CN113395138A - PC-SCMA joint iterative detection decoding method based on deep learning - Google Patents

PC-SCMA joint iterative detection decoding method based on deep learning Download PDF

Info

Publication number
CN113395138A
CN113395138A CN202110660122.3A CN202110660122A CN113395138A CN 113395138 A CN113395138 A CN 113395138A CN 202110660122 A CN202110660122 A CN 202110660122A CN 113395138 A CN113395138 A CN 113395138A
Authority
CN
China
Prior art keywords
information
node
resource
user
factor graph
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202110660122.3A
Other languages
Chinese (zh)
Other versions
CN113395138B (en
Inventor
彭大芹
何彦琦
黄萍
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Chongqing University of Post and Telecommunications
Original Assignee
Chongqing University of Post and Telecommunications
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Chongqing University of Post and Telecommunications filed Critical Chongqing University of Post and Telecommunications
Priority to CN202110660122.3A priority Critical patent/CN113395138B/en
Publication of CN113395138A publication Critical patent/CN113395138A/en
Application granted granted Critical
Publication of CN113395138B publication Critical patent/CN113395138B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L1/00—Arrangements for detecting or preventing errors in the information received
    • H04L1/004—Arrangements for detecting or preventing errors in the information received by using forward error control
    • H04L1/0045—Arrangements at the receiver end
    • H04L1/0047—Decoding adapted to other signal detection operation
    • H04L1/0048—Decoding adapted to other signal detection operation in conjunction with detection of multiuser or interfering signals, e.g. iteration between CDMA or MIMO detector and FEC decoder
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • 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
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L1/00—Arrangements for detecting or preventing errors in the information received
    • H04L1/004—Arrangements for detecting or preventing errors in the information received by using forward error control
    • H04L1/0045—Arrangements at the receiver end
    • H04L1/0047—Decoding adapted to other signal detection operation
    • H04L1/005—Iterative decoding, including iteration between signal detection and decoding operation
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L1/00—Arrangements for detecting or preventing errors in the information received
    • H04L1/004—Arrangements for detecting or preventing errors in the information received by using forward error control
    • H04L1/0056—Systems characterized by the type of code used

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Molecular Biology (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Error Detection And Correction (AREA)

Abstract

The invention belongs to the technical field of computers and communication, and particularly relates to a joint iterative detection decoding method based on deep learning; the method comprises adding a learnable weight factor in a transfer path between a resource node and a user node in a factor graph of a message transfer algorithm of the SCMA; adding a learnable offset which can be learnt off line in the information iteration in a factor graph of the confidence coefficient propagation algorithm; and forming a combined factor graph; using learnable weight factors and learnable offset in the joint factor graph as hidden layer parameters of the deep neural network, and inputting scrambled information bit values to perform iterative updating in the deep neural network; sequentially and iteratively updating resource node information, prior node information and user node information; after updating the resource node information, the prior information of the user node and the user node information, outputting an estimated transmission symbol; the invention can effectively improve the performance of detecting the BER of the decoding within the acceptable range when the operation complexity is increased.

