CN110212920A

CN110212920A - A kind of decoding algorithm of the LDPC code based on deep learning

Info

Publication number: CN110212920A
Application number: CN201910454576.8A
Authority: CN
Inventors: 徐光宪; 郭若蕾
Original assignee: Liaoning Technical University
Current assignee: Liaoning Technical University
Priority date: 2019-05-29
Filing date: 2019-05-29
Publication date: 2019-09-06

Abstract

The invention discloses a kind of decoding algorithms of LDPC code based on deep learning, in conjunction with depth learning technology, respectively to the construction of big girth random LDPC code and structured LDPC code, there are the improvement belief propagation decoding on ring Tanner figure and its low complicated simplified decoding algorithm to be studied, using optimal idea, becate number distribution situation in check matrix where each nonzero element, by optimization when forefront becate number distribution with each column becate array at standard deviation be distributed combine, construct the random LDPC code with superperformance, have studied ring elimination algorithm, simulate the performance for changing the QC-LDPC code of null method construction, pass through improvement, the QC-LDPC code with more excellent big girth is constructed, improve BP decoding algorithm, propose a kind of improvement Min-Sum decoding algorithm, The present invention is based on the ring statistics characteristics of depth learning technology and Tanner figure, have first constructed the random LDPC code of superperformance, then the QC-LDPC code of more excellent big girth has been constructed by emulation, and algorithm is more excellent, and error-correcting performance is more preferable.

Description

A kind of decoding algorithm of the LDPC code based on deep learning

Technical field

The present invention relates to passing need degree field, the decoding algorithm of specially a kind of LDPC code based on deep learning.

Background technique

Since the rediscovering of LDPC code, people have had more than ten years to its research, and main achievement can be concluded Quinquepartite: first is that the graph model of LDPC code, second is that the iterative decoding performance evaluation of LDPC code, third is that the construction of LDPC code, four It is the coding of LDPC code, fifth is that the decoding of LDPC code, by the iterative decoding algorithm of LDPC code, people can be according to application environment Needs compromise between performance and complexity, currently, the ring elimination algorithm of LDPC code more fall behind, do not adjust and decoded Certain factors in journey such as improve the distribution of becate number, the ring statistics in Tanner figure, ring elimination algorithm, BP decoding algorithm, minimum With decoding algorithm, ring extend constraint condition and BP decoding algorithm etc., cause algorithm calculated performance and error-correcting performance compared with Difference.

Summary of the invention

The purpose of the present invention is to solve the ring elimination algorithms of existing LDPC code more to fall behind, during not adjusting decoding Certain factors, such as improve and the distribution of becate number, the ring statistics in Tanner figure, ring elimination algorithm, BP decoding algorithm, minimum and translate Code algorithm, ring extend constraint condition and BP decoding algorithm etc., cause the calculated performance of algorithm and error-correcting performance poor etc. The shortcomings that, and a kind of decoding algorithm of the LDPC code based on deep learning proposed.

To achieve the above object, the invention provides the following technical scheme: including,

Step 1, using gradually optimal idea, the becate number where each nonzero element in check matrix is distributed conduct Optimal design criterion proposes a kind of LDPC code algorithm based on improvement number of rings distributed structure, under the conditions of same code length code rate, Compared with the code of PEG algorithm construction, better performance is obtained.

Step 2, based on the ring statistics characteristic in Tanner figure, current becate number distribution and whole excellent will be optimized by column The becate number standard deviation distribution for changing each column in all column combines, and has constructed a kind of based on the LDPC code for improving ring statistics characteristic Construction algorithm obtains excellent performance compared with the PEG code of same code length code rate.

Step 3, a kind of building method-ring elimination algorithm for having studied big girth quasi-cyclic LDPC code, emulation confirm ring The performance of the QC-LDPC code of elimination algorithm construction, simulation result show that the QC-LDPC code of construction has the becate that cannot be eliminated, Ring extension constraint condition in basic matrix has been generalized to the extension constraint condition of the closed path in basic matrix, one kind has been made and has changed Into ring elimination algorithm, a kind of QC-LDPC code of girth more significantly is constructed, better performance is obtained.

