Communication signal denoising enhancement method based on generative adversarial network

Abstract

The application discloses a communication signal denoising and enhancing method based on a generative adversarial network, and relates to the technical field of communication signal processing.The application adopts a double-discriminator architecture, independently evaluates two different defects of noise residue and signal distortion, identifies dominant defects and secondary defects according to the defects, introduces upper and lower boundaries and an asymmetric calibration mechanism based on historical data for dynamic adjustment, negatively constrains the dominant defects, positively compensates the secondary defects, and intelligently determines a fusion weight.Finally, the method can adaptively weight and fuse two enhanced signals of denoising priority and fidelity priority, effectively solves the problem that a traditional method cannot distinguish defect types and a single compensation strategy leads to difficulty in balancing noise suppression and detail fidelity, and significantly improves signal recovery quality and system robustness in a complex communication environment.

Description

Technical Field

[0001] This method relates to the field of communication signal processing technology, specifically a communication signal noise reduction and enhancement method based on generative adversarial networks. Background Technology

[0002] In wireless communication systems, received signals are superimposed with various noises and interferences during transmission, including thermal noise, multiplicative noise caused by channel fading, and external electromagnetic interference. This leads to a deterioration in signal quality, directly affecting the accuracy of subsequent processing such as demodulation and decoding. Communication signal noise reduction and enhancement technology aims to recover the original transmitted signal as much as possible from the noisy received signal, and is a key preprocessing step for improving the reliability of communication systems. As communication scenarios expand towards high-speed mobility, low signal-to-noise ratio, and complex electromagnetic environments, higher demands are placed on the adaptability and fidelity of noise reduction algorithms.

[0003] Currently, several mature technologies have been developed in the field of communication signal denoising. Traditional methods mainly include linear filtering, nonlinear filtering, and wavelet thresholding, which separate signals and noise based on the differences in statistical characteristics. Traditional methods are effective when the noise model is accurate and the signal is stationary, but their performance degrades significantly in non-stationary, low signal-to-noise ratio, or model mismatch scenarios. Furthermore, adaptive filtering methods require a reference signal or pilot sequence, making them unsuitable for blind denoising scenarios. In recent years, generative adversarial networks (GANs) have been explored for signal denoising. Through adversarial training between the generator and discriminator, they learn the mapping from noisy signals to clean signals. However, existing schemes typically employ a single discriminator structure, where the discriminator outputs only a scalar to determine whether the signal is clean, resulting in a relatively singular evaluation dimension for signal quality.

[0004] The aforementioned existing technologies generally suffer from a significant deficiency: they employ a single compensation strategy for overall signal processing, failing to distinguish between residual noise and signal distortion, two distinct types of defects, and further unable to perform asymmetric adaptive calibration targeting the dominant defect. Specifically, existing methods output only a single signal quality assessment value, which conflates residual noise and signal distortion, failing to quantify which defect is dominant in the current signal. Furthermore, existing methods typically employ a symmetrical adjustment strategy when correcting signals, failing to differentiate the intensity of noise suppression and detail preservation based on the type and severity of the dominant defect. That is, when residual noise is severe, noise should be strongly suppressed; when distortion is severe, fidelity preservation should be prioritized. This makes it difficult to achieve an ideal balance between strong noise suppression and detail preservation in the denoised signal. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this method provides a communication signal noise reduction and enhancement approach based on generative adversarial networks.

[0006] To achieve the above objectives, the technical solution of this method is as follows: a communication signal noise reduction and enhancement method based on generative adversarial networks, comprising:

[0007] S1. Acquire noisy signal data, input the noisy signal data into the generator, and output the first enhanced signal and the second enhanced signal;

[0008] S2. Input the first enhanced signal and the second enhanced signal into the first discriminator and the second discriminator respectively to obtain a first score characterizing the degree of noise residue and a second score characterizing the degree of signal distortion.

[0009] S3. Calculate the absolute value of the difference between the first score and the second score. If the absolute value of the difference exceeds a preset threshold, the larger of the first score and the second score is defined as the dominant defect score, and the smaller one is defined as the minor defect score.

[0010] S4. Based on historical defect scoring data, the upper and lower boundaries of the defect scores are obtained, and the basic adjustment coefficients are mapped based on the normalized distances between the dominant defect score and the upper boundary, and between the secondary defect score and the lower boundary, respectively, to obtain the dominant compensation factor and the secondary compensation factor, wherein the basic adjustment coefficient is the difference between the absolute value of the difference and the preset threshold.

[0011] S5. The dominant defect score and the dominant compensation factor are negatively constrained and fused, and the minor defect score and the minor compensation factor are positively compensated and fused. The first score and the second score after calibration are output, and the fusion weight is determined by the first score and the second score after calibration.

[0012] S6. The first enhanced signal and the second enhanced signal are weighted and fused according to the fusion weight to obtain the final enhanced signal.

[0013] Compared with existing technologies, the beneficial effects of this method are as follows:

[0014] 1. This invention decomposes signal quality into two orthogonal dimensions by outputting noise residue scores and signal distortion scores through two independent discriminators. Based on this, the primary and secondary defects of the current signal are identified by comparing the magnitudes of the two scores, and an asymmetric processing strategy is adopted: negative constraint fusion is performed on the primary defect, and positive compensation fusion is performed on the secondary defect. This differentiated calibration mechanism, which negatively constrains the primary defect and positively compensates for the secondary defect, solves the fundamental contradiction of balancing noise suppression and detail fidelity in traditional single compensation strategies.

[0015] 2. This invention utilizes historical defect scoring data to dynamically generate upper and lower boundaries through weighted probability density estimation, and introduces an adaptive boundary adjustment mechanism to ensure that the normalized distance can truly reflect the severity of the current defect relative to the historical distribution. This ensures that the compensation factor remains at a reasonable level even when the channel drifts slowly or changes abruptly, avoiding excessive suppression or insufficient enhancement.

[0016] 3. This invention employs an exponential function with asymptotic saturation characteristics to map the normalized distance to the compensation factor. This allows for rapid response even with severe defects, while avoiding unlimited compensation. Simultaneously, the sensitivity parameter is automatically adjusted based on the historical variance of the normalized distance, and its rate of change is greater for dominant defects than for secondary defects. This makes compensation more sensitive during periods of severe channel fluctuations, while maintaining stability during periods of calm. This allows the calibration intensity to be intelligently adjusted according to defect severity, historical stability, and defect type, further enhancing the system's robustness.

[0017] 4. The initial fusion weights of this invention are calculated based on the received signal strength indication value, so that the system naturally biases towards the noise reduction branch under extremely weak signal conditions and towards the fidelity branch under strong signal conditions, thereby responding quickly to changes in channel strength. Then, fine-tuning is performed through weight correction based on the calibrated score, so that the final fusion weights can reflect the actual noise reduction effect in real time. This dual adjustment mechanism not only ensures rapid decision-making in extreme cases, but also achieves accurate balance under normal conditions. Finally, the output signal achieves an ideal balance between strong noise suppression and high fidelity. Attached Figure Description

[0018] The disclosure of this method is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this method. In the drawings, the same reference numerals are used to refer to the same components. Wherein:

[0019] Figure 1 This is the flowchart of the method.

[0020] Figure 2 This is a flowchart of the data processing method. Detailed Implementation

[0021] It is readily understood that, based on the technical solution of this method, those skilled in the art can propose various interchangeable structural and implementation methods without altering the essential spirit of this method. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this method and should not be considered as the entirety of this method or as limitations or restrictions on the technical solution of this method.

[0022] like Figure 1 and Figure 2 As shown, the specific implementation steps of this method include the following steps:

[0023] S1. Acquire noisy signal data, input the noisy signal data into the generator, and output the first enhanced signal and the second enhanced signal.

