RU2356064C2 - Method of identifying radio signals - Google Patents

Abstract

FIELD: physics, radio engineering.

SUBSTANCE: invention concerns to radio engineering, namely, to methods of identification of types and parametres of radio signal modulation and can be used for identifying the compound time-frequency structure radio signals. The proposed method consists in the following. Reference radio signals are preset, the frame wavelet-conversion of multiple discrete and quantised samples is performed, matrices of reference radio signal power distribution matrices are formed after excluding less-significant wavelet-factors from successions of normalised and ranged wavelet-factors. It also comprises forming the vectors of reference radio signal attributes via line concatenation of power distribution matrices, carrying out the wavelet-conversion of multiple discrete and quantised samples of the received identified radio signal after excluding less-significant wavelet-factors from successions of normalised and ranged wavelet-factors. Then, follows forming the vectors of reference radio signal attributes via line concatenation of power distribution matrices and comparing vectors of identified radio signal attributes with those of reference signals. Now, the decision is made on belonging of identified radio signal to one of reference radio signals.

EFFECT: increase in efficiency of identification with required probability of correct identification.

3 cl, 16 dwg

Description

The invention relates to pattern recognition, and in particular to methods for recognizing radio signals, in particular to methods for recognizing the form and modulation parameters of radio signals. The method can be used in the technical means of recognition of radio signals with a complex time-frequency structure.

A known method of signal recognition [Omelchenko V. Recognition of signals by the power spectrum in the optimal Karunen-Loev basis. - Proceedings of the universities MB and MTR of the USSR. Ser. Radio electronics, 1980, No. 12, S.11-18], in which the energy spectrum of the signal is calculated, then the Karunen-Loev transform is performed on it and, based on the obtained characteristics, the signals are first selected for useful and interfering, and then, in the case of useful signal, carry out its comparison with the existing reference classes and assignment to one of them.

The disadvantage of this method is the relatively low probability of correct recognition ¹ (The probability of correct recognition is the relative frequency of making the right decision when assigning the received radio signal to one of the reference classes. The correct recognition event is the opposite (additional) to the error recognition event (P _dec = 1-R _osh ) - see J. Tu, R. Gonzalez. Principles of pattern recognition. Translated from English. - M.: Mir, 1978. S.142-152.) of signals having similar spectra due to weak contrast the value of the formed recognition features.

There is also a method of recognizing radio signals based on the singular decomposition of the pseudo-frequency-time-distribution (pseudo-CWR) Wigner - Wille [Marinovic N., Eichmann G. An expansion of Wigner distribution and its applications. - Proc. IEEE ICASSP-85, 1985, pp.1021-1024], in which the energy distribution matrices (REs) of the reference radio signals are preliminarily generated on the basis of the Wigner-Ville pseudo-FEC, their spectral decomposition is performed, the parameters of the reference radio signals are formed, and then the recognized radio signal is received, discretize and quantize it, form the matrix of the energy distribution of the received radio signal, select the signs of the received radio signal, compare them with the parameters of the reference radio signals and identify the received radio signal from the results of the comparison.

The disadvantage of this method is the relatively low probability of correct recognition of radio signals of complex frequency-time structure, as well as radio signals when exposed to noise and interference, which is due to the features used for the recognition of pseudo-FWM Wigner - Ville [Cohen L. Time-frequency distributions. Review // TIIER, 1989, v.77, No. 10. S.72-121]. The probability of correct recognition is reduced due to the appearance of interference background and false power peaks in the Wigner-Ville pseudo-FWM, which distort the real picture of the signal energy distribution in frequency-time coordinates.

