Communication signal enhancement method and device based on gradient feedback closed loop, equipment and medium

Abstract

The invention provides a communication signal enhancement method and device based on a gradient feedback closed loop, equipment and a medium, and the method comprises the steps: processing a target I / Q signal, and outputting a clean baseline signal; inputting the statistical feature vector of the baseline signal into a mixed strategy pool to obtain a target enhancement strategy; dynamically adjusting control parameters of a plurality of differentiable physical effect modules integrated in the physical differentiable generator based on the target enhancement strategy, and applying physically differentiable channel disturbance to the baseline signal to generate an initial enhancement signal; and based on the comprehensive credibility score of each initial enhanced signal, screening out a high-credibility enhanced signal, and based on a joint enhancement module, carrying out joint enhancement processing on the high-credibility enhanced signal to determine a final enhanced signal. The problems of enhancement splitting, generation distortion, screening missing, learning disjunction and the like in the prior art are effectively solved, and the efficiency and accuracy of communication signal enhancement oriented to low signal-to-noise ratio and small sample scenes are improved.

Description

Technical Field

[0001] This application relates to the field of communication signal processing technology, and in particular to communication signal enhancement methods, apparatus, devices and media based on gradient feedback closed loop. Background Technology

[0002] As wireless communication systems evolve towards higher frequencies, higher dynamic ranges, and lower signal-to-noise ratios, electromagnetic signal sensing tasks (such as modulation recognition, spectrum sensing, and radar target classification) face increasingly severe challenges from small sample sizes and strong interference. In real-world scenarios, due to high annotation costs and limited acquisition environments, the number of high-quality signal samples available for training is often extremely limited. Especially under complex channel conditions, models are prone to overfitting and poor generalization. Therefore, effectively improving the robustness and discriminative power of signal feature representation under limited sample conditions has become a key technical challenge in the field of intelligent signal processing.

[0003] To address the aforementioned issues, existing technologies primarily employ data augmentation strategies to expand the training set size. Typical methods include: 1. Simulation augmentation methods based on fixed physical models: These methods apply perturbations to the original signal by constructing AWGN channels, multipath fading models, or ITU standard channel templates to generate diverse samples. However, these methods rely on preset parameter scripts, lack adaptive adjustment capabilities, and cannot dynamically optimize the perturbation type and intensity according to downstream task requirements. Furthermore, their generation process is non-differentiable, making it difficult to participate in end-to-end joint training, resulting in a disconnect between the augmented sample distribution and the real task boundary. 2. Data augmentation methods based on feature space perturbations: These methods perform random translation, scaling, shearing, or noise superposition operations on the time-domain waveform or spectrum. While simple to implement and computationally inexpensive, these operations lack physical consistency constraints, and the generated samples often deviate from the statistical characteristics of the real signal (such as constellation diagram structure and power spectrum morphology), even introducing artifact features that mislead model learning. 3. Data generation methods based on generative adversarial networks (GANs) or diffusion models: These methods utilize deep generative models to learn the signal data manifold and synthesize new samples. While capable of generating signals with a degree of realism, the generation process is often decoupled from the task model, lacking a quality screening mechanism and prone to producing "pseudo-samples" with pattern collapse, category drift, or semantic distortion. Furthermore, traditional GANs do not embed physical law modeling, resulting in poor interpretability of the generated results and failing to meet the controllability requirements of communication systems for channel behavior. Therefore, there is an urgent need for a novel signal enhancement method that can achieve closed-loop optimization of the entire "generation-screening-enhancement-learning" chain while ensuring physical consistency, in order to solve the model generalization problem in low signal-to-noise ratio, small sample scenarios. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide a method, apparatus, device and medium for enhancing communication signals based on gradient feedback closed loop. By feature purification, hybrid strategy pool, physical differentiable generator, quality screening and joint enhancement, it effectively overcomes the problems of enhancement fragmentation, generation distortion, screening omission and learning disconnect in the prior art, and improves the efficiency and accuracy of communication signal enhancement for low signal-to-noise ratio and small sample scenarios.

[0005] This application provides a communication signal enhancement method based on gradient feedback closed loop, the communication signal enhancement method comprising: The target I / Q signal is denoised and enhanced using a feature purifier to output a clean baseline signal. The statistical feature vector of the baseline signal is extracted and input into the hybrid policy pool. The target enhancement policy that matches the baseline signal is selected from multiple reference enhancement policies. The reference enhancement policy is obtained by gradient backpropagation dynamic optimization through the distillation framework of the teacher model and the student model. The baseline signal and the target enhancement strategy are input into the physical differentiable generator. Based on the target enhancement strategy, the control parameters of multiple differentiable physical effect modules integrated inside the physical differentiable generator are dynamically adjusted, and a physically differentiable channel perturbation is applied to the baseline signal to generate an initial enhancement signal. High-confidence enhancement signals are selected based on the comprehensive confidence score of each initial enhancement signal, and the high-confidence enhancement signals are then subjected to joint enhancement processing based on the joint enhancement module to determine the final enhancement signal.

[0006] In one possible implementation, the reference enhancement strategy is determined through the following steps: The original I / Q signal is processed by the feature purifier to output a clean sample baseline signal. The sample statistical feature vector of the sample baseline signal is input into the mixing strategy pool to output the initial sample enhancement strategy. The sample baseline signal and the initial sample enhancement strategy are input into the physical differentiable generator to generate an initial sample enhancement signal. High-confidence sample enhancement signals are selected from multiple initial sample enhancement signals. The high-confidence sample enhancement signals are then subjected to joint enhancement processing to determine the final enhanced sample set. The final augmented sample set and the sample baseline signal are input into the distillation framework of the teacher model and the student model. Knowledge transfer is achieved through consistency loss constraints. Based on the task performance loss of the student model on the validation set, the gradient is calculated and backpropagated to the physical parameter space generated by the physical differentiable and the policy network parameter space of the hybrid policy pool. The parameters of both are updated synchronously, so that the generated reference augmentation policy gradually converges to the optimal direction for improving the robustness of the model.

