A fusion feature extraction method and spectrum sensing method and system
By extracting signal characteristics from the time domain and frequency domain angles in spectrum perception and fusion, the problems of high complexity of time domain characteristics and single information in the prior art are solved, and more efficient spectrum perception performance and wireless spectrum utilization are achieved. Rate.
Patent Information
- Application Number
- CN202210648705.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-06-09
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2042-06-09
AI Technical Summary
The existing spectrum perception algorithms are complex and have a single information when utilizing time domain features, making it difficult to effectively improve spectrum perception performance, especially when the primary user prior information is not required.
Using a spectrum perception method based on feature fusion, the signal features are extracted from the time domain and frequency domain angles and the feature fusion is performed, and the autocorrelation and cross-correlation of the signal in the time dimension and the autocorrelation of the frequency dimension is used. and cross-correlation, construct fusion features to improve spectrum perception performance.
A spectrum perception model with low complexity is realized, channel detection capability is improved, collision probability between different users is reduced, and wireless spectrum utilization is improved.
Smart Images

Figure CN115065975B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of spectrum sensing technology, and more specifically, to a fusion feature extraction method and a spectrum sensing method and system. Background Art
[0002] With the growing demand for wireless communications, radio spectrum has gradually become a valuable resource. In this context, cognitive radio is proposed to solve the problem of insufficient spectrum resource utilization. Spectrum sensing is an important part of radio cognition. Many algorithms have been proposed to improve the performance, reliability, and practicality of spectrum sensing, providing technical support for improving spectrum utilization.
[0003] Traditional non-machine learning spectrum sensing algorithms include matched filter detection, cyclostationary feature detection, energy detection, etc. The core of matched filter detection is to detect the coherence of the signal, that is, the coherence of the received signal and the prior information. This method requires some prior information of the primary user, so it is not used in most cases. In contrast, cyclostationary feature detection does not require prior information. It mainly relies on the periodic characteristics of the signal to judge the channel state, that is, the correlation of the received signal in the previous and subsequent time. Its disadvantage is that the algorithm complexity is relatively high.
[0004] In practice, the most widely used method is the energy detection algorithm, which can avoid the above shortcomings. The energy detection algorithm is a spectrum sensing algorithm based on the frequency domain, which determines the channel state by calculating the energy of the signal in the entire spectrum. Energy detection can be used for both non-machine learning sensing and machine learning sensing. When used for non-machine learning sensing, each sensing node calculates the energy of the signal and derives the discrimination threshold to determine the channel state; when energy detection is used for non-machine learning sensing, the sensing algorithm mostly adopts the form of collaborative sensing, that is, each sensing node sends the energy of the sensed signal to the fusion center, and then uses machine learning methods to determine the channel state. This method does not require the derivation of the discrimination threshold, that is, the energy of the signal is regarded as a feature for machine learning. However, the information extracted by the energy detection algorithm is single, and the performance of spectrum sensing still has room for improvement.
[0005] The prior art cyclostationary characteristic spectrum sensing method based on OFDM frequency domain interpolation pilot uses the time domain smoothing periodogram method to obtain the spectrum correlation function of the OFDM signal, and selects characteristic points with pilot-related data in the spectrum correlation function according to the rules of the present invention, and uses these characteristic points to construct the test statistic according to the method proposed by the present invention and analyze its statistical characteristics, so as to make statistical decisions to achieve the purpose of spectrum sensing. However, this method only uses the time domain characteristics, the algorithm complexity is relatively high, and the performance loss needs to be compensated by extending the sensing time. Summary of the invention
[0006] In order to overcome the above technical problems, the present invention provides a fusion feature extraction method and a spectrum sensing method and system which do not require a priori information of the primary user, can simultaneously utilize the time domain and frequency domain features of the signal and have low algorithm complexity.
[0007] The technical solution of the present invention is as follows:
[0008] A spectrum sensing method based on feature fusion comprises the following steps:
[0009] S1. Establish a communication system model and obtain a signal matrix according to the communication system model;
[0010] S2, extracting the time domain features and frequency domain features of the signal matrix;
[0011] S3. Fusing the time domain features and the frequency domain features to obtain fused features.
[0012] This technical solution proposes a spectrum sensing method based on feature fusion, which extracts signal features of the communication system from two different perspectives, time domain and frequency domain, and then performs feature fusion. Compared with extracting frequency domain features or time domain features separately, this technical solution simultaneously utilizes the autocorrelation and cross-correlation of each sampled signal in the time dimension and the autocorrelation and cross-correlation of each sampled signal in the frequency dimension to obtain a fusion feature that takes into account the complementarity between time domain features and frequency domain features.
