Automatic filling and examination and knowledge base integrated optimization method based on dynamic rule engine

Abstract

The invention provides an automatic filling and reviewing and knowledge base integrated optimization method based on a dynamic rule engine, and relates to the technical field of software engineering, and the method comprises the steps: carrying out the standardization processing of electronic form data, and generating a to-be-reviewed data set based on a preset filling rule. And the constructed dynamic rule engine realizes time sequence tracking and self-adaptive recombination of the rule, a scenarized rule chain is generated for rule matching, and finally an examination result is obtained. According to the method, through multi-level knowledge modeling and knowledge dual network construction, potential knowledge conflicts are detected, the repair strategy is generated, and the accuracy and efficiency of filling and examination are remarkably improved.

Description

Technical Field

[0001] The present invention relates to software engineering technology, and in particular to an automatic reporting review and knowledge base integration optimization method based on a dynamic rule engine. Background Art

[0002] Automatic filling review and knowledge base integration of electronic forms is an important part of modern information management. With the deepening of digital transformation of enterprises, more and more business processes rely on the processing and analysis of electronic form data. Traditional form review methods mainly rely on manual review, which is inefficient and prone to errors. The existing automatic filling review and knowledge base integration technology based on rule engines still has some defects and shortcomings: Insufficient flexibility in rule configuration: Existing rule engines usually use static rule configuration, which is difficult to adapt to complex business scenario changes. When business rules change, rule configuration needs to be manually modified, which has high maintenance costs and is prone to human errors.

[0003] The maintenance cost of the knowledge base is high: The construction and maintenance of the knowledge base requires a lot of manpower and time. As the business continues to develop, the scale of the knowledge base will continue to expand, and the updating and maintenance of knowledge will become more difficult.

[0004] Difficulty in detecting and repairing knowledge conflicts: In practical applications, there may be some contradictory or conflicting knowledge in the knowledge base. These knowledge conflicts will lead to inconsistent execution results of the rule engine and even lead to wrong decisions. Existing technologies lack effective knowledge conflict detection and repair mechanisms, making it difficult to ensure the consistency and integrity of the knowledge base. Summary of the invention

[0005] The embodiment of the present invention provides an automatic reporting review and knowledge base integration optimization method based on a dynamic rule engine, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention, Provide automatic reporting review and knowledge base integration optimization method based on dynamic rule engine, including: Standardize the format of electronic form data to obtain pre-processed data, perform structured analysis on the pre-processed data according to preset reporting rules, and generate a data set to be reviewed; Build a dynamic rule engine based on the rule time sequence graph, track the rules in time sequence, decompose the rules into atomic rule units, adaptively reorganize the atomic rule units according to the business scenario, generate a scenario-based rule chain, use the scenario-based rule chain to match the rules of the data set to be reviewed, and obtain the review result; Perform multi-level knowledge modeling on the review results and construct a vertically layered knowledge system; establish a knowledge dual network based on the knowledge system, calculate the time series distribution of the cross entropy between the original network and the dual network, perform curve fitting and extrapolation on the entropy value change trend, combine with dynamic prediction of network consistency, detect and locate potential knowledge conflicts, generate repair strategies, and update the repaired knowledge to the knowledge base.

[0007] In an optional embodiment, The steps of building a dynamic rule engine based on the rule time sequence graph, tracking the rules in time sequence, decomposing the rules into atomic rule units, and adaptively reorganizing the atomic rule units according to the business scenarios to generate scenario-based rule chains include: Construct a rule time series graph, represent the rule node features based on the spatiotemporal feature tensor; atomically decompose the rules based on semantic integrity; use the scenario feature vector to drive the rule reorganization and generate a scenario-based rule chain, including: Construct a rule timing graph, use business rules as nodes of the rule timing graph, use the timing dependencies between rules as directed edges of the rule timing graph, construct a spatiotemporal feature tensor for the nodes, the spatiotemporal feature tensor includes the timing attributes, spatial constraints and business attributes of the rules, construct a rule timing matrix based on the spatiotemporal feature tensor, perform feature enhancement on the rule timing matrix, construct a multidimensional association matrix including rule timing features and business attributes, and perform timing tracking on the rules according to the time overlap and execution order relationship between the rules in the multidimensional association matrix; Based on the rule time series graph, the rules are atomically decomposed into atomic rule units that maintain semantic integrity, the semantic relevance between the atomic rule units is calculated, and when the semantic relevance is less than a preset boundary threshold, the separation boundary of the atomic rule unit is determined to generate an atomic rule unit set; A scene feature vector is constructed, and the fitness score between each atomic rule unit in the atomic rule unit set and the scene feature vector is calculated. According to the fitness score, a dynamic programming algorithm is used to adaptively reorganize the atomic rule units to generate a scene-based rule chain.

[0008] In an optional embodiment, A spatiotemporal feature tensor is constructed for the node, wherein the spatiotemporal feature tensor includes a regular time series attribute, a spatial constraint, and a business attribute; a regular time series matrix is ​​constructed based on the spatiotemporal feature tensor; and feature enhancement is performed on the regular time series matrix. The steps of constructing a multidimensional association matrix including regular time series features and business attributes include: Constructing a spatiotemporal feature tensor of a rule node, the spatiotemporal feature tensor comprising a time series feature vector, a space feature vector and a business feature vector, wherein the time series feature vector comprises the effective time, expiration time, maximum execution time and minimum execution time of the rule, the space feature vector comprises the scope, geographical location restriction and management level of the rule, and the business feature vector comprises the priority, rule type and business field of the rule; Performing time series dimension projection on the spatiotemporal feature tensor to obtain a time series feature matrix, calculating the time overlap and execution order relationship between the rules to obtain the time series relationship strength, and constructing a rule time series matrix based on the time series feature matrix and the time series relationship strength; Calculating the similarity of the spatial feature vectors to obtain spatial association strength, and calculating the correlation of the service feature vectors to obtain service association strength; Performing feature enhancement on the regular time series matrix, performing multi-dimensional feature fusion on the regular time series matrix, the spatial association intensity and the business association intensity, and generating an enhanced association matrix; Construct a query matrix, a key-value matrix and a numerical matrix, calculate the attention weight based on the enhanced association matrix, perform weighted fusion on the attention weight and the enhanced association matrix, and generate a multidimensional association matrix containing regular high-order association relationships.

[0009] In an optional embodiment, The steps of performing multi-level knowledge modeling on the review results and constructing a vertically hierarchical knowledge system include: Extract business entities, rule entities, result entities and time entities from the review results, build an entity relationship set including trigger relationships, generation relationships and timing relationships, and extract an attribute set including business type, rule identifier, result type, timestamp and confidence level; Based on the entity relationship set and the attribute set, a conceptual hierarchical relationship among business, rules, results and time is constructed, the semantic similarity between concepts is calculated, a concept clustering tree is constructed according to the semantic similarity, and semantic association rules between concepts are extracted; The knowledge system is divided into a data layer, a rule layer, a reasoning layer, and an application layer. The data layer processes time series data streams. The rule layer generates context-aware rules based on the time series data streams. The reasoning layer performs state-dependent reasoning according to the context-aware rules. The application layer builds scenario-based applications based on the state-dependent reasoning. The semantic association rules are used to construct inter-layer context information, and bottom-up knowledge extraction is performed to obtain inter-layer knowledge mapping relationships. Global constraints are determined based on the inter-layer knowledge mapping relationships, and top-down knowledge guidance is performed. A knowledge evolution equation is constructed by combining external environment information and internal consistency requirements, a knowledge update gradient is calculated according to the knowledge evolution equation, and an adaptive knowledge update strategy is determined by combining the knowledge update gradient and the knowledge drift amount; According to the inter-layer knowledge mapping relationship, the difference between the upper and lower layer knowledge representations and the degree of violation of the structural constraints are calculated to obtain the consistency deviation, the knowledge coverage of the target business scenario and the missing of the knowledge link are counted to obtain the integrity index, and the decision accuracy of the knowledge application and the degree of adaptability to the scenario are analyzed to obtain the effectiveness index; a weighted objective function of the consistency deviation, integrity index and effectiveness index is constructed, and the knowledge structure integrity and semantic consistency constraints are set based on the weighted objective function. The indicators that violate the constraints are optimized by combining the gradient descent and the coordinate alternating optimization method, and the weights of each indicator are dynamically adjusted during the optimization process, and the optimization step size is adaptively adjusted according to the application feedback; Calculate the accuracy, coverage and timeliness of knowledge to obtain knowledge quality assessment results, and dynamically adjust the knowledge structure based on the assessment results.

