An adaptive integrated feature selection method for modulation recognition of mixed signals

Abstract

The application relates to the technical field of modulation identification, in particular to a self-adaptive integrated feature selection mixed signal modulation identification method, which comprises the following steps: dividing a received signal into a plurality of small signals by using a sliding window; determining the position where the mixed signal occurs and the position where the mixed signal ends based on the power matrix of each small signal, and extracting the mixed signal; performing feature extraction on the mixed signal, and constructing a feature set of the mixed signal; selecting each feature in the feature set by using five feature selection methods, namely, a minimum redundancy maximum correlation criterion selection method, a multi-cluster feature selection method, a correlation feature selection method, a recursive feature elimination selection method and a random forest feature importance selection method, and preliminarily screening five stable feature subsets; performing self-adaptive optimization on the five stable feature subsets by using an HGSO algorithm to obtain an optimal feature subset; and performing modulation identification based on the optimal feature subset to obtain a modulation identification result. The method effectively improves the efficiency and accuracy of mixed signal modulation identification.

Description

TECHNICAL FIELD

[0001] Embodiments of the present application relate to the technical field of modulation recognition of communication signals, and particularly relate to a method for modulation recognition of mixed signals based on adaptive integrated feature selection. BACKGROUND

[0002] Modulation recognition of communication signals is a key technology in non-cooperative communication environment and has a wide range of applications in many fields. Most of the current modulation recognition algorithms are for the recognition of single-user modulation modes. However, with the vigorous development of the wireless communication industry, the rapid increase in the number of access devices and the demand for spectrum resources has caused the current electromagnetic environment to be highly dense and interlaced in time domain, frequency domain and space domain. Signals from different transmitters overlap in time-frequency domain. For example, two parties communicating using paired carrier multiple access (PCMA) technology send signals using the same frequency band, time slot and code word at the same time. A third listening party can only receive mixed signals in time-frequency domain. At this time, the modulation recognition algorithm for single signals cannot be directly applied to the modulation recognition of mixed signals. Therefore, how to design an efficient and low-complexity algorithm to correctly recognize the modulation mode of mixed signals has important engineering practical significance.

[0003] Li, Han and Zaerin et al. analyzed the high-order statistics and instantaneous features of mixed signals, manually selected features with high discrimination according to the calculated theoretical values, and designed a classifier using support vector machine (SVM). The most mixed double signal modulation modes that can be recognized are 21. Wu et al. performed fractional Fourier transform on mixed signals, constructed features according to the detail components of different layer wavelet decomposition, and set appropriate threshold values to recognize the modulation modes of different mixed signals using decision tree. However, the feature calculation complexity of this method is relatively high. Zhang et al. considered the different power ratios between mixed signal components and the case of small sample, added attention layer and LSTM layer based on capsule neural network, so as to better capture the time-frequency characteristics and hierarchical structure of the signal, and achieved higher recognition accuracy under low signal-to-noise ratio. However, the model constructed by this method has high complexity.

[0004] In summary, the inventors of the present application found that the current mixed signal modulation recognition methods still have the following defects and deficiencies.

[0005] First, the mixed signal based on traditional feature extraction analysis is single, and the mixed pattern is relatively fixed. Most methods analyze two mixed signal components, and the maximum number of mixed signal components is three. In addition, the mixed pattern mainly explores complete mixing in the time domain and frequency domain, only a small part considers partial mixing in the frequency domain, and almost no analysis is conducted on partial mixing in the time domain.

[0006] Second, the classifier designed by the recognition method based on traditional feature extraction has poor universality, and there are actually many factors that affect the modulation mode features of mixed signals, including the modulation mode types of component signals, mixed patterns, power ratios between component signals, carrier frequency offsets, phase offsets, and signal-to-noise ratios. Most of the methods proposed at present only consider the influence of a single factor, and the designed classifier can only identify mixed signals of a fixed form.

[0007] Third, the mixed signal analyzed by the recognition method based on deep learning lacks the application of expert knowledge, and most of them consider converting the modulation recognition problem of mixed signals to the image recognition field. SUMMARY

[0008] To solve the above technical problems, the embodiments of the present application propose a mixed signal modulation recognition method based on adaptive integrated feature selection, which can automatically locate the positions where the mixed signal occurs and ends in the time domain, is suitable for modulation mode recognition of time domain mixed signals with multiple mixed patterns and multiple factors, and effectively improves the efficiency and accuracy of mixed signal modulation recognition.

[0009] To achieve the above purpose, the embodiments of the present application propose a mixed signal modulation recognition method based on adaptive integrated feature selection, which comprises the following steps: dividing a received signal into a plurality of small signals by using a sliding window, determining the positions where the received signal occurs and ends based on the power matrix of each small signal, and extracting a mixed signal; performing feature extraction on the mixed signal, constructing a feature set of the mixed signal, wherein the features in the feature set of the mixed signal include 10 combined high-order moment features, 16 combined high-order cumulant features, 8 high-order moment features, 8 high-order cumulant features, and 4 instantaneous features; selecting each feature in the feature set by using five feature selection methods including minimum redundancy maximum correlation criterion selection, multi-cluster feature selection, relevant feature selection, recursive feature elimination selection, and random forest feature importance selection, and preliminarily screening five stable feature subsets; performing adaptive optimization on the five stable feature subsets by using an HGSO algorithm to obtain an optimal feature subset; and performing modulation recognition based on the optimal feature subset to obtain a modulation recognition result.

