Denoising auxiliary decoding method based on IDRSN

Abstract

The invention discloses a de-noising auxiliary decoding method based on IDRSN. The invention relates to the technical field of communication and deep networks, and solves the problem of noise reduction of noise-attached signals of a receiving end by using a deep neural network in a complex communication environment. The method comprises the following steps: superposing an original signal and noise; calculating a log-likelihood ratio; sC decoding is carried out; calculating a residual error; improving deep residual shrinkage network processing; outputting a de-noised signal; the IDRSN integrates a deep residual network, an Inception module, an attention mechanism and a new threshold function, and through the algorithm, the network can effectively estimate received signal noise in a channel noise environment, accurately distinguish signals and noise and extract useful features, so that a prediction result is closer to a real channel gain, interference of the noise on information data is reduced, and the prediction accuracy is improved. And then decoding is carried out in combination with a traditional algorithm, so that the decoding performance can be effectively improved, and a more efficient decoding process is realized.

Description

Technical Field

[0001] The present invention relates to the field of communication and deep network technology, and in particular to an IDRSN-based denoising auxiliary decoding method. Background Art

[0002] In recent years, with the rapid development of fifth-generation mobile communication technology (5G), the requirements for communication systems have become more complex and stringent compared to fourth-generation technology (4G). In particular, in the context of 5G enhanced mobile broadband, polar codes have been identified as the short-code encoding scheme for control channels. This breakthrough has rekindled widespread academic attention and made polar codes a key research topic. However, in real-world communication environments, channels are inevitably affected by signal attenuation and noise interference. These factors often hinder the performance of traditional polar code decoding methods in the face of complex interference conditions. In particular, when the code length is long, existing decoding algorithms face high decoding delays, which poses a challenge for low-latency and high-accuracy communication systems. Therefore, optimization and improvement of polar code decoding algorithms are needed. Summary of the Invention

[0003] The purpose of the present invention is to provide a denoising auxiliary decoding method based on IDRSN to solve the problem of denoising the noisy signal at the receiving end using a deep neural network in a complex communication environment.

[0004] The present invention provides an IDRSN-based denoising auxiliary decoding method, comprising:

[0005] In step 1, noise n is added to the original signal u to generate a noisy signal y = u + n. The noisy signal represents the signal received in an actual communication environment and contains the superposition effect of the information component u and the noise component n. It is the input for the subsequent denoising process.

[0006] In step 2, the received noisy signal y is passed to the LLR calculation module, which analyzes the likelihood ratio of the received signal, extracts the signal reliability information, and generates an LLR value l. The LLR value represents the reliability of each bit in the received signal.

[0007] Step 3: The LLR value l is input to the SC decoder. The SC decoder is based on the principle of polar code and uses the reliability information provided by LLR to decode and generate a preliminary signal estimate by bit-by-bit elimination.

[0008] Step 4: After obtaining a preliminary signal estimate Then, it is subtracted from the noisy signal y to generate residual noise The residual noise Represents the noise portion remaining after removing the preliminary estimated signal from the received signal;

[0009] Step 5: Residual Noise It is input into IDRSN for processing. IDRSN uses the feature extraction and noise suppression capabilities of deep learning to perform further denoising operations on the residual noise and output noise estimation

[0010] Step 6: Noise estimation of IDRSN output As the final estimation result of the noise, the final estimation result is fed back to the whole system, and then the noisy complex signal minus the predicted noise is subtracted. Get the denoised signal Then recalculate the channel LLR value and SC decoding to get the decoding output

[0011] Furthermore, in step 5, IDRSN uses the following threshold function:

[0012]

[0013] Where x represents the input feature; represents the output feature; sgn(x) represents the sign function; τ represents the threshold; and N represents a constant. By introducing the attention mechanism, the DRSN can adaptively assign weights to different feature channels. Typically, it employs a SE module structure to model global information in the feature map and dynamically adjust channel strengths, thereby highlighting key features and suppressing irrelevant or noisy features. Furthermore, the feature contraction module, combined with the attention mechanism, automatically suppresses redundant features through a new threshold function, strengthening the network's ability to focus on valid signals. This design enables the IDRSN to achieve powerful feature extraction and interference resistance in high-noise environments.

[0014] Furthermore, in step 5, IDRSN uses the Inception module to capture multi-scale feature information in the same layer through multiple parallel convolution kernels and pooling layers of different sizes, thereby realizing parallel processing of multi-scale information.

