Digital architecture graph dynamic optimization method based on multi-source data association mining

Abstract

The embodiment of the invention provides a method for dynamically optimizing one graph of a digital architecture based on multi-source data association mining, and belongs to the technical field of digital architecture management and optimization. The method comprises the steps of collecting multi-source architecture data and performing tagging processing; modeling based on the incidence relation of the graph database; dynamically constructing a digital architecture graph; and performing dynamic optimization based on association mining. According to the method, automatic acquisition, association mining and dynamic optimization of architecture assets are realized, and the dynamic management and control capability and lean management level of a digital architecture are improved.

Description

Technical Field

[0001] This invention relates to the field of digital architecture management and optimization technology, specifically to a dynamic optimization method for a single graph of digital architecture based on multi-source data association mining. Background Technology

[0002] Currently, enterprises commonly face challenges in managing their digital architecture, including difficulties in integrating multi-source data, lagging dynamic updates, and insufficient analysis of data relationships. Specifically, these problems manifest as follows: First, the architecture assets are scattered across different systems, such as permission platforms, monitoring tools, and data platforms. The data formats are diverse, including structured database tables, semi-structured XML / Excel, and unstructured documents and drawings. The lack of a unified association model makes data integration difficult. Secondly, traditional architecture diagrams are mostly static displays and cannot reflect the real-time operating status of the system, such as changes in key indicators like service call frequency and data link latency, which often cause architecture optimization decisions to lag behind actual needs. Third, there are many hidden dependencies among architectural elements, such as implicit associations between business processes and data models. These relationships have not been effectively uncovered, affecting the accuracy and comprehensiveness of architectural problem diagnosis.

[0003] While graph database technology can store some relationships, it lacks algorithmic design for dynamic architecture optimization. Natural language processing technology has limited application in tagging and has failed to form an effective collaborative mechanism with graph computing, making it difficult to meet enterprises' needs for intelligent management of the entire lifecycle of digital architecture. Existing technologies for digital architecture management have the following key shortcomings: 1. Limitations of static visualization: Traditional architecture diagrams use static images or fixed topology displays, which cannot synchronize system operation indicators (such as CPU utilization and interface response time) in real time, resulting in a disconnect between the architecture status and the actual operation. Operation and maintenance personnel need to manually compare multiple documents to locate the problem. 2. Low efficiency of manual analysis: Association rule mining relies on expert experience. For example, identifying the dependency relationship between "business process and data table" requires manual sorting, which takes more than 2 hours in a scenario with 1000+ architecture assets, and the accuracy rate is only 65%. 3. Lack of multi-source data fusion standards: There is a lack of unified data fusion specifications. For example, structured ISC permission data and unstructured architecture documents are not effectively linked, resulting in a data utilization rate of less than 40%. 4. Limited dynamic optimization algorithms: Association rule mining often uses the Apriori algorithm, which takes more than 120 seconds in scenarios with 100,000+ records, failing to meet real-time decision-making requirements. Summary of the Invention

[0004] The purpose of this invention is to provide a dynamic optimization method for a single graph of digital architecture based on multi-source data association mining. This method realizes automatic collection, association mining and dynamic optimization of architecture assets, thereby improving the dynamic control capability and lean management level of digital architecture.

[0005] To achieve the above objectives, embodiments of the present invention provide a method for dynamic optimization of a single graph in a digital architecture based on multi-source data association mining, the method comprising: Collect multi-source architecture data and perform tagging processing; Modeling is based on the relationships in a graph database; Dynamically construct a single diagram of a digital architecture; Dynamic optimization is performed based on association mining.

[0006] Preferably, the process of collecting multi-source architecture data and performing tagging includes: Construct a standard meta-model of architectural elements covering business, applications, data, technology, and security domains; Based on this standard meta-model, a differentiated acquisition strategy is adopted to automatically collect architecture data from multi-source heterogeneous systems and perform entity recognition on unstructured data. A multi-level labeling system is constructed based on the recognition results and the domain knowledge graph, and the label weights are calculated.

[0007] Preferably, the standard meta-model is defined using a triplet structure according to formula (1). (1) in, This includes business domains, application domains, data domains, technology domains, and security domains; For structured data, RESTful API combined with CDC technology is used for real-time or near real-time data collection. For semi-structured data, a combination of XPath / XLST parsing and data cleaning is used for data collection. For unstructured documents and drawings, a collaborative link of "web crawling + OCR + entity recognition based on pre-trained language models" is used for collection and structured extraction. The tagging system includes first-level tags based on the core attributes of entities, second-level tags based on the core functions of entities, and third-level tags based on the relationships between entities; and the tag weights are calculated using a weighted method that combines the TF-IDF algorithm and the TextRank algorithm.

[0008] Preferably, when calculating label weights, According to formula (2), the local weights of terms in a single document are calculated using the TF-IDF algorithm. (2) Where TF(t,d) is the frequency of term t in document d, and is calculated according to formula (3). , (3) Where N is the total number of documents collected, DF(t) is the number of documents containing term t, and 1 is added to the denominator to avoid zero values; The importance weights of terms in the global document set are calculated using the TextRank algorithm. A term co-occurrence network is constructed, and the node weights are iteratively calculated according to formula (4). (4) Where (WS(v_i)) is the weight of term (v_i), d is the damping coefficient, (In(v_i)) is the set of terms pointing to (v_i), and (Out(v_j)) is the set of terms pointing to (v_j). The final label weight is the weighted sum of TF-IDF weight and TextRank weight, and the top 3 terms with the highest weights are used as the core entity labels.