[0013] Furthermore, in step S1, the communication system model includes a primary user, a fusion center and M secondary users, the M secondary users all sense the primary user, and the M secondary users are all connected to the fusion center; the method for obtaining the signal matrix X is:
[0014] S11, the secondary user obtains the perception information as a sampling sequence, and then sends the sampling sequence to the fusion center. The length of the sampling sequence is N, and the expression of the sampling sequence is:
[0015]
[0016] Among them, w(n) represents additive Gaussian white noise, s(n) represents the useful signal transmitted by the primary user, and h(n) represents channel fading;
[0017] S12. The fusion center uses the sampling sequence to construct a signal matrix. The expression of the signal matrix is:
[0018]
[0019] Among them, x m =[x m (1)x m (2)...x m(N)] is the sampling sequence of the mth secondary user.
[0020] Furthermore, the time domain features of step S2 include a first time domain feature and a second time domain feature, and the specific method for extracting the time domain features and the frequency domain features of the signal matrix is:
[0021] S21. Calculate the covariance matrix of the signal matrix. The expression of the covariance matrix is:
[0022]
[0023] Where X is the signal matrix and N is the length of the sampling sequence;
[0024] S22, using the covariance matrix to calculate the first time domain feature and the second time domain feature, the expression of the first time domain feature is: The expression of the second time domain feature is: Among them, ∑r ij is the sum of all elements in the covariance matrix;
[0025] S23, obtain the center frequency f of the sampling sequence in the signal matrix c , take f c A neighborhood (f c -f 0 , f c +f 0 ) is used to construct the frequency domain features by taking the average amplitude of the amplitude spectrum in the inner part. The expression of the frequency domain features is:
[0026]
[0027] in, M represents the number of sensing nodes, F m (u) represents the Fourier transform sequence of the mth node sampling sequence, with a center frequency of f c The corresponding index in the Fourier transform sequence is F c ; 2f 0 Indicates the spectrum bandwidth to be counted, the spectrum bandwidth is less than or equal to the channel bandwidth; f s represents the sampling frequency, f s / N represents the spectrum resolution; 2F 0 +1 is an integer representing the index of the spectrum bandwidth in the Fourier transform sequence; this dimension feature is equivalent to the statistical reception signal at f c A neighborhood (f c -f 0 ,f c +f 0 ) is the average amplitude of the amplitude spectrum within.
[0028] Furthermore, step S3 uses a vector concatenation method to perform feature fusion to obtain fused features. The feature fusion formula is:
[0029] λ=(λ 1 ,λ 2 ,λ 3 ) T
[0030] Among them, λ is the fusion feature, λ 1 is the first time domain feature, λ 2 is the second time domain feature, λ 3 is the frequency domain feature.
[0031] A spectrum sensing method, using the fusion feature extracted by the fusion feature extraction method, comprises the steps of:
[0032] S4, using the fusion features as training samples to train the classifier to obtain a spectrum sensing model;
[0033] S5. Execute spectrum sensing tasks through the spectrum sensing model.
[0034] Furthermore, the classifier in step S4 is a K-means classifier or a GMM classifier.
[0035] Furthermore, the classifier in step S4 is a K-means classifier, and the specific method of training the classifier is:
[0036] S41, obtaining S fusion features as training samples;
[0037] S42. Randomly select two training samples from the S training samples as initial samples, and initialize the initial samples to class 0 and class 1. The class center C of class 0 0 and the class center C of class 1 1 Corresponding to the centers of the two initial samples respectively;
[0038] S43. For all training samples, calculate their class center C 0 and class center C 1 The Euclidean distance of , and divide them into class 0 and class 1 according to the minimum Euclidean distance criterion;
[0039] S44. Recalculate the class center C of class 0 0 and the class center C of class 1 1 ;
[0040] S45, if the center C 0 and class center C 1 If there is no change or the preset number of iterations is reached, step S46 is executed; otherwise, the cluster center C is updated. 0 and class center C 1And return to execute S43;
[0041] S46, judgment class center C 0 and class center C 1 Does it meet||C 0 || 2 >||C 1 || 2 , if so, then exchange the class center class center C 0 and class center C 1 , then execute step S47, if not, directly execute step S47;
[0042] S47, Conservation Center C 0 and class center C 1 , using the class center C 0 and class center C 1 As the final spectrum sensing model.
