A residual degradation LDPC hierarchical decoding working method based on AGDO optimization algorithm and dynamic quantization technology

Abstract

The application relates to a residual degradation LDPC hierarchical decoding working method based on an AGDO optimization algorithm and a dynamic quantization technology, and belongs to the technical field of wireless communication. In the framework of a residual degradation hierarchical decoding algorithm (RB_DLNMS), an AGDO intelligent optimization algorithm is combined to adaptively search for a quantization parameter, an optimized quantization parameter set suitable for hierarchical decoding is obtained, and a dynamic quantization structure of in-layer 8-bit intermediate confidence storage and updating and inter-layer normalization compression to 6-bit is introduced in the decoding process. On the premise of keeping a low quantization bit width and a low complexity, the dynamic quantization structure can effectively improve overflow and saturation phenomena in the hierarchical quantization decoding process and reduce the bit error rate of 6-bit quantization hierarchical decoding.

Description

Technical Field

[0001] This invention relates to a residual degradation LDPC hierarchical decoding method based on AGDO optimization algorithm and dynamic quantization technology, belonging to the field of wireless communication technology. Background Technology

[0002] In recent years, the demand for high-speed, low-latency, and highly reliable wireless communication has continued to grow, placing higher requirements on channel coding schemes in 5G systems. Against this backdrop, low-density parity-check codes (LDPC), with their error correction performance approaching the Shannon limit and good parallel decoding capabilities, have been adopted by the 3GPP standard as one of the data channel coding schemes for 5G enhanced mobile broadband (eMBB) scenarios.

[0003] LDPC code decoding methods can generally be divided into two main categories: hard decision decoding and soft decision decoding. Hard decision methods are simple to implement but have limited decoding performance; soft decision methods can make full use of the receiver's probability information and have better decoding performance, but have higher implementation complexity.

[0004] In practical engineering implementations, to meet requirements such as hardware resources, power consumption, and throughput, it is usually necessary to quantize the soft information in the BP or NMS algorithms. However, when traditional uniform quantization strategies are combined with the LNMS algorithm, they are prone to insufficient dynamic range and numerical overflow, leading to significant decoding errors. In the prior art, Chinese patent document CN120675574A discloses an improved Dragonfly algorithm combined with residual degradation in LDPC quantization layered decoding. This method uses the improved Dragonfly algorithm for quantization parameter search and combines it with residual degradation layered decoding to explore LDPC quantization layered decoding. However, its quantization parameter optimization mainly relies on the improved Dragonfly algorithm, and there is still room for further improvement in search efficiency, parameter update stability, and integration with dynamic quantization structures. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a residual degenerate LDPC layered decoding method based on the AGDO optimization algorithm and dynamic quantization technology. Within the framework of the residual degenerate layered decoding algorithm (RB_DLNMS), the AGDO intelligent optimization algorithm is used to adaptively search for quantization parameters, resulting in an optimized set of quantization parameters suitable for layered decoding. Furthermore, during the decoding process, an 8-bit intermediate confidence storage and update within each layer and a dynamic quantization structure that is normalized and compressed to 6 bits between layers are introduced. While maintaining a low quantization bit width and low complexity, this method effectively improves the overflow and saturation phenomena in the layered quantization decoding process, reducing the bit error rate of 6-bit quantization layered decoding.

[0006] This invention employs AGDO to search for quantization parameters, which can simultaneously utilize gradient momentum information and adaptive step size information during parameter update, thereby improving the stability and convergence efficiency of quantization parameter optimization. Furthermore, while maintaining the basic framework of residual degradation hierarchical decoding, this invention combines quantization parameter optimization with dynamic quantization structure, which is beneficial for further improving performance indicators such as bit error rate, frame error rate, decoding time, and average number of iterations, while saving certain computational resources and improving the feasibility of engineering applications.

[0007] The technical solution of the present invention is as follows: A residual-degraded LDPC hierarchical decoding method based on AGDO optimization algorithm and dynamic quantization technique is proposed, with the following steps: (1) Send training signals; (2) Estimate channel state information; (3) Determine the noise situation and calculate the soft information in the logarithmic domain; (4) Intelligent optimization training: The training signal is introduced into the AGDO algorithm to train the quantization parameter set of quantization methods Q1 and Q2 (Q1 and Q2 are referred to in document CN120675574A An improved dragonfly algorithm combined with residual degradation LDPC quantization layered decoding method). The AGDO algorithm relies on reading a large number of sample signals and decoding performance feedback to perform global search and adaptive adjustment of quantization parameters, and imports candidate quantization parameters into the RB_DLNMS decoding chain with dynamic quantization structure for performance evaluation to obtain the optimal quantization parameter set suitable for layered quantization decoding. When the signal-to-noise ratio is small, the quantization value density at both ends and the quantization interval density near zero end are adaptively increased to adapt to the signal inversion and low sensitivity of quantization values ​​caused by large noise. (5) Decoding: The optimal set of quantization parameters obtained from training is applied to the residual degradation decoding algorithm that introduces a dynamic quantization structure.

[0008] According to a preferred embodiment of the present invention, in step (1), specifically, a D-frame original information bit sequence is generated by the signal source, and the single-frame original information bit sequence is denoted as... M represents the number of effective information bits per frame. The original information bit sequence is input into the LDPC encoder for encoding to obtain the codewords for the single frame. Where N is the code length, NM is the number of parity bits, and the corresponding parity check matrix is ​​denoted as . Then, BPSK modulation is applied to the codeword x to obtain the transmitted signal. During the training phase, D-frame modulated signals are transmitted as input samples for subsequent channel state information estimation, noise power estimation, and offline optimization of quantization parameters.

[0009] According to a preferred embodiment of the present invention, in step (2), the received signal y is represented in the time domain as follows: Where h is the equivalent channel response and w is the noise signal; under AWGN channel conditions, h=1; under fading channel conditions, h is the equivalent response of the corresponding fading channel. Transforming the above relationship to the frequency domain, it is expressed as Y=GX+W, where G is the Fast Fourier Transform of the channel system response function, X is the Fast Fourier Transform of the transmitted signal, and W is the Fast Fourier Transform of the noise w. To perform a Fast Fourier Transform on the received signal y, statistical analysis of the D-frame training signal yields the average least squares estimate of the channel state information. , where i represents the i-th frame of data.