Description

PC-SCMA joint iterative detection decoding method based on deep learning
Technical Field
The invention relates to a Deep Learning (DL) -based PC-SCMA (Polar Coded-Sparse Code Multiple Access) Joint Iterative Detection Decoding (JIDD) method, which has the characteristics of low operation complexity, high Decoding and Detection accuracy and the like, can be applied to scenes with the requirements of high frequency spectrum utilization rate, massive user Access and the like in a 5 th generation mobile communication system, and belongs to the technical field of computers and communication.
Background
With the continuous development of mobile communication, the performance requirements of diversified application scenarios on communication technology are higher and higher. The 5G mobile communication network is expected to achieve higher spectral efficiency, which puts demands on channel coding and decoding technology and improvement of multiple access technology, the channel coding and decoding technology is directly related to the quality of system error rate performance, and the multiple access technology aims to provide shared carrier resources for a plurality of user terminals, so as to meet the demand of higher spectral efficiency.
At present, compared with other channel coding techniques, the Polar code (Polar) technique is the only channel coding technique which is theoretically proven to reach the shannon limit in the memoryless channel, and is recognized as a 5G control channel coding scheme by the 3GPP organization. And the complexity of coding and decoding is low, and the method has wide application prospect. On the other hand, the Sparse Code Multiple Access (SCMA) technology is mainly developed by hua as a company, has the advantages of high spectrum utilization rate, high user overload rate, low time delay and the like, and is widely concerned by the industrial and academic circles. In addition, in recent years, deep learning techniques have shown significant optimization effects in many fields, including optimization of algorithms by deep learning techniques in the field of communications.
However, as the length of the polar code and the number of iterations of the decoding algorithm increase, a large number of multiplication operations and storage space occupation are increased by combining a large number of learnable parameters introduced after the deep learning technique, and therefore how to improve the performance of the decoding BER without increasing too high operation complexity becomes a current popular research direction. On the other hand, the SCMA system multi-user detection algorithm combined with deep learning also has the problems of high operation complexity or unsatisfactory detection performance and the like and needs to be solved urgently.
Disclosure of Invention
The invention aims to provide a joint iterative detection decoding method for improving the performance of multi-user detection and decoding error rate aiming at a PC-SCMA system. The invention firstly uses an Offset Min-Sum (OMS) algorithm which can be learned to make up for the loss caused by the minimum Sum approximation in a Belief Propagation (BP) decoding algorithm, improves the BER performance of the BP decoding algorithm on the premise of not increasing additional multiplication operation, and improves the convergence speed. Besides, a Recurrent Neural Network (RNN) is adopted to realize parameter sharing of a multilayer Neural Network, a Linear rectification function (ReLU) is added as an activation function of a hidden layer, and a Hinge loss function is adopted to replace a Cross Entropy (Cross Entropy) loss function to further reduce off-line training overhead, and the decoding algorithm is named as RNN-OMS-BP decoding algorithm. Then adding a Weight factor capable of off-line learning into a Message Passing Algorithm (MPA) of an SCMA system, realizing the capability of off-line training to improve the Message Passing accuracy by adding the Weight factor, and expanding a combined factor graph framework of a decoding Algorithm and a Weight-MPA Algorithm into a deep neural network framework by combining the RNN-OMS-BP decoding Algorithm provided above, thereby providing a PC-SCMA combined iterative detection decoding method based on deep learning, and having the cost of slightly increasing the operation complexity.
The invention adopts the following technical scheme:
a PC-SCMA joint iterative detection decoding method based on deep learning, the method comprises the following steps:
s1, performing multi-user detection by using a message passing algorithm of SCMA, and adding learnable weight factors in a passing path between a resource node and a user node in a factor graph;
s2, decoding the polarization code by using a confidence coefficient propagation algorithm, and adding a learnable offset which can be learnt off line in the information iteration in the factor graph;
s3, cascading the factor graph of multi-user detection and the factor graph of polar code decoding to form a combined factor graph;
s4, using learnable weight factors and learnable offsets in the joint factor graph as hidden layer parameters of a deep neural network, and inputting scrambled information bit values to perform iterative updating in the deep neural network;
s5, updating the resource node information, and transmitting the resource node information into a decoding part of the joint factor graph after the updating is finished;
s6, converting the resource node information into prior information for a decoding algorithm to execute an information iteration process of a polar code decoding algorithm;