The closest analogue in technical essence to the claimed one is the method for recognizing radio signals according to RF patent No. 2261476, IPC ⁷ G06K 9/00 of 09/27/2005. Reference radio signals are preliminarily set in the closest analogue. Then, the reference radio signals are discretized, quantized, and a continuous wavelet transform (VP) operation is performed on them in order to obtain an RE matrix. After that, for each matrix RE form the vector RE. Then, for all obtained RE vectors, the total covariance matrix is calculated. After that, spectral decomposition of the RE matrices of the reference radio signals is performed by calculating the eigenvectors and eigenvalues of the total covariance matrix. Then a truncated matrix of eigenvectors is formed by selecting the eigenvectors of the general covariance matrix corresponding to its maximum eigenvalues. When forming the parameters of the reference radio signals, a truncated matrix of eigenvectors is multiplied by the RE vectors of the reference radio signals, and the average values of the obtained products are used as the parameters of the standards. After that, a recognizable radio signal is received, it is sampled, quantized, and then the continuous VP operation of its quantized samples is performed. Then, an RE vector is formed from the RE matrix, and a truncated matrix of eigenvectors is multiplied by its RE vector to extract features of the received radio signal. The calculation results are taken as signs of recognition of the received radio signal, which are sequentially compared with the parameters of the previously obtained standards. The comparison results serve as the basis for deciding on the correlation of the recognized radio signal to one or another class.

The disadvantage of the prototype method is the relatively low efficiency (speed ² ) (Speed is the time it takes for the system to transition from a certain initial state to the required final state; one of the system quality ratings is see in the Cybernetics Dictionary. Kiev: Ukr. Sov. Encyclopedia, 1979, 623 p. // P.89.) of the recognition process, due to the need to perform continuous VP operations, perform spectral decomposition of RE matrices and form a truncated matrix of eigenvectors that are associated with a significant amount of computer ny operations.

The purpose of the claimed technical solution is to develop a method for recognizing radio signals, providing increased recognition efficiency while maintaining the required probability of correct recognition, by eliminating the redundancy of the initial description of radio signals and the formation of feature vectors without spectral decomposition of RE matrices.

The claimed technical solution expands the arsenal of funds for a similar purpose.

l-го эталонного радиосигнала формируют путем построчной конкатенации всех ВК М_l-й матрицы РЭ. После квантования принятого распознаваемого радиосигнала его матрицу РЭ

и вектор признаков

формируют аналогично как и для l-го эталонного радиосигнала.This goal is achieved by the fact that in the known method of recognizing radio signals, namely, that L≥2 reference radio signals are pre-set, form for the l-th reference radio signal, where l = 1, ..., L, the energy distribution matrix M _l , for which it is sampled, quantized, and then the VP operation of the sequence of its quantized samples is performed, for this purpose, the quantized samples are filtered using K≥2 filters and a feature vector of the l-th reference radio signal is generated, after which a recognized radio signal is received discretize it, quantize it, and then perform the VP operation of the sequence of its quantized samples, for which its quantized samples are filtered using K≥2 filters and form a vector of signs of the received radio signal, then the received radio signal is identified by comparing its signs with the characteristics of the reference radio signals, for each reference radio signals perform frame ³ operation (Frame VP is a wavelet transform using a multiple of two scaling (in frequency), and continuous shifts (in time). - see Dyakonov V. Wavelets. From theory to practice. - M .: SOLON-R, - 2002, 448 p. // S.106.) VP. To this end, the quantized samples are filtered using filters. The filtering result is wavelet coefficients (VK). The passband ΔФ _{k of the} k-th filter, where k = 1, ..., K, is selected from the condition ΔФ _k = 2 ^(k-1) ΔФ, where ΔФ is the bandwidth of the first filter. Then, the VCs of the lth reference radio signal obtained in each kth frequency band ΔФ _k are normalized, ranked and excluded insignificant VCs. Then, from the remaining VCs, a matrix of RE М _{l of the} lth reference radio signal is formed, and the rows of the matrix of RE M _{l of the} lth reference radio signal are VCs obtained in each k-th frequency band ΔФ _{k of a} sequence of its ranked and without meaningful VCs. Sign vector

l-th reference radio signal is formed by line-wise concatenation of all VK M _l- th matrix RE. After quantization of the received recognizable radio signal, its RE matrix

and feature vector

form in the same way as for the l-th reference radio signal.

