An intelligent supervision method and system based on a data visualization platform

Abstract

The present application relates to the technical field of data processing, and more particularly to an intelligent monitoring method and system based on a data visualization platform, comprising collecting multi-source water conservancy data through double-channel redundancy checking and converting the data into a standardized space-time matrix, using containerization adaptation and a variational autoencoder to clean the data and mark the confidence; storing the data based on a confidence threshold, generating a routing strategy to distribute data fragments through reinforcement learning; using space-time graph convolution to calculate flood peak evolution prediction results, combining Bayesian networks and long short-term memory to calculate output risk assessment values, and activating federal learning parameter updates when the confidence is insufficient; loading physical constraint parameters into a digital twin model to deduce a flood control scenario, generating a set of scheduling instructions and a pre-play animation through multi-objective optimization; dynamically rendering an interactive visualization interface and associating a cause-and-effect deduction path, triggering federal model parameter updates through user feedback, and realizing a dynamic collaborative function chain of forecasting, early warning, pre-play and pre-plan.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to an intelligent supervision method and system based on a data visualization platform. BACKGROUND

[0002] The water conservancy data visualization platform integrates various water conservancy monitoring data sources, clearly displays the dynamic changes of hydrological elements through a graphical interface, and continuously analyzes real-time data streams through an intelligent supervision system integrated with machine learning models to automatically detect flow anomalies or rainfall deviations, and further predict potential disasters such as flood or drought risks. This capability assists water resource managers in making quick response decisions, optimizing reservoir scheduling or water-saving measures, ultimately improving water resource management efficiency and reducing disaster impact probability.

[0003] The existing water conservancy data visualization system lacks module decoupling design and standardized data conversion mechanism, resulting in the specific performance of basin full-element data intelligent linkage failure: when the reservoir scheduling module needs to urgently call meteorological rainfall data, river water level data and soil moisture data for flood risk pre-play, the existing architecture requires manual reconstruction of data pipeline adaptation. At this time, satellite remote sensing image data with inconsistent structure cannot be converted into standardized grid data that can be analyzed by hydrological models in real time, and the downstream disaster damage assessment module is continuously blocked due to its dependence on upstream preprocessing results. For example, in a sudden heavy rainfall event in a certain basin, format conflicts caused by diversified data sources (such as misalignment of time stamps of unmanned aerial vehicle aerial images and ground sensor data) force the system to interrupt the "forecast-pre-play" business chain, directly delaying the flood control scheduling decision window period, and causing the overall decline of disaster prevention efficiency. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides an intelligent supervision method and system based on a data visualization platform, which solves the technical problems of hindering independent expansion due to strong correlation between functional modules, difficulty in real-time synchronization of multi-dimensional water conservancy data, and blocking effect caused by linear dependence of analysis process, resulting in dynamic collaboration failure of basin-level forecasting, early warning, pre-play and pre-plan function chain.

[0005] To solve the above technical problems, the specific application of the present application is as follows:

[0006] In a first aspect, the present application provides an intelligent supervision method based on a data visualization platform, comprising:

[0007] Step 1: Collect meteorological radar data, river water level sensor data, soil moisture monitoring data and satellite remote sensing images, and output time-aligned initial data sets through double-channel redundancy verification;

[0008] Step 2: converting the initial data set into a standardized spatio-temporal matrix, including: containerizing the adaptive unified data format, pre-training the variational autoencoder to calculate the reconstruction error, generating scan anomaly data points, and outputting the cleaned spatio-temporal matrix with confidence labels;

[0009] Step 3: receiving the cleaned spatio-temporal matrix, storing the matrix data according to the confidence threshold, storing data with confidence greater than or equal to the threshold in the cache library, storing historical data in the spatio-temporal index library, monitoring real-time business scenario state and data storage location, generating routing strategy table through reinforcement learning decision, and extracting data fragments from the storage library according to the routing strategy table and distributing them to the message queue;

[0010] Step 4: extracting data fragments from the message queue, inputting parallel processing, and parallel processing including: spatio-temporal graph convolution calculation to generate flood peak evolution prediction results, Bayesian network and long short-term memory calculation to output risk assessment values, activating federal learning parameter update when integrated voting confidence is lower than the threshold, and outputting analysis results with reliability indicators;

[0011] Step 5: receiving the analysis results output by step 4, loading the pre-stored physical constraint parameters of the digital twin model, deducing the flood control scene according to the analysis results, outputting the flood peak evolution dynamic parameters, inputting the flood peak evolution dynamic parameters into multi-objective optimization calculation, iteratively generating scheduling instruction set, and generating pre-play animation data according to the scheduling instruction set;

[0012] Step 6: dynamically rendering the scheduling instruction set and pre-play animation data, generating an interactive visualization interface and associating the causal reasoning path, receiving user feedback instructions to trigger federal model parameter update.

[0013] Further, the intelligent supervision method based on the data visualization platform provided by the application comprises the following steps:

[0014] The main channel transmits the original data stream to the platform entrance, and the auxiliary channel transmits the feature hash value to the platform entrance at the same time;

[0015] The platform entrance receives the main channel data and calculates the real-time hash value of the main channel data;

[0016] The feature hash value transmitted by the auxiliary channel is extracted, and the real-time hash value and the feature hash value are compared. If they are inconsistent, the data source node is triggered to retransmit the main channel data until the hash values are consistent, and the initial data set with integrity mark is output.

[0017] Further, the intelligent supervision method based on the data visualization platform provided by the application comprises the following steps:

[0018] Receive the initial dataset, containerize and adapt the initial dataset, parse the meteorological data to obtain the spatiotemporal matrix of rainfall intensity, perform band calculations on the remote sensing data to obtain the elevation raster, and integrate the spatiotemporal matrix of rainfall intensity and the elevation raster into a unified format matrix.

[0019] The format-unified matrix is ​​input into the variational autoencoder to calculate the reconstruction error. The format-unified matrix is ​​input into the generative adversarial network to scan abnormal data points and output anomaly scores. The reconstruction error and anomaly scores are merged to generate data confidence labels.

[0020] Data confidence labels are mapped to a format-unified matrix, and the cleaned spatiotemporal matrix with confidence labels is output.

[0021] Furthermore, in the intelligent supervision method based on a data visualization platform described in this invention, step 3 includes:

[0022] After receiving the cleaned spatiotemporal matrix, the data is classified and stored according to the confidence threshold. Data with a confidence level greater than or equal to the threshold is stored in the high-speed cache and a spatiotemporal index is created. The remaining data is stored in the historical database.

[0023] Based on the stored data location, read the location status, capture the real-time business scenario code and data volume, and input the location status, business scenario code and data volume into the reinforcement learning decision.

[0024] The reward function calculates the optimal routing action, generates a routing policy table, retrieves data shards from the repository based on the routing policy table, and distributes the data shards to the message queue.

[0025] Furthermore, in the intelligent supervision method based on a data visualization platform described in this invention, step 4 includes:

[0026] Receive the data fragments distributed in step 3, and construct a watershed topology graph based on the water conservancy facility relationships in the data fragments. Nodes are water conservancy facility entities, and edges are water flow directions and distances.

[0027] The watershed topology map is input into the spatiotemporal graph for convolution calculation, and the flood peak evolution prediction results are output.

[0028] Input the flood peak evolution prediction results into the risk assessment and output the risk assessment value;

[0029] When the confidence level of the integrated voting for the risk assessment value is lower than the threshold, the federated learning parameter update is activated.

[0030] The system integrates flood peak evolution prediction results, risk assessment values, and federated learning update signals to output analysis results with reliability indicators.

[0031] Furthermore, in the intelligent supervision method based on a data visualization platform described in this invention, step 5 includes:

[0032] Receiving the analysis result output in step 4, the digital twin model loads the pre-stored physical constraint parameters, deduces the flood control scene according to the analysis result, and outputs the dynamic parameters of flood peak evolution;

[0033] Input the dynamic parameters of flood peak evolution into multi-objective optimization calculation, and generate the gate opening sequence by chromosome coding;

[0034] The fitness calculation synchronously evaluates the flood control safety degree and the water supply guarantee degree, iteratively optimizes the gate opening sequence, outputs the Pareto optimal scheduling scheme, and generates the pre-visualization animation data according to the Pareto optimal scheduling scheme.

[0035] Further, the intelligent supervision method based on the data visualization platform provided by the application comprises the following steps:

[0036] Receiving the initial gate opening sequence, the tournament selection filters the parent chromosomes;

[0037] Input the parent chromosomes into the simulated binary crossover to generate the child chromosome sequence;

[0038] Dynamically adjusting the weight coefficients α, β and γ of the child chromosome sequence, updating the fitness calculation rule, iteratively optimizing the child chromosome sequence based on the updated rule, and outputting the Pareto optimal scheduling data;

[0039] Generating personnel transfer path data based on the Pareto optimal scheduling data.

