DECODING METHOD OF LDPC CODES BASED ON PARTIAL AVERAGE RESIDUAL BELIEF PROPAGATION

Abstract

A decoding method of low-density parity-check (LDPC) codes based on partial average residual belief propagation includes the following steps: S1: calculating a size of a cluster π in a protograph based on a code length m and a code rate of a target codeword; S2: pre-computing an edge residual rci→vj corresponding to each edge from a variable node to a check node in a check matrix H; S3: calculating, based on π, a partial average residual (PAR) value corresponding to each cluster in the check matrix H; S4: sorting m/π clusters in descending order of corresponding PAR values, and updating an edge with a largest edge residual in each cluster; S5: updating edge information mc→vi from a check node ci to a variable node vj, and then updating a log-likelihood ratio (LLR) value L(vj) of the variable node vj; and S6: after the updating, making a decoding decision.

Description

CROSS-REFERENCE TO RELATED APPLICATIONS

This application claims priority to Chinese Patent Application No. 202110358266.3 with a filing date of Apr. 1, 2021. The content of the aforementioned application, including any intervening amendments thereto, is incorporated herein by reference.

TECHNICAL FIELD

The present disclosure relates to the field of communications technologies, and more specifically, to a decoding method of low-density parity-check (LDPC) codes based on partial average residual belief propagation (PARBP).

BACKGROUND ART

Decoding algorithms of LDPC codes are divided into hard decision decoding and soft decision decoding. The hard decision decoding algorithm has the advantages of low decoding complexity and easy hardware implementation, but its error correction performance is far inferior to the soft decision decoding algorithm. The soft decision decoding algorithm can be divided into three categories according to message update policies: parallel scheduling, serial scheduling, and dynamic scheduling. Different from parallel scheduling and serial scheduling that have a fixed update order, dynamic scheduling can dynamically select an appropriate update order based on a current decoding state. Therefore, compared with parallel scheduling and serial scheduling, dynamic scheduling has faster convergence speed and better error correction performance, and is quite suitable for ultra-reliable and low-latency decoding scenarios. A. I. Casado et al. proposed an edge residual—based dynamic scheduling decoding algorithm [A. I. V. Casado, M. Griot and R. D. Wesel. Informed Dynamic Scheduling for Belief-Propagation Decoding of LDPC Codes [C]. IEEE International Conference on Communications (ICC 2007). Glasgow, UK. 932-937, June 2007.].

The dynamic scheduling decoding algorithm is a residual belief propagation (RBP) algorithm [A. I. V. Casado, M. Griot and R. D. Wesel. Informed Dynamic Scheduling for Belief-Propagation Decoding of LDPC Codes [C]. IEEE International Conference on Communications (ICC 2007). Glasgow, UK. 932-937, June 2007.]. In the RBP algorithm, a message update order is dynamically selected based on check-to-variable (C2V) edge residuals of all edges. For an edge connecting a variable node v_j and a check node c_i, its edge residual r_ci→vj is calculated according to formula (1):

r _{c i →v j} =∥m _{c i →v j} ^pre −m _{c i →v j} ∥  (1)

In the formula, m_ci→vj is current C2V edge information from the check node c_i to the variable node v_j, and m_ci→vj^pre is a pre-computed value of to-be-updated C2V edge information. It can be learned from the calculation formula of the edge residual that a value of the edge residual is always greater than 0. Therefore, an edge with an edge residual of 0 will not be selected. The edge residual is a quantity that describes a difference between a current edge information value and a to-be-updated edge information value. A greater edge residual means a greater difference between the current edge information value and the to-be-updated edge information value, a greater adjustment of an updated edge information value, and a more reliable message update.

Therefore, before each edge information update, the RBP algorithm firstly finds an edge with a largest edge residual from all edges, and then updates its edge information value. After the edge information with the largest edge residual is updated, the residual value is set to 0. Then a decoding decision is made. A message update calculation method of the dynamic scheduling decoding algorithm is consistent with those of the serial scheduling decoding algorithm and the parallel scheduling decoding algorithm.

