CN103763230A - Improved self-adaption blind source separation method - Google Patents

Improved self-adaption blind source separation method Download PDF

Info

Publication number
CN103763230A
CN103763230A CN201410022179.0A CN201410022179A CN103763230A CN 103763230 A CN103763230 A CN 103763230A CN 201410022179 A CN201410022179 A CN 201410022179A CN 103763230 A CN103763230 A CN 103763230A
Authority
CN
China
Prior art keywords
matrix
separation
blind source
improved
piece
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201410022179.0A
Other languages
Chinese (zh)
Other versions
CN103763230B (en
Inventor
郭业才
张政
柏鹤
黄友锐
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University of Information Science and Technology
Original Assignee
Nanjing University of Information Science and Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Nanjing University of Information Science and Technology filed Critical Nanjing University of Information Science and Technology
Priority to CN201410022179.0A priority Critical patent/CN103763230B/en
Publication of CN103763230A publication Critical patent/CN103763230A/en
Application granted granted Critical
Publication of CN103763230B publication Critical patent/CN103763230B/en
Expired - Fee Related legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Landscapes

  • Filters That Use Time-Delay Elements (AREA)

Abstract

The invention provides an improved self-adaption blind source separation method NVS-NGA, and the improved self-adaption blind source separation method overcomes the defect that no blind source separation scheme capable of effectively handling the contradiction between the rate of convergence and babbling errors exists in the prior art. Firstly, a traditional separation system structure is improved, the performance index of blind source separation is improved, the improved separation performance index serves as an independent variable, and a Rayleigh distribution function serves as a dependent variable and a variable step size function. Compared with a traditional blind source separation method, the improved self-adaption blind source separation method can rapidly and effectively separate an original signal from a received mixed signal, can effectively handle the contradiction between the rate of convergence and the babbling errors, is high in rate of convergence, small in babbling error and high in stability, and has wide application prospects in the aspects of wireless communication, image processing, voice signal processing and the like.

