An edge parallel distribution system for weather radar echo field data

Abstract

This invention relates to the field of edge computing technology and discloses an edge parallel distribution system for weather radar echo field data. The system includes an edge control unit, a heterogeneous computing unit, and a synchronization monitoring unit. The edge control unit extracts amplitude gradient operators from sampling grid points to calculate information entropy weights and generates a computational cost estimation tensor. The heterogeneous computing unit adjusts the stream processor thread block size based on the information entropy weights and issues memory pre-occupancy instructions to offset memory access latency. The synchronization monitoring unit monitors feedback latency in real time and triggers mask dimensionality reduction for non-regional areas when the latency exceeds a preset threshold for the link transmission cycle. This invention solves the computational surge problem caused by non-uniform data spatial topology by constructing computational resource-aware adaptive distribution rules, achieving deterministic throughput of high-density data streams at the edge.

Description

Technical Field

[0001] This invention relates to an edge parallel distribution system for weather radar echo field data, belonging to the field of edge computing technology. Background Technology

[0002] Current weather radar generates high-resolution echo field data through multi-beam scanning. The mainstream approach adopts a centralized architecture, which pushes the raw echo data to the central node for processing and storage in real time. The data stream push enables centralized analysis of observation data, which is the basis for meteorological network data sharing.

[0003] The spatial distribution of echo field data is non-uniform, and the computational task density in strong convection regions increases non-linearly. Edge nodes face instantaneous computational surges during processing. Conventional buffering and peak smoothing introduces memory access delays, sacrificing observation timing alignment accuracy. Increasing the sampling rate easily leads to bandwidth overflow. Increasing bandwidth or the inability of computing units to match the computational command stream with the data topology from the underlying logic means that existing scheduling schemes are mostly post-task allocations. When processing GHz-level data streams, command queuing occurs, inducing frame synchronization failures and becoming a bottleneck limiting the real-time performance of radar networks. In addition to hardware architecture and data topology mismatch, the data... The distribution and processing logic also suffers from a lack of deep perception. For example, Chinese invention patent CN115877386B discloses a radar echo extrapolation method that introduces three-dimensional environmental field data. It integrates multi-dimensional environmental field information to optimize the extrapolation prediction accuracy. Such methods focus on improving the interpretability of the algorithm. When faced with the GHz-level raw non-uniform data stream at the edge, it is regarded as a static sequence input. It lacks prior recognition of data features. When the high information entropy echo region bursts, it cannot pre-set computing resources according to data features. The instruction queuing induces memory access jitter, causing the frame synchronization failure.

[0004] Therefore, how to perform predictive resource pre-configuration based on the non-uniform characteristics of the echo field space and achieve adaptive alignment of computing load and distribution bandwidth has become the technical problem to be solved by this invention. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: an edge parallel distribution system for weather radar echo field data, comprising: The system comprises an edge control unit, a heterogeneous computing unit, and a synchronization monitoring unit. The edge control unit is connected to the heterogeneous computing unit, and the synchronization monitoring unit is connected to both the edge control unit and the heterogeneous computing unit. The edge control unit receives the original non-uniformly distributed grid data stream and performs operations on the original non-uniformly distributed grid data stream with a step size of [missing information]. Sparse sampling is used to extract the amplitude gradient operator between adjacent sampling points, and the mean gradient of the sampling grid points is calculated. The edge control unit is based on the formula Calculate information entropy weights And based on information entropy weights Generate a computational cost prediction tensor; where, To preset the gradient extremum, The constant coefficient is the preset algorithm logic complexity. Heterogeneous computing units are used to perform task mapping based on information entropy weights. Adjust the thread block size of the stream processor in the parallel processing unit, and adjust the information entropy weight. When the preset weight threshold is exceeded, a memory pre-occupancy instruction for the grid tensor is executed to load the grid tensor into the first-level cache; The synchronous monitoring unit is used to obtain the feedback latency of the parallel processing unit. The system performs labeling, uses an attribute recognition algorithm to identify features in the original non-uniformly distributed grid data stream, and divides the original non-uniformly distributed grid data stream into important data streams and background data streams based on these features; during feedback delay... Exceeding the link transmission period of At that time, the system performs mask dimensionality reduction for the corresponding region of the background data stream, and performs distribution between real-time parallel distribution mode and asynchronous buffer distribution mode based on the labeling results.

