5G-based early warning information targeted release method and platform

Abstract

The application discloses a kind of early warning information targeting release method and platform based on 5G, is specifically related to meteorological information targeting release field, including by obtaining multimodal meteorological sensor data, and dynamically allocating edge computing node and cloud server resource carries out data processing and risk assessment, to generate early warning release instruction set, execute early warning release instruction set and start system optimization process, to realize the targeting release of early warning information.A kind of early warning information targeting release method and platform based on 5G by constructing dynamic resource allocation model, and setting the task redistribution mechanism when network slice bandwidth fluctuation, realize the dynamic cooperation of edge and cloud resource, reduce the fragmentation of computing resources;Through edge computing node and cloud server carry out knowledge distillation shared feature weight, and by distillation frequency dynamic adjustment, realize the combination of local risk assessment and global correlation analysis, alleviate the end-to-end reliability decline caused by multiple subsystem call.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of meteorological information targeted release, more particularly, the present application relates to a warning information targeted release method and platform based on 5G. BACKGROUND

[0002] With the growth of meteorological disaster prevention and mitigation demand, real-time accurate release and personalized access of early warning information are particularly urgent. The traditional meteorological warning technology architecture mainly consists of distributed meteorological sensors, communication networks, centralized data processing servers and unified early warning release systems. Through centralized cloud computing processing of meteorological data, regional early warning information is generated by relying on fixed threshold models.

[0003] However, the traditional meteorological warning technology is limited by network bandwidth and time delay, resulting in problems such as delayed response to early warning and low efficiency of emergency response. The existing technology proposes an improved scheme combining distributed edge computing and dynamic sharding transmission. Computing nodes are deployed near the network edge of the sensor, data local preprocessing is used to reduce cloud load, and a compressed neural network is used to realize fast meteorological prediction. This shortens the data collection to early warning generation link delay, and can achieve kilometer-level early warning accuracy, and changes from a centralized single processing mode to a distributed collaborative processing mode, to a certain extent, improving the efficiency and pertinence of meteorological warning.

[0004] However, in actual use, there are still some shortcomings, such as lack of dynamic collaboration mechanism for edge-cloud resources, frequent migration of multi-modal data processing tasks between heterogeneous nodes, causing network load fluctuations and computing resource fragmentation; the lightweight model configured by the computing node is limited by the parameter size, and it is difficult to meet the needs of high-precision risk assessment and personalized content generation at the same time, and the user query response needs to call different subsystems multiple times, resulting in a decrease in end-to-end response reliability; the early warning push strategy relies on a static rule base and cannot be optimized autonomously according to real-time weather evolution and user feedback, and the early warning coverage and accuracy are negatively correlated in extreme scenarios, making it difficult to achieve effective transition from traditional regional broadcast early warning to fine and individualized targeted early warning. SUMMARY

[0005] In order to overcome the above-mentioned defects of the prior art, the present application provides a warning information targeted release method and platform based on 5G, which solves the problems in the background art by the following scheme.

[0006] In order to achieve the above-mentioned purpose, the present application provides the following technical scheme:

[0007] A warning information targeted release method based on 5G, comprising:

[0008] S1: Obtain first early warning targeting information, which includes radar, satellite and ground station data, classify according to the type of the first early warning targeting information, and mark the priority of the classified first early warning targeting information;

[0009] S2: Based on the priority marking, a dynamic resource allocation model is constructed, and the processing task of the first early warning targeting information is allocated to the edge computing node and the cloud server through the reinforcement learning algorithm to generate the first targeting feature;

[0010] S3: The first targeting feature is processed by spatio-temporal alignment and feature fusion to generate second early warning targeting information, and local risk assessment is performed by the edge computing node and global correlation analysis is performed by the cloud server to generate the second targeting feature;

[0011] S4: Based on the second targeting feature, a personalized early warning content is generated using a large language model, and a third targeting feature of the personalized early warning content is calculated;

[0012] S5: Based on the second targeting feature and the third targeting feature, a multi-target pushing strategy is generated through an online learning algorithm, and a warning release instruction set is generated based on the multi-target pushing strategy;

[0013] S6: Execute the early warning release instruction set, and collect user response data to start the optimization process.

[0014] Preferably, the S2, the dynamic resource allocation model is composed of state space, action space and reward function, specifically including:

[0015] The task type of the first early warning targeting information, the real-time resource usage of the edge computing node and the network state corresponding to the network slice channel are input into the state space in the form of a vector;

[0016] The task allocation ratio of the edge computing node and the cloud server is output through the action space in the form of a matrix;

[0017] By maximizing the cumulative reward, the reward function guides the dynamic resource allocation model to select the task allocation action that can maximize the long-term revenue under different system states.

[0018] Preferably, the S2, the task type of the first early warning targeting information includes the source of the first early warning targeting information, the size of the superimposed task data packet, and the timeliness label level of the task;

[0019] The real-time resource usage of the edge computing node includes the CPU utilization, memory occupancy, GPU video memory remaining amount and current pending task queue depth of the node;

[0020] The network state corresponding to the network slice channel includes bandwidth utilization of the slice, end-to-end delay of the task, packet loss rate of data transmission, and channel quality indication;

[0021] The task allocation ratio of the edge computing node to the cloud server is the first target feature.

