An efficient decoding method for large-scale random access without passive address

Abstract

The present invention discloses an efficient decoding method for large-scale non-source random access, which relates to the field of wireless communications. In a wireless network, a multi-antenna base station simultaneously serves a large number of single-antenna mobile devices to access the network. In any given time slot, only a small number of devices are activated, and the other devices are in a dormant state. Based on a non-source authorization-free random access protocol, all activated devices use the same codebook known to the base station to send segmented and mapped codewords to the uplink. The base station uses a Bayesian optimal orthogonal approximate message passing method combined with a maximum expectation algorithm to perform codeword decoding and environmental parameter estimation, and uses a Bayesian classification algorithm to splice the decoded message fragments to restore the original information. The present invention provides an efficient large-scale random access method for wireless networks.

Description

Technical Field

[0001] The present invention relates to the field of wireless communications, and in particular to an efficient decoding method for large-scale random access without a passive address. Background Art

[0002] In recent years, with the advent of the Internet of Everything (IoE) era, the number of mobile devices and the amount of data they generate has exploded. In this context, large-scale machine-to-machine communication (M2M) is considered one of the key application scenarios for next-generation wireless networks. This means that wireless networks need to support the simultaneous access of large numbers of wireless terminals. Furthermore, because only a small number of terminal devices are active within a given time interval, while others are temporarily dormant to conserve energy, data requests from terminal devices in the network are often sporadic.

[0003] These characteristics mean that achieving the Internet of Everything still faces several problems and challenges. On the one hand, faced with the massive number of devices and sporadic data traffic, traditional authorization-based random access schemes result in excessive access latency and signaling overhead. On the other hand, in current active address random access schemes, activated devices must transmit their own unique pilot sequence to the base station at the beginning of each time slot. The base station then uses device activation detection and channel estimation algorithms to determine the device's activation status and obtain the corresponding channel information. To obtain accurate activation status and channel information, very long pilot sequences must be sent, which in turn wastes limited spectrum resources and increases computational complexity.

[0004] Therefore, to address these issues, it is necessary to design an efficient random access method based on the characteristics of the new generation of wireless networks. This paper studies large-scale passive address random access technology. All devices use the same codebook to send data information to the base station. The base station only pays attention to the information sent by the device itself, not the identity information of the activated device. Based on this, the activated device does not need to send a long pilot sequence in advance to detect the status of the device and channel. This can significantly reduce service latency, reduce the consumption of wireless resources, and improve data processing efficiency in the era of the Internet of Things. It is expected to break through the bottleneck problem of large-scale machine communication in wireless networks and solve a series of problems in the access of massive mobile devices to wireless networks. Summary of the Invention

[0005] In order to solve the problems of large connection capacity, high communication delay, low spectrum efficiency and high computational complexity when large-scale mobile devices access wireless networks in the prior art, the present invention proposes an efficient decoding method for large-scale random access without passive address.

[0006] The specific technical solutions adopted in the present invention are as follows:

[0007] The present invention provides an efficient decoding method for large-scale non-source random access, comprising the following steps:

[0008] S1: In a wireless network, a base station with M antennas is pre-deployed, and there are K t potential single-antenna mobile devices access the wireless network through the base station; in any given time slot, only K a Single-antenna mobile devices are active, of which K a <<K t , and remember the activation device set as All activated devices cut the b-bit information to be sent into segments, each segment is J bits long, and is divided into L = b / J sub-blocks in total;

[0009] S2: All active devices map L sub-blocks to codewords in the same codebook known to the base station, and send them to the base station via L sub-time slots.

[0010] S3: In each sub-slot, after receiving the data, the base station uses a Bayesian optimal orthogonal approximate message passing method to estimate the state matrix of the transmitted codeword. At the same time, the base station uses a maximum expectation method to estimate the channel parameters.

[0011] S4: The base station determines the codeword state matrix estimate in step S3 according to a hard threshold method to obtain a list of activated codewords to be sent;

[0012] S5: After obtaining the activation codeword list of each sub-time slot, the base station uses a Bayesian classification method to splice together the codewords from the same activated device in each sub-time slot and demap them to restore the original information sent.

