Real-time interaction and intelligent management method for multi-dimensional data space

Abstract

The invention discloses a real-time interaction and intelligent management method for a multi-dimensional data space, and relates to the technical field of data interaction management, and the method comprises the following steps: I, collecting multi-source data, including structured data, unstructured data and spatio-temporal data, in real time through a diversified collection terminal, and dynamically adjusting the sampling frequency based on a data entropy value, carrying out cross-device space-time synchronization calibration and edge end lightweight processing on the acquired data; iI, preprocessing the data flow in real time, fusing the multi-modal data, and the like. According to the method, the sampling frequency is dynamically adjusted based on the data entropy value, and in combination with edge end lightweight processing, invalid data transmission is reduced, and the key data acquisition efficiency is improved; hot data is stored in a time sequence database and a memory database, real-time monitoring of an emergency scene and second-level response of emergency command are supported, dynamic capacity expansion and shrinkage of computing resources are matched, edge-cloud end collaboratively distributes tasks, network delay is reduced, and system stability and response speed in a high-concurrency scene are guaranteed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data interaction management, in particular to a real-time interaction and intelligent management method of multi-dimensional data space. BACKGROUND

[0002] Multi-dimensional data space refers to organizing data in a structured space composed of multiple independent dimensions to support efficient query, analysis and visualization. With the rapid development of Internet of Things, smart city and other fields, data presents the characteristics of multi-source, heterogeneous and dynamic, forming a complex multi-dimensional data space.

[0003] At present, the traditional multi-dimensional data management method lacks real-time performance and cannot meet the real-time interaction requirements of emergency command and other scenarios. Multi-modal data is stored independently, lacks a unified space-time reference and semantic association, and it is difficult to mine cross-dimensional potential value. Moreover, it relies on manual experience for data query and decision making, and cannot realize automated anomaly detection, trend prediction and resource scheduling. SUMMARY

[0004] The purpose of the present application is to provide a real-time interaction and intelligent management method of multi-dimensional data space to solve the problems of lack of real-time performance, inability to meet the real-time interaction requirements of emergency command and other scenarios, independent storage of multi-modal data, lack of unified space-time reference and semantic association, difficulty in mining cross-dimensional potential value, and reliance on manual experience for data query and decision making, and inability to realize automated anomaly detection, trend prediction and resource scheduling.

[0005] To achieve the above purpose, the present application provides the following technical scheme: a real-time interaction and intelligent management method of multi-dimensional data space, comprising the following steps:

[0006] I. Real-time collection of multi-source data including structured data, unstructured data and space-time data by diversified collection terminals, dynamic adjustment of sampling frequency based on data entropy value, cross-device space-time synchronization calibration and edge lightweight processing of collected data;

[0007] II. Real-time data stream preprocessing and fusion of multi-modal data, storage of multi-dimensional data through hierarchical storage architecture and construction of multi-dimensional index;

[0008] III. Real-time anomaly detection and prediction of real-time data stream using Isolation Forest algorithm, generation of automated strategy through reinforcement learning algorithm, direct execution of high confidence decision, and manual verification of low confidence scheme;

[0009] IV. Construction of digital twin scene based on Unity three-dimensional engine, real-time mapping of physical object state, support for 360° roaming and detail scaling interactive operation; space-time data is displayed through dynamic view, supporting viewing of historical evolution through time axis sliding;

[0010] V. Dynamically adjust the interface layout based on user operation habits and permissions, actively push notifications containing data details, impact analysis, and processing suggestions to the mobile terminal through geo-fencing, time rules, or abnormal event triggers;

[0011] VI. Implement hierarchical decision-making for different scenarios, record the execution results and feedback to the model, and trigger incremental learning of the model through error analysis;

[0012] VII. Perform resource elasticity scheduling, automatically scale up or down according to real-time load, and collaboratively allocate tasks between edge cloud and end.