[0010] According to a preferred embodiment of the present invention, in step (3), the transmitted signal is reconstructed using the channel state information estimate obtained in step (2) to obtain... Thus, the residual is obtained. Based on the uniform distribution characteristics of the noise signal in the frequency domain, the noise power Estimated from the residual average power, i.e. Among them, E ij Let represent the residual value corresponding to the j-th bit of the i-th frame. Based on the noise power estimation result, the received data is converted into logarithmic domain soft information to obtain the log-likelihood ratio information of the j-th bit of the i-th frame. , where y ij L represents the received value corresponding to the j-th bit of the i-th frame. ij This serves as the input soft information for subsequent quantization parameter optimization modules to perform Q1 / Q2 quantization mapping and hierarchical LDPC decoding.

[0011] According to a preferred embodiment of the present invention, in step (4), the optimization process of the AGDO algorithm is as follows: Figure 2 The specific steps are as follows: (41) Optimize parameter initialization assignment, including the maximum number of decoding iterations T, the number of optimization iterations itmax, the population size num, the dimension of the quantization parameter vector Dim, and the upper and lower bound vectors of the parameters. and And the first-order moment decay coefficient β1, the second-order moment decay coefficient β2, and the learning rate of the AGDO algorithm. and constants to prevent the denominator from being zero ; Let the i-th candidate quantization parameter vector be denoted as ,in, Indicates the current optimization iteration number. The initial population is randomly generated within each dimension of the boundary, as shown below: (1) in, , The initial population matrix is ​​composed of the candidate quantization parameter vectors. Then calculate the fitness value corresponding to each candidate vector. Let the individual with the best fitness be the current dominant individual and the current best individual, respectively denoted as... and ,Right now: (2) (42) Perform logarithmic domain transformation on the training signal to obtain the logarithmic domain information of the j-th bit in the i-th frame. Then, the current candidate quantization parameter vector Substitute the input logarithmic field information into quantization methods Q1 or Q2 respectively to quantize the quantized data q1 and q2. (43) The quantized data q1 and q2 are imported into the residual degradation decoding algorithm of the dynamic quantization structure for decoding; (44) Calculate the bit error rate (BER) and frame error rate (FER), and use the decoding performance as the fitness feedback for the AGDO algorithm; among which, BER is given priority as the fitness comparison criterion, and FER is compared when BER is the same; under a given typical Eb / N0 operating point, the fitness function Take the BER value of the corresponding candidate quantization parameter vector after decoding in step (43). If there are multiple candidate quantization parameter vectors with the same BER, select the one with the smaller FER as the better individual. (45) Based on the fitness value of the current population, the AGDO algorithm is used to update and iterate the quantization parameter vector; (46) If the termination condition is met, output the optimal quantization parameter set; otherwise, return to step (42) to continue the next quantization parameter adjustment iteration until the optimal quantization parameter set is obtained. And use it as the input parameter set for the quantization layer decoding in the subsequent step (5).

[0012] According to a preferred embodiment of the present invention, in step (43), the residual degradation decoding algorithm that introduces a dynamic quantization structure refers to: quantized data q1 and q2 first enter the corresponding RB_DLNMS decoding algorithm. During the process of variable node information update and intra-layer accumulation, 8-bit precision is used to store and update the intermediate confidence. Here, 8 bits are only used as temporary storage units in the intra-layer update process and do not occupy too much memory. After each layer update is completed, the maximum absolute value of the intermediate confidence of all variable nodes in that layer is taken as the scaling factor. The 8-bit intermediate result of that layer is normalized and re-compressed and quantized to 6 bits before being passed to the next layer. Thus, without changing the residual degradation criterion structure, the effective dynamic range of the intermediate message is improved and the overflow and saturation problems are improved. For details of the specific decoding implementation, see step (5) below.

[0013] According to a preferred embodiment of the present invention, in step (45), the iterative step includes three parts: asymptotic gradient momentum integration, a dynamic gradient interaction system, and a system optimization operator, and its update process is as follows: First, we perform a progressive gradient momentum integration, where the momentum weight at the t-th iteration is . The random angle mapping coefficient is Then we have: (3) (4) This yields the first update position of the candidate quantization parameter vector during the exploration phase. for: (5) Where k represents the number of consecutive gradient momentum integration operations, used to adjust the search intensity according to the dimensionality of the quantization parameter; Subsequently, a dynamic gradient interaction system update is performed. For each candidate quantization parameter vector, two reference individuals are randomly selected from the current population, denoted as ... and Define direction pointer for: (6) in, This is used to characterize the search direction of the current candidate individual relative to the reference individual, and then a random perturbation vector is defined. for: (7) To enhance local development capabilities, first-order and second-order moment estimations are introduced. Let the gradient information at the current position be... Its first and second moments are denoted as... , Then we have: (8) (9) The forms of deviation correction are as follows: (10) Among them, the gradient information at the current position From the current best individual Current population mean That is, the first update position Determined jointly, and expressed as: (11) Based on this, key reference positions are constructed. : (12) This yields two candidate update positions for the dynamic gradient interaction system: (13) (14) in, This indicates element-wise multiplication in the corresponding dimension, where candidate positions are selected from the two positions above based on a random decision after dynamic gradient interaction. And use a trust domain strategy to determine whether to accept the location; when If the new position is accepted, the current position is retained; otherwise, the current position is retained. If the fitness of the new position is better than that of the current best individual, then the current position is updated. ; After completing the above development and updates, the system optimization operator will be updated. Let the Levy flight control parameters be... The random decay factor is The linear scaling factor is Then the candidate positions given by the system optimization operator are: (15) in, and They are respectively: (16) Levy flight function Represented as: (17) (18) Where r1 and r2 are random numbers between 0 and 1, For the Gamma function, For Levy flight shape parameters, if the system optimization operator triggering condition is met, then... As a new candidate quantization parameter vector, it participates in fitness comparison, thereby enhancing the population's ability to escape local optima.