[0040] Further, the intelligent supervision method based on the data visualization platform provided by the application comprises the following steps:

[0041] Dynamically rendering and loading the scheduling instruction set and the pre-visualization animation data, generating an interactive visualization interface, and superimposing the flood peak risk heat map on the digital elevation model base map;

[0042] The interactive visualization interface receives the click instruction of the user on the high-risk area, triggers the playback of the digital twin model deduction process;

[0043] The visualization interface displays the pre-plan path animation, the animation shows the iterative trajectory of the multi-objective optimization algorithm, the interactive visualization interface receives the user correction instruction, generates a difference vector, encrypts and uploads the difference vector, and triggers the federal model parameter update.

[0044] Further, the intelligent supervision method based on the data visualization platform provided by the application comprises the following steps:

[0045] Receiving the encrypted uploaded difference vector, federated aggregation analyzing the difference vector, and calculating the global model parameter update amount;

[0046] Distributing the global model parameter update amount to independent processing;

[0047] independently process the updated risk assessment model parameters to generate updated risk assessment values;

[0048] input the updated risk assessment values into dynamic rendering to refresh the risk heat map of the visual interface in real time.

[0049] In a second aspect, the present application provides an intelligent supervision system based on a data visualization platform, which is applied to the intelligent supervision method based on the data visualization platform and comprises:

[0050] a data acquisition unit configured to acquire meteorological radar data, river water level sensing data, soil moisture monitoring data and satellite remote sensing images, and execute double-channel redundancy verification to output an initial data set aligned in time;

[0051] a data processing unit configured to convert the initial data set into a standardized space-time matrix, including: containerizing, adapting and unifying a data format, pre-training a variational autoencoder to calculate a reconstruction error, generating scanning abnormal data points in an adversarial manner, and outputting a cleaned space-time matrix with a confidence label;

[0052] a data analysis unit configured to receive the cleaned space-time matrix, store the matrix data in layers according to a confidence threshold, monitor real-time business scenario states and data storage locations, generate a routing strategy table through reinforcement learning decision-making, extract data shards from a storage library according to the routing strategy table and distribute them to a message queue;

[0053] a parallel processing unit configured to extract data shards from the message queue, execute space-time graph convolution calculation to generate a flood peak evolution prediction result, output a risk assessment value through Bayesian network and long short-term memory calculation, activate federal learning parameter updating when integrated voting confidence is lower than a threshold, and output an analysis result with a reliability index;

[0054] a generation unit configured to receive the analysis result, load pre-stored physical constraint parameters through a digital twin model to deduce a flood control scenario, output flood peak evolution dynamic parameters, iteratively generate a scheduling instruction set through multi-objective optimization calculation, and generate pre-play animation data according to the scheduling instruction set;

[0055] an optimization unit configured to dynamically render the scheduling instruction set and the pre-play animation data, generate an interactive visual interface and associate a cause-effect deduction path, and receive user feedback instructions to trigger federal model parameter updating.

[0056] The present application has the following beneficial effects:

[0057] The application effectively eliminates the risk of packet loss in multi-source water conservancy data transmission through a dual-channel redundancy verification mechanism, solves the format conflict problem of satellite images and sensor data by combining containerized adaptive process and intelligent quality evaluation, and supports microservice module hot plug to eliminate the strong correlation between functional modules by strengthening the learning decision-making generated routing strategy table; The space-time graph convolution calculation and Bayesian risk assessment in the parallel processing process are executed synchronously to avoid linear blocking of the analysis process, the federal learning parameter update is automatically activated to improve the model reliability when the integrated voting confidence is insufficient, the digital twin model loads the physical constraint parameters to deduce the flood control scene, and combines the multi-objective optimization calculation to output the Pareto optimal scheduling scheme, realizing the millisecond-level dynamic cooperation of the forecast, early warning, rehearsal and plan function chain; The federal model parameter update driven by user feedback and the real-time refresh of the visual interface form a decision optimization closed loop, which radically solves the coordination failure problem of the basin-level four-pre function chain and significantly improves the efficiency of flood control. BRIEF DESCRIPTION OF DRAWINGS

[0058] In order to more clearly illustrate the technical solutions of the present application, the following will briefly introduce the drawings needed in the embodiments. Obviously, for those skilled in the art, other drawings can also be obtained without creative labor based on the drawings.

[0059] Figure 1 The flowchart of the intelligent monitoring method based on the data visualization platform provided by the embodiment of the present application. DETAILED DESCRIPTION

[0060] In order to make the technical solutions of the present application clearer, the following will combine the specific embodiments of the present application and the corresponding drawings to clearly and completely describe the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application. The following will combine the drawings to specifically describe the present application provided by the embodiments of the present application. In order to better understand the purpose of the present application, the following will further describe the present application in detail.

[0061] In the first aspect, referring to Figure 1 The present application provides an intelligent monitoring method based on a data visualization platform, comprising:

[0062] Step 1: Collect meteorological radar data, river water level sensor data, soil moisture monitoring data and satellite remote sensing images, and output the initial data set aligned in time through dual-channel redundancy verification;

[0063] Step 2: Convert the initial data set into a standardized spatio-temporal matrix, including: containerized adaptation of unified data format, pre-training of variational autoencoder to calculate reconstruction error, generation of adversarial scan anomaly data points, and output of cleaned spatio-temporal matrix with confidence labels;

[0064] Step 3: Receive the cleaned spatio-temporal matrix, store the matrix data in layers according to the confidence threshold, store data with confidence greater than or equal to the threshold in the cache library, store historical data in the spatio-temporal index library, monitor real-time business scenario state and data storage location, generate routing strategy table through reinforcement learning decision, and extract data shards from the storage library according to the routing strategy table and distribute them to the message queue;

[0065] Step 4: Extract data shards from the message queue, input parallel processing, which includes: spatio-temporal graph convolution calculation to generate flood peak evolution prediction results, Bayesian network and long short-term memory calculation to output risk assessment values, activation of federal learning parameter update when integrated voting confidence is below the threshold, and output of analysis results with reliability indicators;

[0066] Step 5: Receive the analysis results output by step 4, load the pre-stored physical constraint parameters of the digital twin model, deduce the flood control scenario according to the analysis results, output the flood peak evolution dynamic parameters, input the flood peak evolution dynamic parameters into multi-objective optimization calculation, iteratively generate a set of scheduling instructions, and generate pre-play animation data according to the scheduling instruction set;

[0067] Step 6: Dynamically render the scheduling instruction set and pre-play animation data, generate an interactive visualization interface and associate the causal reasoning path, and receive user feedback instructions to trigger federal model parameter updates.

[0068] In the data collection stage, the system synchronously acquires meteorological radar data, river water level sensing data, soil moisture monitoring data, and satellite remote sensing images through a distributed sensor network. The dual-channel redundancy check mechanism uses a parallel transmission method with a main channel and an auxiliary channel. The main channel transmits the original data stream, and the auxiliary channel synchronously transmits the feature hash value. After receiving the data at the platform entrance, the real-time hash value is calculated and compared with the auxiliary channel hash value. If they are not consistent, the data source is triggered to retransmit, thereby ensuring data integrity and time alignment. This design solves the synchronization failure problem caused by packet loss or time sequence misalignment of multi-source water conservancy data, providing reliable input for subsequent processing.

[0069] The initial data set enters the data conversion link, and the containerized adaptation process uniformly processes the format of heterogeneous data sources. Meteorological data is parsed into a rainfall intensity spatio-temporal matrix, and remote sensing data is processed through band operation to generate an elevation raster; the variational autoencoder calculates the reconstruction error to quantify data quality, and the adversarial generation network scans sensor anomaly points to output anomaly scores, and the two are combined to generate data confidence labels. The cleaned spatio-temporal matrix with confidence labels is output, providing data quality basis for layered storage.

[0070] The post-cleaning space-time matrix implements a tiered storage strategy according to a confidence threshold. High-confidence data is stored in a cache library and indexed in space-time, and historical data is archived in a space-time index library; the system monitors the business scene state and data storage location in real time, and a reinforcement learning decision module generates a routing strategy table based on scene encoding, data magnitude and storage state. According to the strategy table, data shards are extracted from the corresponding storage library and distributed to the message queue, realizing dynamic scheduling of data resources and decoupling of modules.