[0013] Preferably, in step I, the dynamic adjustment of sampling frequency based on data entropy value is achieved by calculating the entropy value of data in the sliding window in real time, dynamically adjusting the sampling frequency of the sensor or device, and implementing an adaptive strategy of high-value data high-frequency sampling and low-value data low-frequency sampling, including the following steps:

[0014] ① Divide the continuous numerical data into K equidistant intervals, and in the time window [t-W, t], W is the size of the sliding window, count the number of data occurrences c k in each interval, and calculate the frequency

[0015] ② Calculate the data entropy value H t in the window:

[0016] ③ When the dynamic threshold is exceeded and the entropy value is continuously greater than the threshold for N windows, adjust the sampling frequency;

[0017] ④ Adjust the sampling frequency, the adjustment strategy is as follows:

[0018] Define the sampling frequency set F = {f1, f2, f3}, corresponding to the sampling interval, f1 = 1 Hz, f2 = 0.1 Hz, f3 = 0.01 Hz;

[0019] If the current frequency is f1: entropy value > 1.5: maintain f1; entropy value < 0.5: reduce to f2;

[0020] If the current frequency is f2: entropy value > 1.5: increase to f1; entropy value < 0.5: reduce to f3;

[0021] If the current frequency is f3: entropy value > 1.5: increase to f2; - entropy value < 0.5: maintain f3;

[0022] Entropy value between 0.5 and 1.5: maintain the current frequency;

[0023] ⑤ Each time a new data point is collected, the sliding window is updated and the entropy value H t is recalculated.

[0024] Preferably, in step II, the real-time data stream preprocessing includes data stream standardization and feature enhancement;

[0025] Data stream standardization includes data alignment and normalization processing, data alignment is achieved by building a distributed stream processing engine to split data streams, using watermark processing for out-of-order data within 2 seconds to ensure timestamp consistency;

[0026] Feature enhancement includes time series feature extraction and spatial feature fusion, time series feature extraction generates derived features by calculating statistics within a 1-minute window, and spatial feature fusion generates spatial weight features by combining device location information through a Gaussian kernel function.

[0027] Preferably, in step II, the multi-dimensional index construction includes spatial index and business index;

[0028] Spatial index: R-tree index is established for geographic data to support regional range query; TemporalR-Tree index structure is used for spatio-temporal index to support spatial range retrieval within a time window;

[0029] Business index: based on the snowflake model, the dimension table is associated with the fact table to provide fast drill analysis.

[0030] Preferably, in step III, the real-time data stream real-time anomaly detection and prediction includes the following steps:

[0031] A1, based on historical sample verification and calculation of time-consuming fixed construction of 100 isolated trees, each isolated tree randomly selects 1 / 3 features and samples to construct an isolated forest model;

[0032] A2, calculate the average path length h(x) of the sample, the anomaly score s(x) = 2 -h(x) / c(n) , where c(n) is the harmonic number, if the anomaly score s(x) > 0.8, it is determined to be abnormal;

[0033] A3, based on the past 24 hours of data, use the long short-term memory neural network LSTM prediction model to output the future 1-hour prediction value;

[0034] A4, generate a 95% confidence interval through Monte Carlo simulation, and trigger trend warning when the prediction value exceeds the interval.

[0035] Preferably, in step V, the active push rule is as follows:

[0036] Geofencing trigger: fence creation draws the area through GIS tools, and coordinates are converted to the coordinate system; positioning detection mobile terminal acquires real-time location and uses R-tree spatial index to determine whether to enter the fence;

[0037] Time rule trigger: periodic rule: push energy efficiency analysis of the previous day at a preset time every day; time period preference: adjust the push time according to the historical active period of the user;

[0038] Abnormal event trigger: define abnormal level: first-class abnormality (emergency): push immediately; second-class abnormality (warning): system crash, connection timeout, push within 5 minutes; correlation analysis: the push content contains the abnormal influence range and the historical similar event processing scheme.

[0039] Preferably, in step VI, the decision level division and trigger condition are as follows:

[0040] Full-automatic decision: model confidence is greater than or equal to 90%, and the risk level is low;

[0041] Man-machine co-decision: model confidence is 60% to 90%, and the risk level is medium;

[0042] Manual decision: model confidence is less than 60%, and the risk level is high.

[0043] Preferably, in step VI, the incremental learning update rule is: θ t+1 = θ t - η · ▽ J ( θ t );

[0044] In the formula: θ t is the model parameter of the tth iteration, η is the learning rate, and ▽ J ( θ t ) is the gradient of the loss function J to the parameter θ t .