[0010] To achieve the above object, the embodiment of the present application further provides a mixed signal modulation recognition system based on adaptive integrated feature selection, which comprises a sliding window positioning module, which is used for segmenting a received signal into a plurality of small signals by using a sliding window, determining the position where the mixed signal occurs and the position where the mixed signal ends based on a power matrix of each small signal, and extracting the mixed signal; a feature set construction module, which is used for extracting features of the mixed signal and constructing a feature set of the mixed signal, wherein the features in the feature set of the mixed signal comprise 10 combined high-order moment features, 16 combined high-order cumulant features, 8 high-order moment features, 8 high-order cumulant features and 4 instantaneous features; an integrated feature selection module, which is used for selecting each feature in the feature set by using five feature selection methods, i.e., a minimum redundancy maximum relevance criterion selection, a multi-cluster feature selection, a relevant feature selection, a recursive feature elimination selection and a random forest feature importance selection, and preliminarily screening five stable feature subsets; an adaptive optimization module, which is used for adaptively optimizing the five stable feature subsets by using an HGSO algorithm to obtain an optimal feature subset; and a modulation recognition module, which is used for performing modulation recognition based on the optimal feature subset and obtaining a modulation recognition result.

[0011] To achieve the above object, the embodiment of the present application further provides an electronic device, which comprises at least one processor and a memory connected with the at least one processor in communication, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the mixed signal modulation recognition method based on adaptive integrated feature selection.

[0012] To achieve the above object, the embodiment of the present application further provides a computer readable storage medium, which stores a computer program, and the computer program is executable by a processor to implement the mixed signal modulation recognition method based on adaptive integrated feature selection.

[0013] This application proposes an adaptive integrated feature selection method for aliasing signal modulation identification. First, a sliding window is used to accurately locate the aliasing occurrence and termination points of the received signal, thereby accurately extracting the aliased signal. Next, a feature set containing 46 different features is constructed for the aliased signal. Higher-order moment features and higher-order cumulants are used to improve discriminative power, and combined higher-order moment features and combined higher-order cumulants are used to eliminate the influence of power factor and phase jitter on higher-order moments and higher-order cumulants. Then, considering the advantages of three feature selection methods—filtering, wrapping, and embedded—five feature selection methods are employed: minimum redundancy maximum correlation criterion selection, multi-cluster feature selection, correlation feature selection, recursive feature elimination selection, and random forest feature importance selection. The HGSO algorithm is used for adaptive optimization to obtain the optimal feature subset. Finally, modulation identification is performed based on the optimal feature subset to obtain the modulation identification result. This aliasing signal modulation identification method, designed in this way, can quickly and accurately identify aliasing signal modulation when factors such as the type of modulation method, aliasing pattern, power ratio between component signals, carrier frequency offset, phase offset, and signal-to-noise ratio of the aliasing signal change simultaneously. It has high stability and universality.

[0014] Optionally, the received signal is divided into several small signals using a sliding window. Based on the power matrix of each small signal, the positions where aliasing occurs and where aliasing ends are determined, and the aliased signal is extracted, including:

[0015] The window size and sliding step size of the sliding window are set according to the length of the received signal, thereby dividing the received signal into n parts using the set sliding window. win A small signal;

[0016] Calculate the power matrix of each small signal separately, and then calculate the power matrix of the number of small signals n. win Reduced to a preset fixed value n fixed , every n win / n fixed Taking the upper quantile of each small signal, a new power matrix is ​​obtained;

[0017] Unsupervised clustering was performed on the new power matrix using a spectral density clustering algorithm;

[0018] Determine the location of the category change, i.e., the location where the power changes in the new power matrix, Loc_p. Based on Loc_p, deduce the original window position, Loc_win, to obtain the location where aliasing occurs and the location where aliasing ends in the received signal.

[0019] Based on the location where aliasing occurs and the location where aliasing ends, the aliasing signal is extracted from the received signal.

[0020] Optionally, the eight higher-order moment features are m 20, m 40 , m 60 , m 80 , m 21 , m 42 , m 63 and m 84 , 8 high-order cumulant features are c 20 , c 21 , c 40 , c 41 , c 42 , c 60 , c 63 and c 8o ;

[0021] The calculation formulas of the high-order moment features and the high-order cumulant features are respectively:

[0022] m kx (τ1,…,τ k-1 )=E{x(t)x(t+τ1)…x(t+τ k-1 )};

[0023] c kx (τ1,…,τ k-1 )=cum[x(t),x(t+τ1),…,x(t+τ k-1 )];

[0024] Wherein, x(t) represents a stationary random process, k represents an order, τ represents a time delay, and cum(·) represents a cumulant function.

[0025] Optionally, the combined high-order moment features are denoted as F moma , a=1,2,…,10, and the combined high-order cumulant features are denoted as F cumb , b=1,2,…,16.

[0026] The calculation formulas of the 10 combined high-order moment features are respectively:

[0027]

[0028] The calculation formulas of the 16 combined high-order cumulant features are respectively:

[0029] F cum1 =|c 40 | / |c 41 |; F cum2 =|c 40 | / |c 42 |; F cum3 =|c 41 | / |c 42 |;

[0030]

[0031] Optionally, the four instantaneous features are respectively the second-order statistical feature m a , the maximum value feature γ max of the zero-centered normalized non-weak signal instantaneous amplitude, the standard deviation feature σ da of the zero-centered normalized instantaneous amplitude absolute value, and the standard deviation feature σ aa of the zero-centered normalized non-weak signal instantaneous amplitude.

