Weather radar coverage performance analysis method and system based on deep learning

Abstract

The embodiment of the invention provides a weather radar coverage performance analysis method and system based on deep learning, and the method comprises the steps: firstly, extracting a model through deep learning, automatically extracting key features from a large amount of complex weather radar data, generating a target radar coverage representation vector, and effectively avoiding the subjectivity and low efficiency of conventional manual analysis. And then, in combination with the geographic position information of the radar, the spatial relation characteristics are further fused by using a spatial correlation fusion model, and a fusion vector containing spatial correlation information is generated, so that an evaluation result is closer to an actual detection environment. The fusion vector is subjected to intelligent clustering through the clustering model, and the key weather radar set is determined based on the clustering concentration degree, so that a scientific and accurate basis is provided for meteorological monitoring, early warning and resource scheduling. Therefore, not only is the analysis process of the weather radar coverage efficiency optimized, but also the depth and the breadth of analysis are remarkably improved, so that the meteorological service level is improved, and the disaster response capability is enhanced.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to a weather radar coverage effectiveness analysis method and system based on deep learning. Background Art

[0002] In the field of meteorological monitoring, weather radar, as an important observation tool, plays a vital role in real-time tracking and forecasting of weather changes. However, with the continuous expansion of the meteorological monitoring network, the number of weather radars has increased dramatically. How to effectively evaluate the coverage efficiency of these radars and ensure the comprehensiveness and accuracy of monitoring data has become an urgent problem to be solved.

[0003] Traditionally, the evaluation of weather radar coverage effectiveness mainly relies on manual analysis and empirical judgment. This method is not only time-consuming and labor-intensive, but also easily affected by subjective factors, resulting in instability and inconsistency in the evaluation results. In addition, with the continuous development of weather radar technology, the amount and complexity of radar data are also increasing. Traditional methods are unable to cope with large-scale, multi-dimensional data.

[0004] In recent years, with the rise of artificial intelligence technology, especially the rapid development of deep learning technology, new possibilities have been provided for the automated evaluation of weather radar coverage effectiveness. Deep learning models have powerful feature extraction and pattern recognition capabilities, and can automatically learn and extract useful information from large amounts of complex data, providing strong support for intelligent analysis and processing of data.

[0005] However, the relevant technologies still have certain limitations. On the one hand, most methods only focus on the intrinsic characteristics of radar data, ignoring the important impact of radar geographic location information on coverage efficiency; on the other hand, these methods often lack in-depth analysis and utilization of spatial correlation, resulting in evaluation results that are difficult to fully reflect the overall effectiveness of the radar network. Summary of the invention

[0006] In order to at least overcome the above-mentioned deficiencies in the prior art, an object of the embodiments of the present application is to provide a weather radar coverage effectiveness analysis method and system based on deep learning.

[0007] According to one aspect of the present application, a weather radar coverage effectiveness analysis method based on deep learning is provided, the method comprising:

[0008] Acquire X weather radar data records and X geographic location information of X target weather radars, wherein each weather radar data record in the X weather radar data records includes coverage effectiveness statistical data and radar echo sequence data, the coverage effectiveness statistical data reflects data generated after performing detection coverage performance statistics on the target weather radar in a target test time domain interval, and the radar echo sequence data reflects radar echo data generated based on target radar detection activities of the target weather radar in the target test time domain interval, and X is a positive integer not less than 1;

[0009] Inputting the X weather radar data records into a deep learning extraction model to generate X target radar coverage representation vectors;

[0010] Inputting the X target radar coverage characterization vectors and the X geographic location information into a spatial correlation fusion model to generate X spatial correlation fusion vectors;

[0011] Clustering the X spatial correlation fusion vectors based on a clustering model to generate Y feature vector clusters, where Y is a positive integer not less than 1;

[0012] The concentration levels of the Y feature vector clusters are calculated, and a target weather radar set is determined based on the concentration levels of the Y feature vector clusters, wherein the target weather radar set includes at least one of the target weather radars.

[0013] In a possible implementation of the first aspect, the X weather radar data records include X coverage effectiveness statistical data and X radar echo sequence data, and the deep learning extraction model includes an effectiveness feature encoder, an echo feature encoder, and a fusion control unit;

[0014] The step of inputting the X weather radar data records into a deep learning extraction model to generate X target radar coverage characterization vectors includes:

[0015] Inputting the X coverage performance statistical data into the performance feature encoder to generate X performance feature encoding results;

[0016] Inputting the X radar echo sequence data into the echo feature encoder to generate X echo feature encoding results;

[0017] The X effectiveness feature encoding results and the X echo feature encoding results are input into the fusion control unit to generate X target radar coverage characterization vectors.

[0018] For example, in a possible implementation of the first aspect, the performance feature encoder includes a mapping unit and an interaction unit;

[0019] The step of inputting the X coverage performance statistical data into the performance feature encoder to generate X performance feature encoding results includes:

[0020] generating X performance feature fragments based on the feature distribution of the X coverage performance statistical data and the X coverage performance statistical data;

[0021] Inputting the X performance feature segments into the mapping unit to generate X basic embedding representation results, wherein the mapping unit is used to perform a vector-level summation operation on each of the X performance feature segments;

[0022] Inputting the X performance feature segments into the interaction unit to generate X advanced embedding representation results, wherein the interaction unit is used to perform a vector product operation on each of the X performance feature segments and a summation operation on the vector product operation results;

[0023] The X basic embedding representation results are integrated with the X advanced embedding representation results to generate X performance feature encoding results.

[0024] For example, in a possible implementation of the first aspect, the echo feature encoder includes a waveform conversion unit and a screening unit;

[0025] The step of inputting the X radar echo sequence data into the echo feature encoder to generate X echo feature encoding results comprises:

[0026] Inputting the X radar echo sequence data into the waveform transformation unit to generate X×W echo feature distributions, wherein the echo feature distribution is composed of at least one radar echo result in the radar echo sequence data, and W is a positive integer greater than 1;

[0027] Integrating the X×W echo feature distributions to generate X echo integration results;

[0028] The X echo integration results are input into a screening unit to generate X echo feature encoding results.

[0029] For example, in a possible implementation of the first aspect, the fusion regulation unit includes a fusion unit and a mapping output unit;

[0030] The step of inputting the X performance feature encoding results and the X echo feature encoding results into the fusion control unit to generate X target radar coverage characterization vectors includes:

[0031] Integrating each of the X performance feature encoding results with each of the X echo feature encoding results to generate X integrated encoding results;

[0032] Inputting the X integrated coding results into the fusion unit to generate X fused coding vectors;

[0033] The X fused coding vectors are input into the mapping output unit to generate X target radar coverage characterization vectors.

[0034] According to one aspect of an embodiment of the present application, a deep learning system is provided, which includes a processor and a machine-readable storage medium, wherein the machine-readable storage medium stores machine executable instructions, and the machine executable instructions are loaded and executed by the processor to implement a weather radar coverage effectiveness analysis method based on deep learning in any one of the possible implementations described above.

[0035] According to one aspect of the embodiments of the present application, a computer program product or a computer program is provided, the computer program product or the computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the methods provided in various optional implementations of the above three aspects.

[0036] In the technical solutions provided in some embodiments of the present application, the embodiments of the present application significantly improve the accuracy and efficiency of weather radar coverage effectiveness evaluation by comprehensively applying deep learning technology and spatial correlation analysis. Specifically, firstly, through the deep learning extraction model, key features are automatically extracted from a large amount of complex weather radar data to generate a target radar coverage characterization vector, which effectively avoids the subjectivity and inefficiency of traditional manual analysis. Subsequently, combined with the geographic location information of the radar, the spatial correlation fusion model is used to further fuse the spatial relationship features to generate a fusion vector containing spatial correlation information, so that the evaluation result is closer to the actual detection environment. The fusion vector is intelligently clustered through the clustering model, and the key weather radar set is determined based on the concentration of the clusters, which provides a scientific and accurate basis for meteorological monitoring, early warning and resource scheduling. As a result, not only the analysis process of weather radar coverage effectiveness is optimized, but also the depth and breadth of the analysis are significantly improved, so as to improve the level of meteorological services and enhance disaster response capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required to be used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be extracted in combination with these drawings without creative work.

[0038] Figure 1 A flowchart of a weather radar coverage effectiveness analysis method based on deep learning provided in an embodiment of the present application;

[0039] Figure 2 A schematic block diagram of the architecture of a deep learning system for implementing the above-mentioned deep learning-based weather radar coverage effectiveness analysis method provided in an embodiment of the present application. DETAILED DESCRIPTION

[0040] The following description is intended to enable one of ordinary skill in the art to implement and combine the present application, and the description is provided in the context of a specific application scenario and its requirements. It will be apparent to one of ordinary skill in the art that various changes may be made to the disclosed embodiments, and that the general principles defined in the present application may be applied to other embodiments and application scenarios without departing from the principles and scope of the present application. Therefore, the present application is not limited to the described embodiments, but should be given the broadest scope consistent with the claims.

[0041] Figure 1 This is a flow chart of a weather radar coverage effectiveness analysis method based on deep learning provided by an embodiment of the present application. The weather radar coverage effectiveness analysis method based on deep learning is introduced in detail below.

[0042] Step S110, obtaining X weather radar data records and X geographic location information of X target weather radars, wherein each weather radar data record of the X weather radar data records includes coverage effectiveness statistical data and radar echo sequence data, the coverage effectiveness statistical data reflects data generated after performing detection coverage performance statistics on the target weather radar within a target test time domain interval, and the radar echo sequence data reflects radar echo data generated based on the target radar detection activity of the target weather radar within the target test time domain interval, and X is a positive integer not less than 1.

[0043] In detail, the target weather radar refers to a radar device used for meteorological observation, which can detect meteorological information such as precipitation, wind speed, wind direction, etc., and is usually distributed in different geographical locations to cover a wider area. In actual applications, the target weather radar detects meteorological information by emitting electromagnetic waves and receiving their echoes. The radar data is transmitted to the data center through a dedicated network for processing and analysis.

[0044] The weather radar data record contains all relevant information collected by the radar within a specific time period (target test time domain interval), mainly including coverage efficiency statistics and radar echo sequence data. Among them, the target radar equipment regularly records and transmits its working status and detection results. Coverage efficiency statistics such as detection range, accuracy, update frequency, etc. are automatically calculated by the radar system; radar echo sequence data records the detailed echo information of each detection activity.