[0010] In an optional embodiment, The steps of establishing a knowledge dual network based on the knowledge system, calculating the time series distribution of the cross entropy between the original network and the dual network, performing curve fitting and extrapolation on the entropy value change trend, combining dynamic prediction of network consistency, detecting and locating potential knowledge conflicts, generating a repair strategy, and updating the repaired knowledge to the knowledge base include: Construct a knowledge dual network; calculate dynamic cross entropy based on the time decay factor; use the attention mechanism to fuse multi-dimensional prediction features, and evaluate network consistency through local to global hierarchical calculations; dynamically adjust the warning threshold based on historical trigger frequency and activity; locate the conflict source and generate a repair strategy based on the betweenness centrality, including: The association relationship between knowledge nodes is mapped into a knowledge graph structure. The association strength between nodes is calculated based on the feature vector representation of the nodes and the node association set. The dual network is constructed by reversing the association direction between nodes in the original network and recalculating the weights. Obtain the probability distribution of nodes in the original network and the dual network to calculate the network cross entropy, introduce a time decay factor to dynamically weight the network cross entropy, and obtain the cross entropy sequence of the original network and the dual network within a preset time window; construct a polynomial kernel function to map the cross entropy sequence to a high-dimensional feature space, extract a time-varying feature vector, and the time-varying feature vector includes a feature mapping component calculated by the kernel function; perform curve fitting on the time series of the cross entropy based on the time-varying feature vector, and perform extrapolation prediction on the fitting curve; Based on the historical data of the cross entropy sequence, network topology features and node semantic similarity, a multidimensional prediction feature is constructed, and the multidimensional prediction feature is dynamically fused by using an attention mechanism to obtain a fusion feature. According to the fusion feature, hierarchical prediction is performed according to the knowledge field, and network consistency is evaluated through local to global hierarchical calculation. Based on the historical trigger frequency and knowledge update activity, an adaptive threshold is calculated for early warning judgment, and conflict detection of the corresponding level is triggered according to the early warning result; Calculate the betweenness centrality of nodes in the original network and the dual network, locate the conflict source node based on the change in the betweenness centrality, generate a repair strategy based on the graph edit distance for the conflict source node; and update the repair strategy to the knowledge base.

[0011] In an optional embodiment, The steps of performing hierarchical prediction according to the knowledge domain based on the fusion features, performing network consistency evaluation through local to global hierarchical calculation, and calculating an adaptive threshold based on historical trigger frequency and knowledge update activity for early warning judgment include: Construct a multi-level fractal computing framework, use a bidirectional mapping matrix to characterize the knowledge association within the layer; integrate semantic features and topological features to construct a fractal recursive function, and quantify the complexity of hierarchical knowledge through iterative calculation; perform multi-level joint prediction driven by fractal features; evaluate global consistency based on the geometric mean of inter-level consistency covariance and fractal features; dynamically adjust the warning threshold according to the influence of knowledge dissemination, including: Construct a multi-level fractal computing framework, divide the knowledge into multiple levels according to the fusion features and the degree of expertise in the knowledge field and the degree of knowledge relevance, and construct a bidirectional mapping matrix at each level to characterize the association features between knowledge units in the layer; extract multi-dimensional semantic features from the node set of each level, and fuse the multi-dimensional semantic features with the topological features of the knowledge graph to generate a node relationship strength matrix; construct a fractal recursive function based on the relationship strength matrix, and obtain fractal features reflecting the complexity of hierarchical knowledge through iterative calculation; Perform fractal feature-driven hierarchical prediction on the nodes in each layer, perform feature splicing on the fused feature vector of the node, the fractal feature of the layer to which it belongs, and the prediction information of the adjacent layers, and calculate the prediction feature of the node through the parameter matrix and activation function; The local consistency index is calculated based on the similarity of the prediction features of nodes between adjacent levels, and the global consistency index is calculated by combining the weights of each level, the consistency covariance between levels, and the geometric mean of the fractal features; An adaptive threshold is constructed based on the historical trigger frequency and knowledge update activity, and the knowledge verification pass rate, knowledge dissemination coverage and knowledge conversion rate are introduced to calculate the knowledge dissemination influence. The adaptive threshold is dynamically adjusted based on the change rate of the knowledge dissemination influence and the global consistency index; when the global consistency index is lower than the dynamically adjusted adaptive threshold, an early warning is triggered.

[0012] In an optional embodiment, The steps to build a multi-level fractal computing framework include: Collecting node semantic vectors and multidimensional relationship strengths, performing nonlinear mapping of the node semantic vectors and the multidimensional relationship strengths through a bidirectional mapping matrix to obtain node initial features, constructing a fractal recursive function based on the node initial features, the fractal recursive function including the topological association and semantic similarity of the nodes; obtaining hierarchical fractal features through iterative recursive calculations, the recursive calculations including the propagation and diffusion of fractal features and the aggregation of features between hierarchies, and optimizing the global fractal dimension through a dynamic programming method. A multi-scale sliding time window is constructed for the hierarchical fractal feature, and an adaptive time decay factor is introduced for windows of different scales, and the time decay factor is dynamically adjusted according to the knowledge update frequency and the correlation strength; the dynamic fractal dimension of the node and the level is calculated in each time window, and the dynamic fractal dimension is integrated with the historical dimension information through an exponential weighting method; a time series fractal feature sequence is constructed based on the dynamic fractal dimension, and the time series fractal feature sequence contains the evolution trajectory and change pattern of the knowledge structure; trend decomposition and periodicity analysis are performed on the time series fractal feature sequence to extract the long-term trend and short-term fluctuation characteristics of knowledge evolution, and the adjacent moment change rate of the time series fractal feature is calculated, and a multi-scale stability evaluation function is constructed in combination with the variance of the time series fractal feature sequence, and the stability index of the knowledge structure is calculated based on the multi-scale stability evaluation function; An update threshold function is constructed according to the stability index. When the change rate of the time series fractal feature exceeds the update threshold function, the knowledge update mechanism is triggered. The strength of the knowledge update strategy is determined based on the change rate of the time series fractal feature and the stability index. The execution result of the knowledge update strategy is fed back to the fractal feature calculation link for the next round of dynamic update of fractal features.

[0013] This invention innovatively introduces fractal theory into the field of knowledge modeling and establishes a complete set of knowledge evolution perception and early warning methods. First, through the dual structure of generative mind map and dual network, dynamic representation and verification of knowledge are realized; second, a multi-level fractal computing framework is creatively constructed, and the complexity of knowledge is quantified through fractal recursive function, which realizes the accurate description of knowledge structure; third, a hierarchical consistency evaluation mechanism from local to global is innovatively designed, and a dynamic early warning system is established in combination with the influence of knowledge dissemination; finally, adaptive adjustment of knowledge update strategy is realized through multi-scale time window analysis and stability evaluation.

[0014] The present invention can more accurately capture the dynamic evolution characteristics of knowledge structure, more effectively identify and warn of knowledge conflicts, provide more precise knowledge repair strategies, and achieve adaptive updating and optimization of the knowledge base.

