Data requirement dynamic management method, system and device based on scenario modeling

Abstract

The application discloses a data requirement dynamic management method, system and equipment based on scenario modeling, relates to the technical field of data requirement dynamic management, and comprises the following steps: constructing a plurality of heterogeneous data nodes, and performing initial weak connection association; obtaining a multidimensional scene correlation matrix; performing hierarchical semantic analysis, and determining a dynamic scene situation; reconstructing the multidimensional scene correlation matrix into a dynamic correlation matrix; reconstructing a scene contraction topology; performing cross-node correlation fusion to obtain integrated scenario data; and performing scenario aggregation and screen display. The application solves the technical problems of the prior art, such as fixed correlation of multi-source heterogeneous sensing data being unable to adapt to a dynamic scene, large data redundancy calculation, low scene research and judgment efficiency, and poor visual display effect, achieves adaptive topology simplification and cross-node fusion visual output of multi-modal sensing data according to a dynamic scene, and improves the real-time performance and display intuitiveness of multi-source heterogeneous data scene research and judgment.

Description

Technical Field

[0001] This invention relates to the field of dynamic data demand management technology, specifically to a method, system, and device for dynamic data demand management based on scenario-based modeling. Background Technology

[0002] Existing multimodal business perception systems typically use fixed static network topologies to perform multi-source heterogeneous data association operations. They cannot dynamically adjust node association weights and topology structures according to real-time scenario situations. All types of perception data streams participate in the calculation without distinction, resulting in high data redundancy and high computing power consumption. At the same time, the relationships between nodes are fixed and unchanging, and weakly correlated data continues to occupy computing resources. Data fusion lacks scenario-differentiated weight constraints, resulting in low scenario matching accuracy and high judgment delay. In addition, data processing and visualization are disconnected. The integrated multi-dimensional scenario data cannot be uniformly output according to the corresponding display templates for the current dynamic scenario. The information displayed on the screen is messy and lacks hierarchy, making it difficult for staff to quickly and accurately conduct scenario situation analysis and judgment.

[0003] Existing technologies suffer from technical problems such as fixed associations of multi-source heterogeneous sensing data, inability to adapt to dynamic scenarios, large computational load due to data redundancy, low efficiency in scenario analysis, and poor visualization effects. Summary of the Invention

[0004] This application provides a method, system, and device for dynamic management of data requirements based on scenario-based modeling, which addresses the technical problems of existing technologies, such as fixed associations of multi-source heterogeneous sensing data that cannot adapt to dynamic scenarios, large amounts of data redundancy and computational burden, low efficiency in scenario analysis, and poor visualization effects.

[0005] In view of the above problems, this application provides a method, system and device for dynamic management of data requirements based on scenario-based modeling.

[0006] A first aspect of this application provides a method for dynamic management of data requirements based on scenario-based modeling, the method comprising: Multiple heterogeneous data nodes are constructed based on data type mapping in the multimodal perception layer, and initial weak connections are established. Scene association modeling is performed on the multiple heterogeneous data nodes to obtain a multidimensional scene association matrix. Hierarchical semantic analysis is performed on multiple real-time perception data streams of the multiple heterogeneous data nodes to determine the dynamic scene situation. The multidimensional scene association matrix is ​​reconstructed into a dynamic association matrix based on the dynamic scene situation. K key business nodes and K sets of strong association relationships are extracted from the dynamic association matrix, and topology pruning and strengthening of weakly associated nodes are performed to reconstruct the scene shrinking topology. Using the scene shrinking topology as a data interaction guide, cross-node association and fusion are performed on the K business perception data streams of the K key business nodes to obtain scene-integrated data. Scene presentation templates are called according to the dynamic scene situation, and the scene-integrated data is mapped into scene view data for scene-integrated screen display.

[0007] A second aspect of this application provides a dynamic data demand management system based on scenario-based modeling, the system comprising: The system comprises the following modules: an initial weak connection module for mapping and constructing multiple heterogeneous data nodes in the multimodal perception layer based on data type, and performing initial weak connection associations; a multidimensional scene association matrix acquisition module for performing scene association modeling on the multiple heterogeneous data nodes to obtain a multidimensional scene association matrix; a dynamic scene situation determination module for performing hierarchical semantic analysis on multiple real-time perception data streams of the multiple heterogeneous data nodes to determine the dynamic scene situation; a dynamic association matrix acquisition module for reconstructing the multidimensional scene association matrix into a dynamic association matrix based on the dynamic scene situation; a scene shrinking topology reconstruction module for extracting K key business nodes and K sets of strong association relationships from the dynamic association matrix, performing topology pruning and strengthening of weakly associated nodes, and reconstructing the scene shrinking topology; a cross-node association fusion module for using the scene shrinking topology as a data interaction guide to perform cross-node association fusion on the K business perception data streams of the K key business nodes to obtain scene-integrated data; and a scene-integrated aggregation display module for calling scene presentation templates based on the dynamic scene situation, mapping the scene-integrated data into scene view data, and performing scene-integrated aggregation display.

[0008] A third aspect of this application provides an electronic device comprising: a processor; and a memory for storing processor-executable instructions; wherein the processor is configured to execute the data demand dynamic management method based on scenario-based modeling provided in this application.

[0009] One or more technical solutions provided in this application have at least the following technical effects or advantages: Based on data type mapping, multiple heterogeneous data nodes are constructed in the multimodal perception layer for initial weak connection association. Scene association modeling is performed to obtain a multidimensional scene association matrix. Hierarchical semantic analysis is conducted to determine the dynamic scene situation. Based on the dynamic scene situation, the multidimensional scene association matrix is ​​reconstructed into a dynamic association matrix. K key business nodes and K sets of strong association relationships are extracted from the dynamic association matrix, and topology pruning and strengthening of weakly associated nodes are performed to reconstruct the scene shrinking topology. Using the scene shrinking topology as a data interaction guide, cross-node association and fusion are performed on the K business perception data streams of the K key business nodes to obtain scene-integrated data. Based on the dynamic scene situation, a scene presentation template is invoked to map the scene-integrated data into scene view data for scene-integrated screen display. This achieves the technical effect of realizing adaptive topology simplification and cross-node fusion visualization output of multimodal perception data with dynamic scenes, improving the real-time performance and intuitiveness of multi-source heterogeneous data scene analysis. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic diagram of the dynamic data demand management method based on scenario-based modeling provided in an embodiment of this application.

[0012] Figure 2 This is a schematic diagram of the data demand dynamic management system structure based on scenario-based modeling provided in an embodiment of this application.

[0013] Figure 3 This is a schematic diagram of the structure of an electronic device provided in this application.

[0014] Figure reference numerals: Initial weak connection module 10, multi-dimensional scene association matrix acquisition module 20, dynamic scene situation determination module 30, dynamic association matrix acquisition module 40, scene shrinkage topology reconstruction module 50, cross-node association fusion module 60, cross-node association fusion module 70, processor 21, memory 22, input device 23, output device 24. Detailed Implementation

[0015] This application provides a method, system, and device for dynamic management of data requirements based on scenario-based modeling, which addresses the technical problems of existing technologies, such as fixed associations of multi-source heterogeneous sensing data that cannot adapt to dynamic scenarios, large amounts of data redundancy and computational burden, low efficiency in scenario analysis, and poor visualization effects.

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0017] Example 1, as Figure 1 As shown, this application provides a method for dynamic management of data requirements based on scenario-based modeling, the method comprising: Step S100: Based on the data type, construct multiple heterogeneous data nodes in the multimodal perception layer and perform initial weak connection association.