Description

A kind of improved self-adaptive blind source separation method
Technical field
The invention belongs to signal processing technology field, especially relate to a kind of improved self-adaptive blind source separation method.
Background technology
Blind source separates and refers in the case of the theoretical model of signal and source signal cannot accurately be known, how from mixed repeatedly signal (observation signal), to isolate the process of each source signal.Blind source separates (Blind Source Separation, BSS) basic task is in the case of the hybrid mode of source signal the unknown and source signal is also unknown, from one group of observation signal receiving, recover source signal, at aspects such as recognition of face, voice signal processing, processing of biomedical signals, satellite and microwave communications, there is huge application potential.Blind source separation method can be divided into batch processing method and the large class of adaptive approach two.Compared with batch processing method, adaptive approach can real-time tracking signal intensity.
Tradition adaptive blind source separation system, referred to as traditional piece-rate system, as shown in Figure 1.Tradition piece-rate system is in series by the nonsingular hybrid matrix A of the unknown and separation matrix W (k); By mutually source signal S (k)=[s independently of M 1(k), s 2(k) ..., s m(k)] tthrough a nonsingular hybrid matrix A of the unknown, be mixed to get observation signal X (k)=[x 1(k), x 2(k) ..., x m(k)] t, x m(k) be M observation signal.When ignoring transmission delay effect and noise,
X(k)=AS(k) (1)
In formula, A is that M × M ties up matrix.The target that blind source separates is when only knowing observation signal X (k), obtains, after the separation matrix W (k) of a full rank, obtaining separation signal Y (k) from observation signal X (k) by iteration
Y(k)=W(k)X(k) (2)
In formula, Y (k) is an estimation to source signal S (k), is M × 1 dimension matrix; W (k) is that M × M ties up matrix.
Utilize the relation of mutual information and comentropy, the cost function of piece-rate system is defined as
J ( k ) = H ( y m ( k ) ) - ln | det ( W ( k ) ) | = - Σ m = 1 M E ( ln p y ( y m ( k ) ) ) - ln | det ( W ( k ) ) | - - - ( 3 )
In formula, H (y m(k)) be m separation signal y in separation signal vector Y (k) m(k) entropy, p y(y m(k)) be m separation signal y in separation signal vector Y (k) m(k) marginal probability density, E represents mathematic expectaion computing, and ln represents the natural logrithm take e the end of as, and det represents to get the determinant of W (k).When the cost function J of piece-rate system (k) is while being minimum, best separation matrix W (k) can make in separation signal vector Y (k) each component independent of one another.Utilize the Edgeworth of probability density to launch and get fourth order cumulant,
H ( y m ( k ) ) ≈ H N ( y m ( k ) ) - [ 1 12 cum 2 ( y m 3 ( k ) ) + 1 48 cum 2 ( y m 4 ( k ) ) + 7 48 cum 4 ( y m 3 ( k ) ) - 1 8 cum 2 ( y m 3 ( k ) ) cum ( y m 4 ( k ) ) - - - ( 4 )
In formula, H n(y m(k)) Normal Distribution, with y m(k) there is identical average and variance, establish y m(k) average is zero, variance is 1, H n(y m(k) the subscript N) represents normal distribution, cum represents y m(k) cumulant computing;
H (y i(k)) substitution J (k), J (k) is capable and j column element w to i in W (k) ij(k) random gradient is
( ▿ R J ) ij = ∂ J ( k ) ∂ w ij ( k ) = - ( W - 1 ( k ) ) ij T + [ 1 2 cum 2 ( y i 3 ( k ) ) - 1 6 cum ( y i 4 ( k ) ) ] E [ y i 3 ( k ) x j ( k ) ] + [ 3 4 cum ( y i 3 ( k ) ) cum ( y i 4 ( k ) ) - 1 2 cum ( y i 3 ( k ) ) - 7 4 cum 3 ( y i 3 ( k ) ) ] E [ y i 2 ( k ) x j ( k ) ] - - - ( 5 )
In formula, w ij(k) be the capable and j column element of the i of W (k);
Figure BDA0000458465520000024
represent W -1(k) transposition obtains the capable and j column element of the i of matrix; Cum represents cumulant computing, and E represents mathematic expectaion computing;
Figure BDA0000458465520000029
in
Figure BDA00004584655200000210
represent random gradient,
Figure BDA00004584655200000211
for J (k) to W (k) i capable and j column element w ij(k) random gradient,
Figure BDA00004584655200000212
transient expression formula by formula (5), can be obtained
( ▿ R J ) ij = - ( W - 1 ( k ) ) ij T + { [ 1 2 cum 2 ( y i 3 ( k ) ) - 1 6 cum ( y i 4 ( k ) ) ] y i 3 ( k ) + [ 3 4 cum ( y i 3 ( k ) ) cum ( y i 4 ( k ) ) - 1 2 cum ( y i 3 ( k ) ) - 7 4 cum 3 ( y i 3 ( k ) ) ] y i 2 ( k ) } x j ( k ) - - - ( 6 )
In formula (6), the item in braces is only y i(k) function, makes it be
f ( y i ( k ) ) = [ 1 2 cum 2 ( y i 3 ( k ) ) - 1 6 cum ( y i 4 ( k ) ) ] y i 3 ( k ) + [ 3 4 cum ( y i 3 ( k ) ) cum ( y i 4 ( k ) ) - 1 2 cum ( y i 3 ( k ) ) - 7 4 cum 3 ( y i 3 ( k ) ) ] y i 2 ( k ) - - - ( 7 )
At this moment, formula (6) can be written as
( ▿ R J ) ij = - ( W - 1 ( k ) ) ij T + f ( y i ( k ) ) x j ( k ) - - - ( 8 )
Here it should be noted that the f (y that formula (7) provides i(k) being) very concrete functional form, is that the Edgeworth of probability density launches and get fourth order cumulant to obtain, and is referred to as excitation function.In practice, excitation function form can be determined as required, is also just to say, f (y i(k)) be y i(k) which kind of concrete form, can choose according to actual needs.
J (k) asks instantaneous gradient to all elements in W (k), obtains J (k) and to the instantaneous gradient of W (k) is
▿ R J = ∂ J ( k ) ∂ W ( k ) = - [ ( W - 1 ( k ) ) T - f ( Y ( k ) ) X T ( k ) ] - - - ( 9 )
Utilize random gradient and natural gradient relation, obtain J (k) and to the natural gradient of W (k) be
▿ J = ▿ R J · W T ( k ) W ( k ) = - [ ( W - 1 ( k ) ) T - f ( Y ( k ) ) X T ( k ) ] W T ( k ) W ( k ) = - [ I - f ( Y ( k ) ) Y T ( k ) ] W ( k ) - - - ( 10 )
In formula, represent the natural gradient of J (k) to W (k); I representation unit matrix; F (Y (k)) is actual is the non-linear activation primitive being determined by probability density characteristics.Under natural gradient criterion (Natural Gradient Algorithm, NGA), the more new formula of separation matrix W (k) is
W ( k + 1 ) = W ( k ) - μ · ▿ J = W ( k ) + μ [ I - f ( Y ( k ) ) Y T ( k ) ] W ( k ) - - - ( 11 )