[0071] The data shards in the message queue trigger a parallel processing flow. The space-time graph convolution calculation is based on the weighted aggregation of adjacent node features based on the basin topological structure, captures the space-time law of flood peak evolution and outputs the prediction results; the Bayesian network quantifies the uncertainty factors, and the long short-term memory network learns the time series dependence relationship, and jointly generates the risk assessment value. When the integrated voting confidence is lower than the threshold, the federated learning parameter update mechanism is activated, and the model reliability is improved through multi-node collaborative training, and finally the analysis results with reliability indicators are output.

[0072] The analysis results drive the digital twin model to perform flood control scene deduction. The model loads the pre-stored physical constraint parameters, such as hydro-mechanical equations and terrain data, and simulates the evolution of the flood control scene in combination with the dynamic parameters of the flood peak evolution; the multi-objective optimization calculation adopts chromosome coding to generate a gate opening sequence, and the fitness function simultaneously evaluates the flood control safety and water supply security, and outputs the Pareto optimal scheduling scheme through iterative optimization, and generates pre-play animation data to visually display the scheduling effect.

[0073] The dynamic rendering module loads the scheduling instruction set and pre-play animation data, and superimposes the flood peak risk heat map on the digital elevation model base map to generate an interactive visualization interface. Users can trigger the digital twin deduction process playback by clicking on high-risk areas through the interface, view the pre-plan path animation and optimization algorithm iteration trajectory; user correction instructions generate difference vectors, which are uploaded after encryption to trigger federated model parameter updates, forming a closed-loop collaborative mechanism from decision deduction to feedback optimization.

[0074] Specifically, the intelligent supervision method based on the data visualization platform comprises the following steps:

[0075] The main channel transmits the original data stream to the platform entrance, and the auxiliary channel transmits the feature hash value to the platform entrance at the same time;

[0076] The platform entrance receives the main channel data and calculates the real-time hash value of the main channel data;

[0077] The feature hash value transmitted by the auxiliary channel is extracted, and the real-time hash value and the feature hash value are compared. If they are inconsistent, the data source node is triggered to retransmit the main channel data until the hash values are consistent, and the initial data set with the integrity mark is output.

[0078] In the data collection phase, the dual-channel redundancy verification mechanism transmits the original data stream to the platform entrance through the main channel, while the auxiliary channel transmits the feature hash value in parallel. The main channel carries the complete data packets of meteorological radar data, river water level sensing data, soil moisture monitoring data and satellite remote sensing images, and the auxiliary channel transmits the lightweight feature digest generated based on the SHA-256 algorithm. This kind of dual-channel parallel transmission architecture effectively avoids the risk of data packet loss caused by network jitter or device failure in single-channel transmission.

[0079] After receiving the main channel data, the platform entrance immediately calculates the real-time hash value. The calculation process uses the same hash algorithm as the data source end to ensure consistency of the comparison benchmark. After generating the real-time hash value, the system automatically extracts the feature hash value pre-transmitted by the auxiliary channel for comparison. The hash comparison module uses direct memory comparison to avoid the delay caused by disk reading and writing, meeting the real-time requirements of water conservancy monitoring scenarios.

[0080] When the real-time hash value and the feature hash value are inconsistent, the system determines that the data transmission integrity is damaged, and immediately triggers the data source node retransmission mechanism. The retransmission instruction is sent to the edge data collection node through a dedicated control channel, requiring the main channel data stream to be retransmitted. Hash value comparison is continuously performed during the retransmission process until the two hash values are completely consistent. The final output is an initial data set with an integrity label, which includes the timestamp and data volume information of the hash verification passing, providing a reliability basis for subsequent processing links.

[0081] This verification mechanism is particularly suitable for cross-regional water conservancy monitoring networks. For example, when a mountain rain station transmits data through a wireless network, signal interference may cause data packet loss. The dual-channel design enables the system to quickly detect abnormalities through hash value comparison of the auxiliary channel even if part of the main channel data is lost, and triggers the retransmission of specified data segments, avoiding the bandwidth waste caused by existing end-to-end retransmission.

[0082] Specifically, the intelligent supervision method based on the data visualization platform comprises the following steps:

[0083] The initial data set is received, the initial data set is containerized and adapted, the meteorological data is parsed to obtain a rainfall intensity spatio-temporal matrix, the remote sensing data band operation is performed to obtain an elevation grid, and the rainfall intensity spatio-temporal matrix and the elevation grid are integrated into a format-unified matrix;

[0084] The format-unified matrix is input into a variational autoencoder to calculate a reconstruction error, the format-unified matrix is input into an adversarial generative network to scan an abnormal data point and output an abnormal score, and the reconstruction error and the abnormal score are merged to generate a data confidence label;

[0085] The data confidence label is mapped to the format-unified matrix, and a cleaned spatio-temporal matrix with a confidence label is output.

[0086] In the data conversion link, the system starts the containerized adaptive conversion process after receiving the initial data set. This process deploys special parsing containers for meteorological radar data, river water level sensor data, and other heterogeneous data sources. The meteorological data container parses GRIB format files to generate rainfall intensity spatiotemporal matrices, and the remote sensing data container extracts elevation grid information through multi-spectral band operations. The outputs of each container follow a unified coordinate system and timestamp specification, and are finally integrated into a dimensionally consistent format-unified matrix. This design solves the integration problem caused by format differences between satellite images and sensor data. For example, when there is a spatial scale inconsistency between unmanned aerial vehicle aerial photography data and ground monitoring data, the containerized adaptive conversion automatically performs coordinate conversion and data resampling to ensure matrix element alignment.

[0087] The format-unified matrix is then input into two quality assessment channels in parallel. The variational autoencoder calculates reconstruction error based on a pre-trained model, quantifies the degree of data distortion by comparing the difference between the input data and the decoded output, and the adversarial generative network identifies abnormal patterns through a discriminator, outputting an abnormality score for each data point. The two assessment results are combined in a confidence fusion module, with reconstruction error reflecting system error level and abnormality score detecting sudden interference, together forming the basis for generating data confidence labels. For example, in the water conservancy monitoring scenario, the variational autoencoder can mark abnormal values in the rainfall intensity matrix that exceed historical patterns, while the adversarial generative network can identify sudden changes in the elevation grid.

[0088] The generated confidence labels are associated with the format-unified matrix through a spatial mapping module. Each grid cell is attached to the corresponding confidence value, forming a labeled spatiotemporal matrix data structure. This design allows downstream processing modules to flexibly adjust data processing strategies based on confidence thresholds, such as activating a review mechanism or allocating more computing resources for verification in low-confidence areas. The output of the cleaned spatiotemporal matrix with confidence labels not only includes standardized monitoring data, but also carries complete quality assessment information, providing structured input for subsequent hierarchical storage and parallel processing.

[0089] Specifically, the intelligent supervision method based on the data visualization platform comprises the following steps:

[0090] Receiving the cleaned spatiotemporal matrix, storing the data according to the confidence threshold, and storing the data with confidence greater than or equal to the threshold in the cache library and establishing a spatiotemporal index, and storing the remaining data in the historical database;

[0091] Based on the stored data location, reading the location state, capturing the real-time business scenario code and data magnitude, and inputting the location state, business scenario code and data magnitude into the reinforcement learning decision;

[0092] The reward function calculates the optimal routing action, generates a routing policy table, extracts data shards from the repository according to the routing policy table, and distributes the data shards to the message queue.

[0093] In the data storage link, after receiving the cleaned spatio-temporal matrix, the system implements a hierarchical storage strategy according to the preset confidence threshold. High-confidence data is stored in the cache library and indexed in space and time, and historical data is stored in the spatio-temporal index library. This design optimizes resource allocation based on data quality. For example, when real-time rainfall data is needed for flood peak prediction, the system prioritizes high-confidence data from the cache library, reducing access latency.

[0094] The system continuously monitors the real-time business scenario state and data storage location, capturing key parameters such as business scenario encoding and data magnitude. The monitoring module obtains the business scenario encoding through a sensor network, such as a typhoon warning or drought assessment scenario, while reading the data storage location state, including the load of the cache library and index library. The parameters input the reinforcement learning decision module, which calculates the optimal routing action based on the Q-learning algorithm and generates a routing policy table.

[0095] The reinforcement learning decision module evaluates the utility of different routing actions through a reward function, such as routing high-priority data to parallel processing flows. The routing policy table explicitly specifies which storage repository to extract data shards from and how to distribute them to the message queue. According to the policy table, the system extracts data shards from the corresponding storage repository and distributes them to downstream processing modules through the message queue, achieving dynamic scheduling of data flow.