[0018] Specifically, according to different data types such as images, text, sensor values, and device signaling, multiple heterogeneous data nodes belonging to different acquisition devices and different business dimensions are created one by one in the multimodal perception layer. Basic association links between nodes are established only based on shallow static attributes such as data source and basic business affiliation, generating initial weak connection relationships with low association strength without considering business scenario constraints, thus completing the initial construction of the multi-source perception data underlying node network.

[0019] Step S200: Perform scene association modeling on the multiple heterogeneous data nodes to obtain a multi-dimensional scene association matrix.

[0020] Specifically, the system first retrieves all historical event logs stored in the system, performs time-series aggregation processing on the log data according to the business scenario dimension, and generates N scenario instance sequences corresponding to N typical business scenarios. Then, node mining is performed on each scenario instance sequence to extract N sets of key business nodes that play a core role in each scenario. Subsequently, based on the event sequence logic within each set of scenario instance sequences, directed association analysis is performed on the N sets of key business nodes to sort out the data flow and business trigger sequence relationship between nodes, and build N single-scenario topologies corresponding to each other. A standardized association matrix architecture adapted to all heterogeneous data nodes is pre-constructed, with the matrix rows and columns corresponding to all heterogeneous data nodes, and the matrix elements reserving positions for association strength filling. Then, the aforementioned N scenario topologies, the node association strengths corresponding to the N sets of key business nodes, and the directed association relationships are filled into the corresponding positions of the matrix for each scenario to complete the scenario-based matrix filling. By integrating the node association features under multiple scenarios, a multi-dimensional scenario association matrix that can simultaneously carry information on multiple scenarios, multiple nodes, and directed association strength is finally generated.

[0021] Step S300: Perform hierarchical semantic analysis on the multiple real-time sensing data streams of the multiple heterogeneous data nodes to determine the dynamic scene situation.

[0022] Specifically, the system synchronously collects multiple real-time sensing data streams output from heterogeneous data nodes and performs noise reduction and alignment preprocessing. Multimodal temporal features, including images, numerical values, and text, are extracted from these streams. All modal temporal features are then input into a pre-constructed cascaded hierarchical semantic analysis model. This model consists of an atomic event detection layer, an event combination analysis layer, and a scene situation determination layer, connected in series. First, in the atomic event detection layer, according to the pre-configured mapping relationship between each atomic event detection head and the heterogeneous data nodes, the multimodal temporal features are assigned to corresponding parallel atomic event detection heads to perform atomic event recognition, outputting multiple independent atomic event labels. Then, in the event combination analysis layer, P parallel event combination analysis heads uniformly perform time axis alignment and spatiotemporal window aggregation and segmentation on all atomic event labels to obtain Q spatiotemporal data. For each event cluster, the analysis heads use differentiated attention weight vectors to weight and encode the spatiotemporal event clusters to generate candidate event representations. Then, they superimpose three types of spatiotemporal combination constraints—temporal, spatial, and confidence threshold—to complete pattern matching within the composite event pattern library, outputting P groups of independent composite event fragments in parallel. Finally, the scene situation determination layer enters the scene situation determination layer. It first performs conflict resolution to eliminate contradictory events on the P composite event fragments and performs spatiotemporal semantic splicing to generate a global event representation. It then retrieves the scene situation rule templates corresponding to N typical scenarios, iterates through and matches the global event representations to calculate N groups of scene matching degrees, and then arbitrates and filters all matching degrees according to preset scene priority rules, outputting W candidate scene situation labels and their corresponding scene matching degrees. All candidate labels and matching degrees are integrated as the dynamic scene situation corresponding to the current real-time business.

[0023] Step S400: Reconstruct the multidimensional scene association matrix into a dynamic association matrix based on the dynamic scene situation.

[0024] Specifically, W candidate scenario status labels are extracted from the dynamic scenario status. Based on the business characteristics corresponding to each candidate scenario status label, node association parameters, node sets, and association strength values ​​matching the scenario are extracted hierarchically from the pre-constructed multidimensional scenario association matrix. W-layer dynamic association sub-matrices are generated one-to-one. Each layer of dynamic association sub-matrices retains only the heterogeneous data nodes and the association relationships between nodes that are effective under the corresponding candidate scenario, and removes redundant association dimensions of irrelevant scenarios. All W-layer dynamic association sub-matrices extracted hierarchically are uniformly encapsulated and integrated to form a dynamic association matrix that can be dynamically switched with real-time business scenarios and take effect independently at each layer. This completes the reconstruction of the static multidimensional scenario association matrix into a hierarchical dynamic matrix that adapts to the real-time situation.

[0025] Step S500: Extract K key business nodes and K groups of strong associations from the dynamic association matrix, perform topology pruning and strengthening of weak association nodes, and reconstruct the scene shrinkage topology.

[0026] Specifically, the dynamic association matrix is ​​composed of W layers of dynamic association sub-matrices. Each layer of dynamic association sub-matrices is analyzed layer by layer. H key business nodes with association weights higher than a preset threshold are selected from the single-layer sub-matrices and matched with H groups of strong associations. At the same time, weak association nodes with association weights lower than the threshold and no business interaction value and their corresponding weak association links are identified and removed. After single-layer topology pruning and strengthening, the first scenario shrinkage sub-topology is generated. All remaining dynamic association sub-matrices are processed in a loop according to the same logic to obtain W independent scenario shrinkage sub-topologies. Then, node consistency verification and fusion are performed based on the unique identifier of heterogeneous data nodes. Duplicate key business nodes are merged, strong associations between nodes are unified, and redundant and duplicate topology links between multiple layers of sub-topologies are removed. Finally, a simplified scenario shrinkage topology containing K key business nodes and K groups of strong associations and removing invalid weak association links is obtained.

[0027] Step S600: Using the scenario shrinking topology as the data interaction guide, perform cross-node association and fusion of the K business perception data streams of the K key business nodes to obtain scenario-based integrated data.

[0028] Specifically, a basic data processing framework is first built based on N typical scenario topologies obtained from previous modeling. N candidate processing architectures are then constructed, and training and parameter fine-tuning are performed using N sets of historical scenario instance sequences to obtain N candidate processing networks adapted to the characteristics of each typical scenario. Next, the simplified scenario shrinkage topology is used as a global constraint framework for data flow and computational interaction. Based on the distribution of key nodes within the topology and strong node correlation links, the structure of the N candidate processing networks is fused and the parameters are reconstructed to generate a scenario-based aggregation network adapted to real-time dynamic scenarios. Subsequently, K business-aware data streams corresponding to K key business nodes are input into heterogeneous computing processing nodes bound to each key business node within the scenario-based aggregation network. Combined with W scenario matching degrees output from the dynamic scenario situation, differentiated computing power allocation constraints are assigned. Cross-node data flow linkage and feature correlation fusion computation are carried out according to the directed strong correlation links recorded in the scenario shrinkage topology to eliminate temporal deviations and dimensional conflicts between multi-node data streams, unify and standardize data dimensions, correlation logic, and business features, and finally output scenario-based integrated data adapted to the current real-time business scenario and with unified linkage of multi-source data.

[0029] Step S700: Based on the dynamic scene situation, call the scene presentation template, map the scene-integrated data into scene view data, and perform scene-integrated screen display.