s7, after the user node receives the prior information, the updated information is transmitted to the adjacent resource node, and the user node information is updated in an iterative manner;
and S8, outputting the estimated transmission symbol after updating the resource node information, the prior information of the user node and the user node information.
The invention has the beneficial effects that:
the method firstly utilizes a message transfer algorithm based on a weight factor and a confidence coefficient propagation algorithm based on learnable offset minimum sum to replace an original message transfer algorithm and a confidence coefficient propagation algorithm adopted in the traditional combined iterative detection decoding algorithm, and develops a combined factor graph architecture in the iterative detection decoding algorithm into a deep neural network. The invention adopts the setting of learnable parameters and the selection of a reasonable gradient descent algorithm and a loss function, thereby improving the overall performance of the PC-SCMA combined iterative detection decoding method based on deep learning through an off-line training process. The PC-SCMA joint iterative detection decoding method based on deep learning realizes certain improvement on BER performance, increases the operation complexity within an acceptable range, and has certain practical significance and inspiration on the design idea of a multi-user detection and decoding method. The invention can effectively improve the performance of detecting the BER in decoding when the operation complexity is increased within an acceptable range. The introduction of the deep learning technology and the neural network architecture enables the massive parameter adjustment and calculation simplification in the invention to bring more possibilities, and the method has wide application prospect in a fifth generation mobile communication network with larger requirements on the frequency spectrum utilization rate, the user access quantity and the throughput.
Drawings
FIG. 1 is a block diagram of PC-SCMA joint iterative detection decoding according to an embodiment of the present invention;
FIG. 2 is a flowchart of a PC-SCMA joint iterative detection decoding method based on deep learning according to an embodiment of the present invention;
FIG. 3 is a joint factor graph of a joint iterative detection decoding algorithm according to an embodiment of the present invention;
FIG. 4 is a deep neural network architecture corresponding to a single iteration process in an embodiment of the present invention;
fig. 5 is a factor graph corresponding to the right information log-likelihood ratio and the left information log-likelihood ratio in the embodiment of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be described clearly and completely with reference to the accompanying drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
Fig. 1 is a frame diagram of PC-SCMA joint iterative detection decoding adopted in the embodiment of the present invention, and as shown in fig. 1, the number of users is denoted as J, and each user sends an information bit to a polar code encoder; the polar code encoder outputs coded bits to the random interleaver; the random interleaver scrambles the coded bits and forms out-of-order coded bits; the method comprises the steps that disordered coded bits are sent to an SCMA multi-user detector through a channel, the SCMA multi-user detector sends out bit-outside information to a de-interleaver, the de-interleaver sends decoding first-check information to a polar code decoder after de-interleaving, on one hand, the polar code decoder outputs estimated information bits of a user, on the other hand, the polar code decoder outputs symbol-outside information, a priori information is detected through soft information conversion, and the priori information is sent to the SCMA multi-user detector to finish joint iterative detection decoding.
In the invention, the length of a polarization code is recorded as N, the SCMA constellation dimension is recorded as M, the number of orthogonal resources is recorded as K, the number of users is recorded as J, and the original information bit set of the user is recorded as U ═{u1,u2,…,uJAnd the code word information set after being coded by the polarization code coder is recorded as C ═ C1,c2,…,cJAnd recording as B ═ B after interleaving is finished1,b2,…,bJThe random interleaver can reduce the correlation between adjacent bits, and then the correlation is mapped by an SCMA codebook to be marked as a complex codeword set X ═ X1,x2,…,xL},L＝N/log2(M), the received codeword set after channel transmission is denoted as Y ═ Y1,y2,…,yL},L＝N/log2(M)，
Figure BDA0003114874420000041
And
Figure BDA0003114874420000042
respectively is prior information and external information transmitted between the detection end and the decoding end for the jth user.
Fig. 2 is a flowchart of a PC-SCMA joint iterative detection decoding method based on deep learning in an embodiment of the present invention, and as shown in fig. 2, the joint iterative detection decoding method includes:
s1, performing multi-user detection by using a message passing algorithm of SCMA, and adding learnable weight factors in a passing path between a resource node and a user node in a factor graph;
