Business process push-down voucher visual editing system

Abstract

The invention provides a business process push-down voucher visual editing system, and relates to the technical field of communication security, and the system comprises a business document push-down generation module which is used for constructing a business voucher triple structure based on a business process push-down voucher and original business data; the knowledge graph construction module is used for constructing an initial business voucher knowledge graph based on the business voucher triple structure, the business rule entity and the corresponding entity association relationship; the visual analysis module is used for carrying out topology display on the initial business voucher knowledge graph to obtain interactive behavior data; the anomaly prediction module is used for analyzing the interactive behavior data by using an anomaly prediction model to obtain an early warning result; and the updating module is used for optimizing and updating the business voucher knowledge graph based on the early warning result to obtain an updated business voucher knowledge graph and complete visual editing of the business process push-down voucher. According to the invention, the problem that the push-down voucher of the business process is difficult to edit visually is solved.

Description

TECHNICAL FIELD

[0001] The present specification relates to the technical field of communication security, in particular to a business process push certificate visual editing system. BACKGROUND

[0002] In an information system, the association of business documents and certificates is a core requirement for improving enterprise process efficiency and compliance. However, the existing technology has the following problems in the process of pushing the certificate: 1. Low efficiency of manual operation, the traditional certificate generation and processing method mainly depends on manual operation, and the business personnel need to manually create, edit and maintain a large amount of certificate information, which not only consumes time and effort, but also is prone to human error. Especially in complex business scenarios, a single business process may involve the generation and association of multiple certificates, and manual processing is difficult to ensure accuracy and consistency. 2. Insufficient data association, the existing system lacks effective automatic association mechanism between business documents and certificates, and often needs business personnel to manually establish association relationship. This method is prone to data island phenomenon, affecting the integrity and traceability of data, and it is difficult to form a complete business process chain. 3. Limited visualization, most existing systems lack intuitive visual editing interface, and users cannot clearly see the overall state of the business process and the flow of the certificate. This not only affects the user experience, but also increases the difficulty of business analysis and decision-making. 4. Insufficient real-time and accuracy, the traditional system often has problems of data update delay and synchronization not in time when processing the business process push, which affects the real-time and accuracy of the certificate information, and affects the timeliness of business decision-making. 5. Poor scalability and flexibility, existing technology is usually designed based on fixed business mode, when the enterprise business process changes or needs to adapt to new business demands, the system has limited expansion and adjustment capability, and it is difficult to quickly respond to business changes. 6. Difficulty in compliance supervision, in the context of increasingly stringent financial compliance and audit requirements, existing systems lack effective compliance checking and supervision mechanisms, making it difficult to automatically identify and prevent potential compliance risks, increasing the compliance cost and risk of enterprises. 7. Insufficient integration capability, existing systems often run independently, and have limited integration capability with other enterprise information systems, making it difficult to achieve unified management and circulation of data, and affecting the overall improvement of informationization level. SUMMARY

[0003] In view of the above problems in the prior art, the business process push certificate visual editing system provided by the present application solves the problem of difficult visual editing of business process push certificate.

[0004] In order to achieve the above-mentioned purpose of the application, the technical scheme adopted by the present application is as follows: a business process push certificate visual editing system, comprising: The business document push generation module is configured to push the voucher and the original business data based on the business process, utilize multi-element data integration and abnormality detection, and construct a business voucher triple structure. The knowledge graph construction module is configured to construct an initial business voucher knowledge graph based on the business voucher triple structure, business rule entities, and corresponding entity association relationships. The visual analysis module is configured to perform topological display on the initial business voucher knowledge graph, extract structure and attribute information, and obtain interaction behavior data. The abnormality prediction module is configured to analyze the interaction behavior data by using an abnormality prediction model to obtain an early warning result, wherein the abnormality prediction model is obtained by training. The update module is configured to optimize and update the business document push generation module and the knowledge graph construction module based on the early warning result, obtain an updated business voucher knowledge graph, and complete visual editing of the voucher pushed by the business process.

[0005] The business process voucher visual editing system has the advantages that the initial business voucher knowledge graph is constructed by analyzing the voucher pushed by the business process and the original business data, the initial business voucher knowledge graph is optimized and updated, the updated business voucher knowledge graph is obtained, and the visual editing of the voucher pushed by the business process is completed. The visual editing efficiency of the voucher pushed by the business process is greatly improved, and the risk of voucher logic errors is reduced. The risk of voucher logic errors is significantly reduced through intelligent processing. A standardized business voucher knowledge system is established, the accuracy and consistency of financial data are improved, cross-system data is uniformly managed and visually displayed, and decision support capability is improved. Real-time monitoring and early warning are supported, and abnormal situations can be discovered and handled in a timely manner to ensure business continuity.

[0006] Further, the business document push generation module comprises: The multi-element data integration submodule is configured to utilize a mapping matrix to convert multi-element data sets to obtain the original business data. The intelligent rule engine submodule is configured to obtain the business voucher triple structure by calculating rule weights based on the voucher pushed by the business process and the original business data.

[0007] Further, the intelligent rule engine submodule comprises: The construction unit is configured to construct the business voucher triple structure based on the voucher pushed by the business process and the original business data. The weighting unit is configured to analyze rule weights of the business voucher triple structure based on a judgment matrix to obtain a weighted triple structure, wherein the expression of the judgment matrix is: ; wherein, represents legal compliance, represents business priority, represents financial impact degree, represents the importance of index i relative to index j; a decision unit configured to select an optimal path by using a decision tree algorithm based on the weighted triple structure, to obtain a business voucher triple structure.