Step 4, for the scheduling decoding algorithm having on ring Tanner figure, propose improved BP decoding algorithm, to scheduling calculate Method optimizes, and designs two kinds of prioritization schemes, continues additional interpretations using the message after optimization, obtains and calculates than scheduling decoding The excellent performance of method.

Step 5 compares and analyzes Min-Sum decoding algorithm and BP decoding algorithm, calculates for minimum and decoding Method check-node message reliably spends estimation, proposes that a kind of compensation is minimum and decoding check-node message reliably spends estimation Improved Min-Sum decoding algorithm.

Preferably, PEG algorithm is the size m × n and variable node degree series D for giving check matrix H in the step 1 =Dv1, Dv2 ... Dvn }, with PEG algorithm, the corresponding Tanner figure constitution step of check matrix is as follows, one, initialization: Setting v1 is current variable node and enables k=0；Two, current variable node and Dvj check-node are connected, first school is connected It tests node: choosing the smallest row of row weight in current check matrix and place non-zero entry k=k+1；It connects remaining check-node: being with vj Root node extends tree graph, chooses the verification the smallest check-node of degree and is connected with vj, k=k+1；If k=Dvj enables j=j+1 And k=0 and be arranged vj be current variable node, be transferred to step 3, be otherwise transferred to step 2；If three, j=n+1, terminate construction, Otherwise it is transferred to step 2, one of described step 1 is given verification square based on the LDPC code algorithm for improving number of rings distributed structure Size m × the n and variable node degree series D={ Dv1, Dv2 ... Dvn } of battle array H, based on the LDPC code structure for improving number of rings distribution It makes algorithm to be summarized as follows, one, initialization；To all i, j (1≤i≤m, 1≤i≤n), hij=0 is enabled, setting jth=1, which is classified as, works as Forefront simultaneously enables k=0；Two, it when Dvj non-zero entry of forefront placement, places first non-zero entry: choosing row weight in current check matrix The smallest row places non-zero entry, k=k+1；Place remaining non-zero entry；Count each alternative line position set including Fourth Ring, six The number of ring and eight rings, by the long priority level sequence from small to large of ring, step-sizing provides the alternate location of minimum number of rings Place non-zero entry, k=k+1；If k=Dvj, enables j=j+1 and k=0 and jth is set is classified as and work as forefront, be transferred to step 3, it is no Then it is transferred to step 2；If j=n+1, terminate construction；Otherwise it is transferred to step 2, the becate number in the step 1 is to influence LDPC code One key factor of performance, we construct the binary system that four code lengths are respectively the code rate 0.5 of n=504 and n=1008 LDPC code, wherein number of the n=504 based on 8 ring of LDPC code and 10 rings that improve number of rings distributed structure is respectively 403 and 12251, The number of corresponding 8 ring of PEG code and 20 rings is respectively 813 and 11345；Based on improvement number of rings distributed structure when n=1008 The number of 8 ring of LDPC code and 10 rings is respectively 46 and 11410, and the number of corresponding 8 ring of PEG code and 20 rings is respectively 54 and 11086, Obviously, LDPC code of the n=504 and n=1008 based on improvement number of rings distributed structure, institute fewer than corresponding PEG code on number of rings mesh With better performance.The big of becate number distributed structure for being included based on each non-zero entry in successive optimization check matrix is enclosed Long LDPC code has better error-correcting performance compared with PEG code.

Preferably, Tanner figure can explain the operation of iterative decoder in the step 2, and each node is an independence Message handling device, each edge transmit message from given node toward adjacent node, originate in node u1, terminate at the k long of node vk Path is oriented edge sequence e1=(u1, v1) ..., ek=(uk, vk), wherein for all i=1,2 ..., k-1, vi= Ui+1, starting point are closed path with the path that terminal is overlapped, i.e. u1=vk, one of described step 2 is based on improvement ring statistics The LDPC code construction algorithm of characteristic be given check matrix H size m × n and variable node degree series D=Dv1, Dv2 ... Dvn }, it is summarized as follows based on the LDPC code construction algorithm for improving ring statistics characteristic, one, initialization: to all i, j (1 ≤ i≤m, 1≤i≤n), hij=0 is enabled, setting jth=1 is classified as when forefront and enables l=0；Two, when Dvj non-zero is placed in forefront Initialization: member to all k (k=4,6,8), enables nk=∞, for all t (1≤t≤Dvj), enables L (t)=0；Count l Group non-zero combines relevant ring statistics characteristic: statistics is as becate number Ck (the j) (k=that the combination of l group non-zero entry includes in forefront 4,6,8) mean value and standard deviation Dk (k=4,6,8) of becate number where each column in j column before corresponding to H, are counted；By k value from it is small to Big sequence successively compares the size of Ck (j) Yu nk；If j=n+1, terminates construction, be otherwise transferred to step 2.By will by column Optimization is combined when the becate number distribution in forefront is distributed with the becate number standard deviation of each column in all column of global optimization, gives the The construction algorithm of the big girth LDPC code of two classes, the LDPC code that this method constructs has lower error floor, in low errored bit Rate region has better error-correcting performance.