in this embodiment, a Message Passing Algorithm (MPA) in a conventional joint iterative detection decoding Algorithm is replaced with a Weight-MPA Algorithm to perform multi-user detection; two kinds of information are transmitted in the factor graph of multi-user detection, and the information is respectively K resource nodes rkK is to J user nodes u, 1,2jJ1, 2.. J conveys information
Figure BDA0003114874420000051
And with user node ujTo the resource node rkInformation to be transferred
Figure BDA0003114874420000052
Therefore, the process of adding the learnable weight factor in the transmission path between the resource node and the user node in the factor graph of the multi-user detection is represented as:
Figure BDA0003114874420000053
wherein the content of the first and second substances,
Figure BDA0003114874420000054
indicating by the user node u during i iterationsjTo the resource node rkTransmitting the message;
Figure BDA0003114874420000055
indicating the node of the resource in the course of i iterations
Figure BDA0003114874420000056
To user node ujTransmitting the message;
Figure BDA0003114874420000057
representing resource nodes in the course of i iterations
Figure BDA0003114874420000058
To user node ujA learnable weight factor on the propagation path; { zeta-meterjV k represents resource-removing node rkForeign and user node ujAll resource nodes connected.
S2, decoding the polarization code by using a confidence coefficient propagation algorithm, and adding a learnable offset which can be learnt off line in the information iteration in the factor graph;
in this embodiment, a learnable Offset Min-Sum (OMS) algorithm is used to compensate for the loss caused by the minimum Sum approximation in the polar code Belief Propagation (BP) decoding algorithm.
The process of adding the learnable offset amount capable of being learnt off line in the information iteration in the factor graph of the belief propagation algorithm comprises the step of adding the learnable offset amount beta on the operation function g (a, b) of the right information log-likelihood ratio and the left information log-likelihood ratio on the factor graph nodes of each stage (stage) in the factor graph in the hidden layer of the deep neural network corresponding to each iteration process, wherein the operation function can be rewritten as follows:
g'(a,b,β)＝sign(a)sign(b)fReLU(min(|a|,|b|)-β,0) (2)
wherein g' (a, b, β) represents an information iteration function; sign (a) represents a sign function of signal a; sign (b) represents a sign function of signal b; f. ofReLUAn activation function representing a deep neural network; β represents a learnable offset of offline learning; min (| a |, | b |) represents the minimum of the absolute value for signal a and the absolute value for signal b.
S3, cascading the factor graph of multi-user detection and the factor graph of polar code decoding to form a combined factor graph;
the method comprises the following steps of combining factor graphs of an SCMA (sparse code multiple access) multi-user detection algorithm and a polar code decoding algorithm into a joint factor graph, wherein the joint factor graph of the polar code length N being 8, the number J of user nodes being 6 and the number K of orthogonal resources being 4 is shown in FIG. 3, wherein a JIDD algorithm message transmission process executed in the joint factor graph is mainly realized through external iteration of a multi-user detection end and a decoding end, the algorithm can be divided into six steps of (i) - (sixth), and the specific execution process is as follows:
in the resource node updating process, the resource node updates own information according to the signals received from the channel and transmits the information to the adjacent user nodes. This process is shown as step (r) in fig. 3. The prior information updating process is shown as steps from (c) - (v) in fig. 3. In the process, the output external information of the resource node updating process is transmitted to the polar code decoder, and the polar code decoder performs a complete left-right information transmission process. And finally, the user node updates, namely the step sixthly in fig. 3, and the user node calculates the prior information according to the external information output by the polar code decoder. The user node updating process is the last process of a JIDD algorithm single iteration, and if the next iteration is to be started, the user node updates the information of the user node and transmits the information to the adjacent resource node.
S4, using the learnable weight factor and the learnable offset in the joint factor graph as hidden layer parameters of the deep neural network, and inputting the scrambled information bit values to perform iterative updating in the deep neural network.
Deep learning is further classified into different models, such as a feedforward Neural Network (DNN), a Convolutional Neural Network (CNN), a Recurrent Neural Network (RNN), and the like; the invention can select one or more neural networks to be applied as the deep neural network of the invention.
The joint factor graph framework in the invention is expanded into a larger deep neural network, and joint information iteration added with learnable weight factors and learnable offsets is performed in the neural network. The hidden layer in the network architecture is developed by a joint factor graph, the number of users J is 6, the number of orthogonal resources K is 4, the length of a polarization code N is 8, and N is log2The deep neural network architecture with N-3 is shown in fig. 4. In the deep neural network architecture of the DNN-JIDD method, a neural network formed by I times of joint information iterative conversion consists of a 1-layer input layer, a 1-layer output layer and an I (2n +3) layer hidden layer, and in the invention, the neural network hidden layer expanded by a single complete joint information iterative process also consists of three parts: a resource node updating layer, a prior information updating layer and a user node updating layer.
