CN110113057A - A kind of polarization code decoder using deep learning - Google Patents
A kind of polarization code decoder using deep learning Download PDFInfo
- Publication number
- CN110113057A CN110113057A CN201910505885.3A CN201910505885A CN110113057A CN 110113057 A CN110113057 A CN 110113057A CN 201910505885 A CN201910505885 A CN 201910505885A CN 110113057 A CN110113057 A CN 110113057A
- Authority
- CN
- China
- Prior art keywords
- decoding
- algorithm
- scan
- neural network
- decoder
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- H—ELECTRICITY
- H03—ELECTRONIC CIRCUITRY
- H03M—CODING; DECODING; CODE CONVERSION IN GENERAL
- H03M13/00—Coding, decoding or code conversion, for error detection or error correction; Coding theory basic assumptions; Coding bounds; Error probability evaluation methods; Channel models; Simulation or testing of codes
- H03M13/03—Error detection or forward error correction by redundancy in data representation, i.e. code words containing more digits than the source words
- H03M13/05—Error detection or forward error correction by redundancy in data representation, i.e. code words containing more digits than the source words using block codes, i.e. a predetermined number of check bits joined to a predetermined number of information bits
- H03M13/13—Linear codes
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Abstract
The present invention provides the decoding algorithm decoders that a kind of soft elimination (soft cancellation, the SCAN) decoding algorithm of polarization code and deep learning combine.According to the operation rule of SCAN algorithm it is found that the number of nodes of parallel computation is fewer, and the consumed time cycle is also longer when the information update position of SCAN algorithm is closer to output end.The present invention is in order to improve not short enough the problem of decoding delay existing for existing SCAN decoding algorithm, provide a kind of neural network and polarization code decoder that SCAN decoding algorithm combines, which is made of several parts such as soft elimination decoding module, Processing with Neural Network module, activation primitive module and loss function feedback modules.The present invention is replaced on the basis of former SCAN algorithm decodes factor graph, the fourth from the last layer of SCAN algorithm decoding factor graph and later part with deep learning neural network.In this way compared with originally using SCAN algorithm merely, a large amount of calculating cycle can be saved.
Description
Technical field
The invention belongs to field of channel coding, are related to a kind of new neural network and former soft elimination (soft
Cancellation, SCAN) the polarization code decoder that combines of decoding algorithm, especially this algorithm can reduce largely
Time cycle needed for decoding.
Background technique
In recent years, as the rise of artificial intelligence, neural network and deep learning also have been greatly developed, and very
The every field being applied in scientific research and life fastly.And after the concept that Arikan in 2009 proposes polarization code for the first time,
Polarization code just becomes channel coding research direction very powerful and exceedingly arrogant in recent years.Founder's Shannon Zeng Qi of information theory be " communication
Mathematical theory " it proposes in this this works, when transmitting information, if the rate of information throughput is lower than channel capacity, can look for
Make transmission error infinitely small to a kind of coding mode information.However polarization code is that a kind of currently the only mathematics that passes through strictly derives
It proves, in the sufficiently long situation of code length, a kind of coding of shannon capacity can be reached.Wherein there is scholar to find polarization code
Decoding algorithm structure have significantly similar, then a large amount of scientific research in connection type to the network structure of neural network
Personnel begin one's study the application of neural network and deep learning in polarization code decoding.