[0032] The calculation formulas of the four instantaneous features are respectively:

[0033]

[0034] γ max = max |DFT[A cn (i)]| / N; 2

[0035]

[0036] wherein n is the number of sampling points, A(i) is the instantaneous amplitude of the signal, A cn (i)=A(i)-1, A cn (i) is the zero-centered normalized instantaneous amplitude of the signal, DFT(·) represents the discrete Fourier transform, a i is a preset threshold of the non-weak signal, and C is the number of non-weak signals.

[0037] Optionally, the stable feature subset S mRMR is screened by using a minimum redundancy maximum correlation criterion selection method, the stable feature subset S MCFS is screened by using a multi-cluster feature selection method, the stable feature subset S CBFS is screened by using a correlation feature selection method, the stable feature subset S RFE is screened by using a recursive feature elimination selection method, and the stable feature subset S RT is screened by using a random forest feature importance selection method.

[0038] After the five stable feature subsets are preliminarily screened, the method further comprises:

[0039] The validation accuracies of the five stable feature subsets are determined by using a random forest algorithm, the validation accuracy of S mRMR is Acc mRMR , the validation accuracy of S MCFS is Acc MCFS , the validation accuracy of S CBFS is Acc CBFS , the validation accuracy of S RFE ​The verification accuracy of S RFE is recorded as Acc RT . The verification accuracy of S RT is recorded as Acc mRMR .

[0040] Optionally, before the 5 stable feature subsets are adaptively optimized by the HGSO algorithm to obtain the optimal feature subset, the method further comprises:

[0041] Based on the verification accuracy of each stable feature subset, an initial weight is assigned to each stable feature subset, S mRMR The initial weight of S MCFS is recorded as w MCFS The initial weight of S CBFS is recorded as w CBFS The initial weight of S RFE is recorded as w RFE The initial weight of S RT is recorded as w RT ;

[0042] The calculation formula of the initial weight of each stable feature subset is respectively:

[0043] w mRMR = Acc mRMR / (Acc mRMR +Acc MCFS +Acc CBFS +Acc RFE +Acc RT );

[0044] w MCFS = Acc MCFS / (Acc mRmR +Acc MCFS +Acc CBFS +Acc RFE +Acc RT );

[0045] w CBFS = Acc CBFS / (Acc mRMR +Acc MCFS +Acc CBFS +Acc RFE +Acc RT );

[0046] w RFE = Acc RFE / (Acc mRMR +Acc MCFS +Acc CBFS +Acc RFE +AccRT );

[0047] w RT = Acc RT / (Acc mRMR + Acc MCFS + Acc CBFS + Acc RFE + Acc RT )。 BRIEF DESCRIPTION OF DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the drawings needed to be used in the description of the embodiments of the present application or the related art will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort on the basis of these drawings.

[0049] Figure 1 is a flowchart of a self-adaptive integrated feature selection method for mixed signal modulation recognition provided in an embodiment of the present application;

[0050] Figure 2 is a schematic diagram of a mixed scenario provided in an embodiment of the present application;

[0051] Figure 3 is a design diagram of a self-adaptive integrated feature selection algorithm provided in an embodiment of the present application;

[0052] Figure 4 is a schematic diagram of the recognition accuracy of 28 mixed signals using the method provided in the present application varying with the signal-to-noise ratio provided in an embodiment of the present application;

[0053] Figure 5 is a schematic diagram of the average recognition accuracy of the self-adaptive integrated feature selection algorithm for 10 different component signal power ratios provided in an embodiment of the present application;

[0054] Figure 6 is a schematic diagram of the average recognition accuracy of the self-adaptive integrated feature selection algorithm for 10 different component signal power ratios compared with the commonly used algorithm provided in an embodiment of the present application;

[0055] Figure 7 is a schematic diagram of the recognition accuracy of the method provided in the present application when the number of mixed signal components varies provided in an embodiment of the present application;

[0056] Figure 8 is a structural schematic diagram of a self-adaptive integrated feature selection mixed signal modulation recognition system provided in another embodiment of the present application;

[0057] Figure 9 This is a schematic diagram of the structure of an electronic device provided in another embodiment of this application. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the various embodiments of this application will be described in detail below with reference to the accompanying drawings. In the various embodiments of this application, many technical details are presented to enable the reader to better understand this application. However, even without these technical details and various variations and modifications based on the following embodiments, the technical solutions claimed in this application can be implemented. The division of the following embodiments is only for convenience of description and should not constitute any limitation on the specific implementation of this application. The various embodiments can be combined with and referenced by each other without contradiction.

[0059] To more clearly illustrate the technical solution proposed in this application, the relevant content of aliasing signals will be introduced first.

[0060] Generally speaking, when there is only one signal in a single channel, the signal is a single-carrier signal, and its mathematical model is as follows:

[0061]

[0062] Where, N v P is the number of symbols. i For the power of the independently modulated signal, a i (n) is the information symbol sequence, q i (t) is the shaping pulse, T si For the symbol period, f ci φ is the carrier frequency of the modulated signal. i For the initial phase of the carrier, ξT si This is for timing error.