In the embodiment of the present invention, the deep neural network is trained by inputting the scrambled information bit values as a training set.
S5, updating the resource node information, and transmitting the resource node information into a decoding part of the joint factor graph after the updating is finished;
in this embodiment, node updating is performed in a hidden layer serving as a resource node updating layer, where the number of hidden layer layers corresponding to a single joint iteration is 2n +3, where the hidden layer includes 2 resource node updating layers, resource node information is updated in the resource node updating layer, and mathematical expressions adopted in the executed steps are as shown in the following formulas (3) and (4):
before the iteration starts, because all the transmitted symbols are equal in probability, a message sent by the user node to the resource node is initialized to:
Figure BDA0003114874420000071
wherein the content of the first and second substances,
Figure BDA0003114874420000072
indicating the user node u in the initial iteration processjTo the resource node rkTransmitting the message; m denotes the SCMA constellation dimension, J1, …, J denotes the number of users; k is 1, …, and K represents the number of orthogonal resources.
After the initial message sending, the resource node information needs to be updated, and after the receiving end receives the signal, the resource node r of the SCMA detectorkWill update its information and transfer it to the connected user node ujAnd can be expressed as:
Figure BDA0003114874420000073
wherein the content of the first and second substances,
Figure BDA0003114874420000074
indicating by the user node u during i iterationsjTo the resource node rkTransmitting the message; wherein
Figure BDA0003114874420000081
A conditional probability density function representing all codeword combinations under a gaussian white noise channel,
Figure BDA0003114874420000082
Figure BDA0003114874420000083
indicating by the user node during i-1 iterations
Figure BDA0003114874420000084
To the resource node rkTransmitting the message; { zeta-meterkW represents user node ujForeign and resource nodes rkA set of connected user nodes;
Figure BDA0003114874420000085
representing predicted symbol values
Figure BDA0003114874420000086
And the actual transmission symbol value xjIdentical case where xjRepresenting by user node ujThe symbol value passed.
In some preferred embodiments, considering that the resource node needs to calculate the off-symbol information through the updated information after updating, the process may be expressed as:
Figure BDA0003114874420000087
wherein the content of the first and second substances,
Figure BDA0003114874420000088
representing SCMA symbol extrinsic information; l represents the SCMA complex field codeword number. { zeta-meterjDenotes a user node ujAll resource nodes connected; the invention also needs to convert the off-symbol information
Figure BDA0003114874420000089
Reconverting to extra-bit information
Figure BDA00031148744200000810
For the polar code decoding algorithm to perform the operation, the conversion process can be expressed as:
Figure BDA00031148744200000811
since the input value of the polar code decoding algorithm is generally in the LLR form, the above bit extrinsic information needs to be converted into the LLR form again, which can be expressed as:
Figure BDA00031148744200000812
wherein m is more than or equal to 1 and less than or equal to Q, and Q is log2(M)。
Inputting the log-likelihood ratio information into a corresponding polarization code decoding part in a joint factor graph, wherein the last step of the resource node updating process is de-interleaving, and can be represented as:
Figure BDA00031148744200000813
wherein the content of the first and second substances,
Figure BDA00031148744200000814
representing the deinterleaved log-likelihood ratio information.
S6, converting the resource node information into prior information for a decoding algorithm to execute an information iteration process of a polar code decoding algorithm;
in this embodiment, node update is performed on a hidden layer serving as a priori information update layer, a Recurrent Neural Network (RNN) may be used as an architecture base in the deep Neural Network of the present invention, but considering that a joint iterative detection decoding method of a feedforward Neural Network (DNN) can improve BER performance by priority and consider operation complexity secondarily, a cyclic architecture is not used in the embodiment of the present invention to implement multiplexing of parameters in the Network. The number of the prior information updating layers is 2n-1, wherein the executed operation steps comprise mutual conversion of sign information and bit information and an information iteration process of an OMS-BP decoding algorithm, a calculation process of the OMS-BP decoding algorithm is executed based on a joint factor graph in fig. 3, and a learnable function g (a, b) of a right information log-likelihood ratio and a left information log-likelihood ratio is added on a factor graph node of each stage (stage) in the factor graph shown in fig. 5An offset β; subscripts i, j denote the number of stages and the number of nodes, respectively, superscript t denotes the current iteration number, and each node (i, j) delivers two types of LLR information: right information log-likelihood ratio passed from left to right