[0007] In one possible implementation, the step of inputting the sample statistical feature vector of the sample baseline signal into the mixing strategy pool and outputting an initial sample augmentation strategy includes: The hybrid strategy pool dynamically determines the activation probability and effect strength of different physical disturbances based on the sample statistical feature vectors of the instantaneous signal-to-noise ratio, power spectral density shape, symbol rate, and envelope stability of the sample baseline signal, and outputs the sample augmentation strategy; wherein, the hybrid strategy pool is implemented using a neural network structure.

[0008] In one possible implementation, the step of dynamically adjusting the control parameters of multiple differentiable physical effect modules integrated within the physically differentiable generator based on the target enhancement strategy, and applying physically differentiable channel perturbations to the baseline signal to generate an initial enhanced signal includes: Based on the enhancement strategy weights and perturbation strength parameters of each physical effect module in the target enhancement strategy, the control parameters of the carrier frequency offset module, multipath convolution module, IQ imbalance module, power amplifier nonlinearity module, and phase noise module inside the physical differentiable generator are adjusted. According to the differentiable physical effect module after adjusting the control parameters, a physically differentiable channel perturbation is applied to the baseline signal to generate an initial enhancement signal.

[0009] In one possible implementation, the adjustment of control parameters for the carrier frequency offset module, multipath convolution module, IQ imbalance module, power amplifier nonlinearity module, and phase noise module within the physical differentiable generator based on the enhancement strategy weights and perturbation intensity parameters of each physical effect module in the target enhancement strategy includes: The multipath convolution module uses a finite impulse response filter with learnable impulse response coefficients to simulate the multipath propagation effect. The IQ unbalanced module uses an adjustable gain ratio and phase difference to construct a 2×2 linear transformation matrix; The power amplifier nonlinear module uses a polynomial approximation or activation function to simulate the amplifier compression characteristics. The phase noise module simulates time-varying phase jitter by introducing a Gaussian random process with controllable variance.

[0010] In one possible implementation, the step of selecting high-confidence enhancement signals based on the comprehensive confidence score of each of the initial enhancement signals, and performing joint enhancement processing on the high-confidence enhancement signals based on the joint enhancement module to determine the final enhancement signal includes: For each of the initial enhanced signals, the reconstruction similarity, adversarial confidence and task consistency indices are calculated, and the scores of each dimension are weighted and fused to generate a comprehensive credibility score; The high-confidence enhancement signals are selected based on the preset threshold and the comprehensive confidence score. Based on the joint enhancement module, a task-oriented learnable joint enhancement transformation is performed on the high-confidence enhancement signal to expand the diversity of data distribution in the time domain, frequency domain, and amplitude space, thereby obtaining the final enhanced signal.

[0011] In one possible implementation, the learnable joint enhancement transform includes the following: The transformation parameters of the learnable joint enhancement transform are generated by the hybrid strategy pool and are subject to physical rationality constraints. The transformation parameters are generated by the hybrid strategy pool and are subject to physical rationality constraints. The transformation parameters of the learnable joint enhancement transform are generated by the hybrid strategy pool and are subject to physical rationality constraints.

[0012] This application also provides a communication signal enhancement device based on gradient feedback closed loop, the communication signal enhancement device comprising: The feature extraction module is used to denoise and enhance the features of the target I / Q signal, and output a clean baseline signal. The hybrid strategy pool module is used to extract the statistical feature vector of the baseline signal, input the statistical feature vector into the hybrid strategy pool, and select the target enhancement strategy that matches the baseline signal from multiple reference enhancement strategies; wherein, the reference enhancement strategy is obtained by gradient backpropagation dynamic optimization through the distillation framework of the teacher model and the student model; The generator module is used to dynamically adjust the baseline signal and the target enhancement strategy into the physically differentiable generator, dynamically adjust the control parameters of multiple differentiable physical effect modules integrated inside the physically differentiable generator based on the target enhancement strategy, and apply physically differentiable channel perturbation to the baseline signal to generate an initial enhancement signal. The joint enhancement module is used to filter out high-confidence enhancement signals based on the comprehensive confidence score of each initial enhancement signal, and to perform joint enhancement processing on the high-confidence enhancement signals based on the joint enhancement module to determine the final enhancement signal.

[0013] This application also provides an electronic device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, the steps of the communication signal enhancement method based on gradient feedback closed loop described above are performed.

[0014] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the above-described gradient feedback closed-loop communication signal enhancement method.

[0015] This application provides a communication signal enhancement method, apparatus, device, and medium based on gradient feedback closed loop. The communication signal enhancement method includes: performing denoising and feature enhancement processing on a target I / Q signal based on a feature purifier to output a clean baseline signal; extracting the statistical feature vector of the baseline signal and inputting the statistical feature vector into a hybrid strategy pool to select a target enhancement strategy that conforms to the baseline signal from multiple reference enhancement strategies; wherein the reference enhancement strategy is obtained through gradient backpropagation dynamic optimization using a distillation framework of a teacher model and a student model; inputting the baseline signal and the target enhancement strategy into a physically differentiable generator, dynamically adjusting the control parameters of multiple differentiable physical effect modules integrated within the physically differentiable generator based on the target enhancement strategy, and applying physically differentiable channel perturbations to the baseline signal to generate an initial enhancement signal; selecting high-confidence enhancement signals based on the comprehensive confidence score of each initial enhancement signal, and performing joint enhancement processing on the high-confidence enhancement signals based on a joint enhancement module to determine the final enhancement signal. By employing feature purification, a hybrid policy pool, a physically differentiable generator, quality screening, and joint enhancement, this approach effectively overcomes existing problems such as enhancement fragmentation, generation distortion, missing screening, and learning disconnect, thereby improving the efficiency and accuracy of communication signal enhancement for low signal-to-noise ratio and small sample scenarios.

[0016] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a communication signal enhancement method based on gradient feedback closed loop provided in this application embodiment; Figure 2 A schematic diagram of a communication signal enhancement method based on gradient feedback closed loop provided in an embodiment of this application; Figure 3 A schematic diagram of a communication signal enhancement device based on gradient feedback closed loop provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. Based on the embodiments of this application, every other embodiment obtained by those skilled in the art without inventive effort falls within the scope of protection of this application.