[0096] In water conservancy monitoring applications, when the system detects a flood warning scenario, the reinforcement learning decision prioritizes routing real-time water level data shards to the flood peak evolution prediction module, ensuring that critical data is processed in a timely manner. This dynamic routing mechanism effectively avoids data congestion and improves system response speed.

[0097] Specifically, the intelligent supervision method based on the data visualization platform comprises the following steps:

[0098] Receiving the data shards distributed in step 3, constructing a river basin topology graph based on the water conservancy facility relationship in the data shards, with nodes representing water conservancy facility entities and edges representing water flow direction and distance;

[0099] Inputting the river basin topology graph into a spatio-temporal graph convolution calculation to output a flood peak evolution prediction result;

[0100] Inputting the flood peak evolution prediction result into risk assessment to output a risk assessment value;

[0101] When the integrated voting confidence of the risk assessment value is lower than the threshold, activate the federated learning parameter update;

[0102] Integrate the flood peak evolution prediction result, the risk assessment value and the federal learning update signal, and output the analysis result with a reliability index.

[0103] Receiving the data fragments distributed in step 3, analyzing the spatial correlation attributes of water conservancy facilities in the data fragments, taking reservoirs and gates as node entities, and taking river flow direction and hydraulic connection distance as edge relationship, a basin topological graph is constructed;

[0104] The basin topological graph is input into a spatio-temporal graph convolution calculation module, the hydrological characteristics of adjacent nodes are aggregated through a graph convolution layer, the time sequence law of flood peak propagation is captured through a time convolution layer, and a flood peak evolution prediction result is output;

[0105] The flood peak evolution prediction result is input into a risk assessment module, a Bayesian network calculates a disaster occurrence probability distribution, a long short-term memory network learns a historical flood sequence pattern, and a risk assessment value is fused and output;

[0106] When the integrated voting confidence of the risk assessment value is lower than a preset threshold, a federal learning parameter update mechanism is activated, and encrypted model parameters of multiple local nodes are updated;

[0107] Integrating the flood peak evolution prediction result, the risk assessment value and the federal learning parameter update signal, and adding a timestamp and a credibility label, an analysis result with a reliability index is output.

[0108] Specifically, the intelligent supervision method based on the data visualization platform comprises the following steps:

[0109] Receiving the analysis result output in step 4, a digital twin model loads pre-stored physical constraint parameters, a flood control scene is deduced according to the analysis result, and flood peak evolution dynamic parameters are output;

[0110] The flood peak evolution dynamic parameters are input into a multi-objective optimization calculation, and a chromosome code generates a gate opening sequence.

[0111] The fitness calculation synchronously evaluates the flood control safety degree and the water supply guarantee degree, iteratively optimizes the gate opening sequence, outputs a Pareto optimal scheduling scheme, and generates a pre-play animation data according to the Pareto optimal scheduling scheme.

[0112] In the flood control decision-making link, after the system receives the analysis result output in step 4, the digital twin model loads pre-stored physical constraint parameters. The parameters include water conservancy mechanics equation coefficients, terrain elevation data, river roughness coefficients and other basic physical rules. The model is based on the flood peak evolution prediction and the risk assessment value in the analysis result, and deduces the evolution process of the flood control scene in a virtual environment. The deduction process simulates the reservoir storage and discharge dynamics and the river flood propagation under different scheduling strategies, and outputs the flood peak evolution dynamic parameters including water level change rate, flood peak propagation time and other key indicators.

[0113] The dynamic parameter input multi-objective optimization calculation module of flood peak evolution adopts a chromosome coding technology to generate a gate opening sequence. Each chromosome individual represents a set of gate scheduling schemes, and a gene site corresponds to a gate opening value in a specific time period. The fitness calculation module synchronously evaluates two objective functions of flood control safety and water supply guarantee. The flood control safety quantifies the risk of river water level exceeding the warning level, and the water supply guarantee evaluates the ability of reservoir storage capacity to meet the demand.

[0114] The gate opening sequence is iteratively optimized through selection, crossover, and mutation operations. The tournament selection mechanism selects parent individuals with higher fitness, the binary crossover operation generates offspring populations, and the dynamic mutation operation maintains population diversity. The individual fitness is recalculated in each iteration to gradually approach the Pareto optimal solution set. When the maximum number of iterations or the convergence threshold is reached, the Pareto optimal scheduling scheme set is output.

[0115] Pre-visualization animation data is generated based on the Pareto optimal scheduling scheme. The animation sequence shows the flood evolution process under different scheduling strategies, including reservoir water level change curves, river flood peak propagation trajectories, and dynamic changes in flooded areas. The animation data is associated with the scheduling instruction set to provide basic materials for subsequent visualization presentation.

[0116] In specific application scenarios, when the system processes a storm event in a watershed, the digital twin model will deduce the future 72-hour flood evolution process based on real-time rainfall forecasts. The multi-objective optimization module considers both downstream flood control safety and urban water supply demand to generate the best gate control sequence that balances flood discharge and water storage. The final output pre-visualization animation can visually demonstrate the implementation effects of different decision-making schemes, supporting flood control consultation decisions.

[0117] Specifically, the intelligent supervision method based on the data visualization platform includes the following steps:

[0118] Receive the initial gate opening sequence, and select the parent chromosomes through tournament selection;

[0119] Input the parent chromosomes into simulated binary crossover to generate offspring chromosome sequences;

[0120] Dynamically adjust the weight coefficients α, β, and γ of the offspring chromosome sequences, update the fitness calculation rules, iteratively optimize the offspring chromosome sequences based on the updated rules, and output the Pareto optimal scheduling data;

[0121] Generate personnel transfer path data based on the Pareto optimal scheduling data.

[0122] In the optimization calculation link, the system receives the initial gate opening sequence and starts the genetic algorithm optimization process. The tournament selection mechanism selects parent chromosomes with higher fitness from the population. Each chromosome individual represents a set of gate scheduling scheme. The selection process simulates the competition mechanism, randomly selects multiple individuals to compare the fitness values, and the winning individuals are used as parents for reproduction, maintaining population diversity and avoiding premature convergence.

[0123] The parent chromosome input simulates binary crossover operation to generate offspring chromosome sequence by gene bit exchange. The crossover operation randomly selects a crossover point based on a probability threshold, exchanges parent chromosome fragments, and generates new individuals. The simulated binary crossover uses continuous variable processing method to keep the gene value within a reasonable range, avoid illegal solutions, and enhance the global search ability of the algorithm.

[0124] The weight coefficients α, β and γ of the offspring chromosome sequence are dynamically adjusted. The coefficients correspond to the weight proportion of the flood safety degree and water supply security degree target functions. The weight adjustment module adaptively updates the coefficient values according to the iteration progress and population distribution state, balancing the conflict relationship in multi-objective optimization. The updated weight coefficients are input into the fitness calculation rule to re-evaluate the comprehensive fitness of each chromosome.

[0125] Based on the updated fitness calculation rule, the system iteratively optimizes the offspring chromosome sequence. Each iteration performs selection, crossover and mutation operations to gradually approach the Pareto optimal solution set. When the maximum iteration number or solution set convergence threshold is reached, the Pareto optimal scheduling data set is output, including multiple non-dominated solution schemes.

[0126] Based on the Pareto optimal scheduling data, the system generates personnel transfer path data. The path planning module combines the flood inundation range prediction results, traffic network data and shelter location information to calculate the optimal evacuation route. The output data includes transfer time window, path capacity limit and emergency resource allocation scheme, forming a complete personnel scheduling plan.

[0127] Specifically, the intelligent supervision method based on the data visualization platform comprises the following steps:

[0128] The dynamic rendering loads the scheduling instruction set and the pre-animation data to generate an interactive visualization interface, which superimposes the flood peak risk heat map on the digital elevation model base map.

[0129] The interactive visualization interface receives user click instructions on high-risk areas to trigger the playback of the digital twin model deduction process.

[0130] The visualization interface displays the pre-plan path animation, which shows the iteration trajectory of the multi-objective optimization algorithm. The interactive visualization interface receives user correction instructions, generates a difference vector, encrypts and uploads the difference vector, and triggers the federal model parameter update.

[0131] In the visualization presentation link, the system dynamic rendering module loads the scheduling instruction set and the pre-play animation data, and generates an interactive visualization interface through the WebGL graphics engine. The interface takes a digital elevation model as a base map, and superimposes a flood peak risk heat map. The color gradient of the heat map reflects the distribution of risk assessment values, and the red area identifies the high-risk flooded area. The rendering process real-time fuses the terrain elevation data and the dynamic parameters of the flood peak evolution, generates a three-dimensional scene visualization effect, and supports users to zoom in and rotate to view the details of the basin.