Preferably, the QC-LDPC code of one kind girth more significantly is by a length of 2g in matrix H=[hij] in the step 3 Ring is defined as the 2g long ordered sequence for meeting following condition being made of the 2g position hig=1, one, two adjacent hij=1 Position is in same a line different lines or in same a line difference row；Two, the position all 2g hij=1 is different；Three, hij is originated =1 position with terminate the position hij=1 and do not go together in same a line different lines or in same row, in the step 3 basic matrix be by Each L rank girth square matrix I (aij) more significantly in the QC-LDPC code check matrix H of girth more significantly based on L rank uses aij generation It replaces, each zero square matrix of L rank is known as Hb with what ∞ was replaced, and the closed path in the step 3 is by 2g hij=1 element The 2g long ordered sequence of composition, and meet two adjacent hij=1 elements in same a line different lines or in same row difference Row, the starting position hij=1 are not gone together with the end position hij=1 in same a line different lines or in same row, give group moment A length of 6 closed path that one a length of 6 ordered sequence is constituted in battle array Hb, the ring in the step 3, which promotes constraint condition, is In check matrix in the ring of a length of 2g, 2g must satisfy two conditions: 2g 1 is not being gone together, and every row includes two 1,2g 1 In different lines, and each column includes two 1, and it is given basic matrix Hb=that one of described step 3, which improves ring elimination algorithm, Under the conditions of [Hbij], initialization: the initialization value of total all 1 elements of basic matrix Hb is 0, for all Hbij=1 (1≤i≤m, 1≤j≤n), enable aij=0, successive optimization process: by the sequence arranged from 1 column to n in basic matrix H, forefront is worked as in selection, and foundation is closed Combining diameter constraint list: the closed path process on side is not continuously repeated in detection basic matrix, shift value: current 1 element is set Corresponding shift value value range is [0, L-1].The ring being extended to by the ring for including in research basic matrix in check matrix Necessary and sufficient condition and ring elimination algorithm on the basis of this necessary and sufficient condition, the constraint to being extended in check matrix from the ring of basic matrix Condition is promoted, and is corrected to elimination algorithm, and the QC-LDPC code for improving ring elimination algorithm construction corresponds to Tanner The girth of figure shows that the code of improved ring elimination algorithm construction has better error-correcting performance by emulating.

Preferably, improved BP decoding algorithm is the parallel decorrelation process of all variable nodes in the step 4, including is looked into Ask tree graph number and correspondence mappings relationship that each check-node participates in；Initialize all vj；Carry out check-node update and Become node updates；If reaching maximum number of iterations g/2, then output message is calculated；The total message of variable node is calculated, emulation is passed through As a result, it was confirmed that compared with existing related algorithm, it is with good performance in high s/n ratio region.

Preferably, improved Min-Sum decoding algorithm is Ji Yu when (2) ∣≤ln ∣ Zp in the step 5, expression formula max { 0 , ∣ Zp ∣-ln (2) } slows down amplitude in BP iterative decoding and is less than ln2 transmission of news , Dang ∣ Zp ∣ >=ln (2), expression formula max { 0 , ∣ Zp ∣-ln (2) } compensates for estimation and this expression formula excessively of Min-Sum decoding algorithm check-node output message reliability For linear representation, the biasing coefficient of BP decoding algorithm and improved Min-Sum decoding algorithm in the step 5 is all It is invariable, and the codomain for biasing coefficient is (0 ,+∞), shows that improved Min-Sum decoding algorithm ratio is set by simulation result Believe that propagation decoding algorithm has better error-correcting performance, and there is decoding performance more better than BP under high s/n ratio.