And the mainstream decoding algorithm of polarization code has serial offset (successive cancellation, SC) to decode at present
Algorithm and confidence spread (belief propagation, BP) decoding algorithm.SC decoding algorithm is a kind of decoding by bit
Algorithm, whenever translating to an information code word, it needs each the information code word decoded before using.This be by
It is determined in the construction of polarization code, the formation of polarization code needs the identical channel of multiple parameters to carry out combining channel and channel point
It splits, it is not mutually independent that this just determines that polarization code divides each channel to be formed, but mutually relevant.Also just because of
Such decoded mode, polarization code are just proved to be a kind of unique linear block codes up to Shannon Channel Capacity Limit, so
It may be said that SC decoding algorithm is a kind of decoded mode for being most suitable for polarization code structure.But such decoded mode is using string
Row decoding, thus when postpone a meeting or conference a certain bit decoding mistake that is relatively high, and working as front, codeword decoding after will lead to is wrong
It misses, that is, so-called error propagation.For this problem, correlation scholar proposed a kind of entitled serial counteracting list later
The decoding algorithm of (successive cancellation list, SCL).SCL decoding algorithm is unlike SC decoding algorithm
SCL decoding algorithm remains mulitpath in decoding, and each path can be regarded as a kind of individual SC decoding algorithm, this
Sample can avoid error code diffusion phenomena to a certain extent.But since SCL decoding algorithm saves mulitpath in decoding,
So the decoding delay ratio SC decoding algorithm of SCL decoding algorithm is higher.And BP decoding algorithm and SC and SCL decoding algorithm are not
Together, it is a kind of decoding algorithm of parallel decoding.BP decoding algorithm is constantly iterated to calculate by left information and right information, when reaching
It is made decisions after maximum number of iterations according to the left information of the leftmost side.Since it is parallel computation, institute in an iterative process
With compared with SC and SCL decoding algorithm, the decoding delay of BP is shorter and throughput is higher.And SCAN decoding algorithm can be described as SC
The combination of decoding algorithm and BP decoding algorithm.SCAN decoding algorithm has α from left to right as BP decoding algorithm and from right past
Left two kinds of transmitting information of β, and both transmit L information in the recurrence formula and BP iterative formula of information and R information is complete
It is complete consistent.Difference is that SCAN algorithm introduces the sequencing rule an of information update, as shown in formula (1).
αλ-1(ψ) < αλ(φ) < αλ(φ+1) < βλ-1(ψ) < αλ-1(ψ+1) < βλ-1(ψ+1) (1)
Wherein λ is column index, and ψ, φ are the block index in each column, while the node sequence rope in piecemeal is indicated with ω
Draw, it is evident that compared to BP algorithm indexed mode SCAN algorithm more than a kind of block index in column.And the piecemeal in each column
It is calculated according to the numeric order of piecemeal, parallel computation between each node in each piecemeal, therefore SCAN algorithm is existing parallel
Feature has serial feature again.
But with neural network polarization code decoding in extensive use, it can be found that neural network can be used to it is excellent
Change SCAN decoding algorithm.
Summary of the invention
The present invention provides a kind of mind to improve not short enough the problem of decoding delay existing for existing SCAN decoding algorithm
The polarization code decoder combined through network and SCAN decoding algorithm, the decoder is by soft elimination decoding module, neural network
A few part compositions such as module, activation primitive module and loss function feedback module are managed, as shown in Figure 1.The decoder is original SCAN
Four layers of inverse of algorithm are substituted with neural network, which largely reduces the time cycle of decoding.
Based on the above technical problem, the technical scheme adopted by the invention is as follows: the present invention provides it is a kind of can be nerve net
The decoder that network is combined with SCAN decoding algorithm.Because the operation rule according to SCAN algorithm is it is found that work as the letter of SCAN algorithm
Breath updates position closer to output end, and the number of nodes of parallel computation is fewer, and the consumed time cycle is also longer.So this
Invention concentrate SCAN decoding algorithm calculation amount last four layers are replaced with neural network, and the calculating of a neural network
Period is that the number of plies of neural network adds 1, so this makes it possible to the decoding delays for reducing polarization code, and code length is longer, such
The effect that decoder reduces time delay is more obvious.The input of DNN is the α input of all nodes of a piecemeal of the 5th column of inverse,
The output of DNN is the corresponding β output of the piecemeal.When being trained with neural network, training set is one of the 5th column of inverse
The α input of all nodes of piecemeal arranges the code word obtained after corresponding correct code word is overturn with this.At this time to code
The reason of word is overturn is if the hard decision according to SCAN algorithm is regular, if the value of information bit log-likelihood ratio is greater than
Zero, then carrying out the code word that hard decision obtains at this time should be zero, however when all sections of a piecemeal to the 5th column of inverse
When the α input of point is normalized with Sigmoid activation primitive, the α value the big more is possible to be normalized to 1.And inverse the 5th
Arranging corresponding correct code word can be multiplied to obtain by generator matrix with the correct code word of last column decoding, and reciprocal the
The α input of all nodes of one piecemeal of five column can be obtained by the log-likelihood ratio matrix of SCAN decoding algorithm.When with
When neural network is trained, need that neural network is exported β to the range of quantization to (0,1) with Sigmoid activation primitive
It is interior.Then the code word after correct code word is overturn is decoded as standard using this layer, continues to optimize mind with gradient descent method
Weighted value through network.After the output β value for finally obtaining neural network, this value can be returned to SCAN decoding algorithm, in this way
It can be carried out iterative decoding, and with the increase of the number of iterations, the decoding performance of polarization code is continuously available promotion.