[0009] Preferably, the modeling of relationships based on the graph database includes: An attribute graph model is used, with the processed entities as nodes and the relationships between entities as edges, to construct a multi-source data association model; Design an entity linking and attribute fusion algorithm to solve the problems of entity duplication and attribute conflict in multi-source data.

[0010] Preferably, the attribute graph model includes entity nodes, relationship edges and their attributes, wherein entity nodes have unique ID codes, attributes include core attributes and management attributes, and relationship edges define various association types including invocation, dependency, inclusion, and authorization, and include relationship strength attributes. The entity linking algorithm calculates the name similarity and attribute similarity between entities, and then weights them to obtain a comprehensive similarity. When the comprehensive similarity exceeds a preset threshold, the entities are determined to be the same entity and are merged. The attribute fusion algorithm uses a weighted average fusion for numerical attributes and a voting fusion for enumerated attributes.

[0011] Preferably, the dynamically constructed digital architecture diagram includes: A static architecture topology diagram is constructed using a hierarchical layout and force-directed layout algorithm; By accessing and processing real-time running data through a stream processing framework, the calculated heat values ​​and rules are dynamically superimposed onto a static topology map to generate an architecture map that includes dynamic heat maps and indicator labels.

[0012] Preferably, the static topology graph is vertically divided into multiple levels according to domains, the node positions are calculated using a force-directed layout algorithm, and local optimization is performed in conjunction with a community detection algorithm. Interactive rendering is then performed using WebGL technology. Real-time running data is accessed and windowed aggregated through a stream processing framework and associated with entity nodes in the graph database. A heat value calculation model is built based on CPU utilization, interface response time, and error rate indicators. The heat values ​​are mapped to color gradients and rendered on entity nodes to form a dynamic heat map. When key operational indicators exceed preset thresholds, dynamic alarm icons are overlaid on the corresponding entity nodes, accompanied by a flashing effect for proactive alerting.

[0013] Preferably, the dynamic optimization based on association mining includes: Based on the relationships and historical operation logs in the graph database, an improved association rule mining algorithm is used to extract strong association rules. An architecture health prediction model based on LSTM neural network is constructed, and intelligent diagnosis is performed by combining the strong correlation rules, and hierarchical optimization suggestions are automatically generated.

[0014] Preferably, the improved association rule mining algorithm is the Apriori-TID algorithm, which is used to mine frequent itemsets from the set of architecturally associated transactions and to filter strong association rules based on support and confidence. The input features of the architecture health prediction model include operational indicator features, association rule features, label matching features, and historical health features; intelligent diagnosis is triggered when the predicted health score is lower than a preset threshold. The intelligent diagnostic process includes: locating abnormal features through feature contribution analysis, matching relevant strong correlation rules to determine the abnormal propagation path, and conducting root cause analysis in conjunction with a historical fault case library; Based on the root cause analysis results, hierarchical optimization suggestions are automatically generated from the technical, architectural, and management levels, and the optimization suggestions are managed, allocated, and their effects are tracked and verified.

[0015] Based on the above technical solution, a unified standard meta-model of architectural elements is constructed to realize the automatic collection and labeling of multi-source heterogeneous architectural data; a multi-source data association model is established based on graph database, and graph computing algorithms are used to mine hidden associations between architectural assets; a multi-dimensional and polymorphic digital architecture graph is formed by combining static topology construction and dynamic data update technologies; finally, intelligent diagnosis and dynamic optimization of architectural problems are achieved through deep learning models.

[0016] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating the dynamic optimization method of a single graph for digital architecture based on multi-source data association mining provided by the present invention. Figure 2 This is a schematic diagram of the standard meta-model in the digital architecture one-graph dynamic optimization method based on multi-source data association mining provided by the present invention; Figure 3 This is a schematic diagram of the graph database-based association relationship modeling in the digital architecture dynamic optimization method based on multi-source data association mining provided by the present invention. Figure 4 This is a flowchart illustrating the end-to-end dynamic optimization mechanism in the digital architecture dynamic optimization method based on multi-source data association mining provided by the present invention. Detailed Implementation

[0018] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0019] It should be noted that the acquisition, transmission, storage, use, and processing of data in the technical solution of this application all comply with relevant laws and regulations. In the embodiments of this application, certain existing industry solutions such as software, components, and models may be mentioned. These should be considered exemplary, intended only to illustrate the feasibility of implementing the technical solution of this application, and do not imply that the applicant has already used or necessarily used such solutions.

