CN103929210B - Hard decision decoding method based on genetic algorithm and neural network - Google Patents

Abstract

The invention relates to the field of signal processing in communication, in particular to a hard decision decoding method based on a genetic algorithm and neural network, namely, a genetic neural-network decoding (GND) method. The method makes full use of self-optimizing capability of the genetic algorithm and a model classifying function of the neural network to perform optimizing processing on output of hard decision quantization of a matching filter in order to make up for a reliability loss of decoding caused by transmission errors of a channel and hard decision quantization, and therefore, code words more similar to a transmission sequence are recovered out to serve as input of a hard decision decoder in order to obtain a better decoding result. As you can see from theoretical analysis and computer analog simulation, error correction performance of the GND method is similar to that of a traditional soft decision decoding method. Due to the fact that no soft information is needed to be calculated through the channel during decoding, compared with the traditional soft decision decoding method, complexity of the GND method is reduced greatly.

Description

A kind of Hard decision decoding method based on genetic algorithm Yu neutral net

Technical field

The present invention relates to signal processing field in communication, particularly to Hard decision decoding method, the method is based on heredity Algorithm (Genetic Algorithm, GA) realizes with neutral net (Neural Network, NN).

Background technology

At present, error-correcting code technique has had become as indispensable means and the method for the timely reliable communication of realization.But, There is the problems such as the scope of application is little, computation complexity is high in the Soft decision decoding technology of error correcting code, it is difficult in existing skill the most always Under the conditions of art, it is settled properly in more rational finite time.In addition, general decoding algorithm is all serial Process, be appropriate only for low, the digital communication system of middling speed.Digital communication at present and information storage system are just towards high speed, height Bandwidth, high reliability direction are developed, and error correcting code proposes new requirement, and decoding problem has become the one of error correcting code development the most Big bottleneck.

It is a class NP (Non-deterministic that Berlekamp et al. has been proved the decoding problem of general error correcting code Polynomial) challenge, can be equivalent to combinatorial optimization problem and process.Intelligent algorithm (Intelligent Algorithm, IA) as a kind of information being obtained solution complex combination optimization problem by the inherent adaptive optimization mechanism in the natural imitation world Treatment technology has been introduced in error-correcting code technique.Utilize the mechanism such as the adaptive optimization of intelligent algorithm and fast parallel process The difficulty that solution error correcting code decoding technique is faced has most important theories meaning and practical value.

Consult pertinent literature to understand, at present GA decoding algorithm and neural network algorithm are each used alone in error correcting code The research carrying out Hard decision decoding is the most, but on the whole, its decoding combination property is the best, or complexity is high, or entangles Wrong performance is the best.This is because for complicated optimum problem, the intelligent algorithm of single mechanism is difficulty with global optimization, and efficiency Low.Maximize favourable factors and minimize unfavourable ones by mixing different intelligent algorithms, the NP double linear problems of difficulty for solving in science and technology and engineering field can be efficiently solved. For genetic algorithm, its ability of searching optimum is good, but when being used alone, is difficult to accomplish between convergence rate and constringency performance Good is compromise, and the another kind of popular algorithm of neural network algorithm the convergence speed in intelligent algorithm is quickly, but selects parameter The harshness selected.Therefore, if the two combination being applied in decoding algorithm, drawback of can making the best use of the advantages, obtaining and preferably solving knot Really.

Summary of the invention

In order to reduce the complexity of tradition Soft decision decoding, improve decoding speed simultaneously, the present invention provide a kind of based on Genetic algorithm and the Hard decision decoding method of neutral net, to make up because signal transmission errors and hard decision quantify to bring to decoding Reliability loss, thus recover one with transmit the code word of sequence more likelihood as the input of Hard decision decoding device to obtain One more preferable decoding result, improves its error correction.