[0024] The noisy signal data refers to the digital baseband signal obtained by the wireless communication receiver after antenna reception, radio frequency downconversion, and analog-to-digital conversion. Its sampling rate is determined according to the specific communication system. The noisy signal data is usually represented in complex form, containing real and imaginary parts, which correspond to in-phase and quadrature components, respectively.

[0025] After acquiring the noisy signal data, to facilitate processing by the generator network, the continuous signal is divided into signal frames of fixed length. Each frame contains a predetermined number of sampling points. Frame division is only to adapt to the fixed input size of the network and does not change the physical nature of the signal. In this method, one such signal frame is processed at a time and fed into the generator.

[0026] The preprocessed noisy signal data is input into the generator network. After training, the generator network can simultaneously output two enhanced signals with different optimization objectives: a first enhanced signal and a second enhanced signal.

[0027] The generator is the core module in a Generative Adversarial Network (GAN). In traditional GANs, the generator is responsible for producing fake data from random noise, while the discriminator judges the authenticity of the data. In this method, the generator's task is no longer to generate signals out of thin air, but to restore a clean signal from a noisy signal, i.e., to learn a mapping from the noisy domain to the clean domain. The generator is composed of multiple stacked neural network layers and is trained with a large number of pairs of noisy and clean signals, so that its output gradually approximates the real clean signal. After training, the generator can perform forward computation on any input noisy signal and output an enhanced signal. The generator in this method is designed to have two output branches, each focusing on a different optimization objective.

[0028] Obtain the received signal strength indication value corresponding to the noisy signal data; perform sliding median filtering on a predetermined number of consecutive received signal strength indication values ​​to obtain a smoothed received signal strength indication value; compare the smoothed received signal strength indication value with a preset first strength threshold and a second strength threshold.

[0029] If the smoothed received signal strength indication value is lower than the first strength threshold, the first enhanced signal is directly used as the final enhanced signal.

[0030] If the smoothed received signal strength indication value is higher than the second strength threshold, then the second enhanced signal is directly used as the final enhanced signal;

[0031] If the smoothed received signal strength indication value is between the first strength threshold and the second strength threshold, then steps S2 to S6 are executed to determine the final enhanced signal.

[0032] While acquiring noisy signal data, that is, within the same reception time window, the receiver front end measures the energy intensity of the current signal and outputs a value representing the signal power, namely the received signal strength indicator value. The larger the value, the stronger the signal.

[0033] Since each measurement may be subject to severe fluctuations due to transient fading or sudden interference, directly using the raw received signal strength indication value for mode determination would cause frequent jumps. To eliminate the influence of transient fluctuations and measurement noise, this method performs sliding median filtering on a predetermined number of consecutive received signal strength indication values, specifically:

[0034] Maintain a fixed-length sliding window. Each time the window slides, remove the oldest received signal strength indicator (RSSI) value and add the latest RSSI value. Then sort all the RSSI values ​​in the window and take the median as the smoothed RSSI value.

[0035] Two intensity thresholds are preset: a first intensity threshold and a second intensity threshold. The first intensity threshold ranges from -110dBm to -90dBm, with a preferred value of -95dBm. The second intensity threshold ranges from -80dBm to -70dBm, with a preferred value of -75dBm. These two thresholds can be adjusted within the above ranges according to the actual device noise floor and deployment environment.

[0036] The smoothed received signal strength indicator value is compared with two thresholds:

[0037] If the smoothed received signal strength indicator value is lower than the first strength threshold, the current signal mode is determined to be weak, indicating that the received signal strength is extremely weak. At this time, noise interference dominates, and signal quality is mainly affected by noise. The details of the fidelity priority branch output are overwhelmed by noise and have no value for preservation. Therefore, directly using the first enhanced signal as the final enhanced signal can suppress noise to the maximum extent and improve signal detectability. Steps S2 to S6 are skipped to save computational resources.

[0038] If the smoothed received signal strength indicator value is higher than the second strength threshold, it is determined to be a strong signal mode. At this time, the signal itself has rich detail information and relatively weak noise interference. Excessive noise reduction will introduce unnecessary smoothing distortion. Therefore, directly using the second enhanced signal as the final enhanced signal can preserve the signal's detailed features to the maximum extent and improve signal fidelity. Subsequent calibration and fusion steps are also skipped.

[0039] If the smoothed received signal strength indicator value is between two thresholds, it is determined to be in fusion mode. At this time, the signal has both noise and distortion. It is not possible to simply select one enhancement signal. The complete processing flow of steps S2 to S6 needs to be executed. The two enhancement signals are adaptively fused through the dual-channel compensation mechanism to obtain the optimal final enhancement signal.

[0040] The generator adopts an encoder-dual decoder architecture, specifically:

[0041] The noisy signal data is encoded into a high-dimensional feature representation by a shared encoder, and the high-dimensional feature representation is simultaneously input into the first decoder and the second decoder.

[0042] During the training phase, the first decoder optimizes by minimizing the mean square error and outputs a first enhanced signal with priority given to noise suppression.

[0043] During the training phase, the second decoder optimizes by maximizing structural similarity and outputs a second enhanced signal that prioritizes detail preservation.

[0044] During training, depth-constrained feature alignment is applied to the first decoder and the second decoder, and the difference between the two feature representations is constrained by the feature synchronization loss function.

[0045] The encoder-dual decoder architecture, through the design of a shared encoder and independent dual decoders, enables the generation of two enhanced signals with different optimization objectives in the same feature space. The encoder is responsible for compressing the input noisy signal data into a high-dimensional feature representation, which contains the main structural information and noise features of the signal. The first decoder and the second decoder reconstruct the signal based on the same high-dimensional feature representation, but use different optimization objectives, thereby generating enhanced signals with different characteristics.

[0046] The shared encoder employs a multi-layer convolutional neural network structure, compressing the input noisy signal data into a high-dimensional feature representation through layer-by-layer downsampling. Each layer of the encoder contains a certain number of convolutional kernels, the size and number of which vary with increasing network depth to progressively extract multi-level features of the signal. During encoding, batch normalization and activation function processing are performed after each convolutional layer to enhance the expressive power of the features. After encoding, the high-dimensional feature representation is simultaneously input into the first and second decoders, ensuring that both decoders reconstruct the signal based on the same feature foundation.

[0047] The first decoder employs a transposed convolutional network structure, reconstructing high-dimensional feature representations into time-domain signals through layer-by-layer upsampling. During training, the first decoder's optimization objective is to minimize the mean squared error (MSE) between the output signal and the clean reference signal. MSE is a commonly used regression loss function that measures the average of the squared errors between the predicted and true values. By minimizing the MSE, the first decoder learns a reconstruction strategy that tends to smooth the signal and effectively suppress noise, but may lose some signal details. This optimization objective allows the first decoder to prioritize noise suppression when processing noisy signals, resulting in enhanced signals with lower noise levels.

[0048] The second decoder also employs a transposed convolutional network structure, reconstructing high-dimensional feature representations into temporal signals through layer-by-layer upsampling. During training, the optimization objective of the second decoder is to maximize the structural similarity between the output signal and the clean reference signal. Structural similarity is an image quality assessment metric that comprehensively considers brightness, contrast, and structure, and it better reflects the human eye's perception of image quality. By maximizing structural similarity, the reconstruction strategy learned by the second decoder tends to preserve the structural details of the signal, but may not be sufficiently effective in suppressing noise. This optimization objective allows the second decoder to prioritize detail preservation when processing noisy signals, resulting in enhanced signals with high structural fidelity.

[0049] To ensure feature consistency between the outputs of the two decoders and avoid feature space splitting due to different optimization objectives, a deep-constrained feature alignment is applied to the first and second decoders during training. Specifically, at each layer output of the encoder, feature maps of corresponding layers in the first and second decoders are extracted, their similarity is calculated, and the difference between the two feature representations is constrained by a feature synchronization loss function. The goal of the feature synchronization loss function is to minimize the difference between the two feature representations, ensuring high consistency in the feature space between the two decoders. Through this deep-constrained feature alignment mechanism, the two decoders maintain consistency in feature extraction and representation despite their different optimization objectives.