[0043] Furthermore, step S5 first sets the classification threshold T, and then cyclically performs the spectrum sensing task through the spectrum sensing model. The specific method is:
[0044] S51, sampling to obtain a signal matrix X;
[0045] S52, extracting the signal matrix X as the fusion feature λ;
[0046] S53, calculate the fusion feature λ to the class center C 0 and class center C 1 The Euclidean distance d0, d1;
[0047] S54. If d1 / d0>T, the fused feature λ is classified as class 1, that is, the perception result is that the channel is occupied; otherwise, the fused feature λ is classified as class 0, that is, the channel is not occupied.
[0048] A spectrum sensing system based on feature fusion, comprising: a signal matrix construction unit, a feature extraction unit, a feature fusion unit, a classifier training unit, and a spectrum sensing unit;
[0049] The signal matrix construction unit establishes a communication system model and obtains a signal matrix according to the communication system model; the feature extraction unit extracts the time domain features and frequency domain features of the signal matrix; the feature fusion unit fuses the time domain features and the frequency domain features to obtain fused features; the classifier training unit uses the fused features as training samples to train the classifier to obtain a spectrum sensing model; the spectrum sensing unit performs spectrum sensing tasks through the spectrum sensing model.
[0050] Furthermore, the communication system model includes a primary user, a fusion center and M secondary users, the M secondary users all sense the primary user, and the M secondary users are all connected to the fusion center; the method for the signal matrix construction unit to obtain the signal matrix X is:
[0051] S11, the secondary user obtains the perception information as a sampling sequence, and then sends the sampling sequence to the fusion center. The length of the sampling sequence is N, and the expression of the sampling sequence is:
[0052]
[0053] Among them, w(n) represents additive Gaussian white noise, s(n) represents the useful signal transmitted by the primary user, and h(n) represents channel fading;
[0054] S12. The fusion center uses the sampling sequence to construct a signal matrix. The expression of the signal matrix is:
[0055]
[0056] Among them, x m =[x m (1)x m (2)...x m (N)] is the sampling sequence of the mth secondary user.
[0057] The present invention proposes a spectrum sensing method and system based on feature fusion. Compared with the prior art, the beneficial effects of the technical solution of the present invention are: extracting the signal features of the communication system from two different perspectives, time domain and frequency domain, and then performing feature fusion. Compared with extracting frequency domain features or time domain features separately, the present invention simultaneously utilizes the autocorrelation and cross-correlation of each sampled signal in the time dimension and the autocorrelation and cross-correlation of each sampled signal in the frequency dimension, and obtains fusion features that consider the complementarity between time domain features and frequency domain features. The fused features are used for spectrum sensing to improve the performance of spectrum sensing and obtain a spectrum sensing model with lower complexity; the fused features are applied to machine learning classifiers, so that the classifier has better sample separability during machine learning, obtains higher channel detection capabilities, reduces the probability of collision between different users, and improves the utilization rate of wireless spectrum. BRIEF DESCRIPTION OF THE DRAWINGS
[0058] Figure 1 It is a schematic diagram of the steps of the fusion feature extraction method;
[0059] Figure 2 Schematic diagram of the steps of spectrum sensing method;
[0060] Figure 3 It is a schematic diagram of the communication system model structure;
[0061] Figure 4 Amplitude spectrogram of the fast Fourier transform for classification of sampled sequence signals;
[0062] Figure 5 is the amplitude spectrum of the fast Fourier transform of the noise component of the sampling sequence;
[0063] Figure 6 ROC curves of different features when SNR is -14dB;
[0064] Figure 7 ROC curves of different features when SNR is -16dB;
[0065] Figure 8 ROC curves of different features when SNR is -14dB;
[0066] Fig. 9 ROC curves of different features when SNR is -16dB;
[0067] Fig.10 Schematic diagram of the spectrum sensing system structure based on feature fusion. DETAILED DESCRIPTION
[0068] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.
[0069] Example 1
[0070] A fusion feature extraction method, such as Figure 1 As shown, the steps include:
[0071] S1. Establish a communication system model and obtain a signal matrix according to the communication system model;
[0072] S2, extracting the time domain features and frequency domain features of the signal matrix;
[0073] S3. Fusing the time domain features and the frequency domain features to obtain fused features.
[0074] This embodiment extracts signal features of the communication system from two different perspectives, time domain and frequency domain, and then performs feature fusion. Compared with extracting frequency domain features or time domain features separately, this embodiment simultaneously utilizes the autocorrelation and cross-correlation of each sampled signal in the time dimension and the autocorrelation and cross-correlation of each sampled signal in the frequency dimension to obtain fusion features that take into account the complementarity between time domain features and frequency domain features.