In formula, μ represents step-length, is constant; I represents a unit matrix.The conventional cross-talk error criterion of blind source separation method performance of this traditional piece-rate system is weighed, and its expression is
PI ( C ( k ) ) = Σ i = 1 M [ ( Σ l = 1 M | c il | max j = 1 M ( c ij ) - 1 ) ] + Σ i = 1 M [ ( Σ l = 1 M | c li | max j = 1 M ( c ji ) - 1 ) ] - - - ( 12 )
In formula, PI represents cross-talk error; Max represents to get maxima operation; C (k)=W (k) A=P (k) Λ is the composite matrix of traditional piece-rate system, c ilwhat represent is the element of the capable l row of Matrix C (k) i, and P (k) and Λ represent respectively a permutation matrix and diagonal matrix, and P (k), Λ, C (k) are M × Metzler matrix.If hybrid system matrix A is known, the more new formula of taking advantage of A to obtain overall Matrix C (k) by formula (11) right side is
C(k+1)=C(k)+μ[I-f(Y(k))Y T(k)]C(k) (13)
But in actual conditions, because hybrid system matrix A is unknown, more new formula that therefore can not through type (13) directly obtains C (k), also just can not get PI (C (k)).
In addition, traditional adaptive blind source separation algorithm adopts fixed step size, therefore has the contradiction between convergence rate and cross-talk error.Chinese scholars has successively proposed some variable step size methods, but also there is no a kind of blind source separation scheme that very effectively solves contradiction between convergence rate and cross-talk error at present.
Summary of the invention
For addressing the above problem, the invention discloses a kind of improved self-adaptive blind source separation method NVS-NGA, effectively solved the contradiction between convergence rate and cross-talk error, obtained good separating effect.
In order to achieve the above object, the invention provides following technical scheme:
A kind of improved self-adaptive blind source separation method, based on improved adaptive blind source separation system, realize, described improved adaptive blind source separation system comprises hybrid matrix A, separation matrix W (k) and the inverse system W in parallel with separation matrix W (k) a(k), described W a(k) with the contrary A of nonsingular hybrid matrix A -1approximate, described blind source separation method comprises the steps:
Steps A, M the unknown and source signal S independent of each other (k)=[s 1(k), s 2(k) ..., s m(k)] tthrough improving unknown nonsingular hybrid matrix A in piece-rate system, be mixed to get observation signal X (k)=[x 1(k), x 2(k) ..., x m(k)] t; When ignoring transmission delay effect and noise, obtain X (k)=AS (k), k is time series, subscript T represents conjugate transpose; M is positive integer, represents the number of component in S (k); A is that M × M ties up matrix;
Step B, observation signal X (k)=[x that steps A is obtained 1(k), x 2(k) ..., x m(k)] tsend into simultaneously and improve separation matrix W (k) and W in piece-rate system a(k), obtain respectively separation signal Y (k)=W (k) X (k) and Y a(k)=W a(k) X (k), wherein Y (k) is M × 1 dimensional vector, is an estimation of source signal S (k), its component is separate; W (k) is the final separation matrix that improves M × M dimension full rank in piece-rate system; W a(k) contrary
Figure BDA0000458465520000041
it is the final estimation to improving unknown nonsingular hybrid matrix A in piece-rate system; W (k) and W a(k) dimension is identical, and subscript " 1 " represents to get inverse operation;
Wherein separation matrix W (k) and W a(k) more new formula is:
W ( k + 1 ) = W ( k ) + μ ( k ) [ I - f ( Y ( k ) ) Y T ( k ) ] W ( k ) W a ( k + 1 ) = W a ( k ) + μ ( k ) [ I - f ( Y ( k ) ) Y T ( k ) ] W a ( k ) ,
Wherein I representation unit matrix; F (y (k)) is nonlinear activation primitive, and μ (k) is improved step factor, and its formula is:
μ(k)=β{|PI(C G(k))|/α 2}exp{-[PI(C G(k))] 2/(2α 2)},
Wherein, β and α are the control parameters of μ (k); Exp represents the exponential function take e the end of as, wherein PI (C g(k) be) improved separating property index, its formula is:
PI ( C G ( k ) ) = Σ i = 1 M [ ( Σ l = 1 M | c Gil | max j = 1 M ( c Gij ) - 1 ) ] + Σ i = 1 M [ ( Σ l = 1 M | c Gli | max j = 1 M ( c Gji ) - 1 ) ] ,
Wherein C g(k) for improving the overall matrix of piece-rate system,
Figure BDA0000458465520000047
c ilfor the element that the capable l of i of Matrix C (k) is listed as, c gilit is Matrix C g(k) the capable l column element of i, max represents to get maxima operation.
The initial matrix that improves piece-rate system is W (0) and W a(0), wherein
Figure BDA0000458465520000044
improve the initial overall matrix of piece-rate system C G ( 0 ) = W ( 0 ) W a - 1 ( 0 ) .
Preferred as one of the present invention, described f (Y (k))=Y 3(k).
Concrete, described separation matrix W (k) and W a(k) more new formula obtains as follows:
Step a, utilizes the relation of mutual information and comentropy, and the cost function of piece-rate system is defined as
J ( k ) = H ( y m ( k ) ) - ln | det ( W ( k ) ) | = - Σ m = 1 M E ( ln p y ( y m ( k ) ) ) - ln | det ( W ( k ) ) | ,
In formula, H (y m(k)) be m separation signal y in separation signal vector Y (k) m(k) entropy, p y(y m(k)) be m separation signal y in separation signal vector Y (k) m(k) marginal probability density, E represents mathematic expectaion computing, and ln represents the natural logrithm take e the end of as, and det represents to get the determinant of W (k);
Step b, calculates the natural gradient of J (k) to W (k)
▿ J = ∂ J ( k ) ∂ W ( k ) · W T ( k ) W ( k ) = - [ I - f ( Y ( k ) ) Y T ( k ) ] W ( k ) ;
Step c, obtains separation matrix W (k) and W by J (k) to the natural gradient of W (k) a(k) more new formula
W ( k + 1 ) = W ( k ) + μ ( k ) [ I - f ( Y ( k ) ) Y T ( k ) ] W ( k ) W a ( k + 1 ) = W a ( k ) + μ ( k ) [ I - f ( Y ( k ) ) Y T ( k ) ] W a ( k ) .
Beneficial effect: first the present invention improves the structure of traditional piece-rate system, the performance index that blinder source separated are improved, take improved separating property index as independent variable, take Rayleigh Distribution Function as dependent variable, take Rayleigh Distribution Function as variable step function.Compared with traditional blind source separation method, this method can fast and effeciently be isolated primary signal from the mixed signal receiving, and has effectively solved the contradiction between convergence rate and cross-talk error; Not only fast convergence rate, cross-talk error are little, and stability is strong; At aspects such as radio communication, image processing, voice signal processing, all have wide practical use.