Figure BDA0003114874420000091
And left information log-likelihood ratio passing from right to left
Figure BDA0003114874420000092
The operation process can be expressed as:
Figure BDA0003114874420000093
g'(a,b,β)＝sign(a)sign(b)fReLU(min(|a|,|b|)-β,0) (10)
fReLU(x)＝max(0,x) (11)
wherein the content of the first and second substances,
Figure BDA0003114874420000094
representing the log-likelihood ratio of left information transferred from right to left by j factor graph nodes in the i stage in the t iteration processes;
Figure BDA0003114874420000095
representing the right information log-likelihood ratio of the j factor graph nodes in the i stage passing from left to right in the t iteration processes;
Figure BDA0003114874420000096
representing nodes of an i-phase j-factor graph in the hidden layer of a deep neural network during t iterations
Figure BDA0003114874420000097
A learnable offset of; n represents a polarization code length; g' (a, b, β) represents an information iteration function; sign (a) represents a sign function value of the signal a; sign (b) represents a sign function value of the signal b; f. ofReLUAn activation function representing a deep neural network; β represents a learnable offset of offline learning; parameter L and R are indicated by superscriptThe number of iterations, parameter L, R, and β are shown with the left-hand subscript indicating the number of stages performed and the right-hand subscript indicating the number of nodes.
After completing an iterative process, the left information LLR passed to the leftmost needs to be converted into probability domain information and deinterleaved to provide a priori information for the multi-user detector, which can be expressed as:
Figure BDA0003114874420000101
Figure BDA0003114874420000102
s7, after the user node receives the prior information, the updated information is transmitted to the adjacent resource node, and the user node information is updated in an iterative manner;
in this embodiment, the hidden layer serving as the user information update layer is used to update the user node, and after the user node receives the prior information, the updated information is transmitted to the resource node adjacent to the user node again, and the next iteration process is started, which may be represented as:
Figure BDA0003114874420000103
wherein the content of the first and second substances,
Figure BDA0003114874420000104
representing by user node ujTo the resource node rkCommunicated user node ujExternal information of the ith SCMA symbol;
Figure BDA0003114874420000105
representing user node ujExternal information of the ith SCMA symbol, i.e. the obtained prior information;
Figure BDA0003114874420000106
representing by resource node rkTo user node ujPassed resource node rkExternal information of the ith SCMA symbol; { zeta-meterjV k represents resource-removing node rkForeign and user node ujAnd (4) collecting all connected resource nodes.
S8, after updating the resource node information, the prior information of the user node, and the user node information, outputting an estimated transmission symbol, where the output process may be represented as:
Figure BDA0003114874420000107
wherein the content of the first and second substances,
Figure BDA0003114874420000111
representing user node ujI.e. the estimated transmission value of the nth bit of user j,
Figure BDA0003114874420000112
and the log-likelihood ratio value after the decoding of the user j is completed after the T iterations are completed.
In some preferred embodiments, after step S8, the method further includes using a loss function to evaluate the performance loss during the training process, and then calculating an optimal set of weighting factors and offsets through a gradient descent algorithm and back propagation; and taking the optimal weight factor and the optimal offset as hidden layer parameters of the deep neural network to finish the updating of the resource node, the prior information and the user node information.
In the description of the present invention, it is to be understood that the terms "coaxial", "bottom", "one end", "top", "middle", "other end", "upper", "one side", "top", "inner", "outer", "front", "center", "two ends", and the like, indicate orientations or positional relationships based on those shown in the drawings, and are only for convenience of description and simplicity of description, and do not indicate or imply that the device or element being referred to must have a particular orientation, be constructed and operated in a particular orientation, and thus, are not to be construed as limiting the present invention.
In the present invention, unless otherwise expressly stated or limited, the terms "mounted," "disposed," "connected," "fixed," "rotated," and the like are to be construed broadly, e.g., as meaning fixedly connected, detachably connected, or integrally formed; can be mechanically or electrically connected; the term "a" or "an" refers to a term that can be used in a generic sense, and includes, but is not limited to, a generic term, a specific term or a specific term.
Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that changes, modifications, substitutions and alterations can be made in these embodiments without departing from the principles and spirit of the invention, the scope of which is defined in the appended claims and their equivalents.

