RU2748935C1 - Method of recognition of new low bit rate coding protocols - Google Patents

Abstract

FIELD: information technology.

SUBSTANCE: invention relates to the field of information technology, namely to the field of digital communication. A recognition method (NSCR) implemented in vocoders is proposed, a binary digital information stream y is received during the time interval ΔT, a normalized autocorrelation function a is formed on the basis of y, a decision is made on the presence of a block structure in the digital information stream, used to describe the input implementation image in the form of a square matrix M of mean values ​​and the corresponding covariance matrices C. Based on the values ​​of _Vjl, a divergence vector _Vj is formed for each j-th reference image. J vectors _Vj form a matrix of divergence values ​​V of dimension J×L, L is the number of bits between the extrema of the autocorrelation function. A decision is made to discover a new previously unknown NSCR protocol and the number J + 1 is assigned to a new reference description.

EFFECT: technical result consists in reducing the probability of a false alarm and, as a consequence, increasing the recognition reliability (the probability of correct recognition) of new protocols (NSCR).

1 cl, 5 dwg

Description

The invention relates to the field of digital communication, namely, to the processing of digital streams of the channel level of the reference model of the interaction of open systems and can be used in a wide class of digital signal processing tools for recognizing new low-speed speech coding (LSCR) protocols used in the transmission of speech messages over digital communication channels.

The claimed technical solution expands the arsenal of tools for a similar purpose due to the ability to recognize new NSCR protocols with a lesser likelihood of false alarms.

A known method for recognizing NSCR protocols (see RF Patent No. RU 2610285 C1, IPC G10L 19/008 (2013.01), H03M 13/03 (2006.01). A method for recognizing low-speed coding protocols. Aladinsky VA, Kuzminsky SV, Smirnov P.L., Chubaty D.N., publ. 02/08/2017), in which, based on a comparison of the reference descriptions represented by sets of averaged values of the redundancy factor (RI), and the investigated image, represented by a set of RR values of the columns of a rectangular matrix formed from the input digital information stream, the values of the deviation between the measured set of CI values and the sets of averaged CI values are determined, in the obtained set of values of the deviation, the minimum is selected, it is compared with the threshold value, based on the comparison results, a decision is made that the investigated digital stream is formed on the basis of one from the known NSCR protocols, otherwise it is believed that for the formation of the analyzed digital stream, a new th (previously unknown) NSCR protocol. The disadvantage of this method is the low noise immunity of recognition, which leads to the appearance of a large number of false decisions about the use of new NSCR protocols on radio communication lines with the probability of a bit error P _osh > 0.03.

Closest to the claimed is a method (prototype) for recognizing low-speed coding protocols (see RF Patent No. RU 2667462 C1, Method for recognizing low-speed coding protocols, IPC G10L 19/008 (2013.01), H03M 13/03 (2006.01). Aladinsky V.A. ., Veshchunin E.A., Kuz'minsky S.V., Smirnov P.L.publ. 19.09.2018), which consists in the fact that they receive a binary digital information stream y during the time interval ΔT, form a normalized autocorrelation function based on y a, make a decision on the presence of a block structure in the digital information stream y according to the extrema of the autocorrelation function a regular with equal intervals Δτ, divide the digital information stream y into information blocks of N _b bits each with intervals Δτ between the extrema of the autocorrelation function a, sequentially assign the information blocks serial numbers k = 1, 2, ..., K, starting from the first information block, form a rectangular information matrix Y of size erov K × L, L = N _b , the rows of which are information blocks sequentially placed under each other in accordance with their ordinal numbers k = 1, 2, ..., K, columns y _z , z = 1, 2, are selected from the matrix Y, ..., L; the value of the mathematical expectation (MO) m _{z is} calculated for each column y _z , based on the obtained values of m _z , a vector of values of MO m ₀ = (m ₁ , m ₂ , ..., m _z , ..., m _L ) is formed by sequentially placing them in in accordance with the ordinal numbers z, on the basis of the obtained vector m ₀ form a square matrix M of MO values of size L, rows m _l , l = 0, 1, 2, ..., L-1, which contains the values of the vector m ₀ , sequentially circularly shifted to the left by the value l; form rectangular reference information matrices on the basis of which

calculate the vectors of reference values MO m _{j et} , j = 1, 2, ..., J, and square reference covariance matrices {C _{j et} } J corresponding to J known NSCR protocols; the values of the probability of correct recognition P _{jl of the} jth NSCR protocol, j = 1, 2, ..., J, _{are calculated for each l-th row m l of the} matrix M, a decision is made in favor of the j-th NSCR protocol for which the maximum value is provided P _jl .