The present invention is compared with the advantages of the prior art: if the length of each block be 16, according to depth be 3 network into
Row training, then it is 4 times that DNN module, which completes the calculating cycle once calculated, and corresponding SCAN algorithm calculation times are 30 times.
Therefore, calculation delay can be saved by this module being replaced with DNN network.Meanwhile because being all between all piecemeals of this column
Serial computing, therefore only need that a DNN network is arranged in hardware realization, which offers a saving resource consumption and occupy face
Product.In performance, the theoretical proof of deep learning, when frequency of training is sufficiently large, decoding performance can converge to maximum likelihood.Cause
This layer is changed into the decoding performance that DNN network is able to ascend SCAN algorithm by this.
Detailed description of the invention
Fig. 1 is neural network polarization code decoder functional diagram；
Fig. 2 is 128 code length SCAN-DNN structure charts；
Fig. 3 is that 8 code length SCAN decode factor graph；
Fig. 4 is DNN neural network structure figure.
Specific embodiment
Present invention is further described in detail with reference to the accompanying drawing, and following embodiment facilitates the understanding of the present invention,
It is relatively good application example, but is not construed as limiting thereof.
As shown in Figure 1, the decoder is by soft elimination decoding module, Processing with Neural Network module, activation primitive module and damage
Lose a few part compositions such as function feedback module.
As shown in Fig. 2, each rectangle represents the node group of 1 16 node, each dotted line frame is represented in a column
Piecemeal.Meaning in order to better understand the present invention, the present invention with code length be 128, code rate 0.5, signal-to-noise ratio 3db, channel
To be illustrated for the polarization code communication system of additive white Gaussian noise channel.It can according to the theory of polarization code decoding aspect
Know, when code length is 128, a total of 8 layers of the decoding factor graph of polarization code.Since the calculation amount of SCAN decoding algorithm is mainly concentrated
In the last several layers of of decoding factor graph, so the present invention determines that the 4th layer of pervious decoding continues to be translated with SCAN decoding algorithm
Code, the 4th layer of each piecemeal of later part are replaced with a size by the three hidden layer DNN networks of 256-128-64,
And network need to be only trained once, just can be suitably used for all decoding blocks.
Fig. 3 illustrates the partitioned mode and above-mentioned each ginseng of SCAN algorithm factor figure so that code length is the polarization code of N=8 as an example
Several corresponding relationships.As can be seen that secondary series namely when λ=1, node is divided into two pieces up and down, and whens third column is then divided into 4 pieces,
It is divided into 2 when kth columnk-1Block.For first piece of λ=1 namely ψ=0, the value of ω is 0,1,2,3.For first piece of λ=2
Namely the value of φ=0, ω are 0,1.Piecemeal in each column is calculated by numeric order, between each node in each piecemeal simultaneously
Row calculates.Therefore SCAN decoding algorithm can regard a kind of decoding algorithm that SC decoding algorithm is combined with BP decoding algorithm as.