[0006] Preferably, the edge control unit extracts the spatial distribution features of computational intensity based on the computational cost estimation tensor and generates a resource allocation mask map; the heterogeneous computing unit determines the activation ratio of computing cores and the memory bandwidth allocation weight in the parallel processing unit based on the resource allocation mask map, so that the hardware resources are spatially topologically aligned with the high-computational-load regions of the original non-uniformly distributed grid data stream. to Eliminate processing jitter caused by instruction queuing within the time frame.

[0007] Preferably, the edge control unit includes a distributed logical buffer module for performing short-term cyclic overwrite storage on important data streams; the distributed logical buffer module is based on information entropy weights. Adjust the storage cycle of the cyclic overwrite.

[0008] Preferably, the attribute recognition algorithm uses a convolutional neural network to extract morphological features from the original non-uniformly distributed grid data stream; the morphological features are used to identify the target grid points, non-background noise, and high-load computational core areas as features.

[0009] Preferably, when the system executes the real-time parallel distribution mode, it forces important data streams to occupy high-speed physical links for priority transmission; when the system executes the asynchronous buffer distribution mode, it stores background data streams in a preset buffer pool and uses the idle time of the high-speed physical link for asynchronous retransmission.

[0010] Preferably, when extracting the amplitude gradient operator, the edge control unit calculates the sampling grid points at... direction and gradient magnitude in the direction : ,in, Coordinates are The amplitude intensity value of the sampling grid points; the mean gradient value of the sampling grid points. For all gradient magnitudes within the preset sampling region The arithmetic mean.

[0011] Preferably, before distribution, the system performs lossless compression based on operator features on important data streams and lossy compression based on reduced quantization accuracy on background data streams.

[0012] Preferably, the system includes a load balancing module for monitoring the overall load rate of the edge control unit and the heterogeneous computing unit, and for migrating the logical judgment task from the parallel processing unit to the general processing unit when the load rate exceeds a preset security threshold.

[0013] Preferably, the synchronous monitoring unit uses preset probe frames to perform real-time calibration of the end-to-end delay of the high-speed physical link to obtain the link transmission period. and link transmission period Feedback is sent to the edge control unit.

[0014] Preferably, the heterogeneous computing unit includes a general-purpose processor and a graphics processor; the general-purpose processor is used to execute the task splitting logic of the original non-uniformly distributed grid data stream, and the graphics processor executes the task mapping corresponding to the computation cost estimation tensor.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. In radar echo field data, the original echo field data stream is sparsely sampled and the amplitude gradient operator between grid points is extracted to generate a computation cost prediction tensor for quantizing spatial non-uniform features. This enables edge nodes to be pre-loaded with computational pressure awareness before task mapping, avoiding instruction queuing and processing jitter caused by the a posteriori feedback of traditional scheduling logic load. High-density data streams are aligned in real time at the edge. 2. The predictive resource pre-configuration mechanism is coupled with the stream processor thread configuration. The scheduling unit adjusts the memory pre-occupancy instructions and thread block size according to the computation cost weight. Data grids in the core echo region are pre-loaded into the cache. The hardware execution level offsets the memory access latency caused by high-density computing. The heterogeneous computing unit exhibits deterministic throughput performance when processing sudden surge data in strong convection regions. 3. Construct a dynamic closed loop linking computational feedback latency and distribution strategy. When processing high information entropy echo regions, adjust the non-regional mask dimensionality reduction depth in real time, release bandwidth resources to compensate for core strong echo streams, and adaptively align edge computing load with link transmission bandwidth to avoid buffer overflows and system crashes caused by instantaneous data accumulation under extreme weather conditions. Information entropy prediction, heterogeneous resource pre-setting, and distribution mask control form a multi-dimensional collaboration, transforming data stream distribution into computational resource-aware adaptive distribution. The edge performs prior identification of computational costs and dynamic matching of physical resource topology, optimizing the processing sequence of complex spatial feature data without changing the inherent computing power of the hardware, resulting in high processing efficiency and engineering reliability. Attached Figure Description

[0016] Figure 1 This is a block diagram illustrating the overall architecture and data interaction principle of the present invention; Figure 2 This is a statistical chart comparing the data volume of important and background data streams under different compression strategies of this invention; Figure 3 This is a schematic diagram of the execution flow of collaborative control and resource distribution of each unit in this invention. Detailed Implementation

[0017] The following embodiments are provided to illustrate the present invention in detail, and are intended to explain the present invention, but do not limit the scope of protection of the present invention.