[0015] Furthermore, in step six, SC decoding first calculates the LLR of each bit in the received sequence based on the model of the transmission channel to measure the probability that the current bit is 0 or 1. Then, starting from the first bit, decoding is carried out step by step. For frozen bits, the preset value is directly filled in. For information bits, the most likely value is determined based on the LLR value. Each decoded bit is used to update the conditional probability of subsequent bits.

[0016] Furthermore, in step 6, when the SC decoding processes the i-th polarization channel, its transition probability is determined by the signal received by the current decoder and the decoding results of the previous i-1 channels. constitute.

[0017] The present invention has the following beneficial effects: The present invention is an IDRSN denoising auxiliary decoding method, the IDRSN predictor integrates a deep residual network, an InceptionV2 module, an attention mechanism and a new threshold function, and introduces a new threshold function to replace the traditional soft threshold function to more effectively enhance the signal processing and noise recognition capabilities. The advantages of the InceptionV2 module over the standard convolution layer in signal noise estimation are mainly reflected in its parallel multi-scale feature extraction, optimization of computational complexity and better feature selection capabilities. Through convolution kernels of different sizes and pooling operations, InceptionV2 can capture features of different scales, effectively distinguish between signals and noise, and enhance robustness to noise. At the same time, it is more efficient and flexible in processing complex signals, and can better adapt to the estimation task of multi-frequency noise. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0019] Figure 1 This is the Inception V2 module structure diagram;

[0020] Figure 2 This is the IDRSN predictor network structure diagram;

[0021] Figure 3 This is the IDRSN-SC decoding framework structure diagram. DETAILED DESCRIPTION

[0022] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present invention and corresponding drawings. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention. The technical solutions provided by each embodiment of the present invention are described in detail below in conjunction with the drawings.

[0023] like Figure 2As shown in the figure, combined with the IDRSN Polar Codec Modulation System Block Diagram, this flowchart demonstrates a framework for signal denoising, integrating the IDRSN network and the SC decoding algorithm to effectively extract and restore clean signals from noisy signals. The entire process is divided into several key steps, and the parameters involved are defined as follows:

[0024] u is the original polar code signal, which has been polarization-encoded and modulated by binary phase shift keying (BPSK).

[0025] n is the real noise extracted from the channel simulation model obeying Gaussian distribution.

[0026] y is a noisy signal y=u+n generated after a channel simulation model obeying a Gaussian distribution.

[0027] l is the received noisy signal y, which is passed to the LLR calculation module. The LLR calculation module extracts the reliability information of the signal by analyzing the likelihood ratio of the received signal.

[0028] To send it to the SC decoder to obtain preliminary decoding results.

[0029] The noise component that is not eliminated by SC decoding is extracted from the received signal after removing the initial estimated signal for further denoising.

[0030] is the residual noise It is input into IDRSN for processing. Through the learning of multi-layer neural network, IDRSN can more effectively identify and suppress noise features, thereby outputting a more accurate noise estimation.

[0031] Output of the IDRSN network As the final estimation result of the noise, this estimation result can be fed back to the entire system, and then the predicted noise is subtracted from the noisy complex signal. Get the denoised signal

[0032] To recalculate the accurate channel LLR value and SC decoding, and obtain more accurate decoding output.

[0033] An embodiment of the present invention provides an IDRSN-based denoising auxiliary decoding method, comprising:

[0034] Step 1: Superposition of original signal and noise: First, noise n is added to the original signal u to generate a noisy signal y = u + n. The noisy signal represents the signal received in an actual communication environment, contains the superposition effect of the information component x and the noise component n, and is the input of the subsequent denoising process.

[0035] Step 2: Log Likelihood Ratio (LLR) calculation: The received noisy signal y is passed to the LLR calculation module, which analyzes the likelihood ratio of the received signal, extracts signal reliability information, and generates an LLR value l. The LLR value represents the reliability of each bit in the received signal and is used to guide subsequent decoding steps to improve decoding accuracy.

[0036] Step 3, Successive Cancellation (SC) decoding: The LLR value l is input to the SC decoder. The SC decoder, based on the principle of polar code, uses the reliability information provided by the LLR to decode and generate a preliminary signal estimate. The process of generating a preliminary signal estimate is an attempt to recover the original signal and can partially offset the effects of noise, but some residual noise may still exist.

[0037] Step 4: Residual calculation: After obtaining the preliminary signal estimate Then, it is subtracted from the noisy signal y to generate residual noise The residual noise Represents the noise remaining after removing the preliminary estimated signal from the received signal; this step extracts the noise components that are not eliminated by SC decoding for further denoising processing.