Accompanying drawing explanation
Fig. 1 is traditional adaptive blind source separation system structural representation;
Fig. 2 is improved self-adaptive blind source separation method schematic diagram provided by the invention;
Fig. 3 is the convergence graph of the lower three kinds of separating property indexs of Stationary Random Environments condition;
Fig. 4 is simulation result figure of the present invention, and wherein (a) is source signal figure, is (b) mixed signal figure;
Fig. 5 is separating resulting figure under Stationary Random Environments condition, wherein (a) is for being used the separation signal figure after NGA tests, (b) for using the separation signal figure after VS-NGA tests, (c) for using the separation signal figure after NVS-NGA tests;
Fig. 6 is PI mean value curve under Stationary Random Environments condition;
Fig. 7 is the PI mean value curve under non-stationary environment condition.
Embodiment
Below with reference to specific embodiment, technical scheme provided by the invention is elaborated, should understands following embodiment and only for the present invention is described, is not used in and limits the scope of the invention.
As shown in Figure 2, first the present invention improves the structure of former piece-rate system, in traditional piece-rate system separation matrix W (k) upper in parallel one with the contrary A of nonsingular hybrid matrix A -1approximate inverse system W a(k), matrix W a(k), W a(k) be called separation matrix in parallel, this improves structure division and is concatenated into nonsingular hybrid matrix A again; And the performance index that blind source is separated are improved, independent variable using improved separating property index as Rayleigh Distribution Function, using Rayleigh Distribution Function as variable step function, thereby invented a kind of improved self-adaptive blind source separation method NVS-NGA, this method comprises the steps:
Steps A, M the unknown and source signal S independent of each other (k)=[s 1(k), s 2(k) ..., s m(k)] tthrough improving unknown nonsingular hybrid matrix A in piece-rate system, be mixed to get observation signal X (k)=[x 1(k), x 2(k) ..., x m(k)] t; When ignoring transmission delay effect and noise, obtain X (k)=AS (k), k is time series, subscript T represents conjugate transpose; M is positive integer, represents the number of component in S (k); A is that M × M ties up matrix.
Step B, observation signal X (k)=[x that steps A is obtained 1(k), x 2(k) ..., x m(k)] tsend into simultaneously and improve separation matrix W (k) and W in piece-rate system a(k), obtain respectively separation signal Y (k)=W (k) X (k) and Y a(k)=W a(k) X (k), wherein Y (k) is M × 1 dimensional vector, is an estimation of source signal S (k), its component is separate; W (k) is the final separation matrix that improves M × M dimension full rank in piece-rate system; W a(k) contrary
Figure BDA0000458465520000061
it is the final estimation to improving unknown nonsingular hybrid matrix A in piece-rate system; W (k) and W a(k) dimension is identical, and subscript " 1 " represents to get inverse operation.
Based on improved adaptive blind source separation system, first we want the blind source separating property index of computed improved:
Owing to improving separating part W (k) and W in piece-rate system (as shown in Figure 2) a(k) be parallel-connection structure, when improving after piece-rate system operation, will obtain best separation matrix W in parallel a(k); Now W a(k) contrary
Figure BDA0000458465520000062
as being similar to of hybrid matrix A in improvement piece-rate system, the overall matrix of the piece-rate system that is improved is thus designated as C g(k), and
C G ( k ) ≈ W ( k ) W a - 1 ( k ) - - - ( 14 )
By the overall Matrix C of improving piece-rate system g(k) C (k) in substituted (12), obtains separating property index parameter and is
PI ( C G ( k ) ) = Σ i = 1 M [ ( Σ l = 1 M | c Gil | max j = 1 M ( c Gij ) - 1 ) ] + Σ i = 1 M [ ( Σ l = 1 M | c Gli | max j = 1 M ( c Gji ) - 1 ) ] - - - ( 15 )
This separating property index is that the one of formula (12) is improved, and is referred to as to improve separating property index.
In traditional self-adaptive blind source separation method, step size mu is fixed value, and tracking performance is poor, is unfavorable for solving blind source and separates the contradiction between convergence rate and cross-talk error.In order effectively to solve this technical problem, the present invention combines PI value and step factor, builds a variable step function, specifically, is with cross-talk error PI (C g(k)) be independent variable, take Rayleigh Distribution Function as dependent variable, obtain variable step formula and be
μ(k)=β{|PI(C G(k))|/α 2}exp{-[PI(C G(k))] 2/(2α 2)} (16)
In formula, β and α are the control parameters of μ (k), can choose by experiment α=10, β=0.1; Exp represents the exponential function take e the end of as.
Based on the variable step function shown in formula (16), adopt natural gradient criterion, separation matrix W (k) and W in the piece-rate system that is improved a(k) more new formula, concrete steps are as follows:
Utilize the relation of mutual information and comentropy, the cost function of piece-rate system is defined as
J ( k ) = H ( y m ( k ) ) - ln | det ( W ( k ) ) | = - Σ m = 1 M E ( ln p y ( y m ( k ) ) ) - ln | det ( W ( k ) ) |
In formula, H (y m(k)) be m separation signal y in separation signal vector Y (k) m(k) entropy, p y(y m(k)) be m separation signal y in separation signal vector Y (k) m(k) marginal probability density, E represents mathematic expectaion computing, and ln represents the natural logrithm take e the end of as, and det represents to get the determinant of W (k);
Calculate the natural gradient of J (k) to W (k)
▿ J = ∂ J ( k ) ∂ W ( k ) · W T ( k ) W ( k ) = - [ I - f ( Y ( k ) ) Y T ( k ) ] W ( k )
