Self-adaptive Turbo code decoding method and system based on neural network assistance

Abstract

The invention relates to the technical field of communication coding and decoding, in particular to an adaptive Turbo code decoding method and system based on neural network assistance, and the method comprises the following steps: S1, receiving a signal, inputting the signal into a CNN channel estimation network after time-frequency transformation, generating a dynamic segmentation parameter and a polynomial coefficient vector, and adaptively adjusting an approximation mode of branch measurement; s2, extracting a continuously iterated state metric entropy sequence, and inputting the sequence into an LSTM iteration control network for predicting a termination probability P; s3, based on a channel feature vector CNNfeed output by the CNN channel estimation network, optimizing a normalization factor Nk of the forward state measurement and the backward state measurement; and S4, performing iterative calculation, calculating external information and executing hard decision according to the branch metric adjusted in S1 and the normalization factor Nk optimized in S3, and outputting a decoding result, thereby solving the problems of high bit error rate, high power consumption, high time delay and the like of a traditional Turbo code decoding algorithm in a time-varying channel environment.

Description

Technical Field

[0001] This invention relates to the field of communication encoding and decoding technology, specifically to an adaptive Turbo code decoding method and system based on neural network assistance. Background Technology

[0002] Turbo codes, as a milestone technology in third-generation channel coding, possess excellent decoding performance approaching the Shannon limit through parallel concatenated recurrent convolutional codes and iterative decoding mechanisms, and are widely used in wireless communication, satellite communication and other fields.

[0003] Traditional Turbo code decoding algorithms achieve decoding by iteratively calculating branch metrics, state metrics, and log-likelihood ratios, but they face the following key problems in practical applications:

[0004] Poor adaptability to time-varying channels: Traditional algorithms have a fixed branch metric calculation mode, which cannot be dynamically adjusted according to channel quality. In complex time-varying channels, they are prone to high bit error rate and unstable decoding performance.

[0005] Low iteration efficiency: Traditional algorithms typically use a fixed number of iterations. Even if the channel quality is good or the decoding has converged, all iterations still need to be completed, leading to increased latency and power consumption.

[0006] High hardware implementation complexity: The exponential operation of branch metrics and the storage and normalization process of state metrics require a large amount of computing resources and storage units. Traditional algorithms lack adaptive optimization mechanisms. When implemented on terminal hardware such as FPGAs, they are easily limited by storage resources and logic units, making it difficult to balance performance and cost.

[0007] State metric normalization error: Forward / backward state metrics are prone to numerical overflow during iteration. Traditional normalization factors are mostly fixed values ​​or simple statistics, which cannot be dynamically optimized in combination with channel characteristics, resulting in the accumulation of normalization errors and affecting decoding accuracy.

[0008] With increasing demands for low latency, high reliability, and low power consumption, the aforementioned shortcomings of traditional Turbo code decoding algorithms have become increasingly prominent, necessitating the introduction of intelligent technologies to achieve adaptive optimization. Summary of the Invention

[0009] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide an adaptive Turbo code decoding method and system based on neural network assistance.

[0010] This invention is achieved through the following technical solution: The adaptive Turbo code decoding method based on neural network assistance includes the following steps:

[0011] S1. Receive the signal, transform it through time and frequency, and input it into the CNN channel estimation network to generate dynamic segmentation parameters and polynomial coefficient vectors, and adaptively adjust the approximation mode of the branch metric.

[0012] S2. Extract the entropy sequence of the state metric from the continuous iterations and input it into the LSTM iterative control network to predict the termination probability P.

[0013] S3. Based on the channel feature vector CNN_feat output by the CNN channel estimation network, optimize the normalization factor N of the forward and backward state metrics. k ;

[0014] S4. Perform iterative calculations based on the branch metric adjusted in S1 and the normalization factor N optimized in S3. k It calculates external information, performs hard decision-making, and outputs the decoding result.

[0015] In S1, the branching metric Γ at time k k The approximate patterns of (s′,s) include:

[0016] When SNR < θ1, the branch metric is calculated by using a 6th-order polynomial fitting exponent operation.