Распознаваемый радиосигнал считают инцидентным эталонному радиосигналу, разница векторов признаков с которым минимальна.The received radio signal is identified by subtracting modulo its feature vector from the feature vectors of each of the L reference radio signals

The recognized radio signal is considered incident with the reference radio signal, the difference of feature vectors with which is minimal.

As the insignificant VCs of the lth reference and recognizable radio signals at the output of the kth filter, a set of VCs is selected, starting with the smallest so that the total energy of the insignificant wavelet coefficients of the _MVC is 10-30% of the total energy of the entire set of VCs at the output k th filter.

Thanks to the new set of essential features in the claimed method, the redundancy of the description of radio signals is reduced due to the frame IP, as well as due to the fact that the feature vectors are formed without spectral decomposition of the matrices. This reduces the recognition process, i.e. higher efficiency of recognition of radio signals.

The claimed method is illustrated by drawings, which show:

figure 1. Generalized block diagram of the recognition process;

figure 2. Examples of reference radio signals;

figure 3. Time-sampled samples of the radio signal;

figure 4. Voltage-quantized samples of the radio signal;

figure 5. The principle of formation of the matrix of the RE of the radio signal based on the frame VP;

Fig.6. The absolute values of the wavelet coefficients forming the RE matrix at the output of the filters performing the frame VP operation;

Fig.7. Matrix RE;

Fig.8. The principle of formation of the matrix of RE radio signal based on continuous airspace;

Fig.9. Three-dimensional representation of the matrix of the absolute values of the RE elements of a fragment of a radio signal with a duration of 64 discrete samples, obtained on the basis of a continuous VP;

figure 10. Three-dimensional representation of the matrix of the absolute values of the RE elements of a fragment of a radio signal with a duration of 64 discrete samples obtained on the basis of a frame VP;

11. Three-dimensional representation of a matrix fragment of a continuous VP test radio signal;

Fig.12. Three-dimensional representation of a fragment of a matrix of a frame VP of a test radio signal;

Fig.13. Signal vectors of reference radio signals;

Fig.14. Recognized radio signal with a duration of 64 counts;

Fig.15. Recognized radio signal feature vector;

Fig.16. A generalized plot of the probability of correct recognition of the signal-to-noise ratio (SNR) for the prototype method and the proposed method.

; формирование матрицы РЭ М распознаваемого радиосигнала; формирование на основе матрицы РЭ М вектора признаков m; сравнение вектора признаков распознаваемого радиосигнала с векторами признаков эталонных радиосигналов

; принятие решения о принадлежности распознаваемого радиосигнала к одному из L эталонных радиосигналов

.In the General case, the recognition process (figure 1) includes the following procedures: the formation of the matrices RE {M ₁ ... M _L } for the set {S ₁ (t) ... S _L (t)} of the reference radio signals; Formation of feature vectors based on RE matrices {M ₁ ... M _L }

; the formation of the matrix RE M recognizable radio signal; formation of a feature vector m based on the matrix RE; comparison of the vector of signs of a recognizable radio signal with the vectors of signs of a reference radio signal

; making a decision on whether the recognized radio signal belongs to one of the L reference radio signals

.

The claimed method allows to reduce the speed of the recognition process by reducing the number of computational procedures and thereby resolve the contradiction caused by the need for high efficiency of solving the recognition problem with an increasing volume of classes recognized radio signals.

The implementation of the claimed method is explained as follows.

Pre-set L reference radio signals, the number and types of which cover the possible number and types of real radio signals to be recognized in accordance with the task of recognizing them. As an example, FIG. 2 shows three types of such reference radio signals.