By J (k), the natural gradient of W (k) is obtained to separation matrix W (k) and W a(k) more new formula
W ( k + 1 ) = W ( k ) + μ ( k ) [ I - f ( Y ( k ) ) Y T ( k ) ] W ( k ) W a ( k + 1 ) = W a ( k ) + μ ( k ) [ I - f ( Y ( k ) ) Y T ( k ) ] W a ( k ) - - - ( 17 )
Wherein, μ (k) is the improved step factor shown in formula (16); I representation unit matrix; F (Y (k)) is nonlinear activation primitive, if determine nonlinear activation primitive f (Y (k)) by formula (7), this function is
f ( Y ( k ) ) = [ 1 2 cum 2 ( Y 3 ( k ) ) - 1 6 cum ( Y 4 ( k ) ) ] Y 3 ( k ) + [ 3 4 cum ( Y 3 ( k ) ) cum ( Y 4 ( k ) ) - 1 2 cum ( Y 3 ( k ) ) - 7 4 cum 3 ( Y 3 ( k ) ) ] Y 2 ( k )
By formula (18), determine f (Y (k)), need to calculate the three rank cumulant cum (y of y (k) 3) and fourth order cumulant cum (y (k) 4(k)), calculate very complicatedly, be unfavorable for engineering application.In order to reduce amount of calculation, nonlinear activation primitive is preferably to the reduced form of formula (18), i.e. f (Y (k))=Y 3(k); By separation matrix W in parallel aand the overall Matrix C that obtains of hybrid matrix A (k) a(k)=W a(k) A=P a(k) Λ abe called the overall matrix of part in parallel, P aand Λ (k) arepresent respectively a permutation matrix and diagonal matrix, to W a(k) invert and can obtain
Figure BDA0000458465520000074
due to P aand Λ (k) arepresent respectively a permutation matrix and diagonal matrix, by matrix properties, can be obtained
Figure BDA0000458465520000075
with
Figure BDA0000458465520000076
also represent a permutation matrix and diagonal matrix.
At the initial matrix W (0) and the W that improve piece-rate system a(0) in choosing, initial matrix W a(0) choose need satisfy condition: choose suitable W a(0), after, improve the initial overall matrix of piece-rate system
Figure BDA0000458465520000078
by PI (C g(0)) the stepsize formula of substitution Rayleigh Distribution Function form, obtains initial step length μ (0).
Up to the present, there are the separating property index PI (C (k)) of traditional piece-rate system, the separating property index PI (C of part in parallel a) and improve the separating property index PI (C of piece-rate system (k) g(k)).As shown in Figure 3, Fig. 3 shows their constringency performance, and in the whole separation process under Stationary Random Environments condition, at the separation initial stage, the value of PI is very large, illustrates that cross-talk error is very large; In the latter stage separating, it is very little that the value of PI can become, and illustrates that cross-talk error is less, and separating effect is better.On the whole, PI value presents a kind of fast-descending to restraining trend stably in iterative process, wherein improved separating property index PI (C g(k) performance) is best, therefore adopts and improves separating property index PI (C g(k)) can better solve the contradiction between convergence rate and cross-talk error.
In order to verify the validity of the inventive method (referred to as NVS-NGA), take traditional blind source separation method VS-NGA of traditional blind source separation method NGA of fixed step size, variable step as comparison other, carry out emulation experiment comparison by Matlab program.In experiment, source signal is: S1=sign (cos (2*pi*155*t/fs)); S2=sin (2*pi*800*t/fs); S3=sin (2*pi*90*t/fs); S4=sin (2*pi*9*t/fs) * sin (2*pi*300*t/fs), sampling number is 5000, sample frequency is 10000Hz, hybrid matrix A 0=[0.3702 0.7143-0.6188-0.3002; 0.0965 0.7408 0.9365 0.0443; 0.3732-0.3762-0.2500-0.6735;-0.6674 0.6747 0.9162-0.7381]; The fixed step size of NGA is μ=0.003; In VS-NGA, utilize sigmoid function to make variable step
Figure BDA0000458465520000081
wherein η (k)=|| I-f (Y (k)) Y t(k) || be I-f (Y (k)) Y t(k) norm, α=10, β=0.1; In NVS-NGA, variable step is formula (16), parameter alpha=10, β=0.15.Wherein source signal waveform is as shown in Fig. 4 (a), and mixed waveform signal is as shown in Fig. 4 (b).
Embodiment 1: under Stationary Random Environments condition, i.e. hybrid matrix A=A 0shi Jinhang contrast experiment.In experiment, all parameters are got above-mentioned set-point, and 100 Monte-Carlo Simulation results are respectively as shown in Fig. 5 (a), Fig. 5 (b), Fig. 5 (c) and Fig. 5 (d).Fig. 5 (a), Fig. 5 (b) and Fig. 5 (c) correspond respectively to NGA method, VS-NGA method and NVS-NGA method gained separation signal figure of the present invention, and Fig. 6 is PI mean value curve chart.Fig. 6 shows, NGA method restrains when 2700 step, VS-NGA restrains about 2300 steps, and the inventive method NVS-NGA restrains about 1500 steps, also shows to have minimum cross-talk error after the inventive method NVS-NGA convergence.Therefore, the separating property of the inventive method NVS-NGA is best, has minimum cross-talk error and rapid convergence speed.
Embodiment 2: test under non-stationary environment, hybrid matrix is A (k)=A 0+ B (k), B (k)=ρ B (k-1)+τ * randn (4), B (0) is the null matrix of 4 × 4, ρ=0.9, τ=0.0001, randn (4) is the random matrix on 4 rank, other condition is constant, the experiment of 100 Monte-Carlo Simulation PI mean value curve, as shown in Figure 7.Fig. 7 shows, compared with NGA, VS-NGA method, the inventive method NVS-NGA still has the fastest convergence rate and minimum cross-talk error, and therefore, the separating property of the inventive method NVS-NGA is also best.
Known from above-described embodiment, improved self-adaptive blind source separation method provided by the invention, compares with existing fixed step size and variable-step self-adaptive blind source separation method, has improved convergence rate, and has reduced cross-talk error, and separating property has obvious lifting.
The disclosed technological means of the present invention program is not limited only to the disclosed technological means of above-mentioned execution mode, also comprises the technical scheme being comprised of above technical characterictic combination in any.It should be pointed out that for those skilled in the art, under the premise without departing from the principles of the invention, can also make some improvements and modifications, these improvements and modifications are also considered as protection scope of the present invention.