[0045] The coverage effectiveness statistics reflect the statistical data of the detection coverage performance of the target weather radar within the target test time domain interval, including the effective detection range, detection accuracy, data update frequency, etc. These data are automatically generated by the radar system and are obtained by analyzing the detection results and working status of the radar through algorithms. For example, the effective detection range can be determined by analyzing the farthest distance that the radar can accurately detect; the detection accuracy is based on the comparison between the radar's recognition results of meteorological phenomena and the actual situation.

[0046] The radar echo sequence data is a radar echo data sequence generated based on the radar's detection activities within the target test time domain interval, and records the echo intensity, distribution and other information detected each time. Among them, the target weather radar detects meteorological targets by emitting electromagnetic waves and receiving the echoes reflected back. Each detection activity will generate a series of echo data, including parameters such as echo intensity, position, and speed. These data are recorded to form radar echo sequence data.

[0047] That is, in this embodiment, it is assumed that there are 5 target weather radars in the meteorological monitoring system connected to the server, that is, X=5. These 5 target weather radars are located in different geographical locations. For example, radar 1 is located in the suburbs of city A, radar 2 is located near the airport of city B, radar 3 is located at a monitoring site in the mountainous area, radar 4 is located at a coastal port, and radar 5 is located in an agricultural area in the plain.

[0048] For radar 1, within the target test time domain interval (for example, the past week), its coverage efficiency statistics include the following: Effective detection range: the maximum distance at which meteorological information can be accurately detected is 200 kilometers; Detection accuracy: the accuracy of precipitation type judgment is 85%; Data update frequency: meteorological data is updated every 5 minutes.

[0049] Its radar echo sequence data reflects the radar echo data generated by each detection activity during this week. For example, at a certain time of a certain day, a strong echo was detected, indicating that there may be heavy precipitation.

[0050] For Radar 2, coverage effectiveness statistics may include: effective detection range: 150 km; detection accuracy: 90% accuracy in judging thunderstorm weather; data update frequency: updated every 3 minutes. The radar echo sequence data may record the echo of low-level wind shear detected multiple times near the airport.

[0051] The coverage effectiveness statistics of radar 3 include: effective detection range: 180 kilometers, but the detection range in some directions is reduced due to the influence of mountainous terrain; detection accuracy: the detection accuracy of fog in mountainous areas is 70%; data update frequency: once every 6 minutes. The echo sequence data includes the detection results of airflow changes in complex mountainous terrain.

[0052] Radar 4’s relevant data include: effective detection range: 250 kilometers, which is larger due to the open coastal terrain; detection accuracy: 88% accuracy for typhoon warnings; data update frequency: once every 4 minutes. The echo sequence data records the echo characteristics of multiple typhoons passing through.

[0053] The data of radar 5 may be: effective detection range: 220 kilometers; detection accuracy: 75% accuracy in monitoring drought weather; data update frequency: once every 5 minutes. The echo sequence data reflects the changes in weak echoes in the absence of significant precipitation for a long time.

[0054] At the same time, each radar also has corresponding geographic location information, including precise latitude and longitude coordinates, administrative region, etc. The server communicates with the data interface of each radar to obtain the weather radar data records and geographic location information of the five target weather radars, preparing for subsequent analysis and processing.

[0055] Step S120: input the X weather radar data records into a deep learning extraction model to generate X target radar coverage characterization vectors.

[0056] In detail, the deep learning extraction model is a model built using deep learning technology to extract useful information from raw radar data and generate a target radar coverage characterization vector.

[0057] The target radar coverage characterization vector is a vector generated by processing radar data through a deep learning extraction model, and is used to represent the coverage characteristics and detection capabilities of the radar. The target radar coverage characterization vector is the result of the output of the deep learning extraction model, and contains comprehensive information about the radar coverage effectiveness and echo characteristics. The target radar coverage characterization vector can be used to compare and analyze the detection capabilities of different radars.

[0058] In this embodiment, after the server obtains the weather radar data records of the five target weather radars, it inputs these weather radar data records into a pre-trained deep learning extraction model.

[0059] Among them, the deep learning extraction model can be a combined model based on a convolutional neural network (CNN) and a recurrent neural network (RNN).

[0060] For the data of radar 1, the deep learning extraction model first processes the coverage effectiveness statistics. The convolution layer extracts the features in the statistics, such as the numerical features of the effective detection range, detection accuracy, and data update frequency. Then, for the radar echo sequence data, the RNN layer can capture the time series features in the echo data, such as the changing trend of the echo intensity and the frequency of occurrence. After multi-layer neural network calculation and feature fusion, a characterization vector that can comprehensively reflect the coverage characteristics of radar 1 is finally generated.

[0061] The same process is performed for the data from Radar 2. The model extracts the corresponding features and generates its unique characterization vector based on the unique coverage effectiveness statistics and echo sequence data of Radar 2.

[0062] By analogy, the server processes the data of the five radars in this way, and finally obtains five target radar coverage characterization vectors.

[0063] Step S130: input the X target radar coverage characterization vectors and the X geographic location information into a spatial correlation fusion model to generate X spatial correlation fusion vectors.

[0064] The spatial correlation fusion model is a model for fusing the radar's coverage characterization vector with geographic location information to generate a fusion vector containing spatial correlation information. The spatial correlation fusion model may be constructed based on a graph convolutional neural network (GCN) or other spatial relationship learning algorithms. The spatial correlation fusion vector is a vector generated after processing the radar coverage characterization vector and geographic location information through the spatial correlation fusion model, and contains the correlation characteristics of the radar in space. The spatial correlation fusion vector is the output result of the spatial correlation fusion model, which integrates the coverage characteristics of the radar and its positional relationship in space. Through the spatial correlation fusion vector, the mutual influence and cooperative relationship of different radars in space can be analyzed.

[0065] That is, in this embodiment, the server inputs the five target radar coverage characterization vectors generated previously and the corresponding five geographic location information into the spatial correlation fusion model. It is assumed that the spatial correlation fusion model is built based on a graph convolutional neural network (GCN). For the characterization vector and geographic location information of radar 1, the spatial correlation fusion model converts its geographic location information into node features in a graph structure, and then performs a graph convolution operation together with the radar coverage characterization vector. In this way, the coverage characteristics of radar 1 itself and its positional relationship in space are fused to generate a fusion vector that integrates spatial correlation information.

[0066] Similar processing is performed for radars 2-5 so that each radar generates a vector containing its coverage characteristics and spatial associations.

[0067] Step S140: clustering the X spatial correlation fusion vectors based on a clustering model to generate Y feature vector clusters, where Y is a positive integer not less than 1.

[0068] In this embodiment, the server uses a clustering model to cluster the five spatial correlation fusion vectors. Assume that the clustering model is implemented based on a clustering algorithm, such as the KMeans algorithm. First, three initial cluster centers are randomly selected. Then, the distance from each spatial correlation fusion vector to the three cluster centers is calculated, and it is assigned to the cluster with the closest cluster center.

[0069] After the first round of allocation is completed, the center of each cluster is recalculated. For the spatial correlation fusion vector of radar 1, it is assumed that it is allocated to cluster 1. For the spatial correlation fusion vector of radar 2, after calculating the distance, it may be allocated to cluster 2. And so on, the allocation of all 5 vectors and the recalculation of cluster centers are completed. After multiple iterations, the 5 spatial correlation fusion vectors are finally divided into 3 feature vector clusters, that is, Y = 3.

[0070] Step S150, calculating the concentration levels of the Y feature vector clusters, and determining a target weather radar set based on the concentration levels of the Y feature vector clusters, wherein the target weather radar set includes at least one of the target weather radars.

[0071] In this embodiment, for the three generated feature vector clusters, the server starts to calculate their concentration levels.

[0072] First, obtain the spatial correlation fusion vector in each feature vector cluster. For the first cluster, assume that it contains 2 vectors, calculate the average value of these 2 vectors respectively, and use them as the reference node of the cluster. Then calculate the degree of deviation of each vector in this cluster from the reference node, calculate it through some distance metric (such as Euclidean distance), and take the average value as the concentration degree of this cluster. For the second cluster, assume that it contains 1 vector, which itself serves as a reference node and has a concentration degree of 0. For the third cluster, assume that it contains 2 vectors, and calculate the concentration degree in the same way. After calculating the concentration degrees of the three clusters, determine the target weather radar set based on these concentration degrees.

[0073] Assume that a rule is set that if the concentration degree of a cluster is greater than a certain threshold (eg, 0.8), the radars in the cluster are determined as a target weather radar set.

[0074] Assuming that the concentration levels of the first cluster and the third cluster are both greater than the threshold, the radars in these two clusters (assuming they are radar 1, radar 3 and radar 4, radar 5) constitute the target weather radar set.

[0075] Thus, the whole process from acquiring weather radar data to determining the target weather radar set is completed. In practical applications, these results can provide important decision support for weather forecasts, disaster warnings, etc.

[0076] Based on the above steps, the embodiment of the present application significantly improves the accuracy and efficiency of weather radar coverage efficiency evaluation by comprehensively applying deep learning technology and spatial correlation analysis. Specifically, first, through the deep learning extraction model, key features are automatically extracted from a large amount of complex weather radar data to generate a target radar coverage characterization vector, which effectively avoids the subjectivity and inefficiency of traditional manual analysis. Subsequently, combined with the geographic location information of the radar, the spatial correlation fusion model is used to further fuse the spatial relationship features to generate a fusion vector containing spatial correlation information, so that the evaluation result is closer to the actual detection environment. The fusion vector is intelligently clustered through the clustering model, and the key weather radar set is determined based on the concentration of the clusters, which provides a scientific and accurate basis for meteorological monitoring, early warning and resource scheduling. As a result, not only the analysis process of weather radar coverage efficiency is optimized, but also the depth and breadth of the analysis are significantly improved, so as to improve the level of meteorological services and enhance disaster response capabilities.

[0077] In a possible implementation, calculating the concentration levels of the Y feature vector clusters includes:

[0078] Step S151, obtaining the spatial correlation fusion vector in each of the Y feature vector clusters.

[0079] In this embodiment, the server first accesses the data warehouse storing the feature vector clusters and retrieves Y feature vector clusters therefrom. Each feature vector cluster contains a certain number of spatial correlation fusion vectors, which are generated in the previous step by inputting the coverage characterization vector and geographic location information of the target weather radar into the spatial correlation fusion model.

[0080] For example, suppose Y=3, that is, there are three feature vector clusters. The server retrieves these three clusters from the data warehouse and marks them as cluster A, cluster B, and cluster C. Next, the server processes each cluster separately to obtain the spatial correlation fusion vector.

