Emergency early warning information sending method and system

Abstract

The invention provides an emergency early warning information sending method and system, and the method comprises the steps: extracting stop point information from user travel data, and carrying out the analysis through combining the historical travel data of a user, and obtaining a user stop point type tag; generating a dynamic warning level according to the risk level in combination with the weather and the people flow density of the stay point, extracting landmarks around the stay point according to the warning level and the risk type, and generating risk avoiding guidance content; according to the risk avoiding guidance content, through a staying point prediction time window and staying point signal intensity analysis, combining staying point latitude and longitude coordinates of hotels or indoor places in a surrounding geographic information database to obtain a personalized safety staying point recommendation list; the method comprises the following steps: sorting recommendation lists by adopting a multi-target optimization method through personalized safety stay point recommendation lists in combination with stay point path preference and stay point equipment electric quantity data, and determining a path recommendation scheme according to a sorted recommendation sequence and the tolerance of a user to severe weather.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method and system for sending early warning information for emergencies. Background Technology

[0002] Research on meteorological disaster early warning is a core topic in the field of public safety and emergency management. Its importance lies in significantly reducing the threat of disasters to human life and property through timely and accurate information dissemination. With the intensification of climate change and the increasing frequency of extreme weather events, optimizing early warning mechanisms to enhance societal response capabilities has become an urgent and undeniable need. Traditional early warning systems primarily rely on regional broadcasts, depending on meteorological data to predict the scope and intensity of disasters, but often overlook the differences in individual user behavior and scenarios, resulting in a lack of targeted information delivery and difficulty in effectively guiding people to take evasive action. The limitations of existing methods lie in the fact that their early warning information is mostly static and uniform template-based, failing to fully integrate users' real-time location, behavioral patterns, and types of locations. For example, users in outdoor locations may be more sensitive to disaster information, but existing systems struggle to dynamically adjust the content delivered. Simultaneously, group behaviors such as evacuation trends are not effectively incorporated, leading to inefficient resource allocation and guidance. These shortcomings limit the practicality of early warning information in complex scenarios, especially in high-risk areas, where users may miss opportunities for evacuation due to a lack of personalized guidance. The core challenges in this research area lie in how to integrate individual travel characteristics with group behavior patterns, and how to dynamically optimize information delivery based on the type of stopover (e.g., indoor / outdoor, work / leisure). Specifically, differences in individual sensitivity to disaster information, quantitative analysis of historical response behavior, and prediction of real-time group numbers and evacuation directions are pressing technical challenges. The lack of sufficient integration of these factors makes it difficult for early warning systems to find a balance between individual needs and group trends, thus affecting the accuracy and effectiveness of the delivered content. Therefore, the key to improving the precision of meteorological disaster early warning lies in how to analyze differences in user sensitivity to disaster information at different stopover points, combine individual travel preferences with predictions of group behavior, dynamically adjust the presentation and content of early warning information, and provide personalized evacuation guidance when users are about to enter high-risk areas. Summary of the Invention

[0003] This invention provides a method for sending early warning information for emergencies, mainly including: The system extracts stop point information from user travel data and analyzes it in conjunction with historical user travel data to obtain stop point type tags. It then extracts stop point frequency and route preferences from historical user travel data, combines these with stop point type tags, calculates the distribution characteristics of user sensitivity to disaster information, and determines the prediction threshold for high-risk areas. If a user is about to enter a high-risk area, it tracks stop point information in real time, determines the proportion of open outdoor spaces, and assesses the risk level based on the high-risk area prediction threshold. Based on the risk level, combined with weather and pedestrian density at the stop point, it generates a dynamic warning level. Based on the warning level and risk type, it extracts surrounding landmarks to generate evacuation guidance content. Based on the evacuation guidance content, it analyzes the stop point prediction time window and signal strength, and integrates the latitude and longitude coordinates of hotels or indoor venues in the surrounding geographic information database to obtain a personalized safe stop point recommendation list. Finally, it uses this personalized safe stop point recommendation list, combined with stop point route preferences and device battery data, employs a multi-objective optimization method to sort the recommendation list. Based on the sorted recommendation sequence and the user's tolerance for severe weather, it determines a route recommendation scheme. Real-time pedestrian density data at rest points is extracted from the route recommendation scheme. By analyzing the movement trajectory at rest points and surrounding landmarks, the evacuation direction of the population in high-risk areas is determined, and dynamic adjustment parameters for the population are obtained. Based on the dynamic adjustment parameters and the route recommendation scheme, personalized early warning push content is generated and transmitted to user devices in combination with the signal strength at rest points.

[0004] Furthermore, stop point information is extracted from user travel data and analyzed in conjunction with historical user travel data to obtain user stop point type tags. This includes: obtaining stop point coordinate sequences and corresponding timestamps from spatiotemporal trajectory data based on latitude and longitude coordinates collected by user devices, and obtaining a set of user stop points by setting stop time and distance thresholds. Based on the distance and time difference between adjacent points in the user stop point set, the transportation mode identifier between stop points is obtained through comparative analysis of transportation mode speed interval thresholds, and a transportation connection matrix for stop points is generated. Based on the transportation connection matrix of stop points and historical user transportation selection data, a deep learning clustering algorithm is used to calculate the location transfer patterns of stop points, obtaining a sequence of user activity pattern point locations. Based on the sequence of user activity pattern point locations, device positioning signal strength data, and location change frequency data, a random forest algorithm is used to train a location scene classifier to obtain the indoor and outdoor scene types of each stop point. For the indoor and outdoor scene types, arrival timestamp data, and duration data of stop points, the location attributes of stop points are obtained by matching with a geographic information database and comparing with a historical location tag database. Based on the location attributes of the stopover points and historical travel patterns, the work and leisure attribute identifiers of the stopover points are obtained by calculating the proportion of time spent at the stopover points on weekdays and weekends. Based on the work and leisure attribute identifiers, indoor and outdoor scene types, and location attributes, a spatiotemporal feature-based labeling classification rule is used to obtain the type labels for the stopover points.

[0005] Furthermore, the frequency of stopover points and path preferences are extracted from users' historical travel data. Combined with stopover point type tags, the sensitivity distribution characteristics of users to disaster information are calculated to determine the prediction threshold for high-risk areas. If a user is about to enter a high-risk area, stopover point information is tracked in real time to determine the proportion of open outdoor venues. Combined with the high-risk area prediction threshold, the risk level is assessed, including: calculating the user's activity frequency and location preferences at different times using a long short-term memory neural network based on the stopover point sequence and timestamp information in the user's historical travel data to obtain the spatiotemporal characteristic data of user activities. Based on the spatiotemporal characteristic data of user activities, a density clustering algorithm is used to calculate the user's high-frequency stopover point set, and the user's preferred path sequence is extracted from historical trajectory data to obtain the user's path selection characteristics. For the high-frequency stopover point set and user path selection characteristics, the spatial distribution data of open venues is extracted from the geographic information database, and the frequency and duration of users passing through open venues are calculated to obtain outdoor activity characteristics. Based on the outdoor activity characteristics and real-time location monitoring data, a geofencing algorithm is used to determine whether a user is about to enter a high-risk area, and the risk warning value of the user's activity area is obtained. For the risk warning values ​​of user activity areas, risk level determination rules are extracted from the disaster warning database, and the risk level result is calculated based on the user's location relocation probability. Based on the risk level result and the user's location relocation probability, records of user avoidance behaviors for different types of disasters are extracted from historical trajectory data to obtain the user's sensitivity characteristics to disaster information. Based on the user's sensitivity characteristics to disaster information and the percentage of open spaces in the current area, a decision tree algorithm is used to calculate the risk assessment value of the user's area, thus obtaining the regional risk level.

[0006] Furthermore, based on the risk level, combined with the weather and crowd density at the stop point, a dynamic warning level is generated. Based on the warning level and risk type, landmarks around the stop point are extracted to generate evacuation guidance content, including: calculating a weather disaster risk index using a deep learning neural network based on real-time meteorological monitoring data and historical weather disaster records; extracting real-time crowd density values ​​from the stop point monitoring data to obtain a regional risk level index. Based on the regional risk level index and crowd density values, regional crowd saturation is calculated by comparing with a preset crowd density threshold, and a warning level matrix is ​​obtained using a support vector regression algorithm. Based on the warning level matrix, real-time meteorological indicators, and regional crowd saturation, the regional warning level is calculated using thresholds in the risk warning rule base to obtain dynamic warning data. For the dynamic warning data and the current crowd distribution, a distribution map of evacuation sites is extracted from a geographic information database, and an initial evacuation route is obtained by calculating the shortest path between evacuation sites and the stop point. Based on the initial evacuation route and real-time capacity data of evacuation sites, the remaining capacity of each evacuation site is calculated using crowd density distribution to obtain a list of available evacuation sites. Based on the list of available safe havens and real-time traffic data, surrounding landmark information is extracted from a geographic information database, and the optimal evacuation route is obtained by calculating the connectivity between landmarks. According to the optimal evacuation route and landmark distribution data, key landmark nodes are extracted by segmenting the route, and evacuation guidance content containing landmark guidance information is generated.

[0007] Furthermore, based on the risk avoidance guidelines, a personalized list of safe rest stops is obtained by analyzing the predicted time window and signal strength of rest stops, and integrating the latitude and longitude coordinates of rest stops from the surrounding geographic information database. This includes: calculating the arrival time and duration of user rest stops using a long short-term memory neural network based on user historical trajectory data and rest stop time records to obtain rest stop prediction time interval data; extracting the location coordinates of surrounding hotels and indoor venues from the geographic information database based on the rest stop prediction time interval data and device signal strength data, and obtaining a candidate rest stop set using a density clustering algorithm; assessing the safety of candidate rest stops based on the candidate rest stop set and real-time pedestrian flow monitoring data by calculating the configuration of building risk avoidance facilities and space capacity, resulting in a rest stop list with a risk avoidance index; calculating the travel time and path distance to the current location based on the rest stop list with the risk avoidance index and surrounding traffic data to obtain a safe and convenient rest stop sequence; and extracting the user's rest patterns in different venue types based on the safe and convenient rest stop sequence and historical rest preferences to obtain personalized venue recommendation weights. Based on the personalized location recommendation weights and building type data, a support vector machine algorithm is used to calculate the location suitability. The optimal arrival path is then matched from the evacuation guidance database to obtain a personalized list of safe resting points. Finally, based on this personalized list and real-time weather data, an indoor resting suitability index is calculated to obtain a personalized list of safe resting points.

[0008] Furthermore, based on the stop point data, a preset time window segmentation method is used to obtain a prediction data set. Signal strength information is extracted from this set, and the regional range of candidate stop points is determined by calculating the intensity change rate. By analyzing the fluctuation amplitude of signal strength within the region and the length of the time window, it is determined whether the region meets the stop point conditions. If it does, the stop point location is determined, and the corresponding time range is obtained. Determining the stop point location and time range includes: based on the stop point data and time stamps collected by the user equipment, extracting time series features using a long short-term memory neural network, and segmenting the continuous time data to obtain a set of predicted time segments. Based on the set of predicted time segments and the location coordinate sequence, the signal strength difference within a fixed sampling interval is calculated to obtain the signal strength change sequence. For the signal strength change sequence, a density clustering algorithm is used to calculate the signal fluctuation characteristics, and after removing outliers, the signal stable period is extracted to obtain the candidate stop area range. Based on the candidate stop area range and the signal stable period, a preset signal fluctuation threshold is compared to determine whether the signal strength change meets the stability condition. If the stability condition is met, the location cluster center point is obtained by calculating the spatial density distribution of sampling points within the region based on the signal strength distribution characteristics. For the location cluster center point, a duration threshold is extracted from a pre-defined dwell time rule base. The initial dwell point location is obtained by calculating the signal stability duration. Based on the initial dwell point location and the predicted time segment set, the start and end time range of the dwell point is obtained by matching the signal stability interval. For the start and end time range of the dwell point and its location coordinates, the region boundary constraints are calculated to obtain the dwell point location and its corresponding time interval.

[0009] Furthermore, based on the real-time signal strength data collected by the user equipment, the ratio of the standard deviation to the mean of the signal strength value per unit time is calculated. If the ratio is greater than a preset threshold, a sliding window method is used to initially segment the time series, resulting in multiple initial time window sequences. By analyzing the distribution characteristics of the signal change rate within each initial time window, the boundary points of the time windows are determined, resulting in adjusted time windows. This includes: calculating the signal mean and standard deviation per unit time using sliding sampling based on the continuous signal strength data sequence collected by the user equipment, obtaining a signal strength standard deviation-to-mean ratio sequence. Based on the signal strength standard deviation-to-mean ratio sequence, data stability is judged by comparing it with a preset fluctuation threshold. If it is greater than the preset threshold, a support vector regression algorithm is used to identify signal abrupt changes, obtaining an initial segmentation point sequence. For the initial segmentation point sequence, a fixed-size time window is used to segment the data. By calculating the distribution density of the signal change rate within the window, a window division sequence is obtained. Based on the window division sequence, the signal strength mean and standard deviation features are extracted from each time window. By calculating the feature difference between adjacent windows, the optimized window boundary positions are obtained. For the optimized window boundary positions, the signal change trends on both sides of the boundary are calculated to identify stable and fluctuating signal intervals, thus obtaining an adjustment scheme for the time window. Based on this adjustment scheme, adjacent windows with similar signal characteristics are merged to fine-tune the window boundaries, resulting in a time window sequence with fused boundaries. For this fused time window sequence, a moving average method is used to smooth the signal change rate within the window, obtaining the time window boundary point positions and forming the adjusted time window.

[0010] Furthermore, time-series features of the signal rate of change are extracted from the adjusted time window. By calculating the peak value and duration of the stationary phase of the rate of change, a preliminary feature set of candidate dwell points is obtained. A density clustering algorithm is used to group the preliminary feature set. By calculating the distance and density distribution between feature points, the regional range of candidate dwell points is determined. This includes: extracting the signal rate of change sequence from the signal intensity difference within a fixed sampling interval based on the adjusted time window data; using a long short-term memory neural network to extract features of the rate of change, resulting in a signal time-series feature set. Based on the signal time-series feature set, by detecting local maxima and continuous stable intervals of the signal rate of change, the peak duration and stationary phase duration of each feature point are calculated, resulting in a feature point sequence. For the feature point sequence, a density clustering algorithm is used to calculate the spatial distribution density of the feature points. By setting a density threshold, the feature points are grouped to obtain initial clustering results. Based on the initial clustering results and a preset distance threshold, the Euclidean distance and point density between point pairs within each cluster are calculated to obtain the core region of each cluster. For the core region of the cluster, the boundary range of the cluster is obtained by calculating the fluctuation period and variation amplitude of the signal intensity within the region. Based on the boundary range of the cluster and the density distribution of adjacent regions, after merging adjacent regions with continuous density, the signal change features within the region are extracted from the time series feature set. By calculating the periodicity and stability of the feature sequence, the regional range of candidate dwell points is obtained.

[0011] Furthermore, a personalized list of safe rest stops is used, combined with rest stop route preferences and device power data. A multi-objective optimization method is employed to rank the list. Based on the ranked recommendation sequence and user tolerance for severe weather, a route recommendation scheme is determined. This includes: modeling user route selection behavior using a long short-term memory neural network based on the safe rest stop recommendation list and historical route data; extracting the frequency of user selection and travel distance on different road segments from historical trajectories to obtain a route preference feature matrix; establishing a route power consumption prediction model using a support vector regression algorithm based on the route preference feature matrix and device power data; extracting the correspondence between route length and power consumption from historical power consumption records to obtain predicted route energy consumption values; calculating the probability of user route adjustment under different weather intensities based on the predicted route energy consumption values ​​and historical selection records; extracting user behavior characteristics for avoiding severe weather to obtain a weather tolerance index; and extracting user evaluation data for different routes from historical rest records based on the weather tolerance index; and obtaining a rest stop rating sequence by calculating the weighted value of rest duration and visit frequency. Based on the stop point rating sequence and route preference feature matrix, a multi-objective optimization algorithm is used to comprehensively weigh route safety, power consumption, and travel distance to obtain a preliminary ranking scheme. According to the preliminary ranking scheme and current weather conditions, the degree of matching between route travel conditions and user weather tolerance is calculated to obtain a route feasibility assessment value. Based on the route feasibility assessment value and the device's remaining battery power, the power threshold required for complete route travel is calculated to obtain a recommended scheme that meets the device's battery life requirements.

[0012] Furthermore, by acquiring users' historical path preference data and device battery information, a clustering algorithm is used to initially screen stop points. Then, combining path preferences and device battery levels, a multi-objective optimization method is employed to sort the stop points in the list, determining the sorted recommended sequence. Environmental noise and light intensity data for each stop point in the sorted recommended sequence are obtained. A preset threshold is used to determine the severe weather index for each stop point, assessing the user's tolerance for severe weather. This includes: calculating the distance characteristics and power consumption ratio of different paths using a support vector machine algorithm based on users' historical path selection records and remaining device battery data; extracting user movement preferences from historical trajectory data to obtain an initial set of stop points. Based on the initial set of stop points and path power consumption data, a multi-objective optimization algorithm is used to comprehensively calculate the safety, distance, and power consumption of stop point paths, resulting in an optimized and sorted stop point sequence. For the optimized and sorted stop point sequence, real-time noise and light intensity data are obtained from an environmental monitoring database, extracting light attenuation rate and noise growth rate to obtain an environmental change characteristic sequence. Based on the environmental change characteristic sequence and preset environmental reference values, the degree of deviation between noise levels and light intensity is calculated to identify the severity range of adverse environments, thus obtaining a weather severity index for the stop point. For the weather severity index of the stop point, the patterns of stay under different weather conditions are extracted from historical user selection data. By calculating the distribution characteristics of selection probability and stay duration, the user's tolerance value for adverse weather is obtained. Based on the user's adverse weather tolerance value and current environmental conditions, the matching degree between path traversal risk and remaining device power is calculated to obtain a recommended plan that meets the user's tolerance. For the recommended plan that meets the user's tolerance, the rate of environmental deterioration is identified by real-time calculation of the changing trends of environmental noise levels and light intensity, thus obtaining an adverse weather index.

[0013] Furthermore, real-time pedestrian density data at rest points is extracted from the route recommendation scheme. By analyzing the movement trajectories of rest points and surrounding landmarks, the evacuation direction of groups in high-risk areas is determined, yielding dynamic adjustment parameters for the population. Based on these dynamic adjustment parameters and the route recommendation scheme, personalized early warning push content is generated and transmitted to user devices, taking into account the signal strength at rest points. This includes: processing real-time pedestrian density data using a long short-term memory neural network based on the rest point coordinates in the final recommendation scheme to extract pedestrian movement trajectories and speed characteristics, resulting in a regional pedestrian distribution map. Based on the regional pedestrian distribution map and landmark location data, a density clustering algorithm is used to calculate the degree of population aggregation within the region. By extracting the flow direction and movement trend of people in high-density areas, the boundary of high-risk areas is obtained. For the high-risk area boundary and geographical topography data, the slope and width characteristics of evacuation routes are calculated, and safe evacuation routes are extracted from the geographic information database to obtain a regional evacuation plan. Based on the regional evacuation plan and campsite level data, the evacuation speed and landmark distribution density are calculated, and group behavior characteristics are extracted from historical evacuation records to obtain dynamic parameters for the population. Based on group dynamic parameters and user location data, a signal transmission range is obtained by calculating a preset signal coverage area and signal strength attenuation curve. According to the signal transmission range and high-risk area data, corresponding warning content is extracted from the warning template library by matching the risk level of the user's location, resulting in tiered warning information. Based on the tiered warning information and real-time signal strength data, a warning push scheme is derived by calculating the signal reception strength and communication status of user equipment.

[0014] This invention provides a system for sending early warning information for emergencies, mainly comprising: a stop point extraction module, used to extract stop point information from user travel data, analyze it in conjunction with user historical travel data, and obtain user stop point type tags; a sensitivity analysis module, used to extract stop point frequency and path preference from user historical travel data, combine it with stop point type tags, calculate the user's sensitivity distribution characteristics to disaster information, determine the prediction threshold for high-risk areas, and if a user is about to enter a high-risk area, track stop point information in real time, determine the proportion of outdoor open spaces, and assess the risk level in conjunction with the high-risk area prediction threshold; a warning level generation module, used to generate a dynamic warning level based on the risk level, combined with the weather and crowd density of the stop point, and extract surrounding landmarks based on the warning level and risk type to generate evacuation guidance content; and an evacuation guidance generation module, used to obtain a personalized safe stop point recommendation list based on the evacuation guidance content, through stop point prediction time window and stop point signal strength analysis, and by integrating the latitude and longitude coordinates of hotels or indoor venues in the surrounding geographic information database. The route recommendation module is used to sort the personalized safe stop point recommendation list by combining stop point route preferences and stop point device power data, and then determine the route recommendation scheme based on the sorted recommendation sequence and the user's tolerance for severe weather. The early warning push module is used to extract real-time stop point pedestrian density data from the route recommendation scheme, analyze the stop point movement trajectory and surrounding landmarks to determine the evacuation direction of groups in high-risk areas, obtain group dynamic adjustment parameters, and generate personalized early warning push content based on the group dynamic adjustment parameters and the route recommendation scheme, which is then transmitted to the user's device based on the stop point signal strength. The technical solution provided by this embodiment of the invention can include the following beneficial effects: This invention discloses a method for sending early warning information for emergencies. The method analyzes travel chain data acquired by user devices, extracts stop point information, and combines this information with GIS data to determine stop point attributes, calculating the distribution characteristics of user sensitivity to disaster information. When a user is about to enter a high-risk area, this invention assesses the risk level in real time, generating dynamic warning levels and evacuation guidance content. Subsequently, based on factors such as user route preferences and device battery level, a multi-objective optimization method is used to recommend personalized safe stop points. Finally, by analyzing the direction of group evacuation, personalized early warning push content is generated and transmitted to the user's device. This invention achieves accurate disaster early warning for outdoor camping scenarios, effectively improving user safety and disaster response capabilities. Attached Figure Description

[0015] Figure 1 This is a flowchart of a method for sending early warning information for emergencies according to the present invention. Detailed Implementation

[0016] The technical solution of the present invention will be clearly and completely described below with reference to embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. The information involved in this disclosure, including but not limited to user data collection, use, and processing, has been authorized by the individual and complies with relevant laws and regulations.

[0017] like Figure 1 This embodiment of a method for sending early warning information for emergencies may specifically include: S101. Obtain travel chain data from user devices, extract the latitude and longitude coordinates and arrival timestamps of the stops, analyze them in conjunction with geographic information system data and user historical travel records, determine the duration and mode of transportation of the stops, judge their indoor and outdoor attributes and work and leisure attributes, and thus generate type tags for user stops.

[0018] S1011. Collect spatiotemporal trajectory data containing latitude and longitude coordinates and timestamps from user devices. Compare and analyze the data by setting a dwell time threshold of 300 seconds and a distance threshold of 50 meters. When the user's location stays continuously within the specified range, it is determined as a dwell point, and a set of dwell points is generated. Then, based on the distance and time difference between adjacent dwell points in the set, combined with the speed range threshold of the transportation mode, the transportation mode is determined. For example, if a user travels from their residence to their company, a distance of 8000 meters, in 25 minutes, at a speed of about 20 kilometers per hour, it is determined to be taking the subway based on historical data. However, if the distance to the restaurant is 300 meters and the speed is 4 kilometers per hour, it is determined to be walking. A transportation connection matrix of dwell points is generated.

[0019] S1012. Based on the traffic connection matrix of the stop point and the user's historical traffic selection data, a deep learning clustering algorithm is used to calculate the location sequence of user activity patterns. For example, it is identified that a user's commuting pattern of leaving home at 7:30 am and arriving at the company at 8:00 am has a probability of 85%, and a leisure pattern of going to the gym at 10:00 am and staying for 90 minutes on weekends is identified. At the same time, the random forest algorithm is used to classify and train the positioning signal strength and the frequency of location changes to determine the indoor and outdoor scene types of the stop point. For example, indoor signal strength is below -85 dB and the location drift is less than 30 meters, while outdoor signal strength is above -75 dB and the location change is continuous. Then, the location attributes and work / leisure attributes of the stop point are determined by matching the geographic information database and the historical tag database.

[0020] S1013. Based on the location attributes of the stop points and historical travel patterns, generate work and leisure attribute labels by calculating the proportion of stay time on weekdays and rest days. For example, if the proportion of stay points in an office building is more than 80% between 9:00 and 18:00 on weekdays, it is labeled as an office place. If the proportion of stay points in a commercial area is high during lunch hours and weekends, it is labeled as a leisure place. Through label classification rules based on spatiotemporal characteristics, the stop points are marked as residential places, workplaces, catering places, or leisure places. For example, catering places within 300 meters of 12:00 on weekdays are marked as dining places, and commercial area stays of more than 90 minutes on weekends are marked as shopping places.

[0021] By extracting travel chain data from user devices and generating stop point type tags through the above steps, it's understandable that the determination and attribute classification methods for stop points can be adjusted according to actual scenarios. For example, threshold settings or algorithm selection can be optimized by technical personnel based on needs. By analyzing user behavior patterns and scenario characteristics, this method lays the foundation for subsequent risk assessment and early warning pushes, ensuring improved accuracy and practicality of information pushes.

[0022] In some alternative embodiments, the identification of stop type can also be combined with more data sources, such as device sensor data or weather information, to further enrich the accuracy of attribute classification, which will not be elaborated here.

[0023] S102. Extract the frequency of stops and route preference features from users' historical travel data, combine the stop type tags, analyze user activity patterns, calculate the distribution characteristics of users' sensitivity to disaster information, and determine the prediction threshold for high-risk areas. At the same time, when users are about to enter high-risk areas, assess the proportion of outdoor open spaces and risk levels by tracking stop information in real time and using geofencing technology.

[0024] S1021. Obtain the stop point sequence and corresponding timestamp information from the user's historical travel data. Model the time series of user activities using a long short-term memory neural network to extract the user's activity frequency and location transfer patterns at different times. For example, taking a user's travel records for 30 consecutive days as an example, record the coordinates and time information of all locations within 24 hours each day. The analysis shows that the probability of the user leaving their residence between 7 and 8 am on weekdays is 90%, and the probability of leaving between 9 and 10 am on weekends is relatively high. This regularity is transformed into spatiotemporal activity data. Subsequently, the spatiotemporal activity data is grouped using a density clustering algorithm to identify the user's high-frequency stop point set, such as five core areas including residence, workplace, and commercial area. Furthermore, extract the user's preferred path sequence from the historical trajectory to reflect their stable path selection tendency.

[0025] S1022. For the set of high-frequency dwell points and user preference path sequences, spatial distribution data of open places are extracted from the geographic information database. Outdoor activity characteristics are calculated by statistically analyzing the frequency and duration of users passing through open places. For example, a user spends 65% of their time in open places such as parks and squares on weekends, while spending as much as 85% of their time in indoor places on weekdays. This differentiated spatial distribution characteristic is used to quantify the user's risk exposure in different scenarios. If the outdoor activity characteristics exceed the preset threshold, the user is combined with real-time location monitoring data and a geofencing algorithm is used to determine whether the user is approaching a high-risk area. For example, when a user is less than 500 meters away from the boundary of a rainstorm flood point and moves at a speed of 4 km / h, it is predicted that they will enter a high-risk area within 10 minutes, triggering the calculation of the risk warning value. The result is a value of 80 points, which reaches the warning threshold.

[0026] S1023. Based on the risk warning value of the user's activity area, obtain the risk level judgment rules from the disaster warning database. By calculating the user's location change probability and historical avoidance behavior records, determine the user's sensitivity characteristics to disaster information. For example, a user has a 75% probability of changing their route after a rainstorm warning, but tends to choose open places away from high-rise buildings during an earthquake warning. After this differentiated response characteristic is quantified, combined with the proportion of open places at the current location, a decision tree algorithm is used to assess the regional risk level. For example, when the proportion of open places in a commercial area is only 20% and the safety facilities are limited, the risk assessment value reaches 90 points, while in open areas such as parks, the assessment value drops to 30 points because there is sufficient evacuation space.

[0027] By modeling user activity patterns using Long Short-Term Memory (LSTM) neural networks and leveraging their long-term dependency characteristics, it is possible to capture the time-series features of user travel, such as the differences in activity between weekdays and weekends, ensuring the accuracy of route preference analysis. Density clustering algorithms, on the other hand, identify high-frequency stop-point sets by calculating the spatial distance and density distribution between stop points. This eliminates the need to pre-specify the number of clusters and is suitable for scenarios with dynamically changing user behavior.

[0028] For calculating the characteristics of outdoor activities, the system not only relies on frequency and duration, but can also be further optimized by combining weather data or equipment sensor information. For example, the outdoor risk weight can be increased in rainy weather to improve the accuracy of the assessment. The details here can be adjusted by technicians according to actual needs, and will not be elaborated here.

[0029] In risk level assessment, the decision tree algorithm generates clear classification rules by analyzing user sensitivity characteristics, open area ratio, and disaster type layer by layer. For example, when the open area ratio is less than 30% and user sensitivity is higher than 70%, it is preferentially marked as a high-risk area. This hierarchical decision-making process improves the interpretability and reliability of the assessment.

[0030] In some alternative embodiments, if there is insufficient historical user data, a collaborative filtering algorithm can be introduced to supplement path preference prediction, or the threshold setting of geofencing can be optimized through real-time crowd density data, thereby enhancing the system's adaptability to new users.

[0031] S103. Adjust the warning level parameters according to the risk level, combine the real-time weather conditions and crowd density data of the stop point to generate a dynamic warning level, and extract surrounding landmark information from the geographic information database based on the warning level and risk type to generate risk avoidance guidance containing guidance content.

[0032] Features are extracted from real-time meteorological monitoring data and historical weather disaster records. Meteorological factors such as rainfall and wind speed are analyzed through deep learning neural networks to calculate the weather disaster risk index. For example, if the rainfall in a certain area reaches 75 mm / hour and the wind speed is 12 m / s, the risk index is 85 when combined with historical water accumulation records. Then, real-time population density values ​​are obtained from monitoring data at rest points, such as 120 people per 100 square meters. The risk level index of the area is determined by a preset area threshold.

[0033] S1031. Based on the regional risk level indicators and the population density value, the algorithm compares the pre-set population density threshold using a support vector regression algorithm. For example, if the safety threshold is set to 80 people / 100 square meters, the algorithm calculates the population saturation of the region. When the population density exceeds the standard by 120%, the algorithm increases the crowding risk weight to 0.7 and adds a weather risk weight of 0.8 to generate a warning level matrix. Then, combined with real-time meteorological indicators such as rainfall intensity increasing to 90 mm / hour and water depth of 15 cm, the algorithm calculates the dynamic warning level as a red warning by using the thresholds in the risk warning rule base, ensuring that the warning level reflects the comprehensive risk of the current environment and population situation.

[0034] S1032. Based on the dynamic warning level and the distribution of people, extract the distribution map of safe havens from the geographic information database. For example, identify 3 safe havens in a certain area. Calculate the shortest path between each safe haven and the point of stay to generate an initial safe haven route. Then, combine real-time road conditions and safe haven capacity data, such as a subway station with a passenger density of 45 people / 100 square meters and a capacity of 800 people, and a shopping mall with only 200 people remaining, to filter out a list of available safe havens. Subsequently, optimize the route through path connectivity analysis. For example, avoid the main road with a water depth of 20 centimeters and choose an alternative route through the second-floor corridor. Although the distance increases by 200 meters, the travel time is shortened to 8 minutes, generating the optimal safe haven route.

[0035] Based on the optimal evacuation route and landmark distribution data, key landmark nodes such as the main entrance of the shopping mall, the entrance of the second-floor connecting corridor, and the B exit of the subway station are extracted. Evacuation guidance content is generated by segmenting the route. The guidance includes the real-world features of the landmarks and the specific location of the commercial corridor on the B1 floor of the subway station. This area is equipped with flood control facilities to facilitate users' quick location and evacuation.

[0036] Deep learning neural networks analyze the spatiotemporal correlation of meteorological data through multi-layer convolution and recurrent structures. For example, by inputting time series data of rainfall and wind speed into the model and training weights using historical disaster records, the model outputs a predicted risk index, ensuring the accuracy and foresight of weather risk assessment. Support vector regression algorithms, on the other hand, construct a high-dimensional hyperplane to fit the non-linear relationship between pedestrian density and risk level. For instance, they automatically adjust weights when pedestrian saturation exceeds the limit, improving the sensitivity of warning levels.

[0037] In some alternative embodiments, if real-time traffic data is insufficient, user feedback data or camera monitoring information can be introduced to optimize path calculation and ensure the feasibility of the avoidance route. The details here can be adjusted by technicians according to the actual scenario and will not be elaborated further.

[0038] By dynamically adjusting warning levels and route planning, the system can update warning content in real time based on changes in weather and pedestrian flow. For example, when rainfall intensity continues to increase, the system can reassess route feasibility and select safer alternative routes to ensure the timeliness and practicality of guidance content. This allows the system to make full use of surrounding resources and effectively improve users' ability to respond to sudden disasters.

[0039] S104. Based on the content of the evacuation guidelines, by analyzing the predicted time window of the stop point and the equipment signal strength data, and combining the coordinates of different locations in the geographic information database, a personalized list of recommended safe stop points is generated.

[0040] Time-series features are extracted from user historical trajectory data and dwell time records. The arrival time and dwell time of user dwell points are calculated through a long short-term memory neural network to generate predicted time interval data. For example, a user has an 85% probability of staying in a business district from 12:00 to 13:00 on a weekday, with an average stay of 45 minutes. This pattern is used to predict future activity locations. Subsequently, combined with device signal strength data, the coordinates of surrounding hotels, shopping malls and other locations are extracted from a geographic information database. A density clustering algorithm is used to identify a set of candidate dwell points to ensure that the recommended locations match the user's location and needs.

[0041] S1041. Based on the candidate stop set and real-time pedestrian flow monitoring data, analyze the building's safety facilities configuration, such as flood control equipment, emergency power supply, and space capacity. For example, a hotel equipped with an emergency escape route and capable of accommodating 600 people has a current pedestrian density of 25 people / 100 square meters, which is below the safety threshold, resulting in a safety index of 90 points. Another hotel, however, only scores 70 points due to insufficient facilities. Combined with traffic data, calculate access time. For example, if it is an 8-minute walk to the preferred hotel with no water accumulation along the route, generate a safe and convenient stop sequence. Further incorporate user historical preferences, such as the user's tendency to choose hotels during severe weather, having 5 similar records in the past 3 months with an average stay of 120 minutes, calculate a personalized recommendation weight of 0.8. Finally, use a support vector machine algorithm to evaluate the site suitability, match the best arrival path, and generate a recommendation list.

[0042] S1042. Based on the continuous signal strength data collected by the user equipment, calculate the mean and standard deviation within a unit time interval of 30 seconds, and generate a standard deviation-to-mean ratio sequence. For example, in an office setting, the mean is -75 dB and the standard deviation is 3 dB, with a ratio of 0.04, which is lower than the fluctuation threshold of 0.1, indicating that the signal is stable. However, when moving, the standard deviation rises to 12 dB, and the ratio exceeds the threshold of 0.14. If the threshold is exceeded, the sudden change moment is identified by the support vector regression algorithm. For example, at 9:05, the signal drops from -65 dB to -82 dB. An initial time window is divided, and then the rate of change distribution is calculated in 180-second windows. The window boundaries are optimized, the signal stable interval is extracted, and the location and time range of the dwell point are determined.

[0043] Long Short-Term Memory (LSTM) neural networks analyze user activity patterns by capturing long-term dependencies in time series. For example, they divide 24 hours into multiple segments and identify a 90% probability of users staying in office buildings between 9 am and 6 pm on weekdays. This predictive ability improves the accuracy of time windows. Density clustering algorithms calculate the spatial density of signal points, such as 25 sampling points per 50 square meters, and select stable signal areas as candidate places to stay. Support Vector Machines (SVMs) optimize recommendation results by classifying location features, such as the openness and facilities of hotels, and calculating a fit of up to 85%.

[0044] S1043. Extract the signal change rate sequence for the adjusted time window, and detect local peaks and stable intervals using a long short-term memory neural network. For example, if the signal in the office area fluctuates between -78 dB and -72 dB for 180 minutes, use a density clustering algorithm to group feature points, set a density threshold, and identify the core area. For example, if the distance between points in the meeting room is less than 10 meters, the density is 3 times that of the edge. Then, by merging continuous density areas, calculate the fluctuation period and amplitude to determine the final stop point area. For example, if the signal in the restaurant during lunchtime shows a U-shaped change, which matches the activity pattern, candidate stop points are generated.

[0045] By combining real-time weather data such as rainfall intensity of 45 mm / hour and humidity of 85%, the suitability of indoor stay is calculated, and hotels with a risk avoidance index of 90 points are recommended. These hotels provide an 8-minute path with full rain protection, making full use of user preferences and facility conditions to improve risk avoidance efficiency and comfort.

[0046] In some alternative embodiments, real-time traffic conditions or user feedback data can be incorporated to further optimize route selection and location recommendations, ensuring that the solution adapts to dynamically changing environments.

[0047] S105. By using a personalized list of safe stops, combined with users' historical route preferences and device battery data, and employing a long short-term memory neural network to model route selection behavior, a comprehensive evaluation of route safety, power consumption, and weather tolerance is conducted to generate a ranked recommendation scheme.

[0048] By extracting path selection patterns from users' historical trajectory data and analyzing time series features through long short-term memory neural networks, a path preference feature matrix is ​​generated. For example, the probability of a user choosing a walking route within 800 meters between 8 and 9 am on weekdays is 85%, while it drops to 45% during lunchtime. After this preference for time and distance is quantified, combined with device power consumption data, a relationship model between path length and power consumption is established through support vector regression algorithm. For example, in a 1000-meter mixed path, the outdoor section consumes 15% of the power and the indoor section consumes 4%, thus obtaining the predicted value of path energy consumption.

[0049] S1051. Based on the predicted energy consumption of the route and historical selection records, calculate the probability of users adjusting their routes under different weather intensities. For example, when the rainfall is 25 mm / hour, the probability of choosing an outdoor route during the weekday morning rush hour is 70%, which drops to 20% on weekends. Extract the weather tolerance index, and then analyze the dwell time and visit frequency from historical dwell records. For example, a coffee shop has a 90% visit rate on rainy days and an average dwell time of 45 minutes. Calculate the dwell point score sequence, and use a multi-objective optimization algorithm to balance safety, power consumption, and travel distance. For example, when the battery is below 30%, the power consumption weight increases to 0.6, and generate a preliminary ranking scheme to ensure that the recommended route takes into account both device battery life and user habits.

[0050] Based on the preliminary sorting scheme and real-time weather conditions, such as rainfall intensity of 35 mm / hour exceeding the user's tolerance threshold of 25 mm / hour, the support vector machine algorithm is used to analyze the path distance and power consumption ratio to filter the initial set of stopping points. For example, paths with a high proportion of indoor locations are prioritized. Then, noise and light data are obtained from the environmental monitoring database. For example, if the light intensity in a certain area decreases from 8000 lux to 2000 lux and the noise increases from 65 dB to 78 dB, the environmental degradation characteristics are calculated, and the priority of indoor paths in the recommended sequence is adjusted.

[0051] S1052. For the optimized and sorted stop point sequence, a weather severity index is generated by calculating the light attenuation rate and noise growth rate. For example, when the light intensity is below 3000 lux and the noise level exceeds 75 decibels, it is judged as a severe environment. Combined with the user's historical behavior, such as when the probability of choosing an indoor route is 90% when the light intensity is below 2500 lux, a tolerance value is obtained. Then, based on the device's remaining battery power of 35% and the route's power consumption requirements, for example, routes with power consumption exceeding 15% are excluded. Finally, an 800-meter composite route is recommended, with 80% of the route being indoors and charging facilities along the way, to meet the needs of disaster avoidance and battery life.

[0052] Long Short-Term Memory (LSTM) neural networks capture the temporal dependencies of path selection, ensuring that the preference matrix reflects users' dynamic habits. Support Vector Regression (SVR) algorithms improve prediction accuracy by fitting the nonlinear relationship between power consumption and distance. Multi-objective optimization algorithms dynamically adjust weights, for example, increasing the safety weight to 0.5 when the weather worsens, making the recommendation scheme closer to the real-time scenario.

[0053] In some alternative embodiments, real-time traffic or environmental sensor data can be incorporated, such as detecting water depth or wind speed, to further optimize route selection and improve the adaptability of the solution.

[0054] By comprehensively analyzing user behavior, device status, and environmental changes, such as when the light intensity decreases by 500 lux per minute or the noise increases by 2 dB per minute, the path is adjusted in a timely manner to ensure that the recommended solution is both safe and efficient in severe weather and fully meets the personalized needs of users.

[0055] S106. Extract real-time crowd density data from the final recommended plan, analyze crowd movement trajectory and speed characteristics through long short-term memory neural network, combine with geographical terrain data to determine the evacuation direction of the group in high-risk areas, generate dynamic adjustment parameters for the group, and based on this and signal strength analysis, generate and transmit personalized early warning push content to user devices.

[0056] Real-time pedestrian density data is obtained from the stop locations in the final recommended plan. The time series is processed using a long short-term memory neural network to extract the characteristics of pedestrian movement trajectory and speed change. For example, before rainfall, the pedestrian density in a camping area increased from 15 people per 100 square meters to 45 people, and the movement speed decreased from 1.2 m / s to 0.4 m / s. A pedestrian distribution map of the area is generated. Then, the degree of aggregation is analyzed by density clustering algorithm to identify the high-risk area boundary where the density at the main entrance and exit reaches 0.8 people per square meter, which exceeds the safety threshold of 0.5 people per square meter, providing a basis for the evacuation plan.

[0057] S1061. Based on the terrain data in the geographic information database and the boundaries of high-risk areas, calculate the slope and width of evacuation routes. For example, if the main route has a slope of more than 15 degrees and a width of only 2 meters, it is not suitable for large-scale evacuation. The alternative mountain trail with a gentle slope and a width of 4 meters is better. By analyzing historical evacuation records, extract group behavior characteristics, such as campers' average reaction time of 90 seconds and their tendency to follow the crowd, generate group dynamic parameters. Combine this with landmarks around the resting point and the direction of crowd flow, such as 75% of the crowd moving in one direction, optimize the regional evacuation plan and ensure that the route selection meets the actual needs.

[0058] Based on the regional evacuation plan and signal coverage area, the signal strength attenuation curve is calculated. For example, under rainy conditions, the signal coverage in the camping area is reduced from 800 meters to 500 meters, and the intensity in the valley area is below -95 decibels. Combined with the reception strength of user equipment, content is matched from the warning template library. For example, users in the valley receive a water accumulation warning, and users on the hillside receive a mudslide warning. Tiered warning information is generated, and the push power and mode are dynamically adjusted. For example, when the signal is below -100 decibels, SMS push is switched to ensure that the information is delivered.

[0059] S1062. For graded early warning information and real-time signal data, analyze the communication status of user equipment. For example, when the signal-to-noise ratio is below 10 dB, increase the transmission power and push repeatedly. Combine group dynamic parameters and user location to customize personalized early warning content. For example, indicate congestion at the main exit and recommend mountain trails. Make full use of landmark distribution and behavioral characteristics to improve the pertinence and effectiveness of early warning.

[0060] Long Short-Term Memory (LSTM) neural networks accurately predict density change trends by capturing the time dependence of crowd movement, while density clustering algorithms automatically identify high-risk areas by setting density thresholds, eliminating the need for manual pre-setting of cluster numbers and adapting to dynamic scenarios. Signal strength analysis optimizes transmission strategies through attenuation models to ensure communication reliability in harsh environments.

[0061] In some alternative embodiments, wind speed or visibility data can be incorporated to further refine pedestrian flow prediction, or signal transmission can be optimized through feedback from device sensors.

[0062] By tracking pedestrian flow and signal status in real time, such as when the speed of the crowd drops to 60% of that during the daytime during nighttime rainfall, the evacuation routes and delivery methods are dynamically adjusted. This not only improves the accuracy of early warnings but also ensures the effective transmission of information in complex terrain, significantly enhancing users' ability to avoid risks.

[0063] This invention provides a system for sending early warning information for emergencies, mainly comprising: a stop point extraction module, used to extract stop point information from user travel data, analyze it in conjunction with user historical travel data, and obtain user stop point type tags; a sensitivity analysis module, used to extract stop point frequency and path preference from user historical travel data, combine it with stop point type tags, calculate the user's sensitivity distribution characteristics to disaster information, determine the prediction threshold for high-risk areas, and if a user is about to enter a high-risk area, track stop point information in real time, determine the proportion of outdoor open spaces, and assess the risk level in conjunction with the high-risk area prediction threshold; a warning level generation module, used to generate a dynamic warning level based on the risk level, combined with the weather and crowd density of the stop point, and extract surrounding landmarks based on the warning level and risk type to generate evacuation guidance content; and an evacuation guidance generation module, used to obtain a personalized safe stop point recommendation list based on the evacuation guidance content, through stop point prediction time window and stop point signal strength analysis, and by integrating the latitude and longitude coordinates of hotels or indoor venues in the surrounding geographic information database. The route recommendation module is used to sort the personalized safe stop list by combining stop route preferences and device power data, using a multi-objective optimization method. Based on the sorted recommendation sequence and the user's tolerance for severe weather, a route recommendation scheme is determined. The early warning push module extracts real-time pedestrian density data from the route recommendation scheme, analyzes the movement trajectory of the stop points and surrounding landmarks to determine the evacuation direction of groups in high-risk areas, obtains dynamic adjustment parameters for the population, and generates personalized early warning push content based on the dynamic adjustment parameters and the route recommendation scheme. This content is then transmitted to the user's device based on the signal strength of the stop points. The above embodiments are merely one preferred embodiment of the present invention and should not be used to limit the scope of protection of the present invention. Any modifications or refinements made to the main design concept and spirit of the present invention without substantial meaning, but which still solve the same technical problem as the present invention, should be included within the scope of protection of the present invention.

Claims

1. A method for sending early warning information for emergencies, characterized in that, The method includes: The system extracts stop point information from user travel data and analyzes it in conjunction with historical user travel data to obtain stop point type tags. It also extracts stop point frequency and route preferences from historical user travel data, combines these with stop point type tags to calculate the distribution characteristics of user sensitivity to disaster information, and determines the prediction threshold for high-risk areas. If a user is about to enter a high-risk area, it tracks stop point information in real time, determines the proportion of open outdoor spaces, and assesses the risk level based on the high-risk area prediction threshold. Based on the risk level, combined with weather and pedestrian density at the stop point, it generates a dynamic warning level. Based on the warning level and risk type, it extracts surrounding landmarks to generate evacuation guidance content. Based on the evacuation guidance content, it analyzes the stop point prediction time window and signal strength, and integrates the latitude and longitude coordinates of stop points from the surrounding geographic information database to obtain a personalized safe stop point recommendation list. Using this personalized safe stop point recommendation list, combined with stop point route preferences and device battery data, it employs a multi-objective optimization method to sort the recommendation list. Based on the sorted recommendation sequence and the user's tolerance for severe weather, it determines a route recommendation scheme. Real-time pedestrian density data at rest points is extracted from the route recommendation scheme. By analyzing the movement trajectory at rest points and surrounding landmarks, the evacuation direction of the population in high-risk areas is determined, and dynamic adjustment parameters for the population are obtained. Based on the dynamic adjustment parameters and the route recommendation scheme, personalized early warning push content is generated and transmitted to user devices in combination with the signal strength at rest points.

2. The method according to claim 1, characterized in that, The process involves extracting stop point information from user travel data, analyzing it in conjunction with historical user travel data, and obtaining user stop point type tags, including: Spatiotemporal trajectory data is acquired from the user device. Based on the latitude and longitude coordinates and timestamp data in the spatiotemporal trajectory data, a set of dwell points is obtained by comparing and analyzing dwell time thresholds and distance thresholds. For the set of stopping points, a traffic connection matrix between stopping points is obtained by judging the speed range threshold of traffic modes based on the distance data and time difference data between adjacent stopping points. Based on the traffic connection matrix and historical traffic selection data, a deep learning clustering algorithm is used to calculate the location sequence of user activity pattern points; For the location sequence of the activity pattern points, the random forest algorithm is used to classify and train the location signal strength data and location change frequency data to obtain the indoor and outdoor scene or work and leisure type identification of the stop point.

3. The method according to claim 1, characterized in that, The frequency of stops and route preferences are extracted from users' historical travel data. Combined with the type of stop labels, the distribution characteristics of users' sensitivity to disaster information are calculated to determine the prediction threshold for high-risk areas. If a user is about to enter a high-risk area, the system tracks their point of stay in real time, determines the percentage of open outdoor spaces, and assesses the risk level based on a high-risk area prediction threshold, including: Based on the stop point sequence and timestamp information in the user's historical travel data, the spatiotemporal feature data of the user's activities are calculated using a long short-term memory neural network. Based on the spatiotemporal feature data of the user activities, a set of high-frequency dwell points for the user is calculated using a density clustering algorithm. This set of high-frequency dwell points is used to extract the user's preferred path sequence. For the user's preferred path sequence, spatial distribution data of open spaces are extracted from the geographic information database, and outdoor activity characteristics are obtained by calculating the frequency and duration of users passing through open spaces. If the outdoor activity characteristics trigger the geofence warning threshold, the regional risk level is calculated using a decision tree algorithm based on the user's location relocation probability and disaster information sensitivity characteristics.

4. The method according to claim 1, characterized in that, The system generates dynamic warning levels based on risk levels, combined with weather and crowd density at the point of stay; and extracts landmarks around the point of stay based on the warning level and risk type to generate evacuation guidance content, including: By processing real-time meteorological monitoring data and historical weather disaster records through deep learning neural networks, a weather disaster risk index is obtained. Real-time population density values ​​are extracted based on monitoring data of rest points, and regional risk level indicators are obtained by using preset regional thresholds. Based on the regional risk level index and the population density value, a warning level matrix is ​​obtained by processing the preset population density threshold through the support vector regression algorithm; Based on the aforementioned warning level matrix and real-time meteorological indicators, the regional warning level is calculated using the thresholds in the risk warning rule base to obtain dynamic warning data; Based on the dynamic early warning data and the distribution of people, a distribution map of safe havens is extracted from the geographic information database, and the safe haven route is obtained by calculating the shortest path between the safe haven and the point of stay. Based on evacuation routes and landmark distribution data, generate evacuation guidance content that includes landmark guidance information.

5. The method according to claim 1, characterized in that, Based on the risk avoidance guidelines, and through analysis of predicted time windows and signal strength at designated stops, the latitude and longitude coordinates of locations in the surrounding geographic information database are integrated to obtain a personalized list of recommended safe stops, including: Receive user historical trajectory data and dwell time records, and calculate the dwell time prediction interval data through a long short-term memory neural network; Based on the predicted time interval data of the stop points and the equipment signal strength data, the coordinates of the surrounding locations are extracted from the geographic information database, and a set of candidate stop points is obtained through a density clustering algorithm. Based on the candidate stop point set and real-time pedestrian flow monitoring data, a stop point list with a risk avoidance index is obtained by calculating the configuration of building safety facilities and space capacity; Based on the list of rest stops with risk aversion index and the personalized venue recommendation weights, the support vector machine algorithm is used to calculate the venue suitability and obtain a personalized safe rest stop recommendation list.

6. The method according to claim 5, characterized in that, Also includes: Based on the stop point data, a predicted data set is obtained using a preset time window division method. Signal strength information is extracted from the data set, and the regional range of candidate stop points is determined by calculating the intensity change rate. By analyzing the fluctuation amplitude of signal strength within the region and the length of the time window, it is determined whether the region meets the stop point conditions. If it does, the stop point location is determined, and the corresponding time range is obtained to determine the stop point location and time range.

7. The method according to claim 5 or 6, characterized in that, Also includes: Based on the real-time signal strength data collected by the user equipment, the ratio of the standard deviation to the mean of the signal strength value per unit time is calculated. If the ratio of the standard deviation to the mean is greater than a preset threshold, the sliding window method is used to initially segment the time series to obtain multiple initial time window sequences. By analyzing the distribution characteristics of the signal change rate within each initial time window, the boundary point positions of the time window are determined, resulting in an adjusted time window. The time series features of the signal change rate are extracted from the adjusted time window. By calculating the peak value and the duration of the stationary period of the change rate, a preliminary feature set of candidate dwell points is obtained. The preliminary feature set is grouped using a density clustering algorithm. By calculating the distance and density distribution between feature points, the regional range of candidate dwell points is determined.

8. The method according to claim 1, characterized in that, The process involves using a personalized list of safe rest stops, combined with rest stop route preferences and device battery data, to sort the list using a multi-objective optimization method. Based on the sorted recommendation sequence and the user's tolerance for severe weather, a route recommendation scheme is determined, including: By modeling user path selection behavior using a long short-term memory neural network, a path preference feature matrix is ​​obtained. The path preference feature matrix is ​​analyzed using the support vector regression algorithm. The relationship between path length and power consumption is extracted from historical power consumption records to obtain the predicted path energy consumption value. Based on the predicted energy consumption of the route, the user's tolerance for severe weather is obtained by calculating the probability of route adjustment under different weather intensities. Based on the user's tolerance for severe weather and the path preference feature matrix, a multi-objective optimization algorithm is used to calculate the trade-offs between path safety, power consumption, and travel distance to obtain a recommended path.

9. The method according to claim 1, characterized in that, The process involves extracting real-time pedestrian density data at rest points from the route recommendation scheme, analyzing the movement trajectories at rest points and surrounding landmarks to determine the evacuation direction of groups in high-risk areas, and obtaining dynamic adjustment parameters for the population. Based on these dynamic adjustment parameters and the route recommendation scheme, personalized early warning push content is generated and transmitted to user devices in conjunction with the signal strength at rest points, including: Receive real-time pedestrian density data, extract pedestrian movement trajectory and speed features through a long short-term memory neural network, and obtain a regional pedestrian distribution map; Based on the regional pedestrian flow distribution map and landmark location data, a density clustering algorithm is used to calculate the degree of population concentration in the region. By extracting the population flow direction data of high-density areas, the boundary of high-risk areas is obtained. Based on the boundary and geographical topography data of the high-risk area, the slope and width characteristics of the escape routes are calculated, and safe evacuation routes are extracted from the geographic information database to obtain the regional evacuation plan. Based on the regional evacuation plan and the preset signal coverage area, the corresponding warning content is extracted from the warning template library by calculating the signal strength attenuation curve and the signal reception strength of the user equipment, and a warning push plan is obtained.

10. A system for sending early warning information for emergencies, characterized in that, The system includes: a stop point extraction module, used to extract stop point information from user travel data, analyze it in conjunction with user historical travel data, and obtain user stop point type tags; a sensitivity analysis module, used to extract stop point frequency and route preference from user historical travel data, combine it with stop point type tags, calculate the user's sensitivity distribution characteristics to disaster information, determine the prediction threshold for high-risk areas, and if a user is about to enter a high-risk area, track stop point information in real time, determine the proportion of outdoor open spaces, and assess the risk level in conjunction with the high-risk area prediction threshold; a warning level generation module, used to generate dynamic warning levels based on the risk level, combined with the weather and crowd density of the stop point, extract surrounding landmarks based on the warning level and risk type, and generate evacuation guidance content; and an evacuation guidance generation module, used to obtain a personalized safe stop point recommendation list based on the evacuation guidance content, through stop point prediction time windows and stop point signal strength analysis, and by integrating the latitude and longitude coordinates of stop points from the surrounding geographic information database. The route recommendation module is used to sort the personalized safe stop list by combining stop route preferences and device battery data, using a multi-objective optimization method. Based on the sorted recommendation sequence and the user's tolerance for severe weather, a route recommendation scheme is determined. The early warning push module is used to extract real-time pedestrian density data at stop points from the route recommendation scheme. By analyzing the movement trajectory of the stop points and surrounding landmarks, it determines the evacuation direction of the group in high-risk areas, obtains the group dynamic adjustment parameters, and generates personalized early warning push content based on the group dynamic adjustment parameters and the route recommendation scheme. This content is then transmitted to the user's device based on the signal strength at the stop point.

Images