[0014] According to a preferred embodiment of the present invention, step (5) is specifically implemented as follows: (51) Initialization: The received 6-bit quantized logarithmic field information is used as the initial input of the decoder and given to layer 0 for the first iteration update. The initial check message is set to 0. The initialization expression is: (19) (20) During initialization, the received logarithmic field information is first quantized to 6 bits and then sent to the layered decoder, and is only used in the first iteration. Subsequent iterations update the posterior information of each layer through inter-layer transmission. This represents a quantization operator that compresses and maps the input result to a 6-bit representable interval. and These represent the 6-bit extrinsic information and 6-bit posterior information during the initialization phase, respectively. Indicates the first 1 variable node Indicates the relationship with the first The first verification node associated with the first Posterior extrinsic information corresponding to each variable node. This indicates an 8-bit checksum update message during the initialization phase.

[0015] (52) Introduce 8-bit intermediate accumulation iterative message updates; In the k-th layer of the l-th iteration, the 6-bit a posteriori information passed from the previous layer is first expanded to an 8-bit representation, and then the corresponding verification message from the previous round is subtracted to obtain the 8-bit intermediate a posteriori information used by the current layer to verify node updates. Its expression is: (twenty one) in, This represents a bit-width expansion operator from 6 bits to 8 bits. This represents the 8-bit intermediate external information passed from the variable node to the verification node in the l-th iteration and k-th layer. This represents the 6-bit a posteriori information passed from the previous layer. This represents the 8-bit checksum update message on the corresponding edge in the previous round; Based on this, the check nodes are updated using the hierarchical normalized minimum sum algorithm, resulting in: (twenty two) in, This indicates that the node is connected to the j-th check node and is excluding the current variable node. The set of all variable nodes other than This represents the minimum sum algorithm correction factor used in the 8-bit intermediate message update stage. The updated check message is then added back to the current layer's external information to obtain the 8-bit intermediate posterior information for this layer. (twenty three) in, This represents the 8-bit intermediate posterior information after the k-th layer update in the l-th iteration; (53) Normalize and compress to 6 bits by layer. After completing the 8-bit intermediate posterior information update of the k-th layer in the l-th iteration, first calculate the maximum absolute value of the intermediate posterior information of all variable nodes in this layer: (twenty four) in, This represents the maximum absolute value of the 8-bit intermediate posterior information of all variable nodes in the k-th layer during the l-th iteration. Then, using this maximum absolute value as the scaling reference factor for this layer, let: (25) in, Indicates by A defined scaling factor; The inter-layer posterior information, after being normalized and compressed to 6 bits, is as follows: (26) This represents the 6-bit posterior information after layer-by-layer normalization and compression. That is, first divide the 8-bit intermediate accumulation result by the scaling factor, and then remap the result to the 6-bit representable range; (54) Reliability layer determination; (55) Residual degradation: When the change in inter-layer posterior information exceeds the residual threshold, or when the number of existing oscillating layers is less than or equal to the minimum number of oscillating layers, the original reliable layer is degraded into an unreliable layer, i.e., an oscillating layer. The criterion is as follows: (28) or (29) in, The number of oscillation layers currently present. To minimize the number of oscillating layers, The residual threshold; (56) Decoding decision: After all layers have been updated, a hard decision is made on the final a posteriori information; (30) When equation (30) is true, the decision is 1; otherwise, the decision is 0, and the decoding result is stored in the matrix. In, among them, This indicates the last layer at the end of the current complete hierarchical iteration; (57) Verification results.

[0016] According to a preferred embodiment of the present invention, in step (54), specifically: Determine the threshold based on the set reliability layer. Determine whether this layer belongs to a reliable layer; (27) When equation (27) is true, the layer is determined to be a reliable layer; otherwise, it is determined to be an unreliable layer. If a layer is determined to be a reliable layer, it is skipped in the next iteration.

[0017] According to a preferred embodiment of the present invention, in step (57), specifically: Substitute the decoding result into the parity check equation for verification. If the following equation is satisfied: (31) If the code is satisfied, the output is decoded; otherwise, the next iteration begins, continuing until all check equations are satisfied or the maximum number of decoding iterations is reached.

[0018] The beneficial effects of this invention are as follows: 1. This invention introduces the AGDO optimization algorithm into the LDPC quantization layered decoding parameter optimization process. Given a quantization bit width and signal-to-noise ratio, it can adaptively search for Q1 and Q2 quantization parameters to obtain a better set of quantization parameters. Simultaneously, within the residual degradation layered decoding framework, it introduces an 8-bit intermediate confidence storage and update structure within each layer, and a dynamic quantization structure that normalizes and compresses to 6 bits between layers. This allows the intermediate message to have a larger dynamic range during the layer update process, effectively improving overflow and saturation phenomena under low bit width conditions and enhancing numerical stability during quantization decoding. Furthermore, it still uses a 6-bit compressed quantization transmission method between layers, with the 8 bits only serving as temporary storage units during the layer update process, thus not occupying excessive memory. Therefore, it improves decoding performance while also considering hardware implementation complexity, which is beneficial for engineering implementation under resource-constrained conditions.

[0019] 2. This invention improves the accuracy of reliable layer identification and oscillation layer recovery without changing the basic criteria for reliable layer determination and residual degradation, thereby enhancing the effectiveness of the layer removal process. Therefore, this invention can improve the bit error rate while maintaining low complexity and fast decoding speed. Frame error rate Average decoding time per frame, layer removal rate and average number of iterations With these performance indicators, it has good engineering application value. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention.

[0021] Figure 2 This is the optimization process of the AGDO algorithm of the present invention.