[0038] Step 5: Improved Deep Residual Shrinkage Networks (IDRSN) processing: residual noise It is input into IDRSN for processing. IDRSN uses the feature extraction and noise suppression capabilities of deep learning to perform further denoising operations on residual noise. Through the learning of multi-layer neural networks, IDRSN can more effectively identify and suppress noise features, thereby outputting a more accurate noise estimate.

[0039] The first improvement to the Deep Residual Shrinkage Network (DRSN) is the adoption of a new threshold function to replace the traditional soft threshold function, effectively addressing the constant deviation problem commonly seen in soft threshold functions during signal processing. This new threshold function avoids weakening important peak features in the signal, thereby better preserving the salient features of key information. This further enhances its ability to more accurately identify and suppress noise, making the signal recovery process more accurate and efficient.

[0040] In step 5, IDRSN uses the following threshold function:

[0041]

[0042] Where x represents the input feature; represents the output feature; sgn(x) represents the sign function; τ represents the threshold; N represents a constant.

[0043] The new threshold function effectively addresses the discontinuity problem at the threshold of the traditional hard threshold function, while also avoiding the drawback of the soft threshold function that causes some high-frequency information to be lost during processing. First, like the soft threshold function, the new threshold function maintains continuity, thus overcoming the discontinuity problem of the hard threshold function at the threshold. Next, when the input feature is greater than the threshold, the output feature gradually approaches the input feature as the input feature increases. When the input feature is less than the negative threshold, the output feature gradually approaches the input feature as the input feature decreases, thus overcoming the constant bias problem of the soft threshold function. Furthermore, the new threshold function flexibly chooses between soft and hard thresholding, and its shape can be adjusted by adjusting the constant N. When N → ∞, it behaves as a soft threshold function; when N → 0, it behaves as a hard threshold function. This design gives the new threshold function greater flexibility in practical applications, allowing it to balance denoising effectiveness and signal preservation accuracy according to different requirements.

[0044] The second improvement to DRSN is to use the Inception module to replace the standard convolution operation in the residual block. The Inception module can capture multi-scale feature information in the same layer through multiple parallel convolution kernels and pooling layers of different sizes, realize parallel processing of multi-scale information, help to effectively distinguish signals from noise, and enhance robustness to noise, while effectively estimating noise. This design can not only greatly improve the feature extraction capability of the model without significantly increasing the computational cost, but also optimize the parameter usage and enhance the adaptability and robustness of the model to inputs of different scales. At the same time, the multi-path feature processing mechanism helps to improve the gradient flow, thereby supporting the effective training of deeper networks, making the model perform better when handling complex denoising tasks. The InceptionV2 module structure is as follows: Figure 1 shown.

[0045] The Inception structure (also known as GoogLeNet) is a convolutional neural network architecture proposed by Google. Its core innovation is to extract multi-scale features by using convolution kernels of different sizes (such as 1x1, 3x3, 5x5) and pooling operations in parallel, while using 1x1 convolution for dimensionality reduction to reduce the amount of computation. Each Inception module consists of multiple convolution and pooling branches, which work in parallel to capture features from different scales and merge the results by splicing. The advantages of Inception are efficient computing and strong flexibility. It can reduce computing costs while capturing local and global features at the same time to further enhance the network's expressive power. This architecture optimizes computational efficiency by effectively reducing the number of parameters, such as Figure 1 As shown in the figure, InceptionV2 is a convolutional neural network improved on the basis of InceptionV1, which mainly accelerates training, stabilizes the training process and improves the generalization ability of the model by introducing batch normalization. In addition, V2 also optimizes the Inception module and adopts factorized convolution, such as decomposing a 5x5 convolution into two 3x3 convolutions to reduce the amount of calculation while maintaining the performance of the network. InceptionV2 further deepens the network layer and adopts more efficient data enhancement and optimization strategies during training, such as using the Adaptive Moment Estimation (Adam) optimizer. Through these optimizations, InceptionV2 significantly reduces the consumption of computing resources while improving accuracy. The network structure of the IDRSN predictor is shown in the figure. Figure 3 shown.

[0046] Step 6: Output denoised signal: Noise estimation of IDRSN output As the final estimation result of the noise, the final estimation result is fed back to the whole system, and then the noisy complex signal minus the predicted noise is subtracted. Get the denoised signal Then recalculate the accurate channel LLR value and SC decoding to obtain a more accurate decoding output

[0047] The core concept of the SC decoding algorithm is to exploit the polarization characteristics of the channel and recover the original information sequence bit by bit through a successive elimination process. Specifically, SC decoding first calculates the log-likelihood ratio (LLR) of each bit in the received sequence based on a transmission channel model, such as an additive white Gaussian noise (AWGN) channel. This LLR measures the likelihood that the current bit is 0 or 1. Next, decoding is performed step by step, starting with the first bit. Frozen bits are directly filled with preset values, while information bits are judged based on the LLR values ​​to determine their most likely values. Each decoded bit is used to update the conditional probability of subsequent bits, thereby gradually improving decoding accuracy.

