Meteorological phased array radar data processing method and system based on edge computing

Abstract

The application discloses a meteorological phased array radar data processing method and system based on edge computing and belongs to the field of radar signal processing. The method comprises the following steps: an edge computing node performs pulse compression, adaptive clutter suppression and Doppler ambiguity resolution processing on original baseband signals to generate basic meteorological parameters; strong weather features are extracted by using computer vision and image recognition technology; a three-dimensional volume scanning data cube is asymmetrically compressed according to the strong weather features, structured messages are generated and reported to a central cloud platform; global meteorological product inversion is performed by the central cloud platform, and scanning strategy control instructions are generated according to the strong weather features and sent to the edge computing node; and the edge computing node drives the radar to adjust scanning parameters according to the instructions. The application realizes real-time processing of meteorological data at the edge and cloud-edge collaborative scheduling, and significantly reduces transmission bandwidth and processing delay.

Description

Technical Field

[0001] This invention relates to the field of radar signal processing, and more specifically to a meteorological phased array radar data processing method and system based on edge computing. Background Technology

[0002] Phased array radar, with its significant advantages such as high spatiotemporal resolution, flexible beam pointing, and rapid scanning, has gradually replaced traditional mechanically scanned radar, becoming a key piece of equipment for detecting sudden meteorological disasters such as severe convective weather, tornadoes, and short-duration heavy rainfall. However, while significantly improving detection accuracy, phased array radar also brings the severe challenge of a geometrically increasing data volume. Under the traditional centralized processing architecture, the massive amounts of raw base data collected by radar stations need to be transmitted to a remote central server for unified processing via a backhaul network. When facing the collaborative detection of ultra-large-scale radar networks, this mode not only puts enormous instantaneous pressure on transmission bandwidth but also causes significant time delays due to long-distance transmission and queuing at central nodes, making it difficult to meet the operational requirements of second-level response for severe convective weather warnings.

[0003] Data processing for meteorological phased array radar involves not only massive computational demands but also extremely high signal processing complexity. Meteorological echoes are often mixed with strong ground clutter, biological clutter, and velocity ambiguity caused by the Doppler effect. Existing edge computing solutions, such as the simple filtering and linear thresholding used in hydrological monitoring, suffer from insufficient algorithmic model depth and are fundamentally incapable of handling the highly nonlinear feature recognition requirements in meteorological data, such as automatic correction of the zero-degree layer bright band and weak signal extraction of tornado vortex features. Under conditions of limited computing resources at edge computing nodes, balancing the precision of the algorithm with real-time processing has become an unresolved inherent contradiction.

[0004] Patent application CN121186786A proposes an adaptive cooperative scanning method for phased array radars in severe convective weather. The method utilizes a central cooperative control module, a data analysis module, and a distributed phased array radar network. Each radar performs a basic scan and identifies regions of interest in severe convection. The data analysis module extracts feature parameters and calculates a dynamic threat index. The central cooperative control module constructs a global state vector and generates scanning task instructions using a multi-agent deep deterministic strategy gradient algorithm. After execution, observational data is fused to generate a three-dimensional meteorological data volume, and algorithm parameters are updated through empirical replay. This scheme achieves adaptive cooperation among multiple radars to a certain extent, but its core processing and decision-making are still centralized in the cloud. The raw echo data still needs to be uploaded to the central module for fusion and identification, failing to solve the bandwidth pressure problem of real-time transmission of massive amounts of data. Furthermore, its scanning strategy optimization mainly focuses on resource allocation and does not involve dynamic adjustment of the original signal processing parameters, thus failing to fundamentally improve data quality under complex meteorological conditions.

[0005] Patent application CN121522577A discloses an adaptive phased array radar and method for detecting severe convective weather. Based on the active electronically scanned array front-end and radar signal and data processing unit, a cognitive control and resource management unit is introduced. By implementing threat assessment (using 3D U-Net to output a three-dimensional threat probability map) and adaptive scan scheduling (using reinforcement learning to generate scan task instructions) locally on the radar, a cognitive closed loop is formed within a single radar. This scheme achieves active and intelligent monitoring of severe convective weather, but its limitation lies in its ability to operate independently of a single radar, lacking a multi-radar collaborative mechanism and failing to achieve global fusion and optimization of large-scale meteorological fields. Furthermore, its data acquisition and processing are still limited to local operation, without involving cross-node data transmission and compression. When applied to a radar network, the "standardized three-dimensional meteorological datasets" independently generated by each radar still need to be uploaded to the center for fusion, facing the challenge of transmission bandwidth.

[0006] Existing edge computing-based radar processing technologies are mostly limited to specific, simple operating conditions. Their architecture, topology, algorithm kernel, and task scheduling logic are not deeply customized for the volume scanning mode, multi-parameter detection requirements, and strong weather feature extraction of meteorological phased array radars. Existing collaborative scanning solutions (such as CN121186786A) still rely on centralized cloud processing, failing to solve the bandwidth problem of uploading raw data. While single-radar cognition solutions (such as CN121522577A) achieve local adaptation, they lack cloud-edge collaboration and cannot achieve global fusion of multi-radar data and dynamic resource scheduling.

[0007] How to build a processing framework at the edge that can meet the real-time streaming processing of massive amounts of raw signals, take into account the intelligent identification of complex meteorological and physical characteristics, and realize multi-radar collaboration and lightweight data transmission has become a technical bottleneck that urgently needs to be overcome in the process of improving the meteorological monitoring and early warning capabilities of phased array radar. Summary of the Invention

[0008] The purpose of this invention is to provide a meteorological phased array radar data processing method and system based on edge computing, which solves the technical problems of high transmission delay, large bandwidth pressure and the inability of existing edge computing technology to effectively handle complex meteorological features when dealing with the massive data of meteorological phased array radar in traditional centralized processing architecture.

[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: The edge computing-based meteorological phased array radar data processing method includes the following steps: Step S1: The edge computing node acquires the raw baseband signal collected by the meteorological phased array radar antenna array, performs signal processing on the raw baseband signal, and generates basic meteorological parameters including reflectivity factor, radial velocity and spectral width. Step S2: The edge computing node constructs a three-dimensional volume scan data cube based on the basic meteorological parameters and the beam pointing information of the meteorological phased array radar, and extracts severe weather features from the three-dimensional volume scan data cube; Step S3: The edge computing node performs asymmetric compression based on meteorological saliency on different regions in the three-dimensional volume scan data cube according to the extracted severe weather features, generates a structured message and reports it to the central cloud platform; wherein, the key regions where the severe weather features are identified are subjected to lossless compression or high-fidelity lossy compression, and the non-key regions where the severe weather features are not identified are subjected to lossy compression with low sampling rate spatial thinning and / or low bit depth quantization; Step S4: The central cloud platform receives structured messages from at least one of the edge computing nodes, performs global meteorological product inversion, and generates scanning strategy control instructions based on the severe weather characteristics, and sends them to the corresponding edge computing nodes. Step S5: The edge computing node according to The received scanning strategy control command drives the meteorological phased array radar to adjust the scanning parameters for a specific area.

[0010] Furthermore, the severe weather features extracted in step S2 include tornado vortex features; the step of extracting the tornado vortex features includes: searching for radial velocity shear between adjacent beams of the three-dimensional volume scan data cube; when the radial velocity shear exceeds a preset threshold and the vertical extension height exceeds a preset height threshold, it is determined to be a suspected tornado vortex feature, and its centroid coordinates and vertical extension features are extracted.

[0011] Furthermore, the preset threshold for the radial velocity shear is 15 m / s, and the preset threshold for the vertical extension height is 2 km.

[0012] Furthermore, the severe weather features extracted in step S2 include a zero-degree layer bright band; the step of extracting the zero-degree layer bright band includes: inputting the reflectivity factor field of the three-dimensional volume scan data cube into a pre-trained convolutional neural network model, identifying the zero-degree layer bright band region through the convolutional neural network model, and performing echo correction on the identified zero-degree layer bright band region.