[0022] Figure 3 The following are simulation results of the decoding performance of the method of this invention and other comparison algorithms, wherein, Figure 3 In the middle (a), the bit error rate (BER) is the BER of different algorithms. The graph shows the relationship between signal-to-noise ratio Eb / N0 (dB) and eleven algorithms, namely: Layered Normalized Minimum Sum Decoding Algorithm (LNMS), Layered Deletion Decoding Algorithm (DLNMS), Residual Degradation Layered Deletion Decoding Algorithm (RB_DLNMS), Q1 Residual Degradation Layered Deletion Decoding Algorithm Optimized by Improved Dragonfly Algorithm (IDA+Q1RB_DLNMS), Q1 Residual Degradation Layered Deletion Decoding Algorithm Optimized by Improved Harmony Search Algorithm Combining Hybrid Cuckoo Search Operator (HS_CS+Q1RB_DLNMS), Q1 Residual Degradation Layered Deletion Decoding Algorithm Optimized by AGDO (AGDO+Q1RB_DLNMS), and Algorithm with Dynamic Quantization Structure. The proposed algorithms are: AGDO-optimized Q1 residual degenerate deletion layered decoding algorithm (N8to6+AGDO+Q1RB_DLNMS), improved dragonfly algorithm-optimized Q2 residual degenerate deletion layered decoding algorithm (IDA+Q2RB_DLNMS), improved harmony search algorithm combined with hybrid cuckoo search operator-optimized Q2 residual degenerate deletion layered decoding algorithm (HS_CS+Q2RB_DLNMS), AGDO-optimized Q2 residual degenerate deletion layered decoding algorithm (AGDO+Q2RB_DLNMS), and AGDO-optimized Q2 residual degenerate deletion layered decoding algorithm with dynamic quantization structure (N8to6+AGDO+Q2RB_DLNMS). Figure 3 In the middle (b), the frame error rate is that of different algorithms. The relationship between signal-to-noise ratio and signal-to-noise ratio is shown in the graph. Figure 3 In the middle (c), the average correct decoding time per frame for different algorithms is represented. The relationship between signal-to-noise ratio and signal-to-noise ratio is shown in the graph. Figure 3 In the middle (d), the deletion stratification rate of different algorithms is ( ) Relationship between signal-to-noise ratio and other parameters.

[0023] Figure 4 The average number of iterations for different algorithms ( ) Relationship between signal-to-noise ratio and other parameters. Detailed Implementation

[0024] The present invention will be further described below with reference to the embodiments and accompanying drawings, but is not limited thereto.

[0025] Example 1: A residual-degraded LDPC hierarchical decoding method based on AGDO optimization algorithm and dynamic quantization technique is proposed, with the following steps: (1) Send a training signal, and the signal source generates a D-frame original information bit sequence. Let the single-frame original information bit sequence be... M represents the number of effective information bits per frame. The original information bit sequence is input into the LDPC encoder for encoding to obtain the codewords for the single frame. Where N is the code length, NM is the number of parity bits, and the corresponding parity check matrix is ​​denoted as . Then, BPSK modulation is applied to the codeword x to obtain the transmitted signal. During the training phase, D-frame modulated signals are transmitted as input samples for subsequent channel state information estimation, noise power estimation, and offline optimization of quantization parameters.

[0026] (2) Estimate the channel state information. The received signal y is represented in the time domain as follows: Where h is the equivalent channel response and w is the noise signal; under AWGN channel conditions, h=1; under fading channel conditions, h is the equivalent response of the corresponding fading channel. Transforming the above relationship to the frequency domain, it is expressed as Y=GX+W, where G is the Fast Fourier Transform of the channel system response function, X is the Fast Fourier Transform of the transmitted signal, and W is the Fast Fourier Transform of the noise w. To perform a Fast Fourier Transform on the received signal y, statistical analysis of the D-frame training signal yields the average least squares estimate of the channel state information. , where i represents the i-th frame of data.

[0027] (3) Determine the noise situation and calculate the logarithmic domain soft information. Use the channel state information estimate obtained in step (2) to reconstruct the transmitted signal to obtain... Thus, the residual is obtained. Based on the uniform distribution characteristics of the noise signal in the frequency domain, the noise power Estimated from the residual average power, i.e. Among them, E ij Let represent the residual value corresponding to the j-th bit of the i-th frame. Based on the noise power estimation result, the received data is converted into logarithmic domain soft information to obtain the log-likelihood ratio information of the j-th bit of the i-th frame. , where y ij L represents the received value corresponding to the j-th bit of the i-th frame. ij This serves as the input soft information for subsequent quantization parameter optimization modules to perform Q1 / Q2 quantization mapping and hierarchical LDPC decoding.

[0028] (4) Intelligent optimization training: The training signal is introduced into the AGDO algorithm to train the quantization parameter set of quantization methods Q1 and Q2 (Q1 and Q2 are referred to in document CN120675574A An improved dragonfly algorithm combined with residual degradation LDPC quantization layered decoding method). The AGDO algorithm relies on reading a large number of sample signals and decoding performance feedback to perform global search and adaptive adjustment of quantization parameters, and imports candidate quantization parameters into the RB_DLNMS decoding chain with dynamic quantization structure for performance evaluation to obtain the optimal quantization parameter set suitable for layered quantization decoding. When the signal-to-noise ratio is small, the quantization value density at both ends and the quantization interval density near zero end are adaptively increased to adapt to the signal inversion and low sensitivity of quantization values ​​caused by large noise. The optimization process for the AGDO algorithm is as follows: Figure 2 The specific steps are as follows: (41) Optimize parameter initialization assignment, including the maximum number of decoding iterations T, the number of optimization iterations itmax, the population size num, the dimension of the quantization parameter vector Dim, and the upper and lower bound vectors of the parameters. and And the first-order moment decay coefficient β1, the second-order moment decay coefficient β2, and the learning rate of the AGDO algorithm. and constants to prevent the denominator from being zero ; Let the i-th candidate quantization parameter vector be denoted as ,in, Indicates the current optimization iteration number. The initial population is randomly generated within each dimension of the boundary, as shown below: (1) in, , The initial population matrix is ​​composed of the candidate quantization parameter vectors. Then calculate the fitness value corresponding to each candidate vector. Let the individual with the best fitness be the current dominant individual and the current best individual, respectively denoted as... and ,Right now: (2) (42) Perform logarithmic domain transformation on the training signal to obtain the logarithmic domain information of the j-th bit in the i-th frame. Then, the current candidate quantization parameter vector Substitute the input logarithmic field information into quantization methods Q1 or Q2 respectively to quantize the quantized data q1 and q2. (43) The quantized data q1 and q2 are imported into the residual degradation decoding algorithm with dynamic quantization structure for decoding. The residual degradation decoding algorithm with dynamic quantization structure refers to the following: the quantized data q1 and q2 first enter the corresponding RB_DLNMS decoding algorithm. During the variable node information update and layer accumulation process, the intermediate confidence is stored and updated with 8-bit precision. The 8 bits are only used as temporary storage units in the layer update process and do not occupy too much memory. After each layer update, the maximum absolute value of the intermediate confidence of all variable nodes in the layer is taken as the scaling factor. The 8-bit intermediate result of the layer is normalized and re-compressed and quantized to 6 bits before being passed to the next layer. This improves the effective dynamic range of the intermediate message and improves the overflow and saturation problems without changing the residual degradation criterion structure. For the specific decoding implementation, see step (5) below.