Compared with prior art, the beneficial effects of the present invention are: the present invention uses depth learning technology, in the process of decoding In, by improving the distribution of becate number, with the ring statistics in Tanner figure, ring elimination algorithm is improved, BP decoding algorithm is improved, changes Into Min-Sum decoding algorithm, optimizes ring and extend constraint condition and BP decoding algorithm etc., shown greatly by simulation result Calculated performance and the error-correcting performance for improving algorithm are poor.

Detailed description of the invention

Fig. 1 is the flow chart of LDPC code decoding algorithm of the present invention.

Fig. 2 is PEG algorithm pattern of the invention.

Fig. 3 is the performance plot of Tanner figure ring statistics of the present invention.

Fig. 4 is that the present invention is based on the LDPC code algorithm patterns for improving number of rings distributed structure.

Fig. 5 is the group moment system of battle formations of the invention.

Fig. 6 is closed path figure of the invention.

Specific embodiment

Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts every other Embodiment shall fall within the protection scope of the present invention.

Fig. 1-6 is please referred to, the present invention provides a kind of technical solution: including:

In the step 1 PEG algorithm be given check matrix H size m × n and variable node degree series D=Dv1, Dv2 ... Dvn }, with PEG algorithm, the corresponding Tanner figure constitution step of check matrix is as follows, one, initialization: setting v1 is Current variable node and enable k=0；Two, current variable node and Dvj check-node are connected, first check-node is connected: choosing The smallest row of row weight in current check matrix is taken to place non-zero entry k=k+1；It connects remaining check-node: expanding by root node of vj Tree graph is opened up, the verification the smallest check-node of degree is chosen and is connected with vj, k=k+1；If k=Dvj, enables j=j+1 and k=0 is simultaneously Setting vj is current variable node, is transferred to step 3, is otherwise transferred to step 2；If three, j=n+1, terminates construction, be otherwise transferred to Step 2, one of described step 1 are the big of given check matrix H based on the LDPC code algorithm for improving number of rings distributed structure Small m × n and variable node degree series D=Dv1, Dv2 ... and Dvn }, based on the LDPC code construction algorithm for improving number of rings distribution It is summarized as follows, one, initialization；To all i, j (1≤i≤m, 1≤i≤n), enable hij=0, setting jth=1 be classified as when forefront simultaneously Enable k=0；Two, it when Dvj non-zero entry of forefront placement, places first non-zero entry: it is the smallest to choose row weight in current check matrix Row places non-zero entry, k=k+1；Place remaining non-zero entry；Count each alternative line position set including Fourth Ring, six rings and eight The number of ring, by the long priority level sequence from small to large of ring, the alternate location that step-sizing provides minimum number of rings is placed non- Null element, k=k+1；If k=Dvj, enables j=j+1 and k=0 and jth is set is classified as and work as forefront, be transferred to step 3, be otherwise transferred to Step 2；If j=n+1, terminate construction；Otherwise it is transferred to step 2, the becate number in the step 1 is to influence LDPC code performance One key factor, we construct the binary system LDPC code that four code lengths are respectively the code rate 0.5 of n=504 and n=1008, Wherein number of the n=504 based on 8 ring of LDPC code and 10 rings that improve number of rings distributed structure is respectively 403 and 12251, corresponding The number of 8 ring of PEG code and 20 rings is respectively 813 and 11345；Based on 8 ring of LDPC code for improving number of rings distributed structure when n=1008 Number with 10 rings is respectively 46 and 11410, and the number for corresponding to 8 ring of PEG code and 20 rings is respectively 54 and 11086, it is clear that n= 504 and n=1008 is fewer than corresponding PEG code on number of rings mesh based on the LDPC code for improving number of rings distributed structure, so having more Good performance.Big girth LDPC code based on the becate number distributed structure that each non-zero entry in successive optimization check matrix is included There is better error-correcting performance compared with PEG code.