[0015] The present invention breaks through the limitations of traditional knowledge management methods, provides a new technical path for the dynamic maintenance and optimization of large-scale knowledge bases, and has important theoretical innovation value and practical application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 The present invention is a flowchart of an automatic reporting review and knowledge base integration optimization method based on a dynamic rule engine according to an embodiment of the present invention. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0019] Figure 1 FIG. 1 is a flow chart of an automatic reporting review and knowledge base integration optimization method based on a dynamic rule engine according to an embodiment of the present invention. Figure 1 As shown, the method includes: S1. Standardize the format of electronic form data to obtain pre-processed data, perform structured analysis on the pre-processed data according to preset reporting rules, and generate a data set to be reviewed; S2. Build a dynamic rule engine based on the rule time sequence graph, track the rules in time sequence, decompose the rules into atomic rule units, adaptively reorganize the atomic rule units according to the business scenario, generate a scenario-based rule chain, use the scenario-based rule chain to match the rules of the data set to be reviewed, and obtain the review result; S3. Carry out multi-level knowledge modeling on the review results to construct a vertically layered knowledge system; establish a knowledge dual network based on the knowledge system, calculate the time series distribution of the cross entropy between the original network and the dual network, perform curve fitting and extrapolation on the entropy value change trend, combine with dynamic prediction of network consistency, detect and locate potential knowledge conflicts, generate repair strategies, and update the repaired knowledge to the knowledge base.

[0020] For example, when formatting electronic form data, the original data is first cleaned, including removing special characters, unifying date formats, standardizing numerical units, etc. For example, dates in different formats are uniformly converted to the "YYYY-MM-DD" format, and values ​​with thousands separators are converted to pure numeric format. For missing values, fill or mark them according to business rules, such as filling "0" with null values ​​in financial data and marking "not filled" with null values ​​in descriptive fields.

[0021] In the structured analysis phase, the preset reporting rules are parsed and a rule index tree is established. The preset reporting rules include data format rules (such as mandatory item checks, field length rules, data type rules), logical relationship rules (such as single table verification, cross-table associations, year-on-year changes) and business rules (such as range constraints, threshold tests, industry characteristics). For example, in the data format rules, the unified social credit code must be 18 digits; in the logical relationship rules, the total assets are equal to the sum of current assets and non-current assets; in the business rules, the inventory turnover rate of manufacturing enterprises below 0.5 needs to be marked as abnormal.

[0022] In the implementation of the dynamic rule engine based on the rule time series graph, complex rules are first decomposed into atomic rule units. Taking the annual report filing of an enterprise as an example, the complex rule "abnormal year-on-year growth rate of enterprise operating income" can be decomposed into atomic rule units such as operating income validity check, historical data extraction, growth rate calculation, and threshold comparison.

[0023] Time-stamp the atomic rule units to record the effective time, expiration time and update history of the rules. This information is stored in a time-series graph. Each rule node contains a timestamp attribute, and the edges between nodes represent the logical relationship and time-series dependency between rules.

[0024] In different business scenarios, atomic rule units are dynamically combined to form scenario-based rule chains. Taking the quarterly report review of manufacturing enterprises as an example, the production and operation indicator rule chain includes: verifying capacity utilization, checking inventory turnover, checking the relationship between raw material purchases and output, and analyzing unit product cost changes. The financial indicator association rule chain includes: checking the matching degree of sales revenue and accounts receivable, verifying the relationship between raw material purchases and accounts payable, checking the correlation between output and operating costs, and analyzing the rationality of period expenses.

[0025] During the rule matching process, the system automatically performs rule verification and generates classification results. For example, in the production cost rationality check, by calculating the unit product cost and comparing it with the industry average, it is judged to be reasonable when the difference rate is less than 10%; in the income matching check, by comparing the theoretical income and the actual income, it is judged to be normal when the income realization rate is within a reasonable range.

[0026] When multi-level knowledge modeling is performed for review results, a vertically layered knowledge system is constructed according to business fields, data features, and rule types. For example, knowledge is divided into data format layer, business rule layer, cross-table association layer, etc., and knowledge nodes are connected within each layer through semantic associations and rule dependencies.

[0027] When building a knowledge dual network, the dual network is constructed by reversing the association direction in the original knowledge network. For example, in the original network, the edge from the "total assets" node to the "asset-liability ratio" node is transformed into the "asset-liability ratio" node pointing to the "total assets" node in the dual network.

[0028] When calculating the time series distribution of the cross entropy of two networks, collect network status data at different time points. For example, calculate the cross entropy value once a day to obtain a time series. By fitting the curve of this series, the evolution trend of the knowledge system can be predicted. When the prediction curve fluctuates abnormally, it indicates that there may be knowledge conflicts.

[0029] After a potential conflict is discovered, the source of the conflict is located based on the centrality index of the node. For example, if the centrality of a rule node suddenly increases and it has new associations with multiple nodes, this node is likely to be the source of the conflict. For the identified conflicts, specific repair suggestions are generated, such as adjusting rule parameters, updating rule dependencies, etc.

[0030] The dynamic rule engine of the present invention can flexibly adjust the review strategy according to the business scenario; the knowledge dual network provides a verification mechanism to improve the accuracy of the knowledge base; the hierarchical knowledge modeling and conflict detection mechanism ensure the consistency and timeliness of the knowledge base.

[0031] In an optional embodiment, The steps of building a dynamic rule engine based on the rule time sequence graph, tracking the rules in time sequence, decomposing the rules into atomic rule units, and adaptively reorganizing the atomic rule units according to the business scenarios to generate scenario-based rule chains include: Construct a rule time series graph, represent the rule node features based on the spatiotemporal feature tensor; atomically decompose the rules based on semantic integrity; use the scenario feature vector to drive the rule reorganization and generate a scenario-based rule chain, including: Construct a rule timing graph, use business rules as nodes of the rule timing graph, use the timing dependencies between rules as directed edges of the rule timing graph, construct a spatiotemporal feature tensor for the nodes, the spatiotemporal feature tensor includes the timing attributes, spatial constraints and business attributes of the rules, construct a rule timing matrix based on the spatiotemporal feature tensor, perform feature enhancement on the rule timing matrix, construct a multidimensional association matrix including rule timing features and business attributes, and perform timing tracking on the rules according to the time overlap and execution order relationship between the rules in the multidimensional association matrix; Based on the rule time series graph, the rules are atomically decomposed into atomic rule units that maintain semantic integrity, the semantic relevance between the atomic rule units is calculated, and when the semantic relevance is less than a preset boundary threshold, the separation boundary of the atomic rule unit is determined to generate an atomic rule unit set; A scene feature vector is constructed, and the fitness score between each atomic rule unit in the atomic rule unit set and the scene feature vector is calculated. According to the fitness score, a dynamic programming algorithm is used to adaptively reorganize the atomic rule units to generate a scene-based rule chain.

[0032] Exemplarily, first, a rule timing graph is constructed. Business rules are used as nodes of the rule timing graph, and the timing dependencies between rules are used as directed edges. The spatiotemporal feature tensor of each node includes the timing attributes, spatial constraints, and business attributes of the rules. By comprehensively analyzing these features, a rule timing matrix is ​​constructed. This matrix not only contains the timing features of the rules, but also combines the business attributes. Next, the rule timing matrix is ​​feature enhanced to generate a multidimensional association matrix. The multidimensional association matrix contains the temporal overlap and execution order relationship between the rules, thereby realizing the timing tracking of the rules.

[0033] Secondly, atomic decomposition is performed based on the rule time series graph. The rules are decomposed into atomic rule units that maintain semantic integrity. The separation boundary of the atomic rule units is determined by calculating the semantic relevance between the atomic rule units. When the semantic relevance is lower than the preset boundary threshold, the separation boundary of the atomic rule unit is identified, and finally a set of atomic rule units is generated.