[0020] See Figure 1 This invention provides a dynamic optimization method for a single digital architecture graph based on multi-source data association mining. The method includes: constructing a unified standard meta-model for architectural elements to achieve automatic collection and labeling of multi-source heterogeneous architectural data; establishing a multi-source data association model based on a graph database and using graph computing algorithms to mine hidden associations between architectural assets; combining static topology construction and dynamic data update techniques to form a multi-dimensional and polymorphic single digital architecture graph; and finally, using a deep learning model to achieve intelligent diagnosis and dynamic optimization of architectural problems. Specifically: First, such as Figure 2As shown, based on the domain-driven design principle, a standard meta-model of scalable architectural elements covering five core domains—business, application, data, technology, and security—is constructed. This forms a unified architectural asset attribute system and semantic specification, resolving the semantic heterogeneity problem of multi-source data. The meta-model adopts a triple abstract structure, and its expression is:

[0021] in: Domain={Business,Application,Data,Technology,Security} Entity={BusinessProcess,ApplicationSystem,DataTable,...} Attribute={ID,Name,Version,Status,...} The specific definitions and constraints of each dimension are shown in Table 1 below: Table 1

[0022] Secondly, in this embodiment, a multimodal data acquisition framework is adopted, and differentiated acquisition schemes are designed for data with different levels of structure to achieve full, real-time, and low-latency data access, specifically including: 1. Structured data acquisition For structured data (such as user permission tables and API call logs) from systems like the ISC permission platform and the data middleware end-to-end monitoring platform, a combination of RESTful API interface integration and Change Data Capture (CDC) technology is used. (1) Establish a secure connection with the data source system based on the OAuth2.0 authorization protocol, and periodically pull basic static data (such as application system configuration information) through REST API. The pull frequency is configurable (such as once per hour). (2) CDC is implemented using the Debezium open-source component to capture incremental changes in the database in real time (such as adding data table fields and modifying permission policies). Data transmission is buffered using Kafka message queue to ensure that peak data is not lost and the transmission delay is controlled within 100ms. (3) The collected data is uniformly converted to JSON-LD (JSON for Linking Data) format, and semantic association fields are added (such as “@context” to define entity type).

[0023] 2. Semi-structured data acquisition For semi-structured data such as XML configuration files and Excel schema lists, a solution combining XPath / XLST parsing and PySpark data cleaning is adopted: (1) For XML files, extract key entities (such as...) using XPath expressions.<ApplicationSystemID="APP002"> ), using XLST to achieve format standardization conversion; (2) For Excel files, after reading the data using Pandas, data quality is ensured through cleaning operations such as data type validation (e.g., system ID format validation), null value filling (filling with default values ​​based on business rules), and deduplication of duplicate data (based on entity ID primary key); (3) The cleaned data is stored in the Data Lake and the Delta Lake format is used to implement data version control and ACID transaction support.

[0024] 3. Unstructured data acquisition For unstructured data such as architecture design documents (PDF format) and system architecture drawings (PNG / JPG format), a collaborative data collection chain of "web crawler + OCR + NLP" is constructed: (1) The Scrapy crawler framework was used to crawl the architecture documents in the internal document management system. The legality and stability of the data collection were ensured through robots protocol compliance verification and IP rotation mechanism. (2) For scanned PDF or image-type drawings, the TesseractOCR engine (combined with industry-customized training sets) is used for text extraction, and the character recognition accuracy rate is over 98%. The extracted text content is used to perform entity recognition using a BERT-BiLSTM-CRF pre-trained model. The model is fine-tuned based on architecture domain corpus (such as power industry architecture documents) and can accurately identify core entities such as "business process ID, application system name, and data table fields". The entity recognition F1 score is 0.92.

[0025] Next, based on entity recognition results and domain knowledge graphs, a multi-level labeling system is constructed. A combination of TF-IDF and TextRank algorithms is used to generate label weights, achieving a balance between label accuracy and coverage. This includes: 1. Labeling system design Tags are divided into three levels, corresponding to the entity's "attribute features, functional features, and association features," such as: First-level tags (attribute features): generated based on the core attributes of the entity, such as the first-level tags "Running, V2.3.1 version, State Grid Cloud deployment" for the application system "APP001"; Secondary tags (functional features): generated based on the core functions of the entity, such as the secondary tags "permission management, user authentication, role assignment" for "APP001"; Level 3 tags (related features): generated based on entity relationships, such as the level 3 tags of "APP001" being "related business process P001, dependent data table user_info".

[0026] 2. Tag weight calculation First, the local weights of terms in a single document are calculated using the TF-IDF algorithm, as shown in the following formula:

[0027] Where TF(t,d) is the frequency of term t in document d (normalized using "term frequency / total number of words in document").

[0028] Where N is the total number of documents collected, DF(t) is the number of documents containing the term t, and 1 is added to the denominator to avoid zero values.

[0029] Next, the importance weights of terms in the global document set are calculated using the TextRank algorithm, and a term co-occurrence network is constructed. The node weights are then calculated iteratively (similar to PageRank), as shown in the following formula:

[0030] Where (WS(v_i)) is the weight of term (v_i), d is the damping coefficient (taken as 0.85), (In(v_i)) is the set of terms pointing to (v_i), and (Out(v_j)) is the set of terms pointing to (v_j).

[0031] The final label weight is the weighted sum of the TF-IDF weight and the TextRank weight (with weight coefficients of 0.6 and 0.4 respectively). The top 3 terms by weight are used as the core entity labels. Example: {"Entity ID":"APP001","Entity Type":"ApplicationSystem","Label":["Running (Weight 0.92)","Permission Management (Weight 0.88)","Associated with P001 (Weight 0.81)"]}.