[0050] Step S1 serves as the input processing and initial signal generation stage for the entire scheme. It acquires noisy signal data and sends it to the generator network. Through one forward propagation, it simultaneously obtains enhanced signals with two different optimization objectives. It synchronously measures the received signal strength indication value, obtains a stable strength value through sliding median filtering, and then compares it with two preset thresholds to classify it into three modes: weak signal mode, strong signal mode, and fusion mode. Based on the mode determination result, it directly outputs the corresponding enhanced signal under weak or strong signal conditions to avoid unnecessary complex calculations. Only in fusion mode are the two enhanced signals transmitted to subsequent steps for more refined adaptive calibration and fusion.

[0051] The sliding median filtering in this step effectively suppresses instantaneous jitter in the received signal strength indicator, making mode switching more stable and avoiding frequent misjudgments caused by a single outlier. In weak signal mode, the noise reduction priority signal is directly used, achieving maximum noise suppression at extremely low signal-to-noise ratios while saving computational resources. In strong signal mode, the fidelity priority signal is directly used, preventing excessive smoothing from introducing unnecessary distortion and preserving the original signal characteristics. The encoder-dual decoder architecture, combined with different training objectives, can generate two complementary enhanced versions, providing rich candidate sources for subsequent dynamic fusion. Deep constraint feature alignment forces the two decoders to maintain feature consistency during training, avoiding mode collapse and improving the performance of both branches. During online inference, the generator's forward computation is executed only once, and the two enhanced signals are output in parallel, resulting in high computational efficiency, making it suitable for real-time communication systems.

[0052] In practical implementation, the generator architecture can be replaced with an encoder-single decoder-dual output head structure, where two output heads branch out from the last layer of a single decoder, each corresponding to one of the two optimization objectives. The filtering method for the received signal strength indication value can be replaced with moving average filtering, Kalman filtering, or low-pass filtering. The first and second intensity thresholds can be fixed values ​​or dynamically adjusted based on the mean and standard deviation of historical received signal strength indication values. The network structures of the encoder and decoder can be replaced with recurrent neural networks, attention mechanism networks, or other deep learning architectures. The mean squared error loss function can be replaced with mean absolute error loss or other regression loss functions. The structural similarity loss function can be replaced with perceptual loss, feature matching loss, or other image quality assessment metrics.

[0053] S2. Input the first enhanced signal and the second enhanced signal into the first discriminator and the second discriminator respectively to obtain a first score characterizing the degree of noise residue and a second score characterizing the degree of signal distortion.

[0054] In generative adversarial networks (GANs), discriminators are typically trained to distinguish between real and generated samples. In this method, however, the discriminators are no longer a single module that only outputs true / false probabilities. Instead, they are designed as two independent networks with specific quality assessment functions. Specifically, the first discriminator evaluates the amount of residual noise in the signal output by the denoising-priority branch, while the second discriminator evaluates the severity of waveform distortion in the signal output by the fidelity-priority branch. These two discriminators are trained adversarially against the generator during the training phase. The generator attempts to produce a cleaner signal to deceive the discriminators, while the discriminators strive to identify defects in the signal. After sufficient training, the discriminators acquire the ability to objectively score. During online inference, the generator's parameters are fixed, and the discriminators are only used to output scores and no longer participate in backpropagation.

[0055] The first discriminator and the second discriminator are two completely independent discriminator networks, specifically:

[0056] The first discriminator is a noise residual detector. It takes the first enhanced signal as input, extracts the noise-related features of the signal layer by layer through a multi-layer convolutional network, and maps them to a first score. The first score is monotonically positively correlated with the degree of noise residual in the first enhanced signal.

[0057] The second discriminator is a signal distortion detector. It takes the second enhanced signal as input, extracts the waveform distortion-related features of the signal layer by layer through a multi-layer convolutional network, and maps them to a second score. The second score is monotonically positively correlated with the degree of signal distortion in the second enhanced signal.

[0058] The first discriminator receives the first enhanced signal, which is the output of the noise reduction priority branch. Since this branch is given a strong mean square error loss during training, its output signal has very little residual noise, but there may still be a small amount of noise. The first discriminator extracts deep features from the input signal through a multi-layer convolutional network. Each convolutional layer uses a small-sized convolutional kernel, gradually increasing the number of channels and compressing the time dimension. The last layer of the network is a fully connected layer, which maps the high-dimensional features to a scalar value, which is the first score.

[0059] During training, the first discriminator is given pairs of noisy and clean signals, with the noisy signals serving as positive samples and the clean signals as negative samples for adversarial training. After adversarial training, the first discriminator learns to distinguish the amount of residual noise in the signal, and its output value is monotonically positively correlated with the degree of residual noise: the larger the output value, the more residual noise there is, and the smaller the output value, the more thorough the noise suppression.

[0060] Specifically, the input of the first discriminator is the first enhanced signal, and the output is the first score, which ranges from 0 to 1. When the noise residue in the input signal is severe, the first score is close to 1; when the noise residue in the input signal is slight, the first score is close to 0.

[0061] The second discriminator receives the second enhanced signal, which is the output of the fidelity-first branch. This branch is given a strong structural similarity loss during training, and its output signal retains rich details but may have waveform distortion. The network structure of the second discriminator is similar to that of the first discriminator. It extracts deep features from the input signal through a multi-layer convolutional network. Each convolutional layer uses a small-sized convolutional kernel, gradually increasing the number of channels and compressing the time dimension. The last layer of the network is a fully connected layer, which maps the high-dimensional features to a scalar value, which is the second score.

[0062] During training, the second discriminator is given pairs of clean signals and artificially constructed distorted signals, with the clean signals serving as negative samples and the distorted signals as positive samples for adversarial training. Through adversarial training, the second discriminator learns to distinguish whether a signal is distorted, and its output value is monotonically positively correlated with the degree of distortion: the larger the output value, the more severe the distortion, and the smaller the output value, the more faithful the waveform.

[0063] Specifically, the input to the second discriminator is the second enhanced signal, and the output is the second score, which also ranges from 0 to 1. When the waveform distortion in the input signal is severe, the second score is close to 1; when the waveform distortion in the input signal is slight, the second score is close to 0.

[0064] Step S2, as the defect assessment stage, independently evaluates the quality of the two enhanced signals generated in step S1. A specialized discriminator network quantifies the degree of noise residue and signal distortion. The first discriminator maps the first enhanced signal to a first score, and the second discriminator maps the second enhanced signal to a second score. These two scores serve as inputs for subsequent steps, identifying the main defect types of the current signal and driving asymmetric adaptive calibration. This step provides objective and orthogonal quality feedback for the entire method and is a key decision-making basis for the entire adaptive fusion mechanism.

[0065] This step employs two completely independent discriminator networks, which can evaluate signal quality from two orthogonal dimensions: noise and distortion. This ensures the independence and accuracy of the evaluation process and avoids mutual interference between evaluations of different types of defects. The monotonic positive correlation of the scoring mapping ensures that the score values ​​can truly reflect the severity of the defects, providing a reliable basis for subsequent compensation decisions. The discriminator automatically extracts features based on deep learning, which can more accurately reflect the true quality of the signal compared to traditional manually designed indicators. The standardized design of the scoring range facilitates subsequent normalization processing and comparative analysis.