[0013] Preferably, in step S2, a method in which all activated devices map sub-blocks to codewords based on the same codebook is:

[0014] S21: Given a sub-slot length of n0, set the codebook matrix in Represents the complex field, each column c of the codebook matrix j All represent a code word, where j∈[1,2 J ], a total of 2 J codewords; each codeword satisfies the constraint Where ||·||2 represents the vector's two-norm, P t is the maximum transmit power of each codeword;

[0015] S22: In the lth sub-time slot, any kth activated device maps the J bits of information of the lth sub-block to be sent to an integer i k,l ∈[1,2 J ]; where l∈[1,L],

[0016] S23: Activate device k to set the i-th k,l The column is sent to the base station as the codeword sent in the lth sub-time slot.

[0017] Preferably, in step S3, the method for estimating the state matrix of the transmitted codeword and estimating the channel parameters is:

[0018] S31: Input codebook matrix C, maximum value of iteration number t T iter and the received data at the base station end in the first sub-time slot in is the channel vector of the kth single-antenna mobile device, is the large-scale fading coefficient, Obeys a complex Gaussian distribution with zero mean and unit variance I represents the channel matrix; is a codeword activation matrix with values ​​of 0 or 1, and satisfies the i-th k,l The element Δ in the row and column k l (i l,k ,k)=1, where The remaining elements are 0; is additive white Gaussian noise, σ 2 is the noise variance; since the detection and estimation methods of each sub-time slot are the same, the subscript l is ignored in the following description; let is the row sparse codeword state matrix to be detected, whose jth row x j Obeying Bernoulli Gaussian distribution where j∈[1,2 J ],ε j is a non-zero probability, δ0 is the zero-point Dirac function, is the codeword variance, and when j=i k Time θ k,j =1, otherwise θ k,j =0;

[0019] S32: Set up three modules: linear estimator γ(·), nonlinear estimator φ(·) and maximum expectation updater ψ(·); let the t-th iteration input mean of the linear estimator be where x m is the mth column of the state matrix X, v t is the t-th iteration input variance, is a Gaussian random vector; the number of initialization iterations is t = 1, and the estimated value v 0 =1, noise variance where ||·|| F Denote the F norm of the matrix and set the SNR 0=100; initialize the row non-zero probability of the codeword state matrix and codeword variance Where cdf(·) and pdf(·) are the cumulative distribution function and probability density function of the standard normal distribution, respectively. H Y] j,: Represents the matrix C H The jth row of Y;

[0020] S33: Execute linear estimator γ(·): for m=1,...,M, calculate intermediate estimates where y m is the mth column of Y, the linear minimum mean square error estimate calculate As the output mean of the orthogonalized linear estimator γ(·), calculate As the orthogonalized output variance, where tr(·) represents the trace of the matrix;

[0021] S34: Execute nonlinear estimator φ(·): for j=1,...,2 J , let the t-th iteration input mean of the nonlinear estimator be in is a matrix The jth row of is the input variance of the tth iteration; calculate the minimum mean square error estimate of the intermediate quantity and the estimated variance in is the posterior row non-zero probability, and are the posterior Gaussian mean and variance respectively; calculate As the output variance of the orthogonalized nonlinear estimator φ(·), we calculate As the orthogonalized output mean, is a matrix The jth row of

[0022] S35: Execute the maximum expectation updater ψ(·): Update the noise variance to in is a matrix The mth column of ; the non-zero probability of the row of the updated codeword state matrix is Update the codeword variance to

[0023] S36: Determine accuracy requirements Is it satisfied, Is a small positive number; if not satisfied, assign At the iteration number t=t+1 and t≤T iterRepeat steps S33-S36 if satisfied; if satisfied, output codeword state matrix estimation and channel fading coefficient estimation

[0024] Preferably, in step S4, the hard threshold decision method for obtaining the activation codeword list is:

[0025] S41: Estimate value based on codeword state matrix Utilization of judgment criteria Get the codeword activation list for the current sub-time slot in is the decision threshold, yes The jth row of

[0026] S42: Repeat steps S3-S41 for all sub-time slots to obtain a list of L codeword activations The lth codeword activation list contains K a Unordered codeword numbers Also comes with K a The channel vector estimates of the activations and K a The corresponding large-scale fading coefficient estimate

[0027] Preferably, in step S5, the codeword splicing method using Bayesian classification is:

[0028] S51: Activate the first codeword list The code words in each class form K a Class Define K a The class labels of the classes are Initialize l=2;