[0132] The interactive visualization interface receives user mouse click instructions on the high-risk area, and the interface event capture module analyzes the geographical location corresponding to the click coordinates. The click instruction triggers the playback function of the digital twin model deduction process, and the playback module loads the historical deduction data to reproduce the whole process of the flood peak evolution in the form of time axis sliding. The playback process synchronously displays the water level change curve and the gate operation sequence to assist users in analyzing the decision basis.

[0133] The visualization interface continuously plays the pre-plan path animation, and the animation sequence shows the iteration trajectory of the multi-objective optimization algorithm. The genetic algorithm evolution process is presented in a dynamic scatter plot, each particle represents a scheduling scheme, and the trajectory reflects the convergence path of the Pareto frontier. At the same time, the personnel transfer path animation is displayed, the arrow flow direction indicates the evacuation direction, and the path color depth represents the priority of passing.

[0134] The interactive visualization interface provides a parameter correction panel to receive user correction instructions in the form of dragging sliders or inputting numerical values. Users can adjust parameters such as gate opening threshold and evacuation time window, and the correction instruction processing module calculates the difference vector between the new scheme and the original scheme in real time. The difference vector includes parameter offset and timing adjustment, which is encrypted through the AES encryption algorithm and uploaded to the cloud server.

[0135] The encrypted uploaded difference vector triggers the federal model parameter update mechanism. The federal aggregation server receives the encrypted data uploaded by each terminal, analyzes the difference vector through homomorphic decryption, and calculates the global model parameter update amount. The updated model parameters are distributed to each edge node to optimize the inference rules of the risk assessment module in real time, forming a user feedback driven model evolution closed loop.

[0136] Specifically, the intelligent supervision method based on the data visualization platform provided by the present application, the step 6 further comprises:

[0137] Receiving the encrypted uploaded difference vector, the federal aggregation analyzes the difference vector, and calculates the global model parameter update amount;

[0138] Distribute the global model parameter update amount to independent processing;

[0139] Update the risk assessment model parameters in the independent processing, and generate the updated risk assessment value;

[0140] The updated risk assessment value is input into dynamic rendering, and the risk heat map of the visual interface is refreshed in real time.

[0141] In the federal model updating link, the system receives encrypted uploaded difference vectors from multiple terminal nodes. The difference vectors are uploaded to the federal aggregation server through a secure transmission protocol, and each vector includes the user's local adjustment information of the model parameters. The encryption layer uses an asymmetric encryption algorithm to protect the data transmission process, preventing sensitive information from being intercepted or tampered with in the transmission link.

[0142] The federal aggregation server starts the parsing process and performs batch decryption operations on the received encrypted difference vectors. The parsing module extracts the parameter offset and adjustment weight in the vector, and calculates the global model parameter update based on the federated averaging algorithm. The calculation process takes into account the data quality weight of each terminal node to avoid excessive influence of individual abnormal data on the global model.

[0143] The calculated global model parameter update is distributed to each independent processing node through a secure channel. The distribution mechanism uses an incremental update strategy, transmitting only the parameter change amount rather than the complete model, reducing network bandwidth occupancy. Each independent processing node receives the update and fuses it with the local model parameters to complete the model version iteration.

[0144] Based on the updated model parameters, the independent processing node re-runs the risk assessment calculation. The node loads the latest water monitoring data and generates updated risk assessment values through the forward propagation algorithm. The calculation process integrates real-time hydrological features and historical patterns to output a timestamped risk probability distribution.

[0145] The updated risk assessment value is input into the dynamic rendering engine in real time. The rendering engine adjusts the color mapping relationship of the heat map based on the new risk assessment value, with high-risk areas displaying warm tones and low-risk areas displaying cool tones. The visual interface refreshes the risk heat map in frame synchronization mode to ensure that the user always sees the latest model evaluation results.

[0146] In the flood control application scenario of a river basin, when multiple flood control command centers make different risk judgments for the same area, the system integrates the local experience of each center through the federal aggregation mechanism to generate a more accurate global risk assessment. The updated risk heat map can display the dynamic changes in risk levels during the flood peak evolution process, providing continuous optimization reference for flood control decision-making.

[0147] In a second aspect, the present application provides an intelligent monitoring system based on a data visualization platform, which is applied to the intelligent monitoring method based on the data visualization platform as described, and comprises:

[0148] The data acquisition unit is configured to collect meteorological radar data, river water level sensor data, soil moisture monitoring data and satellite remote sensing images, and execute double-channel redundant verification to output an initial data set aligned in time;

[0149] The data processing unit is configured to convert the initial data set into a standardized space-time matrix, including: containerized adaptation of a unified data format, pre-training of a variational autoencoder to calculate reconstruction error, generation of scan anomaly data points through adversarial generation, and output of a cleaned space-time matrix with confidence labels;

[0150] The data analysis unit is configured to receive the cleaned space-time matrix, store the matrix data in layers according to a confidence threshold, monitor real-time business scenario states and data storage locations, generate a routing strategy table through reinforcement learning decision-making, and extract data shards from the storage repository according to the routing strategy table and distribute them to a message queue;

[0151] The parallel processing unit is configured to extract data shards from the message queue, perform space-time graph convolution calculation to generate flood peak evolution prediction results, and output risk assessment values through Bayesian network and long short-term memory calculation. When the integrated voting confidence is lower than the threshold, the federal learning parameter update is activated, and the analysis results with reliability indicators are output.

[0153] The generation unit is configured to receive the analysis results, load the pre-stored physical constraint parameters of the digital twin model to deduce the flood control scenario, output the flood peak evolution dynamic parameters, and iteratively generate a set of scheduling instructions through multi-objective optimization calculation. According to the scheduling instruction set, pre-visualization animation data is generated.

[0154] The optimization unit is configured to dynamically render and load the scheduling instruction set and pre-visualization animation data, generate an interactive visualization interface and associate it with the causal reasoning path, and receive user feedback instructions to trigger federal model parameter updates.

[0155] In the specific implementation process, the system is deployed in the basin flood control command center and connected to data sources such as meteorological radar stations, hydrological monitoring stations and satellite receiving stations. In implementation, a distributed architecture is adopted, with a main processing node deployed at the provincial center and edge nodes deployed at city-level sub-centers, forming a hierarchical processing network. The main node is equipped with a high-performance graphics server cluster, and the edge node is configured with a real-time data processing module, interconnected through a fiber optic private network.

[0156] The data acquisition link accesses the base data stream of the meteorological radar, with a radar scanning interval of 6 minutes per group of volume scan data. Minute-level monitoring values from river water level sensors are also received. Soil moisture monitors upload a set of data every half hour, and satellite remote sensing images update elevation information every 3 hours. The main channel uses TCP protocol to transmit raw data packets, and the auxiliary channel uses UDP protocol to synchronously send data feature summaries. A data verification module is set up at the platform entrance to calculate SHA-256 hash values in real time for integrity verification, and if data packets are lost, the specified node is triggered to retransmit.

[0156] The data conversion stage deploys a containerized adaptation cluster, including meteorological data parsing containers, remote sensing image processing containers, sensor data conversion containers, and other independent modules. The meteorological data container parses radar-based data to generate a 1 km x 1 km grid rainfall intensity matrix, and the remote sensing container processes Sentinel satellite images to generate a 30-meter resolution elevation raster. The output results of each container are uniformly converted to the WGS84 coordinate system, time-aligned to the whole point time, and formed into a standard space-time matrix. The quality assessment module uses a pre-trained variational autoencoder to calculate data reconstruction error, and an adversarial generation network to detect sensor outliers, labeling each data point with a confidence score.

[0157] The cleaned space-time matrix enters the hierarchical storage system, and real-time data with a confidence score of 0.9 or higher is stored in the Redis cache library, establishing a space-time joint index. Historical data is stored in a time series database, with index keys including timestamps and spatial coordinates. The routing decision module monitors business scenarios in real time, and when the system identifies a typhoon warning mode, it automatically routes real-time rainfall data to the flood peak prediction pipeline; when a drought warning occurs, soil moisture data is preferentially routed to the drought assessment module.

[0158] After the parallel processing unit extracts data shards from the message queue, it first constructs a watershed water conservancy facility topology graph. With reservoirs as nodes and river channels as edges, node attributes include reservoir capacity curves, and edge attributes include river channel slope and roughness. The space-time graph convolution network uses a 3-layer graph convolution combined with a time convolution structure to aggregate upstream node features to predict downstream flood peak evolution. The Bayesian network calculates conditional probabilities, and the long short-term memory network learns historical flood sequences. When the voting confidence of multiple models is below 0.85, the federated learning mechanism is triggered to update model parameters.