[0081] For cluster A, the server finds that it contains five spatially associated fusion vectors, labeled vector 1, vector 2, vector 3, vector 4, and vector 5. The server loads these vectors into memory for subsequent calculations.

[0082] Similarly, for clusters B and C, the server also obtains the spatial correlation fusion vectors they contain. Assume that cluster B contains 3 vectors (vector 6, vector 7, vector 8), and cluster C contains 4 vectors (vector 9, vector 10, vector 11, vector 12).

[0083] Step S152, calculating the average vector of the spatial correlation fusion vector in each of the feature vector clusters to generate a cluster reference node.

[0084] Next, the server needs to calculate the spatial correlation fusion vector in each feature vector cluster to generate a cluster reference node. The cluster reference node is a virtual node that represents the average position and direction of all vectors in the cluster.

[0085] For cluster A, the server first calculates the average of vectors 1 to 5. This can be done by adding the corresponding components of each vector and then dividing by the number of vectors. For example, if each vector is a three-dimensional vector, the server will calculate the average of the x, y, and z components separately to obtain a three-dimensional average vector as the reference node of cluster A.

[0086] Similarly, the server will calculate the average vector for clusters B and C and generate their reference nodes. In this way, each cluster has a reference node that represents its overall position and direction.

[0087] Step S153, calculating the average deviation between the spatial correlation fusion vector in each of the Y feature vector clusters and the cluster reference node, and generating the concentration degree of the Y feature vector clusters.

[0088] After obtaining the cluster reference node, the server then needs to calculate the average deviation between the spatial correlation fusion vector in each feature vector cluster and the corresponding cluster reference node. This deviation reflects the degree of difference between the vector in the cluster and the reference node, that is, the concentration of the cluster.

[0089] For cluster A, the server calculates the deviations between vectors 1 to 5 and the reference node of cluster A. This is done by calculating the Euclidean distance between each vector and the reference node. The server then adds these deviations together and divides by the number of vectors to get the average deviation for cluster A.

[0090] Similarly, the server will calculate the average deviation for clusters B and C. In this way, each cluster has an average deviation value that represents its degree of centralization.

[0091] Finally, the server generates the concentration of Y feature vector clusters based on the calculated average deviation value. The concentration is a value between 0 and 1, which indicates the compactness of the vectors in the cluster. The higher the concentration, the closer the vectors in the cluster are to the reference node and the tighter the cluster; conversely, the more dispersed the vectors in the cluster are.

[0092] The server can convert the average deviation value into a concentration value. For example, an inverse relationship can be used to achieve this conversion: concentration = 1 / (1 + average deviation). In this way, when the average deviation is 0 (that is, all vectors coincide with the reference node), the concentration is 1; and when the average deviation increases, the concentration will gradually decrease.

[0093] Finally, the concentration values ​​of Y feature vector clusters are obtained, and this information can be stored in the data warehouse for subsequent analysis and decision-making. For example, in a meteorological monitoring system, these concentration values ​​can help identify which areas have more dense and coordinated weather radar coverage, thereby optimizing the layout and scheduling strategy of radars.

[0094] In a possible implementation, determining the target weather radar set based on the concentration degree of the Y feature vector clusters includes:

[0095] Step S154: sorting the Y feature vector clusters in order based on the concentration levels of the Y feature vector clusters to generate a feature vector cluster list.

[0096] Step S155: determining a target weather radar set based on the first Z feature vector clusters in the feature vector cluster list, where Y≤Z≤1.

[0097] In this embodiment, after the server completes processing of X weather radar data records and generates Y feature vector clusters, it is necessary to determine the target weather radar set according to the concentration of these clusters. This process includes two main steps: sorting the feature vector clusters in order, and determining the target weather radar set based on the sorted cluster list.

[0098] The server first obtains the concentration values ​​of the Y feature vector clusters calculated previously. These values ​​reflect the closeness of the spatial correlation fusion vectors in each cluster. In order to compare and analyze these clusters more intuitively, the server decides to sort them according to the concentration value.

[0099] The server creates an empty feature vector cluster list and prepares to add Y clusters to the list in order of concentration from high to low (or from low to high, depending on actual needs).

[0100] Assume that Y = 3, that is, there are 3 feature vector clusters, and their concentration values ​​are 0.9, 0.7, and 0.5 respectively. The server first finds the cluster with the highest concentration (i.e., the cluster with a value of 0.9) and adds it to the top of the feature vector cluster list. Then, the server finds the cluster with the second highest concentration (i.e., the cluster with a value of 0.7) and adds it to the next position in the list. Finally, the server adds the remaining cluster (i.e., the cluster with a value of 0.5) to the bottom of the list.

[0101] After completing this step, the server obtains a list of feature vector clusters arranged in order of concentration, which provides the basis for the subsequent determination of the target weather radar set.

[0102] After having the feature vector cluster list arranged in order of concentration, the server needs to determine the target weather radar set based on this list. This process can adopt different strategies according to actual needs.

[0103] In a possible implementation, the server sets a parameter Z, where Y≤Z≤X. The parameter Z indicates that the server will select the first Z clusters from the feature vector cluster list to form the target weather radar set.

[0104] Continuing with the above example, assume that Y = 3, and the server sets Z = 2. Then the server will select the first two clusters (i.e., the two clusters with the highest concentration) from the feature vector cluster list to form the target weather radar set.

[0105] In a possible implementation, determining the target weather radar set based on the concentration degree of the Y feature vector clusters includes:

[0106] Step S156, obtaining the number of spatial correlation fusion vectors in each of the Y feature vector clusters.

[0107] Step S157: If the number of the spatial correlation fusion vectors in the feature vector cluster is greater than a threshold value and the concentration degree is greater than a set concentration degree, the feature vector cluster is a target weather radar set.

[0108] In another possible implementation, the server not only considers the concentration of feature vector clusters, but also considers the number of spatial correlation fusion vectors in each cluster. The server sets a threshold value, and only when the number of vectors in a cluster exceeds the threshold value and its concentration is greater than the set concentration, the cluster will be selected as part of the target weather radar set.

[0109] For example, the server sets the threshold value to 5 and the concentration to 0.8. Then for each feature vector cluster, the server will check whether the number of its vectors exceeds 5 and whether its concentration is greater than 0.8. Only clusters that meet both conditions will be selected as part of the target weather radar set.

[0110] In this way, it is possible to more flexibly determine the target weather radar set, taking into account both the compactness of the cluster and the number of vectors in the cluster. This helps to better meet the needs of the meteorological monitoring system in practical applications.

[0111] Therefore, the target weather radar set is determined based on the order of the concentration of the characteristic vector clusters. This process takes into account both the compactness of the clusters and the number of vectors in the clusters, providing a flexible and effective method for determining the target weather radar set for the meteorological monitoring system.

[0112] In practical applications, parameters and strategies can be adjusted according to the specific needs and real-time data of the meteorological monitoring system to ensure the accuracy and effectiveness of the target weather radar set. For example, when facing extreme weather events, the server may increase the threshold or adjust the concentration setting to ensure that more weather radars are included in the target weather radar set, thereby improving the accuracy and reliability of meteorological monitoring.

[0113] In a possible implementation, the method further includes:

[0114] Step A110, obtaining the first weather radar data record and first geographic location information of the sample weather radar, Q second weather radar data records and Q second geographic location information corresponding to the Q linked weather radars of the sample weather radar, and the non-linked weather radar of the sample weather radar, wherein the weather radar data record reflects the radar echo data generated by the detection coverage performance statistics and the target radar detection activity of the weather radar in the target test time domain interval, the first geographic location information includes regional associated geographic location information and site associated geographic location information, the regional associated geographic location information reflects the association relationship between the meteorological monitoring area where the sample weather radar is located and other areas, and the site associated geographic location information reflects the connection between the specific installation location of the sample weather radar and other radar sites. The regional associated geographic location information covers the user-related area, and the user-related area refers to the area with the same meteorological monitoring task or user as the area where the sample weather radar is located, and the site associated geographic location information includes other radar sites installed in the same device or the same device network as the sample weather radar, Q is a positive integer not less than 1, and the Q second geographic location information includes linkage regional associated geographic location information and site associated geographic location information.

[0115] In this embodiment, the sample weather radar in the meteorological monitoring system connected to the server is assumed to be radar A. In the target test time domain interval (for example, the past month), the server obtains the first weather radar data record of radar A, where the coverage efficiency statistics show that the effective detection range is 180 kilometers, the detection accuracy for rainstorm weather is 80%, and the data update frequency is once every 4 minutes. The radar echo sequence data reflects the radar echo generated by each detection activity within this month. For example, an echo with gradually increasing intensity is detected in the afternoon of a certain day, indicating that a rainstorm may be coming.

[0116] In the first geographic location information of radar A, the region-related geographic location information shows that the meteorological monitoring area where radar A is located is closely related to the meteorological areas of neighboring cities B and C, especially in the monsoon season, and is affected by similar meteorological conditions. The site-related geographic location information shows that radar A is installed at the mountaintop monitoring site in city A and is connected to the other two radar sites in city A through a high-speed communication network.

[0117] At the same time, Radar A has three linked weather radars, namely Radar B, Radar C and Radar D. The server obtains three second-level weather radar data records corresponding to these three linked weather radars.

[0118] For example, the coverage effectiveness statistics of radar B are: effective detection range of 150 kilometers, detection accuracy of thunderstorm weather is 85%, and data update frequency is once every 3 minutes; the radar echo sequence data records the echo characteristics of multiple thunderstorm weather.

[0119] The coverage efficiency statistics of Radar C are as follows: effective detection range of 200 kilometers, detection accuracy of 75% in foggy weather, data update frequency every 5 minutes; radar echo sequence data includes echo changes in foggy weather.

[0120] The coverage efficiency statistics of Radar D are as follows: effective detection range is 160 kilometers, the accuracy of typhoon weather warning is 90%, and the data is updated every 4 minutes; the radar echo sequence data records the echo conditions when the typhoon passes.

[0121] In the second geographic location information of these three linked weather radars, Radar B is located in the suburbs of City B. Its linked regional associated geographic location information shows that it interacts with the edge areas of City A and City C, and the site associated geographic location information shows that it is connected to the other two radar sites in City B. Radar C is located at a monitoring point in the mountainous area. Its linked regional associated geographic location information reflects the association with the meteorological areas of multiple surrounding mountainous cities, and the site associated geographic location information indicates that it is a core site in the mountainous monitoring network. Radar D is located at the port of a coastal city. Its linked regional associated geographic location information reflects the close connection with the marine meteorological area, and the site associated geographic location information shows that it works in coordination with two small radar sites near the port.

[0122] In addition, the server also obtains weather radars that are not linked to Radar A, such as Radar E, whose relevant data and geographic location information are not temporarily involved in this processing process.

[0123] Step A120, establishing geographic correlation information between the sample weather radar and the non-linked weather radar, and generating constructed geographic location information of the sample weather radar.

[0124] =In this embodiment, the server starts to establish geographic association information between radar A and non-linked weather radar (such as radar E). First, the server analyzes the longitude and latitude coordinates of radar A and radar E and calculates the straight-line distance between them. Assuming that the coordinates of radar A are (longitude A, latitude A) and the coordinates of radar E are (longitude E, latitude E), through the geographic coordinate calculation method, the distance between the two is 300 kilometers.

[0125] Then, the server considers terrain factors. Assuming there is a mountain range between the two, it will have a certain degree of blocking and attenuation effect on the propagation of radar signals. The server obtains information such as the height and direction of the mountain range based on the terrain database to evaluate the specific impact of this blocking on signal propagation.

[0126] Next, the server analyzes the similarity of meteorological conditions in the two regions. For example, the annual average wind speed, wind direction, humidity and other meteorological parameters in their regions. Suppose it is found that the wind direction and wind speed in their regions during the monsoon season are relatively similar.

[0127] Based on the above information, the server generates the structural geographic location information of radar A. The structural geographic location information may include descriptions of the distance from radar E, terrain impact assessment value, similarity of meteorological conditions, etc., to fully reflect the geographical correlation characteristics between radar A and non-linked radar E.

[0128] Step A130, input the first weather radar data record into a deep learning extraction model to generate a sample radar coverage characterization vector, and input the Q second weather radar data records into a deep learning extraction model to generate Q linked radar coverage characterization vectors.

[0129] In this embodiment, the server inputs the first weather radar data record of radar A into a pre-trained deep learning extraction model.

[0130] For coverage performance statistics, the performance feature encoder in the deep learning extraction model first processes them. For example, the mapping unit in the performance feature encoder performs vector-level summation operations on values ​​such as effective detection range, detection accuracy, and data update frequency to generate a basic embedding representation. At the same time, the interaction unit performs vector product operations on these values ​​and sums the product results to generate an advanced embedding representation. Then, the basic embedding representation and the advanced embedding representation are integrated to obtain the performance feature encoding result.

[0131] For radar echo sequence data, the waveform conversion unit in the echo feature encoder converts it into multiple echo feature distributions, which are then integrated to generate echo integration results. The screening unit then screens and processes the echo integration results to generate echo feature encoding results.

[0132] Finally, the fusion control unit integrates, fuses and maps the performance feature coding results and the echo feature coding results to generate a sample radar coverage characterization vector of radar A.

[0133] At the same time, the server also inputs the second weather radar data records of radar B, radar C and radar D into the deep learning extraction model. After a similar processing process, the linked radar coverage characterization vectors of radar B, radar C and radar D are generated respectively.

[0134] Step A140: input the sample radar coverage characterization vector and the first geographic location information into a spatial correlation fusion model to generate a first sample spatial correlation fusion vector.

[0135] In this embodiment, the server inputs the sample radar coverage characterization vector of radar A and its first geographic location information into the spatial association fusion model.

[0136] The regional correlation focusing network in the spatial correlation fusion model receives the regional correlation geographic location information of radar A and the sample radar coverage characterization vector. By analyzing the regional correlation information and interacting with the sample radar coverage characterization vector, a regional correlation focusing vector is generated. This vector reflects the feature fusion of radar A in terms of regional correlation.

[0137] The site-associated focusing network receives the site-associated geographic location information of radar A and the sample radar coverage characterization vector, generates a site-associated focusing vector, and highlights the feature fusion of radar A at the site-associated level.

[0138] The geographic information fusion network fuses the region-related focused vector and the site-related focused vector to generate a high-order site-related fusion vector that fuses the region information and a high-order region-related fusion vector that fuses the site information.

[0139] Finally, the heterogeneous fusion network further processes the above fusion results to generate the first sample spatial correlation fusion vector, which fully reflects the comprehensive spatial correlation fusion characteristics of the sample radar coverage representation of radar A and the first geographic location information.

[0140] Step A150: input the sample radar coverage characterization vector and the structure geographic location information into a spatial correlation fusion model to generate a second sample spatial correlation fusion vector.

[0141] In this embodiment, the server inputs the sample radar coverage characterization vector of radar A and the constructed geographic location information generated in step A120 into the spatial association fusion model.

[0142] Similarly, the region-related focusing network generates a new region-related focusing vector based on the region-related part of the constructed geographic location information and the sample radar coverage characterization vector. The site-related focusing network generates a site-related focusing vector based on the site-related part of the constructed geographic location information and the sample radar coverage characterization vector.

[0143] The geographic information fusion network fuses the newly generated regional association focus vector and site association focus vector to obtain a high-order site association fusion vector of fused regional information and a high-order regional association fusion vector of fused site information.

[0144] The heterogeneous fusion network processes these fusion results and generates a second sample spatial correlation fusion vector, which reflects the fusion characteristics of the sample radar coverage representation and the constructed geographic location information of radar A.

[0145] Step A160: input the Q linked radar coverage characterization vectors and the Q second geographic location information into a spatial correlation fusion model to generate Q third sample spatial correlation fusion vectors.

[0146] The server inputs the linked radar coverage characterization vectors of radar B, radar C, and radar D and their respective second geographic location information into the spatial correlation fusion model.

[0147] For radar B, the linkage area association focusing network receives its linkage area association geographic location information and linkage radar coverage characterization vector, and generates the linkage area association focusing vector of radar B. The site association focusing network generates the site association focusing vector based on the site association geographic location information of radar B and the linkage radar coverage characterization vector.

[0148] The geographic information fusion network fuses the linkage area associated focusing vector and the site associated focusing vector of radar B to generate a high-order site associated fusion vector of fused regional information and a high-order linkage area associated vector of fused site information.

[0149] The heterogeneous fusion network processes the fusion results and obtains the third sample spatial correlation fusion vector of radar B.

[0150] The same process is applied to radar C and radar D to generate their third sample spatial correlation fusion vectors respectively.

[0151] Step A170: Generate a linkage radar spatial correlation fusion vector based on the average vector of the Q third sample spatial correlation fusion vectors.

[0152] After the server obtains the third sample spatial correlation fusion vectors of radar B, radar C, and radar D, it calculates their average vector.

[0153] First, the server adds the corresponding components of each vector and then divides it by the number of vectors (3) to obtain the average vector. Assuming that the third sample spatial correlation fusion vector of radar B is (B1, B2, B3), that of radar C is (C1, C2, C3), and that of radar D is (D1, D2, D3), the calculation of the average vector is:

[0154] Average vector = ((B1+C1+D1) / 3, (B2+C2+D2) / 3, (B3+C3+D3) / 3)

[0155] Through such calculations, the server generates a spatial correlation fusion vector of the linked radars, which comprehensively reflects the spatial correlation fusion characteristics of the three linked radars.

[0156] Step A180: Generate a global training error based on the first sample spatial correlation fusion vector, the second sample spatial correlation fusion vector and the linked radar spatial correlation fusion vector.

[0157] The server generates a global training error based on the first sample spatial correlation fusion vector, the second sample spatial correlation fusion vector, and the linkage radar spatial correlation fusion vector.

[0158] First, the server constructs a benchmark training error based on the first example spatial correlation fusion vector and the linkage radar spatial correlation fusion vector. The difference between the two is calculated, for example, using methods such as Euclidean distance and cosine similarity to measure the degree of difference, thereby obtaining the benchmark training error.

[0159] Then, a comparative training error is constructed based on the second sample spatial correlation fusion vector and the linked radar spatial correlation fusion vector, and a similar difference calculation method is also used.

[0160] Finally, the benchmark training error and the comparative training error are combined, for example, by weighted average or other comprehensive calculation methods to generate a global training error. The global training error reflects the degree of deviation in the spatial correlation fusion between the sample weather radar and the linkage radar under different geographic location information conditions.

[0161] Step A190, using the global training error to train the deep learning extraction model and the spatial association fusion model.

[0162] The server uses the generated global training error to train the deep learning extraction model and the spatial association fusion model.

[0163] When training the deep learning extraction model, the server adjusts the parameters of the efficiency feature encoder, echo feature encoder, and fusion control unit based on the global training error back propagation, so that the model can generate more accurate coverage characterization vectors when processing new weather radar data.

[0164] For the spatial association fusion model, the server adjusts the parameters of the regional association focusing network, site association focusing network, geographic information fusion network and heterogeneous fusion network according to the global training error to improve the model's ability to fuse information from different geographic locations and generate more accurate spatial association fusion vectors.

[0165] Through continuous iterative training, the deep learning extraction model and spatial correlation fusion model can better process weather radar data and geographic location information, and improve the accuracy and reliability of meteorological monitoring and forecasting.

[0166] In a possible implementation, step A180 includes:

[0167] Step A181, constructing a benchmark training error based on the first sample spatial correlation fusion vector and the linkage radar spatial correlation fusion vector.

[0168] In this embodiment, the server first obtains the first example spatial correlation fusion vector and the linkage radar spatial correlation fusion vector. Assume that the first example spatial correlation fusion vector is (1.2, 0.8, 0.5, 0.7, 0.9) and the linkage radar spatial correlation fusion vector is (1.0, 0.6, 0.4, 0.6, 0.8).

[0169] The server may use the Euclidean distance to calculate the benchmark training error. Each component of the first sample spatial correlation fusion vector and the linkage radar spatial correlation fusion vector may be substituted into the Euclidean distance formula, and the benchmark training error constructed based on the first sample spatial correlation fusion vector and the linkage radar spatial correlation fusion vector is approximately 0.33.

[0170] Step A182, constructing a comparative training error based on the second sample spatial correlation fusion vector and the linkage radar spatial correlation fusion vector.

[0171] Step A183, generating a global training error based on the benchmark training error and the comparison training error.

[0172] Assume that the spatial correlation fusion vector of the second sample is (1.1, 0.7, 0.4, 0.6, 0.8).

[0173] The Euclidean distance is also used to calculate the contrast training error, so the constructed contrast training error is about 0.14.

[0174] The server sets the weight of the baseline training error to 0.6 and the weight of the comparison training error to 0.4.

[0175] Global training error = 0.6 × 0.33 + 0.4 × 0.14 = 0.198 + 0.056 = 0.254

[0176] In this way, the server generates a global training error through the above calculation, which comprehensively considers the differences between the first sample spatial correlation fusion vector and the second sample spatial correlation fusion vector and the linkage radar spatial correlation fusion vector respectively.

[0177] In actual meteorological monitoring scenarios, by continuously repeating the above process, the generated global training error is used to optimize and improve the deep learning extraction model and the spatial correlation fusion model, so as to improve the accuracy and reliability of the model when processing complex weather radar data and geographic location information, thereby providing more accurate support for weather forecasts and disaster warnings.

[0178] In a possible implementation, the spatial correlation fusion model includes a regional correlation focus network, a site correlation focus network, a geographic information fusion network, and a heterogeneous fusion network.

[0179] Step A140 includes:

[0180] Step A141: input the region-associated geographic location information and the sample radar coverage characterization vector into the region-associated focusing network to generate a region-associated focusing vector.

[0181] Step A142: input the site-associated geographic location information and the sample radar coverage characterization vector into the site-associated focusing network to generate a site-associated focusing vector.

[0182] Step A143, input the region-associated focusing vector and the site-associated focusing vector into the geographic information fusion network, generate a high-order site-associated fusion vector for fused region information and a high-order region-associated fusion vector for fused site information, wherein the high-order site-associated fusion vector for fused region information is obtained based on the site-associated focusing vector and the site-associated array for fused region-associated focusing vector, and the high-order region-associated fusion vector for fused site information is obtained based on the region-associated focusing vector and the sample radar association array for fused site-associated focusing vector.

[0183] Step A144: input the high-order site-associated fusion vector of the fused region information and the high-order region-associated fusion vector of the fused site information into the heterogeneous fusion network to generate a first example spatial-associated fusion vector.

[0184] In this embodiment, in the meteorological monitoring system, the server undertakes the important task of processing and analyzing a large amount of weather radar data. Taking radar A as an example, the server starts to process its related data.

[0185] First, the server obtains the sample radar coverage characterization vector and detailed first geographic location information of radar A. In the first geographic location information, the regional associated geographic location information shows that the meteorological monitoring area where radar A is located is closely related to the meteorological areas of neighboring cities B and C during the monsoon season, and is affected by similar meteorological conditions; the site associated geographic location information shows that radar A is installed at the mountaintop monitoring site in city A, and is connected to the other two radar sites in city A through a high-speed communication network.

[0186] The server inputs the regional correlation geographic location information and sample radar coverage characterization vectors into the regional correlation focusing network. After receiving these data, the regional correlation focusing network begins to conduct in-depth analysis and processing.

[0187] It is assumed that the regional associated geographic location information contains the specific correlation data with the meteorological regions of city B and city C during the monsoon season, such as the similarity of monsoon wind direction, the similarity ratio of monsoon rainfall, etc. At the same time, the sample radar coverage characterization vector contains the detection capability and accuracy of radar A for different meteorological phenomena.

[0188] The regional correlation focusing network performs feature extraction and correlation analysis on these input data. For example, it extracts the correlation features between the similarity of monsoon wind direction and the relevant features of wind direction detection in the sample radar coverage characterization vector. For the similarity of rainfall brought by the monsoon, the network performs correlation analysis with the accuracy features of rainfall detection in the sample radar coverage characterization vector.

[0189] After a series of complex calculations and processing, the regional correlation focusing network generates a regional correlation focusing vector, which highlights the key features of radar A in regional correlation, such as its key correlation features when coping with monsoon weather with cities B and C.

[0190] Next, the server inputs the site-associated geographic location information and the sample radar coverage characterization vector into the site-associated focusing network.

[0191] The site-associated geographic location information includes specific data such as the distance between radar A and the other two radar sites in city A, the bandwidth and delay of the communication network, etc. The sample radar coverage characterization vector continues to provide relevant features such as the detection performance of radar A.

[0192] The site association focusing network will conduct a comprehensive analysis of this data. For example, it will consider the impact of the bandwidth and latency of the communication network on the data transmission and sharing of radar A, and combine the features of the sample radar coverage characterization vector such as data update frequency and detection accuracy to extract the key features of the association between sites.

[0193] After processing, the site-association focusing network generates a site-association focusing vector, which clearly reflects the important characteristics of radar A at the site-association level, such as its efficiency and advantages when working in collaboration with other sites.

[0194] Then, the server inputs the generated region-related focusing vector and site-related focusing vector into the geographic information fusion network.

[0195] The geographic information fusion network first processes the site-related focusing vector and the site-related array of the regional-related focusing vector. The site-related array may contain the association data between radar A and other sites under different conditions (such as different weather conditions and different times).

[0196] The network will fuse the regional association focus vector with the site association focus vector, and combine the data in the site association array to generate a high-order site association fusion vector that fuses regional information. This vector not only contains the characteristics of radar A in terms of regional association, but also incorporates detailed information on its association with the site, as well as the history and change trend of the association between the sites.

[0197] At the same time, for the generation of high-order regional correlation fusion vectors of fused station information, the geographic information fusion network will fuse the regional correlation focus vector of the station correlation focus vector with the sample radar correlation array. The sample radar correlation array may contain the historical detection data and correlation features of radar A in different regions and under different meteorological conditions.

[0198] By fusing these data, the high-order regional correlation fusion vector generated by the fused site information fully reflects the comprehensive characteristics of radar A in site correlation and regional correlation, as well as the complex relationship and changing law between them.

[0199] Finally, the server inputs the high-order site-associated fusion vector of the fused regional information and the high-order region-associated fusion vector of the fused site information into the heterogeneous fusion network.

[0200] The heterogeneous fusion network performs a more in-depth comprehensive processing on the two high-order fusion vectors. It analyzes the potential relationship between the elements in the two vectors and explores deeper fusion features.

[0201] For example, it may find that some elements in the high-order site-related fusion vector of fused regional information and the elements in the high-order regional-related fusion vector of fused site information have a pattern of coordinated changes under specific meteorological conditions.

[0202] After a complex calculation and fusion process, the heterogeneous fusion network finally generates the first sample spatial correlation fusion vector. This vector comprehensively and accurately reflects the comprehensive spatial correlation fusion characteristics of the sample radar coverage representation of radar A and the first geographic location information, providing important basic data for subsequent meteorological analysis, prediction and decision-making.

[0203] In actual meteorological monitoring and forecasting work, the server analyzes and integrates a large amount of weather radar data by repeating the above-mentioned processing process, continuously improving the accuracy and reliability of meteorological monitoring and forecasting, and providing better meteorological services and guarantees for people's production and life. For example, in the upcoming heavy rain weather warning, the server can more accurately predict the scope, intensity and movement path of the heavy rain through the spatial correlation fusion analysis of multiple radars, and issue warnings in advance, so that relevant areas can take preventive measures and reduce disaster losses.

[0204] Assume that in a specific meteorological event, the server finds that the meteorological correlation between radar A and city B is significantly enhanced in a specific time period through the processing of the above spatial correlation fusion model, and the efficiency of collaborative work with other stations in city A is also improved. Combined with the data and analysis of other radars, the server can more accurately determine the trend of this meteorological change and the scope of possible impact, and timely release accurate meteorological forecasts and warning information to relevant departments and the public, protect people's lives and property, and improve the society's ability to respond to meteorological disasters.

[0205] In long-term operation, the server will continuously optimize and adjust the parameters and algorithms of the spatial correlation fusion model based on new data and actual forecast results to adapt to changing meteorological conditions and monitoring needs, and provide stronger technical support for the development of the meteorological industry.

[0206] In a possible implementation, the spatial correlation fusion model includes a linkage region correlation focus network, a site correlation focus network, a geographic information fusion network, and a heterogeneous fusion network.

[0207] Step A160 includes:

[0208] Step A161, using Q linkage area associated geographic location information and Q linkage radar coverage characterization vectors as model import data of the linkage area associated focusing network, and using the linkage area associated focusing network to generate Q linkage area associated focusing vectors.

[0209] Step A162: inputting the Q site-associated geographic location information and the Q linkage radar coverage characterization vectors into the site-associated focusing network to generate Q site-associated focusing vectors.

[0210] Step A163, input the Q linkage area associated focusing vectors and the Q site associated focusing vectors into the geographic information fusion network, generate Q high-order site associated fusion vectors of fused area information and Q high-order linkage area associated vectors of fused site information, wherein the high-order site associated fusion vector of the fused area information is obtained based on the site associated focusing vector and the site associated array fused with the linkage area associated focusing vector, and the high-order linkage area associated vector of the fused site information is obtained based on the linkage area associated focusing vector and the linkage radar associated array fused with the site associated focusing vector.

[0211] Step A164: input the high-order site-associated fusion vectors of the Q fusion area information and the high-order linkage area-associated vectors of the Q fusion area information into the heterogeneous fusion network to generate Q third sample space-associated fusion vectors.

[0212] In this embodiment, in the meteorological monitoring system, the server is responsible for processing the data of multiple linked weather radars. Assume that there are three linked weather radars, namely radar B, radar C and radar D.

[0213] First, the server obtains the linkage radar coverage characterization vectors of the three radars and their respective second geographic location information. The second geographic location information includes linkage area-related geographic location information and station-related geographic location information.

[0214] For radar B, its linkage area associated geographic location information shows that it interacts with the edge areas of city A and city C, and the site associated geographic location information shows that it is connected to the other two radar sites in city B.

[0215] The server imports the linkage area associated geographic location information and linkage radar coverage characterization vector of radar B as model data of the linkage area associated focusing network. After receiving these data, the linkage area associated focusing network starts processing.

[0216] It is assumed that the geographical location information associated with the linkage area contains detailed data such as the interaction frequency and interaction intensity with the marginal areas of City A and City C under different meteorological conditions, and the linkage radar coverage characterization vector includes the detection capability and characteristics of radar B for various meteorological phenomena.

[0217] The linkage area correlation focus network will deeply analyze these data. For example, it will extract the key features and focus directions of radar B when interacting with the edge areas of city A and city C based on the interaction frequency and intensity data and the detection capability characteristics of radar B. After complex calculations and analysis, the linkage area correlation focus vector of radar B is generated.

[0218] Similarly, for radar C and radar D, the server performs the same processing to generate their respective linkage area associated focusing vectors.

[0219] Next, the server inputs the site-associated geographic location information of radars B, C, and D and the linkage radar coverage characterization vector into the site-associated focusing network.

[0220] Taking radar B as an example, the site-associated geographic location information includes data such as the distance between it and the other two radar sites in city B, the communication method, and the communication stability. The linked radar coverage characterization vector still provides information such as the detection performance of radar B.

[0221] The site-associated focusing network will comprehensively consider these factors, such as the communication mode and stability, combined with the update frequency and accuracy of radar B's detection data, to extract the important characteristics of radar B in terms of site association and generate the site-associated focusing vector of radar B.

[0222] For radar C and radar D, their site-associated focusing vectors are also generated in a similar manner.

[0223] Then, the server inputs the generated three linkage area associated focus vectors and site associated focus vectors into the geographic information fusion network.

[0224] For radar B, the geographic information fusion network first processes its linkage area association focus vector and site association focus vector. It is assumed that the site association array contains the collaborative data between radar B and other sites under different weather conditions.

[0225] The network will fuse the linkage area association focus vector with the site association focus vector, and combine the data in the site association array to generate a high-order site association fusion vector that integrates regional information. This vector not only reflects the characteristics of radar B in terms of regional association, but also incorporates detailed information on its association with the site, as well as the history and changing trends of its collaboration.

[0226] At the same time, for the generation of high-order linkage area correlation vectors that fuse station information, the geographic information fusion network will fuse the station correlation focus vector with the linkage area correlation focus vector and the linkage radar correlation array. The linkage radar correlation array may contain the relevant data of radar B when working with other linkage radars.

[0227] By fusing these data, a high-order linkage regional correlation vector of the fused site information of radar B is generated.

[0228] For radar C and radar D, their respective high-order site-related fusion vectors of fused regional information and high-order linkage regional correlation vectors of fused site information are generated in the same manner.

[0229] Finally, the server inputs the high-order site-associated fusion vectors of the fusion area information of the three radars and the high-order linkage area-associated vectors of the fusion site information into the heterogeneous fusion network.

[0230] The heterogeneous fusion network will perform more in-depth comprehensive processing on these vectors. For example, it will analyze the potential relationship between certain elements in the high-order site-related fusion vectors of the fusion regional information of different radars, as well as their coordinated change rules with the elements in the high-order linkage regional correlation vectors of the fusion site information.

[0231] After complex calculation and fusion, the third sample space correlation fusion vectors of radars B, C, and D are generated.

[0232] In actual meteorological monitoring and forecasting, these generated third-sample spatial correlation fusion vectors provide important data support for more accurate analysis of meteorological conditions and prediction of weather changes. For example, in the case of an approaching typhoon, by analyzing these vectors, the server can more clearly understand the synergistic effect of different linked radars in monitoring the typhoon's path and intensity, thereby more accurately issuing typhoon warning information, preparing response measures in advance, and reducing the possible losses caused by the typhoon.

[0233] During the long-term operation process, the server will continuously optimize the parameters and algorithms of the spatial correlation fusion model based on new data and actual forecast results, improve the accuracy and reliability of the model, and provide better services for meteorological monitoring and forecasting.

[0234] In a possible implementation, the X weather radar data records include X coverage effectiveness statistical data and X radar echo sequence data, and the deep learning extraction model includes an effectiveness feature encoder, an echo feature encoder, and a fusion control unit.

[0235] Step S120 includes:

[0236] Step S121: input the X coverage performance statistical data into the performance feature encoder to generate X performance feature encoding results.

[0237] Step S122: input the X radar echo sequence data into the echo feature encoder to generate X echo feature encoding results.

[0238] Step S123: input the X performance feature encoding results and the X echo feature encoding results into the fusion control unit to generate X target radar coverage characterization vectors.

[0239] In this embodiment, in the meteorological monitoring system, the server needs to process data from multiple weather radars. Assume that there are 5 weather radars, that is, X=5, which are radar 1 to radar 5.

[0240] First, the server obtains the weather radar data records of the five weather radars. The data record of each weather radar includes coverage effectiveness statistics and radar echo sequence data.

[0241] For the coverage effectiveness statistics of radar 1, for example, the effective detection range is 180 kilometers, the detection accuracy for rainstorm weather is 80%, and the data update frequency is once every 5 minutes. For radar 2, its coverage effectiveness statistics may be an effective detection range of 150 kilometers, a detection accuracy for thunderstorm weather of 85%, and a data update frequency of once every 3 minutes. Similarly, the server obtains the coverage effectiveness statistics of radars 3, 4, and 5.

[0242] Next, the server inputs these five coverage performance statistics into the performance feature encoder.

[0243] The performance feature encoder starts working. For the coverage performance statistics of radar 1, firstly, a performance feature segment is generated based on its feature distribution and specific values. It is assumed that this segment contains information such as the relative importance weight of the effective detection range, the weight of the detection accuracy, and the weight of the data update frequency.

[0244] Then, this performance feature fragment is input into the mapping unit. The mapping unit performs a vector-level summation operation on it. For example, the values ​​of the effective detection range, detection accuracy, and data update frequency are multiplied by the corresponding weights and then summed to generate a basic embedding representation result.

[0245] At the same time, this performance feature fragment is input to the interaction unit. The interaction unit performs vector product operations and sums the product results. For example, the product of the effective detection range and the detection accuracy, the product of the effective detection range and the data update frequency, and the product of the detection accuracy and the data update frequency are first calculated, and then these product results are added to generate an advanced embedding representation result.

[0246] Finally, the basic embedding representation results and the advanced embedding representation results are integrated to generate the performance feature encoding results of radar 1.

[0247] For Radar 2 to Radar 5, the server processes their coverage effectiveness statistical data in the same way and generates respective effectiveness feature encoding results.

[0248] Next, the server processes the radar echo sequence data.

[0249] The radar echo sequence data of radar 1 includes the echo conditions generated by each detection activity in the past period of time. The server inputs these data into the waveform conversion unit of the echo feature encoder.

[0250] The waveform conversion unit converts these radar echo sequence data into multiple echo feature distributions. For example, different feature distributions are generated according to factors such as the intensity, frequency, and duration of the echo.

[0251] Then, these echo feature distributions are integrated to generate the echo integration result of radar 1.

[0252] Afterwards, the echo integration result is input to the screening unit, which removes some noise and unimportant features according to preset screening rules and generates the echo feature coding result of the radar 1.

[0253] The server also processes the radar echo sequence data of radar 2 to radar 5 according to the same steps to generate their respective echo feature coding results.

[0254] At this point, the server has obtained 5 performance feature encoding results and 5 echo feature encoding results.

[0255] The server inputs the five sets of performance feature encoding results and echo feature encoding results into the fusion control unit.

[0256] The fusion unit in the fusion control unit first integrates the performance feature coding result and the echo feature coding result of each group, for example, by adding or weighted summing the corresponding elements of the performance feature coding result and the echo feature coding result.

[0257] Then, the integrated result is input into the fusion unit to generate a fused encoding vector. The fusion unit may use some complex fusion algorithms, such as attention mechanism or multi-layer perceptron for deeper fusion processing.

[0258] Finally, the fused coding vector is input to the mapping output unit, which maps and adjusts the fused coding vector to generate the final target radar coverage representation vector.

[0259] The server processes the encoding results of radar 2 to radar 5 in the same way, and finally generates 5 target radar coverage characterization vectors.

[0260] In actual meteorological monitoring scenarios, the server can comprehensively consider the coverage effectiveness statistics of weather radar and radar echo sequence data through such a processing flow to generate characterization vectors that accurately reflect the radar coverage characteristics. These characterization vectors can provide an important basis for subsequent meteorological analysis, prediction and decision-making.

[0261] Assume that in a severe convective weather monitoring, the server processes the data of multiple weather radars through the above process. The coverage efficiency statistics of the area where Radar 1 is located show that its detection accuracy of severe convective weather is high, but the radar echo sequence data shows some abnormal echo fluctuations. After being processed by the deep learning extraction model, the generated target radar coverage representation vector can accurately reflect this complex situation, providing strong support for meteorological experts to judge the weather change trend in the area.

[0262] For example, in another typhoon monitoring, the coverage effectiveness statistics of Radar 3 showed that it had advantages in long-distance detection, but the radar echo sequence data showed that the echo characteristics at the edge of the typhoon were more complex. Through the fine processing of the deep learning extraction model, the generated target radar coverage characterization vector can highlight the unique contribution and potential problems of Radar 3 in typhoon monitoring, which helps to improve the accuracy of typhoon path and intensity prediction.

[0263] With the continuous accumulation of meteorological data and the continuous optimization of models, the processing power and accuracy of the server will continue to improve, providing a more solid guarantee for the quality and reliability of meteorological services.

[0264] For example, in one possible implementation, the performance feature encoder includes a mapping unit and an interaction unit.

[0265] Step S121 includes:

[0266] Step S1211 : generating X performance feature segments based on the feature distribution of the X coverage performance statistical data and the X coverage performance statistical data.

[0267] Step S1212: input the X performance feature segments into the mapping unit to generate X basic embedding representation results, wherein the mapping unit is used to perform a vector-level summation operation on each of the X performance feature segments.

[0268] Step S1213: input the X performance feature segments into the interaction unit to generate X advanced embedding representation results, wherein the interaction unit is used to perform a vector product operation on each of the X performance feature segments and a summation operation on the vector product operation results.

[0269] Step S1214: Integrate the X basic embedding representation results with the X advanced embedding representation results to generate X performance feature encoding results.

[0270] In the meteorological monitoring system, the server is responsible for processing the coverage effectiveness statistics of multiple weather radars. Assume there are 5 weather radars, namely radar 1 to radar 5.

[0271] First, the server obtains the coverage efficiency statistics of the five radars. For example, the coverage efficiency statistics of radar 1 are: effective detection range of 200 kilometers, detection accuracy of precipitation type judgment is 85%, and data update frequency is once every 6 minutes; the data of radar 2 are: effective detection range of 160 kilometers, detection accuracy of thunderstorm weather judgment is 90%, and data update frequency is once every 4 minutes. The data of radar 3, radar 4 and radar 5 are obtained by analogy.

[0272] Based on the characteristic distribution and specific values ​​of these coverage effectiveness statistics, the server generates an effectiveness characteristic fragment for each radar.

[0273] For radar 1, the server analyzes the characteristics and relationships of its coverage performance statistics. For example, the importance of the effective detection range in the overall performance, the weight of the detection accuracy, and the impact of the data update frequency. Based on these analyses, an performance feature fragment containing these feature relationships and weight information is generated.

[0274] Likewise, respective performance characteristic segments are generated for Radar 2 to Radar 5 .

[0275] Next, the server inputs the five performance feature segments into the mapping unit.

[0276] After receiving the performance feature fragment of radar 1, the mapping unit starts to perform vector-level summation operations. Assume that the effective detection range in the performance feature fragment is represented by v1, the detection accuracy is represented by v2, and the data update frequency is represented by v3, and the corresponding weights are w1, w2, and w3 respectively. Then, the mapping unit will calculate w1*v1+w2*v2+w3*v3 to generate the basic embedding representation result of radar 1.

[0277] For the performance feature fragments of Radar 2 to Radar 5, the mapping unit also processes them in the same way to generate their respective basic embedding representation results.

[0278] Then, the server inputs the five performance feature fragments into the interaction unit.

[0279] After receiving the performance feature fragment of radar 1, the interaction unit performs a vector product operation and a summation operation of the vector product operation results. For example, v1*v2, v1*v3, v2*v3 are calculated, and then these product results are added to generate the advanced embedding representation result of radar 1.

[0280] The interactive unit performs the same processing on the performance feature fragments of Radar 2 to Radar 5 to generate their respective advanced embedding representation results.

[0281] At this point, the server has obtained 5 basic embedding representation results and 5 advanced embedding representation results.

[0282] The server integrates the basic embedding representation result and the advanced embedding representation result of radar 1. For example, they may be processed by weighted addition or using a specific fusion algorithm to finally generate the performance feature encoding result of radar 1.

[0283] For Radar 2 to Radar 5, the server is integrated in the same way to generate their respective performance feature coding results.

[0284] In actual meteorological monitoring, such as in a forecast scenario of an upcoming rainstorm, the server can extract richer and more accurate features from coverage effectiveness statistics through such a processing flow.

[0285] Assume that in this rainstorm forecast, the effective detection range of radar 4 is large, but the detection accuracy is relatively low, and the data update frequency is fast. Through the above processing, the generated performance feature coding results can comprehensively reflect these characteristics and provide more valuable information for subsequent meteorological analysis and forecasting.

[0286] For example, in a typhoon warning, the coverage effectiveness statistics of Radar 5 have unique characteristics. The effectiveness feature coding results generated by the server can accurately capture these characteristics, helping meteorological experts to more comprehensively understand the effectiveness of Radar 5 in typhoon monitoring, so as to make more accurate predictions of typhoon paths and intensities.

[0287] By continuously processing coverage efficiency statistical data under various meteorological scenarios, the server can continuously optimize the parameters and algorithms of the efficiency feature encoder, improve the accuracy and reliability of the generated efficiency feature encoding results, and provide stronger support for meteorological monitoring and forecasting.

[0288] For example, in one possible implementation, the echo feature encoder includes a waveform conversion unit and a screening unit.

[0289] Step S122 includes:

[0290] Step S1221: input the X radar echo sequence data into the waveform transformation unit to generate X×W echo feature distributions, wherein the echo feature distribution is composed of at least one radar echo result in the radar echo sequence data, and W is a positive integer greater than 1.

[0291] Step S1222: Integrate the X×W echo feature distributions to generate X echo integration results.

[0292] Step S1223: input the X echo integration results into a screening unit to generate X echo feature coding results.

[0293] In the meteorological monitoring system, the server is responsible for processing radar echo sequence data from multiple weather radars. Assume there are 5 weather radars, namely radar 1 to radar 5.

[0294] The server first obtains the radar echo sequence data of the five radars.

[0295] For radar 1, its radar echo sequence data contains a large amount of echo information within a period of time, such as echo intensity, frequency, phase, etc. at different time points. The server inputs these data into the waveform conversion unit.

[0296] After receiving the radar echo sequence data of radar 1, the waveform conversion unit processes the data according to the preset algorithms and rules. For example, according to the variation range of the echo intensity, it is divided into different intensity intervals, and each interval is used as a feature to generate multiple echo feature distributions. Assume that three echo feature distributions are generated, namely, intensity feature distribution, frequency feature distribution and phase feature distribution.

[0297] For radars 2 to 5, the server also inputs their radar echo sequence data into the waveform conversion unit in the same way, and each generates multiple echo feature distributions. Assume that radar 2 generates 4 echo feature distributions, radar 3 generates 2 echo feature distributions, radar 4 generates 3 echo feature distributions, and radar 5 generates 2 echo feature distributions. In this way, a total of 5×(2+3+4+3+2)=5×14=70 echo feature distributions are generated.

[0298] Next, the server needs to integrate these large number of echo feature distributions.

[0299] For the three echo feature distributions of radar 1, the server will analyze the correlation and commonality between them. For example, it is found that the intensity feature distribution and the frequency feature distribution have a certain correlation in certain time periods, and then these feature distributions are integrated and merged to generate an echo integration result that can fully reflect the echo characteristics of radar 1.

[0300] The server also integrates the echo feature distributions of radar 2 to radar 5 in the same manner and generates respective echo integration results.

[0301] Then, the server inputs the integration results of these five echoes into the screening unit.

[0302] After receiving the echo integration result of radar 1, the screening unit will process it according to the pre-set screening criteria. For example, it will remove some parts with large noise and small contribution to the overall characteristics, or extract the most representative features. After screening and processing, the echo feature encoding result of radar 1 is generated.

[0303] The screening unit also processes the echo integration results of radar 2 to radar 5 in the same way to generate their respective echo feature coding results.

[0304] In actual meteorological monitoring scenarios, such as monitoring severe convective weather.

[0305] The radar echo sequence data of Radar 3 shows complex changes, and the waveform transformation unit generates multiple echo feature distributions reflecting different aspects of the characteristics. In the integration stage, the server can combine these distributions to form a comprehensive echo integration result, highlighting the characteristics of the echo in severe convective weather.

[0306] During the screening stage, some abnormal features that may be caused by accidental factors were removed, so that the final echo feature coding results can more accurately reflect the essential characteristics of severe convective weather.

[0307] Another example is in the monitoring of a large-scale rainfall process.

[0308] The radar echo sequence data of radar 5 is relatively stable, but also contains some subtle changes. Through the waveform transformation unit and the integration process, the server can extract the key echo features. Under the processing of the screening unit, some unimportant details are removed, and the generated echo feature coding results can clearly reflect the overall situation of rainfall.

[0309] By continuously processing radar echo sequence data under various weather conditions, the server can optimize the parameters and algorithms of the echo feature encoder, improve the quality and accuracy of the generated echo feature encoding results, and provide more valuable information for meteorological analysis and forecasting.

[0310] For example, in a possible implementation, the fusion regulation unit includes a fusion unit and a mapping output unit.

[0311] Step S123 includes:

[0312] Step S1231: Integrate each of the X performance feature coding results with each of the X echo feature coding results to generate X integrated coding results.

[0313] Step S1232: input the X integrated coding results into the fusion unit to generate X fused coding vectors.

[0314] Step S1233: input the X fused coding vectors into the mapping output unit to generate X target radar coverage characterization vectors.

[0315] In this embodiment, in the meteorological monitoring system, it is assumed that the server processes data of 5 weather radars, that is, X=5.

[0316] The server has obtained 5 performance feature encoding results and 5 echo feature encoding results.

[0317] For the first weather radar, the server integrates its performance feature coding result with the echo feature coding result. For example, the performance feature coding result contains coding information about the radar detection range, accuracy, and update frequency, and the echo feature coding result contains coding information about the radar echo's intensity, frequency, and duration. The server uses a specific algorithm to add or weighted sum the corresponding elements in the two coding results to generate the first integrated coding result.

[0318] In the same manner, the server integrates the effectiveness characteristic coding results and echo characteristic coding results of the second to fifth weather radars to generate the second, third, fourth and fifth integrated coding results respectively.

[0319] Next, the server inputs the five integrated encoding results into the fusion unit.

[0320] After receiving the first integrated coding result, the fusion unit performs a deeper fusion process on it. It may use methods such as multi-layer neural networks or attention mechanisms to comprehensively consider the complex relationship between the various elements in the integrated coding result and generate the first fused coding vector.

[0321] The fusion unit also processes the second to fifth integrated coding results in the same manner to generate the second, third, fourth and fifth fused coding vectors respectively.

[0322] The server then inputs the five fused coding vectors into the mapping output unit.

[0323] After receiving the first fused coding vector, the mapping output unit maps it into a final target radar coverage representation vector through a series of linear transformations and nonlinear activation functions. This vector can fully and accurately reflect the coverage characteristics of the first weather radar.

[0324] In the same steps, the mapping output unit processes the second to fifth fused coding vectors to generate the second, third, fourth and fifth target radar coverage characterization vectors respectively.

[0325] In actual meteorological monitoring, for example, in a monitoring scenario where a typhoon is about to land.

[0326] For the first weather radar, the results of its performance feature coding show that it performs well in terms of detection range and accuracy, and the results of its echo feature coding show that its echo intensity and frequency changes are more obvious. Through the process of integration, fusion and mapping output, the generated target radar coverage characterization vector can clearly reflect the unique capabilities and characteristics of this radar in monitoring typhoons.

[0327] For the second weather radar, it may behave differently from the first in some aspects, but through the same processing flow, the server can also generate a target radar coverage characterization vector that accurately reflects its characteristics.

[0328] By continuously processing weather radar data under various meteorological conditions, the server continuously optimizes the parameters and algorithms of the fusion control unit, making the generated target radar coverage characterization vector more accurate and reliable, providing stronger support for weather forecasting and disaster warning.

[0329] Figure 2 The hardware structure of the deep learning system 100 provided in the embodiment of the present application for implementing the above-mentioned weather radar coverage effectiveness analysis method based on deep learning is shown, as shown in FIG. Figure 2 As shown, the deep learning system 100 may include a processor 110 , a machine-readable storage medium 120 , a bus 130 , and a communication unit 140 .

[0330] In an alternative embodiment, the deep learning system 100 may be a single server or a server group. The server group may be centralized or distributed (for example, the deep learning system 100 may be a distributed system). In an alternative embodiment, the deep learning system 100 may be local or remote. For example, the deep learning system 100 may access information and / or data stored in a machine-readable storage medium 120 via a network. For another example, the deep learning system 100 may be directly connected to the machine-readable storage medium 120 to access the stored information and / or data. In an alternative embodiment, the deep learning system 100 may be implemented on a cloud platform. By way of example only, the cloud platform may include a private cloud, a public cloud, an aggregated cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, or the like, or any integration thereof.

[0331] The machine-readable storage medium 120 may store data and / or instructions. In an alternative embodiment, the machine-readable storage medium 120 may store data obtained from an external terminal. In an alternative embodiment, the machine-readable storage medium 120 may store data and / or instructions used by the deep learning system 100 to execute or use to complete the exemplary method described in this application. In an alternative embodiment, the machine-readable storage medium 120 may include a mass storage, a removable storage, a volatile read-write memory, a read-only memory, etc. or any integration thereof. Exemplary mass storage may include a magnetic disk, an optical disk, a solid-state disk, etc. Exemplary removable storage may include a flash drive, a floppy disk, an optical disk, a memory card, a compressed disk, a magnetic tape, etc.

[0332] During the specific implementation process, multiple processors 110 execute computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the weather radar coverage effectiveness analysis method based on deep learning in the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processor 110 can be used to describe the sending and receiving actions of the communication unit 140.

[0333] The specific implementation process of the processor 110 can be found in the various method embodiments executed by the above-mentioned deep learning system 100. The implementation principles and technical effects are similar and will not be repeated in this embodiment.

[0334] In addition, an embodiment of the present application also provides a readable storage medium, in which computer executable instructions are pre-installed. When the processor executes the computer executable instructions, the above-mentioned weather radar coverage effectiveness analysis method based on deep learning is implemented.

[0335] Similarly, it should be noted that in order to simplify the description disclosed in this application, thereby helping to understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, multiple features are sometimes combined into one embodiment, drawings, or descriptions thereof. Similarly, it should be noted that in order to simplify the description disclosed in this application, thereby helping to understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, multiple features are sometimes combined into one embodiment, drawings, or descriptions thereof.

Claims

1. A weather radar coverage effectiveness analysis method based on deep learning, characterized in that: The method comprises: Acquire X weather radar data records and X geographic location information of X target weather radars, wherein each weather radar data record in the X weather radar data records includes coverage effectiveness statistical data and radar echo sequence data, the coverage effectiveness statistical data reflects data generated after performing detection coverage performance statistics on the target weather radar in a target test time domain interval, and the radar echo sequence data reflects radar echo data generated based on target radar detection activities of the target weather radar in the target test time domain interval, and X is a positive integer not less than 1; Inputting the X weather radar data records into a deep learning extraction model to generate X target radar coverage representation vectors; Inputting the X target radar coverage characterization vectors and the X geographic location information into a spatial correlation fusion model to generate X spatial correlation fusion vectors; Clustering the X spatial correlation fusion vectors based on a clustering model to generate Y feature vector clusters, where Y is a positive integer not less than 1; The concentration levels of the Y feature vector clusters are calculated, and a target weather radar set is determined based on the concentration levels of the Y feature vector clusters, wherein the target weather radar set includes at least one of the target weather radars.

2. The weather radar coverage effectiveness analysis method based on deep learning according to claim 1, characterized in that: The calculating the concentration degree of the Y feature vector clusters includes: Obtaining a spatial correlation fusion vector in each of the Y feature vector clusters; Calculating the average vector of the spatial correlation fusion vector in each of the feature vector clusters to generate a cluster reference node; The average deviation between the spatial correlation fusion vector in each of the Y feature vector clusters and the reference node of the cluster is calculated to generate the concentration degree of the Y feature vector clusters.

3. The weather radar coverage effectiveness analysis method based on deep learning according to claim 1, characterized in that: The step of determining the target weather radar set based on the concentration degree of the Y feature vector clusters includes: Based on the concentration levels of the Y feature vector clusters, the Y feature vector clusters are sorted in order to generate a feature vector cluster list; The target weather radar set is determined based on the first Z feature vector clusters in the feature vector cluster list, where Y≤Z≤1.

4. The weather radar coverage effectiveness analysis method based on deep learning according to claim 1, characterized in that: The step of determining the target weather radar set based on the concentration degree of the Y feature vector clusters includes: Obtaining the number of spatial correlation fusion vectors in each of the Y feature vector clusters; If the number of the spatial correlation fusion vectors in the feature vector cluster is greater than a threshold value and the concentration degree is greater than a set concentration degree, the feature vector cluster is a target weather radar set.

5. The weather radar coverage effectiveness analysis method based on deep learning according to claim 1, characterized in that: The method further comprises: Obtain a first weather radar data record and first geographic location information of a sample weather radar, Q second weather radar data records and Q second geographic location information corresponding to Q linked weather radars of the sample weather radar, and a non-linked weather radar of the sample weather radar, wherein the weather radar data record reflects the radar echo data generated by the detection coverage performance statistics and target radar detection activities of the weather radar within the target test time domain interval, the first geographic location information includes region-associated geographic location information and site-associated geographic location information, the region-associated geographic location information reflects the association between the meteorological monitoring area where the sample weather radar is located and other areas, and the site-associated geographic location information reflects the connection between the specific installation location of the sample weather radar and other radar sites; the region-associated geographic location information covers user-related areas, and the user-associated area refers to an area with the same meteorological monitoring task or user as the area where the sample weather radar is located, and the site-associated geographic location information includes other radar sites installed in the same device or the same device network as the sample weather radar, Q is a positive integer not less than 1, and the Q second geographic location information includes linkage region-associated geographic location information and site-associated geographic location information; Establishing geographic correlation information between the sample weather radar and the non-linked weather radar, and generating constructed geographic location information of the sample weather radar; Inputting the first weather radar data record into a deep learning extraction model to generate a sample radar coverage representation vector, and inputting the Q second weather radar data records into a deep learning extraction model to generate Q linked radar coverage representation vectors; Inputting the sample radar coverage characterization vector and the first geographic location information into a spatial correlation fusion model to generate a first sample spatial correlation fusion vector; Inputting the sample radar coverage characterization vector and the structure geographic location information into a spatial correlation fusion model to generate a second sample spatial correlation fusion vector; Inputting the Q linkage radar coverage characterization vectors and the Q second geographic location information into a spatial correlation fusion model to generate Q third sample spatial correlation fusion vectors; Generate a linkage radar spatial correlation fusion vector based on the average vector of the Q third sample spatial correlation fusion vectors; Generate a global training error based on the first sample spatial correlation fusion vector, the second sample spatial correlation fusion vector, and the linkage radar spatial correlation fusion vector; The deep learning extraction model and the spatial association fusion model are trained using the global training error.

6. The weather radar coverage effectiveness analysis method based on deep learning according to claim 5, characterized in that: The generating a global training error based on the first sample spatial correlation fusion vector, the second sample spatial correlation fusion vector and the linkage radar spatial correlation fusion vector comprises: Constructing a benchmark training error based on the first sample spatial correlation fusion vector and the linkage radar spatial correlation fusion vector; Constructing a comparative training error based on the second sample spatial correlation fusion vector and the linkage radar spatial correlation fusion vector; A global training error is generated based on the benchmark training error and the comparison training error.

7. The weather radar coverage effectiveness analysis method based on deep learning according to claim 5, characterized in that: The spatial correlation fusion model includes a regional correlation focus network, a site correlation focus network, a geographic information fusion network and a heterogeneous fusion network; The step of inputting the sample radar coverage characterization vector and the first geographic location information into a spatial correlation fusion model to generate a first sample spatial correlation fusion vector includes: Inputting the region-associated geographic location information and the sample radar coverage characterization vector into the region-associated focusing network to generate a region-associated focusing vector; Inputting the site-associated geographic location information and the sample radar coverage characterization vector into the site-associated focusing network to generate a site-associated focusing vector; Inputting the region-associated focusing vector and the site-associated focusing vector into the geographic information fusion network, generating a high-order site-associated fusion vector fused with the region information and a high-order region-associated fusion vector fused with the site information, wherein the high-order site-associated fusion vector fused with the region information is obtained based on the site-associated focusing vector fused with the region-associated focusing vector and the site-associated array, and the high-order region-associated fusion vector fused with the site information is obtained based on the region-associated focusing vector fused with the site-associated focusing vector and the sample radar association array; The high-order site-associated fusion vector of the fused region information and the high-order region-associated fusion vector of the fused site information are input into the heterogeneous fusion network to generate a first sample space-associated fusion vector.

8. The weather radar coverage effectiveness analysis method based on deep learning according to claim 5, characterized in that: The spatial correlation fusion model includes a linkage area correlation focus network, a site correlation focus network, a geographic information fusion network and a heterogeneous fusion network; The step of inputting the Q linkage radar coverage characterization vectors and the Q second geographic location information into a spatial correlation fusion model to generate Q third sample spatial correlation fusion vectors includes: Importing Q linkage area associated geographic location information and Q linkage radar coverage characterization vectors as model data of the linkage area associated focusing network, and generating Q linkage area associated focusing vectors using the linkage area associated focusing network; Inputting Q site-associated geographic location information and Q linkage radar coverage characterization vectors into the site-associated focusing network to generate Q site-associated focusing vectors; Input the Q linkage area associated focusing vectors and the Q site associated focusing vectors into the geographic information fusion network to generate Q high-order site associated fusion vectors of fused regional information and Q high-order linkage area associated vectors of fused site information, wherein the high-order site associated fusion vector of fused regional information is obtained based on the site associated focusing vector and the site associated array fused with the linkage area associated focusing vector, and the high-order linkage area associated vector of fused site information is obtained based on the linkage area associated focusing vector and the linkage radar associated array fused with the site associated focusing vector; The high-order site-associated fusion vectors of the Q fused area information and the high-order linkage area-associated vectors of the Q fused area information are input into the heterogeneous fusion network to generate Q third sample space-associated fusion vectors.

9. The weather radar coverage effectiveness analysis method based on deep learning according to claim 1, characterized in that: The X weather radar data records include X coverage efficiency statistical data and X radar echo sequence data, and the deep learning extraction model includes an efficiency feature encoder, an echo feature encoder and a fusion control unit; The step of inputting the X weather radar data records into a deep learning extraction model to generate X target radar coverage characterization vectors includes: Inputting the X coverage performance statistical data into the performance feature encoder to generate X performance feature encoding results; Inputting the X radar echo sequence data into the echo feature encoder to generate X echo feature encoding results; The X effectiveness feature encoding results and the X echo feature encoding results are input into the fusion control unit to generate X target radar coverage characterization vectors.

10. A deep learning system, characterized in that: The deep learning system includes a processor and a machine-readable storage medium, wherein the machine-readable storage medium stores machine-executable instructions, and the machine-executable instructions are loaded and executed by the processor to implement the deep learning-based weather radar coverage effectiveness analysis method described in any one of claim 19.

Images