The present invention is a kind of hybrid intelligent interpretation method, is properly termed as genetic neural network decoding (Genetic Neural- Network Decoding, GND) method, its method comprises the following steps:

(1) training neutral net:

(1.1) real number symbol sebolic addressing r (r is received₁,r₂,r₃...) firmly sentenced after demodulated device matched filtering tolerance Certainly sequence R；

(1.2) by hard decision sequence R obtained after demodulator matched filtering tolerance respectively with stochastic generation d_h/ 2 n dimensions Binary sequence T adds through mould 2 and produces d afterwards_h/ 2 candidate sequence A；

(1.3) training neutral net: neutral net is made up of three-layer network as a grader, i.e. input layer, implicit Layer and output layer, input layer is made up of individual n-k neuron, and output layer has 1 neuron, and hidden layer includes (2/3) (n-k+t+1) Individual neuron, wherein k is the information bit number of code, and t is the maximum error correction number of this yard；Training process is: by syndrome sequence As input training mode, weight w of corresponding error pattern is exported as target, is allowed to input a syndrome just (w=1,2,3..., n), syndrome S is according to representated by genetic algorithm individuality to obtain weight w of corresponding error pattern Code word B and the check matrix H of code obtain, i.e.

S=B H ' formula (1)；

(2) genetic algorithm optimization is used to obtain one and transmit the code word of sequence more likelihood:

(2.1) initialization of population: generate the binary vector of 2t n position as initial population

(2.1.1) first of population individual member P1: the hard decision sequence R (r that matched filtering device is exported₁,r₂,..., r_n) it is set to first individual member P1 of population:

$P_1=R(r_1,r_2,...,r_n),r_i=\left\{\begin{array}{cc}1,& q_i>0\\ 0,& q_i\&le;0\end{array}\right.0<i<n$

Wherein, Q (q₁,q₂,...,q_n) sequence of real numbers that quantifies for the unmatched wave filter hard decision that receives；

(2.1.2) the individual member Pi of other 2t-1 of population: by by the uniform binary Orders Corrected T randomly generated (t₁,t₂,...,t_2t-1) and hard decision sequence R be added obtain, it may be assumed that

P_i=mod (R+T, 2), 2≤i≤2^t, T=rand [0,1]；

(2.2) ideal adaptation degree is evaluated:

According to following formula, the fitness that genetic algorithm is individual is evaluated

Wherein, λ (P, Q) is correlation function, is used for the Euclidean distance calculating Genetic vectors Pi and receiving between sequence of real numbers Q, It is the most similar that reality that is individual and that receive belongs to sequence, then the value of λ is the biggest,

$\mathrm{\&lambda;}(P,Q)=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{p}_i\&CenterDot;q_i$

Weight (Error class (Indiv.)) is the output result of neutral net, and penalty to be obtained needs elder generation Calculate syndrome sequence S that heredity to be assessed is individual, then by S input neural network；

(2.3) natural selection: select outstanding from initial population based on roulette wheel selection or other systems of selection Body participates in heredity, and the individual selected probability of i-th is:

$p\left(P_i\right)=\frac{a\left(P_i\right)}{\underset{j=0}{\overset{N-1}{\mathrm{\&Sigma;}}}a\left(P_j\right)}$

(2.4) pairing intersect: the individuality chosen will match at random, by by own partial element (code element) with right Side's intersection produces new individual；

(2.5) hereditary variation: produce in random selection process (2.4) is new individual, and it is carried out variation process, process side Method is, by certain individual bit element (code element) upset, i.e. by 0 → 1 or 1 → 0；

(2.6) heredity terminates: heredity will terminate when genetic generation number reaches preset value, and now in population, fitness is the highest Individuality will be output, if generation number is not up to, jump procedure (2.3) continues genetic process.

(3) optimal sequence input hard decision error correcting deocder genetic algorithm exported decodes, and is finally decoded Result.