[0032] Then, based on the relationship modeling of the graph database, such as... Figure 3 As shown, an attribute graph model is used to construct a multi-source data relationship model. Leveraging the efficient relational query capabilities of graph databases, explicit and implicit dependencies between architectural elements are uncovered, addressing the problem of low efficiency in relational queries in traditional relational databases. Specifically, this includes: Based on the RDF triple specification, design a three-layer graph structure of "entity-relationship-attribute": 1. Entity Node Each entity node corresponds to an entity in the standard metamodel (such as BusinessProcess or ApplicationSystem). The node attributes include all the core attributes of the entity (such as ID, Name, and Status), and also add management attributes such as "data source" and "updateTime". Nodes are encoded with unique IDs (such as “BusinessProcess_P001” or “ApplicationSystem_APP001”) to ensure the uniqueness of entities across data sources.

[0033] 2. Relationship edge Relationship edges define the types of associations between entities, covering eight core relationships such as "Call," "DependOn," "Contain," and "Authorize," which conform to the business logic of the architecture domain, such as: Business process → Call → Application system (e.g., P001 → Call → APP002). Application System → Dependencies → Data Tables (e.g., APP002 → DependOn → user_info); Application system → contains → functional modules (e.g., APP001 → Contain → User authentication module). Access control policy → Authorization → Role (e.g., SP001 → Authorize → Administrator role); The relationship edge attributes include "relationship strength (calculated based on historical interaction frequency)", "creation time (createTime)", and "last update time (lastUpdateTime)", such as: (APP002,DependOn,user_info,{strength:0.9,createTime:"2024-01-10",lastUpdateTime:"2024-05-20"}).

[0034] 3. Graphical Model Index Design In this implementation, to improve the efficiency of relational queries, a composite index and a full-text index are created in the graph database: Composite indexes are created based on "entity type + core attribute", such as (EntityType, ID) or (EntityType, Name), and support quick location of entities of specific types. Full-text index: Creates full-text indexes for text attributes such as entity names and tags, supporting fuzzy searches (e.g., "search for application systems whose names contain 'permissions'").

[0035] This invention addresses the problems of entity duplication (such as different IDs for the same application system in ISC and data platform) and attribute conflict (such as different CPU utilization values ​​for the same server in different monitoring tools) in multi-source data by designing a two-stage fusion algorithm of "entity linking + attribute fusion". 1. Entity Linking The following are the specific steps for determining whether entities from different data sources are the same entity using multi-dimensional similarity calculation: (1) Name similarity calculation (Sim_name) The method combines edit distance (LevenshteinDistance) and Jaccard similarity, with the following formula: Edit distance: (Editing operations include insertion, deletion, and replacement; the higher the value, the higher the similarity.) Jaccard similarity: (S1 and S2 are the word segmentation sets of entity names); final

[0036] (2) Attribute similarity calculation (Sim_attr) For numerical attributes (such as CPU utilization), normalized Euclidean distance is used to calculate similarity; for enumerated attributes (such as running status), a method of "1 point for a perfect match and 0 points for a non-match" is used; finally... This is a weighted average of the similarity of each attribute (the weights are configured based on the importance of the attribute, such as ID weight 0.5, state weight 0.3, and version number weight 0.2).

[0037] (3) Comprehensive calculation of entity similarity

[0038] Where α=0.6 and β=0.4 (determined through grid search optimization), when When the confidence level is greater than 0.8, the entities are identified as the same entity and merged. The attributes of the merged entity are selected based on the principle of "prioritizing data sources with higher confidence" (e.g., the application system status confidence of the ISC system is higher than that of the log system).

[0039] 2. Attribute Fusion To address attribute value conflicts for the same entity across different data sources, a weighted average fusion algorithm (for numerical attributes) and a voting fusion algorithm (for enumerated attributes) are employed: Numerical attributes (such as CPU utilization): ,in, The confidence weight for data source i (calculated based on historical data accuracy, such as a confidence score of 0.8 for monitoring tool A and 0.2 for monitoring tool B). Enumerated attributes (such as running status): Count the frequency of attribute values ​​for each data source, and take the attribute value with the highest frequency as the final value. If the frequencies are the same, take the attribute value corresponding to the data source with the highest confidence.