Tanner figure can explain the operation of iterative decoder in the step 2, and each node is at an independent message Device is managed, each edge transmits message from given node toward adjacent node, originates in node u1, the path k long for terminating at node vk is Oriented edge sequence e1=(u1, v1) ..., ek=(uk, vk), wherein for all i=1,2 ..., k-1, vi=ui+1, Starting point is closed path with the path that terminal is overlapped, i.e. u1=vk, one of described step 2 is based on improvement ring statistics characteristic LDPC code construction algorithm is the size m × n and variable node degree series D={ Dv1, Dv2 ... Dvn } of given check matrix H, It is summarized as follows based on the LDPC code construction algorithm for improving ring statistics characteristic, one, initialization: to all i, j (1≤i≤m, 1≤i≤ N), hij=0 is enabled, setting jth=1 is classified as when forefront and enables l=0；Two, when Dvj non-zero entry, initialization: to institute are placed in forefront There is k (k=4,6,8), enable nk=∞, for all t (1≤t≤Dvj), enables L (t)=0；It is related to count the combination of l group non-zero Ring statistics characteristic: statistics l group non-zero entry in the forefront combines the becate number Ck (j) (k=4,6,8) for including, and counts pair Answer the mean value and standard deviation Dk (k=4,6,8) of becate number where each column in j column before H；By the sequence of k value from small to large, successively Compare the size of Ck (j) Yu nk；If j=n+1, terminates construction, be otherwise transferred to step 2.By the way that the short of forefront is worked as in optimization by column Ring number distribution is combined with the becate number standard deviation distribution of each column in all column of global optimization, gives the big girth LDPC of the second class The construction algorithm of code, the LDPC code that this method constructs have lower error floor, have in low bit error rate region more preferable Error-correcting performance.

The QC-LDPC code of one kind girth more significantly is by the ring definition of a length of 2g in matrix H=[hij] in the step 3 For the 2g long ordered sequence for meeting following condition being made of the 2g position hig=1, one, two adjacent positions hij=1 exist With a line different lines or in same a line difference row；Two, the position all 2g hij=1 is different；Three, hij=1 are originated It sets with the end position hij=1 in same a line different lines or does not go together in same row, basic matrix is will to be based on L in the step 3 Each L rank girth square matrix I (aij) more significantly in the QC-LDPC code check matrix H of the girth more significantly of rank is replaced with aij, often A zero square matrix of L rank is known as Hb with what ∞ was replaced, what the closed path in the step 3 was made of 2g hij=1 element 2g long ordered sequence, and meet two adjacent hij=1 elements and do not go together in same a line different lines or in same row, it originates The position hij=1 is not gone together with the end position hij=1 in same a line different lines or in same row, is given one in basic matrix Hb A length of 6 closed path that a a length of 6 ordered sequence is constituted, it is check matrix that the ring in the step 3, which promotes constraint condition, In a length of 2g ring in, 2g must satisfy two conditions: 2g 1 is not being gone together, and every row includes two 1,2g 1 in difference Column, and each column includes two 1, it is given basic matrix Hb=[Hbij that one of described step 3, which improves ring elimination algorithm ,] condition Under, initialization: the initialization value of total all 1 elements of basic matrix Hb is 0, for all Hbij=1 (1≤i≤m, 1≤j≤n), Enable aij=0, successive optimization process: by the sequence arranged from 1 column to n in basic matrix H, forefront is worked as in selection, establishes closed path constraint List: the closed path process on side is not continuously repeated in detection basic matrix, shift value is set: the corresponding displacement of current 1 element Value value range is [0, L-1].By the ring for including in research basic matrix be extended to the necessary and sufficient condition of ring in check matrix with And ring elimination algorithm on the basis of this necessary and sufficient condition, it is carried out to extending in check matrix from the constraint condition of the ring of basic matrix It promotes, elimination algorithm is corrected, the QC-LDPC code for improving ring elimination algorithm construction corresponds to the girth of Tanner figure, Show that the code of improved ring elimination algorithm construction has better error-correcting performance by emulating.

Improved BP decoding algorithm is the parallel decorrelation process of all variable nodes in the step 4, including inquiry is each The tree graph number and correspondence mappings relationship that check-node participates in；Initialize all vj；Carry out check-node update and traitorous point It updates；If reaching maximum number of iterations g/2, then output message is calculated；The total message of variable node is calculated, is demonstrate,proved by simulation result It is real, it is with good performance in high s/n ratio region compared with existing related algorithm.