[0020] First, the applicable scenarios for this application will be introduced. This application can be applied to the field of communication signal processing technology.

[0021] Research has revealed that existing technologies primarily employ data augmentation strategies to expand the training set size. Typical methods include: 1. Simulation augmentation methods based on fixed physical models: These methods apply perturbations to the original signal by constructing AWGN channels, multipath fading models, or ITU standard channel templates to generate diverse samples. However, these methods rely on preset parameter scripts, lack adaptive adjustment capabilities, and cannot dynamically optimize the perturbation type and intensity according to downstream task requirements. Furthermore, their generation process is non-differentiable, making it difficult to participate in end-to-end joint training, resulting in a disconnect between the augmented sample distribution and the real task boundary. 2. Data augmentation methods based on feature space perturbations: These methods perform random translation, scaling, shearing, or noise superposition operations on time-domain waveforms or spectrograms. While simple to implement and computationally inexpensive, these operations lack physical consistency constraints, and the generated samples often deviate from the statistical characteristics of the real signal (such as constellation diagram structure and power spectrum morphology), even introducing artifact features that mislead model learning. 3. Data generation methods based on generative adversarial networks (GANs) or diffusion models: These methods utilize deep generative models to learn the signal data manifold and synthesize new samples. While capable of generating signals with a degree of realism, traditional GANs often decouple the generation process from the task model, lacking a quality screening mechanism and prone to producing "pseudo-samples" exhibiting pattern collapse, category drift, or semantic distortion. Furthermore, traditional GANs do not embed physical law modeling, resulting in poor interpretability of the generated results and failing to meet the controllability requirements of communication systems for channel behavior. Therefore, a novel signal enhancement method is urgently needed that can achieve closed-loop optimization across the entire "generation-screening-enhancement-learning" chain while maintaining physical consistency, in order to address the model generalization challenge in low signal-to-noise ratio, small sample scenarios.

[0022] Based on this, the embodiments of this application provide a communication signal enhancement method based on gradient feedback closed loop. Through feature purification, hybrid policy pool, physical differentiable generator, quality screening and joint enhancement, it effectively overcomes the problems of enhancement fragmentation, generation distortion, screening omission and learning disconnect in the prior art, and improves the efficiency and accuracy of communication signal enhancement for low signal-to-noise ratio and small sample scenarios.

[0023] Please see Figure 1 , Figure 1 This is a flowchart illustrating a communication signal enhancement method based on gradient feedback closed loop, provided as an embodiment of this application. Figure 1 As shown in the figure, the communication signal enhancement method provided in this application embodiment includes: S101: Performs denoising and feature enhancement processing on the target I / Q signal to output a clean baseline signal.

[0024] In this step, a feature purifier is used to preprocess the target I / Q signal. Its core function is to remove noise and anomalous components while retaining effective features useful for subsequent identification and enhancement tasks. This can form a stable and reliable "baseline signal" before enhancement, thereby avoiding unreasonable or excessively distorted perturbations learned by the physically differentiable generator. Here, the feature purifier can employ a one-dimensional convolutional autoencoder, a time-frequency domain U-Net, or a lightweight residual network structure. Its input is the original signal sequence x, and its output is the purified signal x_clean. During training, the following loss combinations can be used: 1. Reconstruction Loss (MSE): ensuring the output signal is close to the input signal in overall waveform. 2. Contrastive Learning Loss: enhancing the model's ability to discriminate useful features by constructing positive / negative sample pairs. 3. Perceptual Loss: constraining the purification result in the time-frequency domain feature space, making it more consistent with the real distribution at the signal structure level. Through the above design, the feature purifier of this application can not only effectively filter out AWGN, simulation noise, and sampling anomalies, but also maintain the consistency of the signal in statistical features such as constellation diagrams and power spectra, becoming a reliable input for subsequent physically differentiable generators and enhancement strategies. Compared with existing technologies, the introduction of the feature purifier can significantly reduce the risk of unexplained perturbations in the generated samples and improve the stability and reliability of the entire enhancement pipeline.

[0025] S102: Extract the statistical feature vector of the baseline signal, input the statistical feature vector into the hybrid policy pool, and select the target enhancement policy that matches the baseline signal from multiple reference enhancement policies; wherein, the reference enhancement policy is obtained by gradient backpropagation dynamic optimization through the distillation framework of the teacher model and the student model.

[0026] In this step, the statistical feature vector of the baseline signal is extracted, and the statistical feature vector is input into the hybrid strategy pool to select the target enhancement strategy that matches the baseline signal from multiple reference enhancement strategies.

[0027] It should be noted that corresponding target enhancement strategies can be selected based on the type of target signal.

[0028] Here, statistical feature vectors include, but are not limited to: instantaneous signal-to-noise ratio, power spectral density shape, instantaneous frequency offset estimation, symbol rate or bandwidth index, and temporal envelope stability.

[0029] In one possible implementation, the reference enhancement strategy is determined through the following steps: A: The original I / Q signal is processed by the feature purifier to output a clean sample baseline signal. The sample statistical feature vector of the sample baseline signal is input into the mixing strategy pool to output the initial sample enhancement strategy.

[0030] The initial sample augmentation strategy is the initial combination of physical disturbances and parameter settings adopted by the MixPool controller before optimization through gradient feedback loop. It serves as the "starting point augmentation scheme" to initiate the entire cascade augmentation process.

[0031] In one possible implementation, the step of inputting the sample statistical feature vector of the sample baseline signal into the hybrid strategy pool and outputting an initial sample augmentation strategy includes: the hybrid strategy pool dynamically decides the activation probability and effect intensity of different physical disturbances based on the sample statistical feature vectors of the instantaneous signal-to-noise ratio, power spectral density shape, symbol rate, and envelope stability of the sample baseline signal and outputs the initial sample augmentation strategy; wherein, the hybrid strategy pool is implemented using a neural network structure.