[0040] Next, a visualization architecture driven by a dual engine of "static topology + dynamic data" is adopted, combining layered rendering and real-time data stream processing technology to achieve a holographic, real-time display of the architecture status, enabling operations and maintenance personnel to quickly grasp the overall architecture and local details. Specifically, this includes: A hybrid layout strategy combining Force-DirectedLayout and HierarchicalLayout is used to construct a structured and readable static architecture topology diagram. 1. Hierarchical division: Vertically, it is divided into five levels: "Business Domain → Application Domain → Data Domain → Technology Domain → Security Domain". Entity nodes in each level adopt the same basic style (e.g., business domain nodes are blue circles, application domain nodes are green squares) to ensure visual consistency. 2. Force-oriented layout calculation: The Fruchterman-Reingold algorithm is used to calculate node positions. By simulating "repulsive forces between nodes" and "pull forces on edges," the topology graph achieves a balance between "compactness" and "edge crossover rate." Repulsive force: Where k is the ideal distance and d(u,v) is the distance between nodes u and v; pull: ; Node displacement: Calculate the x and y displacements of the nodes based on the resultant force, and iterate until the layout is stable (the displacement is less than the threshold, such as 1px). 3. Local optimization: For entity nodes in the same department / region, the community detection algorithm (Louvain algorithm) is used to divide them into communities. Nodes in the same community are identified by the same background color, which improves the business relevance and readability of the topology map. 4. Topology Rendering: Employs WebGL hardware-accelerated rendering technology and uses the D3.js visualization library to achieve interactive rendering of static topology maps, supporting the following core functions: (1) Hierarchical collapse / expansion: Clicking on a certain level title (such as "Application Domain") will collapse / expand all entity nodes under that level, adapting to large screen and small screen display scenarios; (2) Entity filtering: Supports filtering by dimensions such as "entity type (e.g., application system / data table)", "operation status (e.g., running / maintaining)" and "department". The filtering results are updated in real time to update the topology map; (3) Viewing node details: When the mouse hovers over an entity node, a tooltip pop-up window containing all the entity’s attributes (such as system ID, version number, and associated entities) is displayed. Clicking on the node will take you to the entity details page.

[0041] Based on the streaming framework (Apache Flink), real-time data processing of the architecture's runtime is achieved by overlaying the real-time status onto a static topology diagram using heatmap rendering and dynamic metric annotation techniques, thus realizing dynamic visualization of the architecture's status. 1. Real-time data access and processing (1) Data access: Real-time reading of system monitoring tools (such as Zabbix, Prometheus) running data through the FlinkCDC connector, including indicators such as CPU utilization, memory usage, interface response time, and error rate, with data access latency controlled within 500ms; (2) Data preprocessing: Flink’s window function (SlidingWindow, sliding window size 1 minute, step size 10 seconds) is used to aggregate and calculate real-time data to eliminate instantaneous fluctuation interference, such as calculating the average CPU utilization and the 95th percentile of interface response time within one minute. (3) Data association: The preprocessed running data is associated with the entity nodes in the graph database through the "entity ID" to generate "entity-real-time indicator" key-value pairs, which are stored in the Redis cache (expiration time 5 minutes) for quick query by the visualization layer.

[0042] 2. Generation of dynamic heatmaps A thermal value calculation model is built based on real-time operational metrics, and the entity's operational load and health status are intuitively displayed through color gradient changes: (1) The formula for the thermal value is:

[0043] in: The normalized CPU utilization (value range [0,1], calculated using the following formula): , (actual CPU utilization). The normalized interface response time (value range [0,1], calculation formula: , (This refers to the interface response time threshold, such as 500ms). The normalized error rate (value range [0,1], calculated using the following formula): , (actual error rate) w1, w2, and w3 are the indicator weights (determined by the Analytic Hierarchy Process (AHP), such as w1=0.4, w2=0.3, w3=0.3), and satisfy w1+w2+w3=1.

[0044] (2) Color mapping rule: Map the corresponding RGB colors according to the range of HeatValue values ​​to form a gradient change: HeatValue∈[0,0.3]: Cool color (RGB(0,255,0), green), indicating good operating status; HeatValue∈(0.3,0.7]: intermediate color (RGB(255,255,0), yellow), indicating a normal operating status; HeatValue∈(0.7,1]: Warm color (RGB(255,0,0), red), indicating abnormal operation status, requiring close attention.

[0045] (3) Heatmap rendering: The mapping between HeatValue and color is implemented through the scaleLinear() function of D3.js, and the color is assigned to the entity node in the topology map. At the same time, the mouse hover is supported to view specific indicator values ​​(such as "CPU utilization: 35%, interface response time: 200ms, error rate: 0.1%").

[0046] 3. Dynamic indicator labeling For key operational metrics (such as interface response timeouts and sudden increases in error rates), dynamic icon labeling and alarm flashing technology are used to proactively alert users to abnormal states. (1) Icon labeling rules: Interface response timeout (>threshold): Mark a "clock" icon in the upper right corner of the application system node; A sudden increase in error rate (more than 50% month-on-month): mark the application system node with an exclamation mark icon in the upper right corner; CPU utilization is too high (>80%): Mark the "CPU" icon in the upper right corner of the server node; (2) Alarm blinking effect: The icon blinking effect is achieved using the CSS3 animation property (e.g., "show for 0.5 seconds, hide for 0.5 seconds"). It also supports clicking the icon to view alarm details (e.g., alarm time, abnormal indicator value, historical trend).

[0047] See Figure 4In this implementation, a dynamic optimization mechanism is constructed that integrates "association rule mining - health prediction - intelligent diagnosis - optimization suggestions" across the entire chain. By combining graph computing and deep learning technologies, early warning and precise optimization of architectural problems can be achieved, solving the problems of low efficiency and strong lag in traditional manual optimization.