[0029] S52: Activate the list of the first codeword In the kth activation channel vector where k∈[1,K a ], calculate the posterior probability of belonging to the k'th class respectively where k'∈[1,K a ], and find the class that maximizes this probability The judgment and the corresponding codeword numbers belong to the class in is a class The prior probability of yes Belong to class The conditional probability of is calculated as

[0030] S53: Current codeword activation list Classification is completed, update the class label The kth activation channel vector has been judged to belong to the class Let the codeword activation list index l=l+1, and repeat steps S52-S53 until all lists are classified;

[0031] S54: concatenate L activation codeword serial numbers belonging to the same class in chronological order.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] The efficient decoding method for large-scale, passive random access proposed in this paper addresses a series of issues arising from the large number of mobile devices accessing wireless networks, such as high connection capacity, high communication latency, and low spectrum utilization. The codeword detection method based on Bayesian optimal orthogonal approximation message passing, the parameter estimation method based on expectation maximization, and the codeword splicing method based on Bayesian classification proposed in this paper offer advantages such as low access latency and high spectrum efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a system diagram of an efficient decoding method for large-scale passive random access;

[0035] Figure 2 The detection accuracy of the codeword state matrix under different detection methods, that is, the comparison of the normalized mean square error (RMSE) (the methods are respectively the ideal case assuming known channel parameters, the maximum expectation-orthogonal approximate message passing method in the present invention, and the maximum expectation-classical approximate message passing method in the comparative example);

[0036] Figure 3 The error probability performance of the decoding method proposed in the present invention is compared when the number of base station antennas is different and the device transmission power is different (the number of antennas is 8, 32 and 64 respectively). DETAILED DESCRIPTION

[0037] The present invention will be further described and illustrated below with reference to the accompanying drawings and specific embodiments. The technical features of each embodiment of the present invention may be combined accordingly, provided that there is no conflict between them.

[0038] In this embodiment, the system block diagram of the efficient decoding method for large-scale non-source random access is as follows: Figure 1 As shown, the base station has M antennas, and in any given time slot, K aEach active mobile device with a single antenna accesses the wireless network through this base station. Based on a non-passive, unauthorized random access protocol, all active devices fragment their messages and then use the same codebook known to the base station to map these fragmented messages into codewords. These codewords are then sent uplink to the base station in sub-timeslot segments. After receiving the codeword information superimposed over the wireless channel, the base station uses a Bayesian optimal orthogonal approximate message passing method combined with the maximum expectation algorithm for codeword decoding and environmental parameter estimation. A Bayesian classification algorithm is then used to piece together the decoded information fragments to recover the original message from the active device.

[0039] The specific technical solutions adopted in this embodiment are as follows:

[0040] An efficient decoding method for large-scale random access without a passive address comprises the following steps:

[0041] S1: In a wireless network, a base station with M antennas is pre-deployed, and there are K t potential single-antenna mobile devices accessing the wireless network through the base station; in any given time slot, there are only K a Single-antenna mobile devices are active, of which K a K t , and remember the activation device set as All activated devices cut the b-bit information to be sent into segments, each segment is J bits long, and is divided into L=b / J sub-blocks in total.

[0042] S2: All activated devices map L sub-blocks to codewords in the codebook based on the same codebook known to the base station, and send the codewords to the base station via L sub-time slots.

[0043] In this step, all activated devices map sub-blocks to codewords based on the same codebook as follows:

[0044] S21: Given a sub-slot length of n0, set the codebook matrix in Represents the complex field, each column c of the codebook matrix j All represent a code word, where j∈[1,2 J ], a total of 2 J codewords. Each codeword satisfies the constraint Where ·2 represents the vector's two-norm, P t is the maximum transmit power of each codeword;

[0045] S22: In the lth sub-time slot, any kth activated device maps the J bits of information of the lth sub-block to be sent to an integer i k,l ∈[1,2 J ]; where l∈[1,L],

[0046] S23: Activate device k to set the i-th k,l The column is sent to the base station as the codeword sent in the lth sub-time slot.

[0047] S3: In each sub-time slot, after receiving the data, the base station uses a Bayesian optimal orthogonal approximate message passing method to estimate the state matrix of the transmitted codeword; at the same time, the base station uses a maximum expectation method to estimate the channel parameters.