[0066] In practical implementation, the discriminator network structure can be replaced with a fully connected neural network, a recurrent neural network, or an attention mechanism network; the scoring mapping method can be replaced with a linear mapping, a multinomial mapping, or a piecewise linear mapping; the scoring range can be adjusted to other numerical ranges according to actual needs, such as a percentage score from 0 to 100; the training method of the discriminator can be replaced with unsupervised learning or semi-supervised learning to reduce the dependence on labeled data; the first and second discriminators can be merged into a single discriminator network, which outputs two scores simultaneously through a multi-output head structure; the feature extraction method can be replaced with frequency domain analysis, wavelet transform, or other signal processing methods; and time-domain smoothing can be introduced into the score calculation to avoid instantaneous fluctuations in the score.

[0067] S3. Calculate the absolute value of the difference between the first score and the second score. If the absolute value of the difference exceeds a preset threshold, the larger of the first score and the second score is defined as the dominant defect score, and the smaller of the two scores is defined as the minor defect score.

[0068] The calculation of the absolute value of the difference is to determine the degree of difference between the two defect scores. The first discriminator outputs a first score representing the degree of noise residue, and the second discriminator outputs a second score representing the degree of signal distortion. Both scores are between 0 and 1. The absolute value of their difference is calculated. The larger the absolute value of the difference, the more inconsistent the judgments of the two discriminators on the signal quality, that is, the signal may have a serious defect in a certain aspect. The smaller the absolute value of the difference, the more similar the evaluations of the two discriminators are, and the signal performs evenly in terms of noise and distortion.

[0069] The preset threshold is a configurable parameter used to determine whether the two scores are significantly different. The value of the threshold is usually between 0.05 and 0.20, with a preferred value of 0.10. The specific value can be adjusted according to the actual application scenario. For scenarios with relatively stable noise environment, the threshold can be appropriately reduced to improve the compensation accuracy; for scenarios with large changes in noise environment, the threshold can be appropriately increased to enhance the robustness of the system.

[0070] After calculating the absolute value of the difference between the first score and the second score, the method further includes:

[0071] The absolute value of the difference is compared with a preset threshold. If the absolute value of the difference does not exceed the preset threshold, the first score is directly used as the first score after calibration, and the second score is used as the second score after calibration.

[0072] After calculating the absolute value of the difference, the absolute value of the difference is compared with the preset threshold to determine whether it is necessary to distinguish between primary and secondary defects and how to perform scoring calibration.

[0073] If the absolute value of the difference does not exceed the preset threshold, it is considered that the evaluations of the two discriminators are basically consistent, and the noise and distortion levels in the signal are similar. There is no need to perform complex asymmetric calibration. Therefore, the original first score is directly used as the first score after calibration, and the original second score is used as the second score after calibration. The subsequent normalization, compensation factor calculation and other steps are skipped. This can save computing resources and maintain the original evaluation of signal quality.

[0074] If the absolute value of the difference exceeds the preset threshold, it is considered that there is a significant difference between the two evaluations, and the signal has a dominant defect in a certain aspect. Therefore, the larger of the two scores is defined as the dominant defect score, and the smaller one is defined as the secondary defect score. Subsequent steps S4 and S5 will perform asymmetric processing on the dominant and secondary scores respectively: the dominant score will be suppressed, and the secondary score will be enhanced, so that the two will converge.

[0075] Step S3, as the defect primary / secondary differentiation stage, determines whether primary / secondary differentiation is necessary based on the degree of difference between the two defect scores. When the difference is significant, the primary and secondary defects are identified, providing a decision-making basis for subsequent targeted compensation. When the difference is not significant, the original score is used directly, simplifying the processing flow. The output of step S3 directly determines whether differentiated compensation for primary and secondary defects is needed in subsequent steps, making it a key decision-making link in the entire adaptive fusion mechanism, connecting the defect assessment stage and the compensation factor calculation stage.

[0076] This step introduces a preset threshold, which allows for flexible adjustment of the system's sensitivity. A lower threshold enables the system to respond to subtle differences, while a higher threshold makes the system more stable. When two scores are close, the original score is used directly, avoiding complex normalization and nonlinear mapping operations without calibration, thus reducing average computational complexity and improving real-time processing capabilities. By clearly identifying dominant and secondary defects, a clear direction is provided for subsequent differential compensation, effectively improving the overall performance of noise reduction and fidelity. The magnitude of the absolute value of the difference itself also implies information about the degree of conflict, which can be used in subsequent steps to adjust the compensation intensity for more refined adaptation.

[0077] In practice, the preset threshold can be dynamically adjusted based on the statistical characteristics of signal quality or historical score differences, or an adaptive threshold can be adopted based on the specific values ​​of the two scores. For the identification of dominant defects, in addition to direct comparison, the size and confidence of the scores can be comprehensively considered through weighted voting or fuzzy logic. The calculation of the absolute value of the difference can be replaced by the calculation of the relative difference, that is, the difference is divided by the average of the two scores. The primary and secondary distinction mechanism can be replaced by a multi-level distinction mechanism, for example, the difference can be divided into three levels: slight difference, moderate difference and significant difference, based on the size of the difference, and different compensation strategies can be adopted for each level.

[0078] S4. Based on historical defect scoring data, obtain the upper and lower boundaries of the defect scores, and map the basic adjustment coefficients based on the normalized distances between the dominant defect score and the upper boundary, and between the secondary defect score and the lower boundary, respectively, to obtain the dominant compensation factor and the secondary compensation factor, wherein the basic adjustment coefficient is the difference between the absolute value of the difference and the preset threshold.

[0079] First, this step dynamically calculates two thresholds—an upper boundary and a lower boundary—using historically accumulated first and second scores through a weighted statistical method. These boundaries reflect the normal fluctuation range of signal quality. Then, for the dominant and minor defect scores of the current frame, their normalized distances relative to the boundaries are calculated, thus converting absolute scores into relative positions. Next, using the base adjustment coefficient obtained in step S3 as the amplitude benchmark—the difference between the absolute value of the difference and the preset threshold—the normalized distance is mapped to a dominant compensation factor and a minor compensation factor through two independent exponential functions. The dominant compensation factor is used to suppress dominant defects, and the minor compensation factor is used to enhance minor defects.

[0080] The upper and lower boundaries of the defect score obtained based on historical defect scoring data include:

[0081] Construct a time-series queue of historical defect scoring data, wherein the time-series queue stores the first and second scores obtained from each signal data processing in chronological order;

[0082] Each first score and each second score in the time-series queue is assigned a weight that decays exponentially over time;

[0083] Obtain the weighted historical defect score data, and use probability density estimation to determine its high quantile and low quantile, which are respectively used as the upper boundary and the lower boundary; the value range of the high quantile is 90% to 100%, and the value range of the low quantile is 1% to 10%.

[0084] The time-series queue is a first-in, first-out (FIFO) data structure used to store historical defect scoring data. The length of the time-series queue can be set according to actual needs. It stores the first and second scores obtained from each frame of signal processing in chronological order. The latest score data is added to the end of the queue, and the oldest score data is removed from the head of the queue, providing the latest statistical basis for boundary determination.

[0085] To emphasize the importance of recent data, each score in the time-series queue is assigned a weight that decays exponentially over time. The weight is calculated using an exponential function with a decay factor as the base and the time interval as the exponent. The decay factor determines the rate at which the weight decays with increasing frame intervals, ranging from 0.9 to 0.99. A decay factor closer to 1 indicates slower decay and a longer duration of historical data influence; a smaller decay factor indicates faster decay and a stronger dominance of recent data. The time interval represents the number of frames since the current time, determining the degree of weight decay of a specific historical frame relative to the current frame; a larger interval results in a smaller weight; a zero interval corresponds to a weight of 1. This exponentially decaying weight mechanism ensures that recent data has a greater impact on boundary determination, while the influence of older data gradually decreases, allowing the boundaries to better reflect the current signal processing status.