FIG. 1 shows a message update process of the RBP algorithm. The specific decoding operations are as follows:

Step 1: Calculate edge residuals of all C2V edges in a Tanner graph, and find an edge with a largest edge residual. As shown in FIG. 1(a), assuming that an edge connecting a check node c_i and a variable node v_j has a largest edge residual, update C2V edge information m_ci→vj of the edge. After the update is completed, set an edge residual r(m_ci→vj) to 0.

Step 2: As shown in FIG. 1(b), make a decoding decision on the updated variable node v_j, and if the decoding is successful or a maximum number of iterationsIter_max, is reached, stop the decoding. Otherwise, update a V2C (variable-to-check) message m_vj→ca from the variable node v_j to a check node c_a∈N(v_j\c_i.

Step 3: As shown in FIG. 1(c), after the check node c_a∈N(v_j\c_i obtains the updated information, pre-compute a C2V edge residual from a check node C_a to a variable node v_b∈N(C_a)\V_j. Repeat the above decoding process until the decoding is successful or the maximum number of iterations Iter_max, is reached.

Although the conventional dynamic scheduling decoding algorithm is easy to understand, there are several problems.

1. The dynamic scheduling decoding algorithm performs a global search when an edge with a largest edge residual is selected for updating. As a result, some edges with larger residual values are always updated, while other edges with smaller residual values cannot be selected and updated in the case of global search. Therefore, greediness becomes more obvious as the number of iterations increases. In other words, some variable nodes occupy too many computing resources, such that decoding performance of the decoding algorithm cannot be further improved.

2. Because the LDPC codes have a ring structure of many edges, the dynamic scheduling decoding policy of selecting an update order based on a largest edge residual may cause some rings to fail to be decoded during the update process, thereby forming trap sets. Such trap sets may lead to error floors during the decoding process. Various existing decoding algorithms may generate error floors of different depths.

SUMMARY

In order to resolve the problems of occupying too many computing resources and generating error floors of different depths in the prior art, the present disclosure provides a decoding method of LDPC codes based on PARBP, which occupies fewer computing resources, thereby effectively improving a decoding throughput and lowering an error floor.

To achieve the foregoing objectives, the following technical solutions are used in the present disclosure: A decoding method of LDPC codes based on PARBP is provided. The method includes the following steps:

S1: calculating a size π of a cluster of a base matrix in a protograph based on a code length m and a code rate of a target codeword;

S2: pre-computing an edge residual r_ci→vj corresponding to each edge from a variable node to a check node in a check matrix H, where the check matrix H is obtained by lifting the base matrix based on a structural characteristic of the protograph;

S3: calculating, based on π, a partial average residual (PAR) value corresponding to each cluster in the check matrix H;

S4: sorting m/π clusters in descending order of corresponding PAR values, and updating an edge with a largest edge residual in each cluster;

S5: updating edge information m_ci→vj from a check node c_i to a variable node v_j, and then updating a log-likelihood ratio (LLR) value L(v_i) of the variable node v_j; and

S6: after the updating, making a decoding decision; and if the decoding decision is successful or a maximum number of iterations is reached, ending the decoding; or if the decoding decision is unsuccessful and the maximum number of iterations is not reached, going to step S2 to continue decoding.

Preferably, a formula for calculating the size of the cluster it in a protograph in step S1 is as follows:

π=m÷b_v   (2)

wherein b_v is a number of variable nodes in the base matrix.