Claims (8)

1. A PC-SCMA joint iterative detection decoding method based on deep learning is characterized in that the method comprises the following steps:
s1, performing multi-user detection by using a message passing algorithm of SCMA, and adding learnable weight factors in a passing path between a resource node and a user node in a factor graph;
s2, decoding the polarization code by using a confidence coefficient propagation algorithm, and adding a learnable offset which can be learnt off line in information iteration in a factor graph;
s3, cascading the factor graph of multi-user detection and the factor graph of polar code decoding to form a combined factor graph;
s4, using learnable weight factors and learnable offsets in the joint factor graph as hidden layer parameters of a deep neural network, and inputting scrambled information bit values to perform iterative updating in the deep neural network;
s5, updating the resource node information, and transmitting the resource node information into a decoding part of the joint factor graph after the updating is finished;
s6, converting the resource node information into prior information for a decoding algorithm to execute an information iteration process of a polar code decoding algorithm;
s7, after the user node receives the prior information, the updated information is transmitted to the resource node adjacent to the user node, and the user node information is updated in an iterative manner;
and S8, outputting the estimated transmission symbol after updating the resource node information, the prior information of the user node and the user node information.
2. The PC-SCMA joint iterative detection decoding method based on deep learning of claim 1, wherein the process of adding learnable weight factors in the transmission path between the resource nodes and the user nodes in the factor graph of multi-user detection is represented as follows:
Figure FDA0003114874410000011
wherein the content of the first and second substances,
Figure FDA0003114874410000012
indicating by the user node u during i iterationsjTo the resource node rkTransmitting the message;
Figure FDA0003114874410000021
indicating the node of the resource in the course of i iterations
Figure FDA0003114874410000022
To user node ujTransmitting the message;
Figure FDA0003114874410000023
representing resource nodes in the course of i iterations
Figure FDA0003114874410000024
To the user sectionPoint ujA learnable weight factor on the propagation path; { zeta-meterjV k represents resource-removing node rkForeign and user node ujAll resource nodes connected.
3. The PC-SCMA joint iterative detection decoding method based on deep learning of claim 1, wherein the process of adding the learnable offset which can be learned through off-line in the information iteration in the factor graph comprises that in the hidden layer of the deep neural network corresponding to each iteration process, the factor graph nodes of each stage add the offset on the operation functions of the right information log-likelihood ratio and the left information log-likelihood ratio, which are expressed as:
g'(a,b,β)＝sign(a)sign(b)fReLU(min(|a|,|b|)-β,0)
wherein g' (a, b, β) represents an information iteration function; sign (a) represents a sign function of signal a; sign (b) represents a sign function of signal b; f. ofReLUAn activation function representing a deep neural network; β represents a learnable offset of offline learning.
4. The PC-SCMA joint iterative detection decoding method based on deep learning of claim 1, wherein the updating process of the resource node information comprises:
Figure FDA0003114874410000025
wherein the content of the first and second substances,
Figure FDA0003114874410000026
indicating by the user node u during i iterationsjTo the resource node rkTransmitting the message;
Figure FDA0003114874410000027
a conditional probability density function representing all codeword combinations under a Gaussian white noise channel;
Figure FDA0003114874410000028
representing predicted symbol values
Figure FDA0003114874410000029
And the actual transmission symbol value xjA consistent condition;
Figure FDA00031148744100000210
indicating by the user node during i-1 iterations
Figure FDA00031148744100000211
To the resource node rkTransmitting the message; { zeta-meterkW represents user node ujForeign and resource nodes rkAll connected user nodes.
5. The PC-SCMA joint iterative detection decoding method according to claim 1, wherein the step S5 of passing the information of the user node into the decoding portion of the joint factor graph after the updating is completed includes calculating the extra-symbol information from the updated resource node information; and converting the off-symbol information into off-bit information, converting the off-bit information into log-likelihood ratio information, and inputting the log-likelihood ratio information into a corresponding polarization code decoding part in a joint factor graph.
6. The deep learning-based PC-SCMA joint iterative detection decoding method as claimed in claim 1, wherein the information iteration process of the polar code decoding algorithm in step S6 includes:
Figure FDA0003114874410000031
wherein the content of the first and second substances,
Figure FDA0003114874410000032
to representLeft information log-likelihood ratio transmitted from right to left by j factor graph nodes in the i stage in the t iteration processes;
Figure FDA0003114874410000033
representing the right information log-likelihood ratio transmitted from left to right of the j factor graph nodes in the i stage in the t iteration processes;
Figure FDA0003114874410000034
representing i-phase j-factor graph nodes in hidden layers of deep neural networks during t iterations
Figure FDA0003114874410000035
A learnable offset of; n represents a polarization code length; g' (a, b, β) ═ sign (a) sign (b) fReLU(min (| a |, | b |) - β, 0); g' (a, b, β) represents an information iteration function; sign (a) represents a sign function value of the signal a; sign (b) represents a sign function value of the signal b; f. ofReLUAn activation function representing a deep neural network; β represents a learnable offset of offline learning; the parameters L and R are superscripts indicating the number of iterations, the left-hand subscript in the parameters L, R and β indicating the number of phases performed, and the right-hand subscript indicating the number of nodes.
7. The deep learning-based PC-SCMA joint iterative detection decoding method of claim 1, wherein the iteratively updating user node information in step S7 comprises:
Figure FDA0003114874410000036
wherein the content of the first and second substances,
Figure FDA0003114874410000037
representing by user node ujTo the resource node rkTransmitting the message;
Figure FDA0003114874410000038
representing user node ujObtaining prior information;
Figure FDA0003114874410000039
representing a resource node rkTo user node ujInformation of the transmitted first SCMA symbol; { zeta-meterjV k representation and user node ujConnected dividerkAll but resource node sets.
8. The PC-SCMA joint iterative detection decoding method based on deep learning of claim 1, wherein after step S8, the method further comprises using a loss function to evaluate performance loss during training, and then calculating an optimal set of weighting factors and offsets through a gradient descent algorithm and back propagation; and taking the optimal weight factor and the optimal offset as hidden layer parameters of the deep neural network to finish the updating of the resource node, the prior information and the user node information.
CN202110660122.3A 2021-06-15 2021-06-15 PC-SCMA joint iterative detection decoding method based on deep learning Active CN113395138B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110660122.3A CN113395138B (en) 2021-06-15 2021-06-15 PC-SCMA joint iterative detection decoding method based on deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110660122.3A CN113395138B (en) 2021-06-15 2021-06-15 PC-SCMA joint iterative detection decoding method based on deep learning