[0022] Preferably, the S3, the second target feature is a dynamic risk map, and the obtaining step comprises:

[0023] Deploying a lightweight convolutional neural network on the edge computing node extracts a disaster probability map of a target area corresponding to the second early warning target information, and the disaster probability map is a probability of occurrence of a corresponding disaster type in the target area;

[0024] Deploying a Transformer model on the cloud server performs cross-regional correlation analysis to generate a global cross-regional correlation risk heat map, and the cross-regional correlation risk heat map is a risk distribution and mutual influence degree of the target area;

[0025] The Kriging interpolation algorithm is used to perform spatial interpolation on the disaster probability map and the cross-regional correlation risk heat map to generate a dynamic risk map corresponding to the target area.

[0026] Preferably, the S3, the dynamic risk map corresponding to the target area is composed of risk distribution values corresponding to each observation point in the target area, which are spatial grid points formed after the original data source for obtaining the first early warning target information in the S1 is spatiotemporally aligned and gridded, and located in the target area, wherein the risk distribution value in the dynamic risk map corresponding to the disaster probability map is calculated as follows:

[0027] ,

[0028] wherein, represents the observation point to be calculated, represents the number of observation points, represents the index of the observation point, represents the disaster probability at the observation point , and represents the contribution degree of the disaster probability at the observation point to the disaster probability at the observation point ,

[0029] The calculation formula of the contribution degree of the disaster probability at the observation point to the disaster probability at the observation point is specifically represented as:

[0030] ​​ ,

[0031] wherein, is a semivariogram value between observation points and , is a semivariogram value between observation points and observation points , all represent indexes of observation points; the semivariogram value is calculated according to a formula, and specifically represented as:

[0032] ,

[0033] wherein, is a semivariogram value at a distance and a direction , is a number of observation point groups at a distance and a direction , is a disaster probability at an observation point , is a disaster probability at an observation point away from .

[0034] Preferably, the S4, acquisition of the third target feature, specifically comprises:

[0035] generating an individualized early warning vector through feature intersection based on the second early warning target information, the second target feature and the historical operation database;

[0036] adopting a parameter efficient fine-tuning technology and combining a preset prompt template to guide a large language model to generate an individualized early warning text containing a disaster avoidance route suggestion;

[0037] the disaster avoidance route suggestion is based on a position of the user and a safety area marked in a dynamic risk map, and a shortest path from the position of the user to an optimal safety area is calculated by using a Dijkstra algorithm.

[0038] Preferably, the S4, the individualized early warning vector comprises a user attribute vector representing individualized attributes of the user and the position of the user, a real-time vector representing a real-time risk situation around the position of the user and a knowledge vector containing meteorological disaster information;

[0039] the user attribute vector is extracted from the historical operation database by extracting user historical behavior features related to early warning information receiving and responding and real-time geographic position information of the user;

[0040] The real-time vector marks a region with a disaster probability and a risk probability higher than a preset threshold as a high-risk region by analyzing dynamic risk map data in the second targeted feature;

[0041] The knowledge vector includes meteorological disaster information and response measures.

[0042] Preferably, the S5, the acquisition of the third targeted feature, specifically further includes:

[0043] The attention mechanism inside the large language model is used to extract an attention head related to a meteorological entity corresponding to the knowledge vector, and an average attention score of the attention head to a key meteorological entity in the process of generating the early warning text is calculated , specifically represented as:

[0044] ,

[0045] wherein, is the number of attention heads, is the index of the attention head, is a query matrix of each vector in the personalized early warning vector, is a key matrix of each vector in the personalized early warning vector, is the transpose of the key matrix, is the dimension of the query matrix.

[0046] The average attention score , the probability of the large language model corresponding to the generation of the personalized early warning text , and the similarity between the meteorological entity extracted in the personalized early warning text and the corresponding meteorological entity in the knowledge vector are fused and calculated to obtain a semantic confidence index , specifically represented as:

[0047] ,

[0048] wherein, , , are the weights of the average attention score, the probability of the large language model corresponding to the generation of the personalized early warning text, and the similarity between the meteorological entity extracted in the personalized early warning text and the corresponding meteorological entity in the knowledge vector, respectively, is the cosine similarity between the meteorological entity extracted in the personalized early warning text and the corresponding meteorological entity in the knowledge vector.

[0049] To achieve the above object, the application provides the following technical scheme: a 5G-based early warning information targeted publishing platform, comprising a historical operation database, a central processing module and a user information terminal, implementing a 5G-based early warning information targeted publishing method, comprising:

[0050] An early warning information acquisition module: acquires first early warning targeted information, which includes radar, satellite and ground station data, classifies the first early warning targeted information according to the types of the first early warning targeted information, and marks the classified first early warning targeted information with priorities;

[0051] A computing task allocation module: based on the priority marking, constructs a dynamic resource allocation model, allocates the processing tasks of the first early warning targeted information to edge computing nodes and cloud servers through a reinforcement learning algorithm, and generates first targeted features;

[0052] A dynamic risk analysis module: performs spatio-temporal alignment and feature fusion processing on the first targeted features, generates second early warning targeted information, and performs local risk assessment using edge computing nodes and global correlation analysis using cloud servers to generate second targeted features;

[0053] A targeted information analysis module: based on the second targeted features, generates personalized early warning content using a large language model, and calculates third targeted features of the personalized early warning content;

[0054] A targeted instruction generation module: based on the second targeted features and the third targeted features, generates a multi-target pushing strategy through an online learning algorithm, and generates a set of early warning publishing instructions based on the multi-target pushing strategy;

[0055] An iterative optimization module: executes the set of early warning publishing instructions and collects user response data to start an optimization process;

[0056] The historical operation database is all data texts of the 5G-based early warning information targeted publishing platform, and real-time collection of information texts output by each module, the central processing module is used for controlling information text instructions output by each module in the platform, and the user information terminal is an information output device receiving the 5G-based early warning information targeted publishing platform.