[0032] To address the problem that existing enhancement strategies are fixed or random and cannot provide differentiated enhancement for different signal characteristics, this application employs a hybrid strategy pool located at the strategy decision layer of the enhancement process. This pool is used to select and combine different physical perturbation strategies. Its input is the statistical feature vector of the current signal. The output is the weights of the enhancement strategy. and disturbance intensity parameters The MixPool, used to control the combination and intensity of differentiable physical effect modules, can employ a learnable policy network (reinforcement learning agent) that dynamically outputs policy parameters based on input signal characteristics.

[0033] in, These are the internal network parameters for the hybrid policy pool.

[0034] B: Input the sample baseline signal and the initial sample enhancement strategy into the physical differentiable generator to generate an initial sample enhancement signal. Select a high-confidence sample enhancement signal from multiple initial sample enhancement signals, perform joint enhancement processing on the high-confidence sample enhancement signal, and determine the final enhanced sample set.

[0035] Here, the sample baseline signal and the initial sample enhancement strategy are input into the physical differentiable generator to generate the initial sample enhancement signal. High-confidence sample enhancement signals are selected from multiple initial sample enhancement signals. The high-confidence sample enhancement signals are then subjected to joint enhancement processing to determine the final enhanced sample set.

[0036] C: Input the final augmented sample set and the sample baseline signal into the distillation framework of the teacher model and the student model. Achieve knowledge transfer through consistency loss constraints. Based on the task performance loss of the student model on the validation set, calculate the gradient and backpropagate it to the physical parameter space generated by the physical differentiable and the policy network parameter space of the hybrid policy pool. Update the parameters of both synchronously so that the generated reference augmentation policy gradually converges to the optimal direction for improving the robustness of the model.

[0037] Here, the final augmented sample set processed by the joint augmentation module is input into the student model for training. Simultaneously, the corresponding sample baseline signal (i.e., the clean I / Q signal output by the feature purifier) ​​is used as a reference input to train the teacher model. In the early stages of training, the teacher model uses only the original or purified high-quality signals for supervised learning to ensure it possesses high-confidence semantic discriminative ability, and can be considered the "knowledge authority" for downstream tasks. Subsequently, a weighted consistency distillation loss function is constructed to constrain the output distribution of the student model on the augmented data to approximate the prediction results of the teacher model. Specifically, it is defined using the KL divergence form as follows:

[0038] in, For the teacher model, For student models, For the first i The credibility weight of the final enhanced sample is calculated by integrating the three indicators mentioned above: reconstruction similarity, adversarial confidence, and task consistency. This indicates that the teacher model is related to the sample baseline signal. The category probability distribution, The predicted distribution of the student model for the corresponding augmented samples. The loss term serves two purposes: firstly, it preserves the robustness gains from the enhancements, and secondly, it suppresses student model decision drift caused by generation bias, thereby achieving a balance between "physical diversity" and "semantic consistency".

[0039] Here, through multiple rounds of training, the parameters of the hybrid policy pool and the physically differentiable generator are iteratively updated. Updates stop when any of the following termination conditions are met: for example, the improvement in the student model's task performance metrics on the validation set is consistently below a preset threshold across several iterations; or the number of closed-loop iterations reaches a preset maximum limit. At this point, the reference augmentation policy set is considered to have converged in the feature space, and closed-loop training ends. The converged parameters of the hybrid policy pool and the physically differentiable generator are then fixed and used to generate the final augmentation dataset.

[0040] It should be noted that the network parameters of the teacher model are frozen during the closed-loop training phase.

[0041] In each training iteration, the system calculates the overall objective loss based on the student model's classification accuracy, F1 score, or other task performance metrics on the independent validation set. Then, the automatic differentiation engine backpropagates this loss to calculate its gradient with respect to each learnable parameter. The gradient not only updates the student model's own network parameters but also continues forward through the entire enhancement pipeline until it reaches: 1. the physical parameter space in the physically differentiable generator (G_phys), such as the frequency offset of the carrier frequency offset module, the impulse response coefficients of the multipath convolutional layer, and the gain ratio and phase difference of the IQ imbalance module; 2. the policy network parameter space of the MixPool, including the weights and biases of its internal neural networks, and the probability distribution parameters (such as the α / β values ​​of the Beta distribution) for controlling perturbation combinations. This embodiment employs the Gumbel-Softmax approximation or the Straight-Through Estimator (STE) technique to relax these operations, enabling the gradient to be effectively propagated to the front-end modules. Ultimately, by simultaneously updating the physical perturbation parameters of G_phys and the policy network parameters of MixPool using an optimizer (such as Adam or SGD with momentum), the system gradually adjusts its augmentation behavior. This includes reducing the intensity of strong nonlinear perturbations that cause semantic distortion, or enhancing the multipath simulation that facilitates generalization at low signal-to-noise ratios. After multiple rounds of training, the generated reference augmentation policy no longer depends on the initial prior settings but adaptively converges to a set of optimal parameter configurations that best improve the robustness and generalization performance of the student model, achieving a shift from "passive augmentation" to "active evolution."

[0042] S103: Input the baseline signal and the target enhancement strategy into the physical differentiable generator, dynamically adjust the control parameters of multiple differentiable physical effect modules integrated inside the physical differentiable generator based on the target enhancement strategy, and apply a physically differentiable channel perturbation to the baseline signal to generate an initial enhancement signal.

[0043] In this step, the baseline signal and the target enhancement strategy are input to the physical differentiable generator. The control parameters of the multiple differentiable physical effect modules integrated inside the physical differentiable generator are dynamically adjusted according to the target enhancement strategy. Physically differentiable channel perturbations are applied to the baseline signal to generate the initial enhancement signal.

[0044] In one possible implementation, the step of dynamically adjusting the control parameters of multiple differentiable physical effect modules integrated within the physically differentiable generator based on the target enhancement strategy, and applying physically differentiable channel perturbations to the baseline signal to generate an initial enhanced signal includes: Based on the enhancement strategy weights and perturbation strength parameters of each physical effect module in the target enhancement strategy, the control parameters of the carrier frequency offset module, multipath convolution module, IQ imbalance module, power amplifier nonlinearity module, and phase noise module inside the physical differentiable generator are adjusted. According to the differentiable physical effect module after adjusting the control parameters, a physically differentiable channel perturbation is applied to the baseline signal to generate an initial enhancement signal.