[0045] Preferably, in step VII, the cooperative scheduling mechanism is as follows:

[0046] Static allocation strategy: edge priority predefines a list of tasks that can be processed by the edge; tasks that cannot be processed by the edge are automatically routed to the cloud;

[0047] Dynamic load-aware allocation: edge nodes report real-time computing power status; cloud scheduler matches according to task computing power demand;

[0048] Network outage autonomy: edge nodes store local task queues, execute according to priority when the network is interrupted, and synchronize the results to the cloud through the transmission protocol MQTT after recovery.

[0049] Compared with the prior art, the present application has the following advantages:

[0050] 1. In this invention, the sampling frequency is dynamically adjusted based on the data entropy value, combined with lightweight processing at the edge, to reduce invalid data transmission and improve the efficiency of key data acquisition; hot data is stored in a time-series database and an in-memory database, supporting real-time monitoring and second-level response of emergency command in emergency scenarios. With the dynamic scaling of computing resources and the collaborative allocation of tasks between the edge and the cloud, network latency is reduced, ensuring system stability and response speed in high-concurrency scenarios.

[0051] 2. In this invention, text keywords are mapped to structured data entities through entity linking technology, realizing semantic alignment between structured and unstructured data. Spatiotemporal data is rasterized into discrete grid cells, establishing a unified spatiotemporal benchmark, eliminating spatiotemporal deviations of multi-source data, and supporting fast spatiotemporal range queries with multidimensional indexes. The snowflake model constructs business indexes to realize multidimensional drill-down analysis, and hot and cold data are stored in layers, which not only meets the needs of real-time interaction, but also facilitates cross-dimensional historical trend mining.

[0052] 3. In this invention, the automatic replacement of manual inspection with anomaly detection and prediction exposes potential risks in advance, reduces the time for hazard investigation, strengthens learning-driven hierarchical decision-making, and combines the PPO reinforcement learning algorithm iterative optimization strategy to realize the digitization of human experience and improve decision-making efficiency and accuracy.

[0053] 4. In this invention, by automatically triggering model parameter updates based on root mean square error, the system adapts to dynamic environmental changes, avoids detection failures caused by data distribution drift, and continuously optimizes automated decision-making capabilities. Attached Figure Description

[0054] Figure 1 This is a flowchart of a real-time interaction and intelligent management method for a multidimensional data space according to the present invention;

[0055] Figure 2 This is a resource scheduling flowchart for a real-time interaction and intelligent management method for a multi-dimensional data space according to the present invention. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] Please see Figure 1 and Figure 2 This invention provides a technical solution: a method for real-time interaction and intelligent management of a multi-dimensional data space, comprising the following steps:

[0058] Step one, real-time collection of multi-source data through diversified collection terminals, including structured data, unstructured data and spatio-temporal data, dynamic adjustment of sampling frequency based on data entropy value, cross-device spatio-temporal synchronization calibration and edge lightweight processing of collected data, through distributed stream processing engine and watermark technology, ensure the uniformity of multi-device data timestamp, eliminate the interference of spatio-temporal deviation on the analysis result, complete the data processing at the edge, reduce the cloud computing load, reduce the network transmission cost, and improve the overall system response speed;

[0059] Among them, the dynamic adjustment of sampling frequency based on data entropy value adjusts the sampling frequency of sensors or devices by real-time calculation of the entropy value of the data in the sliding window, realizes the adaptive strategy of high-value data high-frequency sampling and low-value data low-frequency sampling, can reduce the redundant collection of low-value data, reduce the energy consumption and data transmission pressure of the device, prolong the service life of the hardware, maintain high-frequency sampling for high fluctuation data, ensure that key information is not lost, and improve the accuracy of subsequent analysis, including the following steps:

[0060] ①Divide the continuous numerical data into K equidistant intervals, count the number of data occurrences c in each interval in the time window [t-W, t], and W is the size of the sliding window k , calculate the frequency

[0061] ②Calculate the data entropy value H in the window t :

[0062] ③When the dynamic threshold is determined and the entropy value exceeds the threshold for N consecutive windows, adjust the sampling frequency;

[0063] ④Adjust the sampling frequency, the adjustment strategy is as follows:

[0064] Define the sampling frequency set F = {f1, f2, f3}, corresponding to the sampling interval, f1 = 1 Hz, f2 = 0.1 Hz, f3 = 0.01 Hz;

[0065] If the current frequency is f1: entropy value > 1.5: maintain f1; entropy value < 0.5: reduce to f2;

[0066] If the current frequency is f2: entropy value > 1.5: increase to f1; entropy value < 0.5: reduce to f3;

[0067] If the current frequency is f3: entropy value > 1.5: increase to f2; - entropy value < 0.5: maintain f3;

[0068] Entropy value between 0.5-1.5: maintain the current frequency;

[0069] ⑤Each time a new data point is collected, the sliding window is updated and the entropy value H is recalculated t .