Improved Min-Sum decoding algorithm is Ji Yu when (2) ∣≤ln ∣ Zp in the step 5, expression formula max { 0 , ∣ Zp ∣-ln (2) } it slows down in BP iterative decoding amplitude and is less than ln2 transmission of news , Dang ∣ Zp ∣ >=ln (2), expression formula max { 0 , ∣ Zp ∣-ln (2) } compensate for Min-Sum decoding algorithm check-node output message reliability cross estimation and this expression formula be linear Expression formula, the biasing coefficient of BP decoding algorithm and improved Min-Sum decoding algorithm in the step 5 all it is constant not Become, and the codomain for biasing coefficient is (0 ,+∞), shows that improved Min-Sum decoding algorithm compares belief propagation by simulation result Decoding algorithm has better error-correcting performance, and has decoding performance more better than BP under high s/n ratio.

Design and realization approach of the invention are as follows: cause the factor of error-correcting performance difference to be analyzed in the process according to decoding, It is following because usually improving algorithm performance by changing.One, change becate number distributed structure, proposes a kind of based on improvement number of rings point The LDPC code algorithm of cloth construction shows that algorithm is based on each non-zero entry in successive optimization check matrix and is wrapped by simulation result The big girth LDPC code of the becate number distributed structure contained has better error-correcting performance compared with PEG code.Two, schemed based on Tanner Statistical property has constructed a kind of LDPC code construction algorithm based on improvement ring statistics characteristic, has shown to pass through by simulation result Optimization by column is combined when the becate number distribution in forefront is distributed with the becate number standard deviation of each column in all column of global optimization, is given The construction algorithm of the big girth LDPC code of the second class is gone out, the LDPC code that this method constructs has lower error floor, low Bit error rate region has better error-correcting performance.Three, optimization closed path extends constraint condition, has made a kind of improvement ring and has disappeared Except algorithm, a kind of QC-LDPC code of girth more significantly has been constructed, has been shown by simulation result by being wrapped in research basic matrix The necessary and sufficient condition and ring elimination algorithm on the basis of this necessary and sufficient condition that the ring contained is extended to the ring in check matrix, to verification It extends in matrix and is promoted from the constraint condition of the ring of basic matrix, elimination algorithm is corrected, improve ring elimination The QC-LDPC code of algorithm construction corresponds to the girth of Tanner figure, shows improved ring elimination algorithm construction by simulation result Code has better error-correcting performance.Four, it proposes improved BP decoding algorithm, dispatching algorithm is optimized, simulation result is passed through Show compared with existing related algorithm, it is with good performance in high s/n ratio region.Five, improved Min-Sum decoding algorithm, Show that improved Min-Sum decoding algorithm has better error-correcting performance than BP decoding algorithm by simulation result, and And there is decoding performance more better than BP under high s/n ratio

It is obvious to a person skilled in the art that invention is not limited to the details of the above exemplary embodiments, Er Qie In the case where without departing substantially from spirit or essential attributes of the invention, the present invention can be realized in other specific forms.Therefore, no matter From the point of view of which point, the present embodiments are to be considered as illustrative and not restrictive, and the scope of the present invention is by appended power Benefit requires rather than above description limits, it is intended that all by what is fallen within the meaning and scope of the equivalent elements of the claims Variation is included within the present invention.Any reference signs in the claims should not be construed as limiting the involved claims.

Claims

1. a kind of decoding algorithm of the LDPC code based on deep learning characterized by comprising