[0030] Specifically, the system reads the candidate scenario status tags (W) and their corresponding scenario matching degrees carried in the dynamic scenario status, filters the candidate scenario status tags with the highest matching degrees, and retrieves the exclusive scenario presentation templates pre-stored in the system that correspond one-to-one with the scenario. These templates have built-in exclusive data display dimensions, chart layouts, field mapping rules, visualization color schemes, and interactive components for the corresponding scenario. The scenario-integrated data obtained from cross-node fusion is processed according to the template's preset field mapping rules to complete field matching, unit conversion, and numerical normalization. The multi-source heterogeneous integrated data is automatically filled into the various visualization components of the template, completing the conversion of structured data into scenario view data in the form of standardized graphics, tables, and indicator cards. Finally, the generated scenario view data is pushed to the front-end large screen display module, and the integrated scenario-integrated rendering display of multi-indicator and multi-dimensional business data is completed according to the template layout. The scenario switching interaction interface is retained simultaneously, and the corresponding template is reloaded and the view content is refreshed in real time as the dynamic scenario status is updated, realizing adaptive visualization screen display output for business scenarios.

[0031] In one possible implementation, step S200 further includes: Step S210: Perform time-series aggregation on historical event logs based on business scenarios to obtain N scenario instance sequences of N typical scenarios, and then perform node mining on the N scenario instance sequences to obtain N sets of key business nodes.

[0032] Step S220: Based on the N scene instance sequences, perform directed association analysis on the N sets of key business nodes to construct N scene topologies.

[0033] Step S230: After constructing the association matrix architecture of the multiple heterogeneous data nodes, the scenario-based matrix is ​​filled according to the N scenario topologies and N sets of key business nodes to obtain the multi-dimensional scenario association matrix.

[0034] Specifically, the system reads all historical event logs in batches. First, it categorizes and aggregates the logs based on the built-in business scenario identifiers. Then, it sorts the scattered log records under the same business scenario according to the event timestamp, aggregates continuous segments, removes isolated and invalid log segments, and generates N complete scenario instance sequences that correspond one-to-one with N typical business scenarios. Subsequently, it traverses and parses each scenario instance sequence, extracts the heterogeneous data node identifiers associated with each log in the sequence, and counts the frequency of each node in the scenario sequence, data output weight, and business trigger priority. Nodes with frequency, weight, and priority all higher than the preset screening threshold are selected as the core carrying nodes of the current scenario. The node mining process is executed on each of the N scenario instance sequences, and finally, N sets of key business nodes belonging to N typical scenarios and independent of each other are obtained.

[0035] Each scenario instance sequence and key business node group corresponding to the same typical scenario are retrieved separately. The occurrence sequence of events, data transmission flow, and business triggering dependencies within the scenario instance sequence are analyzed line by line. The order of data output and reception and the frequency of event-driven associations between any two key business nodes are statistically analyzed. The directed associations of one-way data transmission and two-way interaction between nodes are distinguished and the corresponding association strength weights are assigned. Using a set of key business nodes in the current scenario as topology vertices and the weighted directed associations mined as topology directed edges, the network structure of a single scenario is completed. Following the same processing logic, the remaining N-1 sets of scenario instance sequences and corresponding key business nodes are analyzed and network structures are built. Finally, N scenario topologies that match N typical scenarios one by one and have complete information on nodes, directed edges, and association weights are generated.

[0036] First, a unified, standardized association matrix architecture is built using all heterogeneous data nodes as row and column indices. Each cell of the matrix reserves multi-dimensional storage fields for scene number, directed association type, and association strength weight, forming a blank matrix base that can carry multi-scene association information. Then, N scene topologies and N sets of key business nodes are traversed sequentially. For a single scene topology, all key business nodes and weighted directed association edges between nodes are extracted. The cells corresponding to the starting node row and the ending node column in the matrix are located, and the current scene number, association direction, and association strength weight are written into the multi-dimensional fields of the cells. Cells corresponding to nodes without business association are filled with zero-value association weights, completing the filling of single-scene association information. The directed association data of nodes in the remaining N-1 scene topologies are sequentially filled into the corresponding cells of the matrix. The node association features of all typical scenarios are superimposed and integrated to finally generate a multi-dimensional scene association matrix that is compatible with multi-scene, multi-node, and multi-dimensional information of directed association strength.

[0037] In one possible implementation, step S200 further includes: A pre-constructed hierarchical semantic analysis model is provided, wherein the hierarchical semantic analysis model includes a cascaded atomic event detection layer, an event combination analysis layer, and a scene situation determination layer. The atomic event detection layer has multiple atomic event detection heads connected in parallel, the event combination analysis layer has P event combination analysis heads connected in parallel, and the scene situation determination layer is configured with a scene situation determination head.

[0038] Specifically, a hierarchical semantic analysis model is pre-built. This model employs a three-level cascaded deep multi-branch, multi-head parallel neural network architecture. The three layers are, in order, an atomic event detection layer, an event combination analysis layer, and a scene situation determination layer. Feature transfer channels are set between layers to ensure complete feed-back of preceding outputs. Within the atomic event detection layer, multiple independent atomic event detection heads are deployed in parallel. Each detection head uses a lightweight CNN and Softmax classification algorithm, and each head is individually bound to the modal features corresponding to a class of heterogeneous data nodes. During training, a sample set of labeled single-node raw perceptual data and atomic event labels is input, and the convolution kernel and classification weights are iteratively optimized using a cross-entropy loss function. During inference, the corresponding modal features are independently extracted, and standardized atomic event labels are output. In the event combination analysis layer, P event combination analysis heads are configured in parallel. Each analysis head uses a multi-head spatiotemporal attention encoding algorithm with a built-in learnable differential attention weight vector. During training, atomic event label samples with temporal and spatial coordinate annotations are input, using a temporal loss function... A multi-objective loss function is constructed by jointly using loss, spatial distance loss, and confidence loss to achieve parameter convergence. During inference, atomic event labels are aligned on the time axis, clustered in a spatiotemporal window, and then weighted and encoded. Combined with the triple constraints of temporal, spatial, and confidence, a composite event pattern library is matched to generate independent composite event fragments. The scene situation determination layer is configured with a separate scene situation determination head, which adopts graph matching similarity calculation and multi-rule priority arbitration algorithm. During the training phase, a large number of composite event fragments and corresponding scene situation label samples are input to fit the feature matching thresholds and scene priority ranking rules of various typical scenes. During inference, the conflicts of P-way composite event fragments are first resolved and spatiotemporal semantics are spliced ​​to obtain a global event representation. Then, the scene matching degree is calculated by traversing and matching with N typical scene preset situation rule templates. Based on the preset scene priority rules, situation arbitration and screening are completed, and dynamic scene situation carrying matching degree values ​​is output. After the entire model is trained, the weights and rule templates are solidified, and forward inference is directly performed layer by layer in the real-time data stream parsing stage to achieve end-to-end hierarchical semantic analysis and automatic scene situation determination.

[0039] In one possible implementation, step S300 further includes: Step S310: Perform multimodal feature extraction on the multiple real-time sensing data streams to obtain multiple modal temporal features, and then input them into the hierarchical semantic analysis model, wherein: In the atomic event detection layer, based on the preset mapping relationship between the multiple atomic event detection heads and the multiple heterogeneous data nodes, the multiple modal temporal features are assigned to the multiple atomic event detection heads for atomic event identification, thereby obtaining multiple atomic event labels.

[0040] In the event combination analysis layer, the P event combination analysis heads perform spatiotemporal combination analysis on the multiple atomic event labels under differentiated attention weights, and output P composite event fragments.

