A communication signal adaptive processing method and system based on a model-driven sparse reconstruction network

Abstract

The application provides a communication signal adaptive processing method and system based on a model-driven sparse reconstruction network, which establishes a sparse frequency domain model of a communication signal, constructs a sensing matrix and a real value processing, extracts an observation feature and an environmental state, performs adaptive configuration selection, performs forward inference of a model-driven sparse reconstruction network, and outputs a sparse threshold reconstruction result. An interpretable model-driven expansion network with clear structure and clear physical meaning is constructed, and is deeply combined with an adaptive selection module based on a receiver observable feature for optimization. The system can intelligently and dynamically adjust the reconstruction strategy and network parameters of the signal according to the real-time acquired communication environment feature, so as to effectively cope with the changes of the channel condition, noise level and interference mode. The method can ensure that the overall calculation complexity is maintained at a low level, and significantly enhances the adaptability and robustness to the dynamic time-varying communication environment.

Description

Technical Field

[0001] This invention relates to the fields of wireless communication signal processing and artificial intelligence-assisted communication reception technology, and in particular to a method and system for adaptive processing of communication signals based on model-driven sparse reconstruction networks. Background Technology

[0002] In applications such as broadband communication, spectrum sensing, cognitive radio, sparse multipath channel estimation, and low sampling rate receivers, the communication signals to be processed often exhibit sparsity in a certain transform domain (such as the frequency domain). For example, in orthogonal frequency division multiplexing (OFDM) systems, the actual number of active subcarriers is usually much smaller than the total number of subcarriers in the system; in spectrum sensing, the occupied frequency points to be detected are usually only a small portion of the complete spectrum. Using compressed sensing theory, such sparse signals can be recovered using a small number of observations at a sampling rate far lower than the Nyquist rate, thereby significantly reducing system sampling costs and hardware complexity.

[0003] To address the above problems, existing technologies have mainly proposed the following types of sparse reconstruction methods, but all of them have inherent drawbacks: Greedy pursuit methods, such as the Orthogonal Matching Pursuit (OMP) algorithm, iteratively select the atoms most relevant to the residual to gradually approximate the sparse solution. Their advantages are simplicity and ease of understanding. However, their core drawback lies in their strong reliance on prior knowledge of signal sparsity, and the sharp decline in the correctness of atom selection under high-noise environments, leading to deterioration in reconstruction performance.

[0004] Convex optimization methods, such as basis pursuit (BP) or LASSO, solve the problem by relaxing the l0 norm to the l1 norm, exhibiting good global convergence and theoretical guarantees. However, their main drawback lies in their high computational complexity, especially when the signal dimension is large, limiting their real-time processing capabilities and making it difficult to meet the timeliness requirements of communication reception.

[0005] Iterative thresholding methods, such as the Iterative Soft Throttling Algorithm (ISTA) and the Fast Iterative Soft Throttling Algorithm (FISTA), approximate the optimal solution by iteratively performing gradient descent and soft thresholding operations. These methods are widely used due to their simple structure and controllable computational cost. However, their performance is highly dependent on the manual setting of hyperparameters such as the threshold parameter and step size. In dynamic communication environments where signal-to-noise ratio, compression ratio, or signal sparsity varies, fixed parameter configurations often cannot guarantee optimal reconstruction performance under different conditions, resulting in insufficient noise suppression at low signal-to-noise ratios and inadequate detail recovery at high signal-to-noise ratios.

[0006] Purely data-driven deep learning methods: In recent years, deep learning-based methods (such as the LISTA (LearningISTA) network) have been introduced to accelerate sparse reconstruction. These methods optimize network parameters end-to-end by learning from large amounts of data, achieving excellent performance within a finite number of iterations. However, their core drawback lies in their "black box" nature: the design of the network structure and the generation process of solutions lack physical interpretability; the model's generalization ability depends on the coverage of the training data; and when the training data is insufficient or inconsistent with the distribution of the test data, performance may severely degrade, or even generate solutions that violate physical laws.