[0057] Preferably, the iterative optimization module, the steps of the optimization process specifically include:

[0058] According to the early warning publishing instruction set of the targeted instruction generation module, the 5G network slice interface is called to dynamically adjust QoS parameters, and early warning information targeted distribution is completed;

[0059] Based on the user response data, trigger model parameter update of risk assessment in the dynamic risk analysis module and prompt word optimization of the large language model in the targeted information analysis module after differential privacy cleaning;

[0060] After executing the early warning release instruction set, the network slice channel state data is backflowed to the dynamic resource allocation model in the computing task allocation module for next cycle resource scheduling optimization;And

[0061] Periodically distill the model knowledge deployed by the cloud server to the model deployed by the edge computing node, and inject the user response data into the training set of the large language model in the targeted information analysis module to form a closed loop evolution link;

[0062] Among them, the user response data includes user response delay and click rate data.

[0063] The technical effects and advantages of the present application are:

[0064] 1、The present application solves the problems of load fluctuation and resource fragmentation caused by traditional static allocation by constructing a dynamic resource allocation model and setting a task reallocation mechanism when the network slice bandwidth fluctuates, realizes dynamic collaboration of edge-cloud resources, reduces network load fluctuation, reduces computing resource fragmentation, and improves resource utilization efficiency;

[0065] 2、The present application shares feature weights through knowledge distillation between the lightweight convolutional neural network of the edge computing node and the Transformer model of the cloud server, and dynamically adjusts the distillation frequency, which alleviates the problem that the lightweight model is difficult to meet the demand of high-precision risk assessment and personalized content generation, realizes the combination of local risk assessment and global correlation analysis, improves the risk assessment accuracy and personalized content generation ability, alleviates the decline of end-to-end reliability caused by multiple subsystem calls, and improves the end-to-end response reliability;

[0066] 3、The present application establishes a game revenue function covering rate and false alarm rate, and uses an online gradient descent algorithm to solve the multi-objective push strategy, solves the problem that the early warning push strategy depends on the static rule base, realizes dynamic strategy optimization, and balances the early warning coverage rate and accuracy in extreme scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0067] Figure 1 An implementation step flowchart of a 5G-based early warning information targeted release method according to an embodiment of the present application is provided.

[0068] Figure 2 A module block diagram of a 5G-based early warning information targeted release platform according to an embodiment of the present application is provided. DETAILED DESCRIPTION

[0069] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0070] The terms used in the following embodiments of the present application are only for the purpose of describing the specific embodiments, and are not intended to be limiting on the present application. As used in the specification of the present application, the singular forms "a," "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "and / or" used in the present application, means and includes any or all possible combinations of one or more listed items.

[0071] Hereinafter, the terms "first", "second", and "third" are only for the purpose of description, and cannot be understood as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined with "first", "second", and "third" can explicitly or implicitly include one or more features, and in the description of the embodiments of the present application, the meaning of "a plurality of" is two or more, unless otherwise specified.

[0072] As shown in the accompanying drawings Figure 1 A 5G-based early warning information targeted release method, by acquiring multi-modal meteorological sensor data, and dynamically allocating edge computing nodes and cloud server resources for data processing and risk assessment, to generate early warning release instruction set, execute early warning release instruction set and start system optimization process, to realize the targeted release of early warning information. Specifically, the following steps are included:

[0073] S1: obtaining first early warning targeting information, the first early warning targeting information includes radar, satellite and ground station data, classifying according to the type of the first early warning targeting information, and marking the priority of the classified first early warning targeting information;

[0074] S2: based on the priority marking, a dynamic resource allocation model is constructed, and the processing task of the first early warning targeting information is allocated to the edge computing node and the cloud server through the reinforcement learning algorithm, to generate the first targeting feature;

[0075] S3: performing spatio-temporal alignment and feature fusion processing on the first targeting feature, generating second early warning targeting information, and using the edge computing node for local risk assessment, and the cloud server for global correlation analysis, to generate the second targeting feature;

[0076] S4: generating personalized early warning content by using a large language model based on the second targeting feature, and calculating a third targeting feature of the personalized early warning content;

[0077] S5: generating a multi-target pushing strategy by an online learning algorithm based on the second targeting feature and the third targeting feature, and generating a set of early warning publishing instructions based on the multi-target pushing strategy;

[0078] S6: executing the set of early warning publishing instructions, and collecting user response data to start an optimization process.

[0079] Specifically, in S1, raw heterogeneous sensor data from various types of meteorological sensors, including but not limited to radars, satellites, and ground stations, is obtained and preliminarily processed. Based on 5G network slicing technology, a dedicated network slicing channel is divided and created for each type of sensor on the 5G access network side, and a differentiated QoS strategy is configured. The raw heterogeneous sensor data is converted into structured data packets, and priority labels are marked according to task type and importance.

[0080] In this embodiment, a radar data dedicated channel is created, and a URLLC mode QoS strategy is configured, with a single-frame transmission latency upper limit set and sufficient bandwidth reserved. In combination with the time-sensitive network protocol, periodic synchronous transmission of radar scanning data is ensured. A satellite data dedicated channel is created, and an eMBB mode QoS strategy is enabled to allocate dynamic bandwidth to adapt to data volume changes in different time periods, and block compression encoding is used to further reduce transmission load. A ground station dedicated channel is created, and a data packet aggregation strategy is implemented to effectively reduce signaling overhead and improve network resource utilization to adapt to the connection needs of a large number of terminals.