Knowable to above method, neutral net is joined heredity as to the supplementary of genetic algorithm optimization performance by the present invention In the ideal adaptation degree evaluation mechanism of algorithm, in Fitness analysis mechanism, neutral net serves as the angle of a pattern classifier Color, it according to the code word representated by genetic algorithm individuality and recently Hamming distance between available codeword heredity individuality carried out point Class, the hereditary individuality identical with Hamming distance between nearest code word is divided into a class.This operation utilizes code in coding standards battle array One-to-one relationship between word syndrome and coset leader, is reflected syndrome sequence individual for the heredity of input by neutral net Penetrate the weight for corresponding coset leader (weight of coset leader) to realize.The result that neutral net obtains will as compensate because of Son joins in the evaluation mechanism of genetic algorithm, to strengthen the optimization performance of genetic algorithm further.

Therefore, the present invention makes full use of the self-optimized ability of genetic algorithm and the pattern classification function of neutral net to reception The hard decision of matched filtering device quantifies output and is optimized process, to make up because signal transmission errors and hard decision quantify to decoding Bring reliability loss, thus recover one with transmit sequence more likelihood code word as Hard decision decoding device input with Obtain a more preferable decoding result.Can be seen that from theory analysis and computer simulation emulation, its error-correcting performance is close to traditional soft Judgement decoding；And its decoding process need not utilize channel statistical Soft Inform ation, and the relatively conventional Soft decision decoding of its complexity is big Amplitude reduction.

Accompanying drawing explanation

Fig. 1 is GND algorithm flow chart of the present invention；

Fig. 2 is neural network classifier schematic diagram of the present invention；

Fig. 3 is GND decoding algorithm Performance Simulation Results schematic diagram of the present invention；

Fig. 4 is the GND algorithm corresponding with table 2 and Chase2, GPD algorithm complex

Graphic correlation analysis chart.

Detailed description of the invention

In order to make the purpose of the present invention, technical scheme and advantage clearer, below in conjunction with drawings and Examples, right The present invention is described in further detail.

The present invention is a kind of Hard decision decoding method based on genetic algorithm Yu neutral net, with block code (n, k) as a example by, It realizes flow process as it is shown in figure 1, include:

1, training neutral net:

1) real number symbol sebolic addressing r (r is received₁,r₂,r₃...) and obtain hard decision sequence after demodulated device matched filtering tolerance Row R；

2) by hard decision sequence R obtained after demodulator matched filtering tolerance respectively with stochastic generation d_h/ 2 n dimensions two System sequence T adds through mould 2 and produces d afterwards_h/ 2 candidate sequence A；

3) neutral net used required in training neutral net: GND decoding is as in figure 2 it is shown, it is as a classification Device is made up of three-layer network, i.e. input layer, hidden layer and output layer.Input layer is made up of individual n-k neuron, and output layer has 1 Neuron, hidden layer includes (2/3) (n-k+t+1) individual neuron, and wherein k is the information bit number of code, and t is that the maximum of this yard is entangled Wrong number.This operation realizes according to following steps: using syndrome sequence as input training mode, by corresponding mistake Weight w of pattern exports as target, is allowed to input a syndrome and just can obtain the weight w (w of corresponding error pattern =1,2,3..., n), syndrome S can obtain, i.e. according to the check matrix H of code word B representated by genetic algorithm individuality and code

S=B H ' formula (1)

According to minimal distance docoding criterion and standard array theory, the Hamming distance between code word and recently available codeword The biggest, then in standard array, it corresponding accompanies first collection weight the biggest, and the error bit number that i.e. it may contain is the most, then by it Corresponding syndrome is the biggest as the output valve of the neutral net of input.

2, genetic algorithm optimization is used to obtain one and transmit the code word of sequence more likelihood:

1) initialization of population: generate 2^tThe binary vector of individual n position is as initial population

A. first of population individual member P₁: the hard decision sequence R (r that matched filtering device is exported₁,r₂,...,r_n) set It is set to first individuality of population:

$P_1=R(r_1,r_2,...,r_n),r_i=\left\{\begin{array}{cc}1,& q_i>0\\ 0,& q_i\&le;0\end{array}\right.0<i<n$ Formula (2)

Wherein, Q (q₁,q₂,...,q_n) sequence of real numbers that quantifies for the unmatched wave filter hard decision that receives.