[0063] Assuming the bandwidth f allowed by a single-channel receiver re Within a range, during a reception time T re Within a given range, there are R independent narrowband signals falling within the receiver's bandwidth. These signals are statistically independent of each other, and the duration of each component signal varies throughout the entire reception time, but all are fully received by the receiver. Furthermore, noise contamination is also present. Therefore, the mathematical model for a single-channel time-frequency aliasing signal is:

[0064]

[0065] Among them, s i (t,T si Let T be the i-th component signal in the aliased signal. sidenotes the duration of each component signal, n(t) is stationary additive white Gaussian noise, and each component signal in the aliasing signal and each signal component in the aliasing signal and the noise are mutually statistically independent.

[0066] In order to more comprehensively and intuitively measure the aliasing of the aliasing signal, the influence of aliasing in the time domain, the frequency domain and the spatial domain is comprehensively considered from the perspective of the component signal, wherein the spatial domain is embodied by the power variation, and the time domain aliasing degree, the frequency domain aliasing degree and the power ratio of the aliasing signal are defined.

[0067] First, the time domain aliasing degree of the signal is defined, assuming that the aliasing signal composed of R component signals has a duration of T re , and each component signal has a duration of T i (0 < T i ≤ T re ), T ij denotes the duration of aliasing of the i-th signal and the j-th signal (T ij = T ji ), and the time domain aliasing ratio P t of the R component signals is:

[0068]

[0069] For the definition of the frequency domain aliasing ratio, it is assumed that the bandwidth of each component signal is f i (0 < f i < f re ), f ij denotes the frequency band of aliasing of the i-th signal and the j-th signal (f ij = f ji ), and the frequency domain aliasing ratio P f is:

[0070]

[0071] The definition of the power ratio of the aliasing signal is relatively simple, the power of each component signal is denoted as E i , and it is assumed that the aliasing signal is composed of component signal 1 and component signal 2, and the power ratio between the component signal 1 and the component signal 2 is E1:E2.

[0072] One embodiment of the present application provides an aliasing signal modulation identification method based on adaptive integrated feature selection, which is applied to an electronic device, wherein the electronic device can be a terminal or a server, and the electronic device in the present embodiment and each of the following embodiments is taken as an example to illustrate a server. The implementation details of the aliasing signal modulation identification method based on adaptive integrated feature selection provided in the present embodiment are described in detail below. The following content only provides implementation details for easy understanding, and is not necessary for implementing the present solution.

[0073] The specific flow of the adaptive integrated feature selection and mixed signal modulation recognition method proposed in this embodiment can be as shown in Figure 1 The specific flow of the adaptive integrated feature selection and mixed signal modulation recognition method proposed in this embodiment can be as shown in

[0074] In step 101, the received signal is segmented into a plurality of small signals by using a sliding window. Based on the power matrix of each small signal, the position where the mixed signal occurs and the position where the mixed signal ends are determined, and the mixed signal is extracted.

[0075] In a specific implementation, as shown in Figure 2 The server needs to extract the mixed signal from the received signal, which needs to be achieved by using a sliding window. The server segments the received signal into a plurality of small signals by using a sliding window. Based on the power matrix of each small signal, the position where the mixed signal occurs and the position where the mixed signal ends are determined, and the mixed signal is extracted.

[0076] In one example, the server first needs to set a proper window size win_length and a sliding step step_length according to the length Signal_length of the received signal. Then, the received signal is segmented into N win small signals by using the set sliding window. Subsequently, the power matrix of each small signal is calculated. The number N win of small signals is reduced to a preset fixed value N fixed , and the upper quantile is taken for every N win / N fixed small signals to obtain a new power matrix. Next, the server needs to perform unsupervised clustering on the new power matrix by using a spectral density clustering algorithm. Finally, the position where the class changes, i.e., the position Loc_p where the power in the new power matrix changes, is determined. The position Loc_p is backtracked to the position Loc_win of the original window, so as to obtain the position where the mixed signal occurs and the position where the mixed signal ends. Then, the mixed signal is extracted from the received signal according to the position where the mixed signal occurs and the position where the mixed signal ends.

[0077] In step 102, feature extraction is performed on the mixed signal, and a feature set of the mixed signal is constructed. The features in the feature set of the mixed signal include 10 combined high-order moment features, 16 combined high-order cumulant features, 8 high-order moment features, 8 high-order cumulant features, and 4 instantaneous features.

[0078] In a specific implementation, the instantaneous characteristic statistic of the aliasing signal is essentially the second-order statistical characteristic of the characteristics of the instantaneous amplitude, phase and frequency of the analysis signal in a period of time. The existence of the equivalence and multiplicity characteristics makes it impossible to identify the minimum phase system, and it is sensitive to noise in actual communication, so that the discrimination degree of the characteristics is reduced under non-ideal conditions. In order to overcome the influence of these factors, higher order statistics, collectively referred to as higher order statistics, must be selected. The most commonly used higher order statistics are higher order moments and higher order cumulants. Based on this, the server extracts features of the aliasing signal and constructs a feature set of the aliasing signal. The features in the feature set of the aliasing signal include 10 combined higher order moment features, 16 combined higher order cumulant features, 8 higher order moment features, 8 higher order cumulant features and 4 instantaneous features. It should be noted that the power factor and phase jitter will have an impact on the higher order moments and higher order cumulants, so in this embodiment, different orders of higher order moments and higher order cumulants are used in the form of ratio to construct new combined higher order moments and combined higher order cumulants, thereby eliminating the adverse effects of power factor and phase jitter.