[0017] When θ1≤SNR<θ2, the branch metric is calculated using a 5-segment piecewise linear approximation.

[0018] When SNR≥θ2, the branch metric is calculated by combining linear interpolation with neural network error compensation term;

[0019] SNR stands for Signal-to-Noise Ratio. θ1 and θ2 are the segmented thresholds of the SNR, used for calculating the dynamic switching branch metric. The value range of θ1 is 0 to 3 dB, and the value range of θ2 is 3 to 5 dB.

[0020] The 6th degree polynomial f(x) is expressed as w i The coefficients are those of a 6th-degree polynomial.

[0021] When SNR < θ1, the formula for calculating the branch metric is:

[0022] Γ k (s′,s)=L c ·r k ·u k (s′,s)+L a (k)·u k (s′,s);

[0023] In the formula, L c r is the channel reliability factor. k u is the soft value of the receiver at time k. k(s′,s) are the input bits corresponding to the state transition, L a (k) represents the prior information at time k.

[0024] When θ1≤SNR<θ2, the slope of each segment is dynamically adjusted by CNN_feat, and the specific formula for calculating the branch metric is as follows:

[0025] Γ k (s′,s)=k j ·(L c ·r k +L a (k))+b j ;

[0026] In the formula, k j The slope of the piecewise linear approximation of the j-th segment is dynamically adjusted by CNN_feat, b j Let L be the intercept of the j-th segment. c ·r k +L a (k) represents the merged soft input value.

[0027] When SNR≥θ2, the neural network error compensation term is Δ, Δ=0.1CNN_feat, and the specific calculation formula for the branch metric is:

[0028] Γ k (s′,s)=linear_interp(L c ·r k +L a (k))+Δ;

[0029] In the formula, linear_interp() is the linear interpolation function.

[0030] The formula for calculating the state metric entropy in S2 is as follows:

[0031]

[0032] In the formula, h k Entropy is a state metric. k (s) is the forward state metric of state s at time k, 2 m This represents the number of encoder registers.

[0033] The normalization factor N in S3 k The optimization formula is:

[0034] N k =α·CNN_feat+β·max(A k B k );

[0035] In the formula: α and β are the adaptive coefficients obtained from training the neural network, and max(A k B k ) represents the maximum value of the forward state metric and the backward state metric.

[0036] Step S4 includes the following sub-steps:

[0037] S4-1. Iterate through steps S1 to S3;

[0038] S4-2. Determine whether the set termination iteration condition is met. If yes, proceed to S4-3; otherwise, return to S4-1.

[0039] S4-3. Based on the branch metric adjusted in S1 and the state metric optimized in S3, and the normalization factor N. k It calculates external information and performs hard decision-making, outputting the decoding result. The decision formula is:

[0040]

[0041] LLR i =L c ·r i +L a (i)+L ext (i);

[0042] In the formula: For the estimated value of the i-th bit, LLR i Let L be the log-likelihood ratio of the i-th bit. c r is the channel reliability factor. i To receive soft values, L a (i) represents prior information, L ext (i) refers to external information;

[0043] S4-4. Output the estimated bit values ​​as binary decoding data to complete the decoding.

[0044] In step S4-2, the iteration terminates and proceeds to step S4-3 when any of the following termination conditions are met;

[0045] Condition 1: The maximum number of iterations is reached;

[0046] Condition 2: Set the prediction termination probability P, which is greater than or equal to the expected target and the CRC check passes.

[0047] A neural network-assisted adaptive Turbo code decoding system, applied to the aforementioned neural network-assisted adaptive Turbo code decoding method, includes:

[0048] The CNN channel estimation network takes the time-frequency transformed signal as input, generates dynamic segmentation parameters and polynomial coefficient vectors, and adaptively adjusts the approximation mode of the branch metric.

[0049] The LSTM iterative control network extracts the entropy sequence of the state metric from continuous iterations as input to predict the termination probability P.

[0050] The optimization module is used to optimize the normalization factor N of the forward and backward state metrics. k ;

[0051] The decoding decision module is used to calculate external information and perform hard decisions, outputting the decoding result as the estimated value of the bits as the binary decoded data.