[0048] In this step, the method for estimating the state matrix of the transmitted codeword and the channel parameters is:

[0049] S31: Input codebook matrix C, maximum value of iteration number t T iter and the received data at the base station end in the first sub-time slot in is the channel vector of the kth mobile device, is the large-scale fading coefficient, Obeys a complex Gaussian distribution with zero mean and unit variance I represents the channel matrix; is a codeword activation matrix with values ​​of 0 or 1, and satisfies the i-th k,l The element Δ in the row and column k l (i l,k ,k)=1, where The remaining elements are 0; is additive white Gaussian noise, σ 2 is the noise variance; since the detection and estimation methods of each sub-time slot are the same, the subscript l is ignored in the following description; let is the row sparse codeword state matrix to be detected, its jth row x j Obeying Bernoulli Gaussian distribution where j∈[1,2 J ],ε j is a non-zero probability, δ0 is the zero-point Dirac function, is the codeword variance, and when j=i k Time θ k,j =1, otherwise θ k,j =0;

[0050] S32: Set up three modules: linear estimator γ(·), nonlinear estimator φ(·) and maximum expectation updater ψ(·); let the t-th iteration input mean of the linear estimator be where x m is the mth column of the state matrix X, v t is the t-th iteration input variance, is a Gaussian random vector; the number of initialization iterations is t = 1, and the estimated value v 0 =1, noise variance in· F Denote the F norm of the matrix and set the SNR 0 =100; initialize the row non-zero probability of the codeword state matrix and codeword variance Where cdf(·) and pdf(·) are the cumulative distribution function and probability density function of the standard normal distribution, respectively. H Y] j,: Represents the matrix C H The jth row of Y;

[0051] S33: Execute linear estimator γ(·): for m=1,...,M, calculate intermediate estimates where y m is the mth column of Y, the linear minimum mean square error estimate calculate As the output mean of the orthogonalized linear estimator γ(·), calculate As the orthogonalized output variance, where tr(·) represents the trace of the matrix;

[0052] S34: Execute nonlinear estimator φ(·): for j=1,...,2 J , let the t-th iteration input mean of the nonlinear estimator be in is a matrix The jth row of is the input variance of the tth iteration; calculate the minimum mean square error estimate of the intermediate quantity and the estimated variance in is the posterior row non-zero probability, and are the posterior Gaussian mean and variance respectively; calculate As the output variance of the orthogonalized nonlinear estimator φ(·), we calculate As the orthogonalized output mean, is a matrix The jth row of

[0053] S35: Execute the maximum expectation updater ψ(·): Update the noise variance to in is a matrix The mth column of ; the non-zero probability of the row of the updated codeword state matrix is Update the codeword variance to

[0054] S36: Determine accuracy requirements Is it satisfied, Is a small positive number; if not satisfied, assign At the iteration number t=t+1 and t≤T iter Repeat steps S33-S36 if satisfied; if satisfied, output codeword state matrix estimation and channel fading coefficient estimation

[0055] S4: Based on a hard threshold method, the base station determines the codeword state matrix estimation value in step S3 to obtain a list of activated codewords to be sent.

[0056] In this step, the hard threshold decision method for obtaining the activation codeword list is:

[0057] S41: Estimate value based on codeword state matrix Utilization of judgment criteria Get the codeword activation list for the current sub-time slot in is the decision threshold, yes The jth row of

[0058] S42: Repeat steps S3-S41 for all sub-time slots to obtain a list of L codeword activations The lth codeword activation list contains K a Unordered codeword numbers Also comes with K a The channel vector estimates of the activations and K a The corresponding large-scale fading coefficient estimate

[0059] S5: After obtaining the activation codeword list of each sub-time slot, the base station uses a Bayesian classification method to splice together the codewords from the same activated device in each sub-time slot and demap them to restore the original information sent.

[0060] In this step, the codeword splicing method using Bayesian classification is:

[0061] S51: Activate the first codeword list The code words in each class form K a Class Define K a The class labels of the classes are Initialize l=2;

[0062] S52: Activate the list of the first codeword In the kth activation channel vector where k∈[1,K a ], calculate the posterior probability of belonging to the k'th class respectively where k'∈[1,K a ], and find the class that maximizes this probability The judgment and the corresponding codeword numbers belong to the class in is a class The prior probability of yes Belong to class The conditional probability of is calculated as

[0063] S53: Current codeword activation list Classification is completed, update the class label The kth activation channel vector has been judged to belong to the class Let the codeword activation list index l=l+1, and repeat steps S52-S53 until all lists are classified;

[0064] S54: concatenate L activation codeword serial numbers belonging to the same class in chronological order.