Claims (4)

1. an improved self-adaptive blind source separation method, it is characterized in that: based on improved adaptive blind source separation system, realize, described improved adaptive blind source separation system comprises hybrid matrix A, separation matrix W (k) and the inverse system W in parallel with separation matrix W (k) a(k), described W a(k) with the contrary A of nonsingular hybrid matrix A -1approximate, described blind source separation method comprises the steps:
Steps A, M the unknown and source signal S independent of each other (k)=[s 1(k), s 2(k) ..., s m(k)] tthrough improving unknown nonsingular hybrid matrix A in piece-rate system, be mixed to get observation signal X (k)=[x 1(k), x 2(k) ..., x m(k)] t; When ignoring transmission delay effect and noise, obtain X (k)=AS (k), k is time series, subscript T represents conjugate transpose; M is positive integer, represents the number of component in S (k); A is that M × M ties up matrix;
Step B, observation signal X (k)=[x that steps A is obtained 1(k), x 2(k) ..., x m(k)] tsend into simultaneously and improve separation matrix W (k) and W in piece-rate system a(k), obtain respectively separation signal Y (k)=W (k) X (k) and Y a(k)=W a(k) X (k), wherein Y (k) is M × 1 dimensional vector, is an estimation of source signal S (k), its component is separate; W (k) is the final separation matrix that improves M × M dimension full rank in piece-rate system; W a(k) contrary
Figure FDA0000458465510000011
it is the final estimation to improving unknown nonsingular hybrid matrix A in piece-rate system; W (k) and W a(k) dimension is identical, and subscript " 1 " represents to get inverse operation;
Wherein separation matrix W (k) and W a(k) more new formula is:
W ( k + 1 ) = W ( k ) + μ ( k ) [ I - f ( Y ( k ) ) Y T ( k ) ] W ( k ) W a ( k + 1 ) = W a ( k ) + μ ( k ) [ I - f ( Y ( k ) ) Y T ( k ) ] W a ( k ) ,
Wherein I representation unit matrix; F (Y (k)) is nonlinear activation primitive, and μ (k) is improved step factor, and its formula is:
μ(k)=β{|PI(C G(k))|/α 2}exp{-[PI(C G(k))] 2/(2α 2)},
Wherein, β and α are the control parameters of μ (k), and exp represents the exponential function take e the end of as, wherein PI (C g(k) be) improved separating property index, its formula is:
PI ( C G ( k ) ) = Σ i = 1 M [ ( Σ l = 1 M | c Gil | max j = 1 M ( c Gij ) - 1 ) ] + Σ i = 1 M [ ( Σ l = 1 M | c Gli | max j = 1 M ( c Gji ) - 1 ) ] ,
Wherein C g(k) for improving the overall matrix of piece-rate system,
Figure FDA0000458465510000014
c ilfor the element that the capable l of i of Matrix C (k) is listed as, c gilit is Matrix C g(k) the capable l column element of i, max represents to get maxima operation.
2. improved self-adaptive blind source separation method according to claim 1, is characterized in that: the initial matrix that improves piece-rate system is W (0) and W a(0), wherein
Figure FDA0000458465510000021
improve the initial overall matrix of piece-rate system
Figure FDA0000458465510000022
3. improved self-adaptive blind source separation method according to claim 1 and 2, is characterized in that: described f (Y (k))=Y 3(k).
4. improved self-adaptive blind source separation method according to claim 1 and 2, is characterized in that: described separation matrix W (k) and W a(k) more new formula obtains as follows:
Step a, utilizes the relation of mutual information and comentropy, and the cost function of piece-rate system is defined as
J ( k ) = H ( y m ( k ) ) - ln | det ( W ( k ) ) | = - Σ m = 1 M E ( ln p y ( y m ( k ) ) ) - ln | det ( W ( k ) ) | ,
In formula, H (y m(k)) be m separation signal y in separation signal vector Y (k) m(k) entropy, p y(y m(k)) be m separation signal y in separation signal vector Y (k) m(k) marginal probability density, E represents mathematic expectaion computing, and ln represents the natural logrithm take e the end of as, and det represents to get the determinant of W (k);
Step b, calculates the natural gradient of J (k) to W (k)
▿ J = ∂ J ( k ) ∂ W ( k ) · W T ( k ) W ( k ) = - [ I - f ( Y ( k ) ) Y T ( k ) ] W ( k ) ;
Step c, obtains separation matrix W (k) and W by J (k) to the natural gradient of W (k) a(k) more new formula
W ( k + 1 ) = W ( k ) + μ ( k ) [ I - f ( Y ( k ) ) Y T ( k ) ] W ( k ) W a ( k + 1 ) = W a ( k ) + μ ( k ) [ I - f ( Y ( k ) ) Y T ( k ) ] W a ( k ) .
CN201410022179.0A 2014-01-17 2014-01-17 A kind of improved self-adaptive blind source separation method Expired - Fee Related CN103763230B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201410022179.0A CN103763230B (en) 2014-01-17 2014-01-17 A kind of improved self-adaptive blind source separation method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201410022179.0A CN103763230B (en) 2014-01-17 2014-01-17 A kind of improved self-adaptive blind source separation method