[0159] The digital twin model loads physical parameters such as river terrain and reservoir characteristic curves to simulate flood control scenarios in a virtual environment. The multi-objective optimization module uses the NSGA-II algorithm with a population size of 100 and iterates 200 times to find the Pareto optimal solution. The fitness function considers both downstream flood control safety index and urban water supply guarantee rate, outputting the best gate control strategy that balances flood discharge and water storage.

[0160] The visualization interface uses WebGL technology to render a three-dimensional scene, superimposing a flood peak risk heat map on the digital elevation model. Users can replay the digital twin simulation process and view flood evolution trajectories by clicking on high-risk areas. After receiving user adjustments to gate opening, the system generates a parameter difference vector and updates the risk assessment model through the federated learning mechanism, refreshing the heat map display in real time.

[0161] In response to a basin rainstorm event, the system realizes a complete closed loop from data collection to decision optimization through the above process. For example, when the radar detects the movement of a strong rainfall belt, the system automatically triggers the flood peak prediction process, the digital twin model deduces the future 6-hour flood evolution, the optimization algorithm generates the gate scheduling scheme, and the visualization interface intuitively displays the effects of different decisions, forming a coordinated system of prediction, warning, rehearsal and plan.

[0162] The present application solves the problem of coordination failure of functional chain through systematization. At the module decoupling level, container adaptation process and message queue mechanism are used to realize independent expansion of functional modules. Each data source type corresponds to an independent container instance, such as meteorological data analysis container and remote sensing data processing container running in parallel, and data is exchanged between containers through standardized interfaces. As a data transfer station, the message queue allows the flood peak prediction module and the drought assessment module to communicate asynchronously, and new functional modules can be connected to the system by subscribing to the corresponding message queue topic, eliminating the compilation-level dependency between modules.

[0163] To solve the problem of multi-source data synchronization, a dual-channel redundancy check and intelligent quality evaluation mechanism is designed. The main channel transmits the original data stream, and the auxiliary channel synchronously transmits the lightweight feature hash value. The platform entry end detects data packet loss through real-time hash comparison and triggers retransmission immediately. The variational autoencoder is introduced in the data conversion stage to calculate the reconstruction error, and the generative network detects abnormal patterns and labels each data point with a confidence label. When the timestamps of satellite images and sensor data conflict, the system automatically aligns the time reference and fills in the missing values, ensuring the spatiotemporal consistency of multi-source data.

[0164] To prevent process blocking, a dynamic coordination architecture of parallel processing and federated verification is constructed. The spatiotemporal graph convolutional network processes the topological relationship of the basin to generate flood peak prediction, the Bayesian network and the long short-term memory network are calculated in parallel to generate risk assessment values, and the double-path analysis eliminates serial dependence. When the risk assessment confidence is lower than the threshold, the federated learning mechanism is automatically activated, each edge node updates the model parameters locally, and the coordination server aggregates the gradient to update the global model, avoiding the performance bottleneck caused by centralized processing. The digital twin model loads the physical constraint parameters to deduce multiple schemes in real time, and the multi-objective optimization algorithm iteratively optimizes in parallel, forming a non-blocking pipeline from prediction to plan.

[0165] Specifically in the flood control scene of a river basin, when a typhoon warning is triggered, the system dynamically expands the resources of the flood peak prediction module through the containerized architecture while keeping the drought assessment module running normally. The dual-channel verification mechanism ensures that meteorological radar data and water level monitoring data are synchronized, the intelligent quality assessment module identifies abnormal water level values and marks them with low confidence. The parallel processing procedure synchronously generates flood peak evolution prediction and inundation risk maps, and when multiple model results differ, the federated learning mechanism integrates the local knowledge of various flood control centers to quickly update the model, ensuring that reliable scheduling schemes are output before heavy rain arrives. This design forms a dynamic collaborative closed loop of the functions of forecasting, warning, simulation, and plan, effectively improving emergency response efficiency.

[0166] In a river basin rainstorm event response scenario, the system is deployed in a river flood control command center, connected to meteorological radar stations, water level monitoring stations, soil moisture meters, and remote sensing satellite data sources. When the radar detects the movement of a heavy rain belt, the data acquisition unit starts the dual-channel redundancy verification mechanism: the main channel transmits the radar-based data stream, and the auxiliary channel synchronously sends the data feature hash value. The platform entry end calculates the real-time hash value and compares it with the auxiliary channel summary, and immediately triggers the edge node to retransmit when data packet loss is found, outputting the time-aligned initial data set. The data processing unit parses the radar data through the containerized adaptation process to generate a 1-kilometer grid rainfall intensity matrix, processes satellite images to generate a 30-meter elevation grid, and a variational autoencoder calculates the reconstruction error to mark low-quality data points. An adversarial generative network identifies abnormal water level values and outputs standardized matrices with confidence labels. The data analysis unit stores real-time data with a confidence of 0.9 or above in the cache library, and the reinforcement learning module generates routing strategies based on typhoon warning scenarios to distribute data shards to the flood peak prediction message queue. The parallel processing unit constructs a river basin topology graph, a spatio-temporal graph convolution network aggregates upstream reservoir features to predict flood peak evolution, and a Bayesian network calculates inundation risk probability. When the risk assessment confidence is below the threshold, the federated learning mechanism is activated to update the model parameters. The generation unit loads the river terrain parameters to simulate the flood control scenario, and a multi-objective optimization algorithm iterates 200 times to generate a gate scheduling scheme, outputting a simulation animation to show the flood evolution trajectory. The optimization unit renders a three-dimensional risk heat map through the WebGL engine, the user clicks on the high-risk area to replay the simulation process, adjusts the gate opening parameters to generate a difference vector, encrypts and uploads it to trigger the federated model update, and refreshes the visualization interface in real time.

[0167] Embodiment two of the present application: in the regional drought monitoring scene, the system is deployed in the water resources dispatching center, accesses the soil moisture monitoring network and weather forecast data. When satellite remote sensing shows that the soil humidity is lower than the threshold value, the data acquisition unit synchronously receives the soil moisture data and weather forecast, and the double-channel verification ensures the data integrity. The data processing unit generates a soil humidity grid by analyzing the multispectral remote sensing data through an independent container, containerizes the unified elevation data and sensor data format, identifies abnormal drought spots through a variational autoencoder, detects sensor drift error through a generative adversarial network, and outputs a soil moisture matrix with a confidence label. The data analysis unit stores the data according to the confidence level, and the reinforcement learning module routes the data to the drought assessment pipeline according to the drought warning scene. The parallel processing unit fuses historical drought patterns and real-time data, the long short-term memory network predicts the drought development trend, the Bayesian network calculates the water shortage risk, and the federated learning mechanism integrates the experience of multiple irrigation areas to update the evaluation model. The generation unit loads the reservoir storage capacity curve and water demand parameters, and the multi-objective optimization algorithm generates a water resources dispatching scheme to balance agricultural irrigation and urban water supply demand, and outputs a drought relief plan animation. The optimization unit dynamically displays the reservoir water level change and water delivery route, the user adjusts the water supply proportion to generate parameter differences, the federated aggregation updates the model parameters, and the drought risk assessment results are optimized in real time.

[0168] The technical terms of the present application are explained as follows:

[0169] The variational autoencoder is a generative model that maps input data to a probability distribution in the latent space through an encoder, and then reconstructs the original data by sampling from the distribution through a decoder. In the present application, the variational autoencoder is pre-trained on the standardized spatio-temporal matrix, and the reconstruction error is calculated by comparing the differences between the original matrix and the decoder output matrix. This error value is used to quantify the reliability of each data point, and the larger the reconstruction error, the higher the degree of deviation from the normal distribution, which may include noise or outliers. The reconstruction error calculation result is one of the important indicators for data quality assessment, and provides a basic input for subsequent generation of data confidence labels.

[0170] The adversarial network includes a generator and a discriminator, the generator tries to generate samples similar to the real data distribution, and the discriminator tries to distinguish between real samples and generated samples. The present application uses the discrimination ability of the discriminator in the adversarial network as an anomaly detector to scan the abnormal data points in the spatio-temporal matrix. The discriminator outputs an anomaly score for each data point, and the higher the score, the more likely the point belongs to an abnormal pattern. This mechanism is particularly suitable for detecting outliers caused by sensor failure or environmental mutations, which may not be effectively identified by existing reconstruction error methods.