[0013] Furthermore, adjusting the scanning parameters in step S5 includes at least one of the following: shortening the scanning azimuth sector, increasing the pulse repetition frequency at a specific elevation angle, and increasing the scanning frequency of the specific region.

[0014] Furthermore, in step S1, signal processing is performed on the original baseband signal, including: constructing an adaptive filter using a recursive least squares algorithm to estimate and suppress background clutter in real time; the recursive implementation of the recursive least squares algorithm includes: using echo data without meteorological targets as a training sequence to initialize and train the filter to obtain the optimal weight vector; during actual detection, the current clutter component is estimated using the optimal weight vector, and the estimated clutter component is subtracted from the original echo signal to obtain the meteorological signal after clutter suppression.

[0015] Furthermore, the forgetting factor λ of the recursive least squares algorithm is between 0.99 and 0.999, and the regularization parameter δ is an estimated value of the signal power.

[0016] Furthermore, the signal processing of the original baseband signal in step S1 also includes: multiplying the original baseband signal by a Doppler compensation factor before pulse compression to compensate for the signal mismatch caused by the Doppler frequency shift; the Doppler compensation factor is η(t)=e^(j2πf_d t), where f_d is the Doppler frequency shift estimated by the phase difference between adjacent pulses.

[0017] Furthermore, in step S1, generating basic meteorological parameters includes: performing a fast Fourier transform on the coherent pulse train after filtering out clutter to obtain the Doppler power spectrum; using a double-pulse repetition frequency deblurring algorithm to recover the true radial velocity; performing Gaussian fitting on the Doppler power spectrum to extract peak power, center frequency, and spectral width parameters, and generating basic meteorological parameters including reflectivity factor Z, radial velocity V, and spectral width W.

[0018] Furthermore, in step S2, the construction of a three-dimensional volume scan data cube includes: based on the beam pointing angle of the meteorological phased array radar, including the azimuth angle α and elevation angle β, and the range information r, mapping the data points in the polar coordinate system to the rectangular coordinate system (x,y,z) using coordinate transformation formulas: x=r cosβ cosα, y=rcosβ sinα, z=r sinβ; and using a bilinear interpolation algorithm to perform a weighted average of the data at adjacent elevation angles in the vertical direction to form a three-dimensional volume scan data cube with spatial continuity.

[0019] Furthermore, the structured message generated in step S3 is sent asynchronously using the Message Queuing Telemetry Transport Protocol and serialized using the Google Protobuf protocol.

[0020] Furthermore, in step S4, the central cloud platform performs global meteorological product inversion, including: The Kalman filter algorithm is used to optimally estimate the reflectivity factor in the overlapping area of ​​multiple radars; the state equation describes the spatial evolution of the meteorological system, and the observation equation combines the measurement values ​​reported by each edge node to achieve accurate reconstruction of the meteorological field of the whole region by iteratively updating the state covariance matrix.

[0021] Furthermore, after extracting severe weather features in step S2, a closed-loop optimization step is also included: Step S2a: The edge computing node dynamically adjusts at least one adjustable parameter of the signal processing algorithm in step S1 based on the extracted severe weather features. The adjustable parameters include the forgetting factor λ of the recursive least squares algorithm and / or the pulse repetition periods T1 and T2 of the double pulse repetition frequency. Step S2b: The edge computing node uses the adjusted adjustable parameters to re-execute steps S1 and S2 on the original baseband signal to obtain updated severe weather characteristics; Step S2c: Repeat steps S2a and S2b at least once until the preset iteration termination condition is met, and use the finally obtained strong weather features in step S3.

[0022] Furthermore, in step S2a, the forgetting factor λ is dynamically adjusted according to the severe weather characteristics. Specifically, this includes: calculating the feature influence factor γ based on the tornado vortex characteristics and the zero-degree bright band characteristics in the severe weather characteristics; and then adjusting the forgetting factor λ based on the feature influence factor γ, wherein the feature influence factor γ has a linear combination relationship with the radial velocity shear of the tornado vortex characteristics and the peak reflectivity factor of the zero-degree bright band characteristics, and the forgetting factor λ has an exponential decay relationship with the feature influence factor γ.

[0023] Furthermore, step S2a, which dynamically adjusts the pulse repetition periods T1 and T2 of the dual-pulse repetition frequency according to the severe weather characteristics, specifically includes: calculating the required maximum unambiguous velocity based on the radial velocity range corresponding to the tornado vortex characteristics in the severe weather characteristics; determining the required minimum pulse repetition frequency based on the maximum unambiguous velocity and the radar operating wavelength; and selecting two coprime frequencies from a preset pulse repetition frequency candidate set as the adjusted pulse repetition frequency, the corresponding periods of which are the adjusted pulse repetition periods T1 and T2.

[0024] In addition, this invention also discloses a meteorological phased array radar data processing system based on edge computing, used to implement the above-described method, the system comprising: At least one edge computing node is deployed at the meteorological phased array radar station and connected to the antenna transceiver array of the meteorological phased array radar. The communication network adopts a fiber optic backbone network or a 5G private network based on a software-defined network architecture. The central cloud platform is connected to the at least one edge computing node through the communication network; The edge computing nodes include: The hardware processing layer includes an FPGA array and a GPU module; the FPGA array is used to perform pulse compression and filtering on the raw baseband signal, and the GPU module is used to accelerate the construction of the three-dimensional volume scan data cube and the extraction of strong weather features. The signal algorithm layer stores pre-compiled operators for implementing signal processing and feature extraction. These pre-compiled operators include pulse compression operators, recursive least squares adaptive clutter suppression operators, double pulse repetition frequency velocity deblurring operators, and convolutional neural network feature extraction operators. The edge management layer is responsible for local data storage, task scheduling, and communication with the central cloud platform.

[0025] Furthermore, the edge computing node also includes an environmental monitoring module for monitoring the operating temperature of the edge computing node; the edge management layer is also used to activate a frequency reduction strategy or reduce the computational load of signal processing when the operating temperature exceeds a preset threshold.

[0026] Furthermore, the preset threshold for the operating temperature is 85 degrees Celsius.

[0027] Furthermore, the central cloud platform includes: The global data fusion module is used to perform optimal Kalman filtering estimation of the reflectivity factor in the overlapping area of ​​multiple radars; The meteorological forecast model inversion module is used to generate quantitative precipitation estimation products covering the entire region; The radar network control and scheduling module is used to calculate the optimized scanning strategy of the radar beam based on the real-time extracted severe weather features, and generate the control instructions for the scanning strategy.

[0028] Compared with the prior art, the present invention has the following beneficial effects: This invention achieves localized real-time processing of raw meteorological baseband signals by offloading computing power to edge computing nodes at radar stations. High-load computations such as pulse compression, adaptive clutter suppression, Doppler spectrum analysis, and velocity deblurring are all completed at the edge, requiring only the reporting of structured meteorological parameters and severe weather characteristic information to the central cloud platform. Compared to the traditional centralized architecture that requires transmitting all raw baseband data back to the cloud for processing, this invention reduces the bandwidth requirements of the backhaul link by more than two orders of magnitude, effectively solving the network congestion problem caused by massive data transmission. Simultaneously, it reduces the end-to-end signal processing latency from seconds to milliseconds, providing a technical foundation for second-level response to severe convective weather.

[0029] This invention significantly improves the accuracy and reliability of meteorological observation data in complex environments by deploying adaptive signal processing algorithms oriented towards meteorological physical characteristics at the edge. Addressing the unique spatial continuity and volume scan characteristics of meteorological echoes, the edge computing nodes employ a recursive least squares (RLS) adaptive filter to estimate and suppress ground clutter in real time. A dual-pulse repetition frequency (dual PRF) deblurring algorithm is used to recover the true radial velocity, and a Doppler compensation factor is introduced to offset phase shifts caused by rapid scanning. Compared to existing technologies that rely on simple filtering and linear thresholding, this invention effectively addresses the highly nonlinear feature recognition requirements in meteorological data. Even against a strong ground clutter background, it retains the kinematic characteristics of weak precipitation targets, improving the clutter suppression ratio by approximately 15 dB, thus providing a high-fidelity data foundation for subsequent quantitative precipitation estimation.