[0065] In order to further verify the effect of the method of the present invention, the normalized mean square error of the maximum expectation-orthogonal approximate message passing method of the present invention is compared with the ideal case where the channel parameters are assumed to be known and the maximum expectation-classical approximate message passing method, as shown below:

[0066] The ideal case where the channel parameters are known refers to the non-zero probability ε mentioned in step S31 of the present invention. j , codeword variance g j and noise variance σ 2 The base stations are all known, and there is no need to perform the initialization of the three in step S32 and the maximum expectation updater described in step S35. The remaining steps are consistent with the present invention. The maximum expectation-classical approximate message passing method (as a comparative example) refers to the present invention. Steps S33-S34 are changed to: Calculate the intermediate quantity matched filter estimate R t =C H D t-1 +S t-1 ,in D t-1 is the residual matrix, and in step S32, its initialization setting D is supplemented 0 =Y; Calculate the estimated variance u based on the residual matrix t =tr(D t-1 (D t-1 ) H) / (n0M) as the output variance; calculate the minimum mean square error estimate As the output mean, diag(·) indicates that the vector in the brackets is a matrix consisting of diagonal elements, The calculation method of is the same as that described in step S34 of the present invention; update the residual matrix Wherein mean(·) represents the average. Meanwhile, in step S35 of the present invention, the update of the noise variance is deleted. The remaining steps are consistent with the present invention.

[0067] Computer simulations show that Figure 2 As shown, in the efficient decoding method for large-scale non-passive random access proposed by the present invention, the normalized mean square error (NMSE) of the codeword state matrix estimate decreases with increasing antenna numbers and approaches saturation when the number of antennas is sufficiently large. Furthermore, the maximum expectation-orthogonal approximate message passing method proposed by the present invention outperforms the comparative scheme, namely, the detection performance of the maximum expectation-classical approximate message passing algorithm, which is closer to the ideal case where the environmental parameters are assumed to be known. Figure 3 The results show that the proposed decoding method reduces the probability of error in recovering the original information when the device transmit power increases, and significantly improves decoding performance as the number of antennas increases. Therefore, the present invention provides an efficient access method for wireless networks with large-scale devices.

[0068] The embodiment described above is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Persons skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, any technical solution obtained by equivalent substitution or equivalent transformation falls within the scope of protection of the present invention.

Claims

1. An efficient decoding method for large-scale random access without source address, characterized by The steps include: S1: In a wireless network, a base station with M antennas is pre-deployed, and there are K t potential single-antenna mobile devices access the wireless network through the base station; in any given time slot, only K a Single-antenna mobile devices are active, of which K a <<K t , and remember the activation device set as All activated devices cut the b-bit information to be sent into segments, each segment is J bits long, and is divided into L = b / J sub-blocks in total; S2: All active devices map L sub-blocks to codewords in the same codebook known to the base station, and send them to the base station via L sub-time slots. S3: In each sub-slot, after receiving the data, the base station uses a Bayesian optimal orthogonal approximate message passing method to estimate the state matrix of the transmitted codeword. At the same time, the base station uses a maximum expectation method to estimate the channel parameters. S4: The base station determines the codeword state matrix estimate in step S3 according to a hard threshold method to obtain a list of activated codewords to be sent; S5: After obtaining the activation codeword list of each sub-time slot, the base station uses a Bayesian classification method to splice together the codewords from the same activated device in each sub-time slot and demap them to restore the original information sent.

2. The efficient decoding method for large-scale non-source random access according to claim 1, characterized in that: In step S2, all activated devices map sub-blocks to codewords based on the same codebook as follows: S21: Given a sub-slot length of n0, set the codebook matrix in Represents the complex field, each column c of the codebook matrix j All represent a code word, where j∈[1,2 J ], a total of 2 J codewords; each codeword satisfies the constraint Where ||·||2 represents the vector's second norm, P t is the maximum transmit power of each codeword; S22: In the lth sub-time slot, any kth activated device maps the J bits of information of the lth sub-block to be sent to an integer i k,l ∈[1,2 J ]; where l∈[1,L], S23: Activate device k to set the i-th k,l The column is sent to the base station as the codeword sent in the lth sub-time slot.