[0171] Reinforcement learning is a method of learning optimal strategies through interactions between agents and environments to maximize cumulative rewards. In the data routing link of the invention, the reinforcement learning agent takes the data storage location, real-time business scenario state, and data magnitude as state input, and selects the routing target of data shards as the action space. The agent evaluates the effects of different routing actions through a reward function, for example, routing high-priority data to parallel processing flows can obtain positive rewards. Through continuous trial and error learning, the intelligent learning decision module generates an optimal routing strategy table, achieving dynamic intelligent scheduling of data resources.

[0172] Spatiotemporal graph convolution is a deep learning model specifically designed to handle graph-structured data, capable of capturing both spatial topological relationships and temporal dynamic changes. In the invention, the water basin water conservancy facilities are abstracted as a graph structure, with nodes representing reservoirs and other facilities and edges representing water flow direction and distance. The spatiotemporal graph convolution model aggregates the neighborhood features of each node through a graph convolution layer to capture spatial dependencies, and analyzes the variation of each node feature over time through a temporal convolution layer to capture temporal dependencies. This spatiotemporal joint modeling capability effectively predicts the evolution process of the flood peak in the basin, outputting the propagation trajectory and intensity variation of the flood peak in the future period.

[0173] Bayesian networks are a type of probabilistic graphical model that represents conditional dependencies between variables through directed acyclic graphs and performs probabilistic reasoning. In the risk assessment link of the invention, Bayesian networks take meteorological factors, hydrological data, and terrain features as nodes to construct conditional probability relationships between factors and flood risk. When new observation data is input, the network updates the probability distribution of risk events through a probability propagation algorithm and outputs a probabilistic risk assessment value. This probabilistic-based assessment method can explicitly quantify the uncertainty of the prediction, providing more comprehensive reference information for decision-making.

[0174] Long short-term memory networks are a special type of recurrent neural network that effectively solves the gradient vanishing problem in long sequence training through a gating mechanism. The invention uses long short-term memory networks to learn long-term dependencies in historical disaster events and capture the timing patterns of the occurrence and development of flood disasters. The network learns the complete cycle pattern of flood peak formation, evolution, and recession by inputting continuous time series of hydro-meteorological data. When receiving real-time data streams, the network can predict future risk development trends based on the learned timing patterns, complementing the Bayesian network and improving the accuracy of risk assessment.

[0175] Federated learning is a distributed machine learning paradigm that allows multiple participants to train a model locally without exchanging raw data, and only by exchanging model parameter updates to achieve common modeling. In the present invention, when the risk assessment results of multiple terminal nodes have large differences or insufficient confidence, the system activates the federated learning mechanism. Each edge node calculates the model parameter update on the local data, uploads it to the coordination server after encryption for secure aggregation, and generates an improved global model. This mechanism not only protects the data privacy of each node, but also improves the generalization ability and reliability of the model through knowledge sharing.

[0176] Digital twin model constructs a complete digital mapping of physical entities in virtual space by integrating physical laws, geometric features and behavior rules. The digital twin model in the present invention loads physical constraint parameters such as river terrain, reservoir characteristics and hydrodynamic equations to construct a virtual copy of the watershed. When inputting real-time monitoring data or prediction results, the model simulates and deduces based on physical laws to simulate the flood propagation process under different scenarios. This deduction method based on physical mechanism can effectively make up for the shortcomings of pure data-driven models and improve the interpretability and reliability of prediction results.

[0177] Multi-objective optimization algorithm aims to find a set of solutions that can optimize multiple conflicting objectives simultaneously, i.e. Pareto optimal solution set. The present invention uses a multi-objective optimization method based on genetic algorithm to encode the gate opening sequence as a chromosome, with each gene representing the gate opening at a specific time period. The optimization process simultaneously considers multiple objectives such as flood control safety, water supply guarantee and power generation benefit, and continuously evolves the population through selection, crossover and mutation operations. Finally, a set of non-dominated solutions is output to form multiple scheduling schemes for decision makers to choose from, maximizing comprehensive benefits.

[0178] The construction of digital twin model begins with the comprehensive digital mapping of the physical world. The model first loads pre-stored physical constraint parameters of the watershed, including river terrain elevation data, reservoir characteristic curves, hydrodynamic equation coefficients, soil permeability and water conservancy engineering facility structure parameters, and other basic physical rules. Parameters are obtained through multiple channels such as geographic information systems, engineering design drawings and field survey data, and standardized processing forms a unified digital description. The model architecture adopts hierarchical design, with the bottom layer being a geographic spatial database, the middle layer being a physical law engine, and the upper layer being a dynamic simulation core, forming a complete framework from static data to dynamic deduction.

[0179] On the data processing route, the digital twin model receives the analysis results output from the parallel processing unit, including flood peak evolution prediction data and risk assessment indicators. The model first performs data fusion processing to align the real-time monitoring data and prediction results in space and time, eliminating the scale differences and time sequence deviations between different data sources. During the fusion process, Kalman filtering algorithm is used to smooth the data, reducing the influence of measurement noise on the deduction results while maintaining sensitivity to sudden changes. After data fusion, an input data set with consistent space-time reference is generated, providing a high-quality data foundation for subsequent scenario deduction.

[0180] After completing data fusion, the model starts flood control scenario deduction calculation. The deduction process is based on the basic principles of hydrology and hydraulics, using the Saint-Venant equation set to describe the movement law of flood in the river channel, and solving by finite volume method. The model calculates the water level, flow rate, flow velocity and other hydraulic elements of each river section in real time, simulating the propagation process of flood wave in the basin. At the same time, considering the influence of reservoir regulation, gate control, dike constraint and other engineering measures, through coupled solution, the dynamic interaction of engineering dispatching and flood evolution is realized. The deduction time step is adaptively adjusted according to the rapidity of the flood, with smaller step size for rapid flow area to ensure calculation stability, and larger step size for slow flow area to improve calculation efficiency.

[0181] During the deduction process, the model continuously performs real-time correction. By assimilating the latest monitoring data, comparing the differences between the deduction results and the actual observation values, and dynamically adjusting the model parameters to reduce the deduction error. The correction mechanism uses the ensemble Kalman filter method, through parallel deduction and observation data assimilation of multiple parameter sets, continuously optimizing the model state variables and key parameters. This real-time correction mechanism ensures that the digital twin model can accurately track the actual flood evolution process, timely correct the deduction trajectory, and improve the prediction accuracy and reliability.

[0182] After the deduction is completed, the model outputs the flood peak evolution dynamic parameter set, including the water level, flow process line of each control section at different time periods, flood propagation time, maximum inundation range and water depth distribution, and other key indicators. The parameters not only include the expected value, but also the probability distribution information, quantifying the uncertainty of the deduction results. The output data is packaged in a standardized format, including timestamp, spatial location, parameter type, numerical size and confidence level, etc. metadata, for easy parsing and use by downstream modules.

[0183] The digital twin model is tightly coupled with the visualization module, and the inference results directly drive the dynamic presentation of the three-dimensional scene. The model output interface converts abstract numerical results into graphical instructions that can be parsed by the graphics engine, including mesh vertex coordinates, color mapping relationships, animation keyframes, and other visualization elements. Through the application programming interface, the inference parameters are transmitted to the rendering engine in real time, and the flood inundation animation is dynamically generated on the digital elevation model base map, visually displaying the entire process of flood evolution. The visualization process supports multi-level of detail rendering, displaying the global flood situation in the distance and the local flow field details in the near view, meeting the observation needs of different scales.

[0184] User interaction is an important feature of the digital twin model. Through the visualization interface, users can adjust the initial conditions, modify the engineering parameters, and set different scheduling schemes, and the model responds to the input in real time and re-performs the inference calculation. The interactive instructions are captured through the event listening mechanism and are converted into parameter adjustment instructions that the model can recognize after analysis. The model uses an incremental calculation strategy to update the original inference results locally, quickly generating inference results under new schemes, ensuring the real-time and smoothness of the interaction.

[0185] The digital twin model also has the function of replaying the inference process. All historical inference states are recorded in the time series database, forming a complete inference history track. When the user clicks the replay control in the interface, the model extracts the inference state sequence of the specified time period from the database and re-executes the inference process according to the set speed. Auxiliary information such as water level curves and flow process lines is displayed synchronously during the replay process, helping users analyze key nodes and turning points in the inference process and deeply understand the flood evolution law.

[0186] Finally, the digital twin model passes the inference results to the multi-objective optimization module to provide scene input for scheduling decisions. The output data includes complete scene description information, including not only hydrological and hydraulic parameters but also metadata such as constraint condition satisfaction during the inference process and boundary condition change records. The rich information provides sufficient basis for subsequent optimization calculations, ensuring that the generated scheduling scheme not only meets the optimization objectives but also satisfies the physical laws and engineering constraint conditions, forming a complete closed loop from scene inference to decision optimization.