Then, a set of operations is performed in order to form a feature vector of each l-th reference radio signal, where l = 1, ..., L. For this, each reference radio signal is sampled (figure 3) and quantized (figure 4). The sampling and quantization procedures for analog signals are known and described, for example [V. Grigoriev. Signal transmission in foreign information technology systems. - SPb .: VAS. 1998, // S.83-85], and the quantized samples of the reference sequences of radio signals are formed in accordance with the requirements of the calculation of statistical estimates [Mathematical Encyclopedic Dictionary. - M .: Owls. Encyclopedia, 1988, 847 p .; Korn G., Korn T. Handbook of mathematics. Per. from English - M .: Nauka, 1977, // S.638-643]. The sample length N (with the numbers of discrete samples of radio signals n = 1, ..., N) is selected within 256 ... 163 84, depending on the requirements for the probability of correct recognition and processing time (the sample length is determined by the condition N = 2 ⁱ , where i = 1 , 2, ... is an integer). The larger N, the higher the probability of correct recognition, but the processing time increases.

Then form a set of matrices RE {M ₁ ... M _L }, for which a frame IP operation is performed on the quantized samples of the reference radio signals. The operation of the frame VP, is to filter the samples of the quantized radio signal using a combination of K≥2 filters (figure 5). The total number K of filters in this case is determined according to the condition:

where ΔF is the width of the spectrum of the radio signal; ΔF ₁ - bandwidth of the first filter [Dyakonov V. Wavelets. From theory to practice. - M .: SOLON-R, - 2002, 448 pp. // S.117-121]. In turn, the bandwidth ΔФ _{k of the} k-th filter, where k = 1, ..., K, is selected from the condition

ΔФ _k = 2 ^(k-1) ΔФ ₁ .

Such a choice of filter bandwidths provides a complete overlap in the frequency spectrum of the radio signal with a wavelet filter system and at the same time avoids the description redundancy inherent in a continuous VP [Dyakonov V. Wavelets. From theory to practice. - M .: SOLON-R, - 2002, 448 S. // S.104-107]. The result of a frame VP is a set of VC outputs K filters. The set of VK at the output of each k-th filter is normalized and then ranked. Normalization consists in highlighting in the k-th population the largest VK and dividing all the other VK into the largest VK (see Fig. 6).

The set of VK at the output of the k-th filter contains insignificant VK, which without prejudice to the accuracy of determining the signs of a radio signal can be excluded from further consideration (see Fig.6). This in turn will significantly reduce the amount of further calculations, i.e. reduce the time required for the recognition of the radio signal, and thereby increase the efficiency of recognition.

VCs that are insignificant include VK, the total energy of which is E _MVK = 10-30% of the total energy of the set of all VK at the output of the k-th filter. [Dyakonov V. Wavelets. From theory to practice. - M .: SOLON-R, - 2002, 448 p. // S.127-128]. The specific value of the E _MVC during the selection of insignificant VC depends on the type of signal, the required recognition probability and the time allocated to recognition.

Then, from the obtained “truncated” set of VCs (see FIG. 6), the matrices of RE M _{l are formed} for each l-th reference radio signal (see FIG. 7). Moreover, the rows of the RE matrix М _{l of the} lth standard radio signal are VCs obtained in each kth frequency band ΔФ _k at the output of the corresponding kth filter. Instead of the insignificant VK, zero values are entered in the RE matrix.

The choice of a frame VP is explained by the fact that, in essence, a continuous VP is an excessive description for radio signals in relation to a frame VP [Dyakonov V. Wavelets. From theory to practice. - M .: SOLON-R, - 2002, 448 p. // S. 94-98]. From [Dyakonov V. Wavelets. From theory to practice. - M .: SOLON-R, - 2002, 448 p. // S.104-107] it is known that the redundancy of a continuous airspace does not affect the probability of correct recognition of radio signals.

Consequently, the use of frame VIs will reduce the number of computational operations and thereby increase the efficiency of recognition. For example, if a system consisting of 64 filters (Fig. 8) is required to obtain an RE matrix based on a continuous VP fragment of a radio signal with a duration of 64 discrete samples (Fig. 8). That is, to obtain an RE matrix based on a frame VP of the same fragment, a system consisting of only 5 filters is required (Fig. 5).

Figure 9 shows a continuous VP of a radio signal with a duration of 64 samples, and figure 10 shows a frame VP of the same radio signal with a duration of 64 samples. In Fig.11 shows a fragment of a matrix of continuous VP obtained from the outputs of the filters No. 16-32 of the system shown in Fig.8. In Fig.12 shows a fragment of the matrix frame VI obtained from the output of the 4th filter of the system shown in Fig.5. From the results presented in Fig.11 and Fig.12, it is seen that the fragments of the matrices almost coincide.