The prototype method has a high accuracy of recognition of the NSCR protocols in conditions of a high level of bit errors in the investigated digital information stream y. However, the disadvantage of the prototype is the appearance when calculating the probability of correct recognition of the uncertainty of the form P _jl = 0/0 for all j, l. The latter arises when digital information streams are fed to the input, including those formed on the basis of new NSCR protocols, with the exception of implementations for which there are reference descriptions. As a result, there is an increase in the probability of a false alarm when recognizing new NSCR protocols in conditions of a decrease in the volume K of the input digital information stream y, especially when K <100. The consequence of an increase in the number J of reference descriptions formed due to incorrectly recognized input realizations during training of the NSCR protocol recognition system is a decrease in the probability of correct recognition.

The aim of the claimed technical solution is to develop a method that reduces the probability of false alarms and, as a consequence, increases the reliability of recognition (the probability of correct recognition) of new NSCR protocols.

This goal is achieved by the fact that in the known method for recognizing NSCR protocols, which includes receiving a binary digital information stream y during the time interval ΔT, forming a normalized autocorrelation function a on the basis of y, making a decision on the presence of a block structure in the information stream y at regular intervals Δτ to the extrema of the autocorrelation function a, dividing the digital information flow y into information blocks of N _b bits each according to the intervals between the extrema of the autocorrelation function a, sequentially assigning sequence numbers k = 1, 2, ..., K to the information blocks, starting from the first information block, forming a rectangular information matrix Y of sizes K × L, L = N _b , the rows of which are information blocks sequentially placed under each other in accordance with their ordinal numbers k = 1, 2, ..., K, extraction of columns y _z , z = from the matrix Y 1, 2, ..., L, determination of values of MO m _z for each column y _z , formation of a vector of values of MO m ₀ = (m ₁ , m ₂ , ..., m _z , ..., m _L ) by sequential placement of values of MO m _z , formation of a square matrix of values of MO M size L × L, rows m _l , l = 0, 1, 2, ..., L-1, which contains the values of the vector then, sequentially circularly shifted to the left by the value l, the formation of rectangular reference information matrices {Y _{j et} } _J , calculation based on the reference information matrices {Y _{j et} } _J vectors of reference values of MO m _{j et} , y = 1, 2, ..., J and square reference covariance matrices {C _{j et} } _J , calculating the value of the probability of correct recognition P _jl for each j-th NSCR protocol, j = 1, 2 , ..., J, for each 1st row m _{l of the} matrix M, making a decision in favor of the jth NSCR protocol, for which the maximum value of P _jl is provided, upon obtaining the result of calculating the probability of correct recognition in the form of uncertainty P _jl = 0/0 for all j, l, form a set of rectangular matrices {Y _l } _L , the columns of which rykh are shifted circularly by l = 0, 1, 2, ..., L-1 relative to the columns of the rectangular information matrix Y. For each matrix Y _{l, the} corresponding square covariance matrix C _{l is} calculated.

Determine the value of the divergence v _jl between the image of the input implementation, represented by rows m _{l of the} square matrix of M values of MO and the corresponding covariance matrices C _l , and each j-th reference image represented by the vector m _{j et} and the corresponding square reference covariance matrix C _{j et.} on the basis of the obtained values of v _jl , a vector of divergence values v _j = (v _j0 , v _j1 , ..., v _jl ..., v _{j (L-1)} ) is formed for each j-th reference image. A matrix of values of divergence V of sizes J × L is composed of J vectors v _j.