[0086] After assigning weights to the historical defect scoring data, all weighted first and second scores are combined and treated as a single sample set. A probability density estimation method is then used to determine the high and low quantiles of the weighted data. The high quantile ranges from 90% to 100%, with a preferred value of 95%; the low quantile ranges from 1% to 10%, with a preferred value of 5%. These two quantiles serve as the upper and lower boundaries, respectively. The specific values ​​of the high and low quantiles can be adjusted according to the actual application scenario. For scenarios requiring high compensation accuracy, the high quantile can be increased and the low quantile decreased; for scenarios requiring high processing speed, the high quantile can be decreased and the low quantile increased.

[0087] The probability density estimation is a statistical method that estimates the probability density function of the data by analyzing the distribution characteristics of the data, and then determines the value corresponding to a specified quantile based on the probability density function.

[0088] When the first score or the second score exceeds the upper boundary or the lower boundary, a boundary adaptive adjustment mechanism is triggered, specifically as follows:

[0089] If the first score or the second score exceeds the upper boundary, the current upper boundary is adjusted upward according to the dynamic learning rate; if the first score or the second score exceeds the lower boundary, the current lower boundary is adjusted downward according to the dynamic learning rate; each adjustment does not exceed a predetermined proportion of the current boundary value.

[0090] The dynamic learning rate is negatively correlated with the number of consecutive frames that have not been exceeded from the current boundary.

[0091] Specifically, when a new score exceeds the upper boundary, it means that the quality defect of the current signal is more severe than in most historical cases. If the upper boundary is not adjusted, the normalized distance of this severe defect may be very small in the subsequent normalized distance calculation. However, if the upper boundary remains unchanged, the boundary will not be able to reflect the new, worse situation. Therefore, the upper boundary needs to be raised to accommodate this more extreme defect value so that the subsequent normalized distance can still fairly reflect the severity of the defect.

[0092] When the new score is below the lower boundary, it means that the quality defect of the current signal is less severe than in most historical cases. If the lower boundary is not adjusted, this extremely slight defect will be assigned a large normalized distance in the subsequent normalized distance calculation, which may result in an unnecessary boost. Therefore, the lower boundary needs to be lowered to reflect the new higher quality and make the benchmark of the normalized distance more reasonable.

[0093] The learning rate is equal to a preset baseline learning rate divided by the sum of the number of consecutive frames that have not been exceeded and 1. The baseline learning rate ranges from 0.1 to 0.3, which determines the sensitivity of the boundary to each exceedance. The larger the learning rate, the faster the boundary adjustment; the smaller the learning rate, the smoother the adjustment.

[0094] The number of consecutive frames that have not been exceeded refers to the number of times that neither the first score nor the second score exceeded the boundary in all consecutive signal frames between the last boundary adjustment and the current exceedance. If a boundary has been stable for a long time, then when it is finally exceeded, only a small adjustment is made, because the previous stability indicates that the boundary is reasonable; conversely, if the boundary is frequently exceeded, it indicates that the current boundary may no longer be suitable for the new signal distribution and should be adjusted quickly.

[0095] The predetermined ratio ranges from 5% to 10%.

[0096] The normalized distance includes a first normalized distance and a second normalized distance, specifically:

[0097] The first normalized distance is equal to the upper boundary minus the dominant defect score, divided by the difference between the upper boundary and the lower boundary.

[0098] The second normalized distance is equal to the minor defect score minus the lower boundary, divided by the difference between the upper boundary and the lower boundary.

[0099] The primary and secondary defect scores are absolute values ​​output by the discriminator. Since the upper and lower boundaries are dynamically changing, the same absolute score may correspond to completely different relative severity at different times. Therefore, directly using the difference between the original scores cannot eliminate the influence of boundary changes, leading to unstable compensation levels.

[0100] To address this issue, this step converts the primary defect score and secondary defect score into normalized distances, which map the original scores to a uniform 0-1 range.

[0101] The first normalized distance reflects the suppression space of the dominant defect score, indicating how much of the dominant defect score is still remaining from the upper boundary. When the dominant defect score is closer to the upper boundary, the normalized distance is closer to 0, indicating that the suppression space is very small and strong suppression is required. When the dominant defect score is closer to the lower boundary, the normalized distance is closer to 1, indicating that the dominant defect is minor and the suppression intensity can be reduced.

[0102] The second normalized distance reflects the potential for improvement in the minor defect score, indicating how far the minor defect score is from the lower boundary. When the minor defect score is closer to the lower boundary, the normalized distance is closer to 0, indicating that the minor defect is very minor and has a large potential for improvement, so a larger positive compensation should be applied. When the minor defect score is closer to the upper boundary, the normalized distance is closer to 1, indicating that the minor defect is already quite serious and has little potential for improvement, so the compensation should be reduced accordingly.

[0103] The mapping process is implemented using a dual-channel exponential function with asymptotic saturation characteristics, specifically as follows:

[0104] The dominant compensation factor is calculated by a first exponential function, the output of which increases with the increase of the first normalized distance, and the first exponential function has an adjustable first sensitivity parameter.

[0105] The secondary compensation factor is calculated by a second exponential function whose output increases as 1 minus the second normalized distance increases. The second exponential function has an adjustable second sensitivity parameter.

[0106] The first sensitivity parameter and the second sensitivity parameter are dynamically adjusted based on the historical statistical variances of the first normalized distance and the second normalized distance, respectively, as follows:

[0107] Maintain a first sliding window and a second sliding window, respectively store the first normalized distance and the second normalized distance of a predetermined number of consecutive signal frames, and respectively calculate the first variance in the first sliding window and the second variance in the second sliding window;

[0108] The first sensitivity parameter is positively correlated with the first variance, the second sensitivity parameter is positively correlated with the second variance, and the rate at which the first sensitivity parameter changes with the first variance is greater than the rate at which the second sensitivity parameter changes with the second variance.

[0109] After obtaining the first and second normalized distances, this step requires converting them into compensation factors that are actually used for calibration. The total magnitude of the compensation factors is controlled by the base adjustment coefficient, while the sensitivity of the compensation factors to changes in the normalized distance is determined by the sensitivity parameter of the exponential function.

[0110] The basic adjustment coefficient is the difference between the absolute value of the difference calculated in step S3 and the preset threshold. It represents the degree to which the difference between the scores of the two discriminators exceeds the normal tolerance range. The greater the difference exceeds the threshold, the larger the basic adjustment coefficient, which means that the gap between the dominant and secondary defects in the current signal is more significant, requiring stronger calibration. The basic adjustment coefficient is a non-negative number, ranging from 0 to 1. In actual communication systems, the absolute value of the difference between the first and second scores usually does not reach 1, and the preset threshold is generally taken as 0.05 to 0.15. Therefore, the actual range of the basic adjustment coefficient is usually between 0 and 0.85.

[0111] The exponential function form has an asymptotic saturation characteristic. When the input is small, the output is approximately linear; when the input increases, the output growth rate gradually slows down and eventually approaches 1. This characteristic meets the calibration requirements, that is, when the severity of the defect, i.e., the normalized distance, is very small or very large, the compensation factor should respond quickly; however, when the defect is already very severe, i.e., the normalized distance is close to 0, the compensation factor cannot increase indefinitely, but should tend to an upper limit, i.e., the base adjustment coefficient.

[0112] The formula for calculating the dominant compensation factor is:

[0113] ;

[0114] Where A is the adjustment coefficient; The first normalized distance, ranging from 0 to 1, represents the remaining proportion of the dominant defect score from the upper boundary; The first sensitivity parameter controls the steepness of the exponential function. The larger the value, the better the function. It can achieve higher output values ​​at smaller values, meaning it is more sensitive to suppressing dominant defects; The smaller the value, the higher the output. Growth is more moderate.

[0115] When the dominant defect is severe, that is When it is very small, the value inside the parentheses is approximately equal to Compensation factor and It is approximately linear and can be adjusted proportionally according to the severity; when the dominant defect is minor, i.e. When the value is close to 1, the value inside the parentheses approaches 1, the compensation factor approaches the basic adjustment coefficient A, and the maximum suppression amount is reached.