[0048] As a step-by-step decoding algorithm, SC decoding has two main components when processing the i-th polarization channel: the signal received by the current decoder and the decoding results of the previous i-1 channels. In other words, polarized channels are interdependent. This relationship can be summarized as follows: the decoding results of higher-order polarized channels depend on the decoding outputs of lower-order channels. Therefore, polarized channel decoding is layer-by-layer dependent, and the decoding results of each layer affect the decoding of subsequent channels.

[0049] In summary, in order to meet the requirements of wireless signal channel coding and decoding in a complex and changeable electromagnetic environment, and to improve the problem that deep neural networks are difficult to converge effectively under noisy conditions, based on the powerful adaptability, nonlinearity and fault tolerance of deep neural networks, and on the basis of drawing on existing deep neural network theoretical technology and application scenarios, a method of combining neural networks with traditional decoders is proposed. By introducing a noise reduction mechanism, the system's ability to process noisy codewords is significantly enhanced. This cascaded traditional decoder method gives full play to the advantages of neural networks in noise suppression and feature learning, and effectively improves the overall decoding performance. The present invention improves on the basis of DRSN. IDRSN integrates a deep residual network, an Inception module, an attention mechanism and a new threshold function. Through this algorithm, the network can effectively estimate the received signal noise in a channel noise environment, accurately distinguish between signals and noise and extract useful features, so that the prediction results are closer to the actual channel gain, reducing the interference of noise on information data, and then combined with the traditional algorithm for decoding, which can effectively improve the decoding performance, thereby achieving a more efficient decoding process.

[0050] The above-described embodiments of the present invention do not limit the protection scope of the present invention.

Claims

1. A denoising auxiliary decoding method based on IDRSN, characterized in that: include: In step 1, noise n is added to the original signal u to generate a noisy signal y = u + n. The noisy signal represents the signal received in an actual communication environment and contains the superposition effect of the information component u and the noise component n. It is the input for the subsequent denoising process. In step 2, the received noisy signal y is passed to the LLR calculation module, which analyzes the likelihood ratio of the received signal, extracts the signal reliability information, and generates an LLR value l. The LLR value represents the reliability of each bit in the received signal. Step 3: The LLR value l is input to the SC decoder. The SC decoder is based on the principle of polar code and uses the reliability information provided by LLR to decode and generate a preliminary signal estimate by bit-by-bit elimination. Step 4: After obtaining a preliminary signal estimate Then, it is subtracted from the noisy signal y to generate residual noise The residual noise Represents the noise portion remaining after removing the preliminary estimated signal from the received signal; Step 5: Residual Noise It is input into IDRSN for processing. IDRSN uses the feature extraction and noise suppression capabilities of deep learning to perform further denoising operations on the residual noise and output noise estimation Step 6: Noise estimation of IDRSN output As the final estimation result of the noise, the final estimation result is fed back to the whole system, and then the noisy complex signal minus the predicted noise is subtracted. Get the denoised signal Then recalculate the channel LLR value and SC decoding to get the decoding output 2. The IDRSN-based denoising auxiliary decoding method according to claim 1, wherein: In step 5, IDRSN uses the following threshold function: Where x represents the input feature; represents the output feature; sgn(x) represents the sign function; τ represents the threshold; N represents a constant.

3. The IDRSN-based denoising auxiliary decoding method according to claim 1, wherein: In step 5, IDRSN uses the Inception module to capture multi-scale feature information in the same layer through multiple parallel convolution kernels of different sizes and pooling layers, thus achieving parallel processing of multi-scale information.

4. The IDRSN-based denoising auxiliary decoding method according to claim 1, wherein: In step six, SC decoding first calculates the LLR of each bit in the received sequence based on the transmission channel model to measure the probability that the current bit is 0 or 1. Then, starting from the first bit, decoding is carried out step by step. For frozen bits, the preset value is directly filled in. For information bits, the LLR value is used to determine their most likely value. Each decoded bit is used to update the conditional probability of subsequent bits.

5. The IDRSN-based denoising auxiliary decoding method according to claim 1, wherein: In step 6, when SC decoding processes the i-th polarization channel, its transition probability is determined by the signal received by the current decoder and the decoding results of the previous i-1 channels. constitute.

Images