[0079] In one example, the selected 8 higher order moment features are m 20 , m 40 , m 60 , m 80 , m 21 , m 42 , m 63 and m 84 , and the 8 higher order cumulant features are c 20 , c 21 , c 40 , c 41 , c 42 , c 60 , c 63 and c 80 . The calculation formulas of the higher order moment features and the higher order cumulant features are respectively:

[0080] m kx (τ1,…,τ k-1 )=E{x(t)x(t+τ1)…x(t+τ k-1 )};

[0081] c kx (τ1,…,τ k-1 )=cum[x(t),x(t+τ1),…,x(t+τ k-1 )];

[0082] wherein x(t) represents a stationary random process, k represents the order, τ represents the time delay, and cum(·) represents the cumulant function.

[0083] In one example, the combined higher order moment features are denoted as Fmoma , a = 1, 2, …, 10, the combined high-order cumulant features are denoted as F cumb , b = 1, 2, …, 16.

[0084] The calculation formulas of the 10 combined high-order moment features are respectively:

[0085]

[0086] The calculation formulas of the 16 combined high-order cumulant features are respectively:

[0087] F cum1 = |c 40 | / |c 41 |; F cum2 = |c 40 | / |c 42 |; F cum3 = |c 41 | / |c 42 |;

[0088]

[0089] In an example, the 4 instantaneous features are respectively the second-order statistics of instantaneous amplitude feature m a , the maximum value of instantaneous amplitude spectrum density feature γ max , the standard deviation of zero-centered normalized non-weak signal instantaneous amplitude feature σ da , and the standard deviation of zero-centered normalized instantaneous amplitude absolute value feature σ aa .

[0090] The calculation formulas of the 4 instantaneous features are respectively:

[0091]

[0092] γ max = max |DFT[A cn (i)]| 2 / N;

[0093]

[0094] Wherein, N is the number of sampling points, A(i) is the instantaneous amplitude of the signal, A cn (i) = A(i) - 1, A cn (i) is the zero-centered normalized instantaneous amplitude of the signal, DFT(·) represents the discrete Fourier transform, a i is a preset threshold of non-weak signal, and C is the number of non-weak signals.

[0095] Step 103, using 5 feature selection methods of minimal-Redundancy-Maximal-Relevance criterion selection, multi-cluster feature selection, correlation based feature selection, recursive feature elimination selection and random forest feature importance selection, each feature in the feature set is selected, and 5 stable feature subsets are obtained by preliminary screening.

[0096] There is a conversion relationship between the high-order moment and the high-order cumulant of the signal, so the M feature characteristic set I feature There is a high correlation and redundancy. In addition, when the power ratio of the component signals of the aliasing signal changes, it is equivalent to a function disturbance to the original feature set, so the target classification task c feature of the aliasing signal modulation mode needs a feature subset S feature with high discrimination and high stability.

[0097] Based on this, the embodiment integrates three types of feature selection methods of filtering, wrapping and embedding, uses 5 feature selection methods of minimal-Redundancy-Maximal-Relevance criterion selection (mRMR), multi-cluster feature selection (MCFS), correlation based feature selection (CBFS), recursive feature elimination selection (RFE) and random forest feature importance selection (RFE) to select each feature in the feature set, and 5 stable feature subsets are obtained by preliminary screening. The design of the adaptive integrated feature selection algorithm is shown in Figure 3

[0098] In one example, the stable feature subset obtained by using the minimal-Redundancy-Maximal-Relevance criterion selection method is denoted as S mRMR , the stable feature subset obtained by using the multi-cluster feature selection method is denoted as S MCFS , the stable feature subset obtained by using the correlation based feature selection method is denoted as S CBFS , the stable feature subset obtained by using the recursive feature elimination selection method is denoted as S RFE , and the stable feature subset obtained by using the random forest feature importance selection method is denoted as S RT .

[0099] ​In an example, after the server preliminarily screens the five stable feature subsets, the server also needs to respectively determine the validation accuracies of the five stable feature subsets by using the random forest algorithm, wherein the validation accuracy of S mRMR is denoted as Acc mRMR , the validation accuracy of S MCFS is denoted as Acc MCFS , the validation accuracy of S CBFS is denoted as Acc CBFS , the validation accuracy of S RFE is denoted as Acc RFE , and the validation accuracy of S RT is denoted as Acc RT .

[0100] In step 104, the HGSO algorithm is used to perform adaptive optimization on the five stable feature subsets to obtain an optimal feature subset.

[0101] In a specific implementation, the five stable feature subsets still have relevance and redundancy. Therefore, in order to further perform adaptive optimization on the ranked features to obtain a feature subset with optimal classification effect and minimum features, it is necessary to search the reordered feature set. Compared with global optimal search, heuristic search is more efficient. In this embodiment, the meta-heuristic Henry Gas Solubility Optimization (HGSO) algorithm is used to perform adaptive optimization on the five stable feature subsets to obtain an optimal feature subset.

[0102] When the Henry Gas Solubility Optimization algorithm is used for feature selection, different positions of the gas in the feature space represent different feature subsets. The main idea is to regard the feature set as a given liquid, and different feature subsets represent the partial pressure of the gas. By exploring the solubility of the gas in different feature spaces, the selection of the optimal feature subset is achieved.

[0103] In an example, in order to improve the stability and generalization ability of the extracted features, before the server performs adaptive optimization on the five stable feature subsets by using the HGSO algorithm to obtain an optimal feature subset, the server also needs to assign different initial weights to the stable feature subsets based on the validation accuracies of the stable feature subsets. Wherein the initial weight of S mRMR is denoted as w mRMR , the initial weight of S MCFS is denoted as w MCFS , the initial weight of S CBFS is denoted as w CBFS , the initial weight of S RFE is denoted as w RFE , and the initial weight of S RT is denoted as w RT .