[0008] The mapping matrix realizes the standardized conversion of heterogeneous data sources, improves the data integration efficiency, and guarantees the data quality and integrity. The business voucher triple is automatically generated based on the rule weight calculation, reducing the manual configuration workload and improving the accuracy and consistency of rule execution. The data-driven intelligent decision is realized, reducing the complexity of business rule configuration. Dynamic rule adjustment and optimization are supported to adapt to business change requirements. A traceable data processing link is provided for audit and compliance management.

[0009] Further, the knowledge graph construction module comprises: a knowledge extraction submodule configured to perform deep semantic analysis on the down-pushed text of the business process by using a knowledge acquisition model, to obtain business rule entities and corresponding entity association relationships; a structured submodule configured to perform structured data processing on the business rule entities and corresponding entity association relationships, to obtain a concept data set; a graph construction submodule configured to analyze the concept data set by using an entity linking algorithm, to obtain an initial business voucher knowledge graph through feature vectorization and similarity calculation.

[0010] Multi-dimensional comprehensive evaluation is realized, balancing compliance, efficiency and risk control. An interpretable decision-making process is provided, enhancing the transparency and credibility of the system. Dynamic weight adjustment is supported to adapt to different business scenarios.

[0011] Further, the visualization analysis module comprises: a data fusion submodule configured to perform standardized processing on the initial business voucher knowledge graph, to map multi-modal data to a shared feature space, and to obtain fusion feature data; a visual interactive submodule configured to obtain user interaction based on the fusion feature data by large-scale graph data rendering; an analysis submodule configured to construct a clause change impact directed acyclic graph based on the user interaction, to calculate the impact strength by using probability propagation, and to obtain interaction behavior data.

[0012] We have established a domain-specific business credential knowledge system, providing strong support for intelligent decision-making. We have achieved visual representation and correlation analysis of knowledge, enhancing business understanding and insight. We also support the incremental construction and dynamic updating of knowledge graphs, ensuring the timeliness of knowledge.

[0013] Furthermore, the abnormality prediction model includes: An input layer, configured to normalize the time series features of the interaction behavior data to obtain standardized features; A long short-term memory layer, configured to iteratively update the standardized features to obtain hidden state data; A graph attention layer is configured to perform graph convolution on the standardized features to obtain an entity graph representation; perform attention weighting processing based on the entity graph representation and the hidden state data to obtain an attention-weighted graph representation; The prediction layer is used to perform a fully connected calculation on the attention-weighted graph representation to obtain a warning result output by the abnormal prediction model.

[0014] It enables intuitive visualization of complex business relationships, improving user understanding and operational efficiency. It supports interactive exploration and analysis, helping users quickly identify business patterns and anomalies. It also provides real-time impact analysis and prediction, supporting proactive risk management.

[0015] Furthermore, the expression of the hidden state data is: ; ; ; ; ; ; in, Represents the output value of the forget gate, which determines what information is discarded from the cell state. express, represents the weight matrix of the forget gate, represents the hidden state at time t-1, represents the input vector at time t, represents the bias vector of the forget gate, represents the output value of the input gate, which determines what new information is stored in the cell state, represents the weight matrix of the input gate, represents the bias vector of the input gate, represents the cell state at time t, The weight matrix representing the candidate values ​​is used to generate new candidate cell states, denotes the hidden state output at the t-th time step, denotes the bias vector of the candidate value, denotes the cell state at the t-1-th time step, denotes the output value of the output gate, which decides what part of the cell state to output, denotes the weight matrix of the output gate, denotes the bias vector of the output gate, denotes the hidden state output at the t-th time step; The expression of the entity graph representation is: ; ; wherein, denotes the entity graph at the L-th layer, denotes the feature vector of the n-th entity node at the L-th layer, denotes the entity graph at the L+1-th layer, denotes the degree matrix, a diagonal matrix, denotes the adjacency matrix with self-connection, denotes the learnable weight matrix at the l-th layer; The expression of the attention-weighted graph representation is: ; ; ; wherein, denotes the attention score at the i-th position, denotes the trainable parameter vector of the attention weight, denotes the transpose matrix of the matrix, denotes the weight matrix, denotes the final hidden state output by the LSTM layer, denotes the weight matrix, denotes the representation vector of the i-th entity at the L-th layer, denotes the attention weight (after softmax normalization) of the i-th entity, denotes the attention score of the j-th entity, and n denotes the total number of entities, denotes the context vector; The expression of the early warning result is: ; ; wherein, denotes the output of the fully connected layer, represents the weight matrix of the prediction layer, represents the fused feature vector, represents the bias term, Represents the predicted value.

[0016] It achieves a deep fusion of time series and graph structure information, with prediction performance significantly superior to traditional methods. It supports multi-step prediction and real-time warning, effectively improving the timeliness of warnings.

[0017] Furthermore, the loss function of the abnormality prediction model training is expressed as: ; ; ; ; ; ; in, represents the cross entropy loss, represents the total number of samples, represents the weight of the i-th sample, represents the true label of the i-th sample, represents the predicted value of the i-th sample, represents the number of normal samples, Indicates the number of samples in a game. represents the L2 regularization term, represents the jth parameter, represents the learnable parameters, represents the contrast loss, Represents similarity labels, represents an example between two eigenvectors, represents the fused feature vector, represents the fused feature vector, Indicates the maximum value, represents the margin parameter, defines the example between dissimilar samples, P represents the set of sample pairs, represents the total loss, represents the loss weight of the regularization loss, Represents the loss weight of the contrastive loss.