[0034] Next, construct a scenario feature vector. The scenario feature vector extracts the key attributes of the business scenario (such as industry type, reporting period type, enterprise size, review level, etc.) and quantifies them into numerical features to form a multidimensional vector representation that reflects the characteristics of a specific business scenario. Calculate the fitness score of each atomic rule unit in the atomic rule unit set and the scenario feature vector. The fitness score reflects the applicability of the atomic rule unit in a specific scenario. According to the fitness score, the dynamic programming algorithm is used to adaptively reorganize the atomic rule units to generate a scenario-based rule chain. This process ensures that the generated rule chain can effectively adapt to different business scenarios.

[0035] The present invention can quickly respond to changes in business scenarios and ensure the effectiveness of rules by dynamically reorganizing atomic rule units; through time tracking, the dependencies and execution order between rules can be clearly understood, which is convenient for subsequent rule management and optimization; through atomic decomposition and scenario-based reorganization, the rule management process is simplified, making rule maintenance and updating more efficient.

[0036] In an optional embodiment, A spatiotemporal feature tensor is constructed for the node, wherein the spatiotemporal feature tensor includes a regular time series attribute, a spatial constraint, and a business attribute; a regular time series matrix is ​​constructed based on the spatiotemporal feature tensor; and feature enhancement is performed on the regular time series matrix. The steps of constructing a multidimensional association matrix including regular time series features and business attributes include: Constructing a spatiotemporal feature tensor of a rule node, the spatiotemporal feature tensor comprising a time series feature vector, a space feature vector and a business feature vector, wherein the time series feature vector comprises the effective time, expiration time, maximum execution time and minimum execution time of the rule, the space feature vector comprises the scope, geographical location restriction and management level of the rule, and the business feature vector comprises the priority, rule type and business field of the rule; Performing time series dimension projection on the spatiotemporal feature tensor to obtain a time series feature matrix, calculating the time overlap and execution order relationship between the rules to obtain the time series relationship strength, and constructing a rule time series matrix based on the time series feature matrix and the time series relationship strength; Calculating the similarity of the spatial feature vectors to obtain spatial association strength, and calculating the correlation of the service feature vectors to obtain service association strength; Performing feature enhancement on the regular time series matrix, performing multi-dimensional feature fusion on the regular time series matrix, the spatial association intensity and the business association intensity, and generating an enhanced association matrix; Construct a query matrix, a key-value matrix and a numerical matrix, calculate the attention weight based on the enhanced association matrix, perform weighted fusion on the attention weight and the enhanced association matrix, and generate a multidimensional association matrix containing regular high-order association relationships.

[0037] Exemplarily, first, construct the spatiotemporal feature tensor of the rule node. The tensor consists of three main feature vectors: time series feature vector, space feature vector and business feature vector. The time series feature vector includes the effective time, expiration time, maximum execution time and minimum execution time of the rule, which can reflect the time attribute of the rule. The space feature vector contains the scope, geographical location restriction and management level of the rule to ensure that the scope of application of the rule is clear. The business feature vector includes the priority of the rule, the type of rule and the business field, which helps to identify the importance of the rule in the business process.

[0038] For example, a rule takes effect on January 1, 2023, expires on December 31, 2023, has a maximum execution time of 48 hours, a minimum execution time of 1 hour, a scope of nationwide, a geographic location restriction of a specific city, a priority of high, a rule type of compliance check, and a business area of ​​finance.

[0039] The constructed spatiotemporal feature tensor is projected in the time dimension to obtain the time feature matrix. The time relationship strength is calculated by analyzing the time overlap and execution order relationship between the rules. The time relationship strength reflects the degree of mutual influence between the rules and provides a basis for the subsequent construction of the rule time matrix.

[0040] For example, if the effective time of rule A overlaps with the expiration time of rule B, it can be considered that there is a certain time relationship between the two rules, and then their temporal relationship strength can be calculated.

[0041] Based on the time series feature matrix and the time series relationship strength, a rule time series matrix is ​​constructed. This matrix presents the time series relationship between different rules in a matrix form, which is convenient for subsequent analysis and processing.

[0042] For example, if the temporal relationship strength between rule A and rule B is 0.8, and the temporal relationship strength between rule A and rule C is 0.5, then in the rule temporal matrix, the relationship value between A and B is 0.8, and the relationship value between A and C is 0.5.

[0043] Next, the similarity of the spatial feature vectors is calculated to obtain the spatial correlation strength. At the same time, the correlation of the business feature vectors is calculated to obtain the business correlation strength. These two strength values ​​will be used in the subsequent feature enhancement steps.

[0044] For example, if rule A and rule B have the same scope and similar priority, their spatial association strength is high; if rule A and rule C have the same business domain, their business association strength is also high.

[0045] Enhance the features of the rule time series matrix, fuse it with the spatial correlation strength and business correlation strength in multi-dimensional features, and generate an enhanced correlation matrix. This step aims to improve the correlation between rules and make subsequent analysis more accurate.

[0046] For example, if the time series matrix value of rule A is 0.8, the spatial correlation strength is 0.7, and the business correlation strength is 0.9, then in the enhanced correlation matrix, the final value of rule A will comprehensively consider the influence of these three factors.

[0047] Finally, the query matrix, key-value matrix and value matrix are constructed. The attention weight is calculated based on the enhanced association matrix, and the attention weight is weightedly fused with the enhanced association matrix to generate a multidimensional association matrix containing regular high-order association relationships. This matrix will provide an important basis for subsequent decision support.

[0048] For example, if the attention weight of a query in the query matrix is ​​0.6 and the value of the corresponding rule in the enhanced association matrix is ​​0.8, the value after weighted fusion is 0.48, which reflects the influence of the query on the rule.

[0049] By constructing spatiotemporal feature tensors and multidimensional association matrices, the present invention can comprehensively reflect the relationship between rules and improve the accuracy of analysis results; the enhanced association matrix provides richer information for decision-making, helping business personnel make more reasonable decisions; through comprehensive analysis of the temporal, spatial and business characteristics of the rules, it can better adapt to changes in the business environment and improve the flexibility of rule management.

[0050] In an optional embodiment, The steps of performing multi-level knowledge modeling on the review results and constructing a vertically hierarchical knowledge system include: Extract business entities, rule entities, result entities and time entities from the review results, build an entity relationship set including trigger relationships, generation relationships and timing relationships, and extract an attribute set including business type, rule identifier, result type, timestamp and confidence level; Based on the entity relationship set and the attribute set, a conceptual hierarchical relationship among business, rules, results and time is constructed, the semantic similarity between concepts is calculated, a concept clustering tree is constructed according to the semantic similarity, and semantic association rules between concepts are extracted; The knowledge system is divided into a data layer, a rule layer, a reasoning layer, and an application layer. The data layer processes time series data streams. The rule layer generates context-aware rules based on the time series data streams. The reasoning layer performs state-dependent reasoning according to the context-aware rules. The application layer builds scenario-based applications based on the state-dependent reasoning. The semantic association rules are used to construct inter-layer context information, and bottom-up knowledge extraction is performed to obtain inter-layer knowledge mapping relationships. Global constraints are determined based on the inter-layer knowledge mapping relationships, and top-down knowledge guidance is performed. A knowledge evolution equation is constructed by combining external environment information and internal consistency requirements, a knowledge update gradient is calculated according to the knowledge evolution equation, and an adaptive knowledge update strategy is determined by combining the knowledge update gradient and the knowledge drift amount; According to the inter-layer knowledge mapping relationship, the difference between the upper and lower layer knowledge representations and the degree of violation of the structural constraints are calculated to obtain the consistency deviation, the knowledge coverage of the target business scenario and the missing of the knowledge link are counted to obtain the integrity index, and the decision accuracy of the knowledge application and the degree of adaptability to the scenario are analyzed to obtain the effectiveness index; a weighted objective function of the consistency deviation, integrity index and effectiveness index is constructed, and the knowledge structure integrity and semantic consistency constraints are set based on the weighted objective function. The indicators that violate the constraints are optimized by combining the gradient descent and the coordinate alternating optimization method, and the weights of each indicator are dynamically adjusted during the optimization process, and the optimization step size is adaptively adjusted according to the application feedback; Calculate the accuracy, coverage and timeliness of knowledge to obtain knowledge quality assessment results, and dynamically adjust the knowledge structure based on the assessment results.