[0075] Example 2
[0076] On the basis of Example 1, in this embodiment, the communication system model in step S1 includes a primary user, a fusion center and M secondary users, the M secondary users all sense the primary user, and the M secondary users are all connected to the fusion center; the primary user and the secondary user are connected to a common radio propagation channel, and the secondary user and the fusion center are connected to a dedicated channel for transmitting a sampled signal; the method for obtaining the signal matrix X is:
[0077] S11, the secondary user obtains the perception information as a sampling sequence, and then sends the sampling sequence to the fusion center. The length of the sampling sequence is N, and the expression of the sampling sequence is:
[0078]
[0079] Among them, w(n) represents additive Gaussian white noise, s(n) represents the useful signal transmitted by the primary user, and h(n) represents channel fading;
[0080] S12. The fusion center uses the sampling sequence to construct a signal matrix. The expression of the signal matrix is:
[0081]
[0082] Among them, x m =[x m (1)x m (2)...x m (N)] is the sampling sequence of the mth secondary user.
[0083] The time domain features of step S2 include the first time domain features and the second time domain features. The specific method of extracting the time domain features and the frequency domain features of the signal matrix is:
[0084] S21. Calculate the covariance matrix of the signal matrix. The expression of the covariance matrix is:
[0085]
[0086] Where X is the signal matrix and N is the length of the sampling sequence;
[0087] S22, using the covariance matrix to calculate the first time domain feature and the second time domain feature, the expression of the first time domain feature is: The expression of the second time domain feature is: Among them, ∑r ij is the sum of all elements in the covariance matrix;
[0088] S23, obtain the center frequency f of the sampling sequence in the signal matrix c , take f c A neighborhood (f c -f 0 , fc +f 0 ) is used to construct the frequency domain features by taking the average amplitude of the amplitude spectrum in the inner part. The expression of the frequency domain features is:
[0089]
[0090] in, M represents the number of sensing nodes, F m (u) represents the Fourier transform sequence of the mth node sampling sequence, with a center frequency of f c The corresponding index in the Fourier transform sequence is F c ; 2f 0 Indicates the spectrum bandwidth to be counted, the spectrum bandwidth is less than or equal to the channel bandwidth; f s represents the sampling frequency, f s / N represents the spectrum resolution; 2F 0 +1 is an integer representing the index of the spectrum bandwidth in the Fourier transform sequence; this dimension feature is equivalent to the statistical reception signal at f c A neighborhood (f c -f 0 ,f c +f 0 ) is the average amplitude of the amplitude spectrum within.
[0091] Step S3 uses the vector concatenation method to perform feature fusion to obtain fusion features. The formula for feature fusion is:
[0092] λ=(λ 1 ,λ 2 ,λ 3 ) T
[0093] Among them, λ is the fusion feature, λ 1 is the first time domain feature, λ 2 is the second time domain feature, λ 3 is the frequency domain feature.
[0094] Example 3
[0095] A spectrum sensing method such as Figure 2 As shown, the steps include:
[0096] S1. Establish a communication system model and obtain a signal matrix according to the communication system model;
[0097] S2, extracting the time domain features and frequency domain features of the signal matrix;
[0098] S3, fusing the time domain features and the frequency domain features to obtain fused features;
[0099] S4, using the fusion features as training samples to train the classifier to obtain a spectrum sensing model;
[0100] S5. Execute spectrum sensing tasks through the spectrum sensing model.
[0101] This embodiment extracts signal features of the communication system from two different perspectives, time domain and frequency domain, and then performs feature fusion, utilizing the complementarity between time domain features and frequency domain features, thereby improving the performance of spectrum sensing and obtaining a spectrum sensing model with lower complexity; the fused features are applied to the machine learning classifier, so that the classifier has better sample separability during machine learning, obtains higher channel detection capabilities, reduces the probability of collision between different users, and improves wireless spectrum utilization.
[0102] The classifier in step S4 is a K-means classifier or a GMM classifier.