[0018] Furthermore, the update module includes: The anomaly detection submodule is used to analyze the business credential triple structure through multi-layer verification and anomaly detection algorithms to obtain a corrected business credential triple structure; The atlas updating submodule is configured to construct a business credential knowledge graph based on the business credential triple structure, the business rule entity, and the corresponding entity association relationship; and based on the early warning result, the business credential knowledge graph is optimized and updated by using an incremental updating algorithm to obtain an updated business credential knowledge graph.

[0019] Further, the anomaly detection submodule comprises: The verification unit is configured to analyze the business credential triple structure by using a multi-layer verification and anomaly detection algorithm to obtain an anomaly score. The analysis unit is configured to analyze the candidate correct value corresponding to the anomaly data based on the anomaly score, and select the candidate correct value with the highest probability as a correction suggestion value. The correction unit is configured to correct the business credential triple structure based on the correction suggestion value to obtain a corrected business credential triple structure.

[0020] A complete mathematical modeling framework is provided to ensure the theoretical rigor and reproducibility of the algorithm. Through precise mathematical expressions, the precise control and optimization of model parameters are realized. The quantitative analysis and performance evaluation of the model are supported, which facilitates continuous improvement and optimization. A solid theoretical foundation is provided for the industrial application of related technologies. Through standardized mathematical representation, technology exchange and knowledge inheritance are promoted. BRIEF DESCRIPTION OF DRAWINGS

[0021] The present specification will be further illustrated in the form of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting, and in these embodiments, the same reference numbers represent the same structures, wherein: Figure 1 is a module schematic diagram of a business process push-down credential visual editing system according to some embodiments of the present specification. DETAILED DESCRIPTION

[0022] The specific embodiments of the present application are described below to facilitate understanding of the present application by those skilled in the art, but it should be clear that the present application is not limited to the scope of the specific embodiments. For those skilled in the art, it is obvious that various changes are within the spirit and scope of the present application as defined in the appended claims, and all applications utilizing the concept of the present application are within the scope of protection.

[0023] EMBODIMENT Figure 1 is a module schematic diagram of a business process push-down credential visual editing system according to some embodiments of the present specification.

[0024] In some embodiments, the business process push-down credential visualized editing system can include a business document push-down generation module, a knowledge graph construction module, a visualized analysis module, an anomaly prediction module, and an updating module.

[0025] The business document push-down generation module is configured to construct a business credential triple structure based on the business process push-down credential and the original business data by using multi-dimensional data integration and anomaly detection.

[0026] The business process push-down credential is a financial term related to credentials. For example, the business process push-down credential can include “receivable credential generation” and “income confirmation”.

[0027] The original business data is a business event and object. For example, the original business data can include contract signing, prepayment receipt, and acceptance completion.

[0028] The business credential triple structure is a mapping rule of the original business data and the business process push-down credential, and is a constructed triple.

[0029] In some embodiments, the business document push-down generation module can include a multi-dimensional data integration submodule and an intelligent rule engine submodule.

[0030] The multi-dimensional data integration submodule is configured to convert multi-dimensional data sets by using a mapping matrix to obtain the original business data.

[0031] The mapping matrix is a five-tuple structure used for intelligent conversion and standardization of heterogeneous system data. For example, the expression of the mapping matrix can be: wherein M represents the mapping matrix, S represents the source system metadata set, T represents the target standard model metadata set, R represents the mapping relationship set from S to T, F represents the conversion function set, and C represents the constraint condition set.

[0032] In some embodiments, the multi-dimensional data integration submodule can use the mapping matrix to map the source system data item s∈S to the target data item t∈T through the mapping relationship r∈R, satisfy the constraint condition c∈C, and apply the conversion function f∈F to obtain the converted value f(s).

[0033] The multi-dimensional data set is a distributed micro-service architecture that realizes data interconnection between heterogeneous systems by combining adapter services and message middleware. For example, the multi-dimensional data set can include key information of documents (such as contract subjects, amounts, and performance clauses) connected to ERP, contract management systems, etc.

[0034] In some embodiments, each source system (such as an ERP, a contract management system) is equipped with a dedicated adapter microservice responsible for data extraction and standardized conversion, and asynchronous communication is realized through a message queue to ensure decoupling and high availability between systems. Workflow: The adapter service accesses the source system API or database through connection parameters, extracts business data according to the change capture strategy (CDC, timestamp comparison or snapshot comparison), converts the extracted data into a standard message format containing metadata and business data, and sends the message to the message queue through a partitioned topic to ensure processing order and reliability. Downstream processing services subscribe to the corresponding topic, obtain and standardize the data, and the data standardization uses a mapping matrix model to realize the mapping of source system data to a standard model through predefined conversion functions. In the conversion function F(x) = M[x], M represents the mapping matrix, which contains the correspondence and conversion rules between the source system, the source field and the target field.

[0035] The intelligent rule engine submodule is configured to obtain a business voucher triple structure by calculating rule weights based on the pushed-down voucher and the original business data.

[0036] In some embodiments, the intelligent rule engine submodule can include a construction unit, a weighting unit and a decision unit.

[0037] The construction unit is configured to construct a business voucher triple structure based on the pushed-down voucher and the original business data.