Publications (2)

Publication Number Publication Date
CN103763230A true CN103763230A (en) 2014-04-30
CN103763230B CN103763230B (en) 2017-03-01

Family

ID=50530392

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201410022179.0A Expired - Fee Related CN103763230B (en) 2014-01-17 2014-01-17 A kind of improved self-adaptive blind source separation method

Country Status (1)

Country Link
CN (1) CN103763230B (en)

Cited By (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104333523A (en) * 2014-10-14 2015-02-04 集美大学 NPCA-based post nonlinear blind source separation method
CN104464749A (en) * 2014-11-24 2015-03-25 沈阳化工大学 Blind-extraction-based method for treating cooling tower noise
CN104660530A (en) * 2015-02-02 2015-05-27 中国人民解放军理工大学 Method for separating time-frequency domain mixed signals in noise background
CN104866866A (en) * 2015-05-08 2015-08-26 太原理工大学 Improved natural gradient variable step-size blind source separation algorithm
CN106534009A (en) * 2016-11-29 2017-03-22 安徽理工大学 Improved variable step size equivariant adaptive blind source separation method
CN106803089A (en) * 2016-12-15 2017-06-06 南京邮电大学 Method based on non-linear principal component analysis separate picture information from image sequence
CN106887238A (en) * 2017-03-01 2017-06-23 中国科学院上海微***与信息技术研究所 A kind of acoustical signal blind separating method based on improvement Independent Vector Analysis algorithm
CN108574525A (en) * 2018-04-16 2018-09-25 南京理工大学 A kind of spaceborne AIS collision alarms separation method based on adaptability moments estimation
CN109086686A (en) * 2018-07-12 2018-12-25 西安电子科技大学 Blind source separation method under time varying channel based on self-adapted momentum factor
CN112637742A (en) * 2020-12-29 2021-04-09 北京安声浩朗科技有限公司 Signal processing method and signal processing device, storage medium and earphone
CN113096684A (en) * 2021-06-07 2021-07-09 成都启英泰伦科技有限公司 Target voice extraction method based on double-microphone array
CN116776076A (en) * 2023-07-06 2023-09-19 中国人民解放军空军预警学院 Underdetermined blind source separation method and system for third-order cumulant and tensor compression

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2008141306A2 (en) * 2007-05-11 2008-11-20 Sigmed, Inc. Non-invasive characterization of a physiological parameter
CN101819782A (en) * 2010-03-10 2010-09-01 重庆邮电大学 Variable-step self-adaptive blind source separation method and blind source separation system

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2008141306A2 (en) * 2007-05-11 2008-11-20 Sigmed, Inc. Non-invasive characterization of a physiological parameter
CN101819782A (en) * 2010-03-10 2010-09-01 重庆邮电大学 Variable-step self-adaptive blind source separation method and blind source separation system

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
侯瑞玲: "自适应盲源分离与盲多用户检测算法研究", 《中国优秀硕士学位论文全文数据库信息科技辑》 *
张银雪等: "基于改进人工蜂群算法的盲源分离方法", 《电子学报》 *