Publications (2)

Publication Number Publication Date
CN113395138A true CN113395138A (en) 2021-09-14
CN113395138B CN113395138B (en) 2022-05-03

Family

ID=77621058

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110660122.3A Active CN113395138B (en) 2021-06-15 2021-06-15 PC-SCMA joint iterative detection decoding method based on deep learning

Country Status (1)

Country Link
CN (1) CN113395138B (en)

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105787458A (en) * 2016-03-11 2016-07-20 重庆邮电大学 Infrared behavior identification method based on adaptive fusion of artificial design feature and depth learning feature
CN106941394A (en) * 2017-03-06 2017-07-11 东南大学 The SCMA of polarization code coding joint-detection decoding algorithm and device
CN107508657A (en) * 2017-08-08 2017-12-22 重庆邮电大学 A kind of SCMA multi-user test methods based on weight factor message transmission
CN110113057A (en) * 2019-06-12 2019-08-09 中国计量大学 A kind of polarization code decoder using deep learning
WO2020159154A1 (en) * 2018-02-08 2020-08-06 Samsung Electronics Co., Ltd. Method for encoding images and corresponding terminals

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105787458A (en) * 2016-03-11 2016-07-20 重庆邮电大学 Infrared behavior identification method based on adaptive fusion of artificial design feature and depth learning feature
CN106941394A (en) * 2017-03-06 2017-07-11 东南大学 The SCMA of polarization code coding joint-detection decoding algorithm and device
CN107508657A (en) * 2017-08-08 2017-12-22 重庆邮电大学 A kind of SCMA multi-user test methods based on weight factor message transmission
WO2020159154A1 (en) * 2018-02-08 2020-08-06 Samsung Electronics Co., Ltd. Method for encoding images and corresponding terminals
CN110113057A (en) * 2019-06-12 2019-08-09 中国计量大学 A kind of polarization code decoder using deep learning

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
MINHOE KIM: "Deep Learning-Aided SCMA", 《IEEE COMMUNICATIONS LETTERS》 *
戴金晟: "基于广义极化变换的多流信号传输理论与方案研究", 《中国优秀博士论文合辑》 *

Also Published As

Publication number Publication date
CN113395138B (en) 2022-05-03

Similar Documents

Publication Publication Date Title
CN106100794B (en) Coding cooperation method based on punched polarization code
CN107241106B (en) Deep learning-based polar code decoding algorithm
CN108847848B (en) BP decoding algorithm of polarization code based on information post-processing
CN108964669B (en) LDPC code quadratic programming decoding method based on degree decomposition and alternative multiplier method
CN115694518A (en) Convolutional code decoding method and system based on GMM-HMM
CN110971244A (en) Forward error correction decoding decoder based on burst error detection
KR100515472B1 (en) Channel coding and decoding method and multiple-antenna communication systems performing the same
Wu et al. Efficient VLSI implementation of soft-input soft-output fixed-complexity sphere decoder
CN110601699B (en) Method for realizing code rate dynamic variable multi-element LDPC code
CN107094026B (en) Graph merging detection decoding method for NB-LDPC coding
CN113395138B (en) PC-SCMA joint iterative detection decoding method based on deep learning
Huang et al. Weighted online fountain codes with limited buffer size and feedback transmissions
CN113992313B (en) Balanced network assisted SCMA encoding and decoding method based on deep learning
CN114666012B (en) Multi-sampling-point joint detection SCMA detection method based on MPA
Tan et al. Area-efficient pipelined vlsi architecture for polar decoder
Zhang et al. Adaptive iterative decoding for expediting the convergence of unary error correction codes
Wang et al. Pre-trained models based receiver design with natural redundancy for chinese characters
Leduc-Primeau et al. A relaxed half-stochastic iterative decoder for LDPC codes
Berber et al. Convolutional decoders based on artificial neural networks
Song et al. ANN based adaptive successive cancellation list decoder for polar codes
CN115426014B (en) Underwater sound MIMO communication method based on unitary space-time code modulation
TW201503625A (en) Equalization in the receiver of a multiple input multiple output system
Yan et al. Research on polar Decoding Method Based on Convolutional Neural Network
Jiang et al. Low-complexity adaptive reconciliation protocol for continuous-variable quantum key distribution
Chen et al. Sparse factor graph-based expanded iterative detection and decoding scheme for uplink polar coded SCMA system

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