Further, a calculation formula of the edge residual r_ci→vj is as follows:

r _{c i →v j} =∥m _{c i →v j} ^pre −m _{c i →v j} ∥  (4)

wherein m_ci→vi represents the edge information from the check node c_i to the variable node v_j, and is expressed as follows:

$\begin{array}{cc}{m}_{c_{i\to v_j}}=2{\tanh }^{-1}\left\{{\prod }_{v_b\in N\left(c_i\right)\backslash v_j}\tanh \left(\frac{m_{v_{b\to c_i}}}{2}\right)\right\}& \left(5\right)\end{array}$

wherein m_vb→ci represents a posterior probability LLR value.

Still further, a calculation formula of the PAR value corresponding to each cluster is as follows:

$\begin{array}{cc}{}_{}=\frac{{\Sigma }_{k=0}^{z_0}❘r_k❘}{z_0}& \left(6\right)\end{array}$

wherein r_k represents a residual value of a k-th edge.

Still further, an expression for updating the LLR value L(v_j) of the variable node v_j is as follows:

L(v_j)=Σ_ca∈M(vj⁾m_ca→vj+C_vj   (7)

where C_vj represents original channel information of the variable node v_j, and M(v_j) represents all check nodes connected to the variable node v_j.

Still further, in step S6, specifically, if the decoding decision is unsuccessful and the maximum number of iterations is not reached, edge information from the variable node v_j to a check node ca is updated according to formula (8):

m _{c a →v j} =Σ_{ca′∈M(vj)\ca}m_ca′→vjC_vj   (8)

wherein M(v_j)\c_a represents all the check nodes connected to the variable node v_j except the check node c_a; and

edge information is pre-computed for all check nodes C_a∈M(v_j\c_i according to equation (5), and then an edge residual is calculated according to equation (4) for a next update iteration process.

A computer system is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, where the processor implements the steps of the foregoing method when executing the computer program.

A computer-readable storage medium storing a computer program is provided, where the steps of the foregoing method are implemented when the computer program is executed by a processor.

The present disclosure has the following beneficial effects:

Based on structural characteristics of protograph-LDPC (P-LDPC) codes, the present disclosure initially proposes to extend the core of the dynamic scheduling decoding algorithm, that is, a decoding policy of preferentially updating an edge corresponding to a largest edge residual, to be updating a cluster in P-LDPC codes. In addition, the concept of PAR is introduced to describe an update level of each cluster. In this way, a range of spatial search during each iteration process in the soft-decision decoding algorithm can be greatly reduced, thereby improving the decoding throughput. Then an edge with a largest edge residual in each cluster is determined and preferentially updated. This ensures that the edge with the largest edge residual in each cluster can be updated in a determined update order, to minimize generation of silent nodes. This also prevents iterative decoding from falling into possible local optima, to lower error floors.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a flowchart of a prior-art RBP algorithm.

FIG. 2 is a Tanner graph after expansion of a protograph in Embodiment 1.

FIG. 3 is a flowchart of a method according to Embodiment 1.

FIG. 4 is a comparison diagram of bit error rates (BERs) of different decoding algorithms for LDPC codes on a fading channel.

FIG. 5 is a comparison diagram of frame error rate (FER) performance of different decoding algorithms for LDPC codes on a fading channel.

FIG. 6 is a comparison diagram of convergence performance of different decoding algorithms for LDPC codes at a fixed signal-to-noise ratio of 3.0 dB as the number of iterations increases.

DETAILED DESCRIPTION OF THE EMBODIMENTS

The present disclosure will be described in further detail below in conjunction with the accompanying drawings and specific embodiments.

Embodiment 1

In this embodiment, a corresponding check matrix can be obtained by lifting a base matrix based on a structural characteristic of a protograph. The following is an example:

$\begin{array}{cc}H=\left[\begin{array}{cccc}{\pi }_{00}& {\pi }_{01}& 0& {\pi }_{03}\\ 0& {\pi }_{11}& {\pi }_{12}& 0\\ {\pi }_{20}& 0& {\pi }_{22}& 0\end{array}\right]& \left(2\right)\end{array}$

wherein π_ij is a unit permutation matrix or its right-shift matrix, that is, a cluster.