B. population other 2^t-1 individual member P_i: by by the uniform binary Orders Corrected T (t randomly generated₁, t₂,...,t_2t-1) and hard decision sequence R be added obtain, it may be assumed that

P_i=mod (R+T, 2), 2≤i≤2t, T=rand [0,1] formula (3)

2) ideal adaptation degree is evaluated:

According to formula 4.11, the fitness that genetic algorithm is individual is evaluated

Formula (4)

A. λ (P, Q) is correlation function, is used for the Euclidean distance calculating Genetic vectors Pi and receiving between sequence of real numbers Q.Individual Belong to sequence to the reality received the most similar, then the value of λ is the biggest.

$\mathrm{\&lambda;}(P,Q)=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{p}_i\&CenterDot;q_i$ Formula (5)

B.Weight (Error class (Indiv.)) is the output result of neutral net, and it is as a compensating factor, Represent the weight of the most probable error pattern of the corresponding code word of heredity individuality, when the individual contained error bit number of heredity is got over Few, the value of Weight (Error class (Indiv.)) is the least, and the value of final fitness will be the biggest.Obtain Penalty, needs to calculate, first with formula (1), syndrome sequence S that heredity to be assessed is individual, then by S input neural network.

3) natural selection: select outstanding individual ginseng from initial population based on roulette wheel selection or other systems of selection With heredity, the individual probability being chosen to participate in heredity is determined by its fitness, and fitness is the highest, and its selected probability is the biggest. Typically, the individual selected probability of i-th is:

$p\left(P_i\right)=\frac{a\left(P_i\right)}{\underset{j=0}{\overset{N-1}{\mathrm{\&Sigma;}}}a\left(P_j\right)}$ Formula (6)

4) pairing intersects: the individuality chosen will match at random, by by own partial element (code element) and the other side Intersecting and produce new individuality, the method that pairing intersects is multiple, and modal have single-point to intersect and select list in multiple-spot detection, this research Point intersects, and crossover probability is set to 0.9；

5) hereditary variation: random selection process 4) the middle new individuality produced, it is carried out variation process, specifically presses following behaviour Realize: by certain individual bit element (code element) upset, i.e. by 0 → 1 or 1 → 0, the mutation probability in this research is set to 0.025；

6) heredity terminates: heredity will terminate when genetic generation number reaches preset value, and now in population, fitness is the highest Individuality will be output, if generation number is not up to, and jump procedure 3) continue genetic process, in this research, genetic generation number is set to 20。

3, optimal sequence input hard decision error correcting deocder genetic algorithm exported decodes, and is finally decoded knot Really.

It is analyzed from analysis of complexity and error bit ability below, further illustrates advantages of the present invention:

1, analysis of complexity:

The computing cost of GND algorithm mainly includes the optimization of genetic algorithm module, and classification and the hard decision of neutral net are translated The error correction of code device, wherein genetic algorithm module proportion is maximum, for part of neural network, as long as network is trained to, When it uses pattern, the simplest several steps add multiplication and power accumulating operation, and computing cost is the least.This analysis is according to whole The complexity of decoding algorithm is estimated by addition required during decoding and multiplication computation amount.With linear block codes BCH (n, k,d_h, t) as a example by, wherein d_hFor the smallest hamming distance of block code, t is the maximum error correction number of block code.

1) genetic algorithm optimization module: genetic algorithm needs to carry out gen generation heredity, per has 2 (dh/2-1) individual individual need in generation Processed, process every time and will perform (n-1) sub-addition and n multiplication, therefore final needs altogether performs Sub-addition operation andSecondary multiplication operates；

2) neural network classification module: the neural network classifier in this research has (2/3) (n-k+t+1) individual hidden list Unit.To a neutral net trained when the pattern of use, each hidden unit needs to perform (n-k-1) sub-addition and (n- K) secondary multiplication, output unit performs (2/3) (n-k+t+1)-1 sub-addition and (2/3) (n-k+t+1) secondary multiplication.Therefore, for one Individual input, the additive operation number of times carried out needed for the whole center that decoded, nerve net module is { (2/3) (n-k+t+1)+n- k-2}·gen·2^tAnd { (2/3) (n-k+t+1)+n-k} gen 2^tSecondary multiplying.