[0029] (44) Calculate the bit error rate (BER) and frame error rate (FER), and use the decoding performance as the fitness feedback for the AGDO algorithm; among which, BER is given priority as the fitness comparison criterion, and FER is compared when BER is the same; under a given typical Eb / N0 operating point, the fitness function Take the BER value of the corresponding candidate quantization parameter vector after decoding in step (43). If there are multiple candidate quantization parameter vectors with the same BER, select the one with the smaller FER as the better individual. (45) Based on the fitness value of the current population, the quantization parameter vector is updated iteratively using the AGDO algorithm; the iterative steps include three parts: asymptotic gradient momentum integration, dynamic gradient interaction system, and system optimization operator, and the update process is as follows: First, we perform a progressive gradient momentum integration, where the momentum weight at the t-th iteration is . The random angle mapping coefficient is Then we have: (3) (4) This yields the first update position of the candidate quantization parameter vector during the exploration phase. for: (5) Where k represents the number of consecutive gradient momentum integration operations, used to adjust the search intensity according to the dimensionality of the quantization parameter; Subsequently, a dynamic gradient interaction system update is performed. For each candidate quantization parameter vector, two reference individuals are randomly selected from the current population, denoted as ... and Define direction pointer for: (6) in, This is used to characterize the search direction of the current candidate individual relative to the reference individual, and then a random perturbation vector is defined. for: (7) To enhance local development capabilities, first-order and second-order moment estimations are introduced. Let the gradient information at the current position be... Its first and second moments are denoted as... , Then we have: (8) (9) The forms of deviation correction are as follows: (10) Among them, the gradient information at the current position From the current best individual Current population mean That is, the first update position Determined jointly, and expressed as: (11) Based on this, key reference positions are constructed. : (12) This yields two candidate update positions for the dynamic gradient interaction system: (13) (14) in, This indicates element-wise multiplication in the corresponding dimension, where candidate positions are selected from the two positions above based on a random decision after dynamic gradient interaction. And use a trust domain strategy to determine whether to accept the location; when If the new position is accepted, the current position is retained; otherwise, the current position is retained. If the fitness of the new position is better than that of the current best individual, then the current position is updated. ; After completing the above development and updates, the system optimization operator will be updated. Let the Levy flight control parameters be... The random decay factor is The linear scaling factor is Then the candidate positions given by the system optimization operator are: (15) in, and They are respectively: (16) Levy flight function Represented as: (17) (18) Where r1 and r2 are random numbers between 0 and 1, For the Gamma function, For Levy flight shape parameters, if the system optimization operator triggering condition is met, then... As a new candidate quantization parameter vector, it participates in fitness comparison, thereby enhancing the population's ability to escape local optima.

[0030] (46) If the termination condition is met, output the optimal quantization parameter set; otherwise, return to step (42) to continue the next quantization parameter adjustment iteration until the optimal quantization parameter set is obtained. And use it as the input parameter set for the quantization layer decoding in the subsequent step (5).

[0031] (5) Decoding: The optimal set of quantization parameters obtained from training is applied to the residual degradation decoding algorithm that introduces a dynamic quantization structure.

[0032] The specific implementation is as follows: (51) Initialization: The received 6-bit quantized logarithmic field information is used as the initial input of the decoder and given to layer 0 for the first iteration update. The initial check message is set to 0. The initialization expression is: (19) (20) During initialization, the received logarithmic field information is first quantized to 6 bits and then sent to the layered decoder, and is only used in the first iteration. Subsequent iterations update the posterior information of each layer through inter-layer transmission. This represents a quantization operator that compresses and maps the input result to a 6-bit representable interval. and These represent the 6-bit extrinsic information and 6-bit posterior information during the initialization phase, respectively. Indicates the first 1 variable node Indicates the relationship with the first The first verification node associated with the first Posterior extrinsic information corresponding to each variable node. This indicates an 8-bit checksum update message during the initialization phase.

[0033] (52) Introduce 8-bit intermediate accumulation iterative message updates; In the k-th layer of the l-th iteration, the 6-bit a posteriori information passed from the previous layer is first expanded to an 8-bit representation, and then the corresponding verification message from the previous round is subtracted to obtain the 8-bit intermediate a posteriori information used by the current layer to verify node updates. Its expression is: (twenty one) in, This represents a bit-width expansion operator from 6 bits to 8 bits. This represents the 8-bit intermediate external information passed from the variable node to the verification node in the l-th iteration and k-th layer. This represents the 6-bit a posteriori information passed from the previous layer. This represents the 8-bit checksum update message on the corresponding edge in the previous round; Based on this, the check nodes are updated using the hierarchical normalized minimum sum algorithm, resulting in: (twenty two) in, This indicates that the node is connected to the j-th check node and is excluding the current variable node. The set of all variable nodes other than This represents the minimum sum algorithm correction factor used in the 8-bit intermediate message update stage. The updated check message is then added back to the current layer's external information to obtain the 8-bit intermediate posterior information for this layer. (twenty three) in, This represents the 8-bit intermediate posterior information after the k-th layer update in the l-th iteration; (53) Normalize and compress to 6 bits by layer. After completing the 8-bit intermediate posterior information update of the k-th layer in the l-th iteration, first calculate the maximum absolute value of the intermediate posterior information of all variable nodes in this layer: (twenty four) in, This represents the maximum absolute value of the 8-bit intermediate posterior information of all variable nodes in the k-th layer during the l-th iteration. Then, using this maximum absolute value as the scaling reference factor for this layer, let: (25) in, Indicates by A defined scaling factor; The inter-layer posterior information, after being normalized and compressed to 6 bits, is as follows: (26) This represents the 6-bit posterior information after layer-by-layer normalization and compression. That is, first divide the 8-bit intermediate accumulation result by the scaling factor, and then remap the result to the 6-bit representable range; (54) Reliability layer determination, specifically: Determine the threshold based on the set reliability layer. Determine whether this layer belongs to a reliable layer; (27) When equation (27) is true, the layer is determined to be a reliable layer; otherwise, it is determined to be an unreliable layer. If a layer is determined to be a reliable layer, it is skipped in the next iteration.