FIG. 2 shows a Tanner graph after expansion of the protograph in the above example, where v_i is a variable node of the base matrix, and c_j is a check node of the base matrix.

Combined with the Tanner graph structural characteristics of the base matrix shown in FIG. 2, this embodiment provides a decoding method of LDPC codes based on partial average residual belief propagation (PARBP). As shown in FIG. 3, the method includes the following steps.

S1: Calculate a size of a cluster it in a protograph based on a code length m and a code rate of a target codeword. A calculation formula is as follows:

π=m÷b_v   (2)

wherein b_v is a number of variable nodes in the base matrix.

S2: Pre-compute an edge residual r_ci→vj corresponding to each edge from a variable node to a check node in a check matrix H.

A calculation formula of the edge residual r_ci→vj is as follows:

r _{c i →v j} =∥m _{c i →v j} ^pre −m _{c i →v j} ∥  (4)

wherein m_ci=vi represents edge information from a check node c_i to a variable node v_j, and is expressed as follows:

$\begin{array}{cc}{m}_{c_{i\to v_j}}=2{\tanh }^{-1}\left\{{\prod }_{v_b\in N\left(c_i\right)\backslash v_j}\tanh \left(\frac{m_{v_{b\to c_i}}}{2}\right)\right\}& \left(5\right)\end{array}$

wherein m_vb=ci represents a posterior probability LLR value.

S3: Calculate, based on π, a PAR value corresponding to each cluster in the check matrix H.

A formula for calculating the PAR value corresponding to each cluster is as follows:

$\begin{array}{cc}=\frac{{\Sigma }_{k=0}^{z_0}❘r_k❘}{z_0}& \left(6\right)\end{array}$

wherein r_k represents a residual value of a k-th edge.

S4: Sort min clusters in descending order of corresponding PAR values, and update an edge with a largest edge residual in each cluster according to formula (5).

S5: Update the edge information m_ci→vj from the check node c_i to the variable node v_j, and then update an LLR value L(v_j) of the variable node v_j.

An expression for updating the LLR value L(v_j) of the variable node v_j is as follows:

L(v_j)=Σ_ca∈M(vj)m_ca→vj+C_vj   (7)

wherein C_vj represents original channel information of the variable node v_j, and M(V_j) represents all check nodes connected to the variable node v_j.

S6: After the updating, making a decoding decision; and if the decoding decision is successful or a maximum number of iterations is reached, end the decoding; or if the decoding decision is unsuccessful and the maximum number of iterations is not reached, go to step S2 to continue decoding.

Specifically, if the decoding decision is unsuccessful and the maximum number of iterations is not reached, edge information from the variable node v_j to a check node ca is updated according to formula (8):

m _{c a →v j} =Σ_{ca′∈M(vj)\ca}m_ca′→vj+C_vj   (8)

wherein M(v_j)\c_a represents all the check nodes connected to the variable node v_j except the check node c_a.

Edge information is pre-computed for all check nodes c_a∈M(v_j)\c_i according to equation (5), and then an edge residual is calculated according to equation (4) for a next update iteration process.

Compared with the conventional dynamic scheduling decoding algorithm, the method in this embodiment gives greater consideration to the structural characteristics of codewords. For single edge protograph LDPC codes, a degree distribution and ring structure of the check matrix H largely depend on a corresponding base matrix. In the base matrix, each cluster definitely has no ring structure. An update order is measured by using a cluster residual, and an edge with a largest edge residual is found in the cluster and updated, so as to ensure that each update can achieve highly reliable information and a largest amount of information, and minimize a range of finding the edges with the largest edge residual. In this way, the computing speed can be increased locally and the spatial complexity can be reduced, thereby greatly improving the convergence speed and avoiding the greediness in global decoding to the greatest extent.