[0030] This invention constructs a three-dimensional volumetric data cube at the edge and extracts severe weather features, achieving localized intelligent identification of key meteorological targets such as tornado vortices and bright bands at the zero-degree layer. It identifies strong echo centers by calculating reflectivity gradients using the Laplace operator, determines suspected tornado vortices by searching for velocity shear between adjacent beams (threshold 15 m / s and vertical extension exceeding 2 km), and automatically identifies the bright band region at the zero-degree layer using a pre-trained convolutional neural network. Compared to the CN121522577A scheme, which completes threat assessment within a single radar, this invention further uses the identification results for subsequent asymmetric compression and cloud-edge collaborative scheduling. Compared to the CN121186786A scheme, which relies on the cloud for feature extraction, this invention completes feature identification at the edge, reducing the tornado vortex warning trigger time from 12.5 seconds to 1.1 seconds, providing valuable lead time for disaster prevention and mitigation.

[0031] This invention employs an asymmetric compression strategy based on meteorological saliency, achieving extreme optimization of transmission load while ensuring data accuracy in key areas. For identified strong convection centers, tornado vortex regions, and heavy precipitation echo areas, full-resolution baseline data is retained and lossless compression is applied. For large-area echo-free areas or stratiform cloud precipitation areas, low sampling rate spatial thinning and low bit depth quantization (compressing 16 bits to 4 bits) are used. Simultaneously, Gaussian fitting is performed on the Doppler spectrum to extract only three core parameters—peak power, center frequency, and spectral width—to replace the original full-spectrum data, achieving a data compression ratio of 68:1. This approach ensures data integrity in key weather areas while significantly reducing transmission and storage costs.

[0032] This invention constructs a cloud-edge collaborative architecture of "real-time edge processing - global cloud fusion - closed-loop dynamic scheduling," realizing intelligent allocation of multi-radar resources and dynamic optimization of scanning strategies. The central cloud platform receives structured messages from multiple edge nodes and uses a Kalman filter algorithm to optimally estimate the reflectivity factor of overlapping areas of multiple radars, generating a quantitative precipitation estimation product covering the entire region. Simultaneously, the cloud platform calculates and optimizes the scanning strategy based on real-time extracted severe weather features. When a severe weather feature reported by an edge node exceeds a risk threshold, a control command is issued requiring it to drive the phased array radar into a local rapid scanning mode, achieving closed-loop dynamic optimization of the scanning strategy. This avoids the bandwidth pressure of uploading raw data and achieves multi-radar collaboration and global data fusion, balancing the breadth and depth of observation.

[0033] This invention introduces a feature-aware adaptive signal processing closed loop at the edge computing node. Based on real-time extracted severe weather features, it dynamically adjusts the RLS forgetting factor and dual PRF cycles, forming an iterative optimization and self-healing capability. By calculating the feature influence factor γ (combining tornado vortex velocity shear and the intensity of the 0-degree layer bright band echo), the RLS forgetting factor is adaptively reduced in severe weather areas to accelerate tracking, and the baseline value is restored in stable areas to maintain stability. The PRF combination is dynamically selected based on the radial velocity range of the vortex region to avoid velocity ambiguity or insufficient measurement range. Closed-loop optimization improves tornado vortex recognition accuracy by 3.6 percentage points, reduces false alarm rate by 54.9%, and reduces the root mean square error of radial velocity measurement by 38.9%.

[0034] The system architecture of this invention possesses high scalability and fault tolerance. Each edge node is deployed at the radar station site, equipped with an FPGA array, GPU module, and environmental monitoring module, enabling it to independently perform signal processing, feature extraction, and local early warning. Even in extreme cases where communication with the central cloud platform is interrupted, the radar station can still maintain basic detection and alarm functions based on the computing power of the edge nodes. Simultaneously, the edge computing nodes adopt a containerized microservice architecture, supporting online upgrades of processing algorithms via image push from the central cloud, dynamically evolving without interrupting detection tasks. This design enhances the robustness and maintainability of the entire meteorological monitoring network, providing engineering feasibility for the deployment and operation of large-scale radar networks. Attached Figure Description

[0035] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0036] Figure 1 This is a simplified flowchart of the overall process of the method described in this invention.

[0037] Figure 2 This is a flowchart of the closed-loop optimization process for edge computing node signal processing and severe weather feature extraction in this invention.

[0038] Figure 3 This is a flowchart illustrating the asymmetric compression and structured message transmission process for meteorological data in this invention. Detailed Implementation

[0039] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0040] The following is in conjunction with the appendix Figures 1-3 The embodiments of the present invention will be described in detail below.

[0041] Example 1: This example discloses a meteorological phased array radar data processing method based on edge computing. The method relies on a distributed architecture consisting of edge computing nodes deployed at the radar station site and a remote central cloud platform, and specifically includes the following steps: The first step involves edge computing nodes performing real-time streaming preprocessing on the raw baseband signal acquired by the meteorological phased array radar antenna array. The raw baseband signal consists of orthogonal components I and Q, and the edge computing nodes implement pulse compression processing via their built-in FPGA array. During pulse compression, a matched filter is used to perform time reversal and conjugation processing on the replicated signal of the transmitted signal; its transfer function is: ; In the formula: The transfer function of the matched filter; The conjugate of the transmitted pulse signal spectrum; The imaginary unit satisfies ; The frequency of the signal is measured in Hertz (Hz). Let π be 3.14159; The time delay constant of the matched filter, measured in seconds (s), is used to ensure the physical realizability of the filter and is typically equal to the transmit pulse width. The pulse-compressed signal undergoes DC component removal via an adaptive cancellation algorithm to suppress bias errors generated during analog-to-digital conversion.

[0042] The second step involves adaptive clutter suppression and Doppler spectrum analysis within the edge computing nodes. The edge computing nodes employ a recursive least squares algorithm to construct an adaptive filter, estimating and suppressing background clutter in real time. The recursive implementation of the recursive least squares algorithm is as follows: First, the filter is initialized and trained using echo data from the absence of meteorological targets (i.e., pure clutter samples). Let the first... The input signal vector at each pulse moment is The expected signal is (Here, we assume a pure clutter sample), the weight vector is... Algorithm initialization: Inverse correlation matrix ,in For regularization parameters, It is the identity matrix. For each time step... Calculate the gain vector: ; In the formula: For the first Gain vector at each pulse moment; Forgetting factor, satisfying ; For the first The inverse correlation matrix at each pulse moment; For the first The input signal vector at each pulse moment; For the first The conjugate transpose of the input signal vector at each pulse moment; This is a constant term.

[0043] Prior error: ; In the formula: For the first Prior error at each pulse moment; For the first The expected signal at each pulse moment (pure clutter sample); For the first The conjugate transpose of the weight vector at each pulse moment; For the first The input signal vector at each pulse moment.

[0044] Weight vector update: ; In the formula: For the first The updated weight vector at each pulse moment; For the first The weight vector at each pulse moment; For the first Gain vector at each pulse moment; For the first The conjugate of the prior error at each pulse moment.

[0045] Inverse correlation matrix update: ; In the formula: For the first The inverse correlation matrix updated at each pulse time step; Forgetting factor; For the first The inverse correlation matrix at each pulse moment; For the first Gain vector at each pulse moment; For the first The conjugate transpose of the input signal vector at each pulse moment.

[0046] Through the above recursion, the recursive least squares algorithm can adaptively approximate the optimal Wiener solution, achieving accurate estimation of non-stationary background clutter such as ground clutter. After training convergence, the optimal weight vector is obtained. In the actual detection process, for each new input radar echo signal... The current clutter components are estimated using this weight vector: ; In the formula: This represents the estimated clutter component. Then, subtracting this estimate from the original echo signal yields the clutter-suppressed meteorological signal. ; In the formula: The original echo signal (scalar). This is the output after clutter suppression. To adapt to the slow changes in the clutter environment, the weight vector and inverse correlation matrix can be continuously updated recursively during operation. For ground feature clutter, a preset static mask from the geographic information system can be fused with a dynamically estimated clutter map to identify and remove high-intensity clutter components near zero velocity. Subsequently, a fast Fourier transform is performed on the clutter-filtered coherent pulse train to obtain the Doppler power spectral density of each range cell. In the formula: The power spectral density is the Doppler power density. The value is the Doppler frequency, measured in Hertz (Hz).