[0034] (55) Residual degradation: When the change in inter-layer posterior information exceeds the residual threshold, or when the number of existing oscillating layers is less than or equal to the minimum number of oscillating layers, the original reliable layer is degraded into an unreliable layer, i.e., an oscillating layer. The criterion is as follows: (28) or (29) in, The number of oscillation layers currently present. To minimize the number of oscillating layers, The residual threshold; (56) Decoding decision: After all layers have been updated, a hard decision is made on the final a posteriori information; (30) When equation (30) is true, the decision is 1; otherwise, the decision is 0, and the decoding result is stored in the matrix. In, among them, This indicates the last layer at the end of the current complete hierarchical iteration; (57) Verification results, specifically: Substitute the decoding result into the parity check equation for verification. If the following equation is satisfied: (31) If the code is satisfied, the output is decoded; otherwise, the next iteration begins, continuing until all check equations are satisfied or the maximum number of decoding iterations is reached.

[0035] use The optimal quantization parameter set obtained from AGDO was applied to the residual degradation decoding algorithm for simulation. The test parameters are as follows: Set the check matrix The size is That is, the length of a single frame transmitted signal is The effective signal length is The length of the verification signal is The bit rate is The number of rows in each layer is The modulation method is BPSK, the channel is white Gaussian noise, and the maximum number of frames is... Quantization bit width .

[0036] The specific steps are as follows: ① Set the simulation signal-to-noise ratio to Eb / N0

[0037] ② Generate 1,000,000 frames of random binary sequence As training signals, the quantization parameter set is trained offline using the methods described in steps (1) to (4) to obtain the optimal quantization parameter sets for the Q1 quantization method and the Q2 quantization method under the conditions of the improved Dragonfly Algorithm (IDA), the improved harmony search algorithm (HS_CS) combined with the hybrid cuckoo search operator, AGDO, and N8to6+AGDO.

[0038] ③ After obtaining the optimal quantization parameter set, a random binary sequence is regenerated as the information sequence to be transmitted, and it is LDPC encoded and BPSK modulated. After being transmitted through a Gaussian white noise channel, the logarithmic domain information is calculated by the receiver. Then, the quantized soft information is imported into Q1RB_DLNMS, Q2RB_DLNMS and residual degradation decoding algorithms with dynamic quantization structure for decoding. At the same time, the unquantized layered decoding algorithms LNMS, DLNMS and RB_DLNMS are set as comparison algorithms.

[0039] ④ During the quantization decoding process, the input information at the receiving end is first represented by 6-bit quantization; during the variable node information update and intra-layer accumulation process within the layered decoding, the intermediate confidence is stored and updated with 8-bit precision; after each layer update is completed, the maximum absolute value of the intermediate confidence of all variable nodes in that layer is taken as the scaling factor, the intermediate result of that layer is normalized, and then compressed back to 6 bits before being transmitted between layers.

[0040] ⑤ Statistical analysis of the bit error rate of each algorithm under different signal-to-noise ratios. Frame error rate Average correct decoding time per frame Deletion of layering rate and average number of iterations ,in , , The statistical results of different algorithms at each signal-to-noise ratio point are plotted as follows: Figure 3 and Figure 4 .

[0041] The minimum number of oscillation layers exists. for residual threshold for Reliability layer determination threshold for The correction factor for the minimum sum algorithm for Under these conditions, the decoding results are analyzed.

[0042] a, Figure 3 (a) and Figure 3 Figure (b) presents the BER and FER variation curves of 11 decoding algorithms under different signal-to-noise ratio conditions. The algorithms compared include the unquantized layered decoding algorithms LNMS, DLNMS, and RB_DLNMS, as well as the residual degradation layered decoding algorithms IDA, AGDO, HS_CS, and N8to6+AGDO based on Q1 and Q2 quantization methods. Figure 3 (a) and Figure 3 As shown in (b), the BER and FER of each algorithm gradually decrease with the increase of the signal-to-noise ratio. Unquantized LNMS has the best overall performance in terms of BER and FER, DLNMS has the weakest performance, and RB_DLNMS is in between. Among the quantized layered decoding algorithms, Q2-class algorithms are generally better than Q1-class algorithms. Among the Q2-class algorithms, the N8to6+AGDO+Q2RB_DLNMS algorithm, which is optimized with AGDO and introduces a dynamic quantization structure, exhibits lower BER and FER. Figure 3 The results show that, while maintaining low-bit-width quantization, optimizing the quantization parameters through AGDO and introducing a dynamic quantization structure within the decoding process can further improve the performance of quantization layer decoding.

[0043] b、 Figure 3 (c) and Figure 4 The average correct decoding time (TIME) and average number of iterations (ITE) for different algorithms under various signal-to-noise ratio (SNR) conditions are presented. As shown in the figures, the average correct decoding time and average number of iterations for each algorithm gradually decrease with increasing SNR. The unquantized LNMS and DLNMS algorithms generally have larger TIME values, while the quantized layered decoding algorithms show reductions in both TIME and ITE. Among these, the N8to6+AGDO layered decoding algorithm, which incorporates a dynamic quantization structure, further reduces the average correct decoding time and average number of iterations while maintaining superior decoding performance. Figure 3 (c) and Figure 4 This demonstrates that the quantization parameter optimization and dynamic quantization structure proposed in this invention can effectively improve decoding efficiency. Figure 3 As can be seen in (d), as the signal-to-noise ratio increases, the layer deletion rate of each layer deletion decoding algorithm gradually increases; among them, the N8to6+AGDO quantization layer decoding algorithm can achieve a high layer deletion rate while maintaining better decoding performance and a lower average number of iterations, indicating that the method achieves a good balance between layer deletion efficiency and decoding performance.

[0044] c. Comprehensive Figure 3 , Figure 4It can be seen that, under a given quantization bit width, using AGDO to optimize the quantization parameters and introducing a dynamic quantization structure in the residual degradation layered decoding process can further improve the bit error rate and frame error rate performance of quantization layered decoding while maintaining low complexity, shorten the average correct decoding time per frame, reduce the average number of iterations, and obtain a high deletion layer rate. Therefore, the proposed method has good comprehensive performance and engineering application value.