To better illustrate the advantages of the decoding method of LDPC codes proposed in this embodiment, performance comparison is performed between this method and the current mainstream dynamic scheduling decoding algorithms. To ensure performance comparison fairness, all simulations were performed under the same hardware configurations. A computer used for the simulation employs Intel(R) Core™ CPU i7-8700, a clock speed of 3.20 GHz, a memory of 16.0 GB, and a Windows 10 64-bit operating system. Specifically, randomly generated LDPC codes are transmitted on an additive white Gaussian noise (AWGN) channel, and a variety of different decoding algorithms including the algorithm in this embodiment are used for decoding, where the maximum number of iterations is set to 5, the maximum number of error frames is set to 100, a modulation coding scheme is set to binary phase shift keying (BPSK), and E_b/N₀ represents a normalized signal-to-noise ratio in decibels (dB).

FIG. 4 and FIG. 5 are comparison diagrams of BERs and FER performance of different decoding algorithms for (576, 288) LDPC codes on a fading channel. In the figures, at a low signal-to-noise ratio, for example, from 1.0 dB to 1.6 dB, error correction performance curves of different dynamic scheduling decoding algorithms almost overlap, meaning that the error correction performance of different algorithms has little difference. However, the error correction performance of the algorithms changes significantly at a signal-to-noise ratio above 1.6 dB. At 1.6 dB or below, performance curves of a residual-decaying-based RBP (RD-RBP) algorithm, a Lazy Queue Residual Decoding (LQRD) algorithm, and a dynamic silent-variable-node-free scheduling (D-SVNFS) algorithm almost overlap, but their performance declines much faster than other algorithms. Starting from 1.8 dB, as the signal-to-noise ratio increases, the PARBP algorithm gradually shows its advantages over the RD-RBP algorithm, the D-SVNFS algorithm, and the LQRD algorithm, and becomes the algorithm with the optimal decoding performance. When the BER is 1.0×10⁻⁴, compared with the suboptimal D-SVNFS algorithm, the PARBP algorithm in this embodiment increases the performance by about 0.28 dB. When the BER is 1.0×10⁻³, compared with the suboptimal D-SVNFS algorithm, the PARBP algorithm in this embodiment increases the performance by about 0.22 dB.

FIG. 6 is a comparison diagram of convergence performance of different decoding algorithms for (576, 288) LDPC codes at a fixed signal-to-noise ratio of 3.0 dB as the number of iterations increases. As shown in FIG. 5, the PARBP algorithm in this embodiment has a good convergence performance. Compared with other decoding algorithms, the PARBP algorithm achieves a nearly converged BER after the 5-th iteration. During the 5-th to 20-th iteration, the BER performance of other decoding algorithms are still converging, while the BER performance of the PARBP algorithm in this embodiment has converged as early as before the 10-th iteration. During the 5-th to 50-th iteration, the BER performance of the PARBP algorithm in this embodiment is significantly better than the BER performance of other algorithms, meaning that the PARBP algorithm in this embodiment has better convergence performance than other algorithms, especially in the speed of convergence.

Embodiment 2

A computer system is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, where the processor implements the following method steps when executing the computer program:

S1: calculating the size π of a cluster of a base matrix in a protograph based on a code length m and a code rate of a target codeword:

S2: pre-computing an edge residual r_ci→vj corresponding to each edge from a variable node to a check node in a check matrix H, where the check matrix H is obtained by lifting the base matrix based on a structural characteristic of the protograph;

S3: calculating, based on π, a PAR value corresponding to each cluster in the check matrix H;

S4: sorting m/π clusters in descending order of corresponding PAR values, and updating an edge with a largest edge residual in each cluster;

S5: updating edge information m_ci→vj from a check node c_i to a variable node v_j, and then updating an LLR value L(v_j) of the variable node v_j; and

S6: after the updating, making a decoding decision; and if the decoding decision is successful or a maximum number of iterations is reached, ending the decoding; or if the decoding decision is unsuccessful and the maximum number of iterations is not reached, going to step S2 to continue decoding.