[0052] Compared with the prior art, the beneficial effects of the present invention are:

[0053] This application addresses the problems of high bit error rate, high power consumption, and high latency in traditional Turbo code decoding algorithms under time-varying channel environments by using deep learning technology to achieve intelligent and adaptive optimization of Turbo decoding.

[0054] This application utilizes LSTM to predict convergence trends based on state metric entropy sequences, reducing the average number of iterations from the traditional 8 to 4.8, decreasing latency by 40%, and power consumption by 40%. In stable channel scenarios, the number of iterations can be further reduced to 3-4, significantly reducing decoding power consumption and improving efficiency.

[0055] This application achieves storage optimization by reducing storage resource consumption by more than 40% through dynamic segmentation approximation and tail bit compression storage technology.

[0056] In this application, the parameter size of the lightweight CNN architecture is only 1 / 5 of that of the traditional deep learning model. When implemented on FPGA, the newly added module occupies 512KB of storage, and the LUT consumption increases by less than 5%, which adapts to the hardware constraints of terminal devices and reduces complexity. Attached Figure Description

[0057] Figure 1 This is a flowchart based on the method described in this application. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] Reference Figure 1The neural network-assisted adaptive Turbo code decoding method includes the following steps:

[0060] S1. Receive the signal, transform it through time and frequency, and input it into the CNN channel estimation network to generate dynamic segmentation parameters and polynomial coefficient vectors, and adaptively adjust the approximation mode of the branch metric.

[0061] In S1, the branching metric for the process of the state transitioning from s′ to s at time k is denoted as Γ. k (s′,s), then the branch metric Γ at time k k The adaptive approximation modes of (s′,s) include:

[0062] When SNR < θ1, it means the signal-to-noise ratio is low and the channel quality is poor. A 6th-order polynomial is used to fit the exponential operation to calculate the branch metric. The 6th-order polynomial is generated through a neural network; the 6th-order polynomial f(x) is expressed as... w i The coefficients are those of a 6th-degree polynomial.

[0063] When SNR < θ1, the formula for calculating the branch metric is:

[0064] Γ k (s′,s)=L c ·r k ·u k (s′,s)+L a (k)·u k (s′,s);

[0065] In the formula, L c r is the channel reliability factor. k u is the soft value of the receiver at time k. k (s′,s) are the input bits corresponding to the state transition, L a (k) represents the prior information at time k.

[0066] When θ1 ≤ SNR < θ2, it means the signal-to-noise ratio is moderate and the channel quality is moderate. A 5-segment piecewise linear approximation is used to calculate the branch metric. When θ1 ≤ SNR < θ2, the slope of each segment is dynamically adjusted by CNN_feat. The specific formula for calculating the branch metric is as follows:

[0067] Γ k (s′,s)=k j ·(L c ·r k +L a (k))+b j ;

[0068] In the formula, k j The slope of the piecewise linear approximation of the j-th segment is dynamically adjusted by CNN_feat, bj Let L be the intercept of the j-th segment. c ·r k +L a (k) represents the merged soft input value.

[0069] When SNR≥θ2, it means that the signal-to-noise ratio is high and the channel quality is good. The branch metric is calculated by using linear interpolation combined with the neural network error compensation term.

[0070] When SNR≥θ2, the neural network error compensation term is Δ, Δ=0.1CNN_feat, and the specific calculation formula for the branch metric is:

[0071] Γ k (s′,s)=linear_interp(L c ·r k +L a (k))+Δ;

[0072] In the formula, linear_interp() is the linear interpolation function.

[0073] Wherein, SNR is the signal-to-noise ratio, and θ1 and θ2 are the segmented thresholds of the signal-to-noise ratio, used for the calculation of the dynamic switching branch metric. θ1 and θ2 are dynamically adjusted by the real-time channel state parameters output by the CNN channel estimation network. The value range of θ1 is 0 to 3 dB, and the value range of θ2 is 3 to 5 dB.