[0041] In the scene situation determination layer, the scene situation determination head performs global scene situation aggregation based on the P composite event fragments and outputs the dynamic scene situation.

[0042] Specifically, multimodal feature extraction is performed on the multiple real-time sensing data streams to obtain multiple modal temporal features, which are then input into the hierarchical semantic analysis model, wherein: A pre-stored mapping table is used to bind heterogeneous data node numbers, data modality types, and each atomic event detection head. The mapping table records the node data source and corresponding modality that each detection head is specifically adapted to. After receiving multimodal temporal features, the hierarchical semantic analysis model first reads the source heterogeneous data node identifiers carried by each modality temporal feature, traverses the matching mapping table to determine the target atomic event detection head corresponding to the set of features, and distributes the features to the corresponding detection heads in parallel. Each atomic event detection head adopts a lightweight convolutional classification network. The network has built-in convolutional extraction kernels and a Softmax classifier that have been trained and converged with labeled samples. After inputting the matched modality temporal features, it first extracts fine-grained event features through convolutional layers, and then maps them to the pre-set atomic event classification space through fully connected layers. The Softmax classifier calculates the confidence probability of each type of atomic event, selects the event category with a confidence level higher than a pre-set threshold as the recognition result, and encodes the category as a standardized atomic event label for output. After all parallel detection heads complete feature inference synchronously, all output results are summarized to obtain multiple atomic event labels that correspond one-to-one with each real-time perception data stream.

[0043] In the event combination analysis layer, the P event combination analysis heads perform spatiotemporal combination analysis on the multiple atomic event tags under differentiated attention weights, and output P composite event fragments. Specifically, the layer synchronously receives all atomic event tags output by the atomic event detection layer, first extracts the timestamp and spatial location coordinates attached to each tag to complete global time axis alignment, and completes aggregation and segmentation by dividing windows according to fixed duration and spatial range to generate several independent spatiotemporal event clusters. Each of the P parallel-running event combination analysis heads loads a set of exclusive and different differentiated attention weight vectors, performs weighted encoding on all spatiotemporal event clusters, and generates multiple sets of weighted event features as candidate event representations. Subsequently, each analysis head introduces three spatiotemporal combination constraints: chronological order, spatial distance, and minimum confidence. The candidate event representations are matched and searched against the system's pre-stored composite event pattern library one by one, and event combination structures that meet all constraints are selected. The structured data containing event type, occurrence time, spatial range, associated nodes, and matching confidence is encapsulated, and P composite event fragments with different structures and focuses are output in parallel.

[0044] The scenario situation determination head first receives P composite event fragments output by the event combination analysis layer. It then iterates through and compares the event occurrence sequence, spatial range, and business logic of each fragment. For composite event fragments with temporal conflicts, spatial contradictions, or business mutual exclusions, it performs conflict resolution, removing invalid contradictory fragments and retaining P valid event fragments. Next, it performs spatiotemporal semantic splicing of all valid event fragments according to their temporal order and spatial correlation, integrating them to generate a global event representation that can fully represent the current business situation. It then retrieves N scenario situation rule templates corresponding to N pre-constructed typical scenarios, and uses a similarity matching algorithm to iterate through and compare the global event representation with each template, calculating the scenario matching degree between the global event representation and each typical scenario. Subsequently, it sorts and arbitrates the N scenario matching degrees according to preset scenario priority rules, selecting W candidate scenario situation tags that meet the matching degree criteria and have the highest priority. At the same time, it binds the scenario matching degree value corresponding to each tag for unified encapsulation, and finally integrates the tag and matching degree information to form a complete dynamic scenario situation and outputs it outward.

[0045] In one possible implementation, step S310 further includes: Step S311: Align the multiple atomic event labels with the time axis, perform spatiotemporal window aggregation and segmentation, and obtain Q spatiotemporal event clusters.

[0046] Step S312: The first event combination analysis head performs attention weighting encoding on the Q spatiotemporal event clusters using a preset composite event attention weight vector to obtain Q weighted event features, which constitute the first group of candidate event representations.

[0047] Step S313: Apply spatiotemporal combination constraints to the first group of candidate event representations, perform combination pattern matching in the composite event pattern library, and output the first composite event fragment, wherein the spatiotemporal combination constraints include temporal constraints, spatial constraints, and confidence threshold constraints.

[0048] Specifically, the unified timestamp and spatial coordinate information built into all atomic event tags are extracted, and the time axis alignment and calibration of all tags are completed using the system's global standard time reference to eliminate the sampling time offset caused by the data streams collected by different heterogeneous data nodes. A fixed-duration time sliding window threshold and a spatial neighborhood distance threshold are preset, and atomic event tags within a continuous time period are continuously slid along the aligned complete time axis. Then, the tags within the same time window are divided into neighborhood clusters according to spatial coordinates. A group of atomic event tags with continuous time sequence and spatial distance less than the threshold are grouped into a group of independent units. All atomic event tags are traversed to complete the global segmentation, and invalid units containing only a single isolated tag and without business relevance are removed. Finally, Q spatiotemporal event clusters with multi-event relevance characteristics are summarized and retained.

[0049] The first event combination analysis head reads a pre-configured composite event attention weight vector adapted to the corresponding business analysis dimensions. Each weight value in the vector corresponds to one of four dimensions: atomic event type, event occurrence sequence, spatial distance, and node association strength. It then sequentially traverses Q spatiotemporal event clusters, extracts the original feature vectors corresponding to all atomic event tags within each spatiotemporal event cluster, multiplies the original feature vectors with the head's dedicated attention weight vector dimension by dimension to complete the weighted assignment, performs pooling aggregation operation on all weighted single event features within the cluster, compresses and fuses them to obtain the weighted event features corresponding to a single spatiotemporal event cluster, processes all Q spatiotemporal event clusters in traversal order to generate Q corresponding weighted event features, and then combines and encapsulates all weighted event features in an orderly manner to form a complete first set of candidate event representations.

[0050] First, load the three types of globally preset constraint judgment rules. The temporal constraint stipulates that the atomic events constituting the composite event must meet a fixed order and the time difference between adjacent events must not be less than the minimum interval. The spatial constraint restricts the Euclidean distance between the coordinates of the events participating in the combination to not exceed a preset spatial threshold. The confidence threshold constraint sets the minimum effective score for feature matching. Traverse each weighted event feature in the first group of candidate event representations, first extract the event temporal sequence, spatial coordinate set, and original confidence score carried by the feature, and then check whether the three conditions of temporal order, upper limit of spatial distance, and lower limit of confidence score are met simultaneously. Eliminate invalid candidate features that do not meet any of the conditions, and retain all compliant weighted event features. Then read the local pre-defined rules. The constructed composite event pattern library stores standardized event feature vectors, event time sequence templates, spatial range templates, and event hierarchy structures for each standard pattern. A cosine similarity algorithm is used to compare the selected compliant candidate feature set with each standard pattern in the library, calculating the overall matching similarity between the candidate features and the standard patterns. The target composite event pattern with the highest similarity exceeding the confidence threshold is selected. The event name, business meaning, covered spatiotemporal interval, associated heterogeneous data nodes, underlying atomic event set, and matching similarity score corresponding to this standard pattern are extracted and uniformly encapsulated into a data packet containing complete structured information, serving as the first composite event fragment output by the first event combination analysis header.

[0051] In one possible implementation, step S310 further includes: Step S314: After resolving conflicts in the P composite event fragments to obtain P valid event fragments, spatiotemporal semantic splicing is performed to obtain a global event representation.