[0103] Taking the K-means classifier as an example, the specific method of training the classifier is:
[0104] S41, obtaining S fusion features as training samples;
[0105] S42. Randomly select two training samples from the S training samples as initial samples, and initialize the initial samples to class 0 and class 1. The class center C of class 0 0 and the class center C of class 1 1 Corresponding to the centers of the two initial samples respectively;
[0106] S43. For all training samples, calculate their class center C 0 and class center C 1 The Euclidean distance of , and divide them into class 0 and class 1 according to the minimum Euclidean distance criterion;
[0107] S44. Recalculate the class center C of class 0 0 and the class center C of class 1 1 ;
[0108] S45, if the center C 0 and class center C 1 If there is no change or the preset number of iterations is reached, step S46 is executed; otherwise, the cluster center C is updated. 0 and class center C 1 And return to execute S43;
[0109] S46, judgment class center C 0 and class center C 1 Does it meet||C 0 || 2 >||C 1 || 2, if so, then exchange the class center class center C 0 and class center C 1 , then execute step S47, if not, directly execute step S47;
[0110] S47, Conservation Center C 0 and class center C 1 , using the class center C 0 and class center C 1 As the final spectrum sensing model.
[0111] Step S5 first sets the classification threshold T, and then executes the spectrum sensing task cyclically through the spectrum sensing model. The specific method is:
[0112] S51, sampling to obtain a signal matrix X;
[0113] S52, extracting the signal matrix X as the fusion feature λ;
[0114] S53, calculate the fusion feature λ to the class center C 0 and class center C 1 The Euclidean distance d0, d1;
[0115] S54. If d1 / d0>T, the fused feature λ is classified as class 1, that is, the perception result is that the channel is occupied; otherwise, the fused feature λ is classified as class 0, that is, the channel is not occupied.
[0116] Example 4
[0117] A spectrum sensing method comprises the steps of:
[0118] S1. Establish a communication system model and obtain a signal matrix according to the communication system model;
[0119] Establish a communication system model where a single primary user occupies a fixed channel, such as Figure 3 As shown in Figure 1, there are M sensing nodes in the communication system model. Each sensing node sends the sensed information to the fusion center, which then determines the channel status. Spectrum sensing can be described as a two-classification problem: judging the channel status as occupied (using H 1 Indicated by) or unoccupied (indicated by H 0 For each sensor node, the length of their sampling sequence is N. Then the sampling sequence of each sensor node can be expressed as the following formula:
[0120]
[0121] Where w(n) represents additive Gaussian white noise, s(n) represents the useful signal transmitted by the primary user, and h(n) represents channel fading. Each sensing node transmits the received sampling sequence to the fusion center, and the fusion center obtains the matrix X:
[0122]
[0123] S2. Extracting the time domain features and frequency domain features of the signal matrix; the time domain features include the first time domain features and the second time domain features. The specific method for extracting the time domain features and frequency domain features of the signal matrix is:
[0124] For a general RF signal without a DC component, its mean is 0. Then the covariance matrix of the sampling sequence can be expressed as the following formula:
[0125]
[0126] The covariance matrix is a semi-positive definite matrix. According to the definition of covariance, the elements on the diagonal of the covariance matrix represent the degree of autocorrelation of the sensing nodes, and the elements on the off-diagonal represent the degree of association between different sensing nodes. From the sampling sequence expression, it can be obtained that for the state H 0 The variance of the sampling sequence corresponding to each sensing node is equal to the variance σ of the additive Gaussian white noise. 2 , their covariance is close to 0 due to the lack of correlation. 1 The variance of the sampling sequence will be the superposition of the noise variance and the signal variance, their average variance will be greater than the average variance of the noise, and due to the correlation of the nodes, the covariance will not be equal to 0 but greater than 0. Based on this qualitative analysis, the average variance of the sensing node is taken as the first time domain feature λ 1 , take the average covariance of the sensing node as the second time domain feature λ 2 , as shown in the following formula:
[0127]
[0128] Among them, ∑r ij is the sum of all elements in the covariance matrix.
[0129] The energy of an unmodulated sinusoidal carrier is concentrated at the carrier frequency f in the frequency domain. c , when the Nyquist sampling theorem is satisfied, the amplitude spectrum of the fast Fourier transform of the received signal is only when f = f c The amplitude at other frequencies is 0. However, due to signal modulation or spectrum leakage, the energy of the carrier will not be concentrated on a single frequency, but will be expressed as f cThe spectrum diffusion trend is centered on the Gaussian white noise. In contrast, the spectrum of Gaussian white noise follows a uniform distribution. Assume that the primary user uses 2PSK modulation with a center frequency of 80MHz. When the sensing node sampling length is 1000, the amplitude spectrum of the fast Fourier transform of the sampling sequence signal component is as follows: Figure 4 Compared with this, the amplitude spectrum of the noise component fast Fourier transform is shown as Figure 5 shown.