[0045] Here, a physically differentiable generator This is the core module in the entire enhancement pipeline, used to generate physically interpretable enhancement samples. It combines traditional Generative Adversarial Networks (GANs) with a differentiable physical channel model, ensuring that the generated perturbations conform to the physical laws of wireless communication (such as electromagnetic wave propagation and hardware non-ideals). This channel layer consists of several differentiable physical modules connected in series and parallel, each corresponding to a typical channel effect. These include, but are not limited to: carrier frequency offset module, multipath convolution module, IQ imbalance module, power amplifier nonlinearity module, and phase noise module. All physical modules are implemented as differentiable functions, and their parameter vectors... It can be initialized from physical priors (e.g., CFO range, path decay distribution) or automatically optimized during training via backpropagation gradients. When generating augmented samples, this channel layer is embedded into the generative model or augmentation process, processing the input baseline signal... Apply physical channel transformation:

[0046] Training objectives may include loss minimizing distribution gap (e.g., MMD, KL divergence) and task loss (e.g., classification cross-entropy), thereby achieving coordinated convergence of channel parameters and task performance.

[0047] In one possible implementation, the adjustment of control parameters for the carrier frequency offset module, multipath convolution module, IQ imbalance module, power amplifier nonlinearity module, and phase noise module within the physical differentiable generator based on the enhancement strategy weights and perturbation intensity parameters of each physical effect module in the target enhancement strategy includes: The multipath convolution module uses a finite impulse response filter with learnable impulse response coefficients to simulate multipath propagation effects; the IQ imbalance module uses adjustable gain ratio and phase difference to construct a 2×2 linear transformation matrix; the power amplifier nonlinear module uses polynomial approximation or activation function to simulate amplifier compression characteristics; and the phase noise module simulates time-varying phase jitter by introducing a Gaussian random process with controllable variance.

[0048] Traditional channel simulations typically use fixed-parameter scripts (such as ITU multipath models and AWGN simulations) or non-physical GAN ​​generators, which cannot adaptively adjust based on task feedback. The channel layer of this invention can participate in backpropagation, allowing channel parameters to dynamically adjust during training, thereby gradually approximating the task-favorable real distribution while maintaining physical plausibility. Through differentiable modeling and trainable parameters, the channel layer of this invention can: achieve physically interpretable augmentation generation; avoid class drift and spurious patterns in GAN-like models; reduce the distributional differences between generated samples and real signals in spectral and phase statistics; and make the augmentation sample distribution adaptive to the task boundary, thereby improving the model's generalization performance and robustness.

[0049] S104: Based on the comprehensive confidence score of each initial enhancement signal, high confidence enhancement signals are selected, and the high confidence enhancement signals are jointly enhanced based on the joint enhancement module to determine the final enhancement signal.

[0050] In this step, high-confidence enhancement signals are selected based on the overall confidence score of each initial enhancement signal, and the high-confidence enhancement signals are jointly enhanced by the joint enhancement module to determine the final enhancement signal. It should be noted that the pre-generated enhancement strategies are associated with and stored with the corresponding I / Q signals. If an enhancement strategy for the I / Q signal to be enhanced exists, the signal enhancement process can be performed directly according to the pre-generated enhancement strategy. If it does not exist, the corresponding enhancement strategy is obtained by dynamically optimizing the hybrid strategy pool and the physically differentiable generator through gradient backpropagation using the distillation framework of the teacher model and the student model.

[0051] In one possible implementation, the step of selecting high-confidence enhancement signals based on the comprehensive confidence score of each of the initial enhancement signals, and performing joint enhancement processing on the high-confidence enhancement signals based on the joint enhancement module to determine the final enhancement signal includes: a: For each of the initial enhanced signals, the reconstruction similarity, adversarial confidence and task consistency indices are calculated, and the scores of each dimension are weighted and fused to generate a comprehensive credibility score.

[0052] Here, although the samples generated by G_phys are physically reasonable, there may still be samples that deviate from the task semantics. A multi-dimensional credibility evaluation mechanism is used to select the optimal samples. Reconstruction similarity: PSNR or MSE measures the structural difference between the generated samples and the real samples; Adversarial confidence: the discriminator's truth score (reflecting the realism of the generated samples); Task consistency: the difference between the teacher model's predicted distributions of the generated samples and the real samples (KL divergence).

[0053] b: Based on the preset threshold and the comprehensive credibility score, the high credibility enhancement signal is selected.

[0054] c: Based on the joint enhancement module, a task-oriented learnable joint enhancement transformation is performed on the high-confidence enhancement signal to expand the diversity of data distribution in the time domain, frequency domain, and amplitude space, thereby obtaining the final enhanced signal.

[0055] Here, based on the joint enhancement module performing a task-oriented learnable joint enhancement transformation on the high-confidence enhancement signal, the diversity of data distribution in the time domain, frequency domain, and amplitude space is expanded to obtain the final enhanced signal.

[0056] In one possible implementation, the learnable joint enhancement transform includes the following: time-sequence flipping or cyclic shifting to enhance time invariance, amplitude scaling or signal-to-noise ratio modulation to simulate power fluctuations, frequency domain distortion or mirror transformation to expand spectral coverage, subband rearrangement or short-time Fourier coefficient perturbation to increase local frequency variations, and splicing or fusing multi-source signals to construct composite interference samples; wherein the transform parameters of the learnable joint enhancement transform are generated by the hybrid strategy pool and are subject to physical rationality constraints.

[0057] Here, the JointAug module further enhances the filtered data. It performs secondary enhancement on high-confidence samples that pass the initial screening, expanding data diversity. Unlike ordinary data augmentation, JointAug's parameters are learnable and task-dependent. Key operations include: time-sequence flipping / cyclic shifting to enhance time invariance; amplitude scaling / signal-to-noise modulation to simulate power fluctuations; frequency domain distortion / mirror transformation to expand the spectral distribution; and time-frequency mixing / stitching to construct multi-source composite samples. JointAug dynamically allocates enhancement strategies from the upstream MixPool controller, with the enhancement ratio and transformation amplitude output by the controller network, achieving task-adaptive enhancement. JointAug introduces "task-related diversity" on top of "physically reasonable" samples, forming a second learnable transformation layer that neither violates physical constraints nor hinders model generalization.