A decision is made to detect a new previously unknown NSCR protocol and a conditional number J = J + 1 is assigned to a new reference description, which includes the mathematical expectation vector m (J + 1) _fl = m ₀ and a square reference covariance matrix C (J + 1) _fl = С ₀ , if for all elements v _{jl of the} matrix of divergence values V the condition v _jl > v _pop = 2000 is satisfied for any j, l, where v _pop is the threshold value, or only for one element v _jl <300, and all the rest v _jl > v _then , it is decided that the input implementation y is formed according to the j-th known NSCR protocol and the information stream y has no initial l bits in the first k = 1 information block. In the case of 300≤v _jl ≤2000, a decision is made that the input implementation y is not recognized.

This solution corresponds, for example, to the situation when the digital information stream y contains less than 100 information blocks (i.e., K <100) or the number of single bit errors in it is more than 10%.

Thanks to the new set of essential features in the claimed method, an increase in the recognition reliability (the probability of correct recognition) of new NSCR protocols is achieved due to a more complete description of the input implementation image represented by rows m _{l of the} square matrix M of MO values and the corresponding covariance matrices C _l , as well as determining the divergence v _jl between this image and the reference images represented by vectors m _{j et} and the corresponding reference covariance matrices C _{j et.} ...

The claimed method is illustrated by drawings, which show:

in fig. 1 - the dependence of the probability of false alarm P _lt on the number of blocks K contained in the digital information stream y, when recognizing the NSCR protocols;

in fig. 2 - the order of formation of a square matrix of values of MO M dimensions Z = L;

in fig. 3 - an algorithm for recognizing new NSCR protocols;

in fig. 4 - dependences of the reliability D _{p of the} decision taken on the number of blocks K contained in the digital information stream y when recognizing the NSCR protocols;

in fig. 5 - dependence of the probability of false alarm P _lt on the probability of a bit error P _osh in the digital information stream y, when recognizing the NSCR protocols.

The development of digital telephone radio communication is carried out in many directions, one of which is the development of NSCR hardware and software, called vocoders. For example, to date, at least 20 different types of vocoders and their corresponding NSCR protocols have been developed for use on high-frequency (HF) radio communication lines (see Aladinsky V.A., Kuzminsky S.V. Analysis of digital streams at the outputs of vocoders, accepted on foreign lines of radio communication of the range of high waves // Successes of modern radio electronics. - M .: Radiotekhnika, No. 7, 2015 - P. 73-76). Known NSCR protocols can be open or closed. Information about closed protocols is confidential. In HF radio communication technology, the open protocols LPC-10-2400 (STANAG 4198) and MELPe-2400 (STANAG 4591), published in the relevant NATO standards, are used. At the same time, the closed protocol MELPe-600 (STANAG 4591) is used. New (unknown, not previously applied) NSCR protocols are closed.

The emergence of new NSCR protocols on radio communication lines is due to the desire of manufacturers of radio communication equipment to increase their market share, commercial profits and protect their competencies, on the one hand, as well as to improve the quality of communication through the introduction of new technical solutions, including modifications of previously known protocols on the other hand. The implementation of the new NSCR protocols in the existing radio communication facilities can be performed at the software level or by connecting an external vocoder using the existing interfaces. Thus, the use of new NSCR protocols when transmitting voice messages using radio communications is not an exceptional event, which determines the need to solve this technical problem using pattern recognition methods.

The use of the prototype method for recognizing NSCR protocols under conditions of a large number of errors (more than 10% of the total number of bits) in the studied digital streams obtained by encoding speech messages using known NSCR protocols, causes the appearance of the results of calculating the probability of correct recognition in the form of uncertainty P _ji = 0/0 for all j, l. Similar results take place with fewer bit errors if K <100. This indicates insufficient noise immunity of the prototype when recognizing new NSCR protocols and insufficient information content of the used set of features to describe the image of the input implementation with a small volume of the investigated information flow y, represented only by the vector MO m ₀ .

A special case is the beginning of the recognition system operation (initial mode), in which the absence of reference descriptions of the NSCR protocols is noted. There is a need to make a decision on whether the input realization y belongs to the class j = 1 based on the results of the analysis of the parameters of the autocorrelation function a, the formation of the 1st reference description and further comparison with the obtained reference description of other input implementations.