The order of formation of the matrix RE M _l is known as the algorithm for the formation of VP and is described, for example, in [Dyakonov V. MATLAB. Wavelets. From theory to practice. - M.: SOLON-R, 2002, 448 p. // S.104-107]. The procedure for calculating the elements of the matrices RE M / is known and described, for example, in [Dyakonov V. MATLAB. Signal and image processing. Special reference. - St. Petersburg: Peter, 2002, 608 p. // S.307-312; Baskakov S. Radio engineering circuits and signals. Textbook for universities on special. "Radio engineering". - 3rd ed., Revised. and add. - M .: Higher. school, 2000, 462 p. // S.65-68].

путем построчной конкатенации всех ВК М_l-й матрицы РЭ с учетом наличия позиций нулевых элементов. Для матрицы РЭ М_l общая длина вектора признаков

равна (К·N). В результате выполнения указанных операций для множества {S₁(t)…S_L(t)}, получают совокупность признаков

эталонных радиосигналов (фиг.13).The procedures for performing a frame VP operation can be implemented using a filter bank (Fig. 5), the amplitude-frequency characteristics of which correspond to the bandwidths of the basic wavelet functions. In this case, the dimension of the RE M matrix is (K × N), where K is the number of filters, N is the number of discrete samples of the signal. After this form the vector of signs of the l-th reference radio signal

by line-by-line concatenation of all VC M _l- th matrix RE, taking into account the presence of positions of zero elements. For the matrix RE M _{l the} total length of the feature vector

equal to (K · N). As a result of performing these operations for the set {S ₁ (t) ... S _L (t)}, we obtain a set of features

reference radio signals (Fig.13).

(t) (фиг.14) и выполняют над ним все описанные действия, которые выполнялись над эталонными радиосигналами. Принятый радиосигнал дискретизируют, квантуют и выполняют операцию фреймового ВП последовательности его квантованных отсчетов, после чего ВК на выходе каждого k -го фильтра нормируют и затем ранжируют, отсекают малозначимые ВК, формируют матрицу РЭ

, из которой построчной конкатенацией формируют вектор признаков принятого радиосигнала

аналогично, как и для тестовых радиосигналов (фиг.15).Then, a recognized radio signal is received.

(t) (Fig) and perform on it all the described actions that were performed on the reference radio signals. The received radio signal is sampled, quantized, and the frame VP operation is performed on the sequence of its quantized samples, after which the VC at the output of each kth filter is normalized and then ranked, cut off the insignificant VC, form the RE matrix

, from which line-by-line concatenation form the vector of signs of the received radio signal

similarly as for test radio signals (Fig.15).

с вектором признаков каждого из эталонных радиосигналов

. Идентификация может быть реализована с использованием различных приемов. Например, путем вычитания по модулю из вектора признаков принятого радиосигнала векторов признаков каждого из L эталонных радиосигналов

Процедуры принятия решения являются известными и описаны, например, в [Фомин Я., Тарловский Г. Статистическая теория распознавания образов. - М.: Радио и связь, 1986 г.// С.30-46; Сато Ю. Обработка сигналов. Первое знакомство. / пер. с яп., под ред. Есифуми Амэмия. - М.: Издательский дом «Додека-XXI», 2002. 176 с.// С.41-54]. Распознаваемый радиосигнал считают инцидентным одному их L эталонных радиосигналов, когда суммарная разница между векторами признаков минимальна.Identify a recognizable radio signal by comparing its feature vector

with a feature vector of each of the reference radio signals

. Identification can be implemented using various techniques. For example, by subtracting modulo from the feature vector of the received radio signal the feature vectors of each of the L reference radio signals

Decision-making procedures are well known and are described, for example, in [Y. Fomin, G. Tarlovsky. Statistical theory of pattern recognition. - M.: Radio and communications, 1986 g. // S.30-46; Sato Yu. Signal Processing. First meeting. / per. with Japanese., ed. Yesifumi Amemiya. - M .: Publishing house "Dodeka-XXI", 2002. 176 p. // S.41-54]. A recognized radio signal is considered incident to one of their L reference radio signals when the total difference between the feature vectors is minimal.