[0130] It can be seen that the amplitude spectrum of the two is significantly different. Some other sine carrier modulation methods such as 2ASK and 8QAM have similar results. Only the spectrum of 2FSK is concentrated on two frequency points. Since the amplitude spectrum of Gaussian white noise is evenly distributed, the amplitude spectrum of the main user is concentrated on f c Therefore, taking advantage of this point, we take f c A neighborhood (f c -f 0 , f c +f 0 ) to construct the frequency domain features:
[0131]
[0132] Where M represents the number of sensing nodes, F m (u) represents the Fourier transform sequence of the mth node sampling sequence, with a center frequency of f c The corresponding index in the Fourier transform sequence is F c ; 2f 0 is the spectrum bandwidth to be counted, which should be less than or equal to the channel bandwidth; f s represents the sampling frequency, then f s / N represents the spectrum resolution. 0 +1 is an integer representing the index of the spectrum bandwidth in the Fourier transform sequence. This dimension feature is equivalent to the statistical reception signal at f c A neighborhood (f c -f 0 ,f c +f 0 ) is the average amplitude of the amplitude spectrum within.
[0133] S3, fusing the time domain features and the frequency domain features to obtain fused features;
[0134] Now we explain the meaning of each dimension feature to indicate the complementarity of the features. 1 is the variance of the signal. The mathematical meaning of variance is the degree to which the sample deviates from the mean. Physically, the signal variance is equivalent to the signal energy divided by the number of samples N. The second time domain feature λ 2is the covariance of the signal. Its mathematical meaning is the linear correlation between two variables. Physically speaking, it is the linear correlation between the sampling signals of two nodes at the same time. The frequency domain feature λ 3 It is extracted in the frequency domain, and its physical meaning is the average amplitude of the signal in a certain frequency band. 1 , the second time domain feature λ 2 The complementarity of is that their mathematical and physical meanings are different. The first time domain feature λ 1 , frequency domain characteristics λ 3 The complementarity lies in: the first time domain feature λ 1 The frequency domain feature λ is counted 3 The signal energy that is not counted is broader, and the frequency domain feature λ 3 is more accurate. The second time domain feature λ 2 , frequency domain characteristics λ 3 The complementarity lies in their different extraction domains and meanings. There is a certain connection and difference between the time domain and the frequency domain. The connection is that they can be converted to each other through Fourier transform and inverse Fourier transform. The difference is that the order of appearance of the two signals is important in the time domain, but the order is not important when the signal is transformed to the frequency domain. In general, the above-mentioned dimensional features have certain complementarity and low dimensionality. The above features can be fused using the vector concatenation method. The formula for feature fusion is:
[0135] λ=(λ 1 ,λ 2 ,λ 3 ) T
[0136] After fusion, the fused feature λ is obtained, which is different from the traditional two-dimensional feature.
[0137] S4, using the fusion features as training samples to train the classifier to obtain a spectrum sensing model;
[0138] After obtaining the fused feature λ as a feature sample, the classifier can be trained using machine learning methods. There are many choices for classifiers. A more suitable classifier is the clustering model because they do not require sample labels. Two clustering algorithms are listed here: K-means and GMM clustering. K-means is a centroid-based clustering algorithm. Its training process is as follows:
[0139]
[0140] GMM clustering is also a centroid-based clustering algorithm. The process of fitting GMM using maximum likelihood is as follows:
[0141]
[0142] Using any of the methods listed above, we can get a clustering model that can be used to perform perception tasks. It is easy to get that the clusters with a larger mean vector describe the distribution of H. 1 Therefore, as long as a threshold T is set, the above model can be used to determine the channel state.
[0143] Figure 6 and Figure 7 The ROC curves of different features using the K-means clustering method are shown when the signal-to-noise ratio is -14dB and -16dB; Figure 8 and Fig. 9 The ROC curves of different features using the GMM clustering method are shown when the signal-to-noise ratio is -14dB and -16dB. As can be seen from the figure, the ROC curve of the fusion feature when using the K-means and GMM clustering algorithms is significantly better than the traditional ED feature, IQ feature, and DAR feature.
[0144] Example 5
[0145] A spectrum sensing system, such as Fig. 9 As shown, it includes: a signal matrix construction unit, a feature extraction unit, a feature fusion unit, a classifier training unit, and a spectrum sensing unit;
[0146] The signal matrix construction unit establishes a communication system model and obtains a signal matrix according to the communication system model; the feature extraction unit extracts the time domain features and frequency domain features of the signal matrix; the feature fusion unit fuses the time domain features and the frequency domain features to obtain fused features; the classifier training unit uses the fused features as training samples to train the classifier to obtain a spectrum sensing model; the spectrum sensing unit performs spectrum sensing tasks through the spectrum sensing model.