[0070] Step two, real-time data stream preprocessing and multi-modal data fusion, multi-dimensional data storage through hierarchical storage architecture and multi-dimensional index construction;

[0071] Among them, real-time data stream preprocessing includes data stream standardization and feature enhancement, data stream standardization includes data alignment and normalization processing, data alignment realizes data shunting through the establishment of a distributed stream processing engine, uses watermark processing for 2 seconds of out-of-order data, ensures timestamp consistency, and processes out-of-order data delay ≤50ms; Feature enhancement includes time series feature extraction and spatial feature fusion, time series feature extraction generates derived features by calculating statistical quantities (mean, standard deviation, peak and slope) in a 1-minute sliding window, and spatial feature fusion generates spatial weight features by combining device location information through a Gaussian kernel function;

[0072] Multi-modal data fusion:

[0073] Structured and unstructured fusion: mapping text keywords to structured data entities through entity linking technology;

[0074] Spatial data rasterization: converting continuous spatiotemporal coordinates into discrete grids to generate spatiotemporal cube base units.

[0075] Hierarchical storage includes hot data and cold data:

[0076] Hot data: using a time series database to store real-time data for nearly 30 days, supporting millisecond-level time series queries; High-frequency access dimension tables are stored in an in-memory database to meet real-time monitoring, early warning, and other low-latency requirements;

[0077] Cold data: historical data is archived to a distributed file system or object storage, with a compression ratio of 1:10-1:20 through storage compression, reducing storage costs, and accelerating batch analysis through columnar storage.

[0078] Multi-dimensional index construction includes spatial index and business index:

[0079] Spatial index: R-tree index is established for geographic data to support regional range queries; TemporalR-Tree index structure is used for spatiotemporal index, supporting spatial range retrieval within a time window, improving spatial and spatiotemporal range query efficiency;

[0080] Business index: based on the snowflake model, the association between dimension tables and fact tables is constructed to support fast drill analysis, facilitate multi-dimensional drill analysis, and improve data analysis flexibility.

[0081] Step three, real-time anomaly detection and prediction of real-time data stream, isolation forest is sensitive to a small number of abnormal samples, suitable for real-time detection of sudden anomalies; Long short-term memory neural network (LSTM) captures long-term dependencies of time series data and predicts future trends; Monte Carlo simulation generates confidence intervals to reduce false positive rates of single models. The combination of the three realizes the dual guarantee of real-time detection and trend warning, improves accuracy, and generates automatic strategies through reinforcement learning intelligent algorithm. High confidence decisions are directly executed, and low confidence schemes are pushed for manual verification to improve decision efficiency;

[0082] The real-time anomaly detection and prediction of the real-time data stream comprises the following steps:

[0083] A1, based on historical sample verification and calculation of time-consuming fixed construction of 100 isolated trees, 1 / 3 features and samples are randomly selected for each tree to construct an isolation forest model;

[0084] A2, calculate the average path length h(x) of the sample, and the abnormal score s(x) = 2 -h(x) / c(n) , where c(n) is the harmonic number, and if the abnormal score s(x) > 0.8, the anomaly is determined;

[0085] A3, based on the past 24 hours of data, use the LSTM prediction model to output the predicted value of the next 1 hour, and expose the potential risk in advance to give the maintenance personnel processing time and reduce the loss;

[0086] The past 24 hours of data input into the LSTM need to be excluded from the abnormal values determined in step A2, and the missing values are filled by linear interpolation;

[0087] Prediction output: the predicted value of each time step and the corresponding error are output according to the sampling frequency.

[0088] A4, generate a 95% confidence interval through Monte Carlo simulation, and trigger a trend warning when the predicted value exceeds the interval.

[0089] Decision grading mechanism:

[0090] Full automatic decision (confidence ≥ 90%), reduce manual intervention delay, improve response speed:

[0091] Execution scope: routine maintenance, resource scheduling and parameter adjustment;

[0092] Execution process: anomaly detection → strategy generation → automatic issuance of instructions → execution result feedback.