[0038] In some embodiments, the construction unit can use a triple structure (subject-relation-object), such as a triple structure (subject-relation-object) of (contract signing-trigger-receivable voucher generation), as a rule representation, where the subject is a business event or object, such as "contract signing", "prepayment received", "acceptance completed", the relation is a trigger action or condition, such as "trigger", "result in", "when the condition is met", and the object is a financial operation to be performed, such as "receivable voucher generation", "income recognition"; by extracting key event points from business process and financial requirement analysis, defining standardized business terms as subjects for each event point, defining a limited set of relational verbs to ensure clear semantics, and mapping to financial operation terms as objects, the triple structure is formalized through a rule description language (RDL) to obtain a business voucher triple structure.

[0039] The weighting unit is configured to analyze the rule weights of the business voucher triple structure based on a judgment matrix to obtain a weighted triple structure.

[0040] The judgment matrix is a matrix used as a rule evaluation index.

[0041] In some embodiments, the expression of the judgment matrix can be: ; wherein, represents legal compliance, represents business priority, represents financial impact, represents the importance of index i relative to index j.

[0042] The weighted triple structure is a triple structure in which weights are assigned to each rule in the business credential triple structure.

[0043] In some embodiments, the weighting unit can calculate a feature vector, normalize the judgment matrix A to obtain a normalized matrix B = [bij]n×n, calculate the row vector sum of the normalized matrix, normalize the row vector sum to obtain a weight vector W = [WL, WB, WF], perform consistency checking, calculate the maximum eigenvalue λmax= Σ(i=1 to n) (AW)i / (nWi), the consistency index is CI = (λmax-n) / (n-1), the consistency ratio CR = CI / RI, if CR < set threshold, then pass the test; score each rule on the three indexes S = [SL, SB, SF], and the final rule weight W = WL×SL + WB×SB + WF×SF.

[0044] The decision unit is configured to select an optimal path based on the weighted triple structure using a decision tree algorithm to obtain the business credential triple structure.

[0045] In some embodiments, the decision unit can select an optimal path through a decision tree algorithm, and the selection basis is Max(Wi×Ci), Wi is the rule weight, and Ci is the confidence. The role of the decision path is to determine the order and combination of rule execution to produce the optimal credential generation result, avoid rule conflict and repeated execution, and obtain the business credential triple structure.

[0046] In some embodiments, the intelligent rule engine submodule can define the triple rule through a visual interface, the system automatically performs rule syntax verification and logical conflict detection, and the analytic hierarchy process is used to calculate and store the rule weight. Rule activation and matching stage: listen to business events (such as contract signing), find all potentially applicable rules from the rule base, evaluate the condition part of each rule, and filter out the rule set that meets the conditions. Execution plan generation stage: when multiple rules are applicable, a decision tree is constructed, the Wi x Ci value of each rule is calculated, and the maximum value strategy is used to select the execution path to generate an ordered rule execution plan. Rule execution stage: call the rule action according to the execution plan order, generate the corresponding financial voucher or business instruction, record the execution log and result. Result verification and feedback stage: perform result verification based on business rules, trigger manual review process in abnormal cases, and update rule confidence based on execution results.

[0047] The knowledge graph construction module is configured to construct an initial business voucher knowledge graph based on the business voucher triple structure, the business rule entity, and the corresponding entity association relationship.

[0048] The business rule entity and the corresponding entity association relationship are extracted from the contract text and the business rule.

[0049] In some embodiments, the knowledge graph construction module can use natural language processing technology to extract text from the contract text and the business rule to obtain the business rule entity and the corresponding entity association relationship.

[0050] The initial business voucher knowledge graph is a knowledge graph constructed based on the business rule entity and the corresponding entity association relationship.

[0051] In some embodiments, the knowledge graph construction module can include a knowledge extraction submodule, a structured submodule, and a graph construction submodule.

[0052] The knowledge extraction submodule is configured to use a knowledge acquisition model to perform deep semantic analysis on the business process pushdown text to obtain the business rule entity and the corresponding entity association relationship.

[0053] The knowledge acquisition model is a language model used to acquire business entities and relationships. For example, the knowledge acquisition model can include a BERT layer, a BiLSTM layer, and a CRF layer.

[0054] The BERT layer is configured to analyze the business process pushdown text to capture the dynamic meaning of words in the context and obtain deep semantic representations.

[0055] The BiLSTM layer is configured to capture long-distance dependencies of deep semantic representations to obtain enhanced semantic sequences.

[0056] The CRF layer is used for optimizing the enhanced semantic sequence with labels to obtain the business rule entity and the corresponding entity association relationship.

[0057] In some embodiments, the knowledge extraction submodule can utilize a pre-trained language model to accurately extract business entities and relationships from unstructured text, providing basic data for knowledge graph construction. For example, the BERT layer can provide deep semantic representation, capturing the dynamic meaning of words in context; the BiLSTM layer can enhance sequence processing capability, capturing long-distance dependency; and the CRF layer can optimize the label sequence to ensure the consistency and legality of the output labels.

[0058] In some embodiments, the knowledge extraction submodule can process unstructured contract text, business rule documents, financial system documents, etc., by removing stop words, punctuation processing, and special character unification; converting the segmented text into context-sensitive word embedding representation; capturing long-distance dependencies in the sequence; optimizing the overall label sequence to ensure annotation consistency, obtaining entity annotation results, relationship candidate pairs, and entity attribute value extraction results.

[0059] The structured submodule is used for structured data processing of the business rule entity and the corresponding entity association relationship, to obtain a concept data set.

[0060] The concept data set is a set constructed by concept type data. For example, the concept data set can include a concept class node set, an attribute definition set, a relationship type definition, and an instance data set.