[0048] First, association rule mining is performed. Based on the association data in the graph database and historical operation logs, an improved Apriori algorithm (Apriori-TID algorithm) is used to mine frequent itemsets and strong association rules between architectural elements, identifying hidden dependencies and potential risks. The specific steps are as follows: 1. Data preprocessing and transaction set construction (1) Data filtering: Extract the architecture relationship data of the past three months from the graph database (such as the call relationship of "business process-application system-data table"), and extract the abnormal event data of the same period from the operation log (such as "application system failure-data table synchronization delay"). (2) Transaction set construction: Each "multi-entity association record" or "abnormal event association record" is converted into a transaction, such as "Transaction 1: {P001,APP002,I003,user_info}" and "Transaction 2: {APP002 failure, T003 synchronization delay}", which eventually forms a transaction set (T={T1,T2,...,Tn}) (n is the total number of transactions).

[0049] 2. Frequent itemset mining (Apriori-TID algorithm) Compared to the traditional Apriori algorithm, the Apriori-TID algorithm reduces the number of candidate set generation times by storing a list of transaction IDs (TIDList), thus improving the efficiency of large-scale data processing. The specific steps are as follows: (1) Generate 1-itemsets: Traverse the transaction set, count the frequency of occurrence of each entity (such as P001, APP002), filter out the entities that meet the minimum support (Min_Support=0.3, configured according to business needs) to form 1-frequent itemsets (L1); (2) Generate candidate set and TIDList: Perform a self-join on Lk ((k≥1)) to generate candidate set Ck+1, and record the TIDList (containing the transaction ID set of the candidate set) corresponding to each candidate set. (3) Filter frequent itemsets: Calculate the support of candidate set Ck+1 (support = length of candidate set TIDList / total number of transactions), filter out candidate sets with support ≥ Min_Support, and form k+1-frequent itemsets Lk+1; (4) Termination condition: When no new frequent itemset can be generated (i.e., Lk+1 is empty), stop the iteration and output all frequent itemsets (e.g., L2={P001→APP002,APP002→user_info}, L3={P001→APP002→user_info}).

[0050] 3. Generation of strong association rules The confidence score of association rules is calculated based on frequent itemsets, and strong association rules are selected using the following formula: Support: Where X and Y are entity subsets, X∩Y= count(X∪Y) is the number of transactions containing X∪Y; Confidence level: ; Filtering criteria: When Support(X→Y)≥0.3 and Confidence(X→Y)≥0.7, it is determined to be a strong association rule, for example: Rule 1: {Business Process P001} → {Application System APP002} (Support=0.45, Confidence=0.92); Rule 2: {Application System APP002, Data Table T003} → {Synchronization Delay} (Support=0.32, Confidence=0.85).

[0051] 4. Storage and updating of association rules The mined strong association rules are stored in a graph database and associated with the corresponding entity nodes in the form of "rule node - association edge". At the same time, an incremental mining mechanism is adopted (the association rule mining is re-executed every seven days) to update the support and confidence of the rules and ensure the timeliness of the rules.

[0052] Secondly, intelligent diagnosis and optimization suggestions are provided. An architecture health prediction model is built based on an LSTM (Long Short-Term Memory) neural network. This model, combined with association rules, enables intelligent diagnosis of architectural problems and generates actionable optimization suggestions. The specific steps are as follows: 1. Architecture Health Prediction Model (LSTM) (1) Feature engineering: Select the following four types of features to construct the model input vector Feature(t) (t is the current time step): Operational metrics: Normalized CPU utilization, interface response time, and error rate (consistent with the metrics in the dynamic heatmap). Characteristics of association rules: Support and confidence of strong association rules (e.g., Support=0.32 and Confidence=0.85 for the rule “APP002→T003”). Tag matching features: the degree of matching between entity tags and domain standard tags (calculated using Jaccard similarity, such as the tag matching degree of "APP001" = 0.9). Historical health characteristics: Health scores for the first three time steps: HealthScore(t-1), HealthScore(t-2), and HealthScore(t-3).

[0053] (2) Model training and evaluation: Training data: Six months of architecture operation data of a power grid company (including features and health labels for 100,000+ time steps, with health labels annotated by experts based on actual operating status). Optimizer: Adam optimizer is used, learning rate = 0.001, batch size = 32, training epochs = 50; Evaluation metrics: The model performance was evaluated using MAE (mean absolute error) and R² (coefficient of determination). After training, MAE < 0.05 and R² > 0.9, which met the accuracy requirements for health prediction.

[0054] (3) Health prediction: The model receives the feature vector Feature(t) at the current time step in real time and outputs the health score HealthScore(t+1) for the next time step, as follows: HealthScore(t+1)=LSTM(HealthScore(t),Feature(t)); 2. Intelligent diagnostic mechanism When HealthScore(t+1) < 0.6 (health threshold, configured according to business needs), the intelligent diagnosis process is triggered, and the specific steps are as follows: (1) Abnormal feature localization: Calculate the contribution of each input feature to the health score (using SHAP value, SHapleyAdditiveexPlanations), and filter out the top 3 abnormal features with the highest contribution (such as "interface response time = 0.95 (after normalization)" and "APP002→T003 rule confidence = 0.5 (below the threshold of 0.7)"). (2) Association rule matching: Query the graph database for strong association rules related to the abnormal features, such as the rule "APP002→T003", to locate the abnormal propagation path (such as "T003 synchronization delay → APP002 interface response timeout → P001 business process lag"). (3) Root cause diagnosis: Combining the abnormal propagation path with the historical fault case library (which stores the mapping relationship between "abnormal phenomenon-root cause-solution"), the KNN (K-nearest neighbor) algorithm is used to match the most similar historical cases to determine the root cause of the current abnormality, such as: "The synchronization delay of the data table T003 on which the application system APP002 depends is too high, which causes the APP002 interface to time out, thus affecting the operation of the business process P001."