[0052] Step S315: Preconstruct N scenario situation rule templates for the N typical scenarios, and use association matching to traverse and compare the global event representations to calculate the matching degree of the N scenarios.

[0053] Step S316: Compare the matching degrees of the N scenarios using the scenario priority rule, perform scenario situation arbitration and screening for the N typical scenarios, and obtain W candidate scenario situation labels and corresponding W scenario matching degrees to constitute the dynamic scenario situation.

[0054] Specifically, the process first extracts the event time interval, spatial coordinate range, event business type, and associated heterogeneous node identifiers from each of the P composite event fragments. Then, it compares any two composite event fragments pairwise to identify conflicting fragments that overlap in time and spatial range but have mutually exclusive business logic or contradictory node data conclusions. For these conflicting fragments, based on preset event priorities and data collection reliability weights, high-reliability fragments are retained, while low-reliability fragments that contradict each other are deleted or corrected. This process completes the conflict removal and correction of all fragments, resulting in P valid event fragments without logical contradictions. Subsequently, all valid event fragments are sorted in ascending order by timestamp. The semantics of atomic events, node association information, and spatiotemporal range feature vectors within each fragment are extracted. These vectors are then concatenated along the time dimension and fused along the spatial dimension to perform unified semantic encoding and splicing, eliminating temporal discontinuities and spatial information fragmentation between fragments. Finally, the multi-dimensional features of all valid events are merged into a complete, unified, and comprehensive high-dimensional feature vector covering all current business information, which serves as the global event representation output.

[0055] Based on historical labeled business data, inherent features of N typical scenarios are extracted, generating N scenario situation rule templates. Each template stores the scenario's standard feature vector, a list of necessary events, a spatiotemporal threshold range, node association rules, and corresponding four sets of dimensional weights w1, w2, w3, and w4, where w1+w2+w3+w4=1. The global feature vector, global event set, event spatiotemporal parameters, and node association topology of the global event representation are extracted. Sub-item scores are calculated sequentially with each scenario template. The first item uses the cosine similarity formula to calculate the similarity score S1 between the global feature vector and the template's standard feature vector, with a value ranging from 0 to 1. The second item is the event overlap matching rate S2 obtained by dividing the number of intersection elements by the total number of necessary events in the template. The values ​​range from 0 to 1. The third item determines whether the global spatiotemporal parameters fall within the template spatiotemporal threshold. If they are perfectly matched, S3=1. If they exceed the threshold, the matching score S3 between 0 and 1 is obtained by linearly decreasing the deviation. The fourth item compares the global node association topology with the template node association rules and calculates the proportion of association links that meet the rules to obtain the topology matching score S4, which ranges from 0 to 1. The scores of the four items are multiplied by the corresponding dimension weights and then summed. The matching degree is calculated as S=w1×S1+w2×S2+w3×S3+w4×S4 to obtain the matching degree of a single scene. The same calculation logic is used to iterate through all N scene situation rule templates and calculate the matching degree of N scenes in the interval of 0 to 1.

[0056] The system retrieves a pre-configured scenario priority rule table, categorized by business importance. This table records the priority level and minimum effective matching threshold for each typical scenario. N scenario matching scores are associated with corresponding scenario identifiers and priority levels. Invalid scenarios with matching scores below their respective minimum effective thresholds are first filtered out. The remaining valid scenarios meeting the matching score requirements are then comprehensively sorted based on their matching score values ​​and preset scenario priority levels. Higher matching scores and higher scenario business priorities result in higher rankings. The first W groups of data are extracted in descending order, and the corresponding scenario numbers for each group are extracted and converted into standardized candidate scenario status labels. Simultaneously, the scenario matching score values ​​corresponding to each label are retained. The W groups of candidate scenario status labels and matching score values ​​are integrated and encapsulated into a unified structured data package, forming a complete dynamic scenario status carrying quantitative indicators of scenario adaptation, which is then output downstream.

[0057] To clearly illustrate the complete implementation process of scenario matching degree calculation and candidate scenario arbitration screening in this solution, this embodiment provides measured matching data for N=5 typical scenarios under the park security business, as shown in Table 1. The business scenarios cover multiple heterogeneous perception data fusion scenarios such as personnel gathering, equipment failure, boundary intrusion, vehicle stagnation, and routine inspection. The weights of each dimension are uniformly configured as w1=0.3, w2=0.3, w3=0.2, and w4=0.2.

[0058] Table 1. Actual measured data on global event matching degree in typical business scenarios.

[0059] The processing steps in this step are illustrated using the measured data in Table 1: The current real-time global event is a composite event involving a large gathering of people at the park entrance / exit, synchronous error reporting by the supporting monitoring equipment, and prolonged parking by external vehicles. After obtaining the global event representation through hierarchical semantic analysis, the four-dimensional sub-scores are calculated sequentially with the five scenario situation rule templates. Then, the comprehensive matching degree of each scenario is obtained through weighted summation. Subsequently, the preset scenario priority and the lowest effective matching degree threshold are retrieved for filtering. Scenario No. 3, boundary crossing intrusion, has a matching degree of 0.414, which is lower than the threshold of 0.5. Scenario No. 5, routine equipment inspection, has a matching degree of 0.218, which is lower than the threshold of 0.3. These are directly judged as invalid scenarios and removed. The remaining three groups of effective scenarios, Nos. 1, 2, and 4, are sorted according to the matching degree value priority and the scenario priority priority under the same matching degree, resulting in W=3 candidate scenario situation labels and corresponding matching degrees of 0.880, 0.690, and 0.711. These three are encapsulated and integrated to form the dynamic scenario situation corresponding to the current real-time business, serving as the input basis for subsequent multi-dimensional scenario association matrix layered extraction and dynamic association matrix construction. Based on this dynamic scenario situation, the subsequent step S510 can extract the corresponding three-layer dynamic association sub-matrices from the multi-dimensional scenario association matrix according to the three candidate scenario labels, and fully realize the reconstruction of the static matrix into a dynamic association matrix adapted to real-time events. In real business scenarios, redundant nodes and association links corresponding to irrelevant scenarios can be filtered out significantly, reducing the invalid computing power overhead in the subsequent topology pruning and network fusion stages.

[0060] In one possible implementation, step S500 further includes: Step S510: Based on the W candidate scene situation labels, perform layered extraction of the multidimensional scene association matrix to obtain W layers of dynamic association sub-matrices, which constitute the dynamic association matrix.

[0061] Step S520: Extract H key business nodes and H groups of strong associations from the first-layer dynamic association submatrix, perform topology pruning and strengthening of weak association nodes, and reconstruct the first scenario shrinkage subtopology.

[0062] Step S530: By analogy, process the W-layer dynamic correlation sub-matrix to obtain W scene shrinkage sub-topologies, perform topology fusion based on node consistency, and generate the scene shrinkage topology.

[0063] Specifically, the unique scene codes and scene matching degrees corresponding to the W current valid candidate scene status labels are read. Using these as filtering indexes, a hierarchical and precise retrieval is performed on the pre-constructed multidimensional scene association matrix. From the full N types of scene node association data stored in the multidimensional scene association matrix, the complete node association topology data, node association weights, directed association directions, and scene attribute parameters corresponding to each of the W candidate scenes are extracted one by one. For each type of candidate scene, an independent two-dimensional association sub-matrix containing all heterogeneous data node row and column indices, directed association weight values ​​between nodes, and zero-value padding for unrelated nodes is extracted. A total of W layers of dynamic association sub-matrixes with independent structure, unique scene attributes, and adapted to the current real-time scene status are extracted. Then, according to the priority order of the W candidate scene matching degrees from high to low, the dynamic association sub-matrixes of each layer are hierarchically sorted and structurally encapsulated. The node index arrangement rules and numerical storage formats of each layer of matrix are unified, and the independent scene identifiers and association weight features of each layer of matrix are preserved. Finally, the ordered W layers of dynamic association sub-matrixes are aggregated to construct a complete dynamic association matrix with clear hierarchy, scene adaptation, and layer-by-layer parsing and calculation capabilities.