[0074] S2. Extract the state metric entropy sequence of continuous iterations and input it into the LSTM iterative control network to predict the termination probability P; the formula for calculating the state metric entropy in S2 is:

[0075]

[0076] In the formula, h k Entropy is a state metric. k (s) is the forward state metric of state s at time k, 2 m This represents the number of encoder registers.

[0077] S3. Based on the channel feature vector CNN_feat output by the CNN channel estimation network, optimize the normalization factor N of the forward and backward state metrics. k The normalization factor N in S3 k The optimization formula is:

[0078] N k =α·CNN_feat+β·max(A k B k );

[0079] In the formula: α and β are the adaptive coefficients obtained from training the neural network, and max(A k B k ) represents the maximum value of the forward state metric and the backward state metric.

[0080] S4. Perform iterative calculations, based on the branch metric adjusted in step S1 and the normalization factor N optimized in step S3. k The external information is calculated and a hard decision is performed, outputting the decoding result. Further, step S4 includes the following sub-steps:

[0081] S4-1. Iterate through steps S1 to S3;

[0082] S4-2. Determine whether the set termination iteration condition is met. If yes, proceed to S4-3; otherwise, return to S4-1. In S4-2, if any of the following termination iteration conditions are met, terminate the iteration and proceed to S4-3.

[0083] Condition 1: Reach the set maximum number of iterations, thereby avoiding infinite iterations;

[0084] Condition 2: Set the prediction termination probability P, which is set when the value of P is greater than or equal to the expected target and the CRC check passes. In this embodiment, the expected target is set to 0.8.

[0085] S4-3. Based on the branch metric adjusted in S1 and the normalization factor N optimized in S3. k It calculates external information and performs hard decision-making, outputting the decoding result. The decision formula is:

[0086]

[0087] LLR i =L c ·r i +L a (i)+L ext (i);

[0088] In the formula: For the estimated value of the i-th bit, LLR i Let L be the log-likelihood ratio of the i-th bit. c r is the channel reliability factor. i To receive soft values, L a (i) represents prior information, L ext (i) refers to external information.

[0089] S4-4. Output the estimated bit values ​​as binary decoding data to complete the decoding.

[0090] A neural network-assisted adaptive Turbo code decoding system, applied to the aforementioned neural network-assisted adaptive Turbo code decoding method, includes:

[0091] The CNN channel estimation network takes the time-frequency transformed signal as input, generates dynamic segmentation parameters and polynomial coefficient vectors, and adaptively adjusts the approximation mode of the branch metric.

[0092] The LSTM iterative control network extracts the entropy sequence of the state metric from continuous iterations as input to predict the termination probability P.

[0093] The optimization module is used to optimize the normalization factor N of the forward and backward state metrics. k ;

[0094] The decoding decision module is used to calculate external information and perform hard decisions, outputting the decoding result as the estimated value of the bits as the binary decoded data.

[0095] The above description is merely an optional embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made using the content of the present invention under the concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A neural network-assisted adaptive Turbo code decoding method, characterized in that, Includes the following steps: S1. Receive the signal, transform it through time and frequency, and input it into the CNN channel estimation network to generate dynamic segmentation parameters and polynomial coefficient vectors, and adaptively adjust the approximation mode of the branch metric. S2. Extract the entropy sequence of the state metric from the continuous iterations and input it into the LSTM iterative control network to predict the termination probability P. S3, based on the channel feature vector CNN feat output by the CNN channel estimation network, optimizing the normalization factors N of the forward state metric and the backward state metric k ; S4, perform iterative calculation, adjust the branch metric according to S1 and optimize the normalization factor N according to S3 k , calculate external information and perform hard decision, output decoding result.

2. The adaptive Turbo code decoding method based on neural network assistance according to claim 1, characterized in that, In the S1, the branch metric Γ at time k k The approximation mode of (s', s) includes: When SNR < θ1, the branch metric is calculated by using a 6th-order polynomial fitting exponent operation. When θ1≤SNR<θ2, the branch metric is calculated using a 5-segment piecewise linear approximation. When SNR≥θ2, the branch metric is calculated by combining linear interpolation with neural network error compensation term; Where SNR is the signal-to-noise ratio, and θ1 and θ2 are the segmented thresholds of the signal-to-noise ratio.