3) Hard decision decoding device module: as a example by the BM decoding algorithm of BCH, its each Hard decision decoding needs to perform (2nt + 2t2-t) sub-addition computing and (2nt+2t2) secondary multiplying.

If quantifying to be considered as additive operation by hard decision, GND algorithm the most of the present invention contrasts with Chase2, GPD algorithm complex Situation is shown in Table 1:

As a example by BCH (31,16,7,3) code, use the relative analysis number of GND algorithm and Chase2, GPD algorithm complex According to being shown in Table 2:

In order to contrast, Fig. 4 gives in tradition Soft decision decoding Chase2 and pertinent literature the one that is given based on The complexity contrast situation of the GPD decoding algorithm of genetic algorithm.

Carefully analyzing table 2 data to understand, in three kinds of decodings, the computing cost amount of GND decoding is minimum, next to that GPD decoding, And the amount of calculation of soft-decision CHASE2 is maximum.This is owing to, during decoding, CHASE2 and GPD algorithm make use of extra soft Information produces candidate sequence, and the wave filter of demodulator is directly exported and processes by GND, and not using Soft Inform ation, to generate candidate empty Between.It can thus be seen that GND algorithm is the interpretation method that a kind of complexity is relatively low, workable.

2, error bit ability analysis:

During analogue simulation of the present invention, employ BCH (31,16) as decoding object, parameter setting values such as table 3,

For carrying out performance comparison, applicant simulates MLD optimal decoding algorithm simultaneously, and (" Yuan builds to see document in GPD decoding State, Wang Lin, Huang Sheng, Wang Yong. probability decoding algorithm [J] based on genetic algorithm. Beijing University of Post & Telecommunication reports .2012, and 35 (5): 98 ～101 "), CHASE2 Soft decision decoding and the error bit ability of BM Hard decision decoding, as shown in Figure 3.In figure, R represents reception The matched filtering device output of manipulator, BER represents the bit error rate, and SNR (dB) represents signal to noise ratio.

Available by carefully analyzing Fig. 3, various decoding algorithms are by exporting result R to the matched filtering device of demodulator Take different processing methods can reduce the bit error rate of receiving sequence to some extent.Such as shown in figure, the bit error rate is 10^-4 Time, BM Hard decision decoding can obtain the gain of about 1.5dB on the basis of the output result of matched filtering device more, and GND algorithm is about 2dB, chase2 are about 2.4GPD decoding about 2.6dB, MLD decoding can obtain about 3.8dB.It will thus be seen that GND decoding is gathered around There is preferable error-correcting performance, and close to traditional Soft decision decoding.Although GND decoding obtains not as chase2 and GPD algorithm Gain is big, but analyzes as discussed above, utilizes channel statistical probability Soft Inform ation to generate search volume by need not, and GND algorithm is complicated Spending relative chase2 soft decoding and GPD decoding reduces a lot, its practicality is higher.Can draw from above analysis, GND algorithm is being translated Achieve a good trade-off between code complexity and decoding error-correcting performance, be a kind of superior novel interpretation method.

It will be apparent to those skilled in the art that and understand, the inventive method for embodiment be merely to illustrate this Bright, and it is not limited to the present invention.Although effectively describing the present invention by embodiment, one skilled in the art will appreciate that this There is many changes spirit without deviating from the present invention in invention.Without departing substantially from the present invention spirit and real situation under, this area Technical staff is when making corresponding change or deformation according to the present invention, but these change accordingly or deformation belongs to the present invention's Claims.