[0104] In one example, the initial weight of each stable feature subset is calculated according to the following formula respectively:

[0105] w mRMR = Acc mRMR / (Acc mRMR +Acc MCFS +Acc CBFS +Acc RFE +Acc RT );

[0106] w MCFS = Acc MCFS / (Acc mRMR +Acc MCFS +Acc CBFS +Acc RFE +Acc RT );

[0107] w CBFS = Acc CBFS / (Acc mRMR +Acc MCFS +Acc CBFS +Acc RFE +Acc RT );

[0108] w RFE = Acc RFE / (v mRMR +Acc MCFS +Acc CBFS +Acc RFE +Acc RT );

[0109] w RT = Acc RT / (Acc mRMR +Acc MCFS +Acc CBFS +Acc RFE +Acc RT )。

[0110] Step 105, modulation recognition is performed based on the optimal feature subset, and a modulation recognition result is obtained.

[0111] The adaptive integrated feature selection method for modulated signal recognition in this embodiment firstly uses a sliding window to accurately find the position where the received signal starts to overlap and the position where the received signal ends to overlap, thereby accurately extracting the overlapped signal. Next, a feature set containing 46 different features is constructed for the overlapped signal, high-order moment features and high-order cumulant features are used to improve the discrimination, and combined high-order moment features and combined high-order cumulant features are used to eliminate the influence of power factor and phase jitter on high-order moment and high-order cumulant. Subsequently, the advantages of three feature selection methods, i.e., filtering, wrapping and embedding, are comprehensively considered, five feature selection methods, i.e., minimum redundancy maximum relevance criterion selection, multi-cluster feature selection, correlation feature selection, recursive feature elimination selection and random forest feature importance selection, are used for feature selection, and the HGSO algorithm is used for adaptive optimization to obtain the optimal feature subset. Finally, modulation recognition is performed based on the optimal feature subset, and the modulation recognition result is obtained. The modulated signal recognition method designed in this way can quickly and accurately realize modulated signal recognition when the modulation mode, the overlapping style, the power ratio between component signals, the carrier frequency offset, the phase offset and the signal-to-noise ratio of the overlapped signal change simultaneously, and has high stability and universality.

[0112] The step division of the above methods is only for the purpose of clear description. In implementation, one step can be combined or some steps can be split and decomposed into multiple steps, as long as the same logical relationship is included, which is within the protection scope of the present application. Adding insignificant modifications or introducing insignificant designs in the algorithm or process, but not changing the core design of the algorithm and process, are within the protection scope of the present application.

[0113] In one embodiment, in order to verify the effectiveness of the adaptive integrated feature selection method for modulated signal recognition proposed in the present application, relevant simulation experiments are performed. Figure 4 The schematic diagram of the recognition accuracy of 28 overlapped signals after using the method proposed in the present application is shown. With the improvement of the signal-to-noise ratio, the recognition accuracy is significantly improved. Figure 5 The average recognition accuracy of the adaptive integrated feature selection algorithm for 10 different component signal power ratios is shown. With the improvement of the signal-to-noise ratio, the recognition accuracy is significantly improved. Figure 6 The comparison of the average recognition accuracy of the adaptive integrated feature selection algorithm for 10 different component signal power ratios with that of the commonly used algorithm is shown. The random forest algorithm has the best comprehensive recognition rate. Figure 7 The recognition accuracy of the method proposed in the present application when the number of overlapped signal components changes is shown. When the signal-to-noise ratio is 10 dB, the modulation recognition accuracy for the overlapped signals with 3 and 4 component signals can reach 85% and 90%, respectively.

[0114] Another embodiment of the present application provides a mixed signal modulation recognition system with adaptive integrated feature selection. The following describes the implementation details of the mixed signal modulation recognition system with adaptive integrated feature selection provided by the embodiment. The following implementation details are provided for the convenience of understanding and are not necessary for implementing the embodiment.

[0115] Figure 8 FIG. 1 is a structural diagram of the mixed signal modulation recognition system with adaptive integrated feature selection provided by the embodiment. The system includes a sliding window positioning module 201, a feature set construction module 202, an integrated feature selection module 203, an adaptive optimization module 204, and a modulation recognition module 205.

[0116] The sliding window positioning module 201 is configured to divide the received signal into a plurality of small signals by using a sliding window, determine the position where the mixing occurs and the position where the mixing ends based on the power matrix of each small signal, and extract the mixed signal.

[0117] The feature set construction module 202 is configured to perform feature extraction on the mixed signal and construct a feature set of the mixed signal. The features in the feature set of the mixed signal include 10 combined high-order moment features, 16 combined high-order cumulant features, 8 high-order moment features, 8 high-order cumulant features, and 4 instantaneous features.

[0118] The integrated feature selection module 203 is configured to select each feature in the feature set by using 5 feature selection methods, including the minimum redundancy maximum correlation criterion selection, the multi-cluster feature selection, the relevant feature selection, the recursive feature elimination selection, and the random forest feature importance selection, and preliminarily screen 5 stable feature subsets.

[0119] The adaptive optimization module 204 is configured to perform adaptive optimization on the 5 stable feature subsets by using the HGSO algorithm to obtain an optimal feature subset.