[0093] Manual verification decision (confidence < 90%):

[0094] Push content: abnormal details, recommended strategy (with risk score), impact analysis verification method: mobile app push, SMS notification, manual response time ≤10 minutes.

[0095] Step four, based on Unity three-dimensional engine to build digital twin scene, real-time mapping object state, support 360° roaming and detail zoom; spatiotemporal data through dynamic view display, support time axis sliding to view historical evolution, spatiotemporal tracing facilitates fault review and root cause analysis; spatiotemporal cube visualization helps users quickly locate the spatiotemporal area where the anomaly occurred;

[0096] The dynamic view development includes a timeline component, historical state reproduction, and spatiotemporal cube visualization;

[0097] The timeline component includes UI design and interaction logic. The UI design includes a horizontal sliding bar ranging from the earliest / latest timestamp, with a current time label displayed. The interaction logic triggers events when sliding, loads the scene state corresponding to the timestamp, and improves the efficiency of multi-dimensional data correlation analysis.

[0098] Historical state reproduction includes device state recovery: iterate through all devices, update properties such as position and material based on the most recent snapshot (forward / backward interpolation) according to the timestamp;

[0099] Spatiotemporal cube visualization:

[0100] Three-dimensional timeline: add a model (transparent cube, X / Y axis as spatial coordinates, Z axis as time) to the corner of the scene, and use different colored cubes to represent abnormal events.

[0101] Interactive functions: click on the dynamic view cube to jump to the corresponding time point in the scene, and support spatiotemporal range selection.

[0102] Step five, dynamically adjust the interface layout based on user operation habits and permissions, actively push notifications containing data details, impact analysis, and processing suggestions to the mobile terminal triggered by geographic fences, time rules, or abnormal events, ensuring that information reaches the relevant personnel at the most appropriate time and avoiding information overload.

[0103] The active push rules are as follows:

[0104] Geofence trigger: fence creation: draw the area on the GIS map tool, convert the coordinates to the WGS84 coordinate system; positioning detection: the mobile terminal obtains the real-time position and uses R-tree spatial indexing to determine whether it has entered the fence.

[0105] Time rule trigger: periodic rule: push the energy efficiency analysis of the previous day at the preset time every day; time period preference: adjust the push time based on the user's historical active time period.

[0106] Abnormal event trigger: define abnormal level: first-level abnormality (emergency): push immediately; second-level abnormality (warning): system crash or connection timeout, push within 5 minutes; correlation analysis: push content includes abnormal impact range, historical similar event handling scheme;

[0107] Step six, hierarchical decision-making is performed for different scenarios, and the execution results are recorded and fed back to the model. Incremental learning of the model is triggered through error analysis. Automatic decision-making covers low-risk and high-frequency scenarios. Man-machine co-decision-making handles medium-risk scenarios, and manual decision-making controls high-risk scenarios. Ensure that the decision-making strategy matches the risk level, improve system robustness, evaluate the model performance through root mean square error (RMSE), trigger incremental learning, and dynamically optimize the model with new data to avoid performance degradation due to changes in data distribution.

[0108] Among them, the decision level is divided and the trigger condition is:

[0109] Automatic decision: model confidence ≥ 90%, low risk level;

[0110] Man-machine co-decision: model confidence 60% ~ 90%, medium risk level;

[0111] Manual decision: model confidence < 60%, high risk level.

[0112] The root mean square error calculation formula is as follows:

[0113]

[0114] In the formula: y i is the true value, is the model prediction value, and n is the sample size. When the root mean square error increases by ≥ 10% compared with the previous round, incremental learning is triggered, and historical samples not covered are supplemented.

[0115] Incremental learning update rule: θ t+1 = θ t - η ·▽J(θ t );

[0116] In the formula: θ t is the model parameter of the tth iteration, η is the learning rate, and▽J(θ t ) is the gradient of the loss function J with respect to the parameter θ t .

