Synthetic data generation method and system fusing AI modeling and rule engine

Abstract

The invention belongs to the technical field of synthetic data generation, particularly relates to a synthetic data generation method and system fusing AI modeling and a rule engine, and is suitable for a multi-label and multi-constraint structured data generation scene. The method comprises the steps of constructing a label set, establishing a standardized data table and control constraints, performing label vector clustering, training a basic generator model to generate candidate synthetic data, and dividing the candidate data into partial overrun data and complete overrun data through control constraint judgment; a precision completion generator is constructed for correction according to part of overrun data; and for the complete overrun data, constructing a supplementary label set, training a residual completion generator to generate residual data, and performing weighted fusion to generate final synthetic data. The system comprises a plurality of functional modules, supports data dynamic regulation and control and compliance auditing, and improves the controllability, integrity and applicability of generated data.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of synthetic data generation, and particularly relates to a synthetic data generation method and system fusing AI modeling and a rule engine. BACKGROUND

[0002] With the development of artificial intelligence technology, data-driven models have made significant progress in financial risk control, intelligent manufacturing, medical diagnosis and other fields. High-quality data, as the core basis for model training, has become an important guarantee for improving algorithm performance and adaptability. Due to the problems of difficult data acquisition, sensitive data cannot be used, and sample distribution is not balanced in actual scenarios, synthetic data has gradually become an effective means to replace or supplement real data, especially in the fields of privacy protection, control of experimental variables, and strengthening of specific labels.

[0003] The existing synthetic data generation methods mainly include two technical routes: one is to rely on deep learning models (such as GAN, VAE, etc.) for sample generation, which has the advantages of being able to simulate complex distribution structures, but has the defects of strong unexplainability and lack of control ability over the generated results; the other is to construct a conditional generation process based on a rule engine or an expert system, which has high explainability and domain adaptation ability, but often lacks learning ability, resulting in limited data diversity. Some improved methods try to introduce constraint learning, goal guidance or domain label intervention mechanisms to improve the generation effect, but there are still deficiencies in high-dimensional label control, multi-model completion collaboration and synthetic precision regulation in complex business scenarios. In addition, most of the current methods lack a standardized label structure management mechanism, and the generation process is difficult to achieve traceability and dynamic adjustment, which is not conducive to enterprises to build a controllable and auditable data generation system in actual application.

[0004] Therefore, it is urgent to propose a synthetic data generation method fusing AI modeling ability and a rule engine to meet the actual application requirements of multi-source label driving, index controllability, precision adjustment, and process auditability. SUMMARY

[0005] In view of the above problems, the purpose of the present application is to propose a synthetic data generation method fusing AI modeling and a rule engine, comprising the following steps: S1, constructing a label set: collecting original data and constructing a label set corresponding to a target application scenario, the label set comprising a plurality of label fields for describing key elements; S2, establishing a standardized data table and control constraints: determining control constraint indicators according to the label fields and their associated relationships, and constructing a first standardized data table based on the label set, the fields of the first standardized data table corresponding one by one to the label fields; S3, constructing a label vector set and clustering division: based on the original data and the first standardized data table, a first association relationship is established to construct a label vector set, and a clustering operation is performed on the label vector set to divide it into multiple initial data sets; S4, training a base generator model to generate candidate synthetic data: based on the initial data set, a deep neural network model is trained as a base generator to generate candidate synthetic data, each field of the candidate synthetic data corresponding to a field of the first standardized data table; S5, execution control judgment and abnormal division: according to the control constraint index, the candidate synthetic data is judged, if it does not meet the constraint index, the candidate synthetic data is classified as a partial out-of-limit data set or a complete out-of-limit data set through a judgment module; S6, constructing an accuracy completion model to optimize candidate data: combining the partial out-of-limit data set with the original data to construct a revised data set, and based on the revised data set, a second deep neural network model is trained as an accuracy completion generator to revise and optimize the candidate synthetic data; S7, generating residual completion data and synthesizing final data: based on the difference value of the out-of-limit data set, a supplementary label data set is constructed, a residual completion model is trained to generate residual data, and the final synthetic data is obtained by weighted fusion of the candidate data; S8, output and log record: output the final synthetic data and the second standardized data table, and construct a synthetic data generation log.

[0006] As a preferred technical solution, the label fields in step S1 include at least geographical information label, time label, device information label, user behavior label and time limit control label for measuring the timeliness of synthetic data generation.

[0007] As a preferred technical solution, in step S3, the clustering operation of the label vector set includes: calculating the distribution density of each label vector in the label vector set in the original data, and dividing the label vector set into high-density subsets and low-density subsets to form initial data sets.