[0051] Exemplarily, business entities, rule entities, result entities and time entities are extracted from the review results. Business entities can be customers, products, etc., rule entities can be compliance standards, operating procedures, etc., result entities are the specific manifestations of the review results, and time entities are time information related to the review. Next, a set of entity relationships is constructed, including trigger relationships (such as the operation of a business entity triggers the application of a rule entity), generation relationships (such as a rule entity generates a result entity) and timing relationships (such as the association between a time entity and a result entity). At the same time, a set of attributes is extracted, including information such as business type, rule identifier, result type, timestamp and confidence.

[0052] Based on the aforementioned entity relationship set and attribute set, the conceptual hierarchical relationship of business, rules, results and time is constructed. By analyzing the relationship between these entities, their semantic similarity is calculated, and then a concept clustering tree is constructed. This clustering tree classifies similar concepts into one category and extracts the semantic association rules between concepts for subsequent knowledge reasoning and application.

[0053] The knowledge system is divided into the data layer, rule layer, reasoning layer, and application layer. The data layer is responsible for processing time-series data streams to ensure the real-time and accuracy of data. The rule layer generates context-aware rules based on time-series data streams to ensure the applicability and effectiveness of the rules. The reasoning layer performs state-dependent reasoning based on context-aware rules to derive new knowledge and conclusions. The application layer builds scenario-based applications based on state-dependent reasoning to ensure the actual application effect of knowledge.

[0054] The semantic association rules extracted above are used to construct inter-layer context information. Through bottom-up knowledge extraction, the inter-layer knowledge mapping relationship is obtained. Based on these mapping relationships, global constraints are determined, and top-down knowledge guidance is performed to ensure the overall consistency and effectiveness of the knowledge system.

[0055] Combining external environmental information and internal consistency requirements, a knowledge evolution equation is constructed. The gradient of knowledge update is calculated through this equation, and the adaptive knowledge update strategy is determined by combining the knowledge update gradient and knowledge drift to cope with the ever-changing environment and needs.

[0056] According to the mapping relationship between the layers of knowledge, the difference between the upper and lower layers of knowledge representation and the degree of violation of the structural constraints are calculated to obtain the consistency deviation. At the same time, the knowledge coverage of the target business scenario and the missing of the knowledge link are counted to obtain the integrity index. The decision accuracy and scenario adaptability of the knowledge application are analyzed to obtain the effectiveness index.

[0057] Construct a weighted objective function of consistency deviation, integrity index, and effectiveness index. Set the integrity and semantic consistency constraints of the knowledge structure based on this objective function. Optimize the indicators that violate the constraints by combining gradient descent with coordinate alternation optimization method. During the optimization process, dynamically adjust the weight of each indicator, and adaptively adjust the optimization step size according to application feedback to ensure continuous optimization of the knowledge structure.

[0058] Calculate the accuracy, coverage and timeliness of knowledge to obtain the knowledge quality assessment results. According to the assessment results, dynamically adjust the knowledge structure to ensure the high quality and applicability of knowledge.

[0059] The present invention improves the accuracy and consistency of the knowledge system, ensuring that information between different levels can be effectively transmitted and applied; through adaptive update strategies, it enhances the knowledge system's ability to respond to changes in the external environment and improves the timeliness and applicability of knowledge; by optimizing the knowledge structure, it improves the coverage and effectiveness of knowledge, ensuring that accurate decision support can be provided in practical applications.

[0060] In an optional embodiment, The steps of establishing a knowledge dual network based on the knowledge system, calculating the time series distribution of the cross entropy between the original network and the dual network, performing curve fitting and extrapolation on the entropy value change trend, combining dynamic prediction of network consistency, detecting and locating potential knowledge conflicts, generating a repair strategy, and updating the repaired knowledge to the knowledge base include: Construct a knowledge dual network; calculate dynamic cross entropy based on the time decay factor; use the attention mechanism to fuse multi-dimensional prediction features, and evaluate network consistency through local to global hierarchical calculations; dynamically adjust the warning threshold based on historical trigger frequency and activity; locate the conflict source and generate a repair strategy based on the betweenness centrality, including: The association relationship between knowledge nodes is mapped into a knowledge graph structure. The association strength between nodes is calculated based on the feature vector representation of the nodes and the node association set. The dual network is constructed by reversing the association direction between nodes in the original network and recalculating the weights. Obtain the probability distribution of nodes in the original network and the dual network to calculate the network cross entropy, introduce a time decay factor to dynamically weight the network cross entropy, and obtain the cross entropy sequence of the original network and the dual network within a preset time window; construct a polynomial kernel function to map the cross entropy sequence to a high-dimensional feature space, extract a time-varying feature vector, and the time-varying feature vector includes a feature mapping component calculated by the kernel function; perform curve fitting on the time series of the cross entropy based on the time-varying feature vector, and perform extrapolation prediction on the fitting curve; Based on the historical data of the cross entropy sequence, network topology features and node semantic similarity, a multidimensional prediction feature is constructed, and the multidimensional prediction feature is dynamically fused by using an attention mechanism to obtain a fusion feature. According to the fusion feature, hierarchical prediction is performed according to the knowledge field, and network consistency is evaluated through local to global hierarchical calculation. Based on the historical trigger frequency and knowledge update activity, an adaptive threshold is calculated for early warning judgment, and conflict detection of the corresponding level is triggered according to the early warning result; Calculate the betweenness centrality of nodes in the original network and the dual network, locate the conflict source node based on the change in the betweenness centrality, generate a repair strategy based on the graph edit distance for the conflict source node; and update the repair strategy to the knowledge base.

[0061] Exemplarily, based on the knowledge system, the process of establishing a knowledge dual network includes multiple steps, aiming to detect and locate potential knowledge conflicts and generate corresponding repair strategies through the calculation of dynamic cross entropy and the evaluation of network consistency. The following are the detailed steps for implementing the technical solution.

[0062] First, a knowledge dual network is constructed. This step involves mapping the associations between knowledge nodes into a knowledge graph structure. The strength of the associations between nodes is calculated by analyzing the feature vector representation of the nodes and the association sets between the nodes. Next, the direction of the associations between the nodes in the original network is reversed and the weights are recalculated to construct a dual network.

[0063] Secondly, the probability distribution of nodes in the original network and the dual network is calculated to calculate the cross entropy of the network. The time decay factor is introduced to dynamically weight the cross entropy and obtain the cross entropy sequence of the original network and the dual network within the preset time window. This sequence provides basic data for subsequent analysis.

[0064] Next, a polynomial kernel function is constructed to map the cross entropy sequence to a high-dimensional feature space. By extracting the time-varying feature vector, a feature vector containing the feature mapping components calculated by the kernel function is obtained. Based on these time-varying feature vectors, a curve fitting is performed on the time series of the cross entropy, and the fitted curve is extrapolated and predicted to identify potential trend changes.

[0065] On this basis, multi-dimensional prediction features are constructed based on the historical data of the cross entropy sequence, network topology features and node semantic similarity. The attention mechanism is used to dynamically fuse these multi-dimensional prediction features to obtain fused features. Based on the fused features, hierarchical predictions are performed according to the knowledge domain, and network consistency evaluation is performed through hierarchical calculations from local to global.