[0147] The communication system model includes a primary user, a fusion center and M secondary users. The M secondary users all sense the primary user and are connected to the fusion center. The signal matrix construction unit obtains the signal matrix X by:
[0148] S11, the secondary user obtains the perception information as a sampling sequence, and then sends the sampling sequence to the fusion center. The length of the sampling sequence is N, and the expression of the sampling sequence is:
[0149]
[0150] Among them, w(n) represents additive Gaussian white noise, s(n) represents the useful signal transmitted by the primary user, and h(n) represents channel fading;
[0151] S12. The fusion center uses the sampling sequence to construct a signal matrix. The expression of the signal matrix is:
[0152]
[0153] Among them, x m =[x m (1)x m (2)...x m (N)] is the sampling sequence of the mth secondary user.
[0154] The time domain features include the first time domain features and the second time domain features. The specific method for the feature extraction unit to extract the time domain features and the frequency domain features of the signal matrix is:
[0155] S21. Calculate the covariance matrix of the signal matrix. The expression of the covariance matrix is:
[0156]
[0157] Where X is the signal matrix and N is the length of the sampling sequence;
[0158] S22, using the covariance matrix to calculate the first time domain feature and the second time domain feature, the expression of the first time domain feature is: The expression of the second time domain feature is: Among them, ∑r ij is the sum of all elements in the covariance matrix;
[0159] S23, obtain the center frequency f of the sampling sequence in the signal matrix c , take f c A neighborhood (f c -f 0 , f c +f 0 ) is used to construct the frequency domain features by taking the average amplitude of the amplitude spectrum in the inner part. The expression of the frequency domain features is:
[0160]
[0161] in, M represents the number of sensing nodes, F m (u) represents the Fourier transform sequence of the mth node sampling sequence, with a center frequency of f c The corresponding index in the Fourier transform sequence is F c ; 2f 0 Indicates the spectrum bandwidth to be counted, the spectrum bandwidth is less than or equal to the channel bandwidth; f s represents the sampling frequency, f s / N represents the spectrum resolution; 2F 0 +1 is an integer representing the index of the spectrum bandwidth in the Fourier transform sequence; this dimension feature is equivalent to the statistical reception signal at f c A neighborhood (f c-f 0 ,f c +f 0 ) is the average amplitude of the amplitude spectrum within.
[0162] The feature fusion unit uses the vector concatenation method to fuse features and obtain fused features. The formula for feature fusion is:
[0163] λ=(λ 1 ,λ 2 ,λ 3 ) T
[0164] Among them, λ is the fusion feature, λ 1 is the first time domain feature, λ 2 is the second time domain feature, λ 3 is the frequency domain feature.
[0165] The classifier trained by the classifier training unit is a K-means classifier or a GMM classifier.
[0166] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.
Claims
1. A fusion feature extraction method, characterized in that: Includes steps: S1. Establish a communication system model, and obtain a signal matrix according to the communication system model; the communication system model includes a main user, a fusion center and M secondary users, the M secondary users all sense the main user, and the M secondary users are all connected to the fusion center; the method for obtaining the signal matrix X is: S11, the secondary user obtains the perception information as a sampling sequence, and then sends the sampling sequence to the fusion center. The length of the sampling sequence is N, and the expression of the sampling sequence is: Wherein, w(n) represents additive Gaussian white noise, s(n) represents the useful signal transmitted by the primary user, and h(n) represents channel fading; H0 represents that the state of the channel is unoccupied, and H1 represents that the state of the channel is occupied; S12. The fusion center uses the sampling sequence to construct a signal matrix. The expression of the signal matrix is: Among them, x m =[x m (1)x m (2)...x m (N)] is the sampling sequence of the mth secondary user; S2. Extracting the time domain features and frequency domain features of the signal matrix; the time domain features include the first time domain features and the second time domain features. The specific method for extracting the time domain features and frequency domain features of the signal matrix is: S21. Calculate the covariance matrix of the signal matrix. The expression of the covariance matrix is: Where X is the signal matrix and N is the length of the sampling sequence; S22, using the covariance matrix to calculate the first time domain feature and the second time domain feature, the expression of the first time domain feature is: The expression of the second time domain feature is: Among them, ∑r ij is the sum of all elements in the covariance matrix; S23, obtain the center frequency f of the sampling sequence in the signal matrix c , take f c A neighborhood (f c -f0,f c +f0) to construct the frequency domain feature. The expression of the frequency domain feature is: in, M represents the number of sensing nodes, F m (u) represents the Fourier transform sequence of the mth node sampling sequence, with a center frequency of f c The corresponding index in the Fourier transform sequence is F c ; 2f0 represents the spectrum bandwidth to be counted, and the spectrum bandwidth is less than or equal to the channel bandwidth; f s represents the sampling frequency, f s / N represents the spectrum resolution; 2F0+1 is an integer representing the index of the spectrum bandwidth in the Fourier transform sequence; the frequency domain feature is equivalent to the statistical reception signal at f c A neighborhood (f c -f0,f c +f0) the average amplitude of the amplitude spectrum; S3. Fusing the time domain features and the frequency domain features to obtain fused features. The fused features are obtained by using a vector concatenation method. The formula for feature fusion is: λ=(λ1,λ2,λ3) T Among them, λ is the fusion feature, λ1 is the first time domain feature, λ2 is the second time domain feature, and λ3 is the frequency domain feature.