[0061] In some embodiments, the structured submodule can convert business system metadata into graph nodes. The structured data processing technology converts table structures, field definitions, and association relationships in the relational database into concepts, attributes, and relationships of the knowledge graph through metadata analysis and pattern mapping, realizes automatic conversion from structured data to graph model, and obtains a concept data set. Specifically, the structured submodule can process business system database metadata, table structure definition, field description and constraint condition, and foreign key relationship, extract table, column, and constraint definition, process data format inconsistency, missing values, etc., map the table to a concept class node, map the column to an attribute definition, and map the foreign key to a relationship between concepts, to obtain a concept data set.

[0062] The initial ontology architecture guides the first round of entity extraction, and the extraction results, in turn, enrich and optimize the ontology structure. The optimized ontology then guides the next round of more accurate entity extraction. The neural network's entity classification corresponds to the concept layer of the ontology, the neural network's relationship extraction corresponds to the ontology's relationship layer, and the neural network's attribute extraction corresponds to the ontology's attribute layer. Neural network models excel at extracting explicit knowledge from unstructured text, while the ontology architecture provides implicit knowledge and reasoning capabilities. The combination of the two enables the complete conversion process from text to knowledge. Through this multi-level, multi-dimensional processing architecture, the system can effectively extract knowledge from various data sources and construct a complete knowledge graph that meets the needs of business-financial mapping, providing a solid knowledge foundation for the rule engine.

[0063] The graph construction submodule is used to analyze the concept data set using the entity linking algorithm, and obtain the initial business credential knowledge graph through feature vectorization and similarity calculation.

[0064] In some embodiments, the graph construction submodule can use an entity linking algorithm to achieve alignment and fusion of entities from different sources based on cosine similarity, identify and merge multiple entity nodes representing the same business object through feature vectorization and similarity calculation, ensure the consistency and integrity of the knowledge graph, and obtain the initial business credential knowledge graph. Specifically, the graph construction submodule can be organized in a multi-layer architecture, dividing the knowledge graph into different levels according to the degree of abstraction and function, supporting query requirements of different granularity, and improving system flexibility and performance, including: concept layer: basic concept definitions in the business and financial fields; entity layer: specific business entities and their attribute data; event layer: records of business activities and financial events; rule layer: graph representation of business rules and processing logic; wherein the similarity calculation formula is S(A,B)=A·B / (|A|×|B|), S represents the similarity result, and A and B represent entity nodes.

[0065] The visualization analysis module is used to perform topological display on the initial business credential knowledge graph, extract structure and attribute information, and obtain interactive behavior data.

[0066] Interaction behavior data is information related to user interaction behavior. For example, interaction behavior data can include time series feature vectors, entity association graphs, and business rule features corresponding to user interactions.

[0067] In some embodiments, the visualization analysis module may include a data fusion submodule, a visualization interaction submodule, and an analysis submodule.

[0068] The data fusion submodule is used to standardize the initial business credential knowledge graph, map the multimodal data into a shared feature space, and obtain fused feature data.

[0069] The fusion feature data is feature data fused based on importance weights of different modal data.

[0070] In some embodiments, the data fusion sub-module can standardize the initial business credential knowledge graph, map different modal data to a shared feature space, calculate importance weights of different modal data using self-attention, adopt a late fusion strategy to fuse multi-modal data, and obtain fusion feature data.

[0071] The visualization interaction sub-module is configured to obtain user interaction conditions by large-scale graph data rendering based on the fusion feature data.

[0072] The user interaction conditions are condition data reflecting user interaction behaviors.

[0073] In some embodiments, the visualization interaction sub-module can obtain user interaction conditions by using large-scale graph data rendering as an interaction engine based on the fusion feature data, adopting an interaction scheme based on an observer pattern to realize separation of QUI and a data model and adopting a Linked Views technology to realize multi-dimensional data linkage, using a sliding window to process real-time data streams, updating only changed nodes to reduce rendering load, and using a real-time data processing mode based on user behavior prediction to pre-load possible browsing data.

[0074] The analysis sub-module is configured to construct a clause change influence directed acyclic graph based on the user interaction conditions, calculate influence strength using probability propagation, and obtain interaction behavior data.

[0075] In some embodiments, the analysis sub-module can construct a clause change influence directed acyclic graph based on the user interaction conditions, calculate influence strength using probability propagation, set a threshold, and mark a node as a high-risk node when the influence strength exceeds the threshold, and obtain interaction behavior data.

[0076] The anomaly prediction module is configured to analyze the interaction behavior data using an anomaly prediction model to obtain an early warning result, wherein the anomaly prediction model is obtained by training.

[0077] The anomaly prediction model is used for...... The type of the anomaly prediction model can be various. For example, the type of the anomaly prediction model can include......

[0078] In some embodiments, the input of the anomaly prediction model can be the interaction behavior data, and the output of the anomaly prediction model can be the early warning result.

[0079] In some embodiments, the structure of the anomaly prediction model is as follows: The abnormality prediction model comprises an input layer, a long short-term memory layer, a graph attention layer and a prediction layer. The output of the input layer is taken as the input of the long short-term memory layer and the graph attention layer, the output of the long short-term memory layer is taken as the input of the graph attention layer, and the output of the graph attention layer is taken as the final output of the abnormality prediction model.

[0080] The input layer is configured to perform time series feature standardization on the interaction behavior data to obtain standardized features. The input of the input layer can comprise the interaction behavior data, and the output can comprise the standardized features.

[0081] The standardized features are the interaction behavior data after standardization processing.