[0066] Subsequently, based on the historical trigger frequency and the activity of knowledge update, the adaptive threshold is calculated to make early warning judgments. According to the early warning results, the corresponding level of conflict detection is triggered to timely discover potential knowledge conflicts.

[0067] Finally, the betweenness centrality of nodes in the original network and the dual network is calculated. Based on the change in betweenness centrality, the conflict source nodes are located, and a repair strategy based on graph edit distance is generated for these conflict source nodes. The repair strategy is updated to the knowledge base to ensure the accuracy and consistency of the knowledge base.

[0068] Through the calculation of dynamic cross entropy and time series analysis, the present invention can timely identify the potential risks of knowledge conflicts, thereby improving the efficiency and accuracy of knowledge management; the use of attention mechanism to fuse multi-dimensional prediction features can more comprehensively evaluate network consistency and improve the stability and reliability of the knowledge network; the generated repair strategy is based on data-driven analysis, which can effectively solve the problem of knowledge conflicts and ensure the continuous updating and optimization of the knowledge base.

[0069] In an optional embodiment, The steps of performing hierarchical prediction according to the knowledge domain based on the fusion features, performing network consistency evaluation through local to global hierarchical calculation, and calculating an adaptive threshold based on historical trigger frequency and knowledge update activity for early warning judgment include: Construct a multi-level fractal computing framework, use a bidirectional mapping matrix to characterize the knowledge association within the layer; integrate semantic features and topological features to construct a fractal recursive function, and quantify the complexity of hierarchical knowledge through iterative calculation; perform multi-level joint prediction driven by fractal features; evaluate global consistency based on the geometric mean of inter-level consistency covariance and fractal features; dynamically adjust the warning threshold according to the influence of knowledge dissemination, including: Construct a multi-level fractal computing framework, divide the knowledge into multiple levels according to the fusion features and the degree of expertise in the knowledge field and the degree of knowledge relevance, and construct a bidirectional mapping matrix at each level to characterize the association features between knowledge units in the layer; extract multi-dimensional semantic features from the node set of each level, and fuse the multi-dimensional semantic features with the topological features of the knowledge graph to generate a node relationship strength matrix; construct a fractal recursive function based on the relationship strength matrix, and obtain fractal features reflecting the complexity of hierarchical knowledge through iterative calculation; Perform fractal feature-driven hierarchical prediction on the nodes in each layer, perform feature splicing on the fused feature vector of the node, the fractal feature of the layer to which it belongs, and the prediction information of the adjacent layers, and calculate the prediction feature of the node through the parameter matrix and activation function; The local consistency index is calculated based on the similarity of the prediction features of nodes between adjacent levels, and the global consistency index is calculated by combining the weights of each level, the consistency covariance between levels, and the geometric mean of the fractal features; An adaptive threshold is constructed based on the historical trigger frequency and knowledge update activity, and the knowledge verification pass rate, knowledge dissemination coverage and knowledge conversion rate are introduced to calculate the knowledge dissemination influence. The adaptive threshold is dynamically adjusted based on the change rate of the knowledge dissemination influence and the global consistency index; when the global consistency index is lower than the dynamically adjusted adaptive threshold, an early warning is triggered.

[0070] Exemplarily, first, a multi-level fractal computing framework is constructed. The framework divides the knowledge field into multiple levels according to the degree of expertise and knowledge relevance. In each level, a bidirectional mapping matrix is ​​constructed to characterize the correlation characteristics between knowledge units in the layer. The bidirectional mapping matrix can effectively reflect the mutual influence and relationship between knowledge units, providing a basis for subsequent feature fusion and prediction.

[0071] Next, multidimensional semantic features are extracted from the node set at each level. These features can be extracted from relevant literature, databases or knowledge graphs through natural language processing technology. The extracted multidimensional semantic features are fused with the topological features of the knowledge graph to generate a node relationship strength matrix. The relationship strength matrix not only reflects the direct relationship between nodes, but also reflects the potential path of knowledge dissemination.

[0072] Then, a fractal recursive function is constructed based on the relationship strength matrix. Through iterative calculation, the fractal recursive function can quantify the complexity of hierarchical knowledge and reflect the propagation and evolution of knowledge between different levels. This process will provide the necessary feature support for subsequent hierarchical prediction.

[0073] In each layer, fractal feature driven hierarchical prediction is performed. The fused feature vector of the node, the fractal features of the layer to which it belongs, and the prediction information of the adjacent layers are feature spliced. The prediction features of the node are obtained by calculating the parameter matrix and activation function. This process ensures information sharing and synergy between different layers, thereby improving the accuracy of prediction.

[0074] Next, the local consistency index is calculated based on the similarity of the prediction features of nodes between adjacent levels. The global consistency index is calculated by combining the weights of each level, the consistency covariance between levels, and the geometric mean of the fractal features. The global consistency index can reflect the stability and consistency of the entire knowledge system and provide a basis for the subsequent early warning mechanism.

[0075] Finally, an adaptive threshold is constructed based on the historical trigger frequency and knowledge update activity. The knowledge verification pass rate, knowledge dissemination coverage, and knowledge conversion rate are introduced to calculate the knowledge dissemination impact. The adaptive threshold is dynamically adjusted based on the knowledge dissemination impact and the change rate of the global consistency index. When the global consistency index is lower than the dynamically adjusted adaptive threshold, the early warning mechanism is triggered to respond to potential knowledge dissemination risks in a timely manner.

[0076] The present invention adopts a bidirectional mapping matrix to characterize the knowledge association within the layer, and combines semantic features with topological features to construct a fractal recursive function, so that the quantification of knowledge complexity is more accurate; the multi-level joint prediction mechanism driven by fractal features improves the accuracy of prediction through feature splicing and parameter matrix calculation; the global consistency is evaluated based on the geometric mean of inter-level consistency covariance and fractal features, making the consistency evaluation more comprehensive; multi-dimensional indicators such as knowledge verification pass rate, knowledge dissemination coverage and knowledge conversion rate are introduced to calculate the knowledge dissemination influence, so as to realize the precise dynamic adjustment of the warning threshold.

[0077] In an optional embodiment, The steps to build a multi-level fractal computing framework include: Collecting node semantic vectors and multidimensional relationship strengths, performing nonlinear mapping of the node semantic vectors and the multidimensional relationship strengths through a bidirectional mapping matrix to obtain node initial features, constructing a fractal recursive function based on the node initial features, the fractal recursive function including the topological association and semantic similarity of the nodes; obtaining hierarchical fractal features through iterative recursive calculations, the recursive calculations including the propagation and diffusion of fractal features and the aggregation of features between hierarchies, and optimizing the global fractal dimension through a dynamic programming method. A multi-scale sliding time window is constructed for the hierarchical fractal feature, and an adaptive time decay factor is introduced for windows of different scales, and the time decay factor is dynamically adjusted according to the knowledge update frequency and the correlation strength; the dynamic fractal dimension of the node and the level is calculated in each time window, and the dynamic fractal dimension is integrated with the historical dimension information through an exponential weighting method; a time series fractal feature sequence is constructed based on the dynamic fractal dimension, and the time series fractal feature sequence contains the evolution trajectory and change pattern of the knowledge structure; trend decomposition and periodicity analysis are performed on the time series fractal feature sequence to extract the long-term trend and short-term fluctuation characteristics of knowledge evolution, and the adjacent moment change rate of the time series fractal feature is calculated, and a multi-scale stability evaluation function is constructed in combination with the variance of the time series fractal feature sequence, and the stability index of the knowledge structure is calculated based on the multi-scale stability evaluation function; An update threshold function is constructed according to the stability index. When the change rate of the time series fractal feature exceeds the update threshold function, the knowledge update mechanism is triggered. The strength of the knowledge update strategy is determined based on the change rate of the time series fractal feature and the stability index. The execution result of the knowledge update strategy is fed back to the fractal feature calculation link for the next round of dynamic update of fractal features.