[0055] 3. Optimization suggestion generation Based on the root cause diagnosis results, combined with the domain knowledge graph and optimization strategy library, hierarchical optimization suggestions are automatically generated, specifically: (1) Technical recommendations: Optimize infrastructure and system configuration, such as "upgrade the CPU of the server where data table T003 is located to 16 cores and expand the memory to 64GB to improve data synchronization processing capabilities" and "add a Redis cache layer to cache high-frequency query data of table T003 and reduce the number of direct calls from APP002 to T003 (expected to reduce the number of calls by 40%)". (2) Architecture layer suggestions: Optimize the architecture design and dependencies, such as "migrate the interface I003 of APP002 to the data middle platform, use the middle platform service to encapsulate the data access logic, and reduce the direct dependency on the T003 table" and "reconstruct the business process P001, split non-core steps (such as log recording) into asynchronous processing, and reduce the synchronous calls to the APP002 interface". (3) Management recommendations: Optimize the operation and maintenance process and monitoring mechanism, such as "add a synchronization delay alarm rule for table T003, triggering an SMS alarm when the synchronization delay is >100ms" and "perform an audit of association rules once a week to check the effectiveness of strong association rules and promptly identify potential dependency risks".

[0056] 4. Follow up on the implementation of optimization suggestions The generated optimization suggestions are stored in the optimization task management module, supporting the following functions: (1) Recommendation classification: Based on the urgency of optimization (based on health score and scope of impact), the recommendations are divided into three levels: "urgent (HealthScore<0.4)", "important (0.4≤HealthScore<0.6)" and "general (HealthScore≥0.6)". (2) Task assignment: Supports assigning optimization tasks to designated maintenance personnel, setting completion deadlines, and notifying via email / WeChat Work; (3) Effect verification: After the optimization task is completed, the model automatically recalculates the health score and compares the changes in indicators before and after optimization (such as "after optimization, the T003 synchronization delay is reduced from 300ms to 50ms, the APP002 interface response time is reduced from 800ms to 200ms, and the HealthScore is increased from 0.55 to 0.88"), and generates an optimization effect report.

[0057] In summary, this invention achieves unified management of structured and unstructured data through standard meta-models and entity linking technology, improving data fusion accuracy by 80%; based on graph computing and deep learning-based association mining, the response time for architecture problem diagnosis is shortened to the second level, and the accuracy of optimization suggestions reaches 90%; furthermore, by utilizing multi-layer overlay and heatmap technology, the holographic perception capability of architecture status is improved by 70%.

[0058] The following provides a specific embodiment to illustrate the present invention: Experimental environment: Hardware: Server (8-core CPU, 32GB RAM), Graphics workstation (GPU NVIDIA A100). Software: Neo4j 5.0 graph database, Python 3.9 (TensorFlow 2.8, NetworkX 2.8) Dataset: 1000+ architecture assets of a power grid company (including 200 business processes, 150 application systems, and 500 data tables); The specific implementation steps are as follows: 1. During the data acquisition phase, web crawlers and API acquisition components were deployed to collect data from six data sources, including the ISC system and data platform, with an average daily data processing volume of 10GB and a tagging accuracy rate of 85%. 2. In the graph database modeling stage, a graph model containing 50,000+ nodes and 100,000+ relationships was built in Neo4j, and the duplicate entity recognition accuracy of the entity linking algorithm reached 92%. 3. The architecture is built from a single map, using D3.js to overlay static topology and dynamic heatmaps, supporting 5 levels of zoom and layer switching, with a response time of less than 1 second; 4. In the dynamic optimization and verification phase, 30 strongly correlated rules were discovered using the Apriori algorithm. The health prediction error of the LSTM model was less than 5%, and the adoption rate of optimization suggestions reached 80%.

[0059] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0060] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0061] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0062] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0063] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0064] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0065] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0066] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0067] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for dynamic optimization of a single graph in a digital architecture based on multi-source data association mining, characterized in that, The method includes: Collect multi-source architecture data and perform tagging processing; Modeling is based on the relationships in a graph database; Dynamically construct a single diagram of a digital architecture; Dynamic optimization is performed based on association mining.

2. The method for dynamic optimization of a single graph of a digital architecture based on multi-source data association mining according to claim 1, characterized in that, The process of collecting and tagging multi-source architecture data includes: Construct a standard meta-model of architectural elements covering business, applications, data, technology, and security domains; Based on this standard meta-model, a differentiated acquisition strategy is adopted to automatically collect architecture data from multi-source heterogeneous systems and perform entity recognition on unstructured data. A multi-level labeling system is constructed based on the recognition results and the domain knowledge graph, and the label weights are calculated.