[0116] The formula for calculating the secondary compensation factor is:

[0117] ;

[0118] Where A is the adjustment coefficient; The second normalized distance, ranging from 0 to 1, represents the remaining proportion of the minor defect score from the lower boundary; There is room for improvement because when the minor defect score is close to the lower boundary, there is a large room for improvement. The second sensitivity parameter controls the sensitivity of the secondary compensation factor to changes in the upscaling space: The larger the value, the faster the compensation factor will approach the base adjustment coefficient when there is more room for improvement, meaning the improvement is more sensitive. The smaller the value, the more gradual the increase in potential for improvement.

[0119] When minor defects are slight When the potential for improvement is small, and the value within the parentheses is close to 1, the compensation factor is close to A, indicating a strong improvement; when the secondary defect is significant... When the score approaches 1 and the potential for improvement approaches 0, the compensation factor approaches 0 and stops increasing. Thus, the minor defect score will be increased by an amount proportional to the potential for improvement.

[0120] The first and second sensitivity parameters are dynamically adjusted based on the historical fluctuations of their respective normalized distances. Specifically, the system maintains two sliding windows, storing the most recent first and second normalized distance sequences respectively, and calculates the variance of the samples within each window. A larger variance indicates more drastic changes in the normalized distance and greater channel instability. In this case, the sensitivity parameters should be increased to make the exponential function more sensitive to distance changes, thereby accelerating the compensation response. Simultaneously, the first sensitivity parameter changes with variance at a greater rate than the second sensitivity parameter, meaning the dominant defect is more sensitive to variance, further reinforcing the asymmetry.

[0121] Step S4, as the core step in calculating the compensation factor, inherits the dominant and secondary defect scores output from Step S3, along with the difference between the absolute value of the difference and the preset threshold. It provides two key compensation factors for the score calibration in Step S5: the dominant compensation factor and the secondary compensation factor. This step first dynamically generates upper and lower boundaries using a weighted historical queue and quantile estimation, and employs an adaptive boundary adjustment mechanism to update these boundaries in response to changes in the signal score. Then, the dominant and secondary defect scores are converted into normalized distances to eliminate the influence of different boundary scales. Finally, a dual-channel exponential nonlinear function maps the normalized distances to the dominant and secondary compensation factors, respectively used for negative suppression of the dominant defect score and positive enhancement of the secondary defect score. This step serves as a bridge between defect identification and score calibration; its output directly determines the magnitude and direction of the calibration in Step S5.

[0122] This step, based on a boundary determination mechanism using historical defect scoring data, avoids the limitations of fixed thresholds, allowing the boundary range to adaptively adjust according to the characteristics of the actual application scenario, thus improving the system's adaptability and robustness. The adaptive boundary adjustment mechanism automatically adjusts the boundary range when abnormal scores are detected; it quickly adjusts the boundary when scores frequently exceed the boundary, and slowly adjusts it when scores stabilize, thereby implementing an adaptive boundary adjustment strategy that ensures system stability and response speed. The normalized distance calculation normalizes the distance between the defect score and the boundary to a fixed range of 0 to 1, enabling unified mapping processing of defect scores under different boundary ranges, improving the stability and consistency of the mapping process. The mapping processing using a dual-channel exponential function achieves asymptotic saturation compensation. The compensation effect is as follows: when the defect score is close to the boundary, the compensation factor increases slowly; when the defect score is far from the boundary, the compensation factor increases rapidly. This asymptotic saturation characteristic can avoid overcompensation for extreme defects, ensuring the rationality and effectiveness of the compensation. The dynamic adjustment mechanism of the sensitivity parameter can adaptively adjust the compensation intensity according to the historical fluctuation of the normalized distance. When the normalized distance fluctuation is large, the sensitivity parameter is increased; when the normalized distance fluctuation is small, the sensitivity parameter is decreased, thereby improving the adaptability and stability of the compensation. At the same time, the design that the rate of change of the first sensitivity parameter with the first variance is greater than the rate of change of the second sensitivity parameter with the second variance makes the dominant compensation factor more sensitive to the fluctuation of the normalized distance than the secondary compensation factor, which is in line with the principle that dominant defects need to be compensated in a focused manner.

[0123] In implementation, regarding boundary determination, the probability density estimation method can be replaced with moving average, moving median, kernel density estimation, or other statistical methods; the weight decay method can be replaced with linear decay, polynomial decay, Gaussian decay, or other decay functions; the specific values ​​of the high and low quantiles can be flexibly adjusted within the stated range. Regarding boundary adaptive adjustment, the calculation of the dynamic learning rate can be replaced with sliding window, Kalman filtering, exponential smoothing, or other adaptive algorithms; the adjustment amplitude limitation mechanism can be replaced with dynamic adjustment based on the statistical characteristics of historical adjustment amplitudes. Regarding normalized distance calculation, other normalization methods can be used, such as min-max normalization, Z-score normalization, decimal scaling normalization, etc. Regarding mapping processing, the dual-channel exponential function can be replaced with the Sigmoid function, Tanh function, ReLU function, or other functions with asymptotic saturation characteristics; the basic adjustment coefficient can be replaced with adjustable parameters or dynamic values ​​of other statistics. Regarding sensitivity parameter adjustment, it can be replaced with moving average, exponential smoothing, moving standard deviation or other statistical methods; the length of the sliding window can be adjusted according to actual needs; the relationship between sensitivity parameter and variance can be replaced with linear relationship, logarithmic relationship or other functional relationship.

[0124] S5. The dominant defect score and the dominant compensation factor are negatively constrained and fused, and the minor defect score and the minor compensation factor are positively compensated and fused. The first calibrated score and the second calibrated score are output, and the fusion weight is determined by the first calibrated score and the second calibrated score.

[0125] The negative constraint fusion specifically involves: subtracting the dominant compensation factor from the dominant defect score to obtain a first temporary value; comparing the first temporary value with the lower boundary: if the intermediate value is less than the lower boundary, the lower boundary is directly used as the calibrated dominant defect score; if the first temporary value is greater than or equal to the lower boundary, the first temporary value is retained as the calibrated dominant defect score.

[0126] The positive compensation fusion specifically involves: adding a secondary compensation factor to the secondary defect score to obtain a second temporary value; comparing the second temporary value with the upper boundary: if the intermediate value is greater than the upper boundary, the upper boundary is directly used as the calibrated secondary defect score; if the second temporary value is less than or equal to the upper boundary, the second temporary value is retained as the calibrated secondary defect score.

[0127] If the original first score is the dominant defect score, then the calibrated first score is equal to the value after negative constraint fusion, and the calibrated second score is equal to the value after positive compensation fusion; conversely, if the original second score is the dominant defect score, then the calibrated first score is equal to the value after positive compensation fusion, and the calibrated second score is equal to the value after negative constraint fusion.

[0128] The fusion weight is determined using the first calibrated score and the second calibrated score as follows:

[0129] The initial fusion weight is calculated based on the smoothed received signal strength indication value, the first strength threshold, and the second strength threshold. The initial fusion weight is equal to the second strength threshold minus the smoothed received signal strength indication value, divided by the difference between the second strength threshold and the first strength threshold.

[0130] A first correction coefficient and a second correction coefficient are preset, and a weight correction amount is calculated based on the first score and the second score after calibration. The weight correction amount is positively correlated with the first correction coefficient and the first score after calibration, and negatively correlated with the second correction coefficient and the second score after calibration.

[0131] The initial fusion weight is added to the weight correction amount, and the addition result is restricted within a preset weight range to obtain the final fusion weight; wherein, the weight of the first enhanced signal is the fusion weight, and the weight of the second enhanced signal is one minus the fusion weight.

[0132] In order to make full use of the prior information contained in the received signal strength indication value, this step first calculates an initial fusion weight, which is determined by the smoothed received signal strength indication value, the first strength threshold, and the second strength threshold.