Fig. 4 is the specific structure instead of four layers of neural network reciprocal in SCAN decoding algorithm.Training set number of the invention
According to the α value for all nodes for being fourth from the last layer this layer, these values can decode the logarithm generated in code seemingly by MATLAB
So obtained than matrix.And comparing collection is then this layer of correct code word of decoding.Due to polarization code changed factor figure connection type simultaneously
The final decoding value of polarization code is not influenced, so fourth from the last layer decodes correct code word and can be decoded correctly by the last layer
Code word and generator matrix G8Multiplication obtains.And if entirely decoding right-on, the correct code word of that the last layer decoding
Should be identical with code word is sent, so fourth from the last layer decodes correct code word and can also be gone with originally transmitted code word multiplied by generation
Matrix G8It obtains.And generator matrix is then matrix F=[1,0；1,1] it by Kronecker product three times, is then turned over using bit
Turn to obtain.And this neural network has made 1100000 samples as data set, and degree of parallelism when training is set as 8,
Epoch is 212, and using the innovatory algorithm Adadelta of stochastic gradient descent method as the parameter optimization algorithm pair of the DNN network
Network is trained.According to Fig. 4 as can be seen that the output of neural network should be theoretically the β of all nodes of fourth from the last layer
Output valve, because of β output valve i.e. LLR ratio, its value range and is compared from minus infinity to positive infinity
Value range to collection is { 0,1 }, so the β value exported to network is needed to be normalized, it is normalized to [0,1] range
Interior, this process can be completed with the Sigmoid activation primitive in neural network.Then the network output valve after normalization
It is compared with collection is compared, then calculates loss, loss is the parameter for being used to measure neural network learning effect.Root
According to Fig. 4 as can be seen that feeding back when the neural network output valve that is not normalized into neural network, polarization code is just at this time
It can be iterated decoding, and with the increase of the number of iterations, the decoding performance of polarization code can be significantly improved.
The above embodiments do not limit the invention in any form, all using similar structure of the invention, method and its similar
Variation pattern technical solution obtained, in protection scope of the present invention.
Claims (3)
1. a kind of improved deep neural network (deep neural network, DNN) is in conjunction with polarization code SCAN decoding algorithm
Polarization code decoder.The decoder is by general SCAN decoding algorithm decoder and DNN deep neural network decoder
Two module compositions.And DNN neural network module only need to be trained once, so that it may suitable for all decoding blocks.
2. a kind of training set and the preparation method for comparing collection.The code word of pilot process is to be not readily available when general decoding
, the present invention using the connection type of decoding factor graph have no effect on decoding correctness and generator matrix be can inverse matrix original
Reason decodes correct code word and generator matrix G using the last layer8The mode of multiplication is come comparison needed for obtaining neural metwork training
Collection.
3. one kind can iteration decoded mode.The present invention protect neural network export log-likelihood ratio β value, when the value after
It is continuous feed back to neural network after, this decoder can be carried out iterative decoding, and with the increase of the number of iterations, decoder
Decoding performance is continuously available improvement.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910505885.3A CN110113057B (en) | 2019-06-12 | 2019-06-12 | Polarization code decoder utilizing deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910505885.3A CN110113057B (en) | 2019-06-12 | 2019-06-12 | Polarization code decoder utilizing deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110113057A true CN110113057A (en) | 2019-08-09 |
| CN110113057B CN110113057B (en) | 2023-06-13 |
Family
ID=67494762
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910505885.3A Active CN110113057B (en) | 2019-06-12 | 2019-06-12 | Polarization code decoder utilizing deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110113057B (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110572166A (en) * | 2019-09-19 | 2019-12-13 | 天津大学 | BCH code decoding method based on deep learning |