[0007] In summary, existing technologies share the following common problems: Lacking environmental adaptability, the fixed parameter configuration of traditional algorithms cannot respond to dynamic changes in channel conditions (such as signal-to-noise ratio, compression ratio, and sparsity). Interpretability and performance are difficult to balance. Traditional methods are interpretable but have poor adaptability, while pure data-driven methods have excellent performance but lack physical constraints and generalization guarantees. There is a lack of a dedicated reconstruction framework for communication signals, which fails to effectively utilize the specific sparse structure inherent in communication signals in transform domains such as the frequency domain. Summary of the Invention

[0008] The purpose of this invention is to address the shortcomings of the prior art by providing a communication signal adaptive processing method and system based on model-driven sparse reconstruction networks, which significantly enhances the adaptability and robustness to dynamic time-varying communication environments, and greatly improves the signal reconstruction accuracy and overall system performance.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides an adaptive processing method for communication signals based on model-driven sparse reconstruction networks, comprising the following steps: S1. Establish a sparse frequency domain model for the communication signal: The sparse frequency domain model is expressed as follows: Frequency domain sparse signals: ,only Each non-zero frequency point carries modulation symbols Time-domain signal: ,in, unitary normalization inverse Fourier transform Compressed observation: ,in, For compressed sampling matrix, Additive white Gaussian noise S2. Constructing the perception matrix and performing real-valued processing; S3. Extract observation features and characterize environmental conditions; S4, Adaptive Configuration Selection; S5. Construct an adaptive model to drive the sparse reconstruction network and drive the forward inference of the sparse reconstruction network. S6. Output of sparse domain reconstruction results.

[0010] Furthermore, S2 specifically refers to: complex frequency domain signal vector By separating the real and imaginary parts, we obtain a real-valued sparse vector. ;in Let be a vector with real part. The vector is the imaginary part. This represents the range of values ​​for the virtual and real vectors.

[0011] Construct an equivalent real-valued sensing matrix such that the real-valued observation vector is ; This is the perception matrix.

[0012] Furthermore, S3 specifically refers to: The adaptive selection module utilizes the observation vectors obtained from the receiver. and perception matrix Extract several meta-features that can characterize the current communication environment state; The meta-features include: Compression ratio: ,in, To compress the observation dimensions; Energy proxy quantity: ; Noise proxy: ; Signal-to-noise ratio (SNR) throughput: ,in It is random noise.

[0013] Furthermore, S4 specifically includes: The adaptive selection module automatically selects a set of reconstruction parameters that match the current environment from a preset configuration library based on the meta-features extracted in S3. The configuration library includes the following schemas: Low signal-to-noise ratio robust mode: Suitable for low signal-to-noise ratio or high compression ratio environments, it uses more expansion layers and strong threshold strength to prioritize noise suppression and avoid the occurrence of spurious peaks; Balanced mode: Suitable for environments with medium signal-to-noise ratio and normal compression rate, it uses a moderate number of unfolding layers and threshold strength to achieve a balance between reconstruction accuracy and computational overhead; High-fidelity mode: Suitable for environments with high signal-to-noise ratio and low compression rate. It uses fewer unfolding layers and weaker threshold strength to preserve weak signal details to the maximum extent and improve reconstruction fidelity.

[0014] Furthermore, S5 specifically includes: The sparse reconstruction network is driven by a model MDSR Net, which is derived from an iterative soft thresholding algorithm. This model is rigorously expanded from the classic ISTA (Iterative Soft Thresholding) process. Each layer in the network corresponds to a specific step in mathematical optimization and has explicit physical interpretability.

[0015] The sparse reconstruction network structure has a clear physical meaning. The layer processing procedure is represented as follows: Residual update: ; Gradient calculation: ; Adaptive threshold generation: ; Soft threshold operation: ; in, The threshold control coefficient decreases progressively at each level; The noise intensity of the current residual estimate; These are the energy normalization weighting coefficients; A threshold factor adjusted based on the compression ratio; Step size; This is a soft thresholding function; The sparse reconstruction network performs iterative inference based on the number of layers and threshold strength selected in S4.

[0016] Furthermore, S6 specifically includes: After completing a specified number of iterations, the final real-valued sparse threshold signal estimate is output. And as needed, it is recombined into a complex frequency domain signal or the active frequency index and the modulation symbol it carries are decoded.

[0017] Furthermore, a communication signal adaptive processing system based on model-driven sparse reconstruction network, which is implemented using the aforementioned communication signal adaptive processing method based on model-driven sparse reconstruction network, also includes: a signal observation modeling module: establishing a sparse frequency domain model, converting complex signals to real values, and constructing an equivalent sensing matrix; Adaptive selection module: Extracts environmental features and intelligently selects the optimal reconstruction parameter configuration; Model-driven sparse reconstruction network: performs forward inference to complete sparse signal reconstruction.

[0018] The beneficial effects of this invention are as follows: By constructing a structurally clear and physically meaningful interpretable model to drive the unfolded network, and performing deep joint optimization with an adaptive selection module based on observable features at the receiver, the system can intelligently and dynamically adjust the signal reconstruction strategy and network parameters according to the real-time acquired communication environment characteristics, thereby effectively coping with changes in channel conditions, noise levels, and interference modes. Ultimately, this method significantly enhances the adaptability and robustness to dynamic time-varying communication environments while ensuring that the overall computational complexity remains at a low level, and greatly improves the signal reconstruction accuracy and overall system performance. Significantly improved environmental adaptability: By introducing an adaptive selection module based on receiver observation features, this method can automatically adjust the core parameters of the reconstructed network, such as the number of layers and threshold strength, according to the dynamic changes of environmental parameters such as signal-to-noise ratio and compression ratio, thus avoiding the performance instability problem of traditional fixed threshold algorithms under different environments.

[0019] Strong reconstruction robustness under low signal-to-noise ratio conditions: Benefiting from the adaptive high threshold and the configuration of more unfolded layers, this method can more effectively suppress noise interference in low signal-to-noise ratio environments. Simulation experiments show that, under typical parameter settings (M=64, K=8, QPSK), when SNR=-5dB, the normalized mean square error (NMSE) of this patented method is reduced by approximately 2.43 dB compared to the FISTA algorithm with a fixed threshold.

[0020] Balancing model interpretability with networked reasoning: The MDSR Net network used in this method is rigorously expanded from the classic ISTA iterative process. Each layer in the network corresponds to a specific step in mathematical optimization (residual calculation, gradient descent, soft thresholding). Therefore, it has the physical interpretability of traditional algorithms, while enjoying the advantages of fast forward inference speed and strong parallel computing capability of deep networks. The computational complexity is moderate. Easy to implement in engineering: The main computational load of MDSR Net is concentrated in matrix multiplication and soft thresholding operations. Its computational complexity is comparable to that of the ISTA algorithm with a fixed number of iterations, and far lower than that of convex optimization methods. Moreover, the number of unfolded layers can be adaptively selected according to actual needs, providing a flexible trade-off between performance and speed, and is easy to implement on hardware platforms such as FPGA or DSP.

[0021] Highly versatile and easily expandable: The sparse domain of this method is not limited to the frequency domain, but can be extended to the time delay domain, Doppler domain, etc.; the basic algorithm can also be extended from ISTA to FISTA, AMP, etc.; the modulation methods can cover BPSK, QPSK and even higher-order QAM, etc., with wide applicability and good scalability. Attached Figure Description

[0022] Figure 1 This is a flowchart of a communication signal adaptive processing method based on model-driven sparse reconstruction networks; Figure 2 For MDSR Net Layer calculation process diagram; Figure 3 A comparison curve of normalized mean square error; Figure 4 Example diagram for reconstructing a sparse signal in the frequency domain; Figure 5 Comparison of support recovery F1 scores under different signal-to-noise ratios; Figure 6 A comparison chart of the average running time of each reconstruction algorithm; Figure 7 Ablation experiment diagram to show the effect of layer number on reconstruction error. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0024] Please see Figure 1 An adaptive processing method for communication signals based on model-driven sparse reconstruction networks includes the following steps: S1. Establish a sparse frequency domain model for the communication signal: The sparse frequency domain model is expressed as follows: Frequency domain sparse signals: ,only Each non-zero frequency point carries modulation symbols Time-domain signal: ,in unitary normalization inverse Fourier transform Compressed observation: ,in, For compressed sampling matrix, Additive white Gaussian noise S2. Constructing the perception matrix and performing real-valued processing; S3. Extract observation features and characterize environmental conditions; S4, Adaptive Configuration Selection; S5. Construct an Adaptive Model-Driven Sparse Reconstruction Network (MDSR Net) to drive forward inference of the sparse reconstruction network; S6. Output of sparse domain reconstruction results.

[0025] Specifically, S2 is: complex frequency domain signal vector By separating the real and imaginary parts, we obtain a real-valued sparse vector. ;in Let be a vector with real part. The vector is the imaginary part. This represents the range of values ​​for the virtual and real vectors.

[0026] Construct an equivalent real-valued sensing matrix such that the real-valued observation vector is ; This is the perception matrix.

[0027] Furthermore, S3 specifically refers to: The adaptive selection module utilizes the observation vectors obtained from the receiver. and perception matrix Extract several meta-features that can characterize the current communication environment state; The meta-features include: Compression ratio: ,in, To compress the observation dimensions; Energy proxy quantity: ; Noise proxy: ; Signal-to-noise ratio (SNR) throughput: ,in It is random noise.

[0028] Specifically, S4 is: The adaptive selection module automatically selects a set of reconstruction parameters that match the current environment from a preset configuration library based on the meta-features extracted in S3. The configuration library includes the following schemas: Low signal-to-noise ratio robust mode: Suitable for low signal-to-noise ratio or high compression ratio environments, it uses more expansion layers and strong threshold strength to prioritize noise suppression and avoid the occurrence of spurious peaks; Balanced mode: Suitable for environments with medium signal-to-noise ratio and normal compression rate, it uses a moderate number of unfolding layers and threshold strength to achieve a balance between reconstruction accuracy and computational overhead; High-fidelity mode: Suitable for environments with high signal-to-noise ratio and low compression rate. It uses fewer unfolding layers and weaker threshold strength to preserve weak signal details to the maximum extent and improve reconstruction fidelity.

[0029] Specifically, S5 is: The sparse reconstruction network is driven by a model MDSR Net, which is derived from an iterative soft thresholding algorithm. This model is rigorously expanded from the classic ISTA (Iterative Soft Thresholding) process. Each layer in the network corresponds to a specific step in mathematical optimization and has explicit physical interpretability.

[0030] The sparse reconstruction network structure has a clear physical meaning. The layer processing procedure is represented as follows: Residual update: ; Gradient calculation: ; Adaptive threshold generation: ; Soft threshold operation: ; in, The threshold control coefficient decreases progressively at each level; The noise intensity of the current residual estimate; These are the energy normalization weighting coefficients; A threshold factor adjusted based on the compression ratio; Step size; This is a soft thresholding function; The sparse reconstruction network performs iterative inference based on the number of layers and threshold strength selected in S4.

[0031] like Figure 2 As shown, each layer of MDSRNet is not an arbitrarily stacked neural network module, but rather unfolds from one iteration of the iterative soft thresholding algorithm ISTA. For the ... The layer, whose input includes the sparse frequency domain estimate obtained from the previous layer. Compressed observation vector and perception matrix .

[0032] First, the network calculates the residuals based on the current estimates. The residual measures the degree of inconsistency between the current reconstruction result and the actual received observations in the observation domain. The larger the residual, the more necessary it is to revise the current estimate.

[0033] Subsequently, the network utilizes the transpose of the equivalent sensing matrix. The gradient direction is obtained by back-projecting the observation domain residuals into the sparse frequency domain. This gradient indicates the direction in which each candidate frequency point in the current frequency domain sparse vector should be enhanced or weakened.

[0034] To avoid noise and spurious peaks being mistaken for valid frequencies, the network further estimates the noise intensity based on the current residuals. And combined with the observed signal energy Compression ratio and threshold control coefficients that decrease layer by layer , generate the first Adaptive threshold of the layer .

[0035] Specifically, S6 is: After completing a specified number of iterations, the final real-valued sparse threshold signal estimate is output. And as needed, it is recombined into a complex frequency domain signal or the active frequency index and the modulation symbol it carries are decoded.

[0036] An adaptive communication signal processing system based on a model-driven sparse reconstruction network, implemented using the aforementioned adaptive communication signal processing method based on a model-driven sparse reconstruction network, further includes: Signal observation modeling module: Establish a sparse frequency domain model, convert complex signals to real values, and construct an equivalent sensing matrix; Adaptive selection module: Extracts environmental features and intelligently selects the optimal reconstruction parameter configuration; Model-driven sparse reconstruction network: performs forward inference to complete sparse signal reconstruction.

[0037] The inputs and outputs of the signal observation modeling module, the adaptive selection module, and the model-driven sparse reconstruction network are shown in Table 1: Table 1. Module Input and Output Table

[0038] The signal observation modeling module, adaptive selection module, and model-driven sparse reconstruction network present a serial data processing flow.

[0039] Example 1: In the simulation experiment of this embodiment, the specific system parameters used are set as follows: The frequency domain length is set to N = 128; The dimension of the real-valued sparse domain is D = 2N = 256.

[0040] The dimension of the compressed observation is M = 64.

[0041] The frequency domain sparsity is set to K = 8.

[0042] The signal-to-noise ratio (SNR) test range covers multiple typical values, including -5, 0, 5, 10, 15, and 20 dB.

[0043] The modulation method used is QPSK.

[0044] The algorithms used for comparison include: the OMP algorithm with a fixed threshold, the ISTA algorithm, the FISTA algorithm, and the adaptive MDSR Net method proposed in this patent.

[0045] The specific operational steps of the patented method in this embodiment can be broken down as follows: Step 1: Data Generation First, a sparse characteristic vector satisfying the frequency domain is randomly generated. Specifically: Eight frequency points are randomly selected and QPSK symbols are injected; then, the signal is transformed to the time domain using a unitarily normalized IFFT transform, and then a preset Gaussian random observation matrix is ​​used. Compressed observations are performed to obtain the observation vector. Finally, an equivalent real-valued sensing matrix is ​​constructed based on the system model. .

[0046] Step 2: Feature Extraction The adaptive selection module selects based on the received observation vectors. and perception matrix The calculation was performed strictly according to the formula given in S3, and four key feature quantities were extracted, namely compression ratio. Energy proxy of the observed signal Noise level proxy and signal-to-noise ratio proxy volume ; Step 3: Configuration Selection: This step, based on the feature values ​​extracted in the previous step, follows preset intelligent rules (e.g., when...). When the system detects a low signal-to-noise ratio (SNR) environment, it automatically selects a low SNR robust operating mode to dynamically choose the most suitable reconfiguration network configuration. In the simulation scenario of this embodiment, it is assumed that the environment is in a low SNR state, so the system automatically selects the low SNR robust mode for subsequent processing.

[0047] Step 4, Network Forward Inference: First, the initial estimate of the sparse signal... Set as the zero vector. Then, call the MDSR Net network corresponding to the low signal-to-noise ratio robust type and perform forward inference strictly according to the computation logic described in S5. In this mode, the number of unfolded layers of the network is preset to a large value (e.g., 20 layers), and the strength parameter of the soft threshold function is also set relatively high.

[0048] Step 5, Output the results: After the network completes the forward computation of all preset layers, it finally outputs an estimate of the frequency-domain sparse signal. and the corresponding support set.

[0049] Table 1 Experimental Results 1

[0050] Example 2: Specific application scenarios for low signal-to-noise ratio robust configurations: This embodiment is largely the same as Embodiment 1 in its basic framework. The core difference lies in that, in this embodiment, the adaptive selection module, upon explicitly detecting environmental characteristics of low signal-to-noise ratio or high compression rate, specifically enables a low signal-to-noise ratio robust configuration designed for harsh conditions. Under this configuration, the number of unfolded layers in the MDSR Net is fixed at 20, and the thresholds for all soft-threshold units in the network are... All are amplified by a uniform, high-valued base coefficient.

[0051] Under stringent simulation conditions of SNR=-5dB and M=64, the MDSRNet configured in this embodiment demonstrates a significant advantage in NMSE performance compared to the FISTA algorithm with a fixed threshold (whose threshold parameter is optimized at 0dB SNR). This results in a normalized mean square error (NMSE) performance gain of approximately 2.43dB, while simultaneously improving the F1 score of support recovery by approximately 0.04.

[0052] This result strongly validates that, in harsh high-noise environments, the mechanism of adaptively enhancing threshold strength can effectively identify and suppress noise components, thereby significantly reducing their interference with the signal support set detection process.

[0053] Depend on Figure 3 As can be seen, in the low signal-to-noise ratio (SNR) regions of -5dB and 0dB, the proposed adaptive MDSRNet achieves the lowest NMSE, and its reconstruction accuracy is significantly better than that of the FISA and OMP algorithms using fixed thresholds. When the SNR increases to above 5dB, the NMSE performance of each algorithm gradually becomes similar, while MDSRNet still maintains its competitiveness comparable to the strong baseline algorithm FISTA.

[0054] Figure 4 This presents a concrete reconstruction example with a signal-to-noise ratio of 0dB. The comparison shows that the spectrum reconstructed by MDSRNet best matches the true spectrum, and its reconstruction result also has the lowest noise floor.

[0055] Example 3: Specific applications of high-fidelity configurations: This embodiment is basically the same as Embodiment 1 in structure and process. The core difference is that the adaptive selection module automatically selects a high-fidelity configuration strategy after detecting and identifying the channel characteristics with high signal-to-noise ratio in real time. Under this configuration, the number of expansion iteration layers of the MDSR Net is set to a relatively small value (e.g., 10 layers), and the threshold parameter decays rapidly with iteration. The purpose is to preserve the subtle features of the signal to the maximum extent and avoid suppressing weak but critical frequency signals.

[0056] Table 2 Experimental Results 2

[0057] Experimental data shows that, under high signal-to-noise ratio conditions (e.g., SNR=15 dB or 20 dB), although the MDSR Net with this configuration slightly outperforms the manually tuned fixed-threshold FISTA benchmark algorithm in terms of NMSE, it exhibits superior detail fidelity and waveform preservation when handling complex scenarios with large signal dynamic range and rich detail. This fully demonstrates that the adaptive configuration framework provided by this patent is not a single strategy, but rather offers multiple optional optimization strategies to flexibly address the differentiated performance priorities (such as extreme accuracy or detail fidelity) in different practical application scenarios.

[0058] Figure 5 The algorithm's ability to recover the frequency domain active support set is demonstrated under different signal-to-noise ratio conditions.

[0059] The support set refers to the set of non-zero frequency points in the real frequency domain signal. For applications such as spectrum sensing, sparse channel estimation, and active subcarrier detection, the support recovery capability is very important because it directly reflects whether the algorithm can correctly identify the effective frequency points.

[0060] Figure 5 The F1 score takes into account both accuracy and recall. ; A higher F1 score indicates more accurate identification of active frequency points.

[0061] Figure 5 The technical significance lies in: Comparing only NMSEs is insufficient to comprehensively evaluate the effectiveness of communication signal processing; Support recovery indicators can reflect the actual task requirements such as spectrum occupancy identification and sparse channel path detection; The adaptive threshold may exhibit different performance under low noise and high noise conditions; This indicator can provide a basis for subsequent optimization of the adaptive selection module.

[0062] Example 4: Applications under different modulation methods and compression rates: The scope of this patent is not limited to QPSK modulation.

[0063] As another embodiment, the technical solution of this patent is also applicable to BPSK modulation. When the modulation method of the signal changes, only the mapping rule from symbols to constellation points needs to be changed accordingly in the initial signal generation step S1. The entire subsequent processing flow from S2 to S6, including preprocessing, sparse transformation, network model processing and reconstruction, does not require any adjustment. Simulation results show that, under BPSK modulation, the method proposed in this patent also maintains its significant performance advantage in the low signal-to-noise ratio region.

[0064] As another example, we tested the compressed observation dimension M by adjusting it to 48 and 80 respectively. The results show that the adaptive selection module in the system can respond sensitively to the compression ratio. Changes: When M=48 (i.e., higher compression ratio and fewer observations), the system will automatically tend to choose a "robust" configuration with stronger noise resistance for robustness considerations; When M=80 (i.e., higher observation dimension and more complete information), the system tends to choose the "balanced" configuration that achieves a balance between convergence speed and accuracy, or choose the "high-fidelity" configuration when conditions permit, in order to pursue better reconstruction quality.

[0065] Example 5: Network structures based on other algorithms: The core idea of ​​this patent is to construct a model-driven unfolded network framework, and the underlying optimization algorithm is not limited to ISTA.

[0066] As an alternative, the FISTA algorithm can be used as the basis for network expansion. FISTA accelerates the convergence process of the classic ISTA by introducing a momentum term. When expanded into a network, the computation of each layer, in addition to the original gradient descent and soft thresholding steps, requires additional momentum term calculations and state update steps. This variant network is expected to achieve faster convergence speed or better reconstruction results while maintaining the good interpretability of the model-driven approach.

[0067] As an alternative, the Approximate Message Passing (AMP) algorithm can also be used as the basis for expansion. The AMP algorithm cleverly introduces an "Onsager correction term" during its iteration process, which can provide more accurate iterative state estimates. The network formed by expanding it can usually achieve better theoretical performance and reconstruction accuracy when dealing with specific scenarios where the observation matrix is ​​an independent and identically distributed Gaussian matrix.

[0068] Running time reflects the computational complexity and real-time processing potential of an algorithm when deployed at the communication receiver. Traditional OMP algorithms typically require iterative selection of support sets and multiple solutions to the least squares problem; ISTA and FISA require a fixed number of iterations; MDSRNet achieves forward computation through a fixed-layer unfolded structure.

[0069] like Figure 6 For illustrative purposes: Different sparse reconstruction algorithms have different computational costs; Model-driven unfolding networks can transform the traditional iterative process into a fixed-layer computational process; The adaptive selection module can select the number of layers based on the scenario, thereby balancing reconstruction accuracy and computational cost; For real-time communication reception, runtime is an important evaluation metric besides reconstruction error.

[0070] like Figure 7 The impact of the number of unfolded layers in MDSRNet on the reconstruction error is demonstrated.

[0071] Figure 7 Different curves correspond to different signal-to-noise ratio conditions, which are used to verify that the number of network layers is not arbitrarily set, but will affect the reconstruction performance.

[0072] The experimental results show that: Under certain medium-to-high signal-to-noise ratio conditions, increasing the number of unfolded layers helps to reduce NMSE; Under low signal-to-noise ratio conditions, excessive number of layers may cause noise components to be further propagated or amplified; Therefore, a fixed number of layers is not necessarily suitable for all communication environments; This provides experimental evidence for introducing an "adaptive selection module" into this method.

[0073] The embodiments described above are merely illustrative of implementation methods of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be defined by the appended claims.