[0018] This invention provides an edge parallel distribution system for weather radar echo field data, comprising an edge control unit, a heterogeneous computing unit connected to the edge control unit, and a synchronization monitoring unit connected to both the edge control unit and the heterogeneous computing unit. The edge control unit receives the original non-uniformly distributed grid data stream and extracts the amplitude gradient operator, calculating the average gradient of the sampled grid points to generate a computational cost estimation tensor. The heterogeneous computing unit adjusts the thread block size in the parallel processing unit according to this computational cost estimation tensor and offsets memory access latency through memory pre-occupancy instructions. The synchronization monitoring unit monitors the feedback latency of the parallel processing unit and triggers a mask dimensionality reduction operation for non-regional areas when the latency exceeds a preset threshold of the link transmission cycle. This system achieves deterministic throughput of high-density data streams at the edge through a predictive resource pre-configuration mechanism that senses the spatial topology characteristics of the data. For the spatial non-uniform distribution characteristics generated by multi-beam scanning of weather radar, the edge control unit receives the original non-uniformly distributed grid data stream and performs a step-size operation on the data stream. Sparse sampling is used to extract the amplitude gradient operator between adjacent sampling bits; the edge control unit calculates the amplitude gradient of all sampling grid points within the sampling area. direction and gradient magnitude in the direction Gradient magnitude The calculation uses the following formula: ,in, coordinates Gradient magnitude at point, coordinates The amplitude intensity value of the sampling grid point; the edge control unit obtains the gradient magnitude values ​​of all gradient points within the preset sampling area. The arithmetic mean of the sampled grid points is used as the mean of the gradient. .

[0019] To quantify the instantaneous impact of the data field on computing resources, the edge control unit calculates the information entropy weight according to the following formula. : In the formula, For information entropy weights, The mean gradient of the sampling grid points, To preset the gradient extremum, The algorithm's logical complexity constant coefficient is preset; the edge control unit is based on the information entropy weight. A computational cost estimation tensor is generated to extract spatial distribution features of computational intensity and generate a resource allocation mask map. The edge control unit obtains hardware environment parameters through a cold start calibration program, runs a benchmark instruction set containing 1024 complex multiplication and addition operations to measure the average instruction cycle, calculates the ratio of actual feedback latency to theoretical minimum latency to calibrate the logic complexity constant coefficient, and presets the gradient extremum to three times the standard deviation of the amplitude change of the background noise flow of the radar site environment to perform normalization processing on the gradient mean. The computational cost estimation tensor is mapped to the execution dimension of the stream processor, and the number of single-thread block cores is determined according to the information entropy weight according to the formula, with the value range limited to 32 to 1024. The hardware resource distribution and echo field spatial characteristics are matched within a time sequence of 1μs to 10μs to offset the memory access latency caused by high-density computation.

[0020] The heterogeneous computing unit includes a general-purpose processor and a graphics processing unit (GPU). The general-purpose processor receives task offloading logic instructions, while the GPU executes parallel computation tasks based on a computational cost estimation tensor mapping. Before task mapping, the system determines the activation ratio of computational cores and the memory bandwidth allocation weights in the parallel processing unit according to a resource allocation mask map, ensuring topological alignment between hardware resources and high-computation-load regions of the original non-uniformly distributed grid data stream. The heterogeneous computing unit then uses information entropy weights... Dynamically adjust the thread block size of the stream processor and adjust the information entropy weight. When the preset weight threshold is exceeded, a memory pre-occupancy instruction for the grid tensor is executed, loading the grid tensor into the L1 cache; this procedure is used to offset memory access latency. to Eliminate processing jitter caused by instruction queuing within the time sequence; the synchronization monitoring unit uses preset probe frames to calibrate the end-to-end delay of the high-speed physical link in real time and obtain the link transmission period. and link transmission period Feedback is sent to the edge control unit; the synchronous monitoring unit obtains the feedback latency from the parallel processing unit. The system runs an attribute recognition algorithm, which uses a convolutional neural network to extract morphological features of the original non-uniformly distributed grid data stream. The convolutional neural network includes a feature extraction layer and a global average pooling layer, which are used to identify target grid points, non-background noise, and high-load computing core areas as features. The system divides the original non-uniformly distributed grid data stream into important data streams and background data streams based on the features.