[0117] Step seven, perform resource elasticity scheduling. Automatically scale up or down according to real-time load, edge-cloud collaborative task allocation, dynamically adjust computing resources with real-time load, avoid resource idling or overload, improve system throughput and cost performance, edge processing real-time tasks, cloud responsible for large-scale modeling and long-term storage, reduce network delay and bandwidth consumption;

[0118] In which, the cooperative scheduling rules are as follows:

[0119] Static allocation strategy: predefine the list of tasks that edge can handle, edge node prefer to handle localized tasks; tasks that edge cannot handle are automatically routed to the cloud, and the cloud handles global tasks;

[0120] Dynamic load-aware allocation: edge nodes report computing power status in real time, when the CPU utilization of edge node is greater than or equal to 80%, new tasks are shunted to the cloud; the cloud scheduler matches according to the task computing power demand, when the cloud load is less than or equal to 30%, actively undertake edge overflow tasks, algorithm: matching score = task demand resource node remaining resource x (1-node current load);

[0121] Network outage autonomy: edge node stores local task queue priority rules, executes according to priority when network is interrupted, and synchronizes incremental results to the cloud through MQTT transmission protocol after network recovery, only transmits new data during network outage.

[0122] In this method, the sampling frequency is dynamically adjusted by means of data entropy value, according to the strategy of high-value data high-frequency sampling and low-value data low-frequency sampling, which can reduce data redundancy while ensuring data effectiveness, and using cross-device space-time synchronization calibration and edge lightweight processing can help improve the space-time consistency of data and reduce the computing pressure of the cloud. The use of entity linking realizes the fusion of structured data and unstructured data, and through the grid processing of space-time data, a unified data model can be constructed, laying the foundation for subsequent analysis and visualization. The hierarchical storage strategy: hot data is stored in time series database and memory database, cold data is archived to distributed file system, and columnar storage is used for compression. This cold and hot data separation method can not only meet the performance requirements of real-time query, but also reduce storage costs, which conforms to the best practice of big data storage.

[0123] Isolation forest algorithm is used for anomaly detection, long short-term memory neural network model is used for prediction, and Monte Carlo simulation is used to generate confidence interval. This multi-algorithm fusion method can improve the accuracy and forward-looking of anomaly detection. The state space, action space and reward function are explicitly modeled, and the PPO reinforcement learning algorithm is used to train the agent. At the same time, a hierarchical mechanism of automatic decision and manual verification decision is set, which guarantees the efficiency of decision-making while taking into account the reliability of decision-making.

[0124] The digital twin scene is built based on the Unity engine, supports 360-degree roaming, dynamic view display of space-time data and time axis interaction and the like, can provide an intuitive monitoring and analysis interface for users, pushes through geographic fences, time rules and abnormal event triggering, and performs hierarchical push on the pushed content, can realize accurate notification, improves the response efficiency of users, performs incremental learning of the model according to the root mean square error, implements hierarchical decision-making in combination with confidence and risk levels, can make the model continuously adapt to the change of data, improve the intelligent level of the system, uses elastic expansion and contraction of resources and edge-cloud collaborative task allocation, can ensure that the system can stably run under different load conditions, meets the elastic design principle of cloud computing.

[0125] Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements to part of the technical features, any modification, equivalent replacement, improvement and the like made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for real-time interaction and intelligent management of multi-dimensional data space, characterized in that, Comprise the following steps: I. Real-time collection of multi-source data, including structured data, unstructured data and spatio-temporal data, through diversified collection terminals, dynamic adjustment of sampling frequency based on data entropy value, cross-device spatio-temporal synchronization calibration and edge lightweight processing of collected data; II. Real-time data stream preprocessing and fusion of multi-modal data, storage of multi-dimensional data through hierarchical storage architecture and construction of multi-dimensional index; III. Real-time anomaly detection and prediction of real-time data stream using isolation forest algorithm, generation of automatic strategy through reinforcement learning algorithm, direct execution of high confidence decisions, and manual verification of low confidence solutions; IV. Construction of digital twin scene based on three-dimensional engine, real-time mapping of physical object state, support for 360° roaming and detail scaling interactive operation; spatio-temporal data is displayed through dynamic view, and history evolution can be viewed through time axis sliding; V. Dynamic adjustment of interface layout based on user operation habit and permission, active push of notifications containing data details, impact analysis and processing suggestions to mobile terminal through geographic fence, time rule or abnormal event triggering; VI. Hierarchical decision making for different scenarios, recording of execution results and feedback to model, triggering of model incremental learning through error analysis; VII. Resource elasticity scheduling, automatic scaling of capacity according to real-time load, and collaborative allocation of tasks by edge cloud.