In the general case (including the training mode), the fact of using the new NSCR protocol must be established in the absence of its reference description and the recognition system must be trained on the basis of one obtained implementation of the binary digital information flow y, which is automatically performed when deciding in favor of the new NSCR protocol.

A positive effect in the proposed method is achieved by increasing the dimension of the data to describe the image of the input implementation, represented in the form of a square matrix M of MO values and the corresponding covariance matrices _C1 , using a more informative measure of discrimination, which is divergence. The latter allows comparing the image of the input implementation of increased dimension with the reference descriptions, as a result of which the probability of correct recognition of new NSCR protocols is increased. In addition, a decrease in the influence of the level of bit errors in the input digital information stream y and the values of its volume K on the recognition quality of new NSCR protocols is achieved by introducing an additional gradation "Implementation ownership is not established" in the decision criterion. As a result, the images of input implementations, which for various reasons have significant distortions, are not recognized as new, which significantly reduces the probability of false alarms when recognizing new NSCR protocols.

The implementation of the claimed method can be carried out as follows (see Fig. 3). At the stage of inputting the initial data, it is advisable to determine the values of the parameters ΔT, N _{cp of the} digital information stream y, the duration of the analysis interval y and / or the capacity of the analyzed stream y, respectively, to form a set {N _b }, including all possible values of this parameter, in the general case - to enter reference descriptions J of known NSCR protocols, including the sets {m _{j et} } J, {C _{j et} } _J , in a particular case - set the parameter J = 0.

After that, the normalized autocorrelation function a is calculated, which characterizes the linear dependence of the original sequence y on the sequence y (τ), shifted circularly by τ = 1, 2, ..., N _cp - 1 bit with respect to y. The normalized autocorrelation function a of a digital stream containing a speech message has pronounced local maxima located with a periodicity of Δτ. The presence of local maxima of the autocorrelation function, located with a period of Δτ, is explained by the fact that information symbols describing the parameters of a digital stream with a voice message are located inside the information block strictly in certain places, at intervals Δτ, which are multiples of the length of the information block. The parameters of the speech signal have a significant linear relationship, so the ACF value at these points will be the largest. The sought autocorrelation function is considered to be a set of values of the form

where a (τ) = c (τ) / d (y) is the correlation coefficient; с (τ) - covariance coefficient; d (y) is the variance of the digital information stream y.

By regular with equal intervals Δτ = N _b extrema of the normalized autocorrelation function a, a decision is made on the presence of a block structure in the information flow y. The absence of extrema of the normalized autocorrelation function a with equal intervals Δτ indicates that the digital information stream y does not contain a speech message, and its analysis ends with the output of a result that the input implementation y does not contain a speech message, and its analysis ends with the output of the result that the input implementation y does not contain a speech message.

In the presence of extrema of the normalized autocorrelation function a with equal intervals Δτ, the number of information blocks contained in the digital information stream y is calculated as follows:

- оператор округления величины до меньшего целого значения.Where

- operator of rounding of a value to a smaller integer value.

Divide the information flow from information blocks on a volume N _b bits each. The information blocks are sequentially assigned sequence numbers k = 1, 2, ..., K, starting from the first information block. A rectangular information matrix Y is formed, the rows of which are information blocks sequentially placed one under the other in accordance with their ordinal numbers k = 1, 2, ..., K.

_{Columns y z} are separated from the information matrix Y. The condition J≥1 is checked (i.e., the presence of at least one reference description), which establishes the possibility of introducing the learning mode.

If the condition J≥1 is not met (this is a special case - the initial mode), then using well-known expressions (see Aladinsky V.A., Kuzminsky S.V. : Radio engineering. No. 12, 2015. - S. 20-25) determine the values of MO m _z by the columns y _z . From the set of MO values, a vector of values of the mathematical expectation of the form m ₀ = (m ₁ , m ₂ , ..., m _z , ..., m _L ) of size L is formed. On the basis of a rectangular information matrix Y, a square covariance matrix C _{0 is} calculated (see ibid. ). Next, a reference description is formed, which is assigned a serial number j = 1, and in this case m ₀ = m _{1 floor} , C ₀ = C _{1 floor} .