Claims

1. An intelligent supervision method based on a data visualization platform, characterized in that, include: Step 1: Collect meteorological radar data, river water level sensor data, soil moisture monitoring data and satellite remote sensing images, and output the initial dataset with time alignment using dual-channel redundancy verification; Step 1 includes: The main channel transmits the raw data stream to the platform entry point, while the auxiliary channel simultaneously transmits the feature hash value to the platform entry point. The platform entry point receives main channel data and calculates the real-time hash value of the main channel data. Extract the feature hash value transmitted through the secondary channel, compare the real-time hash value with the feature hash value, and if they are inconsistent, trigger the data source node to retransmit the main channel data until the hash values ​​are consistent, and output the initial dataset with integrity markers. Step 2: Convert the initial dataset into a standardized spatiotemporal matrix, including: containerization to adapt to a unified data format, pre-training a variational autoencoder to calculate reconstruction error, adversarial generation to scan outlier data points, and outputting a cleaned spatiotemporal matrix with confidence labels; Step 2 includes: Receive the initial dataset, containerize and adapt the initial dataset, parse the meteorological data to obtain the spatiotemporal matrix of rainfall intensity, perform band calculations on the remote sensing data to obtain the elevation raster, and integrate the spatiotemporal matrix of rainfall intensity and the elevation raster into a unified format matrix. The format-unified matrix is ​​input into the variational autoencoder to calculate the reconstruction error. The format-unified matrix is ​​input into the generative adversarial network to scan abnormal data points and output anomaly scores. The reconstruction error and anomaly scores are merged to generate data confidence labels. Data confidence labels are mapped to a format-unified matrix, and the cleaned spatiotemporal matrix with confidence labels is output. Step 3: Receive the cleaned spatiotemporal matrix, store the matrix data in layers according to the confidence threshold, store the data with a confidence level greater than or equal to the threshold in the high-speed cache library, and store the remaining data in the spatiotemporal index library, monitor the real-time business scenario status and data storage location, generate a routing strategy table through reinforcement learning decision, and extract data from the storage library and distribute it to the message queue according to the routing strategy table. Step 4: Extract data shards from the message queue and input them into parallel processing. Parallel processing includes: spatiotemporal graph convolution calculation to generate flood peak evolution prediction results, Bayesian network and long short-term memory calculation to output risk assessment values, activation of federated learning parameter updates when the ensemble voting confidence is lower than the threshold, and output of analysis results with reliability indicators. Step 4 includes: Receive the data fragments distributed in step 3, and construct a watershed topology graph based on the water conservancy facility relationships in the data fragments. Nodes are water conservancy facility entities, and edges are water flow directions and distances. The watershed topology map is input into the spatiotemporal graph for convolution calculation, and the flood peak evolution prediction results are output. Input the flood peak evolution prediction results into the risk assessment and output the risk assessment value; When the confidence level of the integrated voting for the risk assessment value is lower than the threshold, the federated learning parameter update is activated. Integrate flood peak evolution prediction results, risk assessment values, and federated learning update signals to output analysis results with reliability indicators; Step 5: Receive the analysis results output in Step 4, load the pre-stored physical constraint parameters into the digital twin model, deduce the flood control scenario based on the analysis results, output the dynamic parameters of flood peak evolution, input the dynamic parameters of flood peak evolution into multi-objective optimization calculation, iteratively generate the scheduling instruction set, and generate the pre-show animation data based on the scheduling instruction set; Step 6: Dynamically render and load the scheduling instruction set and pre-rendering animation data, generate an interactive visualization interface and associate it with the causal inference path, and receive user feedback instructions to trigger the update of federated model parameters.

2. The intelligent supervision method based on a data visualization platform according to claim 1, characterized in that, Step 3 includes: After receiving the cleaned spatiotemporal matrix, the data is classified and stored according to the confidence threshold. Data with a confidence level greater than or equal to the threshold is stored in the high-speed cache library and a spatiotemporal index is created. The remaining data is stored in the spatiotemporal index library. Based on the stored data location, read the location status, capture the real-time business scenario code and data volume, and input the location status, business scenario code and data volume into the reinforcement learning decision. The reward function calculates the optimal routing action, generates a routing policy table, retrieves data shards from the repository based on the routing policy table, and distributes the data shards to the message queue.

3. The intelligent supervision method based on a data visualization platform according to claim 1, characterized in that, Step 5 includes: Receive the analysis results output from step 4, load the pre-stored physical constraint parameters into the digital twin model, deduce the flood control scenario based on the analysis results, and output the dynamic parameters of the flood peak evolution; The dynamic parameters of the flood peak evolution are input into multi-objective optimization calculations, and chromosome encoding is used to generate a gate opening sequence. The fitness calculation simultaneously evaluates flood control safety and water supply security, iteratively optimizes the gate opening sequence, outputs the Pareto optimal scheduling scheme, and generates pre-show animation data based on the Pareto optimal scheduling scheme.

4. The intelligent supervision method based on a data visualization platform according to claim 3, characterized in that, Step 5 includes: Receive the initial gate opening sequence, and use tournament selection to screen parent chromosomes; The parent chromosome is input into a simulated binary crossover to generate the offspring chromosome sequence; The weighting coefficients α, β, and γ of the offspring chromosome sequence are dynamically adjusted, the fitness calculation rules are updated, the offspring chromosome sequence is iteratively optimized based on the update rules, and Pareto optimal scheduling data is output. Personnel transfer path data is generated based on Pareto optimal scheduling data.

5. The intelligent supervision method based on a data visualization platform according to claim 1, characterized in that, Step 6 includes: Dynamically render and load the scheduling instruction set and pre-show animation data to generate an interactive visualization interface, and overlay a flood peak risk heat map on the digital elevation model base map; The interactive visual interface receives user click commands on high-risk areas, triggering a replay of the digital twin model's simulation process; The visualization interface displays the planned path animation, which shows the iterative trajectory of the multi-objective optimization algorithm. The interactive visualization interface receives user correction commands, generates a difference vector, encrypts and uploads the difference vector, and triggers the update of the federated model parameters.

6. The intelligent supervision method based on a data visualization platform according to claim 5, characterized in that, Step 6 also includes: Receive encrypted uploaded difference vectors, perform federated aggregation and parsing of the difference vectors, and calculate the global model parameter update amount; Distribute global model parameter updates to independent processing; Independently process and update the parameters of the risk assessment model to generate updated risk assessment values; The updated risk assessment values ​​are input into the dynamic rendering, and the risk heatmap on the visualization interface is refreshed in real time.

7. An intelligent monitoring system based on a data visualization platform, applied to the intelligent monitoring method based on a data visualization platform as described in any one of claims 1 to 6, characterized in that, include: The data acquisition unit is configured to acquire meteorological radar data, river water level sensor data, soil moisture monitoring data and satellite remote sensing images, and perform dual-channel redundancy check to output time-aligned initial datasets. The data processing unit is configured to convert the initial dataset into a standardized spatiotemporal matrix, including: containerizing and adapting to a unified data format, calculating the reconstruction error using a pre-trained variational autoencoder, generating adversarial data to scan outlier data points, and outputting a cleaned spatiotemporal matrix with confidence labels. The data analysis unit is configured to receive the cleaned spatiotemporal matrix, store the matrix data in layers according to the confidence threshold, store data with a confidence level greater than or equal to the threshold in a high-speed cache library, store historical data in a spatiotemporal index library, monitor the real-time business scenario status and data storage location, generate a routing strategy table through reinforcement learning decision-making, and extract data from the storage library and distribute it to the message queue according to the routing strategy table. The parallel processing unit is configured to extract data shards from the message queue and input them into parallel processing. The parallel processing includes: spatiotemporal graph convolution calculation to generate flood peak evolution prediction results, Bayesian network and long short-term memory calculation to output risk assessment values, activation of federated learning parameter updates when the ensemble voting confidence is lower than the threshold, and output of analysis results with reliability indicators. The generation unit is configured to receive the analysis results output by the parallel processing unit, load the pre-stored physical constraint parameters into the digital twin model, deduce the flood control scenario based on the analysis results, output the dynamic parameters of the flood peak evolution, input the dynamic parameters of the flood peak evolution into the multi-objective optimization calculation, iteratively generate the scheduling instruction set, and generate the pre-show animation data based on the scheduling instruction set. The optimization unit is configured to dynamically render and load the scheduling instruction set and pre-rendering animation data, generate an interactive visualization interface and associate it with the causal inference path, and receive user feedback instructions to trigger the update of federated model parameters.

Images