Claims

1. A method for residual degradation LDPC hierarchical decoding based on AGDO optimization algorithm and dynamic quantization technology, characterized in that, The steps are as follows: (1) Send training signals; (2) Estimate channel state information; (3) Determine the noise situation and calculate the soft information in the logarithmic domain; (4) Intelligent optimization training: The training signal is introduced into the AGDO algorithm to train the quantization parameter set of quantization methods Q1 and Q2. The AGDO algorithm relies on reading a large number of sample signals and decoding performance feedback to perform global search and adaptive adjustment of quantization parameters, and imports candidate quantization parameters into the RB_DLNMS decoding chain with dynamic quantization structure for performance evaluation to obtain the optimal quantization parameter set suitable for hierarchical quantization decoding; when the signal-to-noise ratio is small, the quantization value density at both ends and the quantization interval density near zero end are adaptively increased. (5) Decoding: The optimal set of quantization parameters obtained from training is applied to the residual degradation decoding algorithm that introduces a dynamic quantization structure.

2. The residual degradation LDPC layered decoding method based on AGDO optimization algorithm and dynamic quantization technology as described in claim 1, characterized in that, In step (1), specifically, a D-frame original information bit sequence is generated by the signal source. Let the single-frame original information bit sequence be... M represents the number of effective information bits per frame. The original information bit sequence is input into the LDPC encoder for encoding to obtain the codewords for the single frame. Where N is the code length, NM is the number of parity bits, and the corresponding parity check matrix is ​​denoted as . Then, BPSK modulation is applied to the codeword x to obtain the transmission signal, and D-frame modulated signals are transmitted during the training phase.

3. The residual degradation LDPC layered decoding method based on AGDO optimization algorithm and dynamic quantization technology as described in claim 2, characterized in that, In step (2), the received signal y is represented in the time domain as follows: Where h is the equivalent channel response and w is the noise signal; under AWGN channel conditions, h=1; under fading channel conditions, h is the equivalent response of the corresponding fading channel. Transforming the above relationship to the frequency domain, it is expressed as Y=GX+W, where G is the Fast Fourier Transform of the channel system response function, X is the Fast Fourier Transform of the transmitted signal, and W is the Fast Fourier Transform of the noise w. To perform a Fast Fourier Transform on the received signal y, statistical analysis of the D-frame training signal yields the average least squares estimate of the channel state information. , where i represents the i-th frame of data.

4. The residual degradation LDPC layered decoding method based on AGDO optimization algorithm and dynamic quantization technology as described in claim 3, characterized in that, In step (3), the transmitted signal is reconstructed using the channel state information estimate to obtain... Thus, the residual is obtained. Based on the uniform distribution characteristics of the noise signal in the frequency domain, the noise power Estimated from the residual average power, i.e. Among them, E ij Let represent the residual value corresponding to the j-th bit of the i-th frame. Based on the noise power estimation result, the received data is converted into logarithmic domain soft information to obtain the log-likelihood ratio information of the j-th bit of the i-th frame. , where y ij This represents the received value corresponding to the j-th bit of the i-th frame.

5. The residual degradation LDPC layered decoding method based on AGDO optimization algorithm and dynamic quantization technology as described in claim 4, characterized in that, In step (4), the optimization process of the AGDO algorithm is as follows: (41) Optimize parameter initialization assignment, including the maximum number of decoding iterations T, the number of optimization iterations itmax, the population size num, the dimension of the quantization parameter vector Dim, and the upper and lower bound vectors of the parameters. and And the first-order moment decay coefficient β1, the second-order moment decay coefficient β2, and the learning rate of the AGDO algorithm. and constants to prevent the denominator from being zero ; Let the i-th candidate quantization parameter vector be denoted as ,in, Indicates the current optimization iteration number. The initial population is randomly generated within each dimension of the boundary, as shown below: (1); in, , The initial population matrix is ​​composed of the candidate quantization parameter vectors. Then calculate the fitness value corresponding to each candidate vector. Let the individual with the best fitness be the current dominant individual and the current best individual, respectively denoted as... and ,Right now: (2); (42) Perform logarithmic domain transformation on the training signal to obtain the logarithmic domain information of the j-th bit in the i-th frame. Then, the current candidate quantization parameter vector Substitute the input logarithmic field information into quantization methods Q1 or Q2 respectively to quantize the quantized data q1 and q2. (43) The quantized data q1 and q2 are imported into the residual degradation decoding algorithm of the dynamic quantization structure for decoding; (44) Calculate the bit error rate BER and frame error rate FER The decoding performance is used as the fitness feedback for the AGDO algorithm; among which, priority is given to... BER As a basis for fitness comparison, when BER If they are the same, then compare them. FER ; At a typical Eb / N0 operating point, the fitness function Take the corresponding candidate quantization parameter vector after decoding in step (43) BER If there are multiple candidate quantization parameter vectors corresponding to the value, BER If they are the same, then select. FER Smaller individuals are considered superior. (45) Based on the fitness value of the current population, the AGDO algorithm is used to update and iterate the quantization parameter vector; (46) If the termination condition is met, output the optimal quantization parameter set; otherwise, return to step (42) to continue the next quantization parameter adjustment iteration until the optimal quantization parameter set is obtained. And use it as the input parameter set for the quantization layer decoding in the subsequent step (5).

6. The residual degradation LDPC layered decoding method based on AGDO optimization algorithm and dynamic quantization technology as described in claim 5, characterized in that, In step (43), the residual degradation decoding algorithm that introduces a dynamic quantization structure refers to the following: quantized data q1 and q2 first enter the corresponding RB_DLNMS decoding algorithm. During the process of updating variable node information and accumulating within the layer, 8-bit precision is used to store and update the intermediate confidence. Here, 8 bits are only used as temporary storage units during the update process within the layer. After each layer update is completed, the maximum absolute value of the intermediate confidence of all variable nodes in that layer is taken as the scaling factor. The 8-bit intermediate result of that layer is normalized and then recompressed and quantized to 6 bits before being passed to the next layer. This improves the effective dynamic range of the intermediate message and mitigates the overflow and saturation problems without changing the residual degradation criterion structure.