When the condition J≥1 is fulfilled, the values of MO m _z are also determined, a vector of values of the mathematical expectation of the form m ₀ , a square matrix M = (m ₀ , m ₁ , ..., m _l , ..., m _L-1 ) of sizes L, containing rows m _l obtained on the basis of the circular shift of the elements of the vector m ₀ (see Fig. 2). A set of rectangular matrices {Y _l } _L = {Y ₀ , Y ₁ , ..., Y _l , ..., Y _L-1 } is formed, whose columns are shifted circularly by l = 0, 1, 2, ..., L-1 with respect to columns of the rectangular information matrix Y. For each matrix Y _l from the collection {Y _l } z, calculate the square covariance matrices C _l .

The values of the divergence v _jl between the image of the input implementation, represented by the corresponding vectors m _l and square covariance matrices C _l , and the j-th reference description are determined. The latter is represented by the vector m _{j et} and the corresponding reference covariance matrix C _{j et} . In the general case, the identification of a new protocol assumes the fulfillment of the conditions m _l ≠ m _j and C _l ≠ C _j , then the divergence characterizing the separating surface of the image and the reference description is calculated using a formula of the form (see Tu J., Gonzalez R. Principles of pattern recognition / Per. From English .-- M .: Mir, 1978 .-- 411 p.)

- след матрицы S размеров Z;

- след матрицы В размеров Z; a_zz, b_zz - элементы диагоналей матриц S и В соответственно;Where

- trace of the matrix S of dimensions Z;

- trace of the matrix B of dimensions Z; a _zz , b _zz are the elements of the diagonals of the matrices S and B, respectively;

From the obtained set of values v _jl, for the convenience of further analysis, a matrix of divergence values V of sizes J × L is formed. The value of each element v _{jl of the} matrix V is compared with the threshold value v _pore = 2000.

If all elements v _jl matrix V satisfies the condition v _jl> v _then, a decision is made about the discovery of a new protocol NCIS and assigned identification number J-J + 1 new reference description that includes the expectation vector m (j + 1) _fl = m ₀ and a square reference covariance matrix C (J + 1) _et = C ₀ . The message "New J + 1 NSCR protocol" is displayed. If this condition is not met and if the condition v _jl <300 is satisfied only for one element of the matrix V, and the remaining v _jl > v _pores = 2000, then it is decided that the input implementation y is formed according to the j-th known NSCR protocol and in the information flow y there are no initial l bits in the first k = 1 information block. If necessary, additional training of the recognition system is carried out, the j-th reference description is corrected m _{j et} , C _{j et} . The message "Incomplete first block of input data, j-th NSCR protocol" is displayed.

In other cases, when 300≤v _jl ≤2000 is true for all elements of the matrix V, it is considered that the input implementation y cannot be attributed to any j-th known NSCR protocol or the new NSCR protocol, the message “The input implementation does not belong to installed ".

Simulation modeling of the claimed method for recognizing NSCR protocols is performed using digital information streams formed on the basis of the well-known protocols LPC-10-2400 (STANAG 4197), conditional class j = 1, MELPe-2400 (STANAG 4591), j = 2, used on the lines radio communications of the high frequency range, and digital information streams, generated by the software by the MELP-240Q vocoder according to the NSCR protocol with the class number j = 3, which is considered new, showed the following. The accepted threshold value D _p = 70% is achieved on the basis of the proposed method at values of K = 82, while the prototype method ensures the satisfaction of this requirement at K = 124. The reliability D _{p of} recognition by the proposed method at values of K≥82 is higher than when using the prototype by an amount from (11 to 55)%, which confirms the effectiveness of the proposed technical solutions in a small volume of the investigated digital information flows.