[0133] The difference between the second strength threshold and the smoothed received signal strength indication value reflects the distance between the current signal strength and the strong signal; the weaker the signal, the larger the numerator. The signal strength range of the difference between the second strength threshold and the first strength threshold converts the smoothed received signal strength indication value into an initial fusion weight value in the range of 0 to 1.

[0134] The initial fusion weight indicates that the closer the signal strength is to the low threshold, the closer the initial weight is to 1, indicating that more reliance should be placed on the first enhanced signal with noise reduction priority; the closer the signal strength is to the high threshold, the closer the initial weight is to 0, indicating that more reliance should be placed on the second enhanced signal with fidelity priority.

[0135] While the initial fusion weights can make a basic bias based on the received signal strength, they do not consider the actual residual noise and distortion in the denoised signal. This method introduces a weight correction amount, which is dynamically determined by the first and second calibrated scores, specifically:

[0136] ;

[0137] The first correction factor is a preset positive number used to control the weight of the first score after calibration on the correction amount. The larger the first correction factor, the more significant the positive contribution of noise residue to the correction amount.

[0138] The second correction coefficient is a preset positive number used to control the weight of the second score after calibration on the correction amount. The larger the second correction coefficient is, the more significant the negative contribution of the distortion degree to the correction amount.

[0139] The first and second correction factors can be the same, or different values ​​can be set according to the sensitivity requirements to noise or distortion, with a preferred range of 0.1 to 0.5.

[0140] and The first and second scores after calibration represent the degree of noise residue and distortion after negative / positive fusion, respectively, with values ​​ranging from 0 to 1.

[0141] With this correction, when noise residue is severe, the correction amount is positive, making the final fusion weight greater than the initial weight, thus favoring the first enhanced signal with noise reduction priority; when distortion is severe, the correction amount is negative, making the final fusion weight less than the initial weight, thus favoring the second enhanced signal with fidelity priority.

[0142] Step S5 takes the dominant defect score, secondary defect score, dominant compensation factor, and secondary compensation factor output from steps S3 and S4, as well as the smoothed received signal strength indication value obtained in step S2. It is responsible for generating the calibrated first and second scores and ultimately determining the fusion weights used for weighted fusion. Specifically, firstly, the dominant and secondary defect scores are actually calibrated using negative constraint fusion and positive compensation fusion to obtain the calibrated first and second scores. Then, the initial fusion weights are calculated using the smoothed received signal strength indication value. Next, the weight correction amount is dynamically calculated based on the calibrated scores, and the initial weights are added to the correction amount. Finally, the sum is limited to obtain the final fusion weight. This weight will be directly used for the weighted fusion in step S6, determining the ratio of the first enhanced signal and the second enhanced signal in the final output.

[0143] This step employs a negative constraint fusion mechanism using direct subtraction to effectively control the negative impact of dominant defect scores on system performance. This simple subtraction operation also reduces computational complexity and improves processing efficiency. The positive compensation fusion mechanism uses direct addition to effectively compensate for the impact of secondary defect scores on system performance, thus improving overall system performance. The initial fusion weight calculation mechanism based on signal strength indicators dynamically adjusts the initial values ​​of the fusion weights according to the current signal environment quality, achieving adaptive influence of signal environment quality on the fusion weights. The weight correction calculation mechanism dynamically corrects the fusion weights based on the actual situation of the calibrated first and second scores, ensuring that the fusion weights are adaptively adjusted according to the actual scoring situation. The weight constraint mechanism prevents the fusion weights from exceeding the effective range, thus guaranteeing system stability and reliability. Furthermore, the weight allocation mechanism for the first and second enhanced signals enables reasonable weighted fusion of different enhanced signals, thereby optimizing the overall system performance.

[0144] In practical implementation, for negative constraint fusion, other negative constraint algorithms can be used, such as subtracting the dominant compensation factor from the dominant defect score and dividing by (1 + dominant compensation factor), or multiplying the dominant defect score by (1 - dominant compensation factor). For positive compensation fusion, other positive compensation algorithms can be used, such as multiplying the minor defect score by (1 + minor compensation factor), or adding the minor compensation factor to the minor defect score and dividing by (1 - minor compensation factor). Regarding the initial fusion weight calculation, the formula for calculating the initial fusion weight can be replaced with other mapping functions, such as those based on the Sigmoid function, Tanh function, or other nonlinear mapping functions. Regarding the weight correction calculation, the weight correction can also be in a product form or based on a rule-based table lookup. Regarding weight constraints, the preset weight range can be replaced with other valid ranges; the weight constraint mechanism can be replaced with other constraint methods, such as those based on soft constraint functions or exponential decay functions. Regarding weight allocation, it can be replaced with normalization-based processing or proportional allocation.

[0145] S6. The first enhanced signal and the second enhanced signal are weighted and fused according to the fusion weight to obtain the final enhanced signal.

[0146] Obtain the final fusion weights calculated in step S5, where the fusion weights represent the proportion of the first enhanced signal in the final output signal. The weighted fusion specifically involves:

[0147] The first enhanced signal is multiplied by its corresponding weight, i.e., the fusion weight. The second enhanced signal is multiplied by its corresponding weight, i.e., 1 minus the fusion weight. Then, the two weighted signals are added together to obtain the final enhanced signal. Here, the fusion weight is the final fusion weight output in step S5, the first enhanced signal is the result of the first enhanced signal processing in step S5, and the second enhanced signal is the result of the second enhanced signal processing in step S5.

[0148] When the fusion weight is close to 1, the final signal is mainly composed of the first enhanced signal with noise reduction priority, so the noise suppression effect is outstanding; when the fusion weight is close to 0, the final signal is mainly composed of the second enhanced signal with fidelity priority, so the details are better preserved; when the fusion weight is around 0.5, the two contribute equally.

[0149] Step S6 is the final output stage of the entire noise reduction and enhancement method. It takes the final fusion weight calculated in step S5 and applies it to the first and second enhanced signals generated in step S1. Specifically, this step merges the two enhanced signals into a single output signal through a weighted summation point-by-point. The weight of the first enhanced signal is the fusion weight, and the weight of the second enhanced signal is one minus the fusion weight. Therefore, this step realizes the transformation from two candidate enhanced signals to the final enhanced signal, allowing the results of all previous steps to be combined into the final output. This step is the data output end of the entire method, and its output signal can be directly used for subsequent demodulation, decoding, or signal analysis.

[0150] The weighted fusion mechanism can dynamically adjust the contribution ratio of the first and second enhanced signals in the final output according to the magnitude of the fusion weights, thereby achieving adaptive weighted processing of different enhanced signals. By ensuring that the sum of the weights of the first and second enhanced signals is 1, the weighted fusion mechanism can ensure that the amplitude of the final enhanced signal remains within a reasonable range, avoiding the problem of excessively large or small signal amplitudes, thus improving the stability of the output signal. The calculation of weighted fusion only involves multiplication and addition operations, with extremely low computational load, so it does not introduce perceptible processing delays and fully meets the requirements of real-time communication systems. Since the first and second enhanced signals come from the dual branches of the same generator and are aligned by deep features, the signal after weighted fusion will not produce phase or timing misalignment, ensuring the integrity of the output signal. The fusion weights are strictly limited to the [0,1] interval, ensuring that the fusion result is always within the reasonable value range of the two branch signals, and there will be no abnormal amplitude or saturation distortion.

[0151] In practice, weighted fusion can also be performed in the frequency domain, that is, the spectra of the two enhanced signals are multiplied by their respective weights and then superimposed, and then an inverse Fourier transform is performed to obtain the time domain signal; or an adaptive filtering method can be used to combine the two signals as filter coefficients; or a time smoothing mechanism can be introduced to perform low-pass filtering on the fusion weights calculated frame by frame to avoid auditory or demodulation noise caused by sudden changes in weights between frames.