[0081] Specifically, in S2, the priority label corresponding to the first early warning targeting information is received, and a dynamic resource allocation model is built on the cloud server through a deep reinforcement learning framework. The first targeting feature is represented in matrix form by real-time allocation of computing tasks on the edge node.

[0082] In one possible implementation, a dynamic resource allocation model for 5G network slicing channels is built based on a deep reinforcement learning framework. The dynamic resource allocation model is composed of a state space, an action space, and a reward function.

[0083] Furthermore, constructing a dynamic resource allocation model includes: inputting the task type of the first early warning targeting information, the real-time resource usage of the edge computing nodes, and the network status corresponding to the network slice channel into the state space in vector form; it should be noted that the task type of the first early warning targeting information includes the source of the first early warning targeting information, the size of the superimposed task data packet, and the timeliness label level of the task; the real-time resource usage of the edge computing nodes includes the node's CPU utilization, memory usage, remaining GPU memory, and the depth of the current pending task queue; the network status corresponding to the network slice channel includes the slice's bandwidth utilization, the end-to-end latency of the task, the packet loss rate of data transmission, and the channel quality indicator; the task allocation ratio between the edge computing nodes and the cloud server is output in matrix form through the action space; in this embodiment The matrix's rows correspond to different task types, and its columns correspond to the computing platforms of edge computing nodes and cloud servers. The value range of each element in the matrix is ​​constrained to [0,1], representing the allocation ratio of each task type on its corresponding computing platform. By maximizing cumulative rewards, the reward function guides the dynamic resource allocation model to learn the optimal resource allocation strategy. This means the dynamic resource allocation model can select task allocation actions that maximize long-term benefits under different system states. These rewards include, but are not limited to, positive rewards for correctly identifying disaster types and taking appropriate response measures, negative rewards for incorrect identification or inappropriate responses leading to more serious consequences, positive rewards for complying with privacy and security requirements, and negative rewards for violating privacy and security requirements. In this embodiment, the reward function is a multi-objective optimization function that minimizes latency, maximizes bandwidth efficiency, and achieves load balancing design. Specifically, it is expressed as:

[0084] ,

[0085] in, , , These represent the weighting coefficients for task processing latency, bandwidth utilization efficiency, and edge node load, respectively. This represents the end-to-end processing latency for task processing. Bandwidth utilization efficiency is expressed as the ratio of the effective amount of data transmitted to the total bandwidth of the slice. Represented as the first Real-time load rate of each edge node This is expressed as the average load rate of all edge nodes. This represents the total number of edge nodes.

[0086] It should be noted that the first target feature is the proportion of task allocation of different task types at the edge computing node and the cloud server; when the 5G network slice bandwidth fluctuation value exceeds the preset threshold, the edge and cloud task reallocation process is forcibly started, the original cloud server processing space modeling task is downgraded to the low precision mode of the edge computing node for execution, and the model state is synchronized through the knowledge distillation link of S3.

[0087] Specifically, in S3, the first target feature obtained after S2 processing is processed through space-time alignment and feature fusion to generate second early warning target information, and a lightweight convolutional neural network model is run on the edge computing node for local risk assessment, and a Transformer model is deployed on the cloud server for global correlation analysis, and the two models share feature weights through knowledge distillation to obtain a dynamic risk map, i.e. the second target feature.

[0088] In this embodiment, a sliding window compensation mechanism is implemented, a sliding window of a preset fixed size is set, data in the sliding window is analyzed to arrange the time sequence of the data of the first target feature continuously and synchronously in space; further, after the space-time alignment of the first target feature is completed, deep feature extraction and fusion are performed, a convolutional neural network is used to capture fluctuation features and local patterns in the time dimension; a Transformer model is introduced to model long-distance dependencies and global correlations in the spatial dimension; through a cross-attention mechanism, the first target feature after local and global feature extraction is dynamically weighted to generate a target area climate environment state in the form of a feature vector, i.e. the second early warning target information.

[0089] In one possible implementation, obtaining the second target feature includes: deploying a lightweight convolutional neural network on an edge computing node to extract a disaster probability map of the target area corresponding to the second early warning target information, wherein the disaster probability map represents the probability of the corresponding disaster type occurring in the target area; this embodiment uses a variant of MobileNetV3-Small, utilizing depthwise separable convolution to improve efficiency; deploying a Transformer model on a cloud server to perform cross-regional correlation analysis, generating a global cross-regional correlation risk heatmap, wherein the cross-regional correlation risk heatmap represents the risk distribution and mutual influence degree of the target area; employing a knowledge distillation mechanism to transfer the knowledge of the Transformer model on the cloud server to the lightweight convolutional neural network on the edge computing node; using a Kriging interpolation algorithm to spatially interpolate the disaster probability map and the cross-regional correlation risk heatmap to generate a dynamic risk map corresponding to the target area; in this embodiment, during the interpolation process, the spatial correlation of meteorological parameters is analyzed by combining a semi-variogram function to optimize the interpolation accuracy; simultaneously, for areas where the risk level gradient change exceeds a preset threshold, an adaptive resolution enhancement strategy is implemented, and the resolution of the corresponding risk map is increased.

[0090] Furthermore, the knowledge distillation mechanism minimizes the KL divergence between the disaster probability map corresponding to the edge computing node and the cross-regional associated risk heatmap corresponding to the cloud server by designing feature alignment loss. At the same time, a dynamic distillation frequency mechanism is introduced to dynamically adjust the distillation frequency according to the 5G channel quality index. In this embodiment, when the 5G channel quality index is greater than or equal to 15, it indicates that the channel quality is good, and knowledge distillation is set to be performed once per hour; when the 5G channel quality index is less than 15, it indicates that the channel quality is poor, and the distillation frequency is reduced to once every 3 hours.

[0091] Furthermore, the dynamic risk map corresponding to the target area is composed of risk distribution values ​​corresponding to each pre-set observation point within the target area. The observation point is a spatial grid point located within the target area, formed by spatiotemporal alignment and gridding of the original data source used to obtain the first early warning target information in S1. The risk distribution values ​​in the dynamic risk map corresponding to the disaster probability map... The calculation is specifically expressed as follows:

[0092] ,

[0093] in, The observation points are represented as the target calculation points. This is expressed as the number of observation points. Represented as the index of the observation point, Represented as observation point The probability of disaster at the observation point The contribution degree of the disaster probability at the observation point to the disaster probability at the observation point ;

[0094] It should be noted that the calculation formula of the contribution degree of the disaster probability at the observation point to the disaster probability at the observation point , is specifically represented as:

[0095] ,

[0096] wherein, is the semi-variogram value between the observation points and , is the semi-variogram value between the observation points and the observation point , all represent the index of the observation point; and the calculation formula of the semi-variogram value is specifically represented as:

[0097] ,

[0098] wherein, is the semi-variogram value at the distance and direction , is the number of observation point groups at the distance and direction , is the disaster probability at the observation point , is the disaster probability at the observation point away from ;

[0099] The calculation of the risk distribution value in the dynamic risk atlas corresponding to the cross-regional correlation risk heat map is specifically represented as:

[0100] ,

[0101] wherein, is the observation point for target calculation, is the number of observation points, is the index of the observation point, is the contribution degree of the risk probability at the observation point to the risk probability at the observation point , is the risk probability at the observation point ;

[0102] ​It should be noted that the risk probability at the observation point The contribution of the risk probability at the observation point The calculation formula of the risk probability at the observation point Specifically represented as:

[0103] ,

[0104] Wherein, Indicated as the semi-variogram value between the observation points And Indicated as the semi-variogram value between the observation points And the observation point All represent the index of the observation point; The calculation formula of the semi-variogram value is specifically represented as:

[0105] ,

[0106] Wherein, Indicated as the semi-variogram value at the distance and direction Indicated as the number of observation point groups at the distance and direction Indicated as the risk probability at the observation point Indicated as the risk probability at the observation point Indicated as the risk probability at the observation point Indicated as the risk probability at the observation point Indicated as the risk probability at the observation point

[0107] Specifically, in S4, the feature vector corresponding to the second early warning targeting information is received, and the dynamic risk graph corresponding to the second targeting feature is combined with the large language model stored in the historical operation database to generate personalized early warning text, and the semantic confidence index corresponding to the personalized early warning content is calculated, that is, the third targeting feature.

[0108] ​​​​In a possible implementation, the obtaining of the third targeting feature includes: generating, based on the second early warning targeting information, the second targeting feature, and a historical operation database, an individualized early warning vector through feature intersection, the individualized early warning vector including a user attribute vector representing individualized attributes of the user and a current location of the user, a real-time vector representing a real-time risk situation around the current location of the user, and a knowledge vector containing meteorological disaster information, wherein the user attribute vector is generated by extracting, from the historical operation database, historical behavior features of the user related to early warning information reception and response and real-time geographic location information of the user; in this embodiment, the historical behavior features of the user include but are not limited to click preferences of historical early warning information, commonly used reception device types, response speeds to early warning information, and the like; the real-time vector is generated by analyzing dynamic risk map data in the second targeting feature, and marking areas with disaster probabilities and risk probabilities higher than a preset threshold as high-risk areas; the knowledge vector contains meteorological disaster information and countermeasures; in this embodiment, the meteorological disaster information includes but is not limited to typhoon path history, rainstorm water accumulation point distribution, characteristics of different disaster types, and the like; a parameter efficient fine-tuning technology is used to guide a large language model to generate an individualized early warning text containing disaster avoidance route suggestions, in combination with a preset prompt template; in this embodiment, the preset prompt template adopts a structured form and is specifically expressed as: [user location] is currently facing [disaster type], it is suggested to perform [action], and the disaster avoidance route is [route]; the disaster avoidance route is calculated based on the location of the user and the safe areas marked in the dynamic risk map, using a Dijkstra algorithm to calculate the shortest path from the location of the user to the optimal safe area.

[0109] In a possible implementation, the obtaining of the third targeting feature further includes: extracting, using an attention mechanism inside the large language model, an attention head related to a meteorological entity corresponding to the knowledge vector; calculating an average attention score of the attention head to a key meteorological entity in the process of generating the early warning text , which is specifically expressed as:

[0110] ,

[0111] wherein, is the number of attention heads, is the index of the attention head, is a query matrix of each vector in the individualized early warning vector, is a key matrix of each vector in the individualized early warning vector, is a transpose of the key matrix, is the dimension of the query matrix.

[0112] The average attention score is used to generate a probability of the large language model corresponding to the individualized early warning text. And the meteorological entities extracted from the personalized warning text. Corresponding meteorological entities in the knowledge vector similarity Perform fusion computation to obtain semantic confidence metrics Specifically, it is expressed as:

[0113] ,

[0114] in, , , These represent the average attention score, the probability of generating the large language model corresponding to the personalized warning text, and the weight of the similarity between the meteorological entities extracted from the personalized warning text and the corresponding meteorological entities in the knowledge vector, respectively. This is represented as the cosine similarity between the meteorological entity extracted from the personalized warning text and the corresponding meteorological entity in the knowledge vector; in this embodiment, it is based on the calculated semantic confidence index. Preset tiered thresholds and corresponding push strategies, including: when A value ≥ 0.8 indicates extremely high confidence, corresponding to an emergency warning level, and a multi-channel real-time push strategy is adopted; when 0.6 ≤ A value less than 0.8 indicates a high confidence level, corresponding to a high alert level, and a strategy prioritizing App push notifications and SMS messages is adopted; when... A value less than 0.6 indicates a low confidence level, corresponding to a standard warning level. This level is indicated by an app pop-up or social media push notification and requires manual review.

[0115] Specifically, in S5, the dynamic risk map corresponding to the second target feature and the semantic confidence index of the third target feature are analyzed. A multi-target push strategy is generated through an online learning algorithm, and a corresponding warning release instruction set is generated based on the generated strategy. The compilation and distribution of the instruction set are performed dynamically.

[0116] It should be noted that a game payoff function is constructed using an online learning algorithm to establish a game payoff function with positive payoff for coverage and negative payoff for false positives, in order to quantify the effectiveness of various push strategies. The coverage rate is the ratio of the number of successfully reached users to the total number of users in the target area, and the false positive rate is the ratio of the number of false warnings to the total number of pushes. The push strategy is dynamically adjusted by using an online gradient descent optimization method based on real-time feedback from coverage, false positive rate, and semantic confidence indicators.

[0117] Specifically, in S6, the warning release instruction set generated in S5 is executed. Executing the warning release instruction set includes calling the network slice management interface to dynamically configure the quality of service parameters to achieve the final delivery of the warning information; according to the requirements of the warning release instruction set, the network slice management interface is called to dynamically configure the QoS parameters of the corresponding network slice; in this embodiment, for warning information with high urgency and requiring low latency delivery, the priority, bandwidth guarantee, latency limit, and other parameters of the corresponding network slice are dynamically adjusted to ensure that the information can be delivered to the target user group quickly and reliably through the predetermined channels.

[0118] Furthermore, user response data to published early warning information is collected, including but not limited to user response latency to receiving early warning information, click-through rate of early warning information, and correlation information between user's actual location and the early warning area. The user response data is cleaned and noise-added using a differential privacy mechanism. Based on the cleaned user response data, parameter updates and optimizations of multiple key models are triggered. Specifically, this includes: updating the convolution kernel parameters of the CNN model based on the loss function of the lightweight CNN model fed back from the edge computing node to improve the accuracy of edge-side recognition of local disaster features; correcting the spatial attention weights of the Transformer model deployed on the cloud server according to the offset information between the user's actual location and the early warning area; periodically distilling the model knowledge deployed on the cloud server into the model deployed on the edge computing node, and injecting the collected user response data features into the training set of the large language model to form a closed-loop optimization chain.

[0119] As attached Figure 2 The platform shown is a 5G-based early warning information targeted release platform, which includes a historical operation database, a central processing module and a user information terminal, as well as an early warning information acquisition module, a computing task allocation module, a dynamic risk analysis module, a targeted information analysis module, a targeted instruction generation module and an iterative optimization module.

[0120] Early warning information acquisition module: acquires first early warning target information, which includes radar, satellite and ground station data, classifies the first early warning target information according to its type, and assigns priority to the classified first early warning target information;

[0121] The task allocation module for computing tasks: Based on the priority marker, a dynamic resource allocation model is constructed, and the processing tasks of the first early warning target information are allocated to edge computing nodes and cloud servers through reinforcement learning algorithms to generate the first target feature;

[0122] The dynamic risk analysis module performs spatio-temporal alignment and feature fusion processing on the first target feature, generates second early warning target information, and performs local risk assessment using an edge computing node and global correlation analysis using a cloud server to generate a second target feature;

[0123] The target information analysis module generates personalized early warning content using a large language model based on the second target feature and calculates a third target feature of the personalized early warning content;

[0124] The target instruction generation module generates a multi-target push strategy based on the second target feature and the third target feature through an online learning algorithm and generates an early warning publishing instruction set based on the multi-target push strategy.

[0125] The iterative optimization module executes the early warning publishing instruction set and collects user response data to start an optimization process.

[0126] The historical operation database includes all data texts of the 5G-based early warning information targeted publishing platform, and real-time collection of information texts output by each module, the central processing module is used for controlling information text instructions output by each module in the platform, and the user information terminal is an information output device receiving the 5G-based early warning information targeted publishing platform.

[0127] In this embodiment, the steps of the optimization process include: calling a 5G network slice interface to dynamically adjust QoS parameters according to the early warning publishing instruction set of the target instruction generation module to complete early warning information targeted distribution; based on the user response data, triggering model parameter updating of risk assessment in the dynamic risk analysis module and prompt word optimization of the large language model in the target information analysis module after differential privacy cleaning; flowing the network slice channel state data after execution of the early warning publishing instruction set back to the dynamic resource allocation model in the computing task allocation module for next cycle resource scheduling optimization; and periodically distilling model knowledge deployed by the cloud server to the model deployed by the edge computing node, and injecting the user response data into the training set of the large language model in the target information analysis module to form a closed-loop evolution link; wherein the user response data includes but is not limited to user response latency, click rate data, etc.

[0128] The platform central processor is a core operation and control unit of the whole early warning information targeted release platform, includes one or more processing cores, establishes a hardware level direct connection channel with the 5G baseband chip and the edge computing node through the PCIe bus, executes instructions, programs, code sets or instruction sets stored in the historical operation database, and can call data stored therein, so as to execute various functions of the early warning information targeted release platform, including analyzing 5G network slice bandwidth real-time state data, generating an edge-cloud task allocation matrix, forcibly binding time sequence sensitive tasks to the edge computing core, and distributing spatial modeling tasks to the GPU cluster; calling the space-time encoder model parameter, performing space-time alignment on the heterogeneous data of radar, satellite and ground station, generating a unified feature vector and injecting into the risk assessment pipeline; dynamically adjusting the weight of the large language model prompt word according to the user response behavior data, and synchronously triggering the knowledge distillation of the lightweight CNN model and the gradient back propagation of the Transformer model.

[0129] The communication bus is used to realize the connection and communication between components.

[0130] The historical operation database is used to save a large amount of data related to early warning information targeted release, including mathematical model related data established based on meteorological model basic data, dynamic optimization data set and real-time operation data, a large amount of historical operation data, historical calibration data table of radar reflectivity-precipitation mapping relationship, spatial correlation matrix of satellite infrared cloud image and ground meteorological station observation data, user portrait feature library containing geographical location preference, terminal device type and historical response delay and other parameters; when executing various functions, the platform central processor frequently calls these data from the historical operation database to perform risk assessment, strategy optimization and other operations, so as to realize accurate control and efficient management of early warning information targeted release.

[0131] The user information end provides an interface for the user to interact with the platform by connecting external devices such as display screens and cameras through standard wired or wireless interfaces.

[0132] Secondly, only the structures related to the disclosed embodiments are involved in the drawings of the disclosed embodiments, other structures can be referred to the general design, and the same embodiments and different embodiments of the present application can be combined with each other under the condition of no conflict;

[0133] Finally, the above only describes the preferred embodiments of the present application, and is not used to limit the present application, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A method for targeted distribution of early warning information based on 5G, characterized in that, The method comprises the following steps: S1: obtaining first early warning targeting information, the first early warning targeting information including radar, satellite and ground station data, classifying according to the type of the first early warning targeting information, and marking the priority of the classified first early warning targeting information; S2: based on the priority marking, constructing a dynamic resource allocation model, and assigning the processing task of the first early warning targeting information to an edge computing node and a cloud server through a reinforcement learning algorithm to generate first targeting features; In S2, the priority marking corresponding to the first early warning targeting information is received, and a dynamic resource allocation model is constructed in the cloud server, and the reward function includes a positive reward when the correct disaster type is identified and the correct response measure is taken; S3: performing spatio-temporal alignment and feature fusion processing on the first early warning targeting information to generate second early warning targeting information, and performing local risk assessment by the edge computing node and global correlation analysis by the cloud server to generate second targeting features; In S3, obtaining the second targeting features comprises: the edge computing node extracts a disaster probability map of a target area; the cloud server performs cross-regional correlation analysis to generate a global cross-regional correlation risk heat map; and the disaster probability map and the cross-regional correlation risk heat map are spatially interpolated to generate a dynamic risk map; The dynamic risk map has a risk distribution value The calculation of the dynamic risk map is specifically represented as: wherein, represents an observation point for which a calculation is targeted, represents the number of observation points, represents an index of an observation point, represents a disaster probability for an observation point at which a disaster probability is calculated, represents a degree of contribution of a disaster probability at an observation point to a disaster probability at an observation point corresponding thereto; The observation point The contribution degree of the disaster probability at the observation point The contribution degree of the disaster probability at the observation point The calculation formula is specifically represented as: in, Represented as observation point and The semi-variogram values ​​between Represented as observation point and observation points The semivariogram values ​​between Both are represented as indices of the observation points; S4: based on the second targeting features, generating personalized early warning content using a large language model, and calculating third targeting features of the personalized early warning content; In S4, the steps of calculating the third targeting features are as follows: based on the second early warning targeting information, the second targeting features and a historical operation database, generating a personalized early warning vector through feature cross; real-time vector marks the area with disaster probability and risk probability higher than a preset threshold as a high-risk area by analyzing the dynamic risk map data in the second targeting features; using parameter efficient fine-tuning technology and combining a preset prompt template, guiding the large language model to generate personalized early warning text containing disaster avoidance route suggestions; based on the user's location and the safety area marked in the dynamic risk map, calculating the shortest path from the user's location to the optimal safety area; S5: based on the second targeting features and the third targeting features, generating a multi-objective push strategy through an online learning algorithm, and generating an early warning publishing instruction set based on the multi-objective push strategy; S6: executing the early warning publishing instruction set and collecting user response data to start an optimization process; The optimization process in S6 includes correcting the spatial attention weight of the cloud server according to the offset information between the actual position of the user and the early warning area. 2.The 5G-based early warning information targeted distribution method of claim 1, wherein: In S2, the dynamic resource allocation model is composed of a state space, an action space and a reward function, and specifically includes: The task type of the first early warning targeting information, the real-time resource usage of the edge computing node and the network state corresponding to the network slice channel are input into the state space in the form of a vector; The task allocation proportion of the edge computing node and the cloud server is output through the action space in the form of a matrix; By maximizing the cumulative reward, the reward function guides the dynamic resource allocation model to select task allocation actions in different system states that can maximize long-term returns. 3.The 5G-based early warning information targeted distribution method of claim 2, wherein: The S2, the task type of the first early warning targeting information includes the acquisition source of the first early warning targeting information, the size of the superimposed task data packet, and the timeliness label level of the task. The real-time resource usage of the edge computing node includes the CPU utilization, memory occupancy, GPU video memory remaining amount, and the depth of the current task queue of the node. The network state corresponding to the network slice channel includes the bandwidth utilization of the slice, the end-to-end delay of the task, the packet loss rate of data transmission, and the channel quality indication. The task allocation ratio of the edge computing node to the cloud server is the first targeting feature.

4. The method of claim 1, wherein the method is based on 5G. The S3, the dynamic risk map corresponding to the target area is composed of the risk distribution values of each observation point in the target area, which is formed by the spatial grid points in the target area after the original data source of the first early warning targeting information in S1 is spatio-temporally aligned and gridded. wherein, is represented as a semi-variogram value at distance and direction is represented as a number of observation point groups at distance and direction is represented as a hazard probability at observation point is represented as a hazard probability at observation point is represented as a hazard probability at observation point is represented as a hazard probability at observation point is represented as a hazard probability at observation point is represented as a hazard probability at observation point is represented as a hazard probability at observation point 5. The method of claim 1, wherein the method is based on 5G. The S4, the personalized early warning vector includes a user attribute vector representing the user's personalized attributes and location, a real-time vector representing the real-time risk situation around the user's location, and a knowledge vector containing meteorological disaster information. The user attribute vector is obtained by extracting the user's historical behavior features related to early warning information reception and response and the user's real-time geographic location information from the historical operation database. The real-time vector marks the area with disaster probability and risk probability higher than the preset threshold as a high-risk area by analyzing the dynamic risk map data in the second targeting feature. The knowledge vector contains meteorological disaster information and countermeasures.

6. The method of claim 5, wherein the method further comprises: The S4, the third targeting feature is obtained, which specifically includes: Using the attention mechanism inside the large language model to extract the attention head related to the meteorological entity corresponding to the knowledge vector, and calculating the average attention score A of the key meteorological entity in the process of generating the early warning text, which is specifically represented as: wherein, is represented as a number of attention heads, is represented as an index of an attention head, is represented as a query matrix of each vector in the personalized early warning vector, is represented as a key matrix of each vector in the personalized early warning vector, is represented as a transpose of the key matrix, is represented as a dimension of the query matrix; The average attention score is obtained , a probability of generating a large language model corresponding to the personalized early warning text , and a similarity between the meteorological entity extracted in the personalized early warning text and the corresponding meteorological entity in the knowledge vector The semantic confidence index is obtained by fusion calculation , which is specifically represented as:​​ wherein, , , respectively represent the average attention score, the probability of generating the large language model corresponding to the personalized early warning text, and the weight of the similarity of the meteorological entity extracted in the personalized early warning text and the corresponding meteorological entity in the knowledge vector, represent the cosine similarity of the meteorological entity extracted in the personalized early warning text and the corresponding meteorological entity in the knowledge vector.

7. A 5G-based early warning information targeted release platform employing the early warning information targeted release method of any one of claims 1-6, comprising a historical operation database, a central processing module and a user information terminal, characterized in that, Also includes: Early warning information acquisition module: acquire first early warning targeting information, which includes radar, satellite and ground station data, classify according to the type of the first early warning targeting information, and mark the priority of the classified first early warning targeting information; Computing task allocation module: based on the priority marking, build a dynamic resource allocation model, assign the processing task of the first early warning targeting information to the edge computing node and the cloud server through the reinforcement learning algorithm, and generate the first targeting feature; Dynamic risk analysis module: spatio-temporal alignment and feature fusion processing of the first targeting feature, generation of second early warning targeting information, local risk assessment by the edge computing node, global correlation analysis by the cloud server, and generation of the second targeting feature; Targeting information analysis module: based on the second targeting feature, generate personalized early warning content using a large language model, and calculate the third targeting feature of the personalized early warning content; The target instruction generation module generates a multi-target pushing strategy based on the second target feature and the third target feature through an online learning algorithm, and generates a pre-warning release instruction set based on the multi-target pushing strategy; The iterative optimization module executes the pre-warning release instruction set and collects user response data to start an optimization process; The historical operation database is all data texts of the 5G-based pre-warning information targeted release platform, and real-time collection of information texts output by each module, the central processing module is used for controlling information text instructions output by each module in the platform, and the user information terminal is an information output device receiving the 5G-based pre-warning information targeted release platform. 8.The 5G-based early warning information targeted distribution platform according to claim 7, characterized in that: The iterative optimization module, the steps of the optimization process, specifically include: According to the pre-warning release instruction set of the target instruction generation module, the 5G network slice interface is called to dynamically adjust the QoS parameter, and the pre-warning information targeted distribution is completed; Based on the user response data, the model parameter update of the risk assessment in the dynamic risk analysis module and the prompt word optimization of the large language model in the targeted information analysis module are triggered after differential privacy cleaning; The network slice channel state data after the execution of the pre-warning release instruction set is backflowed to the dynamic resource allocation model in the computing task allocation module, which is used for resource scheduling optimization in the next cycle; and Periodically, the model knowledge deployed by the cloud server is distilled to the model deployed by the edge computing node, and the user response data is injected into the training set of the large language model in the targeted information analysis module, forming a closed-loop evolution link; The user response data includes user response time delay and click rate data.

Images