Comparison of noise immunity of the proposed method and the prototype method is also made on the basis of simulation at K≥100, bit errors were introduced into digital information streams according to a random law with a uniform error probability, which corresponds to modern conditions for the formation of discrete signals for the transmission of speech messages when noise-immune coding is used , interleaving and scrambling of digital information streams prior to signal transmission to the radio communication channel. The accepted maximum value of the probability P _{osh_max} = 15% corresponds to the value at which the recovery of the speech signal is still possible. As a result, it was found that the developed method for recognizing new NSCR protocols provides a lower probability of a false alarm P _osh with an equal value of the probability of a bit error P _osh in a digital information stream y (see Fig. 5) by an amount from (2 to 25)% than with using the prototype method, the condition P _lt ≤30%, which is the boundary, is fulfilled for the prototype at P _osh ≤10%, at the same time for the proposed method - at P _osh ≤14%, which indicates its better noise immunity compared to the prototype ...

Claims (1)

A method for recognizing new low-rate speech coding (NSCR) protocols implemented in vocoders, which consists in the fact that they receive a binary digital information stream y during the time interval ΔT, form a normalized autocorrelation function a on the basis of y, decide on the presence of a block structure in the digital information stream y along the extrema of the autocorrelation function a regular with equal intervals Δτ, divide the digital information stream y into information blocks of N _b bits each according to the intervals between the extrema of the autocorrelation function a, sequentially assign the sequence numbers k = 1, 2, ..., K to the information blocks, starting from the first information block, a rectangular information matrix Y of size K × L, L = N _{b is formed} , the rows of which are information blocks sequentially placed under each other in accordance with their ordinal numbers k = 1, 2, ..., K, are extracted from the matrix Y columns y _z , z = 1, 2,…, L; determine for each column y _{z the} value of its mathematical expectation m _z , form a vector of values of the mathematical expectation m ₀ = (m ₁ , m ₂ , ..., m _z , ..., m _L ) size L, based on the resulting vector m ₀ form a square matrix M values of the mathematical expectation of size L × L, lines m _l , l = 0, 1, 2, ..., L-1, which contains the values of the vector m ₀ , sequentially circularly shifted to the left by the value l, according to the available training samples {y _{j et} } _J (j = 1, 2, ..., J) corresponding to J well-known NSCR protocols, rectangular reference information matrices {Y _{j et} } _{J are formed} , on the basis of which vectors of reference values of the mathematical expectation m _{j et} , j = 1, 2 , ..., J, and square reference covariance matrices {C _{j et} } _J , calculate the values of the probability of correct recognition P _{jl of the} j-th NSCR protocol, j = 1, 2, ..., J, for each l-th row m _{l of the} matrix M make a decision in favor of the jth NSCR protocol, for which the maximum value of P _jl is provided, which differs which is that for the value of the probability of correct recognition P _jl = 0/0 for all j, l, a set of rectangular matrices {Y _l } _{L is formed} , the columns of which are shifted circularly by l = 0, 1, 2, ..., L-1 relative to the columns rectangular information matrix Y, for each matrix Y _l calculate its square covariance matrix C _l , determine the values of the divergence v _jl between the images of the input implementation, represented by rows m _{l of the} square matrix M of the mean values and the corresponding covariance matrices C _l , and each j-th standard image represented by the vector m _{j et} and the corresponding square reference covariance matrix C _{j et} , based on the obtained values of v _jl , a divergence vector v _j = (v _j0 , v _j1 , ..., v _jl ..., v _{j (L-1)} ) for each j-th reference image, make up a _{matrix of divergence values V of size J × L from J vectors v j} , decide to detect a new previously unknown NSCR protocol and assign a conditional number р J = J + 1 to the new reference description, which includes the vector of mathematical expectation m _{(J + 1) fl} = m ₀ and the square reference covariance matrix C _{(J + 1) fl} = С ₀ , if for all elements v _{jl of the} matrix of divergence values V, the condition v _jl > v _pores = 2000 for any j, l, where v _pores is the threshold value, or only for one element v _jl <300, and the rest v _jl > v _pores , make a decision that the input realization y is formed according to The j-th known NSCR protocol and the information stream y lack the initial l bits in the first k = 1 information block, and in the case of 300≤v _jl ≤2000, a decision is made that the input implementation of у is not recognized.

Images