3. The method for dynamic optimization of a single graph of a digital architecture based on multi-source data association mining according to claim 2, characterized in that, The standard meta-model is defined using a triplet structure according to formula (1). ，（1） in, This includes business domains, application domains, data domains, technology domains, and security domains; For structured data, RESTful API combined with CDC technology is used for real-time or near real-time data collection. For semi-structured data, a combination of XPath / XLST parsing and data cleaning is used for data collection. For unstructured documents and drawings, a collaborative link of "web crawling + OCR + entity recognition based on pre-trained language model" is used for collection and structured extraction. The tagging system includes first-level tags based on the core attributes of entities, second-level tags based on the core functions of entities, and third-level tags based on the relationships between entities; and the tag weights are calculated using a weighted method that combines the TF-IDF algorithm and the TextRank algorithm.

4. The method for dynamic optimization of a single graph of a digital architecture based on multi-source data association mining according to claim 3, characterized in that, When calculating label weights, According to formula (2), the local weights of terms in a single document are calculated using the TF-IDF algorithm. ，（2） Where TF(t,d) is the frequency of term t in document d, and is calculated according to formula (3). , ，（3） Where N is the total number of documents collected, DF(t) is the number of documents containing term t, and 1 is added to the denominator to avoid zero values; The importance weights of terms in the global document set are calculated using the TextRank algorithm. A term co-occurrence network is constructed, and the node weights are iteratively calculated according to formula (4). ，（4） Where (WS(v_i)) is the weight of term (v_i), d is the damping coefficient, (In(v_i)) is the set of terms pointing to (v_i), and (Out(v_j)) is the set of terms pointing to (v_j). The final label weight is the weighted sum of TF-IDF weight and TextRank weight, and the top 3 terms with the highest weights are used as the core entity labels.

5. The method for dynamic optimization of a single graph of a digital architecture based on multi-source data association mining according to claim 1, characterized in that, The modeling of relationships based on graph databases includes: An attribute graph model is used, with the processed entities as nodes and the relationships between entities as edges, to construct a multi-source data association model; Design an entity linking and attribute fusion algorithm to solve the problems of entity duplication and attribute conflict in multi-source data.

6. The method for dynamic optimization of a single graph of a digital architecture based on multi-source data association mining according to claim 5, characterized in that, The attribute graph model includes entity nodes, relationship edges and their attributes. Entity nodes have unique ID codes, attributes include core attributes and management attributes, and relationship edges are defined to include various association types such as invocation, dependency, inclusion, and authorization, and include relationship strength attributes. The entity linking algorithm calculates the name similarity and attribute similarity between entities, and then weights them to obtain a comprehensive similarity. When the comprehensive similarity exceeds a preset threshold, the entities are determined to be the same entity and are merged. The attribute fusion algorithm uses a weighted average fusion for numerical attributes and a voting fusion for enumerated attributes.

7. The method for dynamic optimization of a single graph of a digital architecture based on multi-source data association mining according to claim 1, characterized in that, The dynamically constructed digital architecture diagram includes: A static architecture topology diagram is constructed using a hierarchical layout and force-directed layout algorithm; By accessing and processing real-time running data through a stream processing framework, the calculated heat values ​​and rules are dynamically superimposed onto a static topology map to generate an architecture map that includes dynamic heat maps and indicator labels.

8. The method for dynamic optimization of a single graph of a digital architecture based on multi-source data association mining according to claim 7, characterized in that, The static topology graph is vertically divided into multiple levels according to domains. The force-directed layout algorithm is used to calculate the node positions, and the community detection algorithm is combined for local optimization. The interactive rendering is performed using WebGL technology. Real-time running data is accessed and windowed aggregated through a stream processing framework and associated with entity nodes in the graph database. A heat value calculation model is built based on CPU utilization, interface response time, and error rate indicators. The heat values ​​are mapped to color gradients and rendered on entity nodes to form a dynamic heat map. When key operational indicators exceed preset thresholds, dynamic alarm icons are overlaid on the corresponding entity nodes, accompanied by a flashing effect for proactive alerting.

9. The method for dynamic optimization of a single graph of a digital architecture based on multi-source data association mining according to claim 1, characterized in that, The dynamic optimization based on association mining includes: Based on the relationships and historical operation logs in the graph database, an improved association rule mining algorithm is used to extract strong association rules. An architecture health prediction model based on LSTM neural network is constructed, and intelligent diagnosis is performed by combining the strong correlation rules, and hierarchical optimization suggestions are automatically generated.

10. The method for dynamic optimization of a single graph of a digital architecture based on multi-source data association mining according to claim 9, characterized in that, The improved association rule mining algorithm is the Apriori-TID algorithm, which is used to mine frequent itemsets from the set of architecturally associated transactions and to filter strong association rules based on support and confidence. The input features of the architecture health prediction model include operational indicator features, association rule features, label matching features, and historical health features; intelligent diagnosis is triggered when the predicted health score is lower than a preset threshold. The intelligent diagnostic process includes: locating abnormal features through feature contribution analysis, matching relevant strong correlation rules to determine the abnormal propagation path, and conducting root cause analysis in conjunction with a historical fault case library; Based on the root cause analysis results, hierarchical optimization suggestions are automatically generated from the technical, architectural, and management levels, and the optimization suggestions are managed, allocated, and their effects are tracked and verified.

Images