Cited By (16)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104333523A (en) * 2014-10-14 2015-02-04 集美大学 NPCA-based post nonlinear blind source separation method
CN104464749A (en) * 2014-11-24 2015-03-25 沈阳化工大学 Blind-extraction-based method for treating cooling tower noise
CN104660530A (en) * 2015-02-02 2015-05-27 中国人民解放军理工大学 Method for separating time-frequency domain mixed signals in noise background
CN104660530B (en) * 2015-02-02 2017-07-04 中国人民解放军理工大学 The method that time-frequency domain mixed signal is separated under noise background
CN104866866A (en) * 2015-05-08 2015-08-26 太原理工大学 Improved natural gradient variable step-size blind source separation algorithm
CN106534009A (en) * 2016-11-29 2017-03-22 安徽理工大学 Improved variable step size equivariant adaptive blind source separation method
CN106803089B (en) * 2016-12-15 2020-03-31 南京邮电大学 Method for separating image information from image sequence based on nonlinear principal component analysis
CN106803089A (en) * 2016-12-15 2017-06-06 南京邮电大学 Method based on non-linear principal component analysis separate picture information from image sequence
CN106887238A (en) * 2017-03-01 2017-06-23 中国科学院上海微***与信息技术研究所 A kind of acoustical signal blind separating method based on improvement Independent Vector Analysis algorithm
CN108574525A (en) * 2018-04-16 2018-09-25 南京理工大学 A kind of spaceborne AIS collision alarms separation method based on adaptability moments estimation
CN108574525B (en) * 2018-04-16 2020-07-07 南京理工大学 Satellite-borne AIS collision signal separation method based on adaptive moment estimation
CN109086686A (en) * 2018-07-12 2018-12-25 西安电子科技大学 Blind source separation method under time varying channel based on self-adapted momentum factor
CN109086686B (en) * 2018-07-12 2022-09-30 西安电子科技大学 Blind source separation method under time-varying channel based on self-adaptive momentum factor
CN112637742A (en) * 2020-12-29 2021-04-09 北京安声浩朗科技有限公司 Signal processing method and signal processing device, storage medium and earphone
CN113096684A (en) * 2021-06-07 2021-07-09 成都启英泰伦科技有限公司 Target voice extraction method based on double-microphone array
CN116776076A (en) * 2023-07-06 2023-09-19 中国人民解放军空军预警学院 Underdetermined blind source separation method and system for third-order cumulant and tensor compression

Also Published As

Publication number Publication date
CN103763230B (en) 2017-03-01

Similar Documents

Publication Publication Date Title
CN103763230A (en) Improved self-adaption blind source separation method
CN104537249A (en) Direction of arrival angle estimation method based on Sparse Bayesian learning
CN103777190B (en) A kind of angle estimating method of bistatic MIMO radar high speed highly maneuvering target
CN105182293A (en) Method for estimating DOA and DOD of MIMO radar based on co-prime array
CN106021637A (en) DOA estimation method in co-prime array based on iteration sparse reconstruction
CN107678028A (en) Microwave under Low SNR stares relevance imaging method
CN114338301B (en) Channel estimation method of RIS auxiliary millimeter wave system based on compressed sensing
CN104539340B (en) A kind of sane direction of arrival estimation method being fitted based on rarefaction representation and covariance
CN103605121B (en) Wideband radar data fusion method based on rapid sparse Bayesian learning algorithm
CN105069235A (en) Method for extracting equivalent circuit parameter of dual-frequency-band frequency selective surface
CN104869086A (en) MIMO-OFDM communication system downlink channel estimation method based on two-dimensional compression perception and device thereof
CN107576931A (en) A kind of correlation based on the sparse reconstruct of covariance low dimensional iteration/coherent signal Wave arrival direction estimating method
CN106842163A (en) A kind of Ballistic Target echo-signal time-frequency characteristic method of estimation
CN107852387A (en) Reduce the method that pre-coding matrix calculates in extensive multi-input multi-output system and user equipment is grouped complexity
CN107037409A (en) MIMO radar waveform separation method based on compressed sensing
CN107121662A (en) Single passive location method based on spatial domain rarefaction representation
CN109507636A (en) Wave arrival direction estimating method based on virtual Domain signal reconstruction
CN106130936A (en) A kind of non linear channel equalization method under Alpha Stable distritation noise circumstance
CN104898093A (en) Gerchberg-Saxton (GS) algorithm-based multiple-input multiple-output (MIMO) radar quadrature phase coded signal design method
CN111007457A (en) Radiation source direct positioning method based on block sparse Bayesian model
CN105743511A (en) Compressed sensing signal reconstruction method based on conjugate gradient method
CN103475608B (en) Simulated annealing and fruit bat hybrid optimization small echo GENERALIZED DISCRETE LINEAR RANDOM SYSTEM multi-mode blind equalization method
CN105068090A (en) Method for suppressing interference to GNSS antenna array by using single snapshot data
CN104914451A (en) Space-time anti-interference method for block Toeplitz matrix low-complexity inversion
CN105182306A (en) Stable two-dimensional direction of arrival estimation method based on sparse representation

Legal Events

Date Code Title Description
C06 Publication
PB01 Publication
C10 Entry into substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant
CF01 Termination of patent right due to non-payment of annual fee
CF01 Termination of patent right due to non-payment of annual fee

Granted publication date: 20170301

Termination date: 20220117