Claims (5)

1. a Hard decision decoding method based on genetic algorithm and neutral net integrated intelligent algorithm, it is characterised in that include Following steps:

(1) training neutral net:

(1.1) real number symbol sebolic addressing r (r is received₁,r₂,r₃...) and obtain hard decision sequence after demodulated device matched filtering tolerance Row R；

(1.2) by hard decision sequence R obtained after demodulator matched filtering tolerance respectively with stochastic generation d_h/ 2 n dimensions two are entered Sequence T processed adds through mould 2 and produces d afterwards_h/ 2 candidate sequence A；

(1.3) training neutral net: the input layer of neutral net is made up of individual n-k neuron, and output layer has 1 neuron, hidden Including (2/3) (n-k+t+1) individual neuron containing layer, wherein k is the information bit number of code, and t is the maximum error correction number of this yard；Instruction Practicing process is: using syndrome sequence as input training mode, using defeated as target for weight w of corresponding error pattern Go out, be allowed to input a syndrome just can obtain corresponding error pattern weight w (w=1,2,3..., n), syndrome S Check matrix H according to code word B representated by genetic algorithm individuality and code obtains, i.e.

S=B H ' formula (1)；

(2) genetic algorithm optimization is used to obtain one and transmit the code word of sequence more likelihood:

(2.1) initialization of population: generate the binary vector of 2t n position as initial population

(2.1.1) first of population individual member P₁: the hard decision sequence R (r that matched filtering device is exported₁,r₂,...,r_n) set It is set to first individual member P of population₁:

$\begin{array}{cc}{P}_1=R(r_1,r_2,\mathrm{...},r_n),r_i=\left\{\begin{array}{cc}1,& q_i>0\\ 0,& q_i\&le;0\end{array}\right.& 0<i<n\end{array}$

Wherein, Q (q₁,q₂,...,q_n) sequence of real numbers that quantifies for the unmatched wave filter hard decision that receives；

(2.1.2) the individual member P of other 2t-1 of population_i: by by the uniform binary Orders Corrected T (t randomly generated₁, t₂,...,t_2t-1) and hard decision sequence R be added obtain, it may be assumed that

P_i=mod (R+T, 2), 2≤i≤2^t,

T=rand [0,1]；

(2.2) ideal adaptation degree is evaluated:

According to following formula, the fitness that genetic algorithm is individual is evaluated

Wherein, λ (P, Q) is correlation function, is used for calculating Genetic vectors P_iAnd the Euclidean distance received between sequence of real numbers Q, individual with It is the most similar that the reality received belongs to sequence, then the value of λ is the biggest,

$\mathrm{\&lambda;}(P,Q)=\underset{i=1}{\overset{n}{\&Sigma;}}{p}_i\&CenterDot;q_i$

Weight (Error class (Indiv.)) is the output result of neutral net, and penalty to be obtained needs first to calculate Syndrome sequence S that heredity to be assessed is individual, then by S input neural network；

(2.3) natural selection: select outstanding individual ginseng from initial population based on roulette wheel selection or other systems of selection With heredity, the individual selected probability of i-th is:

$p\left(P_i\right)=\frac{a\left(P_i\right)}{\underset{j=0}{\overset{N-1}{\&Sigma;}}a\left(P_j\right)}$

(2.4) pairing intersects: the individuality chosen will match at random, by own partial element i.e. code element being handed over the other side Fork produces new individual；

(2.5) hereditary variation: produce in random selection process (2.4) is new individual, and it is carried out variation process, processing method For, by individual certain bit element i.e. code element upset, i.e. by 0 → 1 or 1 → 0；

(2.6) heredity terminates: heredity will terminate when genetic generation number reaches preset value, now the highest individual of fitness in population Body will be output, if generation number is not up to, jump procedure (2.3) continues genetic process；

(3) optimal sequence input hard decision error correcting deocder genetic algorithm exported decodes, and is finally decoded result.

Method the most according to claim 1, it is characterised in that the method that described step (2.4) pairing intersects has single-point to intersect And multiple-spot detection.

Method the most according to claim 1, it is characterised in that described step (2.4) preferably single-point intersects, and crossover probability is set to 0.9。

Method the most according to claim 1, it is characterised in that the mutation probability of described step (2.5) hereditary variation is set to 0.025。

Method the most according to claim 1, it is characterised in that the genetic generation number that described step (2.6) heredity terminates is arranged It is 20.