Claims

1. A method for adaptive processing of communication signals based on model-driven sparse reconstruction networks, characterized in that, Includes the following steps: S1. Establish a sparse frequency domain model for the communication signal: The sparse frequency domain model is expressed as follows: Frequency domain sparse signals: ,only Each non-zero frequency point carries modulation symbols Time-domain signal: ,in, unitary normalization inverse Fourier transform Compressed observation: ,in, For compressed sampling matrix, Additive white Gaussian noise S2. Constructing the perception matrix and performing real-valued processing; S3. Extract observation features and characterize environmental conditions; S4, Adaptive Configuration Selection; S5. Construct an adaptive model to drive the sparse reconstruction network and drive the forward inference of the sparse reconstruction network. S6. Output of sparse domain reconstruction results.

2. The adaptive processing method for communication signals based on model-driven sparse reconstruction networks according to claim 1, characterized in that, Specifically, S2 is: complex frequency domain signal vector By separating the real and imaginary parts, we obtain a real-valued sparse vector. ;in, Let be a vector with real part. The vector is the imaginary part. The range of values ​​for the real and virtual vectors; Construct an equivalent real-valued sensing matrix such that the real-valued observation vector is ; This is the perception matrix.

3. The adaptive processing method for communication signals based on model-driven sparse reconstruction networks according to claim 2, characterized in that, Specifically, S3 is: The adaptive selection module utilizes the observation vectors obtained from the receiver. and perception matrix Extract several meta-features that can characterize the current communication environment state; The meta-features include: Compression ratio: ,in, To compress the observation dimensions; Energy proxy quantity: ; Noise proxy: ; Signal-to-noise ratio (SNR) throughput: ,in, It is random noise.

4. The adaptive processing method for communication signals based on model-driven sparse reconstruction networks according to claim 3, characterized in that, Specifically, S4 is: The adaptive selection module automatically selects a set of reconstruction parameters that match the current environment from a preset configuration library based on the meta-features extracted in S3. The configuration library includes the following schemas: Low signal-to-noise ratio robust mode: Suitable for low signal-to-noise ratio or high compression ratio environments, it uses more expansion layers and strong threshold strength to prioritize noise suppression and avoid the occurrence of spurious peaks; Balanced mode: Suitable for environments with medium signal-to-noise ratio and normal compression rate, it uses a moderate number of unfolding layers and threshold strength to achieve a balance between reconstruction accuracy and computational overhead; High-fidelity mode: Suitable for environments with high signal-to-noise ratio and low compression rate. It uses fewer unfolding layers and weaker threshold strength to preserve weak signal details to the maximum extent and improve reconstruction fidelity.

5. The adaptive processing method for communication signals based on model-driven sparse reconstruction networks according to claim 4, characterized in that, Specifically, S5 is: An adaptive model MDSR Net, derived from an iterative soft thresholding algorithm, is invoked to drive the sparse reconstruction network. The sparse reconstruction network structure has a clear physical meaning, and the sparse reconstruction network structure... The layer processing procedure is represented as follows: Residual update: ; Gradient calculation: ; Adaptive threshold generation: ; Soft threshold operation: ; in, The threshold control coefficient decreases progressively at each level; The noise intensity of the current residual estimate; These are the energy normalization weighting coefficients; A threshold factor adjusted based on the compression ratio; Step size; This is a soft thresholding function; The sparse reconstruction network performs iterative inference based on the number of layers and threshold strength selected in S4.

6. The adaptive processing method for communication signals based on model-driven sparse reconstruction networks according to claim 5, characterized in that, Specifically, S6 is: After completing a specified number of iterations, the final real-valued sparse threshold signal estimate is output. And as needed, it is recombined into a complex frequency domain signal or the active frequency index and the modulation symbol it carries are decoded.

7. A communication signal adaptive processing system based on model-driven sparse reconstruction networks, characterized in that: The method employs any one of the model-driven sparse reconstruction network adaptive communication signal processing methods as described in any one of claims 1 to 6, and further includes: Signal observation modeling module: Establish a sparse frequency domain model, convert complex signals to real values, and construct an equivalent sensing matrix; Adaptive selection module: Extracts environmental features and intelligently selects the optimal reconstruction parameter configuration; Model-driven sparse reconstruction network: performs forward inference to complete sparse signal reconstruction.

Images