[0064] The first-level dynamic association submatrix is ​​read from the sorted top-ranking column. Using the directed association weights stored in the matrix cells as the criterion, all rows and columns of the matrix are traversed. Cells with association weights greater than the preset strong association threshold are filtered out, and the corresponding row and column indexes are extracted as H key business nodes and H groups of directed strong associations between nodes. A temporary basic topology is built using the H key business nodes as the topology base vertices and the H groups of strong associations as the initial directed edges. All vertices in the temporary basic topology are traversed, and the total weight of the association edges and the frequency of business event triggering for each vertex are counted one by one. Nodes with a total association weight lower than the weak association threshold and no core event carrying capacity are identified as weak association nodes. All inbound and outbound directed edges corresponding to weak association nodes are deleted simultaneously to complete topology pruning. After pruning, the connectivity of the association links between the remaining nodes is verified. For business logic links that are broken due to pruning, the transitional association edges with the next lower weight are retained to strengthen the links. Redundant and isolated nodes and invalid low-weight associations are eliminated. The identifiers of the remaining nodes, the association weights of the directed edges, and the transmission flow are standardized. Finally, the first scenario shrinking subtopology reconstruction is completed, which is simplified, connected, and retains only the core business interaction links.

[0065] Following the topology extraction, weak-association node pruning, and subtopology reconstruction process corresponding to the first-layer dynamic association submatrix, the same processing logic is sequentially executed layer by layer for the remaining W-1 layers of dynamic association submatrix. Each layer filters high-weight key business nodes and strong associations within the corresponding scenario, removes low-weight weak-association nodes and their associated edges, and completes broken business links, outputting independent and streamlined scenario shrinkage subtopologies layer by layer, ultimately obtaining W scenario shrinkage subtopologies belonging to W candidate scenarios. Subsequently, a global topology fusion operation based on node consistency rules is performed. First, the unique identifier index of heterogeneous data nodes in all W subtopologies is unified, and the same business node with the same number in different subtopologies is merged. Points are identified as having the same origin and are merged. For multiple directed edges from multiple topologies between merged nodes with the same origin, the association weights are weighted according to the matching degree of the corresponding scenario. Unique nodes that exist only in a single sub-topology and have no matching origin, along with their associated edges, are completely retained without deletion. Duplicate and redundant directed edges and isolated invalid vertices generated during the fusion process are cleaned up simultaneously. The transmission flow direction of all directed edges, the superimposed comprehensive association weights, and the node business attributes are uniformly calibrated. W scattered and independent scenario shrinkage sub-topologies are integrated into a unified network structure with unified nodes, fused association weights, and coverage of the core interaction logic of all candidate scenarios, thus completing the generation of the scenario shrinkage topology.

[0066] In one possible implementation, step S600 further includes: Step S610: Using the N scene topologies as the data flow framework, construct N candidate processing architectures, and train and fine-tune them using the N scene instance sequences to obtain N candidate processing networks.

[0067] Step S620: Based on the scene shrinking topology, the N candidate processing networks are fused and reconstructed to obtain a scene-based aggregated network.

[0068] Step S630: Inject the K business-aware data streams into the K heterogeneous computing processing nodes corresponding to the K key business nodes in the scenario-based aggregation network, perform directed correlation fusion calculation under the differentiated computing power constraints of the W scenario matching degrees, and output the scenario-based integrated data.

[0069] Specifically, a heterogeneous fusion algorithm based on Graph Convolutional Neural Networks (GCN) and Temporal Coding Networks (TCN) is used to construct the basic architecture. For N scene topologies, the node distribution, directed adjacency weight matrices, and topological connectivity are analyzed. Topological business nodes are mapped to GCN feature aggregation nodes, and the inter-node association weights are assigned as parameters of the graph convolutional adjacency matrix. Combined with temporal one-dimensional convolutional layers, N dedicated candidate processing architectures are built to adapt to the data propagation logic of each scene. Each architecture uniformly includes a graph convolutional aggregation module, a temporal feature embedding module, and a scene feature mapping output module. The architecture structure is adaptively generated according to the differences in scene topology connectivity. During the training phase, labeled scene instance sequences corresponding to each scene are used as training samples, and a joint loss function of cross-entropy classification loss and topological consistency regularization loss is constructed. The algorithm employs mini-batch gradient descent for iterative training. During forward propagation, the graph convolution module aggregates topological node association features, while the temporal coding module captures the temporal evolution of scene data flow. It then fuses and outputs high-dimensional scene feature vectors to predict scene situation results. During backpropagation, the graph convolution weights, temporal coding parameters, and topological adjacency matrix weights are updated layer by layer based on the joint loss gradient. Overfitting is suppressed using learning rate decay and early stopping strategies. Once the loss of each model converges and the validation set accuracy stabilizes, all parameters are fixed, completing fine-tuning training of N candidate processing networks. In the usage phase, the fixed network weights are directly loaded. Using the real-time scene topology as the data flow carrier, inputting real-time data flow features allows end-to-end scene feature parsing and situation inference output based on the corresponding candidate processing networks.

[0070] The system reads and integrates the core nodes and related links of all candidate scenarios into a scenario-shrinking topology. Each key business node within the topology is used as the basic computational unit after network fusion. Based on the topology node matching index, the system aligns and merges the corresponding sub-modules of the same source nodes in N candidate processing networks. A weighted fusion strategy is adopted for the feature extraction, graph convolution, and temporal coding sub-modules from different candidate networks at the same source nodes. The historical matching average of each candidate network's corresponding scenario is used as the fusion weight to superimpose module parameters. Lightweight feature processing sub-modules are added to newly added nodes unique to the scenario-shrinking topology. Simultaneously, the feature transmission paths between modules are reconstructed according to the directed related links within the topology, and path weights are configured. A global aggregation layer is introduced to uniformly receive the fused features output from all nodes. A complete network structure is constructed, including a node sub-module fusion layer, a topology feature transmission layer, and a global feature aggregation layer. Redundant branches unrelated to the scenario-shrinking topology in multiple candidate networks are eliminated. The overall network parameters and topology data flow paths are uniformly calibrated. Finally, a scenario-based aggregation network adapted to real-time integrated scenarios and integrating the feature extraction capabilities of multiple candidate networks is generated.

[0071] First, a one-to-one mapping relationship is established between K service-aware data streams and K heterogeneous computing processing nodes within the network based on the scenario shrinking topology. The collected original service-aware data streams are distributed to heterogeneous computing processing nodes matching the data modality to complete the underlying feature extraction. The scenario matching degree values ​​corresponding to W candidate scenarios are retrieved as differentiated computing power allocation weights. Scenarios with higher matching degrees are allocated more computing power and have larger feature flow scaling coefficients for the corresponding network links. Scenarios with lower matching degrees have reduced computing power allocation and weakened feature transfer weights for the corresponding links. Based on the directed links of the scenario shrinking topology, cross-node feature transfer and weighted fusion operations are carried out layer by layer. Multi-layer node feature aggregation, temporal feature superposition, and multi-scenario feature weighted integration are completed along the directional flow of the topology. Throughout the process, the computing volume and feature contribution ratio of each branch are dynamically constrained based on the matching degree of each scenario. After fusion, the comprehensive features output by all heterogeneous computing nodes in the network are uniformly regulated, redundant information is removed, and spatiotemporal, node, and multi-scenario matching related structured information are encapsulated. Finally, scenario-integrated data carrying multi-service and multi-scenario fusion features is output.