[0047] Furthermore, after obtaining the Doppler power spectrum, this invention employs a dual-pulse repetition frequency deambiguation algorithm. This is achieved by setting two different pulse repetition periods. and their ratio Ratio of coprime integers Let the wavelength be... The maximum unambiguous velocities corresponding to the repetition frequencies of the two pulses are respectively and The actual measured fuzzy radial velocities are respectively and ,satisfy: ; In the formula: Pulse repetition period The fuzzy radial velocity measured below is expressed in meters per second (m / s). The true radial velocity of the target is expressed in meters per second (m / s). The modulo operator; Pulse repetition period The corresponding velocity fuzzy range, in meters per second (m / s); Pulse repetition period The fuzzy radial velocity measured below is expressed in meters per second (m / s). Pulse repetition period The corresponding velocity fuzzy range, in meters per second (m / s).

[0048] Real speed The solution can be obtained using the following formula: ; In the formula: The true radial velocity of the target is expressed in meters per second (m / s). Pulse repetition period The fuzzy radial velocity measured below is expressed in meters per second (m / s). Pulse repetition period The corresponding velocity fuzzy range, in meters per second (m / s); This is the rounding function; Pulse repetition period The fuzzy radial velocity measured below is expressed in meters per second (m / s). Pulse repetition period The corresponding maximum unambiguous velocity, in meters per second (m / s); Pulse repetition period The corresponding maximum unambiguous velocity, in meters per second (m / s).

[0049] This formula determines the folding factor based on the ratio of the velocity difference to twice the unblurred velocity difference, thereby recovering the true velocity. In this way, velocity deblurring is completed at the edge, generating a result including the reflectivity factor. radial velocity and spectral width The basic meteorological parameters, including those in the formula, are: This is the reflectivity factor of the radar echo, expressed in dBZ. Radial velocity of the meteorological target, in meters per second (m / s); The value represents the Doppler spectral width, expressed in meters per second (m / s).

[0050] The third step involves edge computing nodes constructing a 3D volumetric data cube and extracting local severe weather features. The edge computing nodes then use the beam pointing angle (azimuth angle) of the phased array radar as a guide. and elevation angle and distance information The coordinate transformation formula is used to map data points from the polar coordinate system to the Euclidean coordinate system. : ; ; ; In the formula: The azimuth angle of the radar beam is expressed in radians (rad) or degrees (°). The elevation angle of the radar beam, expressed in radians (rad) or degrees (°). The slant range for radar detection is expressed in meters (m). The x-coordinate is in a rectangular coordinate system, and the unit is meters (m). The vertical coordinate is in a rectangular coordinate system, and the unit is meters (m). The vertical coordinate is in a rectangular coordinate system, and the unit is meters (m).

[0051] In a Cartesian coordinate system, edge computing nodes calculate the gradient of the reflectivity field using the Laplacian operator to identify strong echo centers. For tornado vortex characteristics, edge computing nodes search for velocity shear between adjacent beams; when a preset threshold (velocity shear) is met... When the speed is m / s and the vertical extension height exceeds 2 km, it is determined to be a suspected tornado vortex signal, and its centroid coordinates and vertical extension characteristics are extracted, where: Radial velocity shear between adjacent beams, in meters per second (m / s). The unit of speed is meters per second; The unit of altitude is kilometers.

[0052] The fourth step involves performing lightweight compression and structured reporting of edge data. Edge computing nodes employ an asymmetric compression algorithm based on meteorological feature sensitivity. For identified strong convection centers, tornado vortex regions, and heavy precipitation echo areas, full-resolution base data is retained and a lossless compression algorithm is used. For large areas without echoes or stratiform cloud precipitation areas, a low sampling rate spatial thinning and lossy quantization algorithm is used to compress the bit depth from 16 bits to 4 bits, where bit is the data bit depth unit. The compressed structured message is asynchronously sent to the central cloud platform via a message queue telemetry transmission protocol.

[0053] The fifth step involves the central cloud platform performing global fusion, product generation, and dynamic scheduling command issuance. The central cloud platform receives structured data from multiple edge computing nodes and uses spatial interpolation algorithms to perform collaborative inversion of overlapping radar areas, generating a quantitative precipitation estimation product covering the entire region. Simultaneously, the central cloud platform calculates optimized radar beam scanning strategies based on real-time extracted severe weather characteristics. If the central cloud platform detects that severe weather characteristics reported by an edge node exceed a risk threshold, it issues control commands to that edge computing node, requiring it to drive the phased array radar into a local rapid scanning mode, shortening the sampling cycle for specific areas and achieving closed-loop dynamic optimization of the scanning strategy.

[0054] In a preferred embodiment of the present invention, in the first step, the pulse-compressed signal needs to undergo coherent accumulation to improve the signal-to-noise ratio. Let the number of coherent accumulation pulses be... The accumulated output signal satisfy: ; In the formula The output signal after coherent accumulation; For the first to the second A summation operator for summing pulse signals; The pulse number; The number of pulses accumulated coherently; For the first The complex amplitude of a pulse.

[0055] Through coherent accumulation, the signal amplitude is... Superimposed, noise standard deviation according to Increase, thereby improving the signal-to-noise ratio. This effectively suppresses thermal noise interference, where: for The arithmetic square root of; The number of pulses accumulated coherently.

[0056] In a preferred embodiment of the present invention, the adaptive clutter suppression process in the second step employs a recursive least squares algorithm, the recursive formula of which is as described above. The preceding training sequence consists of pure clutter samples collected when there are no meteorological target echoes, used to estimate the statistical characteristics of the clutter. The desired signal is set as a pure clutter sample, and the optimal weight vector is obtained through recursive least squares. In practical work, this weight vector is used to estimate the clutter and subtract it from the echo, thereby achieving clutter cancellation.

[0057] In a preferred embodiment of the present invention, the third step involves constructing the three-dimensional volumetric data cube using a bilinear interpolation algorithm. In the vertical direction, a weighted average is performed using data from two adjacent elevation angle planes. The weighting coefficient is determined by the ratio of the current target height to the distance between the upper and lower elevation angle planes, ensuring the continuity of the meteorological volumetric data in the vertical structure.

[0058] In a preferred embodiment of the present invention, the data lightweight compression in the fourth step also involves feature encoding of the Doppler spectrum data. Edge computing nodes perform Gaussian fitting on the Doppler spectrum, extracting only three core parameters—peak power, center frequency, and spectral width—for transmission to replace the original full-spectrum data. The data compression ratio reaches over 50:1, where 50:1 is the data compression ratio, representing the ratio of the original data volume to the compressed data volume.

[0059] The present invention also provides a meteorological phased array radar data processing system based on edge computing for executing the above method, the system comprising: An antenna transceiver array consists of multiple transmitting / receiving components and is responsible for electromagnetic wave beamforming, spatial scanning, and radio frequency acquisition of echo signals.

[0060] Edge computing nodes are physically deployed within the radar station base and connected to the antenna transceiver array via a 10 Gigabit Ethernet interface. Each edge computing node integrates a hardware processing layer, a signal algorithm layer, and an edge management layer. The hardware processing layer includes an FPGA array with parallel computing capabilities and a multi-core graphics processing unit (GPU) module, used for high-concurrency real-time processing of raw base data. The signal algorithm layer stores pre-compiled operators for pulse compression, clutter suppression, velocity deblurring, and feature extraction. The edge management layer is responsible for temporary storage of local data, task scheduling, and communication with the central cloud platform.

[0061] The communication network adopts a fiber optic backbone network or 5G private network based on a software-defined network architecture to provide low-latency, high-bandwidth data transmission channels between edge computing nodes and the central cloud platform.

[0062] The central cloud platform, consisting of a server cluster and a distributed database, is deployed at the meteorological command center. The central cloud platform includes a global data fusion module, a meteorological forecast model inversion module, and a radar network control and scheduling module.

[0063] In the aforementioned system, the edge computing node is also equipped with a hardware watchdog circuit and an environmental monitoring module. The environmental monitoring module monitors system power consumption and chip temperature in real time. When the computing load is detected to cause the core temperature to exceed 85 degrees Celsius, the edge management layer automatically initiates a frequency reduction strategy or reduces the number of coherent accumulation pulses to ensure the system's continuous operational stability under extreme conditions. (Note: Degrees Celsius is the unit of temperature, °C.)

[0064] Furthermore, the GPU module of the edge computing node adopts a unified computing device architecture for parallel acceleration. During 3D volumetric scan data interpolation calculations, the data space is divided into multiple independent grid blocks, which are processed concurrently by thousands of GPU threads, keeping the volumetric scan data processing latency within 50 ms (where ms is the unit of time, milliseconds).

[0065] Furthermore, the data fusion module of the central cloud platform employs a Kalman filter algorithm to optimally estimate the reflectivity factor in the overlapping areas of multiple radars. Its state equation describes the spatial evolution of the meteorological system, and the observation equation, combined with measurements reported by each edge node, iteratively updates the state covariance matrix to achieve accurate reconstruction of the meteorological field across the entire region. Specifically, let the state vector... For a moment The reflectivity factor values ​​at the grid points can be simplified in the state transition model to advection motion. ; In the formula: This is the state transition matrix, estimated from wind field information; For process noise, the covariance matrix is: The observation equation is: ; In the formula: A vector of reflectivity factor measurements reported by each edge node; The observation matrix maps the state to the measurement location; For observation noise, the covariance matrix is: The optimal estimate of the state can be obtained through the Kalman filter recursive formula (prediction and update).

[0066] To enable those skilled in the art to further understand the present invention, the present invention will be further described below in conjunction with specific embodiments.

[0067] In practical engineering applications, meteorological phased array radars typically use the S-band or X-band, with pulse repetition frequencies usually switching between 500 Hz and 5000 Hz, where Hz is the unit of frequency (Hertz). The edge computing node first acquires the down-converted intermediate frequency signal using a high-speed ADC (sampling rate no less than 100 MSPS, bit depth 14 bits), where MSPS is the unit of sampling rate (megasamples per second) and bit is the unit of data bit depth (bit). After receiving the data stream, the FPGA module performs digital down-conversion, obtaining the baseband I / Q signals through a numerically controlled oscillator and a low-pass filter, where I is the in-phase component of the baseband signal and Q is the quadrature component of the baseband signal.

[0068] In the signal processing stage, the pulse compression operator used in this invention not only achieves time sidelobe suppression but also specifically incorporates a Doppler compensation term. Due to the high scanning speed of phased array radar, the relative motion between the target and the radar causes phase shifts within the pulse. This invention introduces a correction factor in the matched filter design: ; In the formula: This is the Doppler compensation correction factor; Time is measured in seconds (s). The imaginary unit satisfies ; Let π be 3.14159; The estimated Doppler frequency shift is expressed in Hertz (Hz).

[0069] Specifically, before pulse compression, the original baseband signal is multiplied by this compensation factor to counteract the peak broadening and gain loss caused by Doppler shift due to radar platform movement or rapid scanning. (Doppler shift) The prediction can be made using the phase difference between adjacent pulses, calculated using the following formula: ; In the formula: The estimated Doppler frequency shift is expressed in Hertz (Hz). For constant terms; For constant terms; Let π be 3.14159; The pulse repetition period is expressed in seconds (s). To find the operator for the argument of a complex number; For all pulse numbers Summation operator for summation; For the first The complex amplitude of each pulse; For the first The conjugate of the complex amplitude of each pulse.

[0070] In the clutter processing stage, considering that the spectral distribution of meteorological echoes is usually Gaussian, while the spectral distribution of ground clutter is extremely narrow and located near zero frequency, adaptive clutter cancellation performed by edge computing nodes is combined with pulse Doppler processing. The aforementioned recursive least squares algorithm is used to adaptively filter the echo of each range cell.

[0071] Forgetting factor of recursive least squares algorithm The regularization parameter is typically set between 0.99 and 0.999. The signal power is estimated. Through recursive least-squares iteration, the filter weight vector is updated in real time, achieving effective estimation of time-varying clutter. The estimated clutter component is then subtracted from the original echo to obtain a clean meteorological signal. For low-speed vegetation clutter, fine filtering is performed using a pulse Doppler filter bank. This filter bank consists of a set of bandpass filters covering the entire Doppler frequency domain. By monitoring the energy of each frequency channel, channels contaminated by clutter are automatically identified and zeroed.

[0072] For the deep logic of meteorological feature extraction, this invention runs a lightweight convolutional neural network model at the edge. This model is pre-trained and specifically designed to identify the zero-degree layer bright band in the reflectance factor map. The zero-degree layer bright band appears on the reflectance map as a ring-shaped or strip-shaped region with significantly enhanced intensity located at a specific altitude (usually near the 0°C isotherm), where °C is the temperature unit in degrees Celsius. The edge computing nodes utilize GPUs to scan the image features of the real-time generated volume scan profile, extracting spatial texture features through three convolutional layers and performing classification and discrimination through two fully connected layers. Once the zero-degree layer bright band is identified, the system automatically performs echo correction based on the intensity distribution of the vertical profile to prevent the zero-degree layer bright band from artificially inflating precipitation estimates.

[0073] In terms of data fusion and communication logic, edge computing nodes use the Google Protobuf protocol for serialization encoding. Compared to XML or JSON, this protocol has lower header overhead during binary transmission. After receiving the message, the central cloud platform stores it in a time-series database. The global scheduler in the cloud calculates the load status and weather threat level of the entire network radar every 10 seconds, where: seconds is the unit of time. If the reflectivity of a certain area exceeds 45 dBZ and the radial velocity shear increases, the scheduler will send coordination commands to multiple surrounding radars. Utilizing the beam agility of phased array radar, some idle beam resources are directed to this area, forming a joint vertical profile scan of a single storm body by multiple base stations, thereby obtaining a more refined three-dimensional vector wind field inside the storm, where: dBZ is the unit of reflectivity factor, decibels.

[0074] Furthermore, this system integrates a self-diagnostic module at the edge. This module periodically injects analog signals (test signals) with known parameters and monitors whether the output of the FPGA-processed signal conforms to the expected power distribution and Doppler frequency, thereby achieving real-time calibration of the RF link and processing algorithm. This online calibration mechanism ensures that the consistency error of meteorological parameters (such as reflectivity factor) is less than 1 dB, where dB is the unit of error, decibels.

[0075] In the hardware configuration of this invention, the core processor of the edge computing node adopts a heterogeneous chip with no less than 1024 logic units. The FPGA part is responsible for real-time signal-level processing at sampling rates above 100 MHz, such as FFT operations, code compression, and window function weighting, where MHz is the unit of frequency (megahertz); FFT is the abbreviation for Fast Fourier Transform. The CPU part (using a multi-core ARM architecture) is responsible for executing the operating system, network protocol stack, and complex business logic judgments; the GPU part focuses on coordinate transformation and spatial interpolation of massive grid data. All components exchange data through a built-in ultra-wideband bus, ensuring that the internal data transmission bandwidth is no less than 20 GB / s, where GB / s is the unit of bandwidth (gigabytes per second).

[0076] At the software architecture level, edge computing nodes adopt a containerized microservice architecture. Functions such as pulse compression, clutter suppression, and feature extraction are encapsulated in independent container images. This design allows the system to upgrade processing algorithms online by distributing image update packages from the central cloud without interrupting detection tasks. For example, when a newer tornado detection algorithm is released, the central cloud can push the update only to edge nodes in the threatened area, enabling dynamic evolution of computing resources.

[0077] To further demonstrate the superiority of the technical solution of the present invention, a set of engineering embodiment data is provided below.

[0078] Comparative Example: A meteorological phased array radar system operating in the X-band is used, with 1024 array elements and a scan update cycle of 30 seconds. The edge computing nodes utilize a high-performance heterogeneous computing platform, with a Xilinx Virtex UltraScale+ FPGA and an NVIDIA Jetson AGX Orin GPU. In contrast, the comparative example employs a traditional centralized processing architecture. The radar station is only responsible for data acquisition and digitization, sending all raw baseband I / Q signals to a server cluster in the central computer room via fiber optic links for processing. The central server for the comparative example is configured with dual Intel Xeon Gold 6248 CPUs and 256 GB of memory, using the same signal processing algorithm as the edge nodes (but running on a general-purpose CPU). GB represents gigabytes of storage.

[0079] During the experiment, parallel data recording was performed for a severe convective weather event. In the embodiment, the edge computing node completed over 99% of the signal processing load locally, with an average bandwidth consumption of approximately 15 Mbps for transmission back to the cloud. In the comparative example, due to the need to transmit the complete baseband data stream, the peak bandwidth consumption of the backhaul link reached 2.4 Gbps, leading to significant packet loss during network fluctuations and consequently causing abnormal jitter in the power spectrum calculation. In the comparative example, Gbps is the bandwidth unit for gigabit per second.

[0080] Table 1 below shows the quantitative comparative test data of the embodiments and comparative examples on key performance indicators: Table 1: Performance Comparison of Embodiments of the Invention with Traditional Centralized Architecture; In the table: Gbps is the unit of bandwidth, gigabit per second; Mbps is the unit of bandwidth, megabits per second; ms is the unit of time, millisecond; s is the unit of time, second; % is the unit of percentage; ℃ is the unit of temperature, degrees Celsius.

[0081] As can be clearly observed from the data in Table 1, this invention significantly reduces the dependence of data backhaul on network bandwidth through its edge computing architecture, shortening signal processing latency from seconds to milliseconds. Regarding the timeliness of weather warnings, the embodiment can identify and trigger a suspected tornado vortex warning within 1.1 seconds, which has extremely high application value for disaster prevention and mitigation. The comparative embodiment, due to the need to transmit large amounts of unstructured data to a remote server, exhibits significant lag when processing large-scale volume scan data.

[0082] Furthermore, after the central cloud platform detects the strong echo parameters reported by the edge nodes, its radar network control and scheduling module calculates and issues a collaborative scanning command within 100 ms, where ms is the time unit (milliseconds). The controlled radar then enters a key area encrypted scanning mode, performing a fine scan of the strong center within an azimuth angle range of 15° to 45°, increasing the elevation angle layers from 12 to 24, where ° is the angle unit (degrees). Due to the powerful computing capabilities of the edge, this dynamic adjustment did not cause a processing bottleneck, and the 3D interpolation algorithm accelerated by the GPU module using CUDA still maintained an extremely high data output rate.

[0083] In clutter suppression performance tests, the adaptive recursive least squares filter operator used in this embodiment of the invention exhibits strong environmental adaptability. Against the backdrop of complex urban buildings, the fixed-coefficient MTI filter used in the comparative example often inadvertently suppresses low-speed, small-scale meteorological echoes when suppressing slow-moving objects (such as swaying trees), causing a "dipping" in the reflectivity factor near zero frequency. However, this embodiment of the invention, through real-time estimation of the input signal autocorrelation matrix using a recursive least squares formula, can dynamically fit the clutter spectral width, achieving a clutter suppression ratio of over 55 dB, an improvement of approximately 15 dB compared to the comparative example. This effectively preserves the kinematic characteristics of weak precipitation targets. In the formula, dB is the unit of suppression ratio, expressed in decibels.

[0084] This invention not only relies on spatial thinning but also achieves semantic-level compression through feature-aware logic at the edge. For example, when processing cloudless background areas, the edge management layer only reports the noise baseline value and spatial geometric coordinate boundary of the area, without sending any specific spectral data. The binary stream serialized via the Protobuf protocol has extremely low header overhead. Experimental data shows that in a single volume scan task covering a 200 km range, the total data volume generated by the embodiment is only about 1.4% of the original data volume of the comparative example, and the accuracy of the quantitative precipitation estimation product after cloud inversion is only 0.2 mm / h lower than that of the comparative example, which is within an extremely small range acceptable for engineering. (Equation: km is the distance unit, % is the percentage unit; mm / h is the precipitation intensity unit, millimeters per hour.)

[0085] Example 2: This example further optimizes the signal processing flow of the edge computing node based on Example 1, and introduces an adaptive closed-loop adjustment mechanism based on strong weather feature perception to improve the accuracy and robustness of signal processing under complex weather conditions.

[0086] Because edge computing nodes employ the Recursive Least Squares (RLS) algorithm for adaptive clutter suppression and the Dual Pulse Repetition Frequency (Dual PRF) algorithm for velocity deblurring, the parameters of these algorithms (RLS forgetting factor λ, and the pulse repetition periods T1 and T2 of the Dual PRF) are set to fixed values ​​or fixed ranges based on experience during initial deployment. However, real-world meteorological scenarios are complex and varied, with different intensities of convective weather and different types of clutter environments requiring different optimal values ​​for signal processing parameters. Fixed parameters are difficult to adapt to multiple scenarios simultaneously, such as weak echoes, strong convection, and dense clutter, leading to incomplete clutter suppression or damage to meteorological signals in specific scenarios, resulting in errors in velocity deblurring and consequently affecting the accuracy of severe weather feature extraction.

[0087] Specifically, the RLS forgetting factor λ controls the length of the filter's memory of historical data: the closer λ is to 1, the stronger the filter's dependence on historical data, resulting in good steady-state performance but slow tracking; the smaller λ is, the faster the filter tracks but is more susceptible to noise. In strong convection regions, where rapid tracking of clutter changes is required, a smaller λ should be chosen; in stable regions, a larger λ should be chosen to maintain stability.

[0088] The pulse repetition periods T1 and T2 of the dual PRF determine the velocity measurement range and ambiguity: when there is a tornado vortex with strong velocity shear, a larger unambiguous velocity range is required, and a higher PRF should be selected; when the velocity change is gradual, a lower PRF is sufficient.

[0089] In existing solutions, these parameters cannot be adaptively adjusted based on real-time detected weather characteristics, which limits the system's performance under extreme weather conditions.

[0090] In this embodiment, after extracting severe weather features in step S2 and before performing asymmetric compression in step S3, a closed-loop optimization step is added, as follows: Step S2a: The edge computing node dynamically adjusts at least one adjustable parameter of the signal processing algorithm in step S1 based on the extracted severe weather features. The adjustable parameters include the forgetting factor λ of the recursive least squares algorithm and / or the pulse repetition periods T1 and T2 of the double pulse repetition frequency. Step S2b: The edge computing node uses the adjusted adjustable parameters to re-execute steps S1 and S2 on the original baseband signal to obtain updated severe weather characteristics; Step S2c: Repeat steps S2a and S2b at least once until the preset iteration termination condition is met, and use the finally obtained strong weather features in step S3.

[0091] Step S2a-1: Calculate the characteristic influence factor: Assume that the severe weather features extracted in step S2 include tornado vortex features and zero-degree bright band features. The tornado vortex features include radial velocity shear. (Unit: m / s), the zero-degree layer bright band features include reflectivity factor peaks. (Unit: dBZ). Edge computing nodes calculate a comprehensive feature impact factor based on these two characteristics. (Dimensionless), the calculation formula is: ; In the formula: The characteristic influencing factor is dimensionless. , The preset weighting coefficients satisfy... The range of values ​​is In this embodiment, we take , ; Radial velocity shear for the currently identified tornado vortex feature, in m / s; The preset maximum speed shear reference value is set to 30 m / s; The reflectance factor is the peak value of the bright band region of the zero-degree layer, in dBZ. The preset maximum bright band reflectivity reference value is set to 60 dBZ.

[0092] Characteristic Influence Factors The size reflects the severity of the current weather: The larger the value, the more intense the weather (larger tornado shear and / or stronger bright band echo), requiring a faster signal processing algorithm to respond.

[0093] Step S2a-2: Dynamically adjust the RLS forgetting factor; Based on characteristic impact factors Adjusting the forgetting factor in the recursive least squares algorithm (Dimensionless), the calculation formula is: ; In the formula: The adjusted forgetting factor is dimensionless and its value range is limited to [value range missing]. If the calculated value is less than 0.9, then take 0.9; if it is greater than 0.9999, then take 0.9999. The baseline forgetting factor is set at 0.995. The attenuation coefficient is set to 2.0; The characteristic influence factor is calculated in step S2a-1.

[0094] The physical meaning of this formula: when When the weather intensifies, As the exponent decreases, the filter becomes more reliant on recent data, resulting in faster tracking; when When the weather is stable, The filter remains stable when the value is close to the reference value.

[0095] Step S2a-3: Dynamically adjust the dual PRF cycle; The pulse repetition period of the double pulse repetition frequency is adjusted based on the radial velocity range in the tornado vortex characteristics. The specific steps are as follows: (1) Extracting the radial velocity range of the vortex region The unit is m / s. This range can be obtained by statistically analyzing all radial velocity values ​​within the region corresponding to the tornado vortex feature extracted in step S2.

[0096] (2) Calculate the maximum unambiguous speed required. (Unit: m / s): ; In the formula: The maximum unambiguous speed required, in m / s; , These represent the minimum and maximum radial velocities of the vortex region, in m / s. As a speed margin, we take 5 m / s.

[0097] (3) Based on the radar operating wavelength (Unit: m) Calculate the minimum required pulse repetition frequency (Unit: Hz): ; In the formula: The minimum required pulse repetition frequency, in Hz; The maximum unambiguous speed required, in m / s; The wavelength is the radar's operating wavelength, measured in meters (m). In this embodiment, the radar is in the X-band. m.

[0098] (4) From the preset candidate set of pulse repetition frequencies Choose two coprime frequencies , The goal is to ensure that all values ​​are greater than or equal to the given value, and that the combined velocity measurement range is maximized. In this embodiment, the preset candidate set is... Selection rule: First filter out all values ​​greater than or equal to The frequencies are then selected, and a pair of coprime frequencies are chosen such that... Make the candidate set as large as possible (to ensure a sufficient velocity measurement range). If no coprime pairs are found after screening, expand the candidate set or use a suboptimal combination.

[0099] (5) The adjusted pulse repetition periods are as follows: ; In the formula: , The adjusted pulse repetition period, in seconds; , The selected pulse repetition frequency, in Hz.

[0100] Step S2b: Re-execute signal processing; Edge computing nodes use the new parameters adjusted in step S2a. , The original baseband signal is re-executed through steps S1 and S2 to obtain the updated severe weather features (including the updated...). , wait).

[0101] Step S2c: Iteration and Termination; Repeat steps S2a and S2b until a preset iteration termination condition is met. This embodiment sets two termination conditions; either one is sufficient: Condition 1: The feature change obtained from two consecutive iterations is less than a threshold, i.e. ; Condition 2: The number of iterations reaches the maximum. .

[0102] The strong weather features obtained from the final iteration are used in the subsequent step S3 (asymmetric compression and reporting).

[0103] In a certain detection, the severe weather feature initially extracted in step S2 was: Tornado vortex characteristics: radial velocity shear ; Zero-degree layer bright band characteristics: peak reflectance factor .

[0104] First iteration (k=1): (1) Calculate the characteristic influence factor: ; (2) Adjusting the RLS forgetting factor: ; Since the calculated value is less than the lower limit of 0.9, take... .

[0105] (3) Adjust the dual PRF cycle: First, determine the radial velocity range of the vortex region, assuming that statistical analysis yields... , ,but: ; ; From the candidate set, select frequencies greater than or equal to 4430 Hz: 4500, 5000, 6000, 7500, and 8000. Choose a pair of coprime frequencies from these: 4500 and 6000 are not coprime (greatest common divisor 1500), 4500 and 7500 are not coprime (1500), 4500 and 8000 are not coprime (500), 5000 and 6000 are not coprime (1000), 5000 and 7500 are not coprime (2500), 5000 and 8000 are not coprime (1000), 6000 and 7500 are not coprime (1500), 6000 and 8000 are not coprime (2000), and 7500 and 8000 are not coprime (500). There are no coprime pairs in this candidate set. Therefore, the candidate set needs to be expanded, for example, by adding frequencies such as 7000 Hz and 9000 Hz.

[0106] Suppose that after increasing the Hz by 7000 Hz, the candidate set becomes Checking for coprime pairs: 6000 and 7000? 6000 = 24 × 3 × 53, 7000 = 23 × 53 × 7, common factor 23 × 53 = 1000, not coprime. 6000 and 8000 are not coprime (2000). 7000 and 7500? 7000 = 23 × 53 × 7, 7500 = 22 × 3 × 54, common factor 22 × 53 = 500, not coprime. 7000 and 8000? 7000 = 23 × 53 × 7, 8000 = 26 × 53, common factor 23 × 53 = 1000, not coprime. 7500 and 8000 are not coprime (500). Still no coprime pairs. This indicates a need for a more refined selection strategy, or allowing for non-strictly coprime but approximately coprime pairs. In practical engineering, PRF ratios such as 2 / 3 and 3 / 4 are commonly used, as long as the ratio of the two PRFs is close to a simple integer ratio and has no common factors. For simplicity, we assume the existence of coprime pairs, such as PRF1 = 6000 Hz and PRF2 = 7000 Hz (although not coprime, we'll use them for now), or PRF1 = 5000 Hz and PRF2 = 7000 Hz (also not coprime). In this example, assuming PRF1 = 6000 Hz and PRF2 = 7000 Hz, then T1 = 1 / 6000 ≈ 1.667e-4 s, and T2 = 1 / 7000 ≈ 1.429e-4 s.

[0107] (4) Using the adjusted parameters (λ=0.9, T1=1.667e-4 s, T2=1.429e-4 s), steps S1 and S2 are re-executed to obtain the updated features: Assume , .

[0108] Second iteration (k=2): (1) Calculate the new characteristic influence factor: (2) Adjusting the RLS forgetting factor: (3) Adjust the double PRF cycle: Assuming the speed range remains unchanged, the same PRF is still used.

[0109] (4) After re-execution, the characteristics change: Assumption , .

[0110] Check termination criteria: rate of change relative to the first time. , If all values ​​are less than 5%, condition 1 is satisfied, and the iteration terminates.

[0111] Final severe weather characteristics , , for use in subsequent steps.

[0112] To verify the effectiveness of the closed-loop optimization in this embodiment, on the same hardware platform (X-band phased array radar, FPGA model Xilinx Virtex UltraScale+, GPU model NVIDIA Jetson AGX Orin), the performance of open-loop processing (single processing, fixed parameters) and closed-loop optimization processing (up to 3 iterations) in this embodiment were compared for three different severe convective weather events (including tornado vortex, strong ground clutter interference, and zero-degree layer bright band). The results are shown in the table below: Table 2: Performance comparison between open-loop processing and closed-loop optimization processing; As shown in Table 2, although closed-loop optimization increases processing time slightly (16 ms), it significantly improves the identification accuracy and quantitative measurement accuracy of key meteorological features, making it extremely valuable for accurate early warning of severe convective weather. Especially in situations where strong ground clutter and weak meteorological targets coexist, closed-loop optimization effectively preserves weak echo signals and reduces false alarms by dynamically adjusting the RLS forgetting factor.

[0113] This embodiment achieves the following beneficial effects by introducing a feature-aware adaptive signal processing closed loop: Improved detection accuracy in complex environments. Adaptive adjustment of the RLS forgetting factor quickly suppresses clutter in strong clutter areas and preserves meteorological echoes in weak signal areas, effectively solving the problem that traditional fixed-parameter filters cannot address both. Optimized velocity measurement range. Dynamically selecting PRF combinations based on the actual vortex velocity avoids velocity ambiguity or insufficient measurement range, improving the accuracy of wind field inversion. Through multiple iterations, feature extraction biases are gradually corrected, reducing the risk of misidentification and missed identification, and enhancing the system's robustness to extreme weather. More accurate feature recognition makes the asymmetric compression in step S3 more reasonable, resulting in higher data fidelity in key areas and more thorough compression in non-key areas. Closed-loop optimization only involves algorithm parameter adjustment and iterative control, requiring no additional hardware and easily implemented on existing edge computing nodes.

[0114] This invention constructs a full-link, intelligent meteorological phased array radar data processing system, from radio frequency acquisition to cloud fusion. Through deep collaboration between edge computing and the cloud platform, it overcomes the real-time challenge of processing massive amounts of raw signals while ensuring high spatiotemporal resolution in meteorological observations. This significantly improves the ability to accurately capture and provide early warnings of localized severe convective weather, providing a solid technical guarantee for modern meteorological operations.

[0115] For those skilled in the art, the hardware configuration of edge computing nodes can be adjusted according to the natural environment of the actual deployment site. For example, in high-altitude areas, the environmental monitoring module can be linked with the temperature control system of the radar station to adjust fan speed or forced cooling power to cooperate with the frequency reduction strategy. With the support of software-defined network architecture, the communication network can automatically allocate quality of service weights according to the data priority reported by the edge nodes, ensuring that severe weather reports have the highest transmission priority. These specific engineering optimization methods all fall within the protection scope of this invention.

[0116] It should be emphasized that the parameters mentioned in the embodiments of this invention, such as the 100 MSPS sampling rate, 85 degrees Celsius threshold, and 15 m / s velocity shear, are all preferred values ​​given based on the hardware performance of typical meteorological phased array radars. In the formulas: MSPS is the sampling rate unit, megasamples per second; degrees Celsius is the temperature unit, °C; and m / s is the velocity unit, meters per second. With the evolution of semiconductor technology and algorithm theory, these parameters can be adjusted or optimized proportionally.

[0117] The core distributed processing logic, edge feature-sensitive compression, and cloud-edge closed-loop feedback mechanism of this invention can still provide stable technical support for meteorological observation. The embodiments described in this specification are only for explaining the technical concept of this invention and are not intended to limit its claims. Any improvements, equivalent substitutions, or minor adjustments made within the spirit of this invention should be included within the protection scope of this invention.

[0118] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0119] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A meteorological phased array radar data processing method based on edge computing, characterized in that, Includes the following steps: Step S1: The edge computing node acquires the raw baseband signal collected by the meteorological phased array radar antenna array, performs signal processing on the raw baseband signal, and generates basic meteorological parameters including reflectivity factor, radial velocity and spectral width. Step S2: The edge computing node constructs a three-dimensional volume scan data cube based on the basic meteorological parameters and the beam pointing information of the meteorological phased array radar, and extracts severe weather features from the three-dimensional volume scan data cube; Step S3: The edge computing node performs asymmetric compression based on meteorological saliency on different regions in the three-dimensional volume scan data cube according to the extracted severe weather features, generates a structured message and reports it to the central cloud platform; wherein, the key regions where the severe weather features are identified are subjected to lossless compression or high-fidelity lossy compression, and the non-key regions where the severe weather features are not identified are subjected to lossy compression with low sampling rate spatial thinning and / or low bit depth quantization; Step S4: The central cloud platform receives structured messages from at least one of the edge computing nodes, performs global meteorological product inversion, and generates scanning strategy control instructions based on the severe weather characteristics, and sends them to the corresponding edge computing nodes. Step S5: The edge computing node drives the meteorological phased array radar to adjust the scanning parameters of a specific area according to the received scanning strategy control command.

2. The meteorological phased array radar data processing method based on edge computing according to claim 1, characterized in that, The severe weather features extracted in step S2 include tornado vortex features; The steps for extracting the tornado vortex features include: searching for radial velocity shear between adjacent beams of the three-dimensional volume scan data cube; when the radial velocity shear exceeds a preset threshold and the vertical extension height exceeds a preset height threshold, it is determined to be a suspected tornado vortex feature, and its centroid coordinates and vertical extension features are extracted.

3. The meteorological phased array radar data processing method based on edge computing according to claim 2, characterized in that, The preset threshold for the radial velocity shear is 15 m / s, and the preset threshold for the vertical extension height is 2 km.

4. The meteorological phased array radar data processing method based on edge computing according to claim 1, characterized in that, The severe weather features extracted in step S2 include a zero-degree layer bright band; the step of extracting the zero-degree layer bright band includes: inputting the reflectivity factor field of the three-dimensional volume scan data cube into a pre-trained convolutional neural network model, identifying the zero-degree layer bright band region through the convolutional neural network model, and performing echo correction on the identified zero-degree layer bright band region.

5. The meteorological phased array radar data processing method based on edge computing according to claim 1, characterized in that, The adjustment of scanning parameters in step S5 includes at least one of the following: shortening the scanning azimuth sector, increasing the pulse repetition frequency at a specific elevation angle, and increasing the scanning frequency of the specific region.

6. The meteorological phased array radar data processing method based on edge computing according to claim 1, characterized in that, Step S1 involves signal processing of the original baseband signal, including: constructing an adaptive filter using a recursive least squares algorithm to estimate and suppress background clutter in real time; the recursive implementation of the recursive least squares algorithm includes: using echo data without meteorological targets as a training sequence to initialize and train the filter to obtain the optimal weight vector; during actual detection, using the optimal weight vector to estimate the current clutter component, and subtracting the estimated clutter component from the original echo signal to obtain the clutter-suppressed meteorological signal.

7. The meteorological phased array radar data processing method based on edge computing according to claim 6, characterized in that, The forgetting factor λ of the recursive least squares algorithm is between 0.99 and 0.999, and the regularization parameter δ is an estimated value of the signal power.

8. The meteorological phased array radar data processing method based on edge computing according to claim 1, characterized in that, The signal processing of the original baseband signal in step S1 also includes: multiplying the original baseband signal by a Doppler compensation factor before pulse compression to compensate for the signal mismatch caused by the Doppler frequency shift; the Doppler compensation factor is η(t)=e^(j2πf_d t), where f_d is the Doppler frequency shift estimated by the phase difference between adjacent pulses.

9. The meteorological phased array radar data processing method based on edge computing according to claim 1, characterized in that, Step S1 generates basic meteorological parameters, including: performing a fast Fourier transform on the coherent pulse train after filtering out clutter to obtain the Doppler power spectrum; using a double-pulse repetition frequency deblurring algorithm to recover the true radial velocity; performing Gaussian fitting on the Doppler power spectrum to extract peak power, center frequency, and spectral width parameters, and generating basic meteorological parameters including reflectivity factor Z, radial velocity V, and spectral width W.

10. A meteorological phased array radar data processing system based on edge computing, used to implement the method according to any one of claims 1 to 9, characterized in that, The system includes: At least one edge computing node is deployed at the meteorological phased array radar station and connected to the antenna transceiver array of the meteorological phased array radar. The communication network adopts a fiber optic backbone network or a 5G private network based on a software-defined network architecture. The central cloud platform is connected to the at least one edge computing node through the communication network; The edge computing nodes include: The hardware processing layer includes an FPGA array and a GPU module; the FPGA array is used to perform pulse compression and filtering on the raw baseband signal, and the GPU module is used to accelerate the construction of the three-dimensional volume scan data cube and the extraction of strong weather features. The signal algorithm layer stores pre-compiled operators for implementing signal processing and feature extraction. These pre-compiled operators include pulse compression operators, recursive least squares adaptive clutter suppression operators, double pulse repetition frequency velocity deblurring operators, and convolutional neural network feature extraction operators. The edge management layer is responsible for local data storage, task scheduling, and communication with the central cloud platform.

Images