[0021] Feedback delay Exceeding the link transmission period of At that time, the system triggers a mask dimensionality reduction operation for the corresponding region of the background data stream; the system performs distribution between real-time parallel distribution mode and asynchronous buffer distribution mode based on the marking results; in real-time parallel distribution mode, important data streams occupy high-speed physical links for priority transmission; in asynchronous buffer distribution mode, background data streams are stored in a preset buffer pool, and the system uses the idle time of the high-speed physical link for asynchronous retransmission. The system constructs a distribution adjustment closed loop based on the latency overflow rate, defines a latency deviation function, and dynamically matches the mask dimensionality reduction depth according to the value range. If the deviation function is within the first threshold interval, 1:4 spatial downsampling is triggered. This triggers a 1:8 quantization precision truncation. The attribute recognition algorithm uses a convolutional neural network to extract morphological fingerprints, identify target grid points and mark them as important data streams, while the remaining clutter region is marked as background data streams. Important data streams are forcibly pushed into the physical link priority scheduling sequence, while background data streams are written to a preset buffer pool and lossless retransmission is performed using run-length encoding during link idle periods. The resource allocation mask map guides the hardware execution level to shut down low-load computing cores, releasing processing bandwidth to compensate for core echo streams and avoid buffer overflows caused by instantaneous computation surges. The edge control unit includes a distributed logical buffer module, which performs short-term cyclic overwrite storage on important data streams and adjusts the storage according to information entropy weights. The system adjusts the storage cycle of the loop overlay; before distribution, the system performs lossless compression based on operator features for important data streams and lossy compression based on reduced quantization accuracy for background data streams; the system also includes a load balancing module, which monitors the overall load rate of the edge control unit and heterogeneous computing unit. When the load rate exceeds the preset security threshold, the logical judgment task is migrated from the parallel processing unit to the general processing unit, thereby avoiding the risk of downtime caused by the overflow of computing power of edge nodes.

[0022] Example 1: When the system faces the condition of strong convective cells intruding into the radar's core observation area, the original non-uniformly distributed grid data stream exhibits... The instantaneous peak value is at the level of [level missing], and the rate of change of amplitude intensity between grid points exhibits high spatial non-uniformity in the three-dimensional spatial domain. The edge control unit receives the original non-uniformly distributed grid point data stream in real time and executes it with a step size of [missing information]. Sparse sampling is used to extract the amplitude gradient operator between adjacent sampling points, and based on the amplitude intensity values ​​of all sampling grid points in the sampling region. According to the formula The gradient magnitude of each grid point is calculated. , coordinates Gradient magnitude at point, coordinates The edge control unit calculates the gradient magnitude by taking the amplitude intensity value of the sampling grid point. The arithmetic mean of the sampled grid gradients is obtained. And according to the formula The information entropy weight is calculated. , To preset the gradient extremum, To pre-determine the constant coefficient of the algorithm's logical complexity, the final weights are determined based on information entropy. Generate a computational cost prediction tensor for quantizing computational intensity.

[0023] The heterogeneous computing unit receives the aforementioned computational cost estimation tensor, extracts the spatial distribution features of computational intensity to generate a resource allocation mask map, and determines the activation ratio of the computational cores in the parallel processing unit based on the resource allocation mask map before the parallel processing unit executes the echo processing algorithm mapping, and adjusts the information entropy weights accordingly. In regions exceeding a preset weight threshold, the thread block size of the stream processor is dynamically adjusted. Simultaneously, a memory pre-allocation instruction is issued to preload grid tensors into the L1 cache. This predictive approach, leveraging computational cost to estimate tensor size, offsets the memory access latency caused by high-density computation, thereby achieving… to To eliminate processing jitter caused by instruction queuing within the time sequence, the synchronization monitoring unit uses probe frames to calibrate the end-to-end latency of the high-speed physical link in real time to obtain the link transmission cycle. And continuously monitor the feedback latency of the parallel processing unit. When feedback delay Exceeding the link transmission period of At that time, the system initiates a computation and distribution coordination adjustment program, uses a convolutional neural network to extract the morphological features of the original non-uniformly distributed grid data stream, identifies the target grid points and divides them into important data streams, and triggers a mask dimensionality reduction operation for the corresponding region of the background data stream. By reducing the spatial resolution of the background region, the processing bandwidth is released, ensuring that important data streams occupy high-speed physical links for parallel distribution, so that the original non-uniformly distributed grid data stream maintains a stable throughput state at the edge.

[0024] Example 2: Verification system for The test platform was used to evaluate the processing performance of non-uniform data streams. It included an edge control unit and a heterogeneous computing unit consisting of a general-purpose processor and a graphics processor. The data stream originated from the original non-uniformly distributed grid data stream generated by simulated 16-beam concurrent scanning. The grid amplitude intensity value... The dynamic range is to To simulate real electromagnetic disturbances, the system superimposes a signal-to-noise ratio of [value missing] onto the original signal. Gaussian white noise, and introduce frequency of Simulated power frequency interference; step size The design considerations are to balance feature extraction accuracy and operator generation load; when the monitored sampling frequency is higher than... At that time, step length The value is set to 4, meaning sparse sampling is performed once for every 16 adjacent grid points. After the system starts, the edge control unit performs sparse sampling on the noisy, original, non-uniformly distributed grid point data stream, according to the formula... Calculate the gradient magnitude at each grid point. And summed into the average gradient of the sampling grid points. In the control group design, the system used posterior scheduling, triggering resource allocation only after a task entered the graphics processor's execution queue. In the experimental group, however, the system used a formula... Pre-calculate information entropy weights , for , The value is 1.25, thus based on the information entropy weight. Generate the computational cost estimation tensor, see Table 1.

[0025] Table 1: Comparison of System Processing Performance under Different Information Entropy Conditions With information entropy weight Increase the feedback delay of the control group The feedback delay of the control group under the strong convection core operating condition exhibits a non-linear upward trend. The latency reached 242.8ms, accompanied by instruction queuing and frame drops. However, the experimental group, employing a predictive resource pre-allocation mechanism based on computational cost estimation tensors, loaded grid data into the L1 cache before task mapping via memory pre-allocation instructions. This offset memory access latency at the GPU kernel level, reducing feedback latency. It stabilized at 135.4ms; when When the performance inflection point of 0.90 is reached, the system automatically triggers computation and distribution co-anchoring, adjusting the mask dimensionality reduction ratio of the background data stream region from 1:1 to 1:4. By sacrificing the spatial resolution of non-core regions to gain processing bandwidth in the core strong backflow region, this demonstrates that the synergistic effect of mask control and resource pre-configuration can resolve the contradiction between computing power bottlenecks and real-time throughput. When the value exceeds 0.85, the experimental group reduces the feedback latency by executing memory pre-occupancy instructions for grid tensors and adjusting the storage cycle of the distributed logic buffer module. Maintaining at the link transmission cycle of Below, and when When the over-limit state is increased to 1.03, although the system's feedback latency increases, the important data streams still maintain priority transmission in the real-time distribution mode through the 1:8 mask dimensionality reduction depth.

[0026] Example 3: This example combines Figures 1 to 3 This document describes an edge parallel distribution system for weather radar echo field data, such as... Figure 1 As shown, the original non-uniformly distributed grid data stream is transmitted to the edge control unit, where sparse sampling and gradient extraction are performed. Specifically, this includes extracting the amplitude gradient operator of adjacent sampling bits and calculating the average gradient of the sampling grid points, generating a computational cost estimate, calculating the information entropy weight, and generating a computational cost estimate tensor. This computational cost estimate tensor is transmitted to the heterogeneous computing unit, which performs resource pre-configuration and mapping, adjusts the stream processor thread block size according to the information entropy weight, and eliminates memory access latency by issuing memory pre-occupancy instructions and loading the grid tensor into the L1 cache. The feedback latency generated by the heterogeneous computing unit... Feedback is sent to the synchronization monitoring unit, which performs real-time latency monitoring, including monitoring the latency feedback from the parallel processing unit and comparing it with the link transmission cycle threshold. When the conditions are met, dynamic dimensionality reduction is triggered, i.e., non-regional trigger mask dimensionality reduction. The trigger mask dimensionality reduction instruction issued by the synchronization monitoring unit enters the distribution mode execution logic. With the support of attribute recognition, it performs feature recognition and marking, divides the data stream into important data streams and background data streams, and switches between real-time parallel distribution and asynchronous buffer distribution according to the marking results, ultimately achieving deterministic throughput of high-density data streams on the edge side.

[0027] like Figure 2As shown in the chart, this graph displays statistical results with data volume (MB) on the vertical axis and compression strategy on the horizontal axis. The legend distinguishes between the MB of important data streams and the MB of background data streams. Under the original data compression strategy, the MB value of the background data stream is higher than that of the important data stream. Under the lossless compression strategy, the MB values ​​of both the important and background data streams change. Under the lossy compression strategy, the MB value of the background data stream decreases significantly, while the MB of the important data stream remains at a low data volume level. Figure 3 As shown, the process begins with the edge control unit executing the processing logic branches of extracting amplitude gradient operators, calculating information entropy weights, and generating computation cost estimation tensors. These branches include the heterogeneous computing unit executing adjustments to thread block sizes, executing memory pre-occupancy instructions, and offsetting memory access latency. Simultaneously, the synchronous monitoring unit executes monitoring feedback latency, non-regional mask dimensionality reduction, and feature recognition. The system environment support module performs end-to-end latency real-time calibration in parallel. All branch logic converges on the distribution and resource strategy, which includes switching between real-time parallel and asynchronous buffer modes, prioritizing the transmission of important data streams, and aligning computational intensity spatial topology. The final process leads to deterministic throughput of high-density data streams on the edge side.

[0028] Example 4: System cold start calibration program determines the constant coefficient of preset algorithm logic complexity. The initial values ​​are: The hardware environment includes an 8-core general-purpose processor and a graphics processor with 2560 stream processing cores; the edge control unit obtains the theoretical peak floating-point operations and memory bandwidth of the heterogeneous computing unit; the edge control unit runs a benchmark instruction set containing 1024 complex multiplication and addition operations in an idle state and records its average instruction cycle and cache hit rate; the system determines the logic complexity constant coefficient by calculating the ratio of the actual execution feedback latency to the theoretical shortest feedback latency. Under this hardware configuration, after three cyclic samplings and averaging, the constant coefficient of the logical complexity is... The calibration result is 1.25. When the system processes the original non-uniformly distributed grid data stream, the edge control unit converts the three-dimensional spatial echo field into a... and A two-dimensional discrete grid coordinate system with index ; to achieve a step size of For sparse sampling, the system establishes a sampling index vector with a step size of 4, and only performs sampling when the following conditions are met. and Extracting amplitude intensity value at coordinate point When extracting the amplitude gradient operator, for grid points located at the edge of the discrete grid, the system uses the formula... Calculate the gradient magnitude Ensure the gradient magnitude at the data flow boundary The computation is continuous, generating a computational cost prediction tensor for quantifying the computational intensity.

[0029] The system uses a convolutional neural network to extract features, which contains one layer with 16 features. The feature extraction layer of the convolutional kernel uses a modified linear unit as the activation function to convert grid point amplitude intensity values ​​into neuron activation levels; the synchronization monitoring unit monitors the feedback delay. Exceeding the link transmission period of At that time, the system performs mask dimensionality reduction for the corresponding region of the background data stream; and uses a bilinear interpolation downsampling algorithm to reduce the spatial resolution of non-regional areas from... Reduce to With a mask dimensionality reduction depth of 1:4, the average amplitude values ​​of every 16 adjacent background grid points are calculated and then transmitted as the mean operator; the shared memory of the graphics processor reserves a size of The mask mapping table, hardware execution resources and data space topology features in to Alignment is completed within the time frame.

[0030] Example 5: In the scenario of initial power-on deployment of an edge node at a new radar site, the system runs a hardware environment awareness program to calibrate computing power parameters. The edge control unit sends a baseline computation sequence containing 512 samples to the heterogeneous computing unit. The synchronous monitoring unit records the instruction response time of the general-purpose processor in processing task distribution logic and the total number of instruction cycles of the graphics processor in processing grid parallel computation. Based on the hardware pipeline occupancy rate of the graphics processor when processing data streams, the system calculates the computing power parameters of the node. The system correlates these computing power parameters with the standard floating-point arithmetic capability to calculate the preset algorithm logic complexity constant coefficient for the current hardware architecture. Simultaneously, the background noise stream under radar silence conditions is collected and the amplitude change between sampling points is calculated. Three times the standard deviation of this amplitude change is taken as the preset gradient extreme value. The baseline value is used to correct for deviations in the generation of computation cost prediction tensors by different hardware architectures.

[0031] When the system enters the clutter suppression and extraction mode for a specific geographical area, it retrieves 1024 sets of labeled samples containing strong convection echoes, layered cloud echoes, and background clutter from the local storage module. The feature extraction layer of the convolutional neural network extracts the amplitude intensity values ​​of the sampled grid points. The size of the receptive field is The convolution operation is performed and morphological fingerprints are extracted. The obtained morphological fingerprints are stored in a static lookup table. The synchronization monitoring unit uses a 64-byte probe frame to perform an end-to-end bidirectional latency test on the high-speed physical link to obtain the link transmission period. The system calculates the mean and variance of the morphological fingerprint and adjusts the trigger sensitivity of the trigger mask dimensionality reduction operation based on the variance. At this point, the system combines the morphological fingerprint with the feedback delay from real-time monitoring. The association enables resource pre-configuration and distribution path switching to be performed based on the tensor estimated according to the computational cost when the original non-uniformly distributed grid data stream is input.

[0032] Example 6: In a radar station deployment scenario with an integrated edge computing card, the system executes a threshold calibration procedure for the attribute recognition algorithm. The edge control unit inputs an offline dataset containing 256 echo samples into the convolutional neural network. By recording the neuron activation energy values ​​output by the feature extraction layer, the system statistically analyzes the energy distribution and calculates the probability density function, setting the threshold within this distribution. The confidence level location is used to identify high-load computational core areas within the features. Based on these features, the system divides the original non-uniformly distributed grid data stream into important and background data streams, achieving labeling based on objective level indicators. This is relevant when the system faces information entropy weighting. When changes trigger computational surges, the thread block size of the heterogeneous computing unit executes the stream processor. The dynamic mapping is performed by the system to estimate the tensor based on the computational cost, according to the formula. Determine the execution dimension of the stream processor in the parallel processing unit. This refers to the number of cores contained in a single thread block, with a value ranging from 32 to 1024. The system targets... For regions exceeding the preset weight threshold, a memory pre-allocation instruction is issued to preload the relevant grid tensors into the L1 cache. Simultaneously, the distributed logical buffer module sets the storage period of important data streams to their information entropy weight. The inversely proportional values ​​realize the spatial topology distribution of hardware resource distribution and the original non-uniform grid data stream. to Alignment within the time frame.

[0033] In bandwidth-constrained data distribution scenarios, the edge control unit executes differentiated compression procedures for important and background data streams. The system extracts amplitude intensity distribution features from the important data stream and uses a run-length encoding algorithm to convert the feature data into a compact bit stream. For the background data stream, the system employs a lossy compression method that reduces quantization precision, compressing the sampling bit width from 16 bits to 4 bits. When the synchronization monitoring unit detects a link transmission cycle... More than When there is a positive offset, the distributed logical buffer module calculates the information entropy weight. The storage cycle of critical data streams is shortened. Memory page frames corresponding to low-information-entropy regions are reclaimed to ensure data backtracking depth in high-load computing core areas. Priority link allocation is implemented for original, non-uniformly distributed grid data streams under physical bandwidth bottleneck conditions. In redundant protection conditions where edge node computing power is at critical overload, the system runs the task migration procedure of the load balancing module. The overall load rate of edge control units and heterogeneous computing units is obtained through real-time monitoring and this parameter is marked as... When the overall load rate Reaching the preset security threshold At that time, the synchronous monitoring unit issues an interrupt command to suspend the non-real-time logic branches in the parallel processing unit with a priority lower than that of the feature extraction task. The system then migrates the background data filtering task, originally mapped to the graphics processor, to a task with a main frequency of [missing information]. The general-purpose processor cores are executed. This procedure offsets the transient pressure caused by computational surges through dynamic task refactoring without changing the inherent computing power of the hardware, ensuring that the computing core resources of the parallel processing unit are concentrated on executing resource pre-set instructions guided by the computation cost estimation tensor.

[0034] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0035] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An edge parallel distribution system for weather radar echo field data, characterized in that, Includes edge control units, heterogeneous computing units, and synchronous monitoring units: The edge control unit is connected to the heterogeneous computing unit, and the synchronization monitoring unit is connected to both the edge control unit and the heterogeneous computing unit. The edge control unit receives the raw, non-uniformly distributed grid data stream and performs operations on the raw, non-uniformly distributed grid data stream with a step size of [missing information]. Sparse sampling is used to extract the amplitude gradient operator between adjacent sampling points, and the mean gradient of the sampling grid points is calculated. The edge control unit is based on the formula Calculate information entropy weights And based on information entropy weights Generate a computational cost prediction tensor; where, To preset the gradient extremum, The constant coefficient is the preset algorithm logic complexity. Heterogeneous computing units are used to perform task mapping based on information entropy weights. Adjust the thread block size of the stream processor in the parallel processing unit, and adjust the information entropy weight. When the preset weight threshold is exceeded, a memory pre-occupancy instruction for the grid tensor is executed to load the grid tensor into the first-level cache; The synchronous monitoring unit is used to obtain the feedback latency of the parallel processing unit. The system performs labeling, uses an attribute recognition algorithm to identify features in the original non-uniformly distributed grid data stream, and divides the original non-uniformly distributed grid data stream into important data streams and background data streams based on these features; during feedback delay... Exceeding the link transmission period of At that time, the system performs mask dimensionality reduction for the corresponding region of the background data stream, and performs distribution between real-time parallel distribution mode and asynchronous buffer distribution mode based on the labeling results.

2. The edge parallel distribution system for weather radar echo field data according to claim 1, characterized in that, The edge control unit extracts the spatial distribution features of computational intensity based on the computational cost estimate tensor and generates a resource allocation mask map; Heterogeneous computing units determine the activation ratio of computing cores and the memory bandwidth allocation weights in parallel processing units based on resource allocation masks, enabling spatial topology alignment of hardware resources with high-computation-load regions of the original non-uniformly distributed grid data stream. to Eliminate processing jitter caused by instruction queuing within the time frame.

3. The edge parallel distribution system for weather radar echo field data according to claim 1, characterized in that, The edge control unit includes a distributed logical buffer module for performing short-term cyclic overwrite storage on critical data streams; the distributed logical buffer module is based on information entropy weights. Adjust the storage cycle of the cyclic overwrite.

4. The edge parallel distribution system for weather radar echo field data according to claim 1, characterized in that, The attribute recognition algorithm uses a convolutional neural network to extract morphological features from the original non-uniformly distributed grid data stream; the morphological features are used to identify the target grid points, non-background noise, and high-load computational core areas.

5. An edge parallel distribution system for weather radar echo field data according to claim 4, characterized in that, When the system executes the real-time parallel distribution mode, it forces important data streams to occupy high-speed physical links for priority transmission; when the system executes the asynchronous buffer distribution mode, it stores background data streams in a preset buffer pool and uses the idle time of the high-speed physical link for asynchronous retransmission.

6. The edge parallel distribution system for weather radar echo field data according to claim 1, characterized in that, When extracting the amplitude gradient operator, the edge control unit calculates the sampling grid points at... direction and gradient magnitude in the direction : ,in, Coordinates are The amplitude intensity value of the sampling grid points; the mean gradient value of the sampling grid points. For all gradient magnitudes within the preset sampling region The arithmetic mean.

7. The edge parallel distribution system for weather radar echo field data according to claim 1, characterized in that, Before distribution, the system performs lossless compression based on operator features on important data streams and lossy compression based on reduced quantization accuracy on background data streams.

8. The edge parallel distribution system for weather radar echo field data according to claim 1, characterized in that, The system includes a load balancing module, which monitors the overall load rate of the edge control unit and the heterogeneous computing unit, and migrates the logical judgment task from the parallel processing unit to the general processing unit when the load rate exceeds a preset security threshold.

9. An edge parallel distribution system for weather radar echo field data according to claim 1, characterized in that, The synchronous monitoring unit uses preset probe frames to perform real-time calibration of the end-to-end delay of the high-speed physical link to obtain the link transmission period. and link transmission period Feedback is sent to the edge control unit.

10. An edge parallel distribution system for weather radar echo field data according to claim 1, characterized in that, The heterogeneous computing unit includes a general-purpose processor and a graphics processor; the general-purpose processor is used to execute the task routing logic of the original non-uniformly distributed grid data stream, and the graphics processor executes the task mapping corresponding to the computation cost estimation tensor.

Images