2. The method of Claim 1, wherein, In step I, the dynamic adjustment of sampling frequency based on data entropy value is achieved by real-time calculation of the entropy value of the data in the sliding window, dynamic adjustment of the sampling frequency of the sensor or device, comprising the following steps: ①Divide the continuous numerical data into K equidistant intervals, and count the number of data occurrences c in each interval in the time window [t-W, t], where W is the size of the sliding window k , and calculate the frequency ii) calculating the data entropy value H in the window t : ③ Dynamic threshold judgment and entropy value exceeding threshold value for continuous N windows, adjusting the sampling frequency; ④ Adjust the sampling frequency, the adjustment strategy is as follows: Define the sampling frequency set F={f1,f2,f3}, corresponding to the sampling interval, f1=1Hz, f2=0.1Hz, f3=0.01Hz; If the current frequency is f1: entropy value>1.5: maintain f1; entropy value<0.5: reduce to f2; If the current frequency is f2: entropy value>1.5: increase to f1; entropy value<0.5: reduce to f3; If the current frequency is f3: entropy value>1.5: increase to f2; -entropy value<0.5: maintain f3; Entropy value between 0.5 and 1.5: maintain the current frequency; ⑤ Each new data point collected, the sliding window is updated, and the entropy value H is recalculated t .

3. The method of claim 1, wherein, In step II, the data stream preprocessing includes data stream standardization and feature enhancement; Data stream standardization includes data alignment and normalization processing, data alignment is realized by building a distributed stream processing engine to realize data shunting, using watermark processing for out-of-order data within 2 seconds to ensure timestamp consistency; Feature enhancement includes time series feature extraction and spatial feature fusion, time series feature extraction generates derived features by calculating statistical quantities within a 1-minute window, and spatial feature fusion generates spatial weight features by combining device location information through a Gaussian kernel function.

4. The method of Claim 3, wherein, In step II, the multi-dimensional index construction includes spatial index and business index; Spatial index: R-tree index is established for geographic data to support regional range query; spatio-temporal index adopts balanced tree index structure to support spatial range retrieval within time window; Business index: build dimension table and fact table association based on snowflake model.

5. The method of Claim 1, wherein, In step III, the real-time data stream real-time anomaly detection and prediction includes the following steps: A1, based on historical sample verification and calculation time-consuming fixed construction of 100 isolated trees, each tree randomly selects 1 / 3 features and samples to construct an isolated forest model; A2, calculate sample average path length h(x), anomaly score s(x) = 2 -h(x) / c(n) where c(n) is the harmonic number, and if anomaly score s(x) > 0.8 then determine anomaly; A3, based on the past 24 hours data using long short-term memory neural network prediction model output future 1 hour prediction value; A4, through Monte Carlo simulation generates 95% confidence interval, when the prediction value exceeds the interval, trigger trend warning.

6. The method of Claim 1, wherein, In step V, the active push rules are as follows: Geofencing trigger: fence creation through map tool to draw area, coordinate conversion to coordinate system; positioning detection mobile terminal to obtain real-time position, using R-tree spatial index to determine whether to enter the fence; Time rule trigger: periodic rule: daily at the preset time to push the energy efficiency analysis of the previous day; time preference: according to the user's historical active period to adjust the push time; Abnormal event trigger: define abnormal level, execute hierarchical push operation; Push content includes abnormal influence range and historical similar event processing scheme.

7. The method of Claim 1, wherein, In step VI, the decision level division and trigger condition are as follows: Full automatic decision: model confidence ≥ 90%, risk level low; Man-machine co-decision: model confidence 60% ~ 90%, risk level medium; Manual decision: model confidence < 60%, risk level high.

8. The method of Claim 1, wherein, In step VI, the incremental learning update rule formula is: θ t+1 = θ t - η · ▽J(θ t ); where θ t is the model parameters for the tth iteration, η is the learning rate, and ∇J(θ t ) is the gradient of the loss function J with respect to the parameters θ t .

9. The method of Claim 1, wherein, In step VII, the collaborative scheduling mechanism is as follows: Static allocation strategy: edge priority predefines the task list that can be processed by edge; tasks that cannot be processed by edge are automatically routed to cloud; Dynamic load sensing allocation: edge node reports real-time computing power state; cloud scheduler matches according to task computing power demand; Network disruption autonomy: edge node stores local task queue, network interruption executes according to priority, after recovery, synchronizes the result to the cloud through the transmission protocol.

Images