[0082] In some embodiments, the input layer can perform time series feature standardization on the interaction behavior data by using forward padding or average value padding, perform outlier truncation, and obtain the standardized features.

[0083] The long short-term memory layer is configured to perform iterative update processing on the standardized features to obtain hidden state data. The input of the long short-term memory layer can comprise the standardized features, and the output can comprise the hidden state data.

[0084] The hidden state data is the updated standardized features.

[0085] In some embodiments, the expression of the hidden state data can be: ; ; ; ; ; ; wherein, represents the output value of the forget gate, which decides what information is discarded from the cell state, represents, represents the weight matrix of the forget gate, represents the hidden state at the t-1 time, represents the input vector at the t time, represents the bias vector of the forget gate, represents the output value of the input gate, which decides what new information is stored in the cell state, represents the weight matrix of the input gate, represents the bias vector of the input gate, represents the cell state at the t time, represents the weight matrix of the candidate value, which is used to generate a new candidate cell state, denotes the hidden state output at the t-th time point, denotes the bias vector of the candidate value, denotes the cell state at the t-1-th time point, denotes the output value of the output gate, deciding what part of the cell state is output, denotes the weight matrix of the output gate, denotes the bias vector of the output gate, denotes the hidden state output at the t-th time point.

[0086] The graph attention layer is used to perform graph convolution on the standardized features to obtain an entity graph representation. Based on the entity graph representation and the hidden state data, attention weighting processing is performed to obtain an attention-weighted graph representation. The input of the graph attention layer can include the entity graph representation and the hidden state data, and the output can include the attention-weighted graph representation.

[0087] The entity graph representation is a graph representation of the standardized features after convolution.

[0088] In some embodiments, the expression of the entity graph representation can be: ; ; wherein, denotes the entity graph at the L-th layer, denotes the feature vector of the n-th entity node at the L-th layer, denotes the entity graph at the L+1-th layer, denotes the degree matrix, a diagonal matrix, denotes the adjacency matrix with self-connection, denotes the learnable weight matrix at the l-th layer.

[0089] In some embodiments, the expression of the attention-weighted graph representation can be: ; ; ; wherein, denotes the attention score at the i-th position, denotes the trainable parameter vector of the attention weight, denotes the transpose matrix of the matrix, denotes the weight matrix, denotes the final hidden state representing the output of the LSTM layer, denotes the weight matrix, denotes the representation vector of the i-th entity at the L-th layer, attention weight (softmax normalized) of the i-th entity, attention score of the j-th entity, n represents the total number of entities, context vector.

[0090] a prediction layer, configured to perform full connection calculation on the attention-weighted graph representation to obtain an early warning result of an anomaly prediction model output. The input of the prediction layer can include, and the output can include.

[0091] The early warning result is a result of a triplet anomaly probability.

[0092] In some embodiments, the expression of the early warning result can be: ; ; wherein, output of the full connection layer, weight matrix of the prediction layer, fused feature vector, bias term, predicted value.

[0093] In some embodiments, the anomaly prediction model can be trained by a plurality of labeled training samples. For example, a plurality of labeled training samples can be input into an initial anomaly prediction model, a loss function is constructed based on the labels and the results of the initial anomaly prediction model, and the parameters of the initial anomaly prediction model are iteratively updated based on the loss function by gradient descent or other methods. When a preset condition is met, the model training is completed, and a trained anomaly prediction model is obtained. The preset condition can be convergence of the loss function, number of iterations reaching a threshold, etc.

[0094] In some embodiments, the training samples can include historical interaction behavior data. The labels can be corresponding real early warning results. The labels can be manually annotated.

[0095] In some embodiments, the expression of the loss function of the anomaly prediction model training is: ; ; ; ; ; ; wherein, cross-entropy loss, total number of samples, represents the weight of the i-th sample, represents the true label of the i-th sample, represents the predicted value of the i-th sample, represents the number of normal samples, Indicates the number of samples in a game. represents the L2 regularization term, represents the jth parameter, represents the learnable parameters, represents the contrast loss, Represents similarity labels, represents an example between two eigenvectors, represents the fused feature vector, represents the fused feature vector, Indicates the maximum value, represents the margin parameter, defines the example between dissimilar samples, P represents the set of sample pairs, represents the total loss, represents the loss weight of the regularization loss, Represents the loss weight of the contrastive loss.

[0096] In some embodiments, the abnormality prediction model constraints may include time constraints: whether the prediction delay meets real-time requirements; accuracy constraints: false alarm rate <5%, missed alarm rate <1%.

[0097] In some embodiments, the optimization algorithm of the anomaly prediction model may include an Adam optimizer, and training is stopped when the performance of the validation set does not improve over multiple consecutive rounds.

[0098] The update module is used to optimize and update the business document push-down generation module and the knowledge graph construction module based on the warning results, obtain the updated business voucher knowledge graph, and complete the visual editing of the business process push-down voucher.

[0099] The updated business voucher knowledge graph is a business voucher knowledge graph that is updated by correcting abnormal data.

[0100] In some embodiments, the update module may include an anomaly detection submodule and a map update submodule.

[0101] The anomaly detection submodule is used to analyze the business credential triple structure through multi-layer verification and anomaly detection algorithms to obtain a corrected business credential triple structure.

[0102] In some embodiments, the anomaly detection submodule may include a verification unit, an analysis unit, and a correction unit.

[0103] The verification unit is configured to analyze the abnormal score by using a multi-layer verification and an abnormality detection algorithm based on the business voucher triple structure.