[0058] Traditional enhancement strategies (such as random pruning, flipping, and mixup) lack task adaptability and controllability in signal processing tasks. This application proposes a parameterized, learnable joint enhancement module that dynamically selects and combines enhancement methods (such as amplitude scaling, time-series looping, frequency domain distortion, and signal superposition) through a MixPool controller, and uses a Beta distribution to sample enhancement ratios. This means the enhancement strategy is no longer fixed but is automatically optimized by the controller based on task learning; furthermore, the joint transformation can simultaneously expand the time, frequency, and amplitude distribution of the signal, improving sample diversity; and the generated enhanced samples cover more channel distortion types, improving the model's generalization ability.

[0059] For further details, please refer to Figure 2 , Figure 2 This is a schematic diagram of a communication signal enhancement method based on gradient feedback closed loop provided in an embodiment of this application. Figure 2 As shown, the raw I / Q data is input to the feature purifier, which outputs a clean sample baseline signal. The hybrid policy pool provides enhancement policy control for the physically differentiable channel layer. This hybrid pool processes the sample baseline signal and outputs an initial sample enhancement policy. The sample baseline signal and the initial sample enhancement policy are input to the physically differentiable generator to generate an initial sample enhancement signal. The initial sample enhancement signal is evaluated and screened according to the teacher model to obtain a high-confidence sample enhancement signal. The high-confidence sample enhancement signal is processed by the joint enhancement module to determine the final enhanced sample set. The final enhanced sample set and the sample baseline signal are input into the distillation framework of the teacher model and the student model. The gradient is calculated and backpropagated to the physical parameter space of the physically differentiable generator and the policy network parameter space of the hybrid policy pool. This dynamically guides the optimization of the enhancement policy and the adjustment of the physical parameters, enabling the hybrid policy pool to optimize the physically differentiable generator and the joint enhancement processing, allowing the entire system to evolve and adaptively generate the most beneficial enhancement data for the task model.

[0060] Existing data augmentation methods are mostly independent, script-based operations (such as adding noise, time shifting, frequency offsetting, etc.). Augmentation strategies are fixed and unlearnable, and there is a lack of dependency and logical connection between different augmentations, resulting in poor physical consistency and task adaptability of augmented samples. This application designs the augmentation as a sequentially dependent, indivisible, learnable pipeline, including: a physically differentiable generator, sample credibility screening and joint augmentation transformation, and consistency distillation training. The output of each module is the input of the next module, forming a strongly coupled "generation-screening-transformation-distillation" closed loop, optimizing augmentation quality across the entire chain from the source to the task layer. This avoids the fragmentation between traditional augmentation modules; ensures the continuity of augmented samples in semantics, distribution, and physical attributes; and achieves dynamic learnability and end-to-end tuning of augmentation strategies.

[0061] This application provides a communication signal enhancement method based on gradient feedback closed loop. The method includes: denoising and enhancing a target I / Q signal using a feature purifier to output a clean baseline signal; extracting a statistical feature vector from the baseline signal and inputting the statistical feature vector into a hybrid policy pool to select a target enhancement policy that matches the baseline signal from multiple reference enhancement policies; wherein the reference enhancement policy is dynamically optimized through gradient backpropagation using a distillation framework of a teacher model and a student model; inputting the baseline signal and the target enhancement policy into a physically differentiable generator, dynamically adjusting the control parameters of multiple differentiable physical effect modules integrated within the physically differentiable generator based on the target enhancement policy, and applying physically differentiable channel perturbations to the baseline signal to generate an initial enhancement signal; selecting high-confidence enhancement signals based on the comprehensive confidence score of each initial enhancement signal, and performing joint enhancement processing on the high-confidence enhancement signals based on a joint enhancement module to determine the final enhancement signal. By employing feature purification, a hybrid policy pool, a physically differentiable generator, quality screening, and joint enhancement, this approach effectively overcomes existing problems such as enhancement fragmentation, generation distortion, missing screening, and learning disconnect, thereby improving the efficiency and accuracy of communication signal enhancement for low signal-to-noise ratio and small sample scenarios.

[0062] Please see Figure 3 , Figure 3 This is a schematic diagram of a communication signal enhancement device based on gradient feedback closed loop, provided as an embodiment of this application. Figure 3 As shown, the communication signal enhancement device 300 includes: The feature extraction module 310 is used to perform denoising and feature enhancement processing on the target I / Q signal based on the feature purifier, and output a clean baseline signal. The hybrid strategy pool module 320 is used to extract the statistical feature vector of the baseline signal, input the statistical feature vector into the hybrid strategy pool, and select the target enhancement strategy that matches the baseline signal from multiple reference enhancement strategies; wherein, the reference enhancement strategy is obtained by gradient backpropagation dynamic optimization through the distillation framework of the teacher model and the student model; The generator module 330 is used to dynamically adjust the baseline signal and the target enhancement strategy into the physical differentiable generator, dynamically adjust the control parameters of multiple differentiable physical effect modules integrated inside the physical differentiable generator based on the target enhancement strategy, and apply a physically differentiable channel perturbation to the baseline signal to generate an initial enhancement signal. The joint enhancement module 340 is used to filter out high-confidence enhancement signals based on the comprehensive confidence score of each initial enhancement signal, and to perform joint enhancement processing on the high-confidence enhancement signals based on the joint enhancement module to determine the final enhancement signal.

[0063] The hybrid policy pool module 320 determines the reference enhancement policy through the following steps: The original I / Q signal is processed by the feature purifier to output a clean sample baseline signal. The sample statistical feature vector of the sample baseline signal is input into the mixing strategy pool to output the initial sample enhancement strategy. The sample baseline signal and the initial sample enhancement strategy are input into the physical differentiable generator to generate an initial sample enhancement signal. High-confidence sample enhancement signals are selected from multiple initial sample enhancement signals. The high-confidence sample enhancement signals are then subjected to joint enhancement processing to determine the final enhanced sample set. The final augmented sample set and the sample baseline signal are input into the distillation framework of the teacher model and the student model. Knowledge transfer is achieved through consistency loss constraints. Based on the task performance loss of the student model on the validation set, the gradient is calculated and backpropagated to the physical parameter space generated by the physical differentiable and the policy network parameter space of the hybrid policy pool. The parameters of both are updated synchronously, so that the generated reference augmentation policy gradually converges to the optimal direction for improving the robustness of the model.