[0072] Example 2, based on the same inventive concept as the data demand dynamic management method based on scenario-based modeling in the previous examples, such as... Figure 2 As shown, this application provides a dynamic data demand management system based on scenario-based modeling. The system and method embodiments in this application are based on the same inventive concept. The system includes: The initial weak connection association module 10 is used to construct multiple heterogeneous data nodes based on data type in the multimodal perception layer and perform initial weak connection association.

[0073] The multi-dimensional scene association matrix acquisition module 20 is used to perform scene association modeling on the multiple heterogeneous data nodes to obtain a multi-dimensional scene association matrix.

[0074] The dynamic scene situation determination module 30 is used to perform hierarchical semantic analysis on multiple real-time sensing data streams of the multiple heterogeneous data nodes to determine the dynamic scene situation.

[0075] The dynamic correlation matrix acquisition module 40 is used to reconstruct the multidimensional scene correlation matrix into a dynamic correlation matrix based on the dynamic scene situation.

[0076] The scene shrinking topology reconstruction module 50 is used to extract K key business nodes and K groups of strong correlations from the dynamic correlation matrix, perform topology pruning and strengthening of weak correlation nodes, and reconstruct the scene shrinking topology.

[0077] The cross-node association and fusion module 60 is used to perform cross-node association and fusion of the K business perception data streams of the K key business nodes, guided by the scenario shrinking topology, to obtain scenario-based integrated data.

[0078] The scenario-based aggregation display module 70 is used to call the scenario presentation template according to the dynamic scenario situation, map the scenario-based integrated data into scenario view data, and perform scenario-based aggregation display.

[0079] Furthermore, the system is also used to implement the following functions: Historical event logs are aggregated based on business scenarios over time to obtain N scenario instance sequences of N typical scenarios. Then, node mining is performed on the N scenario instance sequences to obtain N sets of key business nodes. Based on the N scenario instance sequences, directed association analysis is performed on the N sets of key business nodes to construct N scenario topologies. After constructing the association matrix architecture of the multiple heterogeneous data nodes, scenario-based matrix filling is performed based on the N scenario topologies and N sets of key business nodes to obtain the multidimensional scenario association matrix.

[0080] Furthermore, the system is also used to implement the following functions: A pre-constructed hierarchical semantic analysis model is provided, wherein the hierarchical semantic analysis model includes a cascaded atomic event detection layer, an event combination analysis layer, and a scene situation determination layer. The atomic event detection layer has multiple atomic event detection heads connected in parallel, the event combination analysis layer has P event combination analysis heads connected in parallel, and the scene situation determination layer is configured with a scene situation determination head.

[0081] Furthermore, the system is also used to implement the following functions: Multimodal feature extraction is performed on the multiple real-time perceived data streams to obtain multiple modal temporal features, which are then input into the hierarchical semantic analysis model. Specifically: in the atomic event detection layer, based on the preset mapping relationship between the multiple atomic event detection heads and multiple heterogeneous data nodes, the multiple modal temporal features are assigned to the multiple atomic event detection heads for atomic event identification, resulting in multiple atomic event labels; in the event combination analysis layer, the P event combination analysis heads perform spatiotemporal combination analysis on the multiple atomic event labels under differentiated attention weights, outputting P composite event fragments; in the scene situation determination layer, the scene situation determination head performs global scene situation aggregation based on the P composite event fragments, outputting the dynamic scene situation.

[0082] Furthermore, the system is also used to implement the following functions: The timeline is aligned with the multiple atomic event labels, and spatiotemporal window aggregation and segmentation are performed to obtain Q spatiotemporal event clusters. The Q spatiotemporal event clusters are then subjected to attention-weighted encoding with a preset composite event attention weight vector in the first event combination analysis head to obtain Q weighted event features, which constitute a first set of candidate event representations. Spatiotemporal combination constraints are applied to the first set of candidate event representations, and combination pattern matching is performed in the composite event pattern library to output a first composite event fragment. The spatiotemporal combination constraints include temporal constraints, spatial constraints, and confidence threshold constraints.

[0083] Furthermore, the system is also used to implement the following functions: After resolving conflicts among the P composite event fragments to obtain P valid event fragments, spatiotemporal semantic splicing is performed to obtain a global event representation. N scene situation rule templates for the N typical scenarios are pre-constructed, and the global event representation is compared using association matching to calculate the N scene matching degrees. The N scene matching degrees are compared using scene priority rules, and scene situation arbitration screening for the N typical scenarios is performed to obtain W candidate scene situation labels and corresponding W scene matching degrees, which constitute the dynamic scene situation.

[0084] Furthermore, the system is also used to implement the following functions: Based on the W candidate scenario situation labels, the multidimensional scenario association matrix is ​​extracted in layers to obtain W layers of dynamic association sub-matrices, which constitute the dynamic association matrix. H key business nodes and H groups of strong association relationships are extracted from the first layer of dynamic association sub-matrices, and topology pruning and strengthening of weak association nodes are performed to reconstruct the first scenario shrinkage sub-topology. The W layers of dynamic association sub-matrices are processed similarly to obtain W scenario shrinkage sub-topologies, and topology fusion based on node consistency is performed to generate the scenario shrinkage topology.

[0085] Furthermore, the system is also used to implement the following functions: Using the N scenario topologies as the data flow framework, N candidate processing architectures are constructed and trained and fine-tuned using the N scenario instance sequences to obtain N candidate processing networks. Based on the scenario shrinkage topology, the N candidate processing networks are fused and reconstructed to obtain a scenario-based aggregation network. The K business-aware data streams are injected into the scenario-based aggregation network into K heterogeneous computing processing nodes corresponding to the K key business nodes, and directed correlation fusion calculations are performed under the differentiated computing power constraints of the W scenario matching degrees to output the scenario-based integrated data.

[0086] Example 3, Figure 3 This is a schematic diagram of the structure of an electronic device provided by the data demand dynamic management method based on scenario modeling of the present invention, showing a block diagram of an exemplary electronic device suitable for implementing the embodiments of the present invention. Figure 3 The electronic device shown is merely an example and should not be construed as limiting the functionality or scope of the embodiments of the present invention. Figure 3 As shown, the electronic device includes a processor 21, a memory 22, an input device 23, and an output device 24; the number of processors 21 in the electronic device can be one or more. Figure 3 Taking a processor 21 as an example, the processor 21, memory 22, input device 23, and output device 24 in an electronic device can be connected via a bus or other means. Figure 3 Taking the example of a connection between China and Israel via a bus.

[0087] It should be noted that the order of the embodiments described above is for descriptive purposes only and does not represent the superiority or inferiority of the embodiments. Specific embodiments of this specification have been described above. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0088] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0089] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application intends to include such modifications and variations.

