing method based on cognitive radio, characterized in that, Includes the following steps: S10, Input OFDM signal, the OFDM signal is windowed and then subjected to Fast Fourier Transform (FFT); S20, the OFDM signal for the FFT is respectively compared with... and Perform convolution to obtain the first convolution result and the second convolution result; S30, the first convolution result and the second convolution result are convolved again and averaged to output the OFDM signal cyclic spectrum; S40, the cyclic spectrum of the OFDM signal is normalized to form a cyclic autocorrelation grayscale image, and then the depth features are extracted layer by layer through the improved AlexNet model. The CNN is trained using the backpropagation BP algorithm to carry out the spectrum perception training process. The training process is divided into forward computation of data, backpropagation of error, and weight update; in the forward computation of data, the hidden layer output is defined as: in, X represents the weights of layer H. i f(·) represents the input of the current node, and f(·) represents the activation function of the current layer. The output value of the output layer is defined as: in, This represents the output value of the hidden layer, and W represents the weight. S50: After the spectrum sensing training process is completed, the spectrum sensing test process is carried out, and the test set is input to verify the trained CNN model. S60, statistical test results, determine whether the main user exists.

2. The method for cyclic autocorrelation spectrum sensing based on cognitive radio according to claim 1, characterized in that, The spectrum sensing training process in S40 includes the following steps: S401, acquire OFDM signal; S402, Extract the cyclic peak characteristics of the OFDM signal; S403, map the cyclic peak features; S404, Train the set of mapped signals; S405, the set obtained by training the AlexNet model in S404.

3. The method for cyclic autocorrelation spectrum sensing based on cognitive radio according to claim 1, characterized in that, The spectrum sensing test process in S50 includes the following steps: S501, real-time acquisition of OFDM signals; S502, Extract the cyclic peak characteristics of the OFDM signal; S503, map the cyclic peak features; S504, Test the set of mapped signals; S505, training the spectrum sensing model; S506, obtain test results.

4. The method for cyclic autocorrelation spectrum sensing based on cognitive radio according to claim 1, characterized in that, The OFDM baseband signal in S10 can be represented in the following form: Where t∈[0,T]; The OFDM signal model can be obtained through formula (4): Where L represents the total length of the OFDM signal, T c T represents chip time. s Let q(t) represent the effective period, and m represent the rectangular pulse. k，l This represents the l-th sampling point when the k-th OFDM symbol is inserted with the cyclic prefix CP.

5. The method for cyclic autocorrelation spectrum sensing based on cognitive radio according to claim 4, characterized in that: The m k，l The time-domain representation is: Where N represents the effective length, D represents the length of insertion of CPl = 0, 1, ..., L-1, and α k，n This represents the modulation data of the k-th OFDM signal on the n-th carrier, where n = 0, 1, ..., N-1.

6. The method for cyclic autocorrelation spectrum sensing based on cognitive radio according to claim 5, characterized in that: In the k-th OFDM symbol, there are n delayed sampling points τ Time-domain data m k，l The autocorrelation function can be expressed as: Where, 0≤n τ ≤N.

7. The method for cyclic autocorrelation spectrum sensing based on cognitive radio according to claim 6, characterized in that: The autocorrelation function of formula (4) can be expressed as: Where, τ N =|τ|-NT c OFDM has 2 cycles T c and T s R x (t, τ) has periodicity.

8. The method for cyclic autocorrelation spectrum sensing based on cognitive radio according to claim 7, characterized in that: The R x The expression for (t, τ) using the Fast Fourier Transform (FFT) can be as follows: Where α represents the cycle frequency; the Fourier transform of t will show discrete spectral lines. and Place.

9. The method for cyclic autocorrelation spectrum sensing based on cognitive radio according to claim 3, characterized in that, In S506, the accuracy of the test set can be expressed as: Where v represents when λ < λ th The number of tests at time U is such that there are U pairs of test sets {(x1, y1), (x2, y2), ..., (x... i y i )},λ th This indicates the test error threshold; the higher the accuracy of the test set, the better the spectrum sensing performance. The difference between the predicted value and the true value can be expressed as: λ=||Y W，b (x i )-and i || (10) Among them, Y W，b (x i ) represents the output value of the last layer of the CNN model, W represents the weight, and b represents the offset.

10. A cyclic autocorrelation spectrum sensing device based on cognitive radio, characterized in that, include: The FFT transformation module is used to input OFDM signals, which are windowed and then subjected to Fast Fourier Transform (FFT). The convolution module is used to convolve the OFDM signal transformed by the FFT with... and Perform convolution to obtain the first convolution result and the second convolution result; then convolve the first convolution result and the second convolution result again and calculate the average to output the OFDM signal cyclic spectrum; The spectrum sensing training module is used to normalize the cyclic spectrum of the OFDM signal to form a cyclic autocorrelation grayscale image, and then extract deep features layer by layer through the improved AlexNet model, and use the backpropagation BP algorithm to train the CNN for spectrum sensing training process. The spectrum sensing test module is used for the spectrum sensing test process, inputting a test set to verify the trained CNN model; The decision module is used to statistically analyze test results and determine whether the main user exists.

Images