[0008] As a preferred technical solution, in step S5, the control constraint index includes a first-level control constraint constructed based on the label field and a second-level control constraint generated based on the supplementary label; The first-level control constraint includes parameters for specifying the upper limit, lower limit or interval range of the label field value; The second-level control constraint is determined by comparing the dispersion of residual data and candidate synthetic data in each label field.

[0009] As a preferred technical solution, in step S6, the precision completion generator is composed of an initial neural network model and a completion neural network model constructed based on a partial over-limit dataset in parallel, and outputs the fused corrected data.

[0010] As a preferred technical solution, step S7 specifically includes: The label vector difference operation is performed between the fully out-of-limit dataset and the partially out-of-limit dataset to obtain a supplementary label set; A second standardized data table is constructed by adding control constraint indicators to the label fields based on the supplementary label set, and a supplementary dataset is constructed using the supplementary label set and the original data. A third deep neural network model is trained using the supplementary dataset and used as a residual completion generator to generate residual data. The residual data is used to supplement this part of the synthetic data, and the residual data and the candidate synthetic data are weighted and averaged to obtain the corrected synthetic data, which is used as the final synthetic data.

[0011] This invention also provides a synthetic data generation system that integrates AI modeling and rule engines, used to implement the aforementioned synthetic data generation method integrating AI modeling and rule engines, including: The first acquisition module is used to collect raw data and construct a tag set corresponding to the target application scenario. The tag set includes tag fields used to characterize key elements. The second acquisition module is used to determine control constraint indicators based on the label fields and construct a first standardized data table, wherein the fields of the first standardized data table correspond one-to-one with the label fields. The clustering module is used to divide the original data into multiple subsets based on the set of label vectors; The training module is used to train a deep neural network model based on the data subset, which serves as a base generator to generate candidate synthetic data. The judgment module is used to judge the candidate synthetic data according to the control constraint index. If the data does not meet the constraint index, the candidate synthetic data is classified as a partially out-of-limit dataset or a completely out-of-limit dataset. The correction module is used to train a precision completion generator based on a portion of the out-of-limit dataset and correct biases in the candidate synthetic data; The supplementary module is used to train a residual completion generator based on the fully out-of-limit dataset and generate residual data to supplement the synthetic data. The residual data and the candidate synthetic data are weighted and averaged to obtain the final synthetic data.

[0012] As a preferred technical solution, the first acquisition module is also used to extract raw data from the database and perform data cleaning operations to remove missing items and outliers.

[0013] As a preferred technical solution, the supplementary module is further used to perform label difference calculation based on the complete out-of-limit dataset and the partial out-of-limit dataset to obtain the missing label field difference of the label vector set on certain indicators, generate residual labels for the difference label field, and obtain a supplementary label set by correcting the partial error.

[0014] As a preferred technical solution, when correcting candidate synthetic data, the correction module generates lower bound completion data, upper bound completion data, and corrected completion data calculated based on the original data for each original data sample, and then calculates the weighted average of the three to generate corrected synthetic data.

[0015] Beneficial effects: This invention introduces a multi-level control constraint mechanism to achieve dynamic controllability and hierarchical intervention in the synthetic data generation process. Specifically, the designed first-level control constraint sets the value range based on the original label fields, effectively ensuring the structural compliance of the generated data in the basic dimensions. The newly added second-level control constraint combines the discrete evolution relationship between candidate data and residual data to achieve dynamic identification and feedback adjustment of abnormal deviation data, thereby significantly improving the adaptability and generation stability to complex label scenarios. This two-level control mechanism establishes an effective balance between generation accuracy and sample distribution quality, effectively avoiding the problems of strong randomness in generation results and difficulty in backtracking anomalies in existing methods.

[0016] In terms of completion strategy, this invention constructs a hierarchical collaborative completion structure of "precision completion - residual completion". The precision completion generator performs fine-grained correction on some out-of-limit data through a parallel neural network to ensure that local errors are controllable; while the residual completion generator focuses on the structural reconstruction and label reconstruction of completely out-of-limit samples, and further generates the final synthetic data through a weighted fusion of residual data and candidate data. This dual-channel completion process not only improves the filling completeness of data in the edge constraint region, but also enhances the coverage of extreme label space, thereby significantly broadening the label distribution range and application boundaries of the synthetic data.