[0064] Furthermore, the hybrid strategy pool module 320 is used to input the sample statistical feature vector of the sample baseline signal into the hybrid strategy pool and output the initial sample augmentation strategy: The hybrid strategy pool dynamically determines the activation probability and effect strength of different physical disturbances based on the sample statistical feature vectors of the instantaneous signal-to-noise ratio, power spectral density shape, symbol rate, and envelope stability of the sample baseline signal, and outputs the initial sample enhancement strategy; wherein, the hybrid strategy pool is implemented using a neural network structure.

[0065] Furthermore, the generator module 330 is used to dynamically adjust the control parameters of multiple differentiable physical effect modules integrated within the physically differentiable generator based on the target enhancement strategy, and to apply physically differentiable channel perturbations to the baseline signal to generate an initial enhanced signal: Based on the enhancement strategy weights and perturbation strength parameters of each physical effect module in the target enhancement strategy, the control parameters of the carrier frequency offset module, multipath convolution module, IQ imbalance module, power amplifier nonlinearity module, and phase noise module inside the physical differentiable generator are adjusted. According to the differentiable physical effect module after adjusting the control parameters, a physically differentiable channel perturbation is applied to the baseline signal to generate an initial enhancement signal.

[0066] Furthermore, the generator module 330 is used to adjust the control parameters of the carrier frequency offset module, multipath convolution module, IQ imbalance module, power amplifier nonlinearity module, and phase noise module inside the physically differentiable generator based on the enhancement strategy weights and perturbation intensity parameters of each physical effect module in the target enhancement strategy. The multipath convolution module uses a finite impulse response filter with learnable impulse response coefficients to simulate the multipath propagation effect. The IQ unbalanced module uses an adjustable gain ratio and phase difference to construct a 2×2 linear transformation matrix; The power amplifier nonlinear module uses a polynomial approximation or activation function to simulate the amplifier compression characteristics. The phase noise module simulates time-varying phase jitter by introducing a Gaussian random process with controllable variance.

[0067] Furthermore, the joint enhancement module 340 is used to filter out high-confidence enhancement signals based on the comprehensive confidence score of each of the initial enhancement signals, and to perform joint enhancement processing on the high-confidence enhancement signals based on the joint enhancement module to determine the final enhancement signal: For each of the initial enhanced signals, the reconstruction similarity, adversarial confidence and task consistency indices are calculated, and the scores of each dimension are weighted and fused to generate a comprehensive credibility score; The high-confidence enhancement signals are selected based on the preset threshold and the comprehensive confidence score. Based on the joint enhancement module, a task-oriented learnable joint enhancement transformation is performed on the high-confidence enhancement signal to expand the diversity of data distribution in the time domain, frequency domain, and amplitude space, thereby obtaining the final enhanced signal.

[0068] This application provides a communication signal enhancement device based on gradient feedback closed loop. The communication signal enhancement device includes: a feature extraction module for denoising and enhancing a target I / Q signal to output a clean baseline signal; a hybrid strategy pool module for extracting statistical feature vectors of the baseline signal, inputting the statistical feature vectors into a hybrid strategy pool, and selecting a target enhancement strategy that conforms to the baseline signal from multiple reference enhancement strategies; wherein the reference enhancement strategies are dynamically optimized through gradient backpropagation using a distillation framework of a teacher model and a student model; a generator module for dynamically adjusting the baseline signal and the target enhancement strategy and inputting them into a physically differentiable generator, dynamically adjusting the control parameters of multiple differentiable physical effect modules integrated within the physically differentiable generator based on the target enhancement strategy, and applying physically differentiable channel perturbations to the baseline signal to generate an initial enhancement signal; and a joint enhancement module for selecting high-confidence enhancement signals based on the comprehensive confidence score of each initial enhancement signal, and performing joint enhancement processing on the high-confidence enhancement signals based on the joint enhancement module to determine the final enhancement signal.

[0069] Please see Figure 4 , Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 4 As shown, the electronic device 400 includes a processor 410, a memory 420, and a bus 430.

[0070] The memory 420 stores machine-readable instructions executable by the processor 410. When the electronic device 400 is running, the processor 410 communicates with the memory 420 via the bus 430. When the machine-readable instructions are executed by the processor 410, they can perform the operations described above. Figure 1 The steps of the communication signal enhancement method based on gradient feedback closed loop in the method embodiment shown are described in detail in the method embodiment, and will not be repeated here.

[0071] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, can perform the above-described actions. Figure 1 The steps of the communication signal enhancement method based on gradient feedback closed loop in the method embodiment shown are described in detail in the method embodiment, and will not be repeated here.

[0072] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0073] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the shown or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0074] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0075] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0076] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0077] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The scope of protection of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this application. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for enhancing a communication signal based on gradient feedback closed loop, characterized in that, The communication signal enhancement method comprises: The feature purifier is used for denoising and feature enhancement processing of the target I / Q signal, and a clean baseline signal is output; The statistical feature vector of the baseline signal is extracted, and the statistical feature vector is input into a mixed strategy pool to screen a target enhancement strategy that meets the baseline signal from a plurality of reference enhancement strategies; wherein the reference enhancement strategy is dynamically optimized by gradient back propagation through a distillation framework of a teacher model and a student model; The baseline signal and the target enhancement strategy are input into a physically differentiable generator, the control parameters of a plurality of differentiable physical effect modules integrated in the physically differentiable generator are dynamically adjusted based on the target enhancement strategy, physical-micro channel disturbance is applied to the baseline signal, and an initial enhanced signal is generated; A high-confidence enhanced signal is screened based on the comprehensive confidence score of each initial enhanced signal, and the high-confidence enhanced signal is processed by a joint enhancement module to determine a final enhanced signal.