Comparative assessment of the probability of correct recognition. P _rasp using the claimed method and the prototype method is performed by modeling the recognition process on a computer.

As recognized signals, 7 amplitude, frequency, and phase telegraphy signal models with different transmission rates were used (AT 50 Baud - an amplitude telegraphy radio signal with a transmission speed of 50 Baud; BW 800 Hz 500 Bd, BW 800 Hz 200 Baud, BW 600 Hz 100 Baud , BT 200 Hz 100 Baud - frequency telegraphy radio signals with a spacing of 800, 600, 200 Hz and a transmission speed of 500, 200, 100 Baud; FT 500 Baud, FT 1200 Baud - phase telegraphy radio signals with a transmission speed of 500 and 1200 Baud). The radio signals are normalized to the average power level. The formation of the parameters of the standards was carried out on 100 samples of each signal. The probability P _{dec of the} correct recognition was estimated by the Monte Carlo method using 100 samples for each signal for different ratios of the signal power P _s and noise power R _w in the range P _s / P _w = 0-20 dB (Fig. 16).

The results of comparative calculations showed that the probability of correct recognition of the radio signals P _{dec by the} claimed method is close to the prototype method, with an almost 2-fold reduction in recognition time due to a significant reduction in the number of computational procedures, which indicates the possibility of achieving a technical result when using the claimed method - to increase efficiency recognition of radio signals.

Claims (3)

1. The method of recognition of radio signals, which consists in pre-setting L≥2 reference radio signals, is formed for the l-th reference radio signal, where l = 1, ..., L, the energy distribution matrix M _l , for which it is sampled, quantized, and then perform the operation of the wavelet transform of the sequence of its quantized reports, for this purpose, the quantized reports are filtered using K≥2 filters and form a feature vector of the l-th reference radio signal, after which a recognized radio signal is received, sampled, quantized, and quantized then, the wavelet transform operation of the sequence of its quantized samples is performed, for which its quantized samples are filtered using K filters and the vector of signs of the received radio signal is formed, then the received radio signal is identified by comparing its signs with the signs of the reference radio signals, characterized in that the wavelet transform is performed as frame wavelet transform, for which the passband ΔФ _{k of the} k-th filter, where k = 1, ..., K, is selected from the condition ΔФ _k = 2 ^(k-1) ΔФ, where ΔФ is the bandwidth the first filter, then the wavelet coefficients of the l-th reference radio signal obtained in each k-th frequency band ΔФ _k are normalized, rank and exclude insignificant wavelet coefficients, then from the remaining wavelet coefficients form the energy distribution matrix M _{l of the} l-th reference radio signal moreover, the rows of the energy distribution matrix M _{l of the} l-th reference radio signal are the wavelet coefficients obtained in each k-th frequency band ΔФ _{k of a} sequence of its ranked and without insignificant wavelet coefficients, and the vector features of the l-th reference radio signal are formed by line-wise concatenation of all wavelet coefficients M _l- th energy distribution matrix, and after quantization of the received recognized radio signal, its energy distribution matrix and feature vector are formed in the same way as for the l-th radio frequency reference signal. 2. The method according to claim 1, characterized in that as the insignificant wavelet coefficients of the l-th reference and recognizable radio signals at the output of the k-th filter, a set of wavelet coefficients is selected, starting from the lowest, the total energy of which is 10-30% of the total energy of the entire set of wavelet coefficients at the output of the kth filter. 3. The method according to claim 1, characterized in that the received radio signal is identified by subtracting modulo its feature vector from the feature vectors of each of the L reference radio signals, and the recognized radio signal is considered incident with the reference radio signal, the difference of the feature vectors with which is minimal.

Images