Step 1, using gradually optimal idea, by the becate number distribution where each nonzero element in check matrix as optimal Design criteria proposes a kind of LDPC code algorithm based on improvement number of rings distributed structure, under the conditions of same code length code rate, with The code of PEG algorithm construction is compared, and better performance is obtained；Step 2, based on the ring statistics characteristic in Tanner figure, will by column Optimize current becate number distribution to combine with the becate number standard deviation distribution of each column in all column of global optimization, construct It is a kind of to be obtained excellent compared with the PEG code of same code length code rate based on the LDPC code construction algorithm for improving ring statistics characteristic Performance；Step 3, a kind of building method-ring elimination algorithm for having studied big girth quasi-cyclic LDPC code, emulation confirm that ring is eliminated The performance of the QC-LDPC code of algorithm construction, simulation result shows that the QC-LDPC code of construction has the becate that cannot be eliminated, by base Ring extension constraint condition in matrix has been generalized to the extension constraint condition of the closed path in basic matrix, has made a kind of improvement ring Elimination algorithm has constructed a kind of QC-LDPC code of girth more significantly, has obtained better performance；Step 4, for there is ring Scheduling decoding algorithm on Tanner figure proposes improved BP decoding algorithm, optimizes to dispatching algorithm, design two kinds it is excellent Change scheme continues additional interpretations using the message after optimization, obtains the performance more excellent than scheduling decoding algorithm；Step 5, will most Small and decoding algorithm is compared and analyzed with BP decoding algorithm, reliable for Min-Sum decoding algorithm check-node message Estimation is spent, a kind of minimum improved minimum and decoding calculation for reliably spending estimation with decoding check-node message of compensation is proposed Method.

2. a kind of decoding algorithm of LDPC code based on deep learning according to claim 1, it is characterised in that: the step PEG algorithm is the size m × n and variable node degree series D={ Dv1, Dv2 ... Dvn } of given check matrix H, fortune in rapid 1 With PEG algorithm, the corresponding Tanner figure constitution step of check matrix is as follows, one, initialization: setting v1 be current variable node and Enable k=0；Two, current variable node and Dvj check-node are connected, first check-node is connected: choosing current check matrix The smallest row of middle row weight places non-zero entry k=k+1；It connects remaining check-node: extending tree graph by root node of vj, choose verification The smallest check-node of degree is connected with vj, k=k+1；If k=Dvj, enable j=j+1 and k=0 and be arranged vj be current variable Node is transferred to step 3, is otherwise transferred to step 2；If three, j=n+1, terminates construction, be otherwise transferred to step 2, the step 1 One of be based on the LDPC code algorithm for improving number of rings distributed structure given check matrix H size m × n and variable node Degree series D=Dv1, Dv2 ... and Dvn }, it is summarized as follows based on the LDPC code construction algorithm for improving number of rings distribution, one, initialization； To all i, j (1≤i≤m, 1≤i≤n), hij=0 is enabled, setting jth=1 is classified as when forefront and enables k=0；Two, when forefront is put Dvj non-zero entry is set, first non-zero entry is placed: choosing the smallest row of row weight in current check matrix and places non-zero entry, k=k+ 1；Place remaining non-zero entry；Count each alternative line position set including Fourth Ring, six rings and eight rings number, it is long from small by ring To big priority level sequence, the alternate location that step-sizing provides minimum number of rings places non-zero entry, k=k+1；If k= Dvj then enables j=j+1 and k=0 and jth is arranged is classified as and work as forefront, is transferred to step 3, is otherwise transferred to step 2；If j=n+1, knot Beam construction；Otherwise it is transferred to step 2, the becate number in the step 1 is a key factor for influencing LDPC code performance, we The binary system LDPC code that four code lengths are respectively the code rate 0.5 of n=504 and n=1008 is constructed, wherein n=504 is based on improving 8 ring of LDPC code of number of rings distributed structure and the number of 10 rings are respectively 403 and 12251, of corresponding 8 ring of PEG code and 20 rings Number is respectively 813 and 11345；It is respectively based on 8 ring of LDPC code and the numbers of 10 rings for improving number of rings distributed structure when n=1008 46 and 11410, the number for corresponding to 8 ring of PEG code and 20 rings is respectively 54 and 11086, it is clear that n=504 and n=1008 is based on changing LDPC code into number of rings distributed structure is fewer than corresponding PEG code on number of rings mesh, so having better performance.