3. The efficient decoding method for large-scale non-passive random access according to claim 1, characterized in that: In step S3, the method for estimating the state matrix of the transmitted codeword and estimating the channel parameters is: S31: Input codebook matrix C, maximum value of iteration number t T iter and the received data of the base station end in the lth sub-time slot in is the channel vector of the kth single-antenna mobile device, is the large-scale fading coefficient, Obeys a complex Gaussian distribution with zero mean and unit variance I represents the channel matrix; is a codeword activation matrix with values ​​of 0 or 1, and satisfies the i-th k,l The element Δ in the row and column k l (i l,k ,k)=1, where The remaining elements are 0; is additive white Gaussian noise, σ 2 is the noise variance; since the detection and estimation methods of each sub-time slot are the same, the subscript l is ignored in the following description; let is the row sparse codeword state matrix to be detected, whose jth row x j Obeying Bernoulli Gaussian distribution where j∈[1,2 J ],ε j is a non-zero probability, δ0 is the zero-point Dirac function, is the codeword variance, and when j=i k Time θ k,j =1, otherwise θ k,j =0; S32: Set up three modules: linear estimator γ(·), nonlinear estimator φ(·) and maximum expectation updater ψ(·); let the t-th iteration input mean of the linear estimator be where x m is the mth column of the state matrix X, v t is the t-th iteration input variance, is a Gaussian random vector; the number of initialization iterations is t = 1, and the estimated value v 0 =1, noise variance where ||·|| F Denote the F norm of the matrix and set the SNR 0 =100; initialize the row non-zero probability of the codeword state matrix and codeword variance Where cdf(·) and pdf(·) are the cumulative distribution function and probability density function of the standard normal distribution, respectively. H Y] j,: Represents the matrix C H The jth row of Y; S33: Execute linear estimator γ(·): for m=1,...,M, calculate intermediate estimates where y m is the mth column of Y, the linear minimum mean square error estimate calculate As the output mean of the orthogonalized linear estimator γ(·), calculate As the orthogonalized output variance, where tr(·) represents the trace of the matrix; S34: Execute nonlinear estimator φ(·): for j=1,...,2 J , let the t-th iteration input mean of the nonlinear estimator be in is a matrix The jth row of is the input variance of the tth iteration; calculate the minimum mean square error estimate of the intermediate quantity and the estimated variance in is the posterior row non-zero probability, and are the posterior Gaussian mean and variance respectively; calculate As the output variance of the orthogonalized nonlinear estimator φ(·), we calculate As the orthogonalized output mean, is a matrix The jth row of S35: Execute the maximum expectation updater ψ(·): Update the noise variance to in is a matrix The mth column of ; the non-zero probability of the row of the updated codeword state matrix is Update the codeword variance to S36: Determine accuracy requirements Is it satisfied, Is a small positive number; if not satisfied, assign At the iteration number t=t+1 and t≤T iter Repeat steps S33-S36 if satisfied; if satisfied, output codeword state matrix estimation and channel fading coefficient estimation 4. The efficient decoding method for large-scale non-source random access according to claim 1, characterized in that: In step S4, the hard threshold decision method for obtaining the activation codeword list is: S41: Estimate value based on codeword state matrix Utilization of judgment criteria Get the codeword activation list for the current sub-time slot in is the decision threshold, yes The jth row of S42: Repeat steps S3-S41 for all sub-time slots to obtain a list of L codeword activations The lth codeword activation list contains K a Unordered codeword numbers Also comes with K a The channel vector estimates of the activations and K a The corresponding large-scale fading coefficient estimate 5. The efficient decoding method for large-scale non-source random access according to claim 1, characterized in that: In step S5, the codeword splicing method using Bayesian classification is: S51: Activate the first codeword list The code words in each class form K a Class Define K a The class labels of the classes are Initialize l=2; S52: Activate the list of the first codeword In the kth activation channel vector where k∈[1,K a ], calculate the posterior probability of belonging to the k'th class respectively where k'∈[1,K a ], and find the class that maximizes this probability The judgment and the corresponding codeword numbers belong to the class in is a class The prior probability of yes Belong to class The conditional probability of is calculated as S53: Current codeword activation list Classification is completed, update the class label The kth activation channel vector has been judged to belong to the class Let the codeword activation list index l=l+1, and repeat steps S52-S53 until all lists are classified; S54: concatenate the L activation codeword serial numbers belonging to the same class in chronological order.

Images