| CN111106839A (en) * | 2019-12-19 | 2020-05-05 | 北京邮电大学 | Polarization code decoding method and device based on neural network |
| CN111697975A (en) * | 2020-06-01 | 2020-09-22 | 西安工业大学 | Polarization code continuous deletion decoding optimization algorithm based on full-connection neural network |
| CN113395138A (en) * | 2021-06-15 | 2021-09-14 | 重庆邮电大学 | PC-SCMA joint iterative detection decoding method based on deep learning |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109450459A (en) * | 2019-01-16 | 2019-03-08 | 中国计量大学 | A kind of polarization code FNSC decoder based on deep learning |
| US20190095798A1 (en) * | 2017-09-28 | 2019-03-28 | D5Ai Llc | Stochastic categorical autoencoder network |
-
2019
- 2019-06-12 CN CN201910505885.3A patent/CN110113057B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20190095798A1 (en) * | 2017-09-28 | 2019-03-28 | D5Ai Llc | Stochastic categorical autoencoder network |
| CN109450459A (en) * | 2019-01-16 | 2019-03-08 | 中国计量大学 | A kind of polarization code FNSC decoder based on deep learning |
Non-Patent Citations (2)
| Title |
|---|
| 王秀敏等: "极化码译码算法研究现状及发展前景", 《中国计量大学学报》 * |
| 肖韬辉等: "深度学习的快速磁共振成像及欠采样轨迹设计", 《中国图象图形学报》 * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110572166A (en) * | 2019-09-19 | 2019-12-13 | 天津大学 | BCH code decoding method based on deep learning |
| CN110572166B (en) * | 2019-09-19 | 2023-04-18 | 天津大学 | BCH code decoding method based on deep learning |
| CN111106839A (en) * | 2019-12-19 | 2020-05-05 | 北京邮电大学 | Polarization code decoding method and device based on neural network |
| CN111697975A (en) * | 2020-06-01 | 2020-09-22 | 西安工业大学 | Polarization code continuous deletion decoding optimization algorithm based on full-connection neural network |
| CN113395138A (en) * | 2021-06-15 | 2021-09-14 | 重庆邮电大学 | PC-SCMA joint iterative detection decoding method based on deep learning |
| CN113395138B (en) * | 2021-06-15 | 2022-05-03 | 重庆邮电大学 | PC-SCMA joint iterative detection decoding method based on deep learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110113057B (en) | 2023-06-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110113057A (en) | A kind of polarization code decoder using deep learning | |
| CN105811998B (en) | A kind of polarization code constructing method and polarization code coding/decoding system based on density evolution | |
| CN107241106B (en) | Deep learning-based polar code decoding algorithm | |
| Nachmani et al. | Hyper-graph-network decoders for block codes | |
| Xu et al. | Polar decoding on sparse graphs with deep learning | |
| CN107204780B (en) | Merging BP decoding algorithm and device of polar-LDPC (Low Density parity check) concatenated code | |
| CN110336567B (en) | Joint iterative decoding method applied to G-LDPC coding cooperation | |
| CN107565978B (en) | BP decoding method based on Tanner graph edge scheduling strategy | |
| CN109039534A (en) | A kind of sparse CDMA signals detection method based on deep neural network | |
| CN105515590A (en) | Successive cancellation list polarization code decoding algorithm with effective low complexity based on random binary data flows and decoding structural frame thereof | |
| CN110233628B (en) | Self-adaptive belief propagation list decoding method for polarization code | |
| CN109450459B (en) | Polarization code FNSC decoder based on deep learning | |
| CN106941394B (en) | Joint detection decoding method and device for SCMA (sparse code multiple access) coded by polarization code | |
| Lian et al. | What can machine learning teach us about communications? | |
| CN110138390A (en) | A kind of polarization code SSCL algorithm decoder based on deep learning | |
| Cyriac et al. | Polar code encoder and decoder implementation | |
| CN114268328A (en) | Convolutional code decoding method based on bidirectional LSTM and convolutional code encoding and decoding method | |
| CN112332863B (en) | Polar code decoding algorithm, receiving end and system under low signal-to-noise ratio scene of low orbit satellite | |
| CN111835364B (en) | Low-complexity nerve BP decoding method of polarization code | |
| CN106998240A (en) | A kind of interpretation method and decoder | |
| Nachmani et al. | Autoregressive belief propagation for decoding block codes | |
| Qingle et al. | A low complexity model-driven deep learning ldpc decoding algorithm | |
| CN116614142A (en) | Combined decoding method based on BPL decoding and OSD decoding | |
| Teng et al. | Convolutional neural network-aided bit-flipping for belief propagation decoding of polar codes | |
| CN106941393B (en) | LDPC interpretation method and system based on SOR |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |