Marine disaster risk prevention and control early warning system and method based on big data

Abstract

This invention discloses a marine disaster risk prevention and early warning system and method based on big data, belonging to the field of marine disaster monitoring and early warning technology. The system and method include a multi-source data fusion module, an intelligent modeling module, and a real-time early warning module. The method cleans and stores multi-source data after wavelet transform downsampling through edge computing nodes; it extracts spatiotemporal features using a spatiotemporal Transformer network, combines the Coriolis force correction equation to calibrate predicted values, and generates the disaster risk probability and key parameters such as tide level and wind speed for the next 24 hours; it triggers graded early warnings based on a dynamic risk map, and pushes them through multiple channels such as mobile networks and emergency broadcasts within seconds. This invention reduces the tide level prediction error to ≤0.2 meters and the early warning response delay to <10 seconds, effectively solving the problems of high false alarm rate and delayed updates in traditional systems.

Description

Technical Field

[0001] This invention belongs to the field of marine disaster monitoring and early warning technology, specifically relating to a marine disaster risk prevention and early warning system and method based on big data. Background Technology

[0002] Marine disasters, such as storm surges, tsunamis, and giant waves, pose serious threats to the safety of coastal cities, marine economic activities, and the ecological environment. With the widespread adoption of technologies such as satellite remote sensing and sensor networks, marine monitoring data is experiencing explosive growth. How to utilize big data technology to improve the accuracy of disaster prediction and response efficiency has become a key research focus in the international field of marine disaster prevention and mitigation. Traditional marine disaster early warning systems suffer from the following bottlenecks: inefficient data processing; heterogeneous formats of multi-source monitoring data and high-frequency sampling lead to heavy transmission and storage pressures; traditional filtering algorithms struggle to effectively remove noise, resulting in data utilization rates of less than 10%; limitations in prediction models; purely physical models rely on idealized assumptions and cannot capture complex spatiotemporal correlations; purely data models lack physical constraints, leading to low reliability of prediction results, with tide level prediction errors often exceeding 0.5 meters; crude early warning responses; limited risk visualization methods; fixed early warning grading strategies that cannot dynamically match risk levels in different areas, resulting in high information delivery delays; and lagging model updates; post-disaster data is not effectively used for model optimization, requiring manual intervention and retraining, making it difficult to adapt to dynamic changes in the marine environment.

[0003] Therefore, there is an urgent need for a marine disaster risk prevention and early warning system and method that integrates big data processing, intelligent modeling and dynamic response to solve the core problems of traditional systems in terms of data efficiency, prediction accuracy, response performance and model adaptability. Summary of the Invention

[0004] To address the aforementioned pain points, this invention provides a big data-based marine disaster risk prevention and early warning system and method. Through multi-source data fusion, spatiotemporal Transformer and physical formula hybrid modeling, dynamic risk maps and graded early warning, and a closed-loop training mechanism driven by post-disaster data, it achieves accurate prediction and efficient response to marine disaster risks.

[0005] The present invention is as follows:

[0006] A big data-based marine disaster risk prevention and early warning system is characterized by including a multi-source data fusion module, an intelligent modeling module, and a real-time early warning module.

[0007] The multi-source data fusion module is used to access real-time monitoring data from multiple sources such as satellite remote sensing and buoy monitoring, as well as historical disaster data. After cleaning and feature engineering processing, the data is stored in a distributed storage system.

[0008] The intelligent modeling module constructs a hybrid prediction model based on spatiotemporal sequence data. The hybrid prediction model combines a spatiotemporal Transformer network with a physical correction formula to predict disaster risk probability and key parameters. Furthermore, the module has a closed-loop training mechanism that can automatically trigger incremental updates of model parameters based on post-disaster assessment data to continuously optimize prediction accuracy.

[0009] The real-time early warning module generates a dynamic risk map and graded early warning information based on the prediction results, and pushes them to the terminal device.

[0010] Preferably, the multi-source data fusion module includes an edge computing node. The edge computing node acquires the 50Hz high-frequency sampled real-time monitoring raw data output from the device, and performs multi-resolution decomposition on the raw data through a wavelet transform algorithm: it performs anomaly filtering on the high-frequency components, removes noise and outliers exceeding a preset physical threshold, retains low-frequency trend characteristics, and verifies the rationality of the low-frequency components by combining historical data statistical patterns; finally, it generates a minute-level downsampling data sequence that meets the needs of marine disaster early warning.

[0011] Preferably, the distributed storage system uses the HBase engine to store time-series monitoring data such as wave height, wind speed, and tide level, and the Elasticsearch engine to store disaster report text, sensor logs, and satellite remote sensing image metadata. Efficient data management is achieved through the following technologies: a unified API gateway: providing standardized data access interfaces and automatically routing query requests to the corresponding storage engine, supporting cross-database join queries based on time and region dimensions; index optimization strategy: establishing a time-plus-region composite index in HBase and constructing a text inverted index and a geospatial index in Elasticsearch to shorten single-database query response time; and a Redis caching mechanism: implementing a caching strategy for frequently accessed data from the past 7 days, reducing the response time for repeated queries from an average of 200ms to ≤20ms. Through the synergy of these technologies, the cross-database query latency of the storage system is ≤500ms, meeting the real-time data retrieval requirements for marine disaster early warning.

[0012] Preferably, the hybrid prediction model combines a spatiotemporal Transformer network with a physical correction formula. The spatiotemporal Transformer network encodes marine disaster-related features such as wind speed, air pressure, and tide level into a spatiotemporal matrix with a 10-minute time step. It captures cross-feature correlations through a multi-head attention mechanism and outputs initial disaster risk probability and time-series feature vectors. The physical correction formula is based on the fundamental equations of tidal theory and uses Coriolis force correction terms and seabed friction parameter correction terms to calibrate the initial disaster risk probability and key parameters such as tide level output by the spatiotemporal Transformer network. This forms a disaster risk probability prediction value and prediction results for key parameters such as tide level, wind speed, and wave height that include physical mechanism constraints, enabling the prediction of disaster risk probability and key parameters for the next 24 hours.

[0013] Preferably, the physical correction formula is based on the fundamental equations of tidal theory:

[0014]

[0015] In the formula, This represents the storm surge increase (unit: meters), a key parameter that needs to be calibrated. This is a correction term for seabed friction parameters, where These represent the east-west and north-south components of the horizontal flow velocity, respectively, in meters per second. The coefficient of friction is related to the roughness of the seabed. This is a Coriolis force correction term. The Coriolis parameter takes a value of approximately It is related to geographical latitude; The velocity vector of the seawater flow. The curl of the velocity vector is used to correct for the influence of the Coriolis force caused by the Earth's rotation. The tide prediction results output by the spatiotemporal Transformer network are calibrated by the above correction term so that the tide prediction error is ≤0.2 meters.

[0016] Preferably, the real-time early warning module is based on Leaflet.js to develop a dynamic risk map, which supports the overlay of coastal population density and important facility location data, and automatically identifies high-risk areas; the graded early warning includes four levels: blue, yellow, orange and red, and the red warning requires pre-verification and secondary manual confirmation, with an early warning response delay of less than 10 seconds.

[0017] Preferably, the closed-loop training mechanism includes: automatically synchronizing disaster assessment data from the distributed storage system after a disaster, wherein the assessment data includes real-time post-disaster monitoring data processed by the multi-source data fusion module and data of the same type of disaster in the historical case library, wherein the real-time monitoring data includes measured values ​​of wind speed, wave height, and tide level; when the effective feature dimensions of the assessment data exceed 95% without missing values ​​and the number of available samples reaches 15% of the total training data, an incremental update of the parameters of the hybrid prediction model is triggered, and the update process performs gradient optimization based on the current parameters of the model.

[0018] Preferably, the real-time early warning module pushes early warning information through channels including mobile communication networks, internet platforms, emergency broadcasting systems, and smart terminal devices. Through a preset hierarchical strategy configuration table, blue warnings are pushed to monitoring personnel via SMS once a day, yellow warnings are accompanied by WeChat public account graphic notifications, orange warnings trigger hourly emergency broadcasts, and red warnings simultaneously trigger strong reminders on smart terminal apps and push notifications to emergency departments and coastal community leaders every 10 minutes, achieving differentiated delivery of different levels of early warnings.

[0019] Preferably, the multi-source data fusion module, intelligent modeling module, and real-time early warning module are deployed using a unified technical architecture, achieving automated connection through the following methods: standardized data interface: data is transmitted between modules through predefined data formats and interaction protocols; task scheduling mechanism: task chains are automatically arranged based on the workflow engine in the order of data processing, model training, and early warning generation; cross-module linkage: early warning results trigger incremental model training requests, forming a closed-loop feedback.

[0020] A big data-based method for marine disaster risk prevention and early warning, the implementation steps of which are as follows:

[0021] S1. Multi-source data fusion processing: Integrate real-time monitoring data from multiple sources such as satellite remote sensing and buoy monitoring, as well as historical disaster data. Through edge computing nodes, use wavelet transform algorithm to perform minute-level downsampling and anomaly filtering on 50Hz high-frequency data, and perform data cleaning, repair missing values, remove outliers, and perform feature engineering to generate time-series data sequences that meet early warning requirements.

[0022] S2. Hybrid Model Prediction Calculation: The processed data is input into the hybrid prediction model. The hybrid prediction model encodes marine disaster-related features such as wind speed, air pressure, and tide level into a spatiotemporal matrix using a spatiotemporal Transformer network with a 10-minute time step. It extracts spatiotemporal correlation features using a multi-head attention mechanism and outputs the initial value of disaster risk probability and time series feature vector. Based on the basic equations of tidal theory, the prediction results are physically calibrated using Coriolis force correction terms and seabed friction parameter correction terms to generate predicted values ​​of disaster risk probability and key parameters such as tide level, wind speed, and wave height for the next 24 hours.

[0023] S3. Dynamic Early Warning and Information Push: Based on the prediction results, a dynamic risk map developed using Leaflet.js is overlaid with coastal population density and important facility location data to automatically identify high-risk areas; graded early warnings are triggered according to four preset thresholds: blue, yellow, orange, and red. Red warnings require pre-verification and secondary manual confirmation. Early warning information is pushed through multiple channels such as mobile communication networks and emergency broadcasts to ensure a response delay of less than 10 seconds.

[0024] S4. Model closed-loop training mechanism: After a disaster, disaster assessment data is automatically synchronized from the distributed storage system. When the effective feature dimension of the assessment data exceeds 95% without missing data and the number of available samples reaches 15% of the total training data, the parameters of the hybrid prediction model are incrementally updated. The update process is based on gradient optimization performed on the current parameters of the model.

[0025] Compared with the prior art, the advantages of the present invention are as follows:

[0026] (1) In this invention, edge computing nodes are combined with wavelet transform algorithm to downsample 50Hz high-frequency data into minute-level sequences, reducing 90% of invalid data transmission and reducing network load and storage costs; a distributed storage system is used to optimize cross-database query latency to ≤500ms through multi-engine collaboration and index optimization, and the response time for querying high-frequency data in the past 7 days is reduced from 200ms to ≤20ms, meeting the data call requirements for real-time early warning.

[0027] (2) In this invention, a spatiotemporal Transformer network is used to extract cross-feature spatiotemporal correlations through a multi-head attention mechanism and output initial risk probability values. Combined with the Coriolis force and seabed friction correction term of the tidal theory equation, the tide level prediction error is controlled to ≤0.2 meters, which is 60% more accurate than the traditional pure data model. A closed-loop training mechanism is used to automatically trigger incremental updates of the model using post-disaster assessment data, without the need for full training, so that the model prediction accuracy continues to improve with the accumulation of disaster cases and adapts to the dynamic changes of the marine environment.

[0028] (3) In this invention, a dynamic risk map is used to support real-time overlay of population density and important facility data, and to automatically identify high-risk areas, providing visual support for emergency decision-making; a graded early warning strategy is adopted to reach people through multiple channels in a differentiated manner, with a red warning response delay of less than 10 seconds, which is 98% better than the traditional system, ensuring that risk information of different levels is accurately and timely transmitted. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the structure of a big data-based marine disaster risk prevention and early warning system.

[0030] Figure 2 This is a flowchart illustrating a big data-based method for marine disaster risk prevention and early warning. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be explained and described below. However, the following embodiments are only preferred embodiments of the present invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments in the implementation methods without creative effort are all within the protection scope of the present invention.

[0032] Example 1

[0033] I. Deployment Environment and Hardware Configuration

[0034] 1. Scenario Location: A marine disaster monitoring center in a coastal province, covering an area of ​​approximately 20,000 square kilometers of sea, deploying 50 buoy monitoring stations and 3 shore-based radar stations, and connecting to the remote sensing data interface of the satellite marine application center.

[0035] 2. Hardware facilities: 2-1. Edge computing nodes: Deploy NVIDIA Jetson Xavier NX edge computing boxes at each buoy station, equipped with ARM architecture processors, supporting 50Hz real-time data processing, and built-in wavelet transform algorithm firmware;

[0036] 2-2. Distributed storage cluster:

[0037] Time-series data cluster: 10 HBase servers: configured with Intel Xeon Silver 4310, 256GB memory, and SSD storage; stored in tables partitioned by "time (day) + region (1°×1° latitude and longitude grid)";

[0038] Unstructured data cluster: 5 Elasticsearch servers (configured as above) to store satellite remote sensing metadata, such as Sentinel-1 SAR image metadata, about 10GB per day;

[0039] Cache layer: 3 Redis servers, 16GB memory, caching high-frequency access tide level and wind speed data for the past 7 days;

[0040] 2-3. Model training server: 2 GPU servers, NVIDIA A100×4, 2TB of video memory, used for spatiotemporal Transformer network training and physics correction calculation.

[0041] 3. Software Stack:

[0042] 3-1. Edge computing: Wavelet transform algorithm implemented in C++ and integrated into JetsonOS system;

[0043] 3-2. Data storage: HBase 2.4.16 + Elasticsearch 7.17.9, provided through a unified API gateway, developed based on Spring Boot, providing access interfaces;

[0044] 3-3. Model Development: Implement the spacetime Transformer using PyTorch 2.1, and use Python scientific computing libraries: NumPy and SciPy to implement the physical correction formula;

[0045] 3-4. Early warning system: The front end is developed based on Leaflet.js 1.9.4, and the back end is built with Go language to create microservices, connecting to SMS gateway and emergency broadcast protocol.

[0046] II. Multi-source data fusion module

[0047] 1. Data Access

[0048] 1-1. Buoy monitoring data: Data from the SBE37-SI temperature, salinity and depth meter is collected via RS-485 interface at a sampling frequency of 50Hz. The data includes temperature, salinity, current velocity (u / v components), and tide level. After processing by edge computing nodes, the data is transmitted to the central terminal via 4G network.

[0049] 1-2. Satellite remote sensing data: Daily scheduled retrieval of HY-2B satellite wave data from the Satellite Ocean Application Center, with a spatial resolution of 10km and a temporal resolution of 6 hours, including significant wave height, wave direction, and wind speed;

[0050] 1-3. Historical data: Synchronize typhoon path data for the past 10 years from the provincial disaster recovery database: in CSV format, including time, center latitude and longitude, wind speed level; and historical storm surge reports.

[0051] 2. Edge Computing: Wavelet Transform: The 50Hz flow velocity data is decomposed into two levels using the db8 wavelet. High-frequency components > 1Hz are filtered with a threshold; for example, the absolute value of the flow velocity > 12m / s is considered noise. Low-frequency components ≤ 1Hz retain trend characteristics. Combined with historical data statistics, if the normal flow velocity in the sea area is ≤ 8m / s, outliers are removed. Finally, 1-minute interval data is generated, reducing the data volume from 288,000 records / day to 3,456 records / day, with a compression rate of 98.8%.

[0052] 3. Storage Verification: Cross-database query execution: Retrieve tidal data (HBase) for the region "August 1, 2024, 00:00-24:00, 29°-30° North latitude, 121°-122° East longitude" and correlate it with the disaster report for the same period (Elasticsearch). A JMeter load test of 100 queries showed an average latency of 420ms, meeting the ≤500ms requirement. Redis cache hit rate: Redis monitoring tools showed a 92% data access hit rate over the past 7 days, with repeated query response time consistently around 18ms.

[0053] III. Intelligent Modeling Module

[0054] 1. Hybrid model construction:

[0055] 1-1. Spatiotemporal Transformer Network Structure:

[0056] Input layer: Encodes wind speed (m / s), air pressure (hPa), tide level (m), and u / v current velocity (m / s) in 10-minute steps into a spatiotemporal matrix: 144 steps in time dimension and a 3×3 grid centered on the buoy station in space dimension, with a total of 9 nodes;

[0057] Encoder: 6 Transformer layers, 8 attention heads per layer, 512 hidden layer dimensions, capturing cross-feature correlations, such as the lag correlation between wind speed and wave height: historical statistics show that for every 1 m / s increase in wind speed, the wave height increases by 0.2 m after 2 hours;

[0058] Output layer: The fully connected layer outputs the risk probability and the time series feature vector, which is the trend of tide changes in the next 24 hours.

[0059] 2. Physical correction implementation:

[0060] Implement tidal theory equations in Python, targeting a coastal province's sea area (typical latitude approximately 30°N, Coriolis parameter C=). Using PyTorch's automatic differentiation function, the correction term is combined with the Transformer output to form an end-to-end trainable model;

[0061] Calibration Case: During a typhoon on September 10, 2024, Transformer predicted a tide level of 3.2m, which was corrected to 3.45m. The actual measured value was 3.52m, with an error of 0.07m (<0.2m).

[0062] 3. Closed-loop training verification:

[0063] Post-disaster synchronization data: After the typhoon, 320 measured data points were automatically acquired (tide level, wind speed, and wave height with 100% completeness), accounting for 17% of the training set (total training set of 1880 data points), triggering incremental updates;

[0064] Training process: The AdamW optimizer was used with a learning rate of 1e-4 and 50 iterations, taking 1 hour and 45 minutes. After the update, the model's accuracy in predicting the risk probability of similar storm surges increased from 85% to 88%.

[0065] IV. Real-time Early Warning Module

[0066] 1. Dynamic Risk Map: Related Overlay Data:

[0067] Population density: Raster data on the coastal population distribution of a coastal province in 2023: resolution 100m, unit: people / hectare;

[0068] Key facilities: Includes shapefile vector data for 3 nuclear power plants and 5 ports;

[0069] High-risk identification: The storm surge risk heat map is generated by Leaflet's Heatmap plugin. When the predicted inundation depth is >1m, the inundation area is marked with a red polygon and automatically associated with the population in the area: a certain area is marked "Affected population: 5000 people".

[0070] 2. Tiered early warning response:

[0071] 2-1. Blue Alert: Risk Probability 30%

[0072] Triggering conditions: The predicted wave height for the next 24 hours is 2.5-3.0m, and the blue alert threshold for this sea area is 2.5-3.5m;

[0073] Push notification content: Send a text message to 200 monitoring personnel via SMS service, with the message: "[Marine Warning] Blue Warning: Wave height in a certain sea area is expected to reach 2.8m in the next 24 hours. Please pay attention to monitoring."

[0074] 2-2. Yellow Alert: Risk probability 50%

[0075] Triggering condition: Wave height forecast of 3.0-3.5m in the next 24 hours, yellow indicates the 3.0-3.5m range;

[0076] Response measures: Send a text message to monitoring personnel once a day (same as the blue alert), and add a WeChat official account graphic notification, including a wave forecast trend chart;

[0077] Push notification content: Published via WeChat official account: "Yellow Alert: Wave Height in a Certain Sea Area Expected to Reach 3.2m in the Next 24 Hours";

[0078] 2-3. Orange Alert: Risk Probability 70%

[0079] Triggering conditions: The predicted tide level for the next 24 hours is 3.5m, reaching the local warning tide level of 3.5m, but not exceeding the red threshold of 3.8m;

[0080] Response measures: Activate emergency broadcasting: The local emergency broadcasting system will broadcast for 3 minutes every hour from 6:00 to 22:00 daily; send a JSON file containing tide data to the emergency department every hour; and automatically push the "Orange Alert Defense Tips" to the leaders of coastal communities, including a list of key areas for dike inspection.

[0081] 2-4. Red Alert: Risk Probability 85%

[0082] Triggering conditions: If the predicted tide level is 3.8m and exceeds the local warning tide level of 3.5m, a pre-verification report will be automatically generated, including model output and comparison of similar historical cases. After manual confirmation by the duty officer, if an early warning is generated at 10:05:12, it will be pushed to the smart terminal at 10:05:20 with a response delay of 8 seconds. At the same time, a JSON file containing the latest prediction data will be sent to the emergency department every 10 minutes.

[0083] 3. Channel Verification: SMS Delivery Rate: According to operator feedback statistics, all 200 blue alert SMS messages were successfully delivered, with an average delivery time of 2.3 seconds; Emergency Broadcast: Connected to the local emergency broadcast system, the hourly broadcast recordings during the orange alert period can be verified by playback of the recording system, with a 7-second interval between the first broadcast time and the alert generation time.

[0084] V. Comparison Table of Marine Disaster Risk Prevention and Early Warning Systems

[0085]

[0086] 1. Summary: This invention surpasses traditional physical models, commercial systems, and existing regional solutions in three dimensions: data processing efficiency, prediction accuracy, and response speed. It is particularly suitable for the current situation of increasingly frequent marine disasters, and can transform disaster early warning "delayed response" into "real-time prevention and control". It is expected to reduce the risk of casualties by more than 30% and reduce economic losses by more than 100 million yuan per year.

[0087] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A marine disaster risk prevention and early warning system based on big data, characterized in that, Includes a multi-source data fusion module, an intelligent modeling module, and a real-time early warning module: The multi-source data fusion module is used to access real-time monitoring data from multiple sources such as satellite remote sensing and buoy monitoring, as well as historical disaster data. After cleaning and feature engineering processing, the data is stored in a distributed storage system. The intelligent modeling module constructs a hybrid prediction model based on spatiotemporal sequence data. The hybrid prediction model combines a spatiotemporal Transformer network with a physical correction formula to predict disaster risk probability and key parameters. Furthermore, the module has a closed-loop training mechanism that can automatically trigger incremental updates of model parameters based on post-disaster assessment data to continuously optimize prediction accuracy. The real-time early warning module generates a dynamic risk map and graded early warning information based on the prediction results, and pushes them to the terminal device; The hybrid prediction model combines a spatiotemporal Transformer network with a physical correction formula. The spatiotemporal Transformer network encodes marine disaster-related features such as wind speed, air pressure, and tide level into a spatiotemporal matrix with a 10-minute time step. It captures cross-feature correlations through a multi-head attention mechanism and outputs initial disaster risk probability values ​​and time-series feature vectors. The physical correction formula is based on the fundamental equations of tidal theory and uses Coriolis force correction terms and seabed friction parameter correction terms to calibrate the initial disaster risk probability values ​​and key parameters such as tide level output by the spatiotemporal Transformer network. This results in a disaster risk probability prediction value and prediction results for key parameters such as tide level, wind speed, and wave height that include physical mechanism constraints, enabling the prediction of disaster risk probability and key parameters for the next 24 hours. The physical correction formula is based on the fundamental equations of tidal theory: In the formula, This represents the storm surge increase in meters and is a key parameter that needs to be calibrated. This is a correction term for seabed friction parameters, where These represent the east-west and north-south components of the horizontal flow velocity, respectively, in meters per second. The coefficient of friction is related to the roughness of the seabed. This is a Coriolis force correction term. The Coriolis parameter takes a value of approximately It is related to geographical latitude; The velocity vector of the seawater flow. The curl of the velocity vector is used to correct for the influence of the Coriolis force caused by the Earth's rotation. The tide prediction results output by the spatiotemporal Transformer network are calibrated by the above correction term so that the tide prediction error is ≤0.2 meters.

2. The marine disaster risk prevention and early warning system based on big data according to claim 1, characterized in that, The multi-source data fusion module includes an edge computing node. The edge computing node acquires the 50Hz high-frequency sampled real-time monitoring raw data output from the device. It performs multi-resolution decomposition on the raw data through a wavelet transform algorithm: it performs anomaly filtering on the high-frequency components, removes noise and outliers that exceed the preset physical threshold, retains low-frequency trend characteristics, and verifies the rationality of the low-frequency components by combining historical data statistical patterns. Finally, it generates a minute-level downsampling data sequence that meets the needs of marine disaster early warning.

3. The marine disaster risk prevention and early warning system based on big data according to claim 1, characterized in that, The distributed storage system uses the HBase engine to store time-series monitoring data such as wave height, wind speed, and tide level, and the Elasticsearch engine to store disaster report text, sensor logs, and satellite remote sensing image metadata. Efficient data management is achieved through the following technologies: a unified API gateway: providing standardized data access interfaces and automatically routing query requests to the corresponding storage engine, supporting cross-database join queries based on time and region dimensions; index optimization strategy: establishing a time-plus-region composite index in HBase, and building a text inverted index and geospatial index in Elasticsearch to shorten single-database query response time; and a Redis caching mechanism: implementing a caching strategy for frequently accessed data from the past 7 days, reducing the response time for repeated queries from an average of 200ms to ≤20ms. Through the synergy of these technologies, the cross-database query latency of the storage system is ≤500ms, meeting the real-time data retrieval requirements for marine disaster early warning.

4. The marine disaster risk prevention and early warning system based on big data according to claim 1, characterized in that, The real-time early warning module is based on Leaflet.js to develop a dynamic risk map, which supports the overlay of coastal population density and important facility location data, and automatically identifies high-risk areas; the tiered early warning includes four levels: blue, yellow, orange, and red. Red warnings require pre-verification and secondary manual confirmation, and the early warning response delay is less than 10 seconds.

5. The marine disaster risk prevention and early warning system based on big data according to claim 1, characterized in that, The closed-loop training mechanism includes: automatically synchronizing disaster assessment data from the distributed storage system after a disaster. The assessment data includes real-time post-disaster monitoring data processed by the multi-source data fusion module and data on similar disasters in the historical case library. The real-time monitoring data includes measured values ​​of wind speed, wave height, and tide level. When the effective feature dimensions of the assessment data exceed 95% without missing values ​​and the number of available samples reaches 15% of the total training data, an incremental update of the parameters of the hybrid prediction model is triggered. The update process performs gradient optimization based on the current parameters of the model.

6. The marine disaster risk prevention and early warning system based on big data according to claim 1, characterized in that, The real-time early warning module pushes early warning information through mobile communication networks, internet platforms, emergency broadcasting systems, and smart terminal devices. Through a preset hierarchical strategy configuration table, blue warnings are pushed to monitoring personnel via SMS once a day, yellow warnings are accompanied by WeChat public account graphic notifications, orange warnings trigger hourly emergency broadcasts, and red warnings simultaneously trigger strong reminders on smart terminal apps and push notifications to emergency departments and coastal community leaders every 10 minutes, achieving differentiated delivery of different levels of early warnings.

7. The marine disaster risk prevention and early warning system based on big data according to claim 1, characterized in that, The multi-source data fusion module, intelligent modeling module, and real-time early warning module are deployed using a unified technical architecture and achieve automated connection through the following methods: Standardized data interface: Data is transmitted between modules through predefined data formats and interaction protocols; Task scheduling mechanism: Based on the workflow engine, tasks are automatically orchestrated in the order of data processing, model training, and early warning generation; Cross-module linkage: The early warning result triggers an incremental training request for the model, forming a closed-loop feedback.

8. A method for marine disaster risk prevention and early warning based on big data, comprising the design of a marine disaster risk prevention and early warning system based on big data according to any one of claims 1-7, characterized in that, Includes the following steps: S1. Multi-source data fusion processing: Integrate real-time monitoring data from multiple sources such as satellite remote sensing and buoy monitoring, as well as historical disaster data. Through edge computing nodes, use wavelet transform algorithm to perform minute-level downsampling and anomaly filtering on 50Hz high-frequency data, and perform data cleaning, repair missing values, remove outliers, and perform feature engineering to generate time-series data sequences that meet early warning requirements. S2. Hybrid Model Prediction Calculation: The processed data is input into the hybrid prediction model. The hybrid prediction model encodes marine disaster-related features such as wind speed, air pressure, and tide level into a spatiotemporal matrix using a spatiotemporal Transformer network with a 10-minute time step. It extracts spatiotemporal correlation features using a multi-head attention mechanism and outputs the initial value of disaster risk probability and time series feature vector. Based on the basic equations of tidal theory, the prediction results are physically calibrated using Coriolis force correction terms and seabed friction parameter correction terms to generate predicted values ​​of disaster risk probability and key parameters such as tide level, wind speed, and wave height for the next 24 hours. S3. Dynamic Early Warning and Information Push: Based on the prediction results, a dynamic risk map developed using Leaflet.js is overlaid with coastal population density and important facility location data to automatically identify high-risk areas; graded early warnings are triggered according to four preset thresholds: blue, yellow, orange, and red. Red warnings require pre-verification and secondary manual confirmation. Early warning information is pushed through multiple channels such as mobile communication networks and emergency broadcasts to ensure a response delay of less than 10 seconds. S4. Model closed-loop training mechanism: After a disaster, disaster assessment data is automatically synchronized from the distributed storage system. When the effective feature dimension of the assessment data exceeds 95% without missing data and the number of available samples reaches 15% of the total training data, the parameters of the hybrid prediction model are incrementally updated. The update process is based on gradient optimization performed on the current parameters of the model.

Images