[0104] In some embodiments, the verification unit can perform field-level verification on the format, value range, and logical validity of a single field, document-level verification on the logical relationship, consistency, and integrity among multiple fields in a document, and associated document-level verification on the business logic consistency among multiple documents associated with each other, thereby completing the multi-layer verification of the business voucher triple structure.

[0105] In some embodiments, the verification unit can use an isolated forest algorithm to train an abnormal value detector based on historical data, randomly subsample to construct multiple random decision trees, randomly select a feature and a split point for each node, and calculate the average path length of each data point being isolated. Finally, the abnormal score is obtained by standardization.

[0106] The analysis unit is configured to analyze the candidate correct value corresponding to the abnormal data based on the abnormal score, and select the candidate correct value with the highest probability as the correction suggestion value.

[0107] In some embodiments, the analysis unit can provide the most likely correct value by applying Bayesian inference to the abnormality. The Bayesian inference automatic correction mechanism is based on Bayesian probability theory, establishes a prior probability distribution based on historical data, calculates a posterior probability in combination with current context information, and infers the most likely correct value of the abnormal data. First, the abnormal field value and its context information are received, and a set of candidate correct values is obtained from the prior distribution. The likelihood probability of each candidate value is calculated, the posterior probability is synthesized and normalized. Finally, the value with the highest posterior probability is selected as the correction suggestion, and the suggestion value and its confidence are output as the correction suggestion value.

[0108] The correction unit is configured to correct the business voucher triple structure based on the correction suggestion value, to obtain a corrected business voucher triple structure.

[0109] The graph updating submodule is configured to construct a business voucher knowledge graph based on the business voucher triple structure, business rule entities, and corresponding entity association relationships, and to optimize and update the business voucher knowledge graph based on the early warning result by using an incremental updating algorithm, to obtain an updated business voucher knowledge graph.

[0110] In some embodiments, the graph updating submodule can track entity attribute changes based on the change log using an incremental update algorithm, enabling efficient dynamic updating of the knowledge graph, avoiding the high cost of full graph reconstruction, and ensuring the real-time and accuracy of the graph data. Specifically, the graph updating submodule can capture source system change events based on the business credential knowledge graph, convert them into graph operation instructions, apply changes and record change logs, and trigger rule reevaluation in the affected areas; implement graph history version management using multi-version concurrency control, including timestamp marking: each change is attached with timestamp information, snapshot storage: complete graph snapshot of key states, difference storage: incremental change records between versions, version chain: establish the front and back association between versions for version control; ensure graph consistency in a distributed environment through a two-phase commit protocol, including a preparation phase and a commit phase. Preparation phase: the coordinator sends preparation messages to all participating nodes, and the participating nodes verify the feasibility of the operation. Commit phase: after all participating nodes are prepared successfully, the coordinator sends a commit command, otherwise an abort command is sent, completing consistency maintenance; through the update impact range evaluation algorithm and graph analysis technology, predict the nodes and rule range that may be affected by a graph update, trigger rule reevaluation accordingly, improve update efficiency and reduce system load. First, build a dependency graph: record entity dependencies and rule dependencies. Second, mark the change source: identify the initial change point. Then, propagation analysis: propagate the impact label along the dependency edge, and finally, generate an impact report: list all affected entities and rules, complete the update propagation, and obtain the updated business credential knowledge graph.

[0111] In some embodiments of the present specification, a business process push-down credential visualization editing system is provided. By analyzing the business process push-down credential and the original business data, an initial business credential knowledge graph is constructed. The initial business credential knowledge graph is optimized and updated to obtain an updated business credential knowledge graph, and the visualization editing of the business process push-down credential is completed. The visualization editing efficiency of the business process push-down credential is greatly improved, and the risk of credential logic error is reduced. The risk of credential logic error is significantly reduced through intelligent processing. A standardized business credential knowledge system is established, improving the accuracy and consistency of financial data. Cross-system data is unified managed and visualized, improving decision support capabilities. Real-time monitoring and early warning are supported, enabling timely detection and handling of abnormal situations to ensure business continuity.

Claims

1. A business process push-down voucher visual editing system, characterized by: include: The business document push-down generation module is used to push down documents and original business data based on business processes, and build a business document triple structure by using multivariate data integration and anomaly detection; A knowledge graph construction module, configured to construct an initial business credential knowledge graph based on the business credential triple structure, business rule entities, and corresponding entity association relationships; A visualization analysis module is used to perform topological display on the initial business credential knowledge graph, extract structure and attribute information, and obtain interaction behavior data; An anomaly prediction module, configured to analyze the interaction behavior data using an anomaly prediction model to obtain an early warning result; wherein the anomaly prediction model is obtained through training; The update module is used to optimize and update the business document push-down generation module and the knowledge graph construction module based on the warning results, obtain the updated business voucher knowledge graph, and complete the visual editing of the business process push-down voucher.

2. The business process push-down voucher visual editing system according to claim 1 is characterized in that: The business document push-down generation module includes: The multivariate data integration submodule is used to transform the multivariate data set using the mapping matrix to obtain the original business data; The intelligent rule engine submodule is used to push down the voucher and the original business data based on the business process, and obtain the business voucher triple structure by calculating the rule weight.

3. The business process push-down voucher visual editing system according to claim 2 is characterized in that: The intelligent rule engine submodule includes: A construction unit, configured to construct a business voucher triplet structure based on the business process push-down voucher and the original business data; The weighting unit is used to analyze the rule weights of the business voucher triple structure based on the judgment matrix to obtain a weighted triple structure; wherein the judgment matrix is ​​expressed as: ； in, Indicates legal compliance, Indicates business priority, Indicates the financial impact, Indicates the importance of indicator i relative to indicator j; The decision unit is used to select an optimal path based on the weighted triple structure using a decision tree algorithm to obtain a business voucher triple structure.