[0017] This invention introduces a dynamic maintenance mechanism for the tag vector set, supplementary tag set, and second standardized data table in the generation process. This ensures that the tag dimensions maintain consistency and structural closure throughout multiple rounds of generation and correction. Furthermore, it generates a synthetic data generation log containing the entire process at the output stage. This log records information such as tag evolution trajectory, control parameter responses, and generation model calls, enabling auditing, traceability, and post-event verification of the synthetic data generation behavior. This significantly enhances the applicability and credibility of this invention in financial regulation, medical data compliance, and security-sensitive scenarios. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0019] To enhance understanding of the present invention, the present invention will be further described in detail below with reference to embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0020] Example 1 This embodiment provides a synthetic data generation method that integrates AI modeling and a rule engine. It aims to achieve controllable labels, adjustable model accuracy, explainable anomalous sample completion, and traceable synthesis logic during the data generation process. It is particularly suitable for data synthesis scenarios with complex label dimensions, heterogeneous sample structures, and dynamically changing application requirements. The following detailed description of the implementation process, key technical aspects, and parameter settings of the method, along with specific embodiments, further clarifies the technical implications of this invention.

[0021] In this embodiment, the target application scenario is the simulation of smart grid control strategies. The required training data includes rich geographical environmental variables, equipment operation indicators, user behavior tags, and control timeliness parameters. The original data sources include power company databases, historical operation records, and tagged data exported from user behavior simulation platforms. Figure 1 As shown, the entire synthetic data generation process includes the following steps: S1. Construct the tag set: First, structured fields and their meanings are extracted from the original dataset. Based on the requirements of the simulation system for input variables, a label set is constructed. The label fields in the label set include, but are not limited to, multiple dimensions such as: substation geographical location code, dispatch area type, load peak and valley cycle, user behavior model number, dispatch response type, load response threshold, operation execution time, and policy execution delay.

[0022] To enhance the scalability of tags in subsequent steps, all tag fields undergo type labeling (e.g., classification, continuous, timestamp, etc.), logical dependency description (e.g., response type must match user model type), and inter-tag constraint configuration during the construction process, such as "response latency" must not be less than "strategy trigger time," and are stored in the tag field metadata table. This table forms the foundational structure for subsequent standardization and constraint modeling in this invention.

[0023] S2. Establish standardized data tables and control constraints: Based on the aforementioned tag fields and their dependencies, initial primary control constraint indicators are defined. These constraints can take the form of numerical ranges, enumeration ranges, Boolean states, regular expressions, and conditional expressions. Taking the "Response Delay" tag as an example, its control range is set to [0, 600] seconds, and the difference between it and the "Policy Trigger Time" must not exceed the preset maximum response period.

[0024] The aforementioned label fields are then mapped to a standardized field structure in a one-to-one correspondence to construct the first standardized data table. This table structurally describes the field order, data type, logical dependencies, constraints, and default values ​​of the labels, and embeds metadata related to control constraints into the table structure, ensuring that the generated data has structural consistency and constraint verifiability.

[0025] The first standardized data table is defined in the form of JSON Schema during implementation, which facilitates the subsequent data generation engine to perform structure verification and field validation.

[0026] S3. Constructing the label vector set and performing clustering: Standardized raw data is input into the label vector generation module. Each data point is encoded into a high-dimensional vector according to the label field order. Label weight factors are introduced to normalize and stretch different fields to more accurately reflect the differences in label distribution. After the label vectors are generated, a density peak clustering (DPC) algorithm is used to perform unsupervised clustering on the sample set.

[0027] In the specific implementation, the local density and relative distance of each label vector in the vector space are first calculated to identify several high-density centroids as initial clustering cores. Then, based on Euclidean distance, the remaining vectors are assigned to the clusters represented by the corresponding centroids. Each cluster constitutes an initial dataset. This process effectively separates sample subgroups with significant differences in label patterns, providing a class basis for subsequent model training.

[0028] S4. Train the base generator model to generate candidate synthetic data: Each initial dataset is used as input to train an independent deep neural network model as a base generator. The network structure adopts a multilayer perceptron (MLP), with the input being the latent space encoding of the label vectors and the output being the complete set of label fields. The training objective is to minimize the KL divergence between the generated data and the original data at the label level. Each generator is only responsible for the synthesis task of its corresponding initial dataset.

[0029] To improve training performance, a label balancing mechanism is employed during training to ensure that samples with different label values ​​are all learned. Simultaneously, an L2 regularization term is introduced to suppress overfitting, ultimately resulting in a base generator model set. The generator's output data serves as the candidate synthesized data, and its field structure remains consistent with the first standardized data table.

[0030] S5. Execution Control Judgment and Exception Classification: The control and judgment module performs control and judgment operations on the candidate synthetic data generated by the basic generator. The judgