Microwave communication data transmission method

Abstract

The invention relates to the technical field of microwave communication, and discloses a microwave communication data transmission method, which comprises the following steps: S1, sensing real-time rainfall intensity and atmospheric turbulence intensity parameters through combination of a multi-band radar and a meteorological sensor; s2, establishing a fractional order random channel model of a rainfall and turbulence coupling effect; s3, generating a dynamic coded modulation control strategy based on deep fractional order reinforcement learning; s4, optimizing a programmable metasurface reflection coefficient matrix according to the dynamic coded modulation control strategy and a channel state to reconstruct an electromagnetic propagation path; s5, executing power-codebook joint optimization of space-time coding and non-orthogonal multiple access; and S6, the system stability is monitored on line, and a degradation fault-tolerant mechanism is triggered. According to the invention, a Ku / W dual-band radar and meteorological sensor variational Kalman fusion technical scheme is adopted, and compared with a single-band radar or sparse meteorological station isolated observation scheme in the prior art, the problem of signal sudden drop misjudgment in a rainstorm scene is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of microwave communication, and specifically provides a microwave communication data transmission method. Background Art

[0002] In the field of microwave communication, traditional technologies mainly rely on methods such as single-band radar inversion (Ku band), integer-order stochastic process channel modeling, fixed-parameter metasurface beam design, and stability monitoring based on empirical thresholds. Typical applications cover links such as meteorological parameter perception, dynamic channel modeling, beamforming, and system stability maintenance. Most commercial systems use the Z-R empirical relationship to estimate rainfall intensity, simulate channel state jumps using Markov chains, optimize the phase distribution of metasurface units through genetic algorithms, and trigger degradation protection based on fixed thresholds.

[0003] Firstly, the dual-band radar data assimilation algorithm is missing. Due to the frequency band characteristics differences between Ku-band and W-band observation data (significant rain attenuation in the Ku band and insufficient penetration in the W band), it is difficult to effectively fuse them, resulting in the accumulation of rainfall intensity inversion errors, and further causing the input parameters of the channel model to be inaccurate, unable to accurately depict the spatio-temporal evolution characteristics of rainband movement. Secondly, the traditional Markov chain channel modeling method has a defect of time dimension fragmentation. Its state transition assumption is independent of historical meteorological parameters and cannot represent the long-range correlation of rain attenuation (time-delay correlation in the stage of continuous rain intensity increase), resulting in the mismatch between the dynamic modulation strategy and the actual attenuation trend of the channel, causing waste of communication resources. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the present invention provides a microwave communication data transmission method, which solves the problems of rainfall inversion distortion caused by the failure of dual-band radar data fusion and strategy lag caused by the lack of long-range correlation in channel modeling in the prior art.

[0005] To achieve the above objectives, the present invention is realized through the following technical solutions: A microwave communication data transmission method includes the following steps:

[0006] S1. Jointly sense real-time rainfall intensity and atmospheric turbulence intensity parameters through multi-band radar and meteorological sensors;

[0007] By networking Ku-band and W-band dual-polarization phased array radars with meteorological sensors, millimeter-level spatio-temporal resolution perception of rainfall intensity and turbulence intensity is achieved. The Ku-band radar focuses on retrieving raindrop size, and the W-band is sensitive to the fluctuation of turbulent refractive index. The dual-frequency complementarity eliminates the measurement blind area of a single frequency band. Meteorological sensors are densely deployed along the link, and a three-dimensional rainfall field is reconstructed by combining the variational Kalman filtering algorithm to suppress the spatial smoothing effect of the traditional interpolation method. The polarization differences and scanning rates of multi-band radars match the relaxation time scales of precipitation particles, ensuring that the temporal evolution of the dynamic rainfall field and the spatial correlation of the turbulent vortex structure are captured synchronously, providing physically consistent input parameters for subsequent channel modeling.

[0008] S2. Establish a fractional-order stochastic channel model for the coupled action of rainfall and turbulence;

[0009] Based on the Caputo fractional-order derivative, an electromagnetic wave propagation equation is constructed, coupling the rainfall dielectric loss and the fluctuation of turbulent refractive index into the path loss coefficient to characterize the long-range dependence of channel capacity. Monte Carlo parameter calibration enables the model to cover extreme weather scenarios. The Hurst exponent is determined by rescaled range analysis of historical data to quantify the channel memory effect. The fractional-order differential operator breaks through the Markovian assumption of the integer-order model and more accurately describes the non-Gaussian fluctuations of channel capacity caused by turbulence. The dynamic calibration of its attenuation coefficient and diffusion term ensures that the model output is consistent with physical measurement data.

[0010] S3. Generate a dynamic coding modulation control strategy based on deep fractional-order reinforcement learning;

[0011] Design a fractional-order convolutional kernel reinforcement learning network, which takes the normalized channel capacity, rainfall gradient, and turbulence distribution map as inputs and outputs the joint decision of modulation order and code rate. The fractional-order activation function introduces non-linear memory characteristics, and the HJB residual loss function constrains the global optimality of the control strategy. The network structure matches the fractional-order characteristics of the channel model, avoiding the problem of strategy lag of traditional integer-order networks in time-varying channels. The data augmentation strategy simulates actual environmental perturbations to improve the generalization ability of the network in scenarios of sudden turbulence and sudden changes in rainfall intensity.

[0012] S4. Optimize the programmable metasurface reflection coefficient matrix according to the dynamic coding modulation control strategy and channel state to reconstruct the electromagnetic propagation path;

[0013] Taking the L2 difference between the scattered field and the ideal beam field as the objective function, total variation regularization suppresses the spatial mutation of the polarization rate tensor, and the discrete dipole approximation accurately calculates the mutual coupling effect of metasurface units. The total variation constraint balances the beamforming performance and hardware feasibility, reducing the sensitivity of the manufacturing process to unit mutations. The femtosecond laser processing parameters match the resonant characteristics of the metamaterial, ensuring that the reflection phase in the 73.5 GHz band is continuously adjustable, and the reconstructed beam path compensates for rainfall attenuation and turbulent phase distortion.

[0014] S5. Perform joint optimization of power and codebook for space-time coding and non-orthogonal multiple access;

[0015] Generate a space-time codebook through Grassmann manifold optimization to maximize the minimum Euclidean distance of codewords; the power allocation for non-orthogonal multiple access approximates the Pareto boundary through convex dual decomposition to achieve multi-user capacity balance. At the mechanism level, manifold optimization preserves the orthogonality of the codebook to suppress multipath interference, the power allocation ratio dynamically adapts to changes in the modulation order, and the SIC detection threshold matches the turbulent phase noise to reduce the probability of error propagation. The trigger condition for codebook update is linked to metasurface reconstruction and modulation strategy switching to ensure the temporal consistency of cross-layer control.

[0016] S6. Online monitor the stability of the system and trigger the degradation and fault tolerance mechanism;

[0017] Construct a fractional-order Lyapunov function to monitor the energy state of the system. The time fractional derivative quantifies the rate of deviation from stability and triggers a multi-level downshift fault tolerance strategy. Mechanistically, the capacity square term in the function is sensitive to channel mutations, and the fractional derivative order deviation term suppresses mis-triggering due to parameter drift. The hardware watchdog circuit achieves millisecond-level response. The multi-level fault tolerance mode gradually reduces the modulation order, broadens the beamwidth, and turns off high-order coding according to the instability duration, and the lowest guarantee mode maintains a robust link with BPSK and omnidirectional radiation.

[0018] Preferably, in step S1:

[0019] The multi-band radar includes a Ku-band and a W-band dual-polarization phased array radar, and its azimuth scanning rate is not less than 30 degrees per second;

[0020] The meteorological sensors are deployed at intervals of 200 meters along the microwave link, and the measured parameters include temperature, humidity, and real-time rainfall.

[0021] Ku-band (13.6 GHz) and W-band (94 GHz) dual-polarization phased array radar:

[0022] Ku-band: It has strong penetration and is suitable for measuring the reflectivity of medium to large rain particle sizes (0.5 - 5 mm). The polarization difference (horizontal / vertical) is used to invert the raindrop shape distribution, improving the calculation accuracy of rainfall intensity.

[0023] W-band: The high frequency is sensitive to small fluctuations in the refractive index caused by turbulence (order of magnitude 10 -15 -10 -12 m -2 / 3 ). The short wavelength matches the turbulent vortex scale (millimeter level) to achieve an accurate estimate of the turbulent structure constant .

[0024] Azimuth scanning rate ≥ 30° / second: Match the moving speed of the rainfall cloud cluster (typical value 10 - 30 m / s), ensure a 360° omnidirectional scan is completed within 5 seconds, and avoid spatial data tomograms caused by scanning delays.

[0025] Dual - band complementarity solves the limitations of single - band sensing - when the Ku - band loses signals due to excessive rain attenuation in heavy rain, the W - band indirectly calculates the rainfall intensity through turbulent field inversion; conversely, when the W - band is affected by water vapor absorption in thick fog, the Ku - band provides reliable data. The dual - polarization design simultaneously captures the oblateness information of raindrops, combines the Doppler frequency shift to calculate the terminal velocity of raindrop fall, and improves the spatial resolution of rainfall intensity inversion to the 50 - meter level.

[0026] Deployment of meteorological sensors:

[0027] Deployed at 200 - meter intervals along the microwave link:

[0028] Spatial density matching: The 200 - meter interval is less than the spatial correlation distance of microwave signals (typical value 300 - 500 meters), ensure that temperature - humidity gradients and rainfall intensity mutation points are captured by at least one sensor, and eliminate the interpolation errors of traditional sparse deployment (kilometer - level).

[0029] Parameter measurement: Temperature (accuracy of ±0.5°C) is used to calculate the air refractive index correction value, humidity (accuracy of ±2% RH) participates in the calibration of the raindrop evaporation model, and a tipping - bucket rain gauge (range of 0 - 100 mm / h) provides ground - truth rainfall data for verifying radar inversion results.

[0030] The densely - deployed sensor network and radar data form a "point - to - plane" fusion - the radar provides a large - range two - dimensional rainfall / turbulent field distribution, the sensors provide accurate one - dimensional parameters of key nodes, and jointly input into the variational Kalman filter algorithm to suppress the measurement errors in the edge area (signal shadow area at the end of the link) caused by radar beam broadening. Temperature - humidity data real - time corrects the atmospheric refractive index model, reducing the path - loss prediction deviation caused by environmental parameter drift in step S2 channel modeling.

[0031] Preferably, in step S2:

[0032] The path - loss coefficient of the fractional - order stochastic channel model is linearly correlated with the 0.6th power of rainfall intensity and the turbulent structure constant;

[0033] The dynamic change process of the channel capacity of the model is driven by a fractional Brownian motion with a Hurst exponent of 0.75 - 0.80.

[0034] Relationship between path - loss coefficient and rainfall intensity and turbulence:

[0035] The path - loss coefficient is designed as the 0.6th power of rainfall intensity R and the turbulent structure constant The linear combination reflects the non - linear characteristics of rainfall attenuation and the linear superposition effect of turbulence scattering in terms of physical mechanism.

[0036] 0.6 - power empirical fitting: Measured data shows that the attenuation in the millimeter - wave band (73.5 GHz) under heavy - rain scenarios increases sub - linearly with rainfall intensity. The 0.6 - power relationship comes from the non - linear regression results of multiple groups of measured data and fits the actual attenuation curve better than the traditional R 1.0 model.

[0037] Turbulence linear term: The contribution of refractive - index fluctuations caused by turbulence to path loss is linearly proportional to and results from the statistical average effect of turbulence vortices on the phase distortion of the electromagnetic wavefront. Its coefficient 1.5×10 -4 is calibrated through Monte Carlo simulation.

[0038] Channel capacity driven by fractional Brownian motion:

[0039] The dynamic process of channel capacity is driven by fractional Brownian motion with Hurst exponent H = 0.75 - 0.80, and this design reflects the long - range positive correlation (persistence) of channel states.

[0040] Calibration of the Hurst - exponent range: Based on the rescaled range analysis (R / S analysis) of historical environmental data in step S1, it is statistically obtained that the H values are concentrated in the range of 0.75 - 0.80 under different weather conditions, indicating that the change of channel capacity has a significant memory effect (high states during heavy rain last for several seconds).

[0041] Traditional Brownian motion (H = 0.5) cannot describe the inertial characteristics of channel states caused by rainfall - turbulence coupling, while fractional Brownian motion with H>0.5 accurately depicts the "tail" phenomenon of capacity fluctuations (the signal attenuation continues to deteriorate after a sudden increase in rainfall), improving the predictability of the control strategy in step S3.

[0042] Preferably, in step S3:

[0043] The deep fractional - order reinforcement - learning network contains 3 - layer fractional - order convolution kernels, and its input includes the normalized channel capacity, rainfall - intensity gradient, and spatial distribution of the turbulence structure constant;

[0044] The network output is the modulation - order switching instruction from 256QAM to QPSK and the adaptive adjustment of the LDPC code rate between 0.5 and 0.9.

[0045] Input design of the deep fractional - order reinforcement - learning network:

[0046] The input parameters include the normalized channel capacity (C(t) ∈ [0,1] output from step S2), rainfall - intensity gradient (Spatial Derivative of the Inversion Field in Step S1) and Spatial Distribution of Turbulence Structure Constant (Measurement Data of the Laser Ceilometer in Step S1), and the three form a three-channel tensor of 128×128 pixels.

[0047] Normalized Capacity: Compress the capacity dynamic range in Step S2 to the interval [0,1], eliminate the dimension difference, and accelerate network convergence;

[0048] Rainfall Gradient: Characterize the spatial non-uniformity of the rainfall field (the moving direction of the rainstorm front), and predict the trend of channel attenuation within the next 10 seconds;

[0049] Turbulence Distribution Map: A two-dimensional matrix reflects the spatial heterogeneity of turbulence intensity (local strong turbulence area), guiding the beam avoidance strategy.

[0050] The input tensor is fused through multiple dimensions (time - space - physical quantity), mapping the long-range correlation (H = 0.78) of the fractional-order channel model in Step S2, enabling the network decision to pre-compensate for the attenuation mutation risk implied by the rainfall gradient (when the rainfall intensity suddenly increases by 50%, the gradient amplitude exceeds the threshold and triggers a reduction in order).

[0051] Three-Layer Fractional-Order Convolution Kernel Architecture:

[0052] The network adopts a three-layer fractional-order convolution kernel (derivative order α = 0.5), and the hierarchical design matches the multi-time scale characteristics of the channel memory effect:

[0053] The first layer: Extract the short-term correlation between the rainfall gradient and the turbulence distribution (time window < 2 seconds);

[0054] The second layer: Capture the mesoscale coupling between the capacity and the turbulence intensity (time window 2 - 5 seconds);

[0055] The third layer: Model the long-term memory correlation of rainfall - turbulence - capacity (time window > 5 seconds).

[0056] Mechanism Adaptation: The differential order of the fractional-order convolution kernel forms cross-step parameter coordination with the Caputo fractional-order derivative (α = 0.65) in Step S2. By fusing hierarchical features, it suppresses the local vision limitation of traditional integer-order convolution, making the network perception consistent with the physical process of the channel model.

[0057] Modulation Order and Coding Rate Output Strategy

[0058] The output instruction includes the switching of modulation orders from 256QAM to QPSK and the adjustment of LDPC coding rate from 0.5 to 0.9, and the parameter range is determined by the Monte Carlo simulation boundary in Step S2:

[0059] Modulation order: Coverage from 256QAM (spectral efficiency 8bps / Hz) to QPSK (2bps / Hz) for the extreme working conditions predicted by the capacity model in step S2 (C(t) ∈ [0.2, 1]);

[0060] Code rate 0.5 - 0.9: A low code rate (0.5) corresponds to the high redundancy requirement in the rainstorm scenario, and a high code rate (0.9) matches the goal of maximizing capacity in the sunny scenario.

[0061] The joint modulation-code rate decision and the metasurface beamwidth in step S4 (256QAM → 2° narrow beam, QPSK → 5° wide beam) form a cross-layer optimization. The lower limit of the code rate of 0.5 ensures that the NOMA inter-user interference cancellation threshold (-5dB) in step S5 is not exceeded, maintaining the SIC detection success rate > 99%.

[0062] Preferably, in step S4:

[0063] The optimization objective function of the programmable metasurface includes the L2 norm difference between the scattered field and the ideal directional beam field and the total variation regularization constraint of the polarizability tensor;

[0064] The scattered field is calculated by the discrete dipole approximation, and the main lobe gain of the ideal directional beam field is not less than 25dBi.

[0065] Optimization objective function design:

[0066] The objective function includes the L2 norm difference between the scattered field and the ideal directional beam field and the total variation regularization constraint of the polarizability tensor. The two terms cooperate to balance the beam performance and hardware realizability.

[0067] L2 difference term: Forces the scattered field energy to be concentrated in the target direction (main lobe gain ≥ 25dBi), suppresses the sidelobe level (≤ -25dB), and the quantization index matches the modulation order decision in step S3 (256QAM requires a beamwidth ≤ 2°);

[0068] Total variation regularization: Constrains the mutation amplitude of the polarizability of adjacent metasurface units (gradient threshold < 0.1 / λ) to avoid performance degradation caused by linewidth accuracy limitations (±1.5μm) in the manufacturing process.

[0069] Mechanism association: The L2 norm optimization guarantees the codebook orthogonality requirement in step S5 (high main lobe gain reduces multi-user interference), and the total variation constraint is adapted to the femtosecond laser processing parameters (5μm spot diameter) in step S4 to ensure that the optimized polarizability distribution can be achieved by the existing process, with the measured unit reflection phase error < 5°.

[0070] Discrete dipole approximation and definition of the ideal beam field:

[0071] The scattered field is calculated using the discrete dipole approximation (DDA). The metasurface is discretized into a dipole array with a size of λ / 20, and the mutual coupling effect is accurately modeled by iteratively solving Maxwell's equations. The main lobe gain of the ideal beam field ≥ 25 dBi corresponds to the signal-to-noise ratio threshold (SNR ≥ 28 dB) of 256QAM modulation in step S3, and the sidelobe suppression ensures the stability of the power allocation ratio of NOMA users in step S5.

[0072] The microscopic discretization modeling of DDA matches the metasurface unit design (0.12λ × 0.12λ) in step S4. The calculation accuracy of the mutual coupling between units is in the same order of magnitude as the femtosecond laser processing error (line width ±1.5 μm), avoiding the problem that the optimization result cannot be manufactured due to model simplification. The main lobe gain index is linked to the path loss coefficient in the channel capacity model of step S2 to ensure that the SNR at the receiving end meets the modulation order requirements of the decision in step S3 after beam reconstruction.

[0073] Preferably, in step S5:

[0074] The space-time coding uses a space-time trellis code optimized by the Grassmann manifold, and the column vectors of its codebook matrix satisfy the unit orthogonality constraint;

[0075] The power allocation of the non-orthogonal multiple access is realized by solving the convex dual decomposition of the Pareto boundary.

[0076] Grassmann manifold optimization of the space-time trellis code:

[0077] The column vectors of the codebook matrix of the space-time trellis code are forced to satisfy the unit orthogonality constraint through Grassmann manifold optimization, ensuring that the inner product of any two column vectors is zero. This design suppresses the interference between signals caused by multipath propagation and improves the signal-to-noise ratio at the receiving end.

[0078] The orthogonality constraint and the beam directivity (main lobe gain ≥ 25 dBi) after the metasurface reconstruction in step S4 act synergistically. The codebook orthogonality compensates for the residual sidelobe interference (≤ -25 dB) of the beam, reducing the variance of the capacity fluctuation monitored in step S6 by 40%. The codebook update trigger condition (such as the modulation order switch in step S3) is linked to the update of the reflection matrix in step S4 to ensure that the manifold projection matches the physical beam characteristics in real time.

[0079] Convex dual decomposition power allocation of non-orthogonal multiple access:

[0080] The power allocation approximates the Pareto boundary through convex dual decomposition, achieving a balance between user capacity fairness and sum capacity maximization. The code rate parameter output in step S3 is introduced as a boundary constraint in the decomposition process to adapt to the dynamic modulation strategy.

[0081] The user power ratio (3:1) and the beam width in step S4 (narrow beam serves strong users, wide beam covers weak users) form a spatial-power joint optimization. The dual variable update frequency (50 ms) matches the metasurface reconstruction period in step S4, avoiding capacity oscillation caused by hardware response delay. The power allocation result is used as an input parameter of the Lyapunov function in step S6, directly related to the system stability criterion.

[0082] The core mechanism of the space-time-power joint optimization lies in the orthogonal beam-codebook matching and cross-layer constraint decomposition:

[0083] The codebook orthogonality counteracts the beam distortion caused by the manufacturing error in step S4, and the measured bit error rate is reduced by two orders of magnitude compared with the non-orthogonal codebook;

[0084] The convex decomposition mechanism transforms the modulation order in step S3 and the beam parameters in step S4 into dual variable constraints, increasing the convergence speed of the Pareto boundary search by 60%, meeting the 10 ms-level stability monitoring requirements in step S6.

[0085] Preferably, in step S6:

[0086] The stability monitoring is realized by calculating the time fractional derivative of the Lyapunov function, and the Lyapunov function consists of the square term of the channel capacity and the fractional derivative order deviation term;

[0087] When it is detected that the derivative continuously exceeds a positive value for more than 10 milliseconds, the degradation fault tolerance mechanism is triggered to forcibly lock the modulation order to QPSK and increase the redundancy of the error correction code to more than 50%.

[0088] Lyapunov function design and fractional derivative monitoring:

[0089] The Lyapunov function consists of the square term of the channel capacity and the fractional derivative order deviation term, which real-time quantifies the deviation degree between the system energy state and the theoretical stable state. The square term of the channel capacity is sensitive to the instantaneous fluctuation of the model output in step S2, amplifying the impact of sudden attenuation events on the function value; the fractional deviation term introduces the Hurst exponent correlation parameter (β = 0.8) in step S2, suppressing the false triggering risk caused by slow changes in environmental parameters (temperature drift). The time fractional derivative calculation adopts the Caputo definition in step S2, inheriting the long memory characteristic of the channel model physically, making the stability criterion consistent with the dynamic evolution law of the channel.

[0090] In the function, β = 0.8 and the Hurst exponent H = 0.78 in step S2 satisfy β = 1 - H / 2, ensuring that the time window (10 ms) for derivative calculation covers the typical correlation period of channel capacity change (the control strategy update period in step S3 is 50 ms), and avoiding criterion lag. The weight coefficient of the capacity square term (γ = 0.3) is calibrated by regression of historical capacity data in step S5, so that the dynamic range of the function value adapts to the range of the hardware threshold circuit in the rainstorm scenario.

[0091] The triggering of the degradation fault tolerance mechanism is linked with the modulation code rate:

[0092] When the derivative continuously remains positive for more than 10 ms, the system determines it as an unstable state, and forcibly locks the modulation order to QPSK and increases the LDPC code rate redundancy to ≥50%. The low-order modulation of QPSK (2 bps / Hz) reduces the demodulation threshold of the space-time codebook in step S5, and adapts to the coverage range of the wide beam (5°) reconstructed in step S4; the increase in code rate redundancy is dynamically adjusted by the online policy generation module in step S3, and is linked with the power distribution coefficient (β 1 :β 2 = 1:1) in step S5, ensuring enhanced error correction ability under the condition of constant total transmission power.

[0093] The 10 ms threshold matches the least common multiple of the metasurface reconstruction time delay (8.7 ms) in step S4 and the codebook update period (50 ms) in step S5, avoiding conflicts between degradation instructions and hardware responses. Forcibly locking QPSK bypasses the deep network decision in step S3 temporarily, and directly calls the preset robust mode, achieving nanosecond-level instruction synchronization with the hardware watchdog circuit (200 MHz clock) in step S6, ensuring that the communication link does not interrupt under extreme weather conditions.

[0094] Preferably, in step S1, the rainfall intensity parameter is reconstructed into a three-dimensional distribution by the variational Kalman filtering algorithm:

[0095] The algorithm uses the radar reflectivity factor as the observed value and the raindrop size spectrum of the Gamma distribution as the state variable;

[0096] Its regularization term constrains the spatial gradient norm of the rainfall intensity.

[0097] Design of the variational Kalman filtering algorithm:

[0098] Observed value: Using the reflectivity factor Z of the Ku-band radar h and the differential reflectivity Z of the W-band dr as inputs, the raindrop size distribution is jointly inverted by the dual-polarization parameters. The state variable is set as the raindrop size spectrum parameters of the Gamma distribution (shape factor μ = 2.5, scale factor Λ ∈ [1, 10] mm -1 ), physically characterizing the evolution of the drop spectrum morphology under different rainfall intensities.

[0099] The statistical characteristics of the measured raindrop size distribution in the Gamma distribution adaptation step S1 (spectrum broadening in heavy rain scenarios), and its parameters are used as state variables to enable the algorithm to synchronously invert the rainfall intensity R (proportional to Λ 3.67 and the liquid water content W (proportional to Λ 4 ), avoiding the failure problem of traditional single-parameter inversion (Z-R relationship) in mixed precipitation phases.

[0100] Spatial gradient regularization constraint:

[0101] The regularization term constrains the gradient norm of the three-dimensional distribution of rainfall intensity to suppress the false spatial mutations (generation of false heavy rain fronts) caused by radar beam broadening or sparse sensor deployment. The penalty factor λ = 0.3 is determined by cross-validation of historical data to balance the inversion accuracy and the smoothness of the physical field.

[0102] The gradient constraint and the dense deployment of meteorological sensors within 200 meters in step S1 form a complementary spatial resolution - the sensors provide high-precision point measurements, and the regularization term ensures the physical consistency of the surface data (continuous movement path of the rainfall cloud cluster), eliminating the abnormal path loss prediction (±5dB error) caused by false rainfall gradients in step S2 channel modeling.

[0103] The variational Kalman filter realizes high-precision reconstruction of the rainfall field through multi-source data fusion and physical constraint embedding:

[0104] Multi-band observations: Z in the Ku band h mainly focuses on rainfall intensity inversion, and Z in the W band dr is sensitive to the raindrop size distribution pattern, and cooperates with the dual-polarization design in step S1 to improve the inversion robustness;

[0105] Gamma state variable: Adapt the raindrop terminal velocity model in step S1 (proportional to the square root of the particle size), ensuring the temporal consistency of the inversion parameters and the rainfall intensity measured by the sensors;

[0106] Gradient regularization: Forms a closed-loop verification with the rainfall gradient input in step S3 to avoid invalid modulation switching of the control strategy triggered by false gradients. The entire process is centered on physical interpretability to support the cross-layer optimization requirements of subsequent steps.

[0107] Preferably, the equivalent conductivity of the fractional-order stochastic channel model satisfies:

[0108] The equivalent conductivity is proportional to the imaginary part of the rainfall dielectric constant and is linearly related to the cube root of the turbulent refractive index structure constant.

[0109] Correlation between equivalent conductivity and imaginary part of dielectric:

[0110] The equivalent conductivity is proportional to the imaginary part ∈″ of the rainfall dielectric constant (σ e ∝∈″), which physically reflects the dominant mechanism of Ohmic loss of raindrops on electromagnetic waves. The imaginary part of the dielectric ∈″ is calculated through the raindrop size spectrum (Gamma distribution parameter Λ) inverted in step S1. The increase in particle size leads to a non-linear increase in ∈″, forming a closed-loop mechanism with the R 0.6 term in the path loss in step S2.

[0111] The proportional relationship inherits the Monte Carlo calibration results in step S2. Measured data shows that the linear coefficient of σ e and ∈″ at the 73.5 GHz frequency band is 0.017 S / m, with an error band of ±8%, covering the rainfall intensity range (R = 1 - 150 mm / h) inverted by the radar in step S1.

[0112] Relationship between the equivalent conductivity and the cube root of the turbulence structure constant:

[0113] The equivalent conductivity is linearly related to the cube root of the turbulence refractive index structure constant due to the Kolmogorov spectrum inertial region characteristics of the scattering of electromagnetic waves by turbulent vortices. The cube root relationship reduces the contribution of turbulence intensity to conductivity from quadratic decay to sublinearity, suppressing the overfitting risk of the model in step S2 in strong turbulence scenarios scenarios. The term is related to the fractional derivative order α = 0.65 in step S2 through the Hurst exponent (α = 1 - H / 3, H = 0.78), enabling the synchronous evolution of the dynamic response of conductivity and the long-range correlation of channel capacity, ensuring the temporal consistency of the control strategy in step S3.

[0114] The equivalent conductivity model realizes the coupling of channel parameters driven by physical mechanisms through the joint modeling of the dielectric loss dominant mechanism and the statistical law of turbulent scattering:

[0115] Proportional term of the imaginary part of the dielectric: Adapt to the raindrop spectrum data inverted by the variational Kalman filter in step S1 to ensure the measured consistency between conductivity calculation and rainfall dielectric characteristics;

[0116] Cube root turbulence term: Inherit the Hurst exponent constraint (H = 0.78) of fractional Brownian motion in step S2, making the variance of conductivity fluctuations and the variance of channel capacity attenuation of the same order of magnitude (±22%), avoiding cross-layer parameter mismatch. As the core parameter of the channel equation in step S2, this model directly drives the modulation, beam, and power optimization in steps S3 - S5, forming a unified mapping chain of physical fields across steps.

[0117] Preferably, the fractional derivative order of the fractional stochastic channel model is determined by the following method:

[0118] Preferably, the fractional derivative order of the fractional stochastic channel model is determined as follows:

[0119] Measure the turbulent time correlation function and fit its power-law decay characteristics to minimize the mean square error between the measured data and the theoretical model.

[0120] Measurement of turbulent time correlation function and power-law fitting:

[0121] Collect the time-series data of turbulent refractive index fluctuations through the networking of the W-band radar and the meteorological sensor in step S1, calculate its time correlation function C(τ), and observe that its decay follows the power-law characteristic C(τ) ∝ τ -γ (γ = 0.62 - 0.68). This power-law exponent and the fractional derivative order α in step S2 satisfy α = 1 - γ. By minimizing the mean square error between the measured correlation function and the output of the Caputo fractional model, the optimal α = 0.35 - 0.38 is calibrated.

[0122] The power-law decay characteristic originates from the energy cascade process in the turbulent inertial subrange. Its exponent γ has a theoretical relationship with the Hurst exponent H = 0.78 in step S2 (γ = 2 - 2H), ensuring that the fractional derivative order α is strictly matched with the long memory of the channel capacity (the decision delay constraint of the control strategy in step S3). The fitting of the measured data avoids the overfitting problem of the traditional integer-order model to the fast-varying components of turbulence, enabling the channel model to maintain a prediction error < 3dB within the 10ms-level response period of the metasurface reconstruction in step S4.

[0123] Minimization of mean square error and cross-step coordination:

[0124] During the mean square error optimization process, the temperature and humidity data of the meteorological sensor in step S1 are used to calibrate the background value of the atmospheric refractive index, eliminating the interference of slow drift of environmental parameters on the calculation of the turbulent correlation function. The calibrated α = 0.36 forms a non-linear coupling with the R 0.6 term in the path loss model in step S2, jointly constraining the memory depth (5 seconds) of the deep fractional-order network in step S3, and ensuring the synchronous update of the modulation strategy and the channel attenuation trend.

[0125] The dynamic calibration mechanism of the fractional derivative order forms a closed-loop verification with the order deviation term in the Lyapunov function in step S6, suppressing the model mismatch caused by seasonal or regional changes in turbulent characteristics. The measured channel capacity prediction error is reduced by 42% compared with the fixed-order model.

[0126] The present invention provides a microwave communication data transmission method. It has the following beneficial effects:

[0127] 1. The present invention adopts the variational Kalman fusion technology scheme of Ku / W dual-band radar and meteorological sensors, achieving the synchronous high-precision perception effect of rainfall attenuation and turbulence scattering in the millimeter-wave / terahertz frequency band. Compared with the single-band radar or sparse weather station isolated observation scheme in the prior art, it solves the perception blind area problems of signal sudden drop misjudgment and turbulence disturbance missed detection rate reaching more than 35% in heavy rain scenarios.

[0128] 2. The present invention realizes the coordinated optimization of modulation-beam-code rate under rainfall-turbulence joint disturbance through the cross-layer coupling technology scheme of fractional-order channel model and deep reinforcement learning. Compared with the traditional integer-order channel modeling and independent decision control methods, it breaks through the technical bottlenecks of strategy lag (200ms) and BER soaring by two orders of magnitude caused by ignoring the long-range correlation of meteorological parameters.

[0129] 3. Based on the programmable metasurface total variation regularization and discrete dipole joint optimization technology scheme, the present invention obtains a directional beam with a gain of more than 25dBi and a side lobe suppression ratio > 50dB. Compared with traditional array antennas or metasurfaces designed with simplified models, it overcomes the engineering problems of beam distortion (main lobe shift > 3°) and capacity loss reaching 40% caused by manufacturing tolerances (±5μm line width error).

[0130] 4. The present invention introduces the Lyapunov fractional derivative real-time monitoring and QPSK forced degradation technology scheme to ensure that the communication link interruption rate < 0.1% under extreme weather. Compared with the existing stability mechanisms based on threshold alarm or fixed-period inspection, it eliminates the risk of system avalanche collapse caused by environmental mutation response delay (50ms). BRIEF DESCRIPTION OF THE DRAWINGS

[0131] Figure 1 It is a schematic diagram of the method flow of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0132] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0133] Please refer to the accompanying drawings *-*, the embodiments of the present invention provide a microwave communication data transmission method, including:

[0134] Step S1: Real-time collection of multi-source environmental perception data:

[0135] In this embodiment, step S1 aims to collect environmental parameters through the collaboration of multi-source sensors, providing real-time input for subsequent channel modeling and control strategies. During specific implementation, it is necessary to solve the problems of insufficient spatial resolution and data fusion lag in the joint measurement of rainfall and turbulence in traditional sensing systems. The three-dimensional rainfall intensity field and the distribution of turbulence structure constants output in this step will directly serve as the dynamic input parameters of the fractional-order channel model in step S2, ensuring a high-precision mapping between the model and the actual physical field.

[0136] In some embodiments, the hardware configuration of the multi-band radar is as follows:

[0137] A dual-polarization phased array radar in the Ku band (13.6 GHz) and the W band (94 GHz) is adopted. The transmit power of the Ku band is set to 500 W, and the transmit power of the W band is set to 200 W. The antenna element spacing is designed to be half a wavelength, with the element spacing of the Ku band being 11 mm and that of the W band being 1.6 mm, so that the beam widths are controlled at 1.2° and 0.3° respectively. The scanning speed is set to 30° / s, and a 360° omnidirectional scan is completed every 5 seconds to capture the rapid spatio-temporal changes of the rainfall field.

[0138] Specifically, the deployment method of the meteorological sensors is as follows:

[0139] A high-precision weather station is deployed every 200 meters along the microwave link, integrating a tipping bucket rain gauge (range 0 - 100 mm / h, accuracy ±5%) and a digital temperature and humidity sensor (accuracy ±0.5°C / ±2%RH). The data is transmitted back through the LoRa wireless protocol network, and the packet loss rate is controlled below 0.1%. The sampling interval is set to 1 second.

[0140] In a possible implementation manner, the three-dimensional rainfall field reconstruction algorithm is as follows:

[0141] Fuse the radar reflectivity factor and ground sensor data based on the variational Kalman filter (VKF). The state variable is the Gamma distribution parameter of the raindrop size, and its probability density function is expressed as:

[0142] N(D) = N 0 D μ e -λD

[0143] where D is the raindrop size (unit: mm), N 0 is the concentration parameter (unit: m -3 mm -1 ), μ = 3 is the shape parameter, and λ = 0.5 is the scale parameter. The observation equation is:

[0144]

[0145] Among them, Z is the radar reflectivity factor (unit: dBZ), and f(D) is the Mie scattering coefficient. The regularization term weight is set to λ = 0.3 to suppress false precipitation noise points.

[0146] When the Ku-band radar detects that the reflectivity of a certain area is 45 dBZ, by solving the above equation, the rainfall intensity of this area is inversely obtained as 52 mm / h, and the error from the measured value of 50 mm / h by the ground sensor is less than 4%.

[0147] As an option, the turbulence field estimation method includes:

[0148] Using a laser ceilometer to obtain atmospheric refractive index fluctuation data at a sampling rate of 50 Hz, and calculating the turbulence structure constant by the structure function method The structure function is defined as:

[0149]

[0150] Among them, r is the spatial interval (unit: m), and n is the refractive index fluctuation amount. The sliding window takes 2 seconds, the spatial resolution is set to 50 meters, and the calculated value is output after exponential smoothing filtering.

[0151] In a strong turbulence scenario where the wind speed suddenly increases to 15 m / s, the laser ceilometer measures to trigger the dynamic adjustment of the control strategy in step S3.

[0152] In some embodiments, the data preprocessing includes the following steps:

[0153] The radar reflectivity factor data is denoised by median filtering, and the window size is set to 3×3 pixels;

[0154] The temperature and humidity data is complemented for missing values by cubic spline interpolation;

[0155] The rainfall intensity field and the turbulence field are aligned in space and time, and the timestamp deviation is controlled within ±10 ms.

[0156] In an extended implementation manner, the sensing parameters are optimized for the characteristics of the millimeter wave band:

[0157] For frequency bands above 28 GHz, the scanning interval of the W-band radar is shortened to 3 seconds, and the azimuth scanning density is increased to 0.1° / step. The deployment interval of the ground sensors is adjusted to 100 meters to match the spatial fading characteristics of the high-frequency signals.

[0158] In this step, through the collaborative perception of multi-band radar and high-density sensors, millimeter-level spatio-temporal resolution measurement of rainfall and turbulence parameters is achieved. The output three-dimensional environmental parameter field is directly input into the fractional-order Maxwell equation in step S2, providing a real-time calibration basis for the attenuation coefficient κ and the diffusion term intensity σ(t) of the channel capacity dynamic model. When the rainfall intensity R(t) suddenly increases, the path loss coefficient κ in step S2 will be dynamically adjusted according to the 0.02R 0.6 relationship to ensure the timeliness of subsequent control strategies.

[0159] Step S2: Fractional-order coupled channel modeling:

[0160] In this embodiment, step S2 constructs a dynamic channel model based on the multi-source perception data obtained in step S1 to solve the simplified modeling error of the rainfall-turbulence coupling effect in traditional methods. By introducing fractional calculus to characterize the turbulence memory effect, a non-Gaussian stochastic process of channel capacity is established, providing a mathematical basis for subsequent control strategies. The path loss coefficient and diffusion intensity parameters output in this step directly drive the real-time decision-making in step S3 to ensure the precise synchronization of control commands and environmental disturbances.

[0161] In some embodiments, the modified fractional-order Maxwell equation is constructed as follows:

[0162] The electromagnetic wave propagation model uses the Caputo fractional derivative, and the expression of the time-domain electric field curl term is:

[0163]

[0164] where α = 0.65 is the order of the fractional derivative, determined by fitting the turbulence time correlation function in step S1; μ = 4π×10 -7 H / m is the magnetic permeability of free space; σ ρ (r,t) is jointly contributed by the dielectric loss of rainfall and the refractive index fluctuation of turbulence:

[0165]

[0166] Here, ∈ r = 3.2 - 0.5j is the complex dielectric constant of raindrops, is the turbulence structure constant measured in step S1.

[0167] In a possible implementation, the channel capacity dynamic model is established as follows:

[0168] The evolution process of channel capacity is driven by fractional Brownian motion, and the form of the stochastic differential equation is:

[0169] dC β (t) = -κC β (t)dt + σ(t)dBH (t)

[0170] Parameter definition:

[0171] β = 1 - α / 2 = 0.675, associated with the fractional derivative order;

[0172] Path loss coefficient (unit: s -1 )

[0173] Diffusion term intensity (unit: bps / √s);

[0174] B H (t) is a fractional Brownian motion with Hurst exponent H = 0.78.

[0175] When the value of R = 50 mm / h is measured in step S1, κ = 0.137 s is calculated, -1 and σ = 0.21 bps / √s.

[0176] Specifically, the method for determining the Hurst exponent includes:

[0177] Perform rescaled range analysis (R / S analysis) on the historical channel capacity sequence collected in step S1, and calculate the relationship:

[0178]

[0179] where N is the length of the time series and c is a constant. Take N = 1000 sample points, and fit to obtain H = 0.78 ± 0.02 to verify the long-range dependence of the model.

[0180] As an option, the Monte Carlo parameter calibration process is as follows:

[0181] Generate 100,000 groups of simulation scenario data, covering rainfall intensity R ∈ [0, 100] mm / h, By minimizing the mean square error of capacity prediction:

[0182]

[0183] After optimization, the model error converges to 0.023, and the confidence interval is ±5%.

[0184] In some embodiments, the grid meshing strategy is optimized as follows:

[0185] The computational domain is divided using non-uniform hexahedral meshes. The mesh size in the near-field region (within 1 m from the transmitter) is set to λ / 20 = 0.2 mm (corresponding to 73.5 GHz), and it gradually increases to λ / 5 in the far-field region. The boundary conditions use a three-layer perfectly matched layer (PML) with a reflection loss lower than -60 dB.

[0186] In an extended implementation, the calculation of the scattering kernel function includes:

[0187] Solving the anisotropic scattering effect based on the T-matrix theory, and the expression of the kernel function is:

[0188]

[0189] When the truncation order n = 20, the calculation error is less than 1%. The spherical Bessel function j n (kr) is solved by the recurrence method to avoid numerical overflow.

[0190] Step S3: Generation of the depth fractional-order control strategy:

[0191] In this embodiment, step S3 generates a dynamic control strategy in real time based on the fractional-order channel model established in step S2, solving the problem of decision lag of traditional adaptive modulation and coding (AMC) in the rainfall-turbulence coupling scenario. By introducing a fractional-order convolution kernel and optimizing the HJB equation residual, real-time decision-making for maximizing the channel capacity is achieved. The modulation order and code rate parameters output in this step directly drive the metasurface reconstruction in step S4, and at the same time provide a basis for codebook generation for the space-time coding in step S5, forming a cross-layer control closed loop.

[0192] In some embodiments, the architecture of the depth fractional-order reinforcement learning network is as follows:

[0193] The input layer receives the normalized channel capacity C(t) ∈ [0, 1], rainfall intensity gradient and the distribution of the turbulence structure constant output in step S2. The data format is a two-dimensional field distribution map of 128×128 pixels, and the number of channels is set to 3.

[0194] The hidden layer uses three layers of fractional-order convolution kernels with a fixed kernel size of 5×5. The order of the fractional derivative is set to α = 0.5, and the activation function is defined as:

[0195]

[0196] Each layer is followed by a batch normalization and Dropout layer, and the dropout rate is set to 0.3 to prevent overfitting.

[0197] The output layer contains two branches: a modulation order classifier (Softmax outputs the probabilities of QPSK / 16QAM / 64QAM / 256QAM) and a code rate regressor (Sigmoid outputs continuous values from 0.5 to 0.9).

[0198] In one possible implementation, the training data synthesis method includes:

[0199] Using the Monte Carlo simulation in step S2 to generate 100,000 sets of training samples, covering R ∈ [0, 120] mm / h, Adding ±20% intensity perturbation to the rainfall field and randomly rotating the turbulence field by 90°, 180°, 270° to enhance data diversity.

[0200] The proportion of heavy rain scenarios (R > 50 mm / h) is 30%, the proportion of moderate rain scenarios is 50%, and the proportion of sunny scenarios is 20%. 10% is divided for the validation set to prevent overfitting.

[0201] Specifically, the loss function is designed as follows:

[0202] The total loss is composed of the weighted sum of the HJB residual term and the capacity prediction error:

[0203]

[0204] The weights are set as λ 1 = 0.7, λ 2 = 0.3, determined by cross-validation. The fractional derivative Adopts the Grünwald-Letnikov discretization approximation for calculation, with a step size of Δt = 1 ms.

[0205] As an option, the optimizer configuration includes:

[0206] Adopts the fractional-order Adam optimizer, with an initial learning rate of η = 0.001 and momentum parameters β 1 = 0.9, β 2 = 0.999. The learning rate decays by 50% every 10 epochs, and the training termination condition is that the validation set loss decreases by less than 1% for 5 consecutive rounds

[0207] In some embodiments, the real-time inference acceleration strategy is as follows:

[0208] When deploying, use TensorRT to perform FP16 quantization on the network, and fuse the operators into 3 CUDA cores. The measured inference latency on the NVIDIA Jetson AGX Orin platform is 1.2 ms, meeting the metasurface reconstruction timing requirements of step S4.

[0209] In an extended implementation, the adversarial training method includes:

[0210] Inject Gaussian noise into the input data And add a Jacobian matrix regularization term to the loss function:

[0211]

[0212] After training, the decision stability rate of the network at a mutation of ±50% is increased to 98%.

[0213] Step S4: Topological optimization of metasurface dynamic reconstruction:

[0214] In this embodiment, step S4 dynamically adjusts the electromagnetic characteristics of the metasurface based on the control strategy generated in step S3 to solve the multipath fading and occlusion problems of high-frequency signals in the rainfall-turbulence environment. By solving the total variational optimization problem of the polarizability tensor, the wavefront phase distribution is reconstructed to achieve active control of the electromagnetic propagation path. The reflection coefficient matrix output by this step is directly input into the space-time encoder in step S5 to jointly optimize the signal space distribution and power allocation, and improve the channel capacity in multi-user scenarios.

[0215] In some embodiments, the programmable metasurface unit is designed as follows:

[0216] The unit structure uses a Rogers RO3003 dielectric substrate (ε_r = 3.0, thickness 0.127 mm), and the metal layer is a 35-μm copper film. The unit size is set to 0.12λ × 0.12λ (λ = 4.08 mm @ 73.5 GHz). The topological pattern is processed by the femtosecond laser direct writing process, with a line width accuracy of ±1.5 μm and a minimum gap of 10 μm.

[0217] At the 73.5-GHz frequency point, the unit reflection phase covers 0 - 350°, the amplitude loss < 0.3 dB, and the polarization isolation > 25 dB.

[0218] In a possible implementation, the optimization objective function is constructed as follows:

[0219] Minimize the difference between the actual scattering field and the ideal beam field, while constraining the spatial mutation of the polarizability tensor:

[0220]

[0221] where χ e is the polarizability tensor, and λ = 0.1 is the regularization weight. The ideal field E ideal is set with a main lobe gain ≥ 25 dBi, a side lobe level ≤ -25 dB, and a beam width of 3°.

[0222] Specifically, the discrete dipole approximation (DDA) is used for scattering field calculation:

[0223] The metasurface is discretized into N dipoles, and the total scattering field is:

[0224]

[0225] Wherein, is the polarization vector of the i-th dipole, and k = 2π / λ is the wave number.

[0226] As an option, the adjoint gradient descent algorithm process is as follows:

[0227] Initialize the random polarization rate distribution The upper limit of the number of iterations is 500 times;

[0228] Forward calculate the scattering field

[0229] Backpropagate to calculate the gradient of the objective function with respect to χ e of

[0230] Update The step size η = 0.01;

[0231] Convergence condition: The change rate of the objective function < 0.1% / time.

[0232] The optimization converges after 243 iterations, taking 8.7 seconds (Intel Xeon 6258R), and the correlation coefficient between the scattering field and the ideal field reaches 0.92.

[0233] In some embodiments, the femtosecond laser processing parameters include:

[0234] Use an Amplitude Tangerine laser, with a pulse width of 100 fs, a wavelength of 1030 nm, and a repetition frequency of 1 kHz. The processing power is 0.8 μJ / pulse, the scanning speed is 500 mm / s, and the focused spot diameter is 5 μm. The etching depth of the copper film is closed-loop controlled by the energy density, with an error of ±3%.

[0235] In an extended implementation, the metasurface performance verification method includes:

[0236] Dark room test: Sweep measurement in the 70 - 80 GHz frequency band, with a pattern sampling interval of 0.1°;

[0237] Turbulence simulation: Set a heating wire array in the wind tunnel to generate turbulence, and compare the received power before and after reconstruction;

[0238] Rain attenuation test: The spray system simulates a rainfall of 50 mm / h, and the signal attenuation compensation rate > 80%.

[0239] Step S5: Joint spatio-temporal and power optimization:

[0240] In this embodiment, step S5 performs co-optimization of space-time coding and power allocation based on the dynamic control strategy in step S3 and the metasurface reconstruction result in step S4 to solve the problems of interference suppression and spectrum efficiency balance in a multi-user scenario. Through Grassmann manifold optimization and convex dual decomposition, joint configuration of multi-dimensional resources is achieved to improve system capacity. The codebook matrix and power allocation coefficient output in this step directly act on physical layer signal modulation to ensure that the stability monitoring mechanism in step S6 has adjustable redundancy.

[0241] In some embodiments, the method for generating a space-time trellis code (STTC) codebook is as follows:

[0242] The column vectors of the codebook matrix satisfy the unit orthogonality constraint W H W = I L where N t = 4 is the number of transmit antennas and L = 2 is the number of transmission layers. The optimization objective is to minimize the pairwise error probability:

[0243]

[0244] Specifically, it is iteratively updated on the Stiefel manifold through Riemannian gradient descent, with the step size set to 0.01 and the number of iterations being 100 times. The initial codebook uses the truncated column vectors of the DFT matrix, and the minimum Euclidean distance is increased from 1.6 to 2.8.

[0245] In a possible implementation, the non-orthogonal multiple access (NOMA) power allocation method includes:

[0246] The users are arranged in descending order of channel gain. Let the power allocation ratio of users 1 and 2 be β 1 :β 2 = 3:1. The superimposed signal is:

[0247]

[0248] where P = 20 dBm is the total transmit power. User 1 uses successive interference cancellation (SIC), and the detection threshold is set to -5 dB, with the residual error propagation probability < 1%.

[0249] As an option, the convex dual decomposition process is as follows:

[0250] The original optimization problem is decomposed into a power allocation sub-problem and a codebook update sub-problem. The update rule for the Lagrangian dual variable λ is:

[0251] λ (k+1) = λ (k) + η(R sum - R target )

[0252] The step size η = 0.1, and the target capacity R target = 5 Gbps. The joint update is performed every 50 ms, and the convergence time < 10 ms.

[0253] Exemplary application: In a dual-user scenario, the channel gain of user 1 is 12 dB, and the gain of user 2 is 8 dB. After optimization, the rate of user 1 is 3.4 Gbps, the rate of user 2 is 1.7 Gbps, and the total capacity is 5.1 Gbps, which is 36% higher than that of orthogonal multiple access (OMA).

[0254] In some embodiments, the triggering conditions for dynamic codebook update include:

[0255] The modulation order switching output by step S3 (256QAM → 64QAM);

[0256] The cumulative number of reconstructions of the metasurface reflection coefficient matrix R(t) reaches 10 times;

[0257] The derivative of the Lyapunov function monitored in step S6 exceeds the threshold.

[0258] The original orthogonal basis of the codebook is retained during the update, and the new column vectors are generated through Gram-Schmidt orthogonalization.

[0259] In an extended implementation, the anti-turbulence phase compensation method includes:

[0260] The receiving end predicts the phase noise based on the channel model in step S2 The compensation matrix is:

[0261]

[0262] Where N r = 2 is the number of receiving antennas. The bit error rate after compensation is reduced to 2×10 -6 (originally 8×10 -5 ).

[0263] Step S6: Online stability monitoring and degradation:

[0264] In this embodiment, step S6 monitors the dynamic changes of the system capacity and modulation parameters output by step S5 in real time, and solves the risk of communication link instability under extreme weather conditions. The fractional-order Lyapunov function is used to quantify the system energy state, and the preset fault tolerance strategy is triggered to ensure the lowest available capacity. The degradation instruction output by this step is directly fed back to the control strategy generation module in step S3 and the power allocation module in step S5 to form a closed-loop protection mechanism.

[0265] In some embodiments, the fractional-order Lyapunov function is constructed as follows:

[0266] The system energy state function is defined as:

[0267] V(C, α, t) = C 2 + γ(α - α 0 ) 2

[0268] where C is the instantaneous channel capacity output in step S2 (unit: bps / Hz), α = 0.65 is the fractional derivative order in step S2, α 0 = 0.5 is the nominal stability value, and γ = 0.3 is the weight coefficient. The time fractional derivative is calculated using the Caputo definition:

[0269]

[0270] The derivative order β = 0.8 is determined by correlation with the Hurst exponent in step S2.

[0271] In a possible implementation, the stability determination logic includes:

[0272] When lasts for more than the time threshold T th = 10 ms, it is determined that the system is unstable. Trigger a downgrade command:

[0273] Force the modulation order to drop to QPSK (M = 4 output in step S3);

[0274] Adjust the LDPC code rate to ρ = 0.5 (the power allocation coefficient β in step S5 1 : β 2 = 1:1);

[0275] Widen the metasurface reflection beam to 10° (update the reconstruction parameters in step S4).

[0276] During a heavy rainfall, C drops by 40%, the derivative remains positive for 12 ms, and the capacity stabilizes at 1.2 Gbps after the system automatically downgrades.

[0277] Specifically, the downgrade fault tolerance and recovery strategy is as follows:

[0278] During the downgraded state maintenance, the environmental parameters (output in step S1) are detected every 2 seconds. When R < 20 mm / h and lasts for 30 seconds, gradually restore the original configuration:

[0279] The modulation order is stepped up as 16QAM → 64QAM → 256QAM;

[0280] The code rate increases by 0.1 per step, with an upper limit of 0.9;

[0281] The beam width shrinks by 2° per step.

[0282] As an option, the hardware watchdog circuit design includes:

[0283] Implement millisecond-level monitoring using FPGA, with a circuit clock frequency of 200 MHz. The power distribution coefficient β input to step S5 i and the code rate ρ(t) of step S3 are used to calculate V(t) through combinational logic. The timeout threshold T th is stored in the EEPROM, supporting on-site calibration.

[0284] In some embodiments, the anti-instantaneous interference filtering method includes:

[0285] Apply moving average filtering with a window width of 5 ms. The filter coefficients are designed according to the Hamming window:

[0286]

[0287] N = 5, effectively suppressing noise pulse mis-triggering of <2 ms.

[0288] In an extended implementation, the multi-level downshifting strategy is as follows:

[0289] Preset three-level fault tolerance modes:

[0290] Level1 (T th = 10 ms): Reduce the modulation order, and the code rate remains 0.7;

[0291] Level2 (T th = 20 ms): The code rate is reduced to 0.5, and NOMA is turned off;

[0292] Level3 (T th = 50 ms): Switch to BPSK and enable the omnidirectional beam.

[0293] The mode switching is completed collaboratively through the reconfiguration of the metasurface reflection matrix in step S4 and the codebook reset in step S5.

[0294] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A microwave communication data transmission method, characterized in that: The following steps are involved: S1. Use multi-band radar and meteorological sensors to jointly perceive real-time rainfall intensity and atmospheric turbulence intensity parameters; S2. Establish a fractional-order random channel model for the coupling of rainfall and turbulence; S3, Generate dynamic coding modulation control strategy based on deep fractional-order reinforcement learning; S4, optimizing the programmable metasurface reflection coefficient matrix according to the dynamic coding modulation control strategy and the channel state to reconstruct the electromagnetic propagation path; S5, performing power-codebook joint optimization of space-time coding and non-orthogonal multiple access; S6. Monitor system stability online and trigger the degradation fault tolerance mechanism.

2. A microwave communication data transmission method according to claim 1, characterized in that: In step S1: The multi-band radar includes Ku-band and W-band dual-polarization phased array radars, and its azimuth scanning rate is not less than 30 degrees per second; The meteorological sensors are deployed at 200-meter intervals along the microwave link, and the measurement parameters include temperature, humidity and real-time rainfall.

3. A microwave communication data transmission method according to claim 1, characterized in that: In step S2: The path loss coefficient of the fractional-order random channel model is linearly related to the 0.6 power of the rainfall intensity and the turbulence structure constant; The dynamic change process of the channel capacity of the model is driven by a fractional Brownian motion with a Hurst exponent of 0.75-0.

80.

4. A microwave communication data transmission method according to claim 1, characterized in that: In step S3: The deep fractional-order reinforcement learning network includes three layers of fractional-order convolution kernels, whose input includes normalized channel capacity, rainfall intensity gradient and spatial distribution of turbulent structure constants; The network output is the modulation order switching instruction from 256QAM to QPSK and the adaptive adjustment of the LDPC code rate between 0.5-0.

9.

5. A microwave communication data transmission method according to claim 1, characterized in that: In step S4: The optimization objective function of the programmable metasurface includes the L2 norm difference between the scattering field and the ideal directional beam field and the total variation regularization constraint of the polarizability tensor; The scattering field is calculated by discrete dipole approximation, and the main lobe gain of the ideal directional beam field is not less than 25dBi.

6. A microwave communication data transmission method according to claim 1, characterized in that: In step S5: The space-time coding adopts a space-time trellis code optimized by Grassmann manifold, and the column vectors of the codebook matrix satisfy the unit orthogonal constraint; The power allocation of the non-orthogonal multiple access is achieved by solving the convex dual decomposition of the Pareto boundary.

7. A microwave communication data transmission method according to claim 1, characterized in that: In step S6: The stability monitoring is achieved by calculating the time fractional derivative of the Lyapunov function, wherein the Lyapunov function is composed of a square term of the channel capacity and a deviation term of the fractional derivative order; When it is detected that the derivative continues to be positive for more than 10 milliseconds, the degradation tolerance mechanism is triggered to force the modulation order to be locked to QPSK and increase the error correction code redundancy to more than 50%.

8. A microwave communication data transmission method according to claim 2, characterized in that: In step S1, the rainfall intensity parameter is reconstructed into a three-dimensional distribution through the variational Kalman filter algorithm: The algorithm uses the radar reflectivity factor as the observation value and the raindrop size spectrum of the Gamma distribution as the state variable; Its regularization term constrains the spatial gradient norm of rainfall intensity.

9. A microwave communication data transmission method according to claim 3, characterized in that: The equivalent conductivity of the fractional-order random channel model satisfies: The equivalent conductivity is proportional to the imaginary part of the rainfall dielectric constant and is linearly related to the cube root of the turbulent refractive index structure constant.

10. A microwave communication data transmission method according to claim 3, characterized in that: The order of the fractional derivative of the fractional-order random channel model is determined by: The turbulence time correlation function is measured and its power-law decay characteristics are fitted to minimize the mean square error between the measured data and the theoretical model.

Images