Low-noise down-conversion signal processing method and system

Abstract

The invention relates to the technical field of wireless communication, and provides a low-noise down-conversion signal processing method and system, and the method comprises the steps: collecting an I / Q baseband signal, carrying out the processing of a signal branch and a mirror image branch of a dual-branch decoupling convolutional neural network, and obtaining a signal feature vector and a mirror image feature vector; decomposing the feature vector into a real part, an imaginary part, amplitude and phase components, and constructing a state vector by combining temperature, frequency and device parameters; inputting the state vector into a radio frequency sensing deep reinforcement learning network to output a gain control strategy; environment change is monitored through the meta-learning network, and network weight is adjusted based on performance feedback; and regenerating a final gain control strategy by using the updated weight, controlling the variable gain amplifier to carry out gain adjustment on the I / Q signal, and outputting a signal after low-noise gain control. According to the invention, low-noise, high-precision and adaptive gain control processing of I / Q baseband signals in a complex radio frequency environment is realized, and the signal-to-noise ratio and control precision of down-conversion signals are improved.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and in particular to a low-noise down-conversion signal processing method and system. Background Technology

[0002] Radio frequency (RF) receiving systems are a core component of modern electronic devices such as wireless communication, radar detection, and satellite navigation. Down-conversion signal processing technology is responsible for converting high-frequency RF signals into baseband signals that are easily processed digitally. During down-conversion, the I / Q baseband signals are susceptible to various noise sources, including mixer nonlinearity, local oscillator phase noise, and image frequency interference, leading to signal quality degradation and limited dynamic range. To obtain high-quality baseband signals, precise gain control and noise suppression are required for the down-converted I / Q signals. However, the complexity and variability of the RF environment, temperature drift, and aging changes in device parameters present numerous challenges to I / Q signal processing. The key lies in effectively separating the signal from image noise, accurately identifying environmental changes, and implementing adaptive gain control strategy optimization.

[0003] In existing technologies, downconversion signal processing mainly employs traditional filter design and fixed gain control methods to achieve basic signal conditioning functions. However, existing methods do not adequately consider the inherent physical correlation mechanism between the complex electromagnetic environment of the RF system and the characteristics of I / Q signals. This makes it difficult to organically integrate objective RF physical constraints with actual signal processing requirements, resulting in the inability to achieve high-precision and highly adaptive low-noise downconversion signal processing and gain control. Summary of the Invention

[0004] In view of this, the present invention proposes a low-noise downconversion signal processing method and system, which solves the problem that the existing technology does not adequately consider the inherent physical correlation mechanism between the complex electromagnetic environment of the radio frequency system and the characteristics of I / Q signals, and is difficult to organically integrate objective radio frequency physical constraints with actual signal processing requirements, resulting in the inability to achieve high-precision and highly adaptive low-noise downconversion signal processing and gain control.

[0005] The technical solution of this invention is implemented as follows: On one hand, this invention provides a low-noise down-conversion signal processing method, comprising the following steps: The down-converted I / Q baseband signal is acquired and input into a dual-branch decoupled convolutional neural network. The dual-branch decoupled convolutional neural network includes a signal branch and a mirror branch. The I / Q baseband signal is processed by the signal branch and the mirror branch respectively to obtain the signal feature vector and the mirror feature vector. The signal feature vector and the mirror feature vector are decomposed into real components, imaginary components, amplitude components and phase components, and a state vector is constructed by combining system temperature parameters, frequency parameters and device parameters; The state vector is input into the radio frequency sensing deep reinforcement learning network, and a gain control strategy is output based on the state vector. The network weights of the radio frequency sensing deep reinforcement learning network are adjusted based on the performance feedback of the gain control strategy by monitoring environmental change parameters through the meta-learning network, and the updated network weights are output. The RF sensing deep reinforcement learning network is adjusted by updating the network weights to obtain an updated RF sensing deep reinforcement learning network. Based on the updated RF sensing deep reinforcement learning network, a final gain control strategy is generated and output to a variable gain amplifier. The variable gain amplifier adjusts the gain of the I / Q baseband signal according to the final gain control strategy and outputs the I / Q baseband signal with low noise gain control.

[0006] Based on the above technical solutions, preferably, the dual-branch decoupled convolutional neural network further includes a correlation suppression layer and a radio frequency physical constraint module; The correlation suppression layer is disposed between the signal branch and the mirror branch, and is used to calculate the feature correlation between the signal feature vector and the mirror feature vector, and to minimize the feature correlation through a learnable suppression parameter. The radio frequency physical constraint module generates a physical constraint loss based on the downconversion mathematical model, and trains the dual-branch decoupled convolutional neural network together with the physical constraint loss and the feature extraction loss. The signal branch extracts the frequency and time domain features of the useful signal through the first convolutional layer group, and the mirror branch extracts the frequency and time domain features of the mirror frequency noise through the second convolutional layer group. The training processes of the convolutional kernel parameters of the first and second convolutional layer groups are independent of each other.

[0007] Based on the above technical solutions, preferably, feature decoupling is performed through the correlation suppression layer, including: Calculate the cross-correlation matrix between the signal feature vector and the mirror feature vector, and perform singular value decomposition on the cross-correlation matrix to obtain the correlation strength index; The learnable suppression parameter is updated by gradient backpropagation so that the correlation strength index converges to below a preset decoupling threshold. The radio frequency physical constraint module generates a complex domain physical constraint loss function based on the nonlinear characteristics and noise figure model of the radio frequency mixer. The complex domain physical constraint loss function includes an amplitude constraint term and a phase constraint term. The first convolutional layer group uses a mirror suppression convolutional kernel, and the weight distribution of the mirror suppression convolutional kernel is established with negative weights based on the mirror frequency during the downconversion process. The second convolutional layer group uses a mirror enhancement convolutional kernel, and the weight distribution of the mirror enhancement convolutional kernel is established with positive weights based on the mirror frequency.

[0008] Based on the above technical solutions, preferably, the step of inputting the state vector into the radio frequency sensing deep reinforcement learning network and outputting a gain control strategy according to the state vector includes: The radio frequency sensing deep reinforcement learning network includes a radio frequency domain knowledge embedding layer, a dual time-scale Actor-Critic structure, and a radio frequency constraint decision mechanism. The radio frequency domain knowledge embedding layer enhances the state vector with radio frequency domain features and outputs an enhanced state vector. The dual-timescale Actor-Critic structure includes a fast-response Actor network and a long-term evaluation Critic network. The fast-response Actor network receives the enhanced state vector and generates a set of candidate gain adjustment strategies. The long-term evaluation Critic network evaluates the set of candidate gain adjustment strategies and calculates the strategy value. The radio frequency constraint decision mechanism selects the optimal gain control strategy based on the candidate gain adjustment strategy set and the corresponding strategy value, combined with the radio frequency system safety boundary conditions, through a differentiable strategy optimization algorithm. The gain control strategy includes gain adjustment magnitude, adjustment timing, and adjustment curve type.

[0009] Based on the above technical solutions, preferably, feature enhancement is performed through the radio frequency domain knowledge embedding layer, including: A radio frequency system simulator was built to perform radio frequency physical modeling of the mixing process and gain change of the I / Q baseband signal; The state vector is input into the radio frequency system simulator to generate a system response prediction. The system response prediction is then concatenated with the state vector to obtain a physical enhancement feature vector. Self-attention weights are calculated on the physical augmentation feature vector. An attention weight distribution is generated by matrix multiplication of the query matrix, key matrix, and value matrix. The attention weight distribution is then weighted and summed with the physical augmentation feature vector to output the augmentation state vector.

[0010] Based on the above technical solutions, preferably, the dual-timescale Actor-Critic structure is used to decouple different decision-making timescales, including: The fast-response Actor network uses a microsecond-level sampling rate for decision-making and provides fast gain control for transient interference and sudden noise. The long-term evaluation Critic network uses a millisecond-level sampling rate for evaluation and constructs a composite reward function based on signal quality indicators, energy efficiency ratio, and stability indicators. The differentiable strategy optimization algorithm employs radio frequency security risk perception search, transforming the linear region constraint, P1dB compression point constraint, and gain step size constraint of the variable gain amplifier into soft constraints for strategy selection, thereby maximizing signal quality.

[0011] Based on the above technical solutions, preferably, the step of decomposing the signal feature vector and the mirror feature vector into real components, imaginary components, amplitude components, and phase components, and constructing a state vector by combining system temperature parameters, frequency parameters, and device parameters, includes: The signal feature vector and the mirror feature vector are decomposed in the complex field to extract the real component, imaginary component, amplitude component and phase component respectively. The weighted feature component is obtained by assigning weight coefficients to each component through an adaptive weight allocation module. The system temperature parameters, frequency parameters, and device parameters of the current radio frequency system are obtained. The weighted feature components are then fused with the system parameters through a multi-dimensional parameter fusion network to construct the state vector containing radio frequency environment information.

[0012] Based on the above technical solutions, preferably, the step of monitoring environmental change parameters through a meta-learning network, adjusting the network weights of the radio frequency sensing deep reinforcement learning network based on the performance feedback of the gain control strategy, and outputting updated network weights includes: The environmental change parameters of the radio frequency system are monitored in real time through a meta-learning network. These environmental change parameters include signal strength change, spectrum interference change, and system noise change. When the environmental change parameters exceed a preset threshold, a performance evaluation mechanism is activated. Based on the actual control effect of the gain control strategy in the current environment, the performance feedback index is calculated through the strategy performance evaluation module. The network weights of the RF sensing deep reinforcement learning network are adjusted according to the performance feedback index using a gradient update algorithm, and the updated network weights are output.

[0013] Based on the above technical solutions, preferably, the step of adjusting the RF sensing deep reinforcement learning network by updating network weights to obtain an updated RF sensing deep reinforcement learning network, generating a final gain control strategy based on the updated RF sensing deep reinforcement learning network, outputting the final gain control strategy to a variable gain amplifier, and the variable gain amplifier adjusting the gain of the I / Q baseband signal according to the final gain control strategy to output a low-noise gain-controlled I / Q baseband signal includes: The parameters of the radio frequency sensing deep reinforcement learning network are updated using a weight smoothing update mechanism, and the updated network weights are fused with the network weights using an exponential weighted average algorithm. The updated RF sensing deep reinforcement learning network generates the final gain control strategy based on the weighted update. The final gain control strategy is converted into the control command format of the variable gain amplifier through the gain strategy output interface, and real-time gain adjustment control is executed.

[0014] On the other hand, the present invention also provides a low-noise down-conversion signal processing system, the system comprising: A dual-branch feature extraction module is used to acquire the down-converted I / Q baseband signal and input the I / Q baseband signal into a dual-branch decoupled convolutional neural network. The dual-branch decoupled convolutional neural network includes a signal branch and a mirror branch. The I / Q baseband signal is processed by the signal branch and the mirror branch respectively to obtain the signal feature vector and the mirror feature vector. The state vector construction module is used to decompose the signal feature vector and the mirror feature vector into real components, imaginary components, amplitude components and phase components, and construct the state vector by combining the system temperature parameters, frequency parameters and device parameters. The radio frequency sensing decision module is used to input the state vector into the radio frequency sensing deep reinforcement learning network and output a gain control strategy based on the state vector. The adaptive learning module is used to monitor environmental change parameters through the meta-learning network, adjust the network weights of the radio frequency sensing deep reinforcement learning network based on the performance feedback of the gain control strategy, and output updated network weights. The gain control execution module is used to adjust the radio frequency sensing deep reinforcement learning network by updating the network weights to obtain an updated radio frequency sensing deep reinforcement learning network. Based on the updated radio frequency sensing deep reinforcement learning network, a final gain control strategy is generated, and the final gain control strategy is output to a variable gain amplifier. The variable gain amplifier adjusts the gain of the I / Q baseband signal according to the final gain control strategy and outputs a low-noise gain-controlled I / Q baseband signal.

[0015] The low-noise down-conversion signal processing method and system of the present invention have the following advantages over the prior art: (1) I / Q signals and image noise are separated by a dual-branch decoupled convolutional neural network. A comprehensive system state characterization is constructed by combining complex domain multidimensional feature decomposition and RF physical parameter fusion. A gain control strategy is generated by using a RF sensing deep reinforcement learning network. Real-time perception of environmental changes and adaptive adjustment of network weights are achieved through a meta-learning network. Finally, low-noise, high-precision, adaptive gain control processing of I / Q baseband signals under complex RF environment is realized, which improves the signal-to-noise ratio and control accuracy of down-conversion signals. (2) By introducing a correlation suppression layer and a radio frequency physical constraint module, the signal branch and the mirror branch are deeply decoupled. Through the singular value decomposition of the cross-correlation matrix and gradient backpropagation optimization, the coupling interference between the signal features and the mirror noise features is eliminated, and the enhanced extraction of useful signals and the accurate capture of mirror frequency noise are realized, thereby improving the signal separation capability and feature extraction quality of the dual-branch network. (3) By embedding knowledge in the radio frequency field, adopting a dual-time-scale Actor-Critic architecture, and combining radio frequency system physical modeling and radio frequency constraint decision-making mechanism, the environmental cognition ability is dynamically improved based on the self-attention feature enhancement algorithm. A dual-time-scale decoupled structure of microsecond-level fast response and millisecond-level long-term evaluation is adopted. Differentiable strategy optimization algorithm and radio frequency safety boundary constraints are introduced to couple and model the radio frequency environment state, signal quality indicators and hardware operating parameters, thereby realizing high-precision intelligent gain control and dynamic environmental adaptation of I / Q baseband signals. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of a low-noise down-conversion signal processing method according to the present invention. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 This invention provides a low-noise down-conversion signal processing method, comprising the following steps: The down-converted I / Q baseband signal is acquired and input into a dual-branch decoupled convolutional neural network. The dual-branch decoupled convolutional neural network includes a signal branch and a mirror branch. The I / Q baseband signal is processed by the signal branch and the mirror branch respectively to obtain the signal feature vector and the mirror feature vector. The signal feature vector and the mirror feature vector are decomposed into real components, imaginary components, amplitude components and phase components, and a state vector is constructed by combining system temperature parameters, frequency parameters and device parameters; The state vector is input into the radio frequency sensing deep reinforcement learning network, and a gain control strategy is output based on the state vector. The network weights of the radio frequency sensing deep reinforcement learning network are adjusted based on the performance feedback of the gain control strategy by monitoring environmental change parameters through the meta-learning network, and the updated network weights are output. The RF sensing deep reinforcement learning network is adjusted by updating the network weights to obtain an updated RF sensing deep reinforcement learning network. Based on the updated RF sensing deep reinforcement learning network, a final gain control strategy is generated and output to a variable gain amplifier. The variable gain amplifier adjusts the gain of the I / Q baseband signal according to the final gain control strategy and outputs the I / Q baseband signal with low noise gain control.

[0020] Specifically, this embodiment separates the I / Q signals from the image noise using a dual-branch decoupled convolutional neural network, constructs a comprehensive system state characterization by combining complex domain multidimensional feature decomposition and RF physical parameter fusion, generates a gain control strategy using an RF sensing deep reinforcement learning network, and achieves real-time perception of environmental changes and adaptive adjustment of network weights through a meta-learning network. Ultimately, it realizes low-noise, high-precision, adaptive gain control processing of I / Q baseband signals under complex RF environments, improving the signal-to-noise ratio and control accuracy of the down-conversion signal.

[0021] The dual-branch decoupled convolutional neural network also includes a correlation suppression layer and a radio frequency physical constraint module; The correlation suppression layer is disposed between the signal branch and the mirror branch, and is used to calculate the feature correlation between the signal feature vector and the mirror feature vector, and to minimize the feature correlation through a learnable suppression parameter. The radio frequency physical constraint module generates a physical constraint loss based on the downconversion mathematical model, and trains the dual-branch decoupled convolutional neural network together with the physical constraint loss and the feature extraction loss. The signal branch extracts the frequency and time domain features of the useful signal through the first convolutional layer group, and the mirror branch extracts the frequency and time domain features of the mirror frequency noise through the second convolutional layer group. The training processes of the convolutional kernel parameters of the first and second convolutional layer groups are independent of each other.

[0022] Feature decoupling through the correlation suppression layer includes: Calculate the cross-correlation matrix between the signal feature vector and the mirror feature vector, perform singular value decomposition on the cross-correlation matrix, extract the principal singular values, and obtain the correlation strength index. The learnable suppression parameter is updated by gradient backpropagation so that the correlation strength index converges to below a preset decoupling threshold. The radio frequency physical constraint module generates a complex domain physical constraint loss function based on the nonlinear characteristics and noise figure model of the radio frequency mixer. The complex domain physical constraint loss function includes an amplitude constraint term and a phase constraint term. The first convolutional layer group uses a mirror suppression convolutional kernel, and the weight distribution of the mirror suppression convolutional kernel is established with negative weights based on the mirror frequency during the downconversion process. The second convolutional layer group uses a mirror enhancement convolutional kernel, and the weight distribution of the mirror enhancement convolutional kernel is established with positive weights based on the mirror frequency.

[0023] In one specific embodiment, the feature decoupling calculation formula of the correlation suppression layer is: ; in, For the first The correlation strength index of the next iteration; The preset decoupling threshold; The number of singular values; For the first Individual singular values For the first One learnable inhibition parameter; This is the ambient temperature normalization factor. For feature-based correlation The nonlinear activation function; For the first The iteration of the ... Relevance of each feature; It is an exponential function.

[0024] The complex domain physical constraint loss function of the radio frequency physical constraint module is calculated as follows: ; in, For the physical constraint loss function in the complex field; These are the weighting coefficients for the amplitude constraint term. For the first Amplitude constraint loss of the layer, These are the weighting coefficients for the phase constraint term. For the first Phase constraint loss of the layer; Weights for the nonlinear characteristics of the mixer. For nonlinear constraint functions, For the mixer nonlinear coefficients, These are the noise figure model parameters.

[0025] Specifically, this embodiment introduces a correlation suppression layer and a radio frequency physical constraint module to deeply decouple the signal branch and the mirror branch. Through singular value decomposition of the cross-correlation matrix and gradient backpropagation optimization, the coupling interference between signal features and mirror noise features is eliminated. The complex domain physical constraint loss function, established based on the nonlinear characteristics of the RF mixer, integrates the objective physical laws of the RF system into the neural network training process, ensuring the physical consistency and accuracy of feature extraction. By using differentiated weights of mirror suppression convolution kernels and mirror enhancement convolution kernels, it achieves enhanced extraction of useful signals and accurate capture of mirror frequency noise, thereby improving the signal separation capability and feature extraction quality of the dual-branch network.

[0026] The step of decomposing the signal feature vector and the mirror feature vector into real components, imaginary components, amplitude components, and phase components, and constructing a state vector by combining system temperature parameters, frequency parameters, and device parameters, includes: The signal feature vector and the mirror feature vector are decomposed in the complex field to extract the real component, imaginary component, amplitude component and phase component respectively. The weighted feature components are obtained by assigning weight coefficients to each component through an adaptive weight allocation module.

[0027] In one specific embodiment, the process of obtaining the weight coefficients of the adaptive weight allocation module includes: The initial weight ratio is calculated based on the signal-to-noise ratio difference between the signal feature vector and the mirror feature vector; Based on the energy distribution of the real and imaginary components, the complex field weight allocation strategy is dynamically adjusted. Based on the stability assessment of the amplitude component and the phase component, an amplitude-phase weight balance factor is generated; The initial weight ratio, the complex domain weight allocation strategy, and the amplitude-phase weight balance factor are weighted and fused to output the weight coefficients corresponding to each component.

[0028] The system temperature parameters, frequency parameters, and device parameters of the current radio frequency system are obtained. The weighted feature components are then fused with the system parameters through a multi-dimensional parameter fusion network to construct the state vector containing radio frequency environment information.

[0029] In one specific embodiment, the multidimensional parameter fusion network includes: A temperature adaptive layer performs temperature compensation processing on the weighted feature components based on the system temperature parameters to generate temperature-normalized features. The frequency mapping layer converts the frequency parameters into frequency feature vectors and performs frequency domain correlation and fusion with the temperature normalization features. The device aging compensation layer establishes a device performance degradation model based on the device parameters and performs aging compensation on the fused features. The state vector generation layer maps the compensated features into a fixed-dimensional state vector through a fully connected network. The state vector includes a signal quality indicator component, an environmental state component, and a system health component.

[0030] Specifically, this embodiment combines adaptive weight allocation, complex domain feature decomposition, multi-dimensional parameter fusion, and temperature-frequency compensation mechanisms. Based on the signal-to-noise ratio difference and energy distribution, it dynamically adjusts the weight allocation strategy and adopts a multi-layer fusion architecture consisting of a temperature adaptive layer, a frequency mapping layer, and a device aging compensation layer. It introduces an amplitude-phase weight balance factor and a device performance degradation model to comprehensively fuse and model the real, imaginary, amplitude, and phase components of the signal features, as well as the system temperature, frequency, and device parameters. This ensures that the state vector is highly consistent with the actual operating state of the RF system, achieving high-precision complex domain feature extraction of I / Q baseband signals, adaptive compensation of multi-dimensional environmental parameters, and comprehensive state characterization. The comprehensive state characterization includes signal quality indication, environmental status, and system health.

[0031] The step of inputting the state vector into the radio frequency sensing deep reinforcement learning network and outputting a gain control strategy based on the state vector includes: The radio frequency sensing deep reinforcement learning network includes a radio frequency domain knowledge embedding layer, a dual time-scale Actor-Critic structure, and a radio frequency constraint decision mechanism. The radio frequency domain knowledge embedding layer enhances the state vector with radio frequency domain features and outputs an enhanced state vector. The dual-timescale Actor-Critic structure includes a fast-response Actor network and a long-term evaluation Critic network. The fast-response Actor network receives the enhanced state vector and generates a set of candidate gain adjustment strategies. The long-term evaluation Critic network evaluates the set of candidate gain adjustment strategies and calculates the strategy value. The radio frequency constraint decision mechanism selects the optimal gain control strategy based on the candidate gain adjustment strategy set and the corresponding strategy value, combined with the radio frequency system safety boundary conditions, through a differentiable strategy optimization algorithm. The gain control strategy includes gain adjustment magnitude, adjustment timing, and adjustment curve type.

[0032] Feature enhancement through the radio frequency domain knowledge embedding layer includes: A radio frequency system simulator was built to perform radio frequency physical modeling of the mixing process and gain change of the I / Q baseband signal; The state vector is input into the radio frequency system simulator to generate a system response prediction. The system response prediction is then concatenated with the state vector to obtain a physical enhancement feature vector. Self-attention weights are calculated on the physical augmentation feature vector. An attention weight distribution is generated by matrix multiplication of the query matrix, key matrix, and value matrix. The attention weight distribution is then weighted and summed with the physical augmentation feature vector to output the augmentation state vector.

[0033] The dual-timescale Actor-Critic structure decouples different decision-making timescales, including: The fast-response Actor network uses a microsecond-level sampling rate for decision-making and provides fast gain control for transient interference and sudden noise. The long-term evaluation Critic network uses a millisecond-level sampling rate for evaluation and constructs a composite reward function based on signal quality indicators, energy efficiency ratio, and stability indicators. The differentiable strategy optimization algorithm adopts radio frequency safety risk perception search, which transforms the linear region constraint, P1dB compression point constraint and gain step size constraint of the variable gain amplifier into soft constraint conditions for strategy selection, so as to maximize signal quality while ensuring system safety.

[0034] In one specific embodiment, the composite reward function of the dual-timescale Actor-Critic structure is calculated as follows: ; in, For a moment The value of the compound reward function; As the weight of the signal quality index, time The output signal-to-noise ratio; As a weight for energy efficiency ratio, For a moment The energy efficiency ratio index; As the weight of the stability index, For a moment System stability indicators; For safety violation penalty coefficient, To constrain the violation penalty function for safety, These are radio frequency safety boundary parameters.

[0035] The differentiable strategy optimization algorithm uses the following calculation formula for radio frequency security risk perception search: ; in, For strategy The optimization objective function; To quickly respond to Actor network parameters, For Actor network policy distribution, This is the expected value based on the policy distribution; This represents the total number of radio frequency constraints. For the first The soft constraint penalty weight of each constraint. For the first One radio frequency safety constraint function (including linear region constraint, P1dB compression point constraint, and gain step constraint). For the first A safety threshold for each constraint.

[0036] Specifically, this embodiment embeds knowledge from the radio frequency (RF) domain, adopts a dual-time-scale Actor-Critic architecture, and combines RF system physical modeling and RF constraint decision-making mechanisms. It dynamically improves environmental awareness based on a self-attention feature enhancement algorithm, employs a dual-time-scale decoupled structure of microsecond-level fast response and millisecond-level long-term evaluation, introduces a differentiable strategy optimization algorithm and RF safety boundary constraints, and couples and models the RF environment state, signal quality indicators, and hardware operating parameters to achieve high-precision intelligent gain control and dynamic environmental adaptation of I / Q baseband signals.

[0037] The process of monitoring environmental change parameters through a meta-learning network, adjusting the network weights of the RF sensing deep reinforcement learning network based on performance feedback from the gain control strategy, and outputting updated network weights includes: The environmental change parameters of the radio frequency system are monitored in real time through a meta-learning network. These environmental change parameters include changes in signal strength, spectral interference, and system noise. When the environmental change parameters exceed a preset threshold, a performance evaluation mechanism is activated.

[0038] In one specific embodiment, environmental monitoring via the meta-learning network includes: Establish an environmental parameter baseline database to store the signal strength baseline, spectral purity baseline, and noise floor baseline under normal operating conditions; A sliding window mechanism is used to sample the current environmental parameters in real time, and to calculate the deviation rate of the signal strength change relative to the signal strength baseline, the degree of degradation of the spectral interference change relative to the spectral purity baseline, and the magnitude of the rise of the system noise change relative to the noise floor baseline. The deviation rate, the degree of deterioration, and the magnitude of increase are compared with corresponding preset thresholds. When any parameter exceeds the threshold, the severity index of environmental change is calculated. The urgency level of network weight adjustment is determined based on the severity index of environmental changes, and an adjustment trigger signal is sent to the strategy performance evaluation module.

[0039] Based on the actual control effect of the gain control strategy in the current environment, the performance feedback index is calculated through the strategy performance evaluation module. The network weights of the RF sensing deep reinforcement learning network are adjusted according to the performance feedback index using a gradient update algorithm, and the updated network weights are output.

[0040] In one specific embodiment, the performance feedback metric is calculated through the strategy performance evaluation module, including: Collect system response data after the gain control strategy is executed, including output signal-to-noise ratio, gain stability, and control delay; A multi-objective performance evaluation function is constructed, which combines the improvement of the output signal-to-noise ratio, the fluctuation variance of the gain stability, and the response time of the control delay in a weighted manner. Calculate the performance gap between the current strategy and the historical best strategy, and generate a performance feedback error signal; The gradient update algorithm calculates the gradient direction and update step size of the network weights based on the performance feedback error signal, and adopts an adaptive learning rate adjustment mechanism to update the Actor network weights and Critic network weights in the radio frequency sensing deep reinforcement learning network respectively, so as to ensure the convergence of the network's policy optimization in the new environment.

[0041] Specifically, this embodiment integrates meta-learning environment monitoring, policy performance evaluation, gradient adaptive update, and weight stability verification mechanisms. Based on sliding window real-time sampling and environmental parameter baseline comparison algorithms, it dynamically identifies changes in the radio frequency environment. It adopts a multi-objective performance evaluation function and an adaptive learning rate adjustment weight update architecture, introduces an environmental change severity index and a network weight adjustment urgency level, comprehensively monitors and evaluates changes in signal strength, spectral interference, system noise, and the actual execution effect of the gain control strategy. It also performs real-time perception and identification of changes in the radio frequency system environment, dynamically evaluates the performance of the gain control strategy, and adaptively optimizes the network weights, thereby improving adaptability and control accuracy in complex electromagnetic environments.

[0042] The process involves adjusting the RF sensing deep reinforcement learning network by updating network weights to obtain an updated RF sensing deep reinforcement learning network. Based on this updated network, a final gain control strategy is generated. This final gain control strategy is then output to a variable gain amplifier. The variable gain amplifier adjusts the gain of the I / Q baseband signal according to the final gain control strategy, outputting a low-noise gain-controlled I / Q baseband signal. This includes: The parameters of the radio frequency sensing deep reinforcement learning network are updated using a weight smoothing update mechanism. The updated network weights are then fused with the network weights using an exponential weighted average algorithm to avoid instability in the control strategy caused by sudden weight changes.

[0043] In one specific embodiment, the weight smoothing update mechanism includes: A weighted fusion coefficient is set, which is dynamically adjusted according to the severity of environmental changes; the more severe the environmental changes, the larger the weighted fusion coefficient. Calculate the weighted average value of the weight parameters of each layer of the radio frequency sensing deep reinforcement learning network; A weight update stability test mechanism is established. The magnitude of weight change is evaluated by calculating the Euclidean distance between the network weights before and after the update. When the magnitude of weight change exceeds the stability threshold, the weight fusion coefficient is reduced and the fusion weight is recalculated. The fusion weights are loaded layer by layer into the radio frequency sensing deep reinforcement learning network to complete the smooth update of network parameters.

[0044] The updated RF sensing deep reinforcement learning network generates the final gain control strategy based on the weighted update. The final gain control strategy is converted into the control command format of the variable gain amplifier through the gain strategy output interface, and real-time gain adjustment control is executed.

[0045] In one specific embodiment, the gain strategy output interface includes: The strategy parsing module parses the gain adjustment magnitude, adjustment timing, and adjustment curve type in the final gain control strategy into numerical parameters; The hardware adaptation module converts the numerical parameters into corresponding control voltage values ​​based on the control voltage range and gain characteristic curve of the variable gain amplifier. The timing control module generates a gain control timing table based on the adjusted timing to ensure that the timing accuracy of the gain adjustment meets the microsecond-level control requirements. Before outputting the control voltage value, the safety protection module checks whether the control command exceeds the safe operating range of the variable gain amplifier. When an out-of-range command is detected, it automatically limits the control command within the safe boundary and records the abnormality log.

[0046] Specifically, this embodiment combines a weight smoothing update mechanism, a gain strategy output interface, hardware security protection, and microsecond-level timing control. Based on the exponential weighted average algorithm and the degree of environmental change, it dynamically adjusts the weight fusion coefficients. It adopts a smoothing update architecture with hierarchical weight loading and stability verification. Utilizing a four-module output interface including strategy parsing, hardware adaptation, timing control, and security protection, it comprehensively monitors and manages the magnitude of network weight changes, the security of control commands, and the timing of gain adjustments. This ensures the smooth and stable update process of the RF sensing deep reinforcement learning network and the safe and reliable execution of the final control strategy. It achieves seamless smoothing updates of network weights and microsecond-level secure control of the variable gain amplifier, improving the stability of weight updates and the accuracy of control command execution.

[0047] The present invention also provides a low-noise down-conversion signal processing system, the system comprising: A dual-branch feature extraction module is used to acquire the down-converted I / Q baseband signal and input the I / Q baseband signal into a dual-branch decoupled convolutional neural network. The dual-branch decoupled convolutional neural network includes a signal branch and a mirror branch. The I / Q baseband signal is processed by the signal branch and the mirror branch respectively to obtain the signal feature vector and the mirror feature vector. The state vector construction module is used to decompose the signal feature vector and the mirror feature vector into real components, imaginary components, amplitude components and phase components, and construct the state vector by combining the system temperature parameters, frequency parameters and device parameters. The radio frequency sensing decision module is used to input the state vector into the radio frequency sensing deep reinforcement learning network and output a gain control strategy based on the state vector. The adaptive learning module is used to monitor environmental change parameters through the meta-learning network, adjust the network weights of the radio frequency sensing deep reinforcement learning network based on the performance feedback of the gain control strategy, and output updated network weights. The gain control execution module is used to adjust the radio frequency sensing deep reinforcement learning network by updating the network weights to obtain an updated radio frequency sensing deep reinforcement learning network. Based on the updated radio frequency sensing deep reinforcement learning network, a final gain control strategy is generated, and the final gain control strategy is output to a variable gain amplifier. The variable gain amplifier adjusts the gain of the I / Q baseband signal according to the final gain control strategy and outputs a low-noise gain-controlled I / Q baseband signal.

[0048] Specifically, this embodiment of a low-noise down-conversion signal processing system includes five functional modules: dual-branch feature extraction, multi-dimensional state construction, radio frequency sensing decision-making, adaptive learning, and gain control execution. It achieves effective separation of signal and image noise through a dual-branch decoupled convolutional neural network, employs a state vector construction architecture based on complex domain full feature decomposition and multi-parameter fusion, and introduces radio frequency sensing deep reinforcement learning and meta-learning environment monitoring mechanisms. It performs end-to-end integrated processing of I / Q baseband signal feature extraction, environmental state perception, intelligent decision generation, and gain execution control, ensuring high coordination between the entire signal processing flow and the dynamic radio frequency environment. This embodiment achieves low-noise, high-precision processing of down-converted signals, real-time sensing and adaptation to complex radio frequency environments, and safe and reliable control of variable gain amplifiers, improving the signal-to-noise ratio, environmental adaptability, and control accuracy of the radio frequency receiving system.

[0049] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A low-noise down-conversion signal processing method, characterized in that, Includes the following steps: The down-converted I / Q baseband signal is acquired and input into a dual-branch decoupled convolutional neural network. The dual-branch decoupled convolutional neural network includes a signal branch and a mirror branch. The I / Q baseband signal is processed by the signal branch and the mirror branch respectively to obtain the signal feature vector and the mirror feature vector. The signal feature vector and the mirror feature vector are decomposed into real components, imaginary components, amplitude components and phase components, and a state vector is constructed by combining system temperature parameters, frequency parameters and device parameters; The state vector is input into the radio frequency sensing deep reinforcement learning network, and a gain control strategy is output based on the state vector. The network weights of the radio frequency sensing deep reinforcement learning network are adjusted based on the performance feedback of the gain control strategy by monitoring environmental change parameters through the meta-learning network, and the updated network weights are output. The RF sensing deep reinforcement learning network is adjusted by updating the network weights to obtain an updated RF sensing deep reinforcement learning network. Based on the updated RF sensing deep reinforcement learning network, a final gain control strategy is generated and output to a variable gain amplifier. The variable gain amplifier adjusts the gain of the I / Q baseband signal according to the final gain control strategy and outputs the I / Q baseband signal with low noise gain control.

2. The low-noise down-conversion signal processing method as described in claim 1, characterized in that, The dual-branch decoupled convolutional neural network also includes a correlation suppression layer and a radio frequency physical constraint module; The correlation suppression layer is disposed between the signal branch and the mirror branch, and is used to calculate the feature correlation between the signal feature vector and the mirror feature vector, and to minimize the feature correlation through a learnable suppression parameter. The radio frequency physical constraint module generates a physical constraint loss based on the downconversion mathematical model, and trains the dual-branch decoupled convolutional neural network together with the physical constraint loss and the feature extraction loss. The signal branch extracts the frequency and time domain features of the useful signal through the first convolutional layer group, and the mirror branch extracts the frequency and time domain features of the mirror frequency noise through the second convolutional layer group. The training processes of the convolutional kernel parameters of the first and second convolutional layer groups are independent of each other.

3. The low-noise down-conversion signal processing method as described in claim 2, characterized in that, Feature decoupling through the correlation suppression layer includes: Calculate the cross-correlation matrix between the signal feature vector and the mirror feature vector, and perform singular value decomposition on the cross-correlation matrix to obtain the correlation strength index; The learnable suppression parameter is updated by gradient backpropagation so that the correlation strength index converges to below a preset decoupling threshold. The radio frequency physical constraint module generates a complex domain physical constraint loss function based on the nonlinear characteristics and noise figure model of the radio frequency mixer. The complex domain physical constraint loss function includes an amplitude constraint term and a phase constraint term. The first convolutional layer group uses a mirror suppression convolutional kernel, and the weight distribution of the mirror suppression convolutional kernel is established with negative weights based on the mirror frequency during the downconversion process. The second convolutional layer group uses a mirror enhancement convolutional kernel, and the weight distribution of the mirror enhancement convolutional kernel is established with positive weights based on the mirror frequency.

4. The low-noise down-conversion signal processing method as described in claim 1, characterized in that, The step of inputting the state vector into the radio frequency sensing deep reinforcement learning network and outputting a gain control strategy based on the state vector includes: The radio frequency sensing deep reinforcement learning network includes a radio frequency domain knowledge embedding layer, a dual time-scale Actor-Critic structure, and a radio frequency constraint decision mechanism. The radio frequency domain knowledge embedding layer enhances the state vector with radio frequency domain features and outputs an enhanced state vector. The dual-timescale Actor-Critic structure includes a fast-response Actor network and a long-term evaluation Critic network. The fast-response Actor network receives the enhanced state vector and generates a set of candidate gain adjustment strategies. The long-term evaluation Critic network evaluates the set of candidate gain adjustment strategies and calculates the strategy value. The radio frequency constraint decision mechanism selects the optimal gain control strategy based on the candidate gain adjustment strategy set and the corresponding strategy value, combined with the radio frequency system safety boundary conditions, through a differentiable strategy optimization algorithm. The gain control strategy includes gain adjustment magnitude, adjustment timing, and adjustment curve type.

5. The low-noise down-conversion signal processing method as described in claim 4, characterized in that, Feature enhancement through the radio frequency domain knowledge embedding layer includes: A radio frequency system simulator was built to perform radio frequency physical modeling of the mixing process and gain change of the I / Q baseband signal; The state vector is input into the radio frequency system simulator to generate a system response prediction. The system response prediction is then concatenated with the state vector to obtain a physical enhancement feature vector. Self-attention weights are calculated on the physical augmentation feature vector. An attention weight distribution is generated by matrix multiplication of the query matrix, key matrix, and value matrix. The attention weight distribution is then weighted and summed with the physical augmentation feature vector to output the augmentation state vector.

6. The low-noise down-conversion signal processing method as described in claim 5, characterized in that, The dual-timescale Actor-Critic structure decouples different decision-making timescales, including: The fast-response Actor network uses a microsecond-level sampling rate for decision-making and provides fast gain control for transient interference and sudden noise. The long-term evaluation Critic network uses a millisecond-level sampling rate for evaluation and constructs a composite reward function based on signal quality indicators, energy efficiency ratio, and stability indicators. The differentiable strategy optimization algorithm employs radio frequency security risk perception search, transforming the linear region constraint, P1dB compression point constraint, and gain step size constraint of the variable gain amplifier into soft constraints for strategy selection, thereby maximizing signal quality.

7. The low-noise down-conversion signal processing method as described in claim 1, characterized in that, The step of decomposing the signal feature vector and the mirror feature vector into real components, imaginary components, amplitude components, and phase components, and constructing a state vector by combining system temperature parameters, frequency parameters, and device parameters, includes: The signal feature vector and the mirror feature vector are decomposed in the complex field to extract the real component, imaginary component, amplitude component and phase component respectively. The weighted feature component is obtained by assigning weight coefficients to each component through an adaptive weight allocation module. The system temperature parameters, frequency parameters, and device parameters of the current radio frequency system are obtained. The weighted feature components are then fused with the system parameters through a multi-dimensional parameter fusion network to construct the state vector containing radio frequency environment information.

8. The low-noise down-conversion signal processing method as described in claim 1, characterized in that, The process of monitoring environmental change parameters through a meta-learning network, adjusting the network weights of the RF sensing deep reinforcement learning network based on performance feedback from the gain control strategy, and outputting updated network weights includes: The environmental change parameters of the radio frequency system are monitored in real time through a meta-learning network. These environmental change parameters include signal strength change, spectrum interference change, and system noise change. When the environmental change parameters exceed a preset threshold, a performance evaluation mechanism is activated. Based on the actual control effect of the gain control strategy in the current environment, the performance feedback index is calculated through the strategy performance evaluation module. The network weights of the RF sensing deep reinforcement learning network are adjusted according to the performance feedback index using a gradient update algorithm, and the updated network weights are output.

9. The low-noise down-conversion signal processing method as described in claim 1, characterized in that, The process involves adjusting the RF sensing deep reinforcement learning network by updating network weights to obtain an updated RF sensing deep reinforcement learning network. Based on this updated network, a final gain control strategy is generated. This final gain control strategy is then output to a variable gain amplifier. The variable gain amplifier adjusts the gain of the I / Q baseband signal according to the final gain control strategy, outputting a low-noise gain-controlled I / Q baseband signal. This includes: The parameters of the radio frequency sensing deep reinforcement learning network are updated using a weight smoothing update mechanism, and the updated network weights are fused with the network weights using an exponential weighted average algorithm. The updated RF sensing deep reinforcement learning network generates the final gain control strategy based on the weighted update. The final gain control strategy is converted into the control command format of the variable gain amplifier through the gain strategy output interface, and real-time gain adjustment control is executed.

10. A low-noise down-conversion signal processing system, used to execute a low-noise down-conversion signal processing method as described in any one of claims 1-9, characterized in that, The system includes: A dual-branch feature extraction module is used to acquire the down-converted I / Q baseband signal and input the I / Q baseband signal into a dual-branch decoupled convolutional neural network. The dual-branch decoupled convolutional neural network includes a signal branch and a mirror branch. The I / Q baseband signal is processed by the signal branch and the mirror branch respectively to obtain the signal feature vector and the mirror feature vector. The state vector construction module is used to decompose the signal feature vector and the mirror feature vector into real components, imaginary components, amplitude components and phase components, and construct the state vector by combining the system temperature parameters, frequency parameters and device parameters. The radio frequency sensing decision module is used to input the state vector into the radio frequency sensing deep reinforcement learning network and output a gain control strategy based on the state vector. The adaptive learning module is used to monitor environmental change parameters through the meta-learning network, adjust the network weights of the radio frequency sensing deep reinforcement learning network based on the performance feedback of the gain control strategy, and output updated network weights. The gain control execution module is used to adjust the radio frequency sensing deep reinforcement learning network by updating the network weights to obtain an updated radio frequency sensing deep reinforcement learning network, generate a final gain control strategy based on the updated radio frequency sensing deep reinforcement learning network, output the final gain control strategy to a variable gain amplifier, and the variable gain amplifier adjusts the gain of the I / Q baseband signal according to the final gain control strategy to output a low-noise gain-controlled I / Q baseband signal.

Images