[0120] The modulation recognition module 205 is configured to perform modulation recognition based on the optimal feature subset to obtain a modulation recognition result.

[0121] It is worth mentioning that each module involved in the embodiment is a logical module. In actual application, one logical unit can be one physical unit, a part of one physical unit, or a combination of multiple physical units. In addition, in order to highlight the innovative part of the present application, units not closely related to solving the technical problems provided by the present application are not introduced in the embodiment, but this does not mean that there are no other units in the embodiment.

[0122] It can be found that the embodiment is a system embodiment corresponding to the method embodiment described above, and the embodiment can be implemented in cooperation with the method embodiment described above. The related technical details and technical effects mentioned in each of the above embodiments are still valid in this embodiment. In order to reduce repetition, they will not be described here. Accordingly, the related technical details mentioned in this embodiment can also be applied to the above embodiments.

[0123] Another embodiment of the present application provides an electronic device, which has a specific structure as shown in Figure 9 The electronic device includes at least one processor 301 and a memory 302 connected with the at least one processor 301, wherein the memory 302 stores instructions executable by the at least one processor 301, and the instructions are executed by the at least one processor 301 to enable the at least one processor 301 to perform the adaptive integrated feature selection and aliasing signal modulation identification method as described in the above method embodiments.

[0124] The memory and the processor can be connected in a bus manner, and the bus can include any number of interconnected buses and bridges. The bus connects various circuits of one or more processors and memories together. The bus can also connect various other circuits such as peripheral devices, voltage stabilizers, and power management circuits together, which are well known in the art and will not be further described herein. The bus interface is responsible for providing an interface between the bus and the transceiver. The transceiver can be one element or multiple elements such as multiple receivers and transmitters, which provide a unit for communicating with various other devices on the transmission medium. The data processed by the processor is transmitted on the wireless medium through the antenna, and further, the antenna also receives data and transmits the data to the processor.

[0125] The processor is responsible for managing the bus and general processing, and can also provide various functions including timing, peripheral interface, voltage regulation, power management, and other control functions. The memory can be used to store data used by the processor during operation.

[0126] Another embodiment of the present application provides a computer readable storage medium storing a computer program, wherein the computer program is executed by a processor to implement the adaptive integrated feature selection and aliasing signal modulation identification method as described in the above method embodiments.

[0127] That is, a person skilled in the art can understand that all or part of the steps in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a program stored in a storage medium, including a plurality of instructions for causing a device (such as a single-chip microcomputer, a chip, etc.) or a processor to execute all or part of the steps of the method described in various embodiments of the present application. The aforementioned storage medium includes a U disk, a mobile hard disk, a ROM (Read-Only Memory), a RAM (Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0128] A person of ordinary skill in the art can understand that the above-mentioned embodiments are specific embodiments for implementing the present application, and in actual applications, various changes can be made in form and details without departing from the spirit and scope of the present application.

Claims

1. A method of mixed signal modulation recognition with adaptive integrated feature selection, characterized in that, The method comprises the following steps: The received signal is divided into a plurality of small signals by using a sliding window, and the positions of the beginning and end of aliasing of the received signal are determined based on the power matrix of each small signal, and the aliasing signal is extracted; Feature extraction is performed on the aliasing signal to construct a feature set of the aliasing signal, and the features in the feature set of the aliasing signal include 10 combined higher-order moment features, 16 combined higher-order cumulant features, 8 higher-order moment features, 8 higher-order cumulant features and 4 instantaneous features; Five feature selection methods, including minimum redundancy maximum correlation criterion selection, multi-cluster feature selection, relevant feature selection, recursive feature elimination selection and random forest feature importance selection, are used to select the features in the feature set, and five stable feature subsets are obtained through preliminary screening; The HGSO algorithm is used to adaptively optimize the five stable feature subsets to obtain an optimal feature subset; Based on the optimal feature subset, modulation recognition is performed to obtain a modulation recognition result.

2. The method of claim 1, wherein, The received signal is divided into a plurality of small signals by using a sliding window, and the positions of the beginning and end of aliasing of the received signal are determined based on the power matrix of each small signal, and the aliasing signal is extracted, comprising the following steps: The window size and sliding step of the sliding window are set according to the length of the received signal, so that the received signal is divided into N win small signals by using the set sliding window. Calculate the power matrix of each small signal respectively, and the number of small signals N win Reduce to a preset fixed value N fixed , and take the upper quantile of N win / N fixed small signals to obtain a new power matrix; Spectral density clustering algorithm is used for unsupervised clustering of the new power matrix; The positions of the class changes, i.e., the positions of the power changes in the new power matrix Loc_p, are determined, and the positions of the beginning and end of aliasing of the original window are deduced from Loc_p, so as to obtain the positions of the beginning and end of aliasing of the received signal; According to the positions of the beginning and end of aliasing, the aliasing signal is extracted from the received signal.

3. The method of claim 1, wherein, 8 high order moment features are m 20 , m 40 , m 60 , m 80 , m 21 , m 42 , m 63 and m 84 , 8 high order cumulant features are c 20 , c 21 , c 40 , c 41 , c 42 , c 60 , c 63 and c 80 ; The calculation formulas of the higher-order moment features and the higher-order cumulant features are as follows: m kx (τ1,…,τ k-1 )=E{x(t)x(t+τ1)…x(t+τ k-1 )}; c kx (τ1,…,τ k-1 )=cum[x(t),x(t+τ1),…,x(t+τ k-1 )]; Wherein, x(t) represents a stationary random process, k represents an order, τ represents a time delay, and cum(·) represents a cumulative function.

4. The method of claim 3, wherein, The combined higher order moment feature is denoted as F moma , a = 1, 2,..., 10, the combined higher order cumulant feature is denoted as F cumb , b = 1, 2,..., 16; The calculation formulas of the 10 combined higher-order moment features are as follows: The calculation formulas of the 16 combined higher-order cumulant features are as follows: F cum1 = |c 40 | / |c 41 |;F cum2 = |c 40 | / |c 42 |;F cum3 = |c 41 | / |c 42 |; 5. The method of claim 4, wherein, 4 instantaneous features are the second order statistics of instantaneous amplitude feature m a , the maximum of the spectrum density of instantaneous amplitude feature γ max , the standard deviation of zero-centered normalized non-weak signal instantaneous amplitude feature σ da and the standard deviation of zero-centered normalized instantaneous amplitude absolute value feature σ aa ; The calculation formulas of the 4 instantaneous features are as follows: gamma max = max | DFT [A cn (i)]| 2 / N; where N is the number of sampling points, A(i) is the instantaneous amplitude of the signal, A cn (i) = A(i) - 1, A cn (i) is the zero-centered normalized instantaneous amplitude of the signal, DFT(·) denotes the discrete Fourier transform, a i is a preset threshold of non-weak signals, and C is the number of non-weak signals.

6. The method of claim 1 to 5, wherein, The stable feature subset is screened by using a minimum redundancy maximum correlation criterion selection method, and is recorded as S mRMR The stable feature subset is screened by using a multi-cluster feature selection method, and is recorded as S MCFS The stable feature subset is screened by using a correlation feature selection method, and is recorded as S CBFS The stable feature subset is screened by using a recursive feature elimination selection method, and is recorded as S RFE The stable feature subset is screened by using a random forest feature importance selection method, and is recorded as S RT ; After the five stable feature subsets are obtained through preliminary screening, the method further comprises the following steps: The validation accuracy of each of the five stable feature subsets determined using the random forest algorithm is recorded as Acc mRMR , S mRMR The validation accuracy of each of the five stable feature subsets determined using the random forest algorithm is recorded as Acc MCFS , S MCFS The validation accuracy of each of the five stable feature subsets determined using the random forest algorithm is recorded as Acc CBFS , S CBFS The validation accuracy of each of the five stable feature subsets determined using the random forest algorithm is recorded as Acc RFE , S RFE The validation accuracy of each of the five stable feature subsets determined using the random forest algorithm is recorded as Acc RT , S RT .

7. The method of claim 6, wherein the method is a method of identifying a mixed signal modulation by adaptive integrated feature selection, characterized in that, Before the HGSO algorithm is used to adaptively optimize the five stable feature subsets to obtain an optimal feature subset, the method further comprises the following steps: Based on the validation accuracy of each stable feature subset, initial weights, S, are assigned to each stable feature subset. mRMR The initial weights are denoted as w. mRMR S MCFS The initial weights are denoted as w. MCFS S CBFS The initial weights are denoted as w. CBFS S RFE The initial weights are denoted as w. RFE S RT The initial weights are denoted as w. RT ; The calculation formulas of the initial weights of the stable feature subsets are as follows: w mRMR = Acc mRMR / (Acc mRMR +Acc MCFS +Acc CBFS +Acc RFE +Acc RT ); w MCFS = Acc MCFS / (Acc mRMR + Acc MCFS + Acc CBFS + Acc RFE + Acc RT ); w CBFS = Acc CBFS / (Acc mRMR +Acc MCFS +Acc CBFS +Acc RFE +Acc RT ); w RFE = Acc RFE / (Acc mRMR + Acc MCFS + Acc CBFS + Acc RFE + Acc RT ); w RT = Acc RT / (Acc mRMR +Acc MCFS +Acc CBFS +Acc RFE +Acc RT ).

8. A system for identifying mixed signal modulations using adaptive integrated feature selection, comprising: The method comprises the following steps: The sliding window positioning module is configured to divide the received signal into a plurality of small signals by using a sliding window, and determine the positions of the beginning and end of aliasing of the received signal based on the power matrix of each small signal, and extract the aliasing signal; The feature set construction module is configured to perform feature extraction on the aliasing signal to construct a feature set of the aliasing signal, and the features in the feature set of the aliasing signal include 10 combined higher-order moment features, 16 combined higher-order cumulant features, 8 higher-order moment features, 8 higher-order cumulant features and 4 instantaneous features; The integrated feature selection module is configured to select the features in the feature set by using five feature selection methods, including minimum redundancy maximum correlation criterion selection, multi-cluster feature selection, relevant feature selection, recursive feature elimination selection and random forest feature importance selection, and obtain five stable feature subsets through preliminary screening. An adaptive optimization module is configured to perform adaptive optimization on the five stable feature subsets by using an HGSO algorithm to obtain an optimal feature subset. A modulation identification module is configured to perform modulation identification based on the optimal feature subset to obtain a modulation identification result.

9. An electronic device, comprising: The method comprises: at least one processor; and a memory connected to the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the adaptive integrated feature selection method for mixed signal modulation identification according to any one of claims 1 to 7.

10. A computer readable storage medium storing a computer program, characterized in that, The computer program is executed by the processor to implement the adaptive integrated feature selection method for mixed signal modulation identification according to any one of claims 1 to 7.

Images