[0078] For example, first, collecting node semantic vectors and multidimensional relationship strengths is the basis for building a fractal computing framework. The node semantic vector is a high-dimensional representation of node features, while the multidimensional relationship strength reflects the degree of association between nodes. Through a bidirectional mapping matrix, the node semantic vector and the multidimensional relationship strength are nonlinearly mapped to obtain the initial features of the node. This process can be achieved through machine learning algorithms, such as using neural networks for feature extraction and mapping.

[0079] Next, based on the initial features of the nodes, a fractal recursive function is constructed. This function not only considers the topological association of the nodes, but also the semantic similarity. Through recursive calculation, hierarchical fractal features can be obtained. At this time, the propagation and diffusion of fractal features and the aggregation of features between levels are key links. Propagation and diffusion refers to the transmission of features between nodes, while the aggregation of features between levels is the integration of features at different levels. The dynamic programming method is applied here to optimize the global fractal dimension and ensure the efficiency and accuracy of the calculation.

[0080] After obtaining the hierarchical fractal features, a multi-scale sliding time window is constructed. An adaptive time decay factor is introduced in each time window, which is dynamically adjusted according to the knowledge update frequency and the association strength. In this way, the dynamic changes of nodes and levels can be better captured. In each time window, the dynamic fractal dimension of nodes and levels is calculated. The calculation of the dynamic fractal dimension adopts an exponential weighting method and integrates historical dimension information to ensure sensitivity to changes.

[0081] Based on the dynamic fractal dimension, a time series fractal feature sequence is constructed. This sequence not only contains the evolution trajectory of the knowledge structure, but also reflects the change pattern. Trend decomposition and periodicity analysis of the time series fractal feature sequence can extract the long-term trend and short-term fluctuation characteristics of knowledge evolution. The rate of change at adjacent moments is calculated, and combined with the variance of the time series fractal feature sequence, a multi-scale stability evaluation function is constructed. This function is used to evaluate the stability index of the knowledge structure.

[0082] According to the stability index, the update threshold function is constructed. When the change rate of the time series fractal feature exceeds the update threshold, the knowledge update mechanism is triggered. At this time, the strength of the knowledge update strategy is determined based on the change rate and stability index. Finally, the execution result of the knowledge update strategy is fed back to the fractal feature calculation link for the next round of dynamic update.

[0083] The present invention adopts a bidirectional mapping matrix combined with a fractal recursive function, and realizes multi-level propagation and aggregation of fractal features through iterative recursive calculation, making feature extraction more accurate; through a multi-scale sliding time window and an adaptive time decay factor, fine-grained tracking of knowledge evolution is achieved; using dynamic fractal dimensions and time-series fractal feature sequences, the evolution trajectory of the knowledge structure is accurately captured; based on the dynamic adjustment mechanism of stability indicators and update threshold functions, the adaptive optimization of the knowledge update strategy is achieved, thereby improving the real-time response capability of the system.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. The automatic filling review and knowledge base integration optimization method based on dynamic rule engine is characterized by: include: Standardize the format of electronic form data to obtain pre-processed data, perform structured analysis on the pre-processed data according to preset reporting rules, and generate a data set to be reviewed; Build a dynamic rule engine based on the rule time sequence graph, track the rules in time sequence, decompose the rules into atomic rule units, adaptively reorganize the atomic rule units according to the business scenario, generate a scenario-based rule chain, use the scenario-based rule chain to match the rules of the data set to be reviewed, and obtain the review result; Perform multi-level knowledge modeling on the review results and construct a vertically layered knowledge system; establish a knowledge dual network based on the knowledge system, calculate the time series distribution of the cross entropy between the original network and the dual network, perform curve fitting and extrapolation on the entropy value change trend, combine with dynamic prediction of network consistency, detect and locate potential knowledge conflicts, generate repair strategies, and update the repaired knowledge to the knowledge base.

2. The method according to claim 1, characterized in that The steps of building a dynamic rule engine based on the rule time sequence graph, tracking the rules in time sequence, decomposing the rules into atomic rule units, and adaptively reorganizing the atomic rule units according to the business scenarios to generate scenario-based rule chains include: Construct a rule time series graph, represent the rule node features based on the spatiotemporal feature tensor; atomically decompose the rules based on semantic integrity; use the scenario feature vector to drive the rule reorganization and generate a scenario-based rule chain, including: Construct a rule timing graph, use business rules as nodes of the rule timing graph, use the timing dependencies between rules as directed edges of the rule timing graph, construct a spatiotemporal feature tensor for the nodes, the spatiotemporal feature tensor includes the timing attributes, spatial constraints and business attributes of the rules, construct a rule timing matrix based on the spatiotemporal feature tensor, perform feature enhancement on the rule timing matrix, construct a multidimensional association matrix including rule timing features and business attributes, and perform timing tracking on the rules according to the time overlap and execution order relationship between the rules in the multidimensional association matrix; Based on the rule time series graph, the rules are atomically decomposed into atomic rule units that maintain semantic integrity, the semantic relevance between the atomic rule units is calculated, and when the semantic relevance is less than a preset boundary threshold, the separation boundary of the atomic rule unit is determined to generate an atomic rule unit set; A scene feature vector is constructed, and the fitness score between each atomic rule unit in the atomic rule unit set and the scene feature vector is calculated. According to the fitness score, a dynamic programming algorithm is used to adaptively reorganize the atomic rule units to generate a scene-based rule chain.

3. The method according to claim 2, characterized in that A spatiotemporal feature tensor is constructed for the node, wherein the spatiotemporal feature tensor includes a regular time series attribute, a spatial constraint, and a business attribute; a regular time series matrix is ​​constructed based on the spatiotemporal feature tensor; and feature enhancement is performed on the regular time series matrix. The steps of constructing a multidimensional association matrix including regular time series features and business attributes include: Constructing a spatiotemporal feature tensor of a rule node, the spatiotemporal feature tensor comprising a time series feature vector, a space feature vector and a business feature vector, wherein the time series feature vector comprises the effective time, expiration time, maximum execution time and minimum execution time of the rule, the space feature vector comprises the scope, geographical location restriction and management level of the rule, and the business feature vector comprises the priority, rule type and business field of the rule; Performing time series dimension projection on the spatiotemporal feature tensor to obtain a time series feature matrix, calculating the time overlap and execution order relationship between the rules to obtain the time series relationship strength, and constructing a rule time series matrix based on the time series feature matrix and the time series relationship strength; Calculating the similarity of the spatial feature vectors to obtain spatial association strength, and calculating the correlation of the service feature vectors to obtain service association strength; Performing feature enhancement on the regular time series matrix, performing multi-dimensional feature fusion on the regular time series matrix, the spatial association intensity and the business association intensity, and generating an enhanced association matrix; Construct a query matrix, a key-value matrix and a numerical matrix, calculate the attention weight based on the enhanced association matrix, perform weighted fusion on the attention weight and the enhanced association matrix, and generate a multidimensional association matrix containing regular high-order association relationships.