3. The adaptive Turbo code decoding method based on neural network assistance according to claim 2, characterized in that, The 6th degree polynomial f(x) is expressed as w i The coefficients of a 6th-degree polynomial; When SNR < θ1, the formula for calculating the branch metric is: Γ k (s',s) = L c ·r k ·u k (s',s) + L a (k) · u k (s',s); where L c is the channel reliability factor, r k is the received soft value at time k, u k (s′,s) is the input bit corresponding to the state transition, L a (k) is the a priori information at time k.

4. The adaptive Turbo code decoding method based on neural network assistance according to claim 2, characterized in that, When θ1≤SNR<θ2, the slope of each segment is dynamically adjusted by CNN_feat, and the specific formula for calculating the branch metric is as follows: Γ k (s',s) = k j ·(L c r k +L a (k))+b j ; In the formula, k j The slope of the piecewise linear approximation of the j-th segment is dynamically adjusted by CNN_feat, b j Let L be the intercept of the j-th segment. c ·r k +L a (k) represents the merged soft input value.

5. The adaptive Turbo code decoding method based on neural network assistance according to claim 2, characterized in that, When SNR≥θ2, the neural network error compensation term is Δ, Δ=0.1CNN_feat, and the specific calculation formula for the branch metric is: C k (s′,s)=linear_interp(L c ·r k +L a (k)0+Δ; In the formula, linear_interp() is the linear interpolation function.

6. The adaptive Turbo code decoding method based on neural network assistance according to claim 1, characterized in that, The formula for calculating the state metric entropy in S2 is as follows: In the formula, h k Entropy is a state metric. k (s) is the forward state metric of state s at time k, 2 m This represents the number of encoder registers.

7. The adaptive Turbo code decoding method based on neural network assistance according to claim 1, characterized in that, The normalization factor N in S3 k The optimization formula is: N k =α·CNN_feat+β·max(A k ,B k ); In the formula: α and β are the adaptive coefficients obtained from training the neural network, and max(A k B k ) represents the maximum value of the forward state metric and the backward state metric.

8. The adaptive Turbo code decoding method based on neural network assistance according to claim 1, characterized in that, Step S4 includes the following sub-steps: S4-1. Iterate through steps S1 to S3; S4-2. Determine whether the set termination iteration condition is met. If yes, proceed to S4-3; otherwise, return to S4-1. S4-3. Based on the branch metric adjusted in S1 and the normalization factor N optimized in S3. k It calculates external information and performs hard decision-making, outputting the decoding result. The decision formula is: LLR i =L c ·r i +L a (i)+L ext (i); In the formula: For the estimated value of the i-th bit, LLR i Let L be the log-likelihood ratio of the i-th bit. c r is the channel reliability factor. i To receive soft values, L a (i) represents prior information, L ext (i) refers to external information; S4-4. Output the estimated bit values ​​as binary decoding data to complete the decoding.

9. The adaptive Turbo code decoding method based on neural network assistance according to claim 8, characterized in that, In step S4-2, the iteration terminates and proceeds to step S4-3 when any of the following termination conditions are met; Condition 1: The maximum number of iterations is reached; Condition 2: Set the prediction termination probability P, which is greater than or equal to the expected target and the CRC check passes.

10. A neural network-assisted adaptive Turbo code decoding system, characterized in that, The method applied to the neural network-assisted adaptive Turbo code decoding method according to any one of claims 1-9 includes: The CNN channel estimation network takes the time-frequency transformed signal as input, generates dynamic segmentation parameters and polynomial coefficient vectors, and adaptively adjusts the approximation mode of the branch metric. The LSTM iterative control network extracts the entropy sequence of the state metric from continuous iterations as input to predict the termination probability P. The optimization module is used to optimize the normalization factor N of the forward and backward state metrics. k ; The decoding decision module is used to calculate external information and perform hard decisions, outputting the decoding result as the estimated value of the bits as the binary decoded data.

Images