2. The method of claim 1, wherein, The reference enhancement strategy is determined by the following steps: The feature purifier is used for processing the original I / Q signal to output a clean sample baseline signal, and the sample statistical feature vector of the sample baseline signal is input into the mixed strategy pool to output an initial sample enhancement strategy; The sample baseline signal and the initial sample enhancement strategy are input into the physically differentiable generator to generate an initial sample enhanced signal, a high-confidence sample enhanced signal is screened from a plurality of initial sample enhanced signals, and the high-confidence sample enhanced signal is processed by joint enhancement to determine a final enhanced sample set; The final enhanced sample set and the sample baseline signal are input into the distillation framework of the teacher model and the student model, knowledge transfer is realized through consistency loss constraint, the gradient is calculated based on the task performance loss of the student model on the validation set and is back propagated to the physical parameter space of the physically differentiable generator and the strategy network parameter space of the mixed strategy pool, and the parameters of the two are updated synchronously, so that the generated reference enhancement strategy gradually converges to the optimal direction of improving the robustness of the model.

3. The method of claim 2, wherein, The sample statistical feature vector of the sample baseline signal is input into the mixed strategy pool to output an initial sample enhancement strategy, comprising: The mixed strategy pool dynamically decides the enabling probability and action strength of different physical disturbances according to the sample statistical feature vector of the instantaneous signal-to-noise ratio, power spectral density shape, symbol rate and envelope stability of the sample baseline signal to output the initial sample enhancement strategy; wherein the mixed strategy pool is realized by a neural network structure.

4. The method for enhancing a communication signal according to claim 1, wherein, The control parameters of a plurality of differentiable physical effect modules integrated in the physically differentiable generator are dynamically adjusted based on the target enhancement strategy, and physical-micro channel disturbance is applied to the baseline signal to generate an initial enhanced signal, comprising: The control parameters of the carrier frequency offset module, the multipath convolution module, the IQ imbalance module, the power amplifier nonlinearity module, and the phase noise module inside the physically differentiable generator are adjusted based on the enhancement strategy weight and the disturbance intensity parameter of each physical effect module in the target enhancement strategy, and the baseline signal is subjected to physically differentiable channel disturbance by the differentiable physical effect module after the control parameters are adjusted, to generate an initial enhanced signal.

5. The method of claim 4, wherein, The adjustment of the control parameters of the carrier frequency offset module, the multipath convolution module, the IQ imbalance module, the power amplifier nonlinearity module, and the phase noise module inside the physically differentiable generator based on the enhancement strategy weight and the disturbance intensity parameter of each physical effect module in the target enhancement strategy comprises: The multipath convolution module uses a finite impulse response filter with learnable impulse response coefficients to simulate the multipath propagation effect; The IQ imbalance module uses a 2x2 linear transformation matrix with adjustable gain ratio and phase difference; The power amplifier nonlinearity module uses a polynomial approximation or an activation function to simulate the amplifier compression characteristics; The phase noise module simulates time-varying phase jitter by introducing a controllable variance Gaussian random process.

6. The method for enhancing a communication signal according to claim 1, wherein, The high-confidence enhanced signal is filtered based on the comprehensive confidence score of each initial enhanced signal, and the high-confidence enhanced signal is subjected to joint enhancement processing based on the joint enhancement module to determine the final enhanced signal, comprising: The comprehensive confidence score is generated by calculating the reconstruction similarity, the adversarial confidence, and the task consistency index of each initial enhanced signal, and by weighting and fusing the scores in each dimension; The high-confidence enhanced signal is filtered according to the preset threshold and the comprehensive confidence score; The task-oriented learnable joint enhancement transformation is performed on the high-confidence enhanced signal based on the joint enhancement module to expand the distribution diversity of data in the time domain, the frequency domain, and the amplitude space, and to obtain the final enhanced signal.

7. The method of claim 6, wherein, The learnable joint enhancement transformation includes the following: Time reversal or cyclic shift to enhance time invariance, amplitude scaling or signal-to-noise ratio modulation to simulate power fluctuations, frequency domain distortion or mirror transformation to expand spectral coverage, sub-band rearrangement or short-time Fourier coefficient disturbance to increase local frequency changes, and multi-source signal splicing or fusion to construct composite interference samples; wherein the transformation parameters of the learnable joint enhancement transformation are generated by the mixed strategy pool and are subject to physical rationality constraints.

8. A communication signal enhancement device based on gradient feedback closed loop, characterized in that, The communication signal enhancement device comprises: A feature extraction module for denoising and feature enhancement processing of a target I / Q signal, and outputting a clean baseline signal; A mixed strategy pool module for extracting a statistical feature vector of the baseline signal, inputting the statistical feature vector into a mixed strategy pool, and filtering a target enhancement strategy that meets the baseline signal from a plurality of reference enhancement strategies; wherein the reference enhancement strategy is dynamically optimized by gradient backpropagation through a distillation framework of a teacher model and a student model. A generator module is configured to input the baseline signal and the target enhancement strategy dynamic adjustment into a physically differentiable generator, dynamically adjust control parameters of a plurality of differentiable physical effect modules integrated inside the physically differentiable generator based on the target enhancement strategy dynamic adjustment, and apply a physically differentiable channel disturbance to the baseline signal to generate an initial enhanced signal. A joint enhancement module is configured to screen out a high-confidence enhanced signal based on a comprehensive confidence score of each of the initial enhanced signals, and determine a final enhanced signal based on joint enhancement processing of the high-confidence enhanced signal by the joint enhancement module.

9. An electronic device, comprising: The method comprises the following steps: A processor, a memory and a bus, the memory stores machine readable instructions executable by the processor, when the electronic device is running, the processor and the memory communicate through the bus, the machine readable instructions are executed by the processor to perform the steps of the communication signal enhancement method based on the gradient feedback closed loop as claimed in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, the computer program is executed by the processor to perform the steps of the communication signal enhancement method based on the gradient feedback closed loop as claimed in any one of claims 1 to 7.

Images