3. a kind of decoding algorithm of LDPC code based on deep learning according to claim 1, it is characterised in that: the step Tanner figure can explain the operation of iterative decoder in rapid 2, and each node is an independent message processor, and each edge is from giving Determine node and transmit message toward adjacent node, originate in node u1, the path k long for terminating at node vk is oriented edge sequence e1= (u1, v1) ..., ek=(uk, vk), wherein 2 ..., k-1, vi=ui+1, starting point are overlapped with terminal for all i=1 Path be closed path, i.e. u1=vk, one of described step 2 based on improve ring statistics characteristic LDPC code construction algorithm It is the size m × n and variable node degree series D={ Dv1, Dv2 ... Dvn } of given check matrix H, based on improving ring statistics The LDPC code construction algorithm of characteristic is summarized as follows, and one, initialization: enabling hij=0 to all i, j (1≤i≤m, 1≤i≤n), if Jth=1 is set to be classified as when forefront and enable l=0；Two, when Dvj non-zero entry is placed in forefront, initialize: to all k (k=4,6,8), Nk=∞ is enabled, for all t (1≤t≤Dvj), enables L (t)=0；It counts l group non-zero and combines relevant ring statistics characteristic: Statistics works as l group non-zero entry in forefront and combines the becate number Ck (j) (k=4,6,8) for including, each column in j column before the corresponding H of statistics The mean value and standard deviation Dk (k=4,6,8) of place becate number；By the sequence of k value from small to large, successively compare Ck (j) and nk Size；If j=n+1, terminates construction, be otherwise transferred to step 2.

4. a kind of decoding algorithm of LDPC code based on deep learning according to claim 1, it is characterised in that: the step The QC-LDPC code of one kind girth more significantly is that the ring of a length of 2g in matrix H=[hij] is defined as by 2g hig=1 in rapid 3 The 2g long ordered sequence for meeting following condition of position composition, one, two adjacent positions hij=1 in same a line different lines or Person is in same a line difference row；Two, the position all 2g hij=1 is different；Three, the starting position hij=1 and end hij=1 Position is not gone together in same a line different lines or in same row, in the step 3 basic matrix be will be based on the girth more significantly of L rank QC-LDPC code check matrix H in each L rank girth square matrix I (aij) more significantly replaced with aij, each zero square matrix of L rank is used What ∞ replaced obtaining is known as Hb, the 2g long ordered sequence that the closed path in the step 3 is made of 2g hij=1 element, And meet two adjacent hij=1 elements and do not go together in same a line different lines or in same row, the starting position hij=1 and knot The position beam hij=1 is not gone together in same a line different lines or in same row, gives in basic matrix Hb one a length of 6 orderly A length of 6 closed path of Sequence composition, the ring in the step 3 promote the ring that constraint condition is a length of 2g in check matrix In, 2g must satisfy two conditions: 2g 1 is not being gone together, and every row includes two 1,2g 1 in different lines, and each column includes Two 1, it is given basic matrix Hb=[Hbij that one of described step 3, which improves ring elimination algorithm ,] under the conditions of, initialization: group moment The initialization value of total all 1 elements of battle array Hb is 0, for all Hbij=1 (1≤i≤m, 1≤j≤n), enables aij=0, gradually excellent Change process: by the sequence arranged from 1 column to n in basic matrix H, forefront is worked as in selection, establishes closed path constraint list: detection basic matrix In do not continuously repeat the closed path process on side, shift value is set: the corresponding shift value value range of current 1 element be [0, L-1】。

5. a kind of decoding algorithm of LDPC code based on deep learning according to claim 1, it is characterised in that: the step Improved BP decoding algorithm is the parallel decorrelation process of all variable nodes in rapid 4, including inquiring each check-node participation Tree graph number and correspondence mappings relationship；Initialize all vj；It carries out check-node update and becomes node updates；If reached most Big the number of iterations g/2, then calculate output message；Calculate the total message of variable node.

6. a kind of decoding algorithm of LDPC code based on deep learning according to claim 1, it is characterised in that: the step Improved Min-Sum decoding algorithm is Ji Yu when (2) ∣≤ln ∣ Zp in rapid 5, and expression formula max { 0 , ∣ Zp ∣-ln (2) } is slowed down Amplitude is less than ln2 transmission of news , Dang ∣ Zp ∣ >=ln (2) in BP iterative decoding, and expression formula max { 0 , ∣ Zp ∣-ln (2) } is compensated for Min-Sum decoding algorithm check-node exports the estimation excessively of message reliability and this expression formula is linear representation, the step The biasing coefficient of BP decoding algorithm and improved Min-Sum decoding algorithm in 5 is all invariable, and biases coefficient Codomain is (0 ,+∞).