4. The method according to claim 1, characterized in that: The steps of performing multi-level knowledge modeling on the review results and constructing a vertically hierarchical knowledge system include: Extract business entities, rule entities, result entities and time entities from the review results, build an entity relationship set including trigger relationships, generation relationships and timing relationships, and extract an attribute set including business type, rule identifier, result type, timestamp and confidence level; Based on the entity relationship set and the attribute set, a conceptual hierarchical relationship among business, rules, results and time is constructed, the semantic similarity between concepts is calculated, a concept clustering tree is constructed according to the semantic similarity, and semantic association rules between concepts are extracted; The knowledge system is divided into a data layer, a rule layer, a reasoning layer, and an application layer. The data layer processes time series data streams. The rule layer generates context-aware rules based on the time series data streams. The reasoning layer performs state-dependent reasoning according to the context-aware rules. The application layer builds scenario-based applications based on the state-dependent reasoning. The semantic association rules are used to construct inter-layer context information, and bottom-up knowledge extraction is performed to obtain inter-layer knowledge mapping relationships. Global constraints are determined based on the inter-layer knowledge mapping relationships, and top-down knowledge guidance is performed. A knowledge evolution equation is constructed by combining external environment information and internal consistency requirements, a knowledge update gradient is calculated according to the knowledge evolution equation, and an adaptive knowledge update strategy is determined by combining the knowledge update gradient and the knowledge drift amount; According to the inter-layer knowledge mapping relationship, the difference between the upper and lower layer knowledge representations and the degree of violation of the structural constraints are calculated to obtain the consistency deviation, the knowledge coverage of the target business scenario and the missing of the knowledge link are counted to obtain the integrity index, and the decision accuracy of the knowledge application and the degree of adaptability to the scenario are analyzed to obtain the effectiveness index; a weighted objective function of the consistency deviation, integrity index and effectiveness index is constructed, and the knowledge structure integrity and semantic consistency constraints are set based on the weighted objective function. The indicators that violate the constraints are optimized by combining the gradient descent and the coordinate alternating optimization method, and the weights of each indicator are dynamically adjusted during the optimization process, and the optimization step size is adaptively adjusted according to the application feedback; Calculate the accuracy, coverage and timeliness of knowledge to obtain knowledge quality assessment results, and dynamically adjust the knowledge structure based on the assessment results.

5. The method according to claim 1, characterized in that The steps of establishing a knowledge dual network based on the knowledge system, calculating the time series distribution of the cross entropy between the original network and the dual network, performing curve fitting and extrapolation on the entropy value change trend, combining dynamic prediction of network consistency, detecting and locating potential knowledge conflicts, generating a repair strategy, and updating the repaired knowledge to the knowledge base include: Construct a knowledge dual network; calculate dynamic cross entropy based on the time decay factor; use the attention mechanism to fuse multi-dimensional prediction features, and evaluate network consistency through local to global hierarchical calculations; dynamically adjust the warning threshold based on historical trigger frequency and activity; locate the conflict source and generate a repair strategy based on the betweenness centrality, including: The association relationship between knowledge nodes is mapped into a knowledge graph structure. The association strength between nodes is calculated based on the feature vector representation of the nodes and the node association set. The dual network is constructed by reversing the association direction between nodes in the original network and recalculating the weights. Obtain the probability distribution of nodes in the original network and the dual network to calculate the network cross entropy, introduce a time decay factor to dynamically weight the network cross entropy, and obtain the cross entropy sequence of the original network and the dual network within a preset time window; construct a polynomial kernel function to map the cross entropy sequence to a high-dimensional feature space, extract a time-varying feature vector, and the time-varying feature vector includes a feature mapping component calculated by the kernel function; perform curve fitting on the time series of the cross entropy based on the time-varying feature vector, and perform extrapolation prediction on the fitting curve; Based on the historical data of the cross entropy sequence, network topology features and node semantic similarity, a multidimensional prediction feature is constructed, and the multidimensional prediction feature is dynamically fused by using an attention mechanism to obtain a fusion feature. According to the fusion feature, hierarchical prediction is performed according to the knowledge field, and network consistency is evaluated through local to global hierarchical calculation. Based on the historical trigger frequency and knowledge update activity, an adaptive threshold is calculated for early warning judgment, and conflict detection of the corresponding level is triggered according to the early warning result; Calculate the betweenness centrality of nodes in the original network and the dual network, locate the conflict source node based on the change in the betweenness centrality, generate a repair strategy based on the graph edit distance for the conflict source node; and update the repair strategy to the knowledge base.

6. The method according to claim 5, characterized in that The steps of performing hierarchical prediction according to the knowledge domain based on the fusion features, performing network consistency evaluation through local to global hierarchical calculation, and calculating an adaptive threshold based on historical trigger frequency and knowledge update activity for early warning judgment include: Construct a multi-level fractal computing framework, use a bidirectional mapping matrix to characterize the knowledge association within the layer; integrate semantic features and topological features to construct a fractal recursive function, and quantify the complexity of hierarchical knowledge through iterative calculation; perform multi-level joint prediction driven by fractal features; evaluate global consistency based on the geometric mean of inter-level consistency covariance and fractal features; dynamically adjust the warning threshold according to the influence of knowledge dissemination, including: Construct a multi-level fractal computing framework, divide the knowledge into multiple levels according to the fusion features and the degree of expertise in the knowledge field and the degree of knowledge relevance, and construct a bidirectional mapping matrix at each level to characterize the association features between knowledge units in the layer; extract multi-dimensional semantic features from the node set of each level, and fuse the multi-dimensional semantic features with the topological features of the knowledge graph to generate a node relationship strength matrix; construct a fractal recursive function based on the relationship strength matrix, and obtain fractal features reflecting the complexity of hierarchical knowledge through iterative calculation; Perform fractal feature-driven hierarchical prediction on the nodes in each layer, perform feature splicing on the fused feature vector of the node, the fractal feature of the layer to which it belongs, and the prediction information of the adjacent layers, and calculate the prediction feature of the node through the parameter matrix and activation function; The local consistency index is calculated based on the similarity of the prediction features of nodes between adjacent levels, and the global consistency index is calculated by combining the weights of each level, the consistency covariance between levels, and the geometric mean of the fractal features; An adaptive threshold is constructed based on the historical trigger frequency and knowledge update activity, and the knowledge verification pass rate, knowledge dissemination coverage and knowledge conversion rate are introduced to calculate the knowledge dissemination influence. The adaptive threshold is dynamically adjusted based on the change rate of the knowledge dissemination influence and the global consistency index; when the global consistency index is lower than the dynamically adjusted adaptive threshold, an early warning is triggered.

7. The method according to claim 6, characterized in that The steps to build a multi-level fractal computing framework include: Collecting node semantic vectors and multidimensional relationship strengths, performing nonlinear mapping of the node semantic vectors and the multidimensional relationship strengths through a bidirectional mapping matrix to obtain node initial features, constructing a fractal recursive function based on the node initial features, the fractal recursive function including the topological association and semantic similarity of the nodes; obtaining hierarchical fractal features through iterative recursive calculations, the recursive calculations including the propagation and diffusion of fractal features and the aggregation of features between hierarchies, and optimizing the global fractal dimension through a dynamic programming method. A multi-scale sliding time window is constructed for the hierarchical fractal feature, and an adaptive time decay factor is introduced for windows of different scales, and the time decay factor is dynamically adjusted according to the knowledge update frequency and the correlation strength; the dynamic fractal dimension of the node and the level is calculated in each time window, and the dynamic fractal dimension is integrated with the historical dimension information through an exponential weighting method; a time series fractal feature sequence is constructed based on the dynamic fractal dimension, and the time series fractal feature sequence contains the evolution trajectory and change pattern of the knowledge structure; trend decomposition and periodicity analysis are performed on the time series fractal feature sequence to extract the long-term trend and short-term fluctuation characteristics of knowledge evolution, and the adjacent moment change rate of the time series fractal feature is calculated, and a multi-scale stability evaluation function is constructed in combination with the variance of the time series fractal feature sequence, and the stability index of the knowledge structure is calculated based on the multi-scale stability evaluation function; An update threshold function is constructed according to the stability index. When the change rate of the time series fractal feature exceeds the update threshold function, the knowledge update mechanism is triggered. The strength of the knowledge update strategy is determined based on the change rate of the time series fractal feature and the stability index. The execution result of the knowledge update strategy is fed back to the fractal feature calculation link for the next round of dynamic update of fractal features.

Images