[0152] The technical scope of this method is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this method, and all such modifications and variations should fall within the protection scope of this method.

Claims

1. A method for denoising and enhancing communication signals based on generative adversarial networks, characterized in that, Specifically, it includes: S1. Acquire noisy signal data, input the noisy signal data into the generator, and output the first enhanced signal and the second enhanced signal; S2. Input the first enhanced signal and the second enhanced signal into the first discriminator and the second discriminator respectively to obtain a first score characterizing the degree of noise residue and a second score characterizing the degree of signal distortion. S3. Calculate the absolute value of the difference between the first score and the second score. If the absolute value of the difference exceeds a preset threshold, the larger of the first score and the second score is defined as the dominant defect score, and the smaller one is defined as the minor defect score. S4. Based on historical defect scoring data, the upper and lower boundaries of the defect scores are obtained, and the basic adjustment coefficients are mapped based on the normalized distances between the dominant defect score and the upper boundary, and between the secondary defect score and the lower boundary, respectively, to obtain the dominant compensation factor and the secondary compensation factor, wherein the basic adjustment coefficient is the difference between the absolute value of the difference and the preset threshold. S5. The dominant defect score and the dominant compensation factor are negatively constrained and fused, and the minor defect score and the minor compensation factor are positively compensated and fused. The first score and the second score after calibration are output, and the fusion weight is determined by the first score and the second score after calibration. S6. The first enhanced signal and the second enhanced signal are weighted and fused according to the fusion weight to obtain the final enhanced signal.

2. The communication signal noise reduction and enhancement method based on generative adversarial networks according to claim 1, characterized in that, Step S1 also includes: Obtain the received signal strength indication value corresponding to the noisy signal data; perform sliding median filtering on a predetermined number of consecutive received signal strength indication values ​​to obtain a smoothed received signal strength indication value; compare the smoothed received signal strength indication value with a preset first strength threshold and a second strength threshold. If the smoothed received signal strength indication value is lower than the first strength threshold, the first enhanced signal is directly used as the final enhanced signal. If the smoothed received signal strength indication value is higher than the second strength threshold, then the second enhanced signal is directly used as the final enhanced signal; If the smoothed received signal strength indication value is between the first strength threshold and the second strength threshold, then steps S2 to S6 are executed to determine the final enhanced signal.

3. The communication signal noise reduction and enhancement method based on generative adversarial networks according to claim 1, characterized in that: The generator adopts an encoder-dual decoder architecture, specifically: The noisy signal data is encoded into a high-dimensional feature representation by a shared encoder, and the high-dimensional feature representation is simultaneously input into the first decoder and the second decoder. During the training phase, the first decoder optimizes by minimizing the mean square error and outputs a first enhanced signal with priority given to noise suppression. During the training phase, the second decoder optimizes by maximizing structural similarity and outputs a second enhanced signal that prioritizes detail preservation. During training, depth-constrained feature alignment is applied to the first decoder and the second decoder, and the difference between the two feature representations is constrained by the feature synchronization loss function.

4. The communication signal noise reduction and enhancement method based on generative adversarial networks according to claim 1, characterized in that: The first discriminator and the second discriminator are two completely independent discriminator networks, specifically: The first discriminator is a noise residual detector. It takes the first enhanced signal as input, extracts the noise-related features of the signal layer by layer through a multi-layer convolutional network, and maps them to a first score. The first score is monotonically positively correlated with the degree of noise residual in the first enhanced signal. The second discriminator is a signal distortion detector. It takes the second enhanced signal as input, extracts the waveform distortion-related features of the signal layer by layer through a multi-layer convolutional network, and maps them to a second score. The second score is monotonically positively correlated with the degree of signal distortion in the second enhanced signal.

5. The communication signal noise reduction and enhancement method based on generative adversarial networks according to claim 1, characterized in that: After calculating the absolute value of the difference between the first score and the second score, the method further includes: The absolute value of the difference is compared with a preset threshold. If the absolute value of the difference does not exceed the preset threshold, the first score is directly used as the first score after calibration, and the second score is used as the second score after calibration.

6. The communication signal noise reduction and enhancement method based on generative adversarial networks according to claim 1, characterized in that: The upper and lower boundaries of the defect score obtained based on historical defect scoring data include: Construct a time-series queue of historical defect scoring data, wherein the time-series queue stores the first and second scores obtained from each signal data processing in chronological order; Each first score and each second score in the time-series queue is assigned a weight that decays exponentially over time; Obtain the weighted historical defect score data, and use probability density estimation to determine its high quantile and low quantile, which are respectively used as the upper boundary and the lower boundary; the value range of the high quantile is 90% to 100%, and the value range of the low quantile is 1% to 10%.

7. The communication signal noise reduction and enhancement method based on generative adversarial networks according to claim 6, characterized in that: When the first score or the second score exceeds the upper boundary or the lower boundary, a boundary adaptive adjustment mechanism is triggered, specifically as follows: If the first score or the second score exceeds the upper boundary, the current upper boundary is adjusted upward according to the dynamic learning rate; if the first score or the second score exceeds the lower boundary, the current lower boundary is adjusted downward according to the dynamic learning rate; each adjustment does not exceed a predetermined proportion of the current boundary value. The dynamic learning rate is negatively correlated with the number of consecutive frames that have not been exceeded from the current boundary.

8. The communication signal noise reduction and enhancement method based on generative adversarial networks according to claim 1, characterized in that: The normalized distance includes a first normalized distance and a second normalized distance, specifically: The first normalized distance is equal to the upper boundary minus the dominant defect score, divided by the difference between the upper boundary and the lower boundary. The second normalized distance is equal to the minor defect score minus the lower boundary, divided by the difference between the upper boundary and the lower boundary.

9. The communication signal noise reduction and enhancement method based on generative adversarial networks according to claim 1, characterized in that: The mapping process is implemented using a dual-channel exponential function with asymptotic saturation characteristics, specifically as follows: The dominant compensation factor is calculated by a first exponential function, the output of which increases with the increase of the first normalized distance, and the first exponential function has an adjustable first sensitivity parameter. The secondary compensation factor is calculated by a second exponential function whose output increases as 1 minus the second normalized distance increases. The second exponential function has an adjustable second sensitivity parameter. The first sensitivity parameter and the second sensitivity parameter are dynamically adjusted based on the historical statistical variances of the first normalized distance and the second normalized distance, respectively, as follows: Maintain a first sliding window and a second sliding window, respectively store the first normalized distance and the second normalized distance of a predetermined number of consecutive signal frames, and respectively calculate the first variance in the first sliding window and the second variance in the second sliding window; The first sensitivity parameter is positively correlated with the first variance, the second sensitivity parameter is positively correlated with the second variance, and the rate at which the first sensitivity parameter changes with the first variance is greater than the rate at which the second sensitivity parameter changes with the second variance.

10. The communication signal noise reduction and enhancement method based on generative adversarial networks according to claim 2, characterized in that: The fusion weight is determined using the first calibrated score and the second calibrated score as follows: The initial fusion weight is calculated based on the smoothed received signal strength indication value, the first strength threshold, and the second strength threshold. The initial fusion weight is equal to the second strength threshold minus the smoothed received signal strength indication value, divided by the difference between the second strength threshold and the first strength threshold. A first correction coefficient and a second correction coefficient are preset, and a weight correction amount is calculated based on the first score and the second score after calibration. The weight correction amount is positively correlated with the first correction coefficient and the first score after calibration, and negatively correlated with the second correction coefficient and the second score after calibration. The initial fusion weight is added to the weight correction amount, and the addition result is restricted within a preset weight range to obtain the final fusion weight; wherein, the weight of the first enhanced signal is the fusion weight, and the weight of the second enhanced signal is one minus the fusion weight.

Images