7. The residual degradation LDPC layered decoding method based on AGDO optimization algorithm and dynamic quantization technology as described in claim 6, characterized in that, In step (45), the iterative steps include three parts: asymptotic gradient momentum integration, dynamic gradient interaction system, and system optimization operator. The update process is as follows: First, we perform a progressive gradient momentum integration, where the momentum weight at the t-th iteration is . The random angle mapping coefficient is Then we have: (3); （4）； This yields the first update position of the candidate quantization parameter vector during the exploration phase. for: （5）； Where k represents the number of consecutive gradient momentum integration operations, used to adjust the search intensity according to the dimensionality of the quantization parameter; Subsequently, a dynamic gradient interaction system update is performed. For each candidate quantization parameter vector, two reference individuals are randomly selected from the current population, denoted as ... and Define direction pointer for: (6); in, This is used to characterize the search direction of the current candidate individual relative to the reference individual, and then a random perturbation vector is defined. for: （7）； To enhance local development capabilities, first-order and second-order moment estimations are introduced. Let the gradient information at the current position be... Its first and second moments are denoted as... , Then we have: (8); （9）； The forms of deviation correction are as follows: （10）； Among them, the gradient information at the current position From the current best individual Current population mean That is, the first update position Determined jointly, and expressed as: （11）； Based on this, key reference positions are constructed. : （12）； This yields two candidate update positions for the dynamic gradient interaction system: （13）； （14）； in, This indicates element-wise multiplication in the corresponding dimension, where candidate positions are selected from the two positions above based on a random decision after dynamic gradient interaction. And use a trust domain strategy to determine whether to accept the location; when If the new position is accepted, the current position is retained; otherwise, the current position is retained. If the fitness of the new position is better than that of the current best individual, then the current position is updated. ; After completing the above development and updates, the system optimization operator will be updated. Let the Levy flight control parameters be... The random decay factor is The linear scaling factor is Then the candidate positions given by the system optimization operator are: (15); in, and They are respectively: （16）； Levy flight function Represented as: (17); （18）； Where r1 and r2 are random numbers between 0 and 1, For the Gamma function, For Levy flight shape parameters, if the system optimization operator triggering condition is met, then... As a new candidate quantization parameter vector, it participates in fitness comparison, thereby enhancing the population's ability to escape local optima.

8. The residual degradation LDPC layered decoding method based on AGDO optimization algorithm and dynamic quantization technology as described in claim 7, characterized in that, In step (5), the specific implementation is as follows: (51) Initialization: The received 6-bit quantized logarithmic field information is used as the initial input of the decoder and given to layer 0 for the first iteration update. The initial check message is set to 0. The initialization expression is: (19); （20）； During initialization, the received logarithmic field information is first quantized to 6 bits and then sent to the layered decoder, and is only used in the first iteration. Subsequent iterations update the posterior information of each layer through inter-layer transmission. This represents a quantization operator that compresses and maps the input result to a 6-bit representable interval. and These represent the 6-bit extrinsic information and 6-bit posterior information during the initialization phase, respectively. Indicates the first 1 variable node Indicates the relationship with the first The first verification node associated with the first Posterior extrinsic information corresponding to each variable node. This represents an 8-bit checksum update message during the initialization phase; (52) Introduce 8-bit intermediate accumulation iterative message updates; In the k-th layer of the l-th iteration, the 6-bit a posteriori information passed from the previous layer is first expanded to an 8-bit representation, and then the corresponding verification message from the previous round is subtracted to obtain the 8-bit intermediate a posteriori information used by the current layer to verify node updates. Its expression is: (twenty one); in, This represents a bit-width expansion operator from 6 bits to 8 bits. This represents the 8-bit intermediate external information passed from the variable node to the verification node in the l-th iteration and k-th layer. This represents the 6-bit a posteriori information passed from the previous layer. This represents the 8-bit checksum update message on the corresponding edge in the previous round; Based on this, the check nodes are updated using the hierarchical normalized minimum sum algorithm, resulting in: (twenty two); in, This indicates that the node is connected to the j-th check node and is excluding the current variable node. The set of all variable nodes other than This represents the minimum sum algorithm correction factor used in the 8-bit intermediate message update stage. The updated check message is then added back to the current layer's external information to obtain the 8-bit intermediate posterior information for this layer. (twenty three); in, This represents the 8-bit intermediate posterior information after the k-th layer update in the l-th iteration; (53) Normalize and compress to 6 bits by layer. After completing the 8-bit intermediate posterior information update of the k-th layer in the l-th iteration, first calculate the maximum absolute value of the intermediate posterior information of all variable nodes in this layer: (twenty four); in, This represents the maximum absolute value of the 8-bit intermediate posterior information of all variable nodes in the k-th layer during the l-th iteration. Then, using this maximum absolute value as the scaling reference factor for this layer, let: （25）； in, Indicates by A defined scaling factor; The inter-layer posterior information, after being normalized and compressed to 6 bits, is as follows: (26); This represents the 6-bit posterior information after layer-by-layer normalization and compression. That is, first divide the 8-bit intermediate accumulation result by the scaling factor, and then remap the result to the 6-bit representable range; (54) Reliability layer determination; (55) Residual degradation: When the change in inter-layer posterior information exceeds the residual threshold, or when the number of existing oscillating layers is less than or equal to the minimum number of oscillating layers, the original reliable layer is degraded into an unreliable layer, i.e., an oscillating layer. The criterion is as follows: (28); or （29）； in, The number of oscillation layers currently present. To minimize the number of oscillating layers, The residual threshold; (56) Decoding decision: After all layers have been updated, a hard decision is made on the final a posteriori information; (30); When equation (30) is true, the decision is 1; otherwise, the decision is 0, and the decoding result is stored in the matrix. In, among them, This indicates the last layer at the end of the current complete hierarchical iteration; (57) Verification results.

9. The residual degradation LDPC layered decoding method based on AGDO optimization algorithm and dynamic quantization technology as described in claim 8, characterized in that, In step (54), specifically: Determine the threshold based on the set reliability layer. Determine whether this layer belongs to a reliable layer; (27); When equation (27) is true, the layer is determined to be a reliable layer; otherwise, it is determined to be an unreliable layer. If a layer is determined to be a reliable layer, it is skipped in the next iteration.

10. The residual degradation LDPC layered decoding method based on AGDO optimization algorithm and dynamic quantization technology as described in claim 9, characterized in that, In step (57), specifically: Substitute the decoding result into the parity check equation for verification. If the following equation is satisfied: (31); If the code is satisfied, the output is decoded; otherwise, the next iteration begins, continuing until all check equations are satisfied or the maximum number of decoding iterations is reached.

Images