Claims

1. A dynamic data demand management method based on scenario-based modeling, characterized in that, The method includes: Based on data type, multiple heterogeneous data nodes are constructed in the multimodal perception layer and initial weak connections are established. A scene association model is performed on the multiple heterogeneous data nodes to obtain a multi-dimensional scene association matrix; Hierarchical semantic analysis is performed on multiple real-time sensing data streams from the multiple heterogeneous data nodes to determine the dynamic scene situation; Based on the dynamic scene situation, the multidimensional scene association matrix is ​​reconstructed into a dynamic association matrix; K key business nodes and K sets of strong associations are extracted from the dynamic association matrix, and the topology of weak association nodes is strengthened by topology pruning and reconstruction to shrink the scene topology. Using the scenario shrinkage topology as a guide for data interaction, cross-node association and fusion are performed on the K business perception data streams of the K key business nodes to obtain scenario-based integrated data; Based on the dynamic scene situation, the scene presentation template is invoked, and the scene-integrated data is mapped into scene view data for scene aggregation and display.

2. The data demand dynamic management method based on scenario-based modeling as described in claim 1, characterized in that, The method involves performing scene association modeling on the multiple heterogeneous data nodes to obtain a multidimensional scene association matrix, the method comprising: After performing time-series aggregation on historical event logs based on business scenarios to obtain N scenario instance sequences of N typical scenarios, node mining is performed on the N scenario instance sequences to obtain N sets of key business nodes. Based on the N scenario instance sequences, perform directed association analysis on the N sets of key business nodes to construct N scenario topologies; After constructing the association matrix architecture of the multiple heterogeneous data nodes, the scenario-based matrix is ​​filled according to the N scenario topologies and N sets of key business nodes to obtain the multi-dimensional scenario association matrix.

3. The data demand dynamic management method based on scenario-based modeling as described in claim 2, characterized in that, A pre-constructed hierarchical semantic analysis model is provided, wherein the hierarchical semantic analysis model includes a cascaded atomic event detection layer, an event combination analysis layer, and a scene situation determination layer. The atomic event detection layer has multiple atomic event detection heads connected in parallel, the event combination analysis layer has P event combination analysis heads connected in parallel, and the scene situation determination layer is configured with a scene situation determination head.

4. The data demand dynamic management method based on scenario-based modeling as described in claim 3, characterized in that, The method involves performing hierarchical semantic analysis on multiple real-time sensing data streams from the multiple heterogeneous data nodes to determine the dynamic scene situation. Multimodal feature extraction is performed on the multiple real-time sensing data streams to obtain multiple modal temporal features, which are then input into the hierarchical semantic analysis model, wherein: In the atomic event detection layer, based on the preset mapping relationship between the multiple atomic event detection heads and the multiple heterogeneous data nodes, the multiple modal temporal features are assigned to the multiple atomic event detection heads for atomic event identification, thereby obtaining multiple atomic event labels; In the event combination analysis layer, the P event combination analysis heads perform spatiotemporal combination analysis on the multiple atomic event labels under differentiated attention weights, and output P composite event fragments; In the scene situation determination layer, the scene situation determination head performs global scene situation aggregation based on the P composite event fragments and outputs the dynamic scene situation.

5. The data demand dynamic management method based on scenario-based modeling as described in claim 4, characterized in that, The P event combination analysis heads perform spatiotemporal combination analysis under differential attention weights on the multiple atomic event labels, outputting P composite event fragments. The method includes: The timeline is aligned with the multiple atomic event labels, and spatiotemporal window aggregation and segmentation are performed to obtain Q spatiotemporal event clusters; The first event combination analysis head performs attention weighting encoding on the Q spatiotemporal event clusters using a preset composite event attention weight vector to obtain Q weighted event features, which constitute the first group of candidate event representations; Spatiotemporal combination constraints are applied to the first group of candidate event representations, and combination pattern matching is performed in the composite event pattern library to output the first composite event fragment. The spatiotemporal combination constraints include temporal constraints, spatial constraints, and confidence threshold constraints.

6. The data demand dynamic management method based on scenario-based modeling as described in claim 4, characterized in that, The scene situation determination head performs global scene situation aggregation based on the P composite event fragments and outputs the dynamic scene situation. The method includes: After resolving conflicts in the P composite event fragments to obtain P valid event fragments, spatiotemporal semantic splicing is performed to obtain a global event representation; N scenario situation rule templates for the N typical scenarios are pre-constructed, and the global event representation is compared by association matching to calculate the matching degree of the N scenarios; The matching degrees of the N scenarios are compared using scenario priority rules, and scenario situation arbitration screening of the N typical scenarios is performed to obtain W candidate scenario situation labels and corresponding W scenario matching degrees, which constitute the dynamic scenario situation.

7. The data demand dynamic management method based on scenario-based modeling as described in claim 6, characterized in that, The method further includes: Based on the W candidate scene situation labels, the multidimensional scene association matrix is ​​extracted in a layered manner to obtain W layers of dynamic association sub-matrices, which constitute the dynamic association matrix; H key business nodes and H groups of strong associations are extracted from the first-layer dynamic association submatrix. The topology of weak association nodes is strengthened by topology pruning, and the shrinking subtopology of the first scenario is reconstructed. By analogy, the W-layer dynamic correlation sub-matrix is ​​processed to obtain W scene shrinkage sub-topologies. Topology fusion based on node consistency is then performed to generate the scene shrinkage topology.

8. The data demand dynamic management method based on scenario-based modeling as described in claim 7, characterized in that, Using the scenario shrinkage topology as a data interaction guide, cross-node correlation and fusion are performed on the K business-aware data streams of the K key business nodes to obtain scenario-based integrated data. The method includes: Using the N scene topologies as the data flow framework, N candidate processing architectures are constructed, and the N scene instance sequences are used for training and fine-tuning to obtain N candidate processing networks. Based on the scene shrinkage topology, the N candidate processing networks are fused and reconstructed to obtain a scene-based aggregation network; The K business-aware data streams are injected into the K heterogeneous computing processing nodes corresponding to the K key business nodes in the scenario-based aggregation network, and directed correlation fusion calculations are performed under the differentiated computing power constraints of the W scenario matching degrees to output the scenario-based integrated data.

9. A dynamic data demand management system based on scenario-based modeling, characterized in that, The system is used to implement the data demand dynamic management method based on scenario-based modeling as described in any one of claims 1-8, and the system includes: The initial weak connection association module is used to construct multiple heterogeneous data nodes based on data type in the multimodal perception layer and perform initial weak connection association. The multi-dimensional scene association matrix acquisition module is used to perform scene association modeling on the multiple heterogeneous data nodes to obtain a multi-dimensional scene association matrix. The dynamic scene situation determination module is used to perform hierarchical semantic analysis on multiple real-time sensing data streams from the multiple heterogeneous data nodes to determine the dynamic scene situation. The dynamic correlation matrix acquisition module is used to reconstruct the multi-dimensional scene correlation matrix into a dynamic correlation matrix based on the dynamic scene situation. The scene shrinking topology reconstruction module is used to extract K key business nodes and K sets of strong correlation relationships from the dynamic correlation matrix, perform topology pruning and strengthening of weak correlation nodes, and reconstruct the scene shrinking topology. The cross-node association and fusion module is used to perform cross-node association and fusion of the K business perception data streams of the K key business nodes, guided by the scenario shrinking topology, to obtain scenario-based integrated data. The scenario-based aggregation display module is used to call the scenario presentation template according to the dynamic scenario situation, map the scenario-based integrated data into scenario view data, and perform scenario-based aggregation display.

10. An electronic device, characterized in that, The electronic device includes: processor; Memory used to store the processor's executable instructions; The processor is used to execute the data demand dynamic management method based on scenario modeling as described in any one of claims 1 to 8.

Images