2. A spectrum sensing method, characterized in that: The fusion features extracted by the fusion feature extraction method according to claim 1 include the steps of: S4, using the fusion features as training samples to train the classifier to obtain a spectrum sensing model; S5. Execute spectrum sensing tasks through the spectrum sensing model.
3. A spectrum sensing method according to claim 2, characterized in that: The classifier in step S4 is a K-means classifier or a GMM classifier.
4. A spectrum sensing method according to claim 3, characterized in that: The classifier in step S4 is a K-means classifier, and the specific method of training the classifier is: S41, obtaining S fusion features as training samples; S42, randomly selecting two training samples from the S training samples as initial samples, and initializing the initial samples to class 0 and class 1, the class center C0 of class 0 and the class center C1 of class 1 correspond to the centers of the two initial samples respectively; S43, for all training samples, calculating their Euclidean distances to the class center C0 and the class center C1, and dividing them into class 0 and class 1 according to the minimum Euclidean distance criterion; S44, recalculate the class center C0 of class 0 and the class center C1 of class 1; S45, if the class center C0 and the class center C1 have not changed or the preset number of iterations has been reached, then execute step S46; otherwise, update the class center C0 and the class center C1 and return to execute S43; S46, determine whether the class center C0 and the class center C1 meet ||C0||2>||C1||2, if yes, exchange the class center class center C0 and the class center C1, and then execute step S47, if no, directly execute step S47; S47 , saving the class center C0 and the class center C1 , and using the class center C0 and the class center C1 as the final spectrum sensing model.
5. A spectrum sensing method according to claim 4, characterized in that: Step S5 first sets the classification threshold T, and then executes the spectrum sensing task cyclically through the spectrum sensing model. The specific method is: S51, sampling to obtain a signal matrix X; S52, extracting the signal matrix X as the fusion feature λ; S53, calculating the Euclidean distances d0 and d1 from the fusion feature λ to the class center C0 and the class center C1; S54. If d1 / d0>T, the fused feature λ is classified as class 1, that is, the perception result is that the channel is occupied; otherwise, the fused feature λ is classified as class 0, that is, the channel is not occupied.
6. A spectrum sensing system, used to execute the spectrum sensing method according to claim 2, characterized in that: include: Signal matrix construction unit, feature extraction unit, feature fusion unit, classifier training unit, spectrum sensing unit; The signal matrix construction unit establishes a communication system model and obtains a signal matrix according to the communication system model; the feature extraction unit extracts time domain features and frequency domain features of the signal matrix; The feature fusion unit fuses the time domain features and the frequency domain features to obtain fused features; The classifier training unit uses the fusion features as training samples to train the classifier to obtain a spectrum sensing model; the spectrum sensing unit performs the spectrum sensing task through the spectrum sensing model.
7. A spectrum sensing system according to claim 6, characterized in that: The communication system model includes a primary user, a fusion center and M secondary users, the M secondary users all sense the primary user, and the M secondary users are all connected to the fusion center; the signal matrix construction unit obtains the signal matrix X by: S11, the secondary user obtains the perception information as a sampling sequence, and then sends the sampling sequence to the fusion center. The length of the sampling sequence is N, and the expression of the sampling sequence is: Among them, w(n) represents additive Gaussian white noise, s(n) represents the useful signal transmitted by the primary user, and h(n) represents channel fading; S12. The fusion center uses the sampling sequence to construct a signal matrix. The expression of the signal matrix is: Among them, x m =[x m (1)x m (2)...x m (N)] is the sampling sequence of the mth secondary user.