Embodiment 3

A computer-readable storage medium storing a computer program is provided, where the following method steps are implemented when the computer program is executed by a processor:

S1: calculating the size π of a cluster of a base matrix in a protograph based on a code length m and a code rate of a target codeword:

S2: pre-computing an edge residual r_ci→vj corresponding to each edge from a variable node to a check node in a check matrix H, where the check matrix H is obtained by lifting the base matrix based on a structural characteristic of the protograph;

S3: calculating, based on π, a PAR value corresponding to each cluster in the check matrix H;

S4: sorting m/π clusters in descending order of corresponding PAR values, and updating an edge with a largest edge residual in each cluster;

S5: updating edge information c_ci→vj from a check node c_i to a variable node v_j, and then updating an LLR value L(v_j) of the variable node v_j; and

S6: after the updating, making a decoding decision; and if the decoding decision is successful or a maximum number of iterations is reached, ending the decoding; or if the decoding decision is unsuccessful and the maximum number of iterations is not reached, going to step S2 to continue decoding.

It is apparent that the above embodiments are merely intended to describe the present disclosure clearly, rather than to limit the implementations of the present disclosure. Any modification, equivalent substitution and improvement made within the spirit and principle of the present disclosure should fall within the protection scope of the claims of the present disclosure.

Claims

1. A decoding method of low-density parity-check (LDPC) codes based on partial average residual belief propagation (PARBP), wherein the method comprises the following steps: S1: calculating a size π of a cluster of a base matrix in a protograph based on a code length m and a code rate of a target codeword;S2: pre-computing an edge residual rci→vj corresponding to each edge from a variable node to a check node in a check matrix H, wherein the check matrix H is obtained by lifting the base matrix based on a structural characteristic of the protograph;S3: calculating, based on π, a partial average residual (PAR) value corresponding to each cluster in the check matrix H;S4: sorting m/π clusters in descending order of corresponding PAR values, and updating an edge with a largest edge residual in each cluster;S5: updating edge information mci→vj from a check node ci to a variable node vj, and then updating a log-likelihood ratio (LLR) value L(vj) of the variable node vj; andS6: after the updating, making a decoding decision; and if the decoding decision is successful or a maximum number of iterations is reached, ending the decoding; or if the decoding decision is unsuccessful and the maximum number of iterations is not reached, going to step S2 to continue decoding.

2. The decoding method according to claim 1, wherein a formula for calculating the size of the cluster it in the protograph in step S1 is as follows: π=m÷bv   (2)wherein bv is a number of variable nodes in the base matrix.

3. The decoding method according to claim 2, wherein a calculation formula of the edge residual rci→vj is as follows rci→vj=∥mci→vjpre−mci→vj∥  (4)wherein mci→vi represents the edge information from the check node ci to the variable node vj, and is expressed as follows:

4. The decoding method according to claim 3, wherein a formula for calculating the PAR value corresponding to each cluster is as follows:

5. The decoding method according to claim 4, wherein an expression for updating the LLR value L(vj) of the variable node vj is as follows: L(vj)=Σca∈M(vj)mca→vj+Cvj   (7)wherein Cvj represents original channel information of the variable node vj, and M(vj) represents all check nodes connected to the variable node vj.

6. The decoding method according to claim 5, wherein in step S6, specifically, if the decoding decision is unsuccessful and the maximum number of iterations is not reached, edge information from the variable node vj to a check node ca is updated according to formula (8): mca→vj=Σca′∈M(vj)\camca′→vj+Cvj   (8)wherein M(vj)\ca represents all the check nodes connected to the variable node vj except the check node ca; andedge information is pre-computed for all check nodes ca∈M(vj)\ci according to equation (5), and then an edge residual is calculated according to equation (4) for a next update iteration process.

7. A computer system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method according to claim 1 when executing the computer program.

8. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the steps of the method according to claim 1 are implemented.