4. The business process push-down voucher visual editing system according to claim 1 is characterized in that: The knowledge graph construction module includes: The knowledge extraction submodule is used to use the knowledge acquisition model to perform deep semantic analysis on the business process push-down text to obtain the business rule entities and the corresponding entity association relationships; A structuring submodule, configured to perform structured data processing on the business rule entities and the corresponding entity association relationships to obtain a conceptual data set; The graph construction submodule is used to analyze the concept data set using the entity linking algorithm, and obtain the initial business credential knowledge graph through feature vectorization and similarity calculation.

5. The business process push-down voucher visual editing system according to claim 1 is characterized in that: The visual analysis module includes: A data fusion submodule is used to standardize the initial business credential knowledge graph, map the multimodal data into a shared feature space, and obtain fused feature data; The visualization interaction submodule is used to obtain user interaction information through large-scale graph data rendering based on fused feature data; The analysis submodule is used to construct a directed acyclic graph of the impact of the term change based on the user interaction situation, calculate the impact intensity using probability propagation, and obtain interaction behavior data.

6. The business process push-down voucher visual editing system according to claim 1 is characterized in that: The abnormality prediction model includes: An input layer, configured to normalize the time series features of the interaction behavior data to obtain standardized features; A long short-term memory layer, configured to iteratively update the standardized features to obtain hidden state data; A graph attention layer is configured to perform graph convolution on the standardized features to obtain an entity graph representation; perform attention weighting processing based on the entity graph representation and the hidden state data to obtain an attention-weighted graph representation; The prediction layer is used to perform a fully connected calculation on the attention-weighted graph representation to obtain a warning result output by the abnormal prediction model.

7. The business process push-down voucher visual editing system according to claim 6 is characterized in that: The expression of the hidden state data is: ； ； ； ； ； ； in, Represents the output value of the forget gate, which determines what information is discarded from the cell state. express, represents the weight matrix of the forget gate, represents the hidden state at time t-1, represents the input vector at time t, represents the bias vector of the forget gate, represents the output value of the input gate, which determines what new information is stored in the cell state, represents the weight matrix of the input gate, represents the bias vector of the input gate, represents the cell state at time t, The weight matrix representing the candidate values ​​is used to generate new candidate cell states, represents the hidden state output at time t, represents the bias vector of candidate values, represents the cell state at time t-1, Represents the output value of the output gate, which determines what part of the cell state is output. represents the weight matrix of the output gate, represents the bias vector of the output gate, Represents the hidden state output at time t; The expression represented by the entity graph is: ； ； in, represents the entity graph of level L, Represents the feature vector of the nth entity node in the Lth layer, represents the entity graph of the L+1th layer, represents the degree matrix, a diagonal matrix, represents the adjacency matrix with self-connection, represents the learnable weight matrix of layer l; The expression of the attention weighted graph representation is: ； ； ； in, represents the attention score of the i-th position, A trainable parameter vector representing the attention weights, express The transpose of a matrix, represents the weight matrix, represents the final hidden state output by the LSTM layer, represents the weight matrix, represents the representation vector of the i-th entity at the L-th layer, represents the attention weight of the i-th entity (normalized by softmax), represents the attention score of the j-th entity, n represents the total number of entities, represents the context vector; The expression of the early warning result is: ； ； in, represents the output of the fully connected layer, represents the weight matrix of the prediction layer, represents the fused feature vector, represents the bias term, Represents the predicted value.

8. The business process push-down voucher visual editing system according to claim 7 is characterized in that: The loss function of the abnormality prediction model training is expressed as: ； ； ； ； ； ； in, represents the cross entropy loss, represents the total number of samples, represents the weight of the i-th sample, represents the true label of the i-th sample, represents the predicted value of the i-th sample, represents the number of normal samples, Indicates the number of samples in a game. represents the L2 regularization term, represents the jth parameter, represents the learnable parameters, represents the contrast loss, represents the similarity label, represents an example between two eigenvectors, represents the fused feature vector, represents the fused feature vector, Indicates the maximum value, represents the margin parameter, defines the example between dissimilar samples, P represents the set of sample pairs, represents the total loss, represents the loss weight of the regularization loss, Represents the loss weight of the contrastive loss.

9. The business process push-down voucher visual editing system according to claim 1 is characterized in that: The update module includes: The anomaly detection submodule is used to analyze the business credential triple structure through multi-layer verification and anomaly detection algorithms to obtain a corrected business credential triple structure; The graph update submodule is used to construct a business credential knowledge graph based on the business credential triple structure, business rule entities and corresponding entity association relationships; based on the early warning results, the business credential knowledge graph is optimized and updated using an incremental update algorithm to obtain an updated business credential knowledge graph.

10. The business process push-down voucher visual editing system according to claim 9 is characterized in that: The anomaly detection submodule includes: The verification unit is used to analyze the business credential triple structure through multi-layer verification and anomaly detection algorithms to obtain an anomaly score; an analyzing unit, configured to analyze candidate correct values ​​corresponding to the abnormal data based on the abnormality score, and select the candidate correct value with the highest probability as the correction suggestion value; The correction unit is used to correct the business voucher triple structure based on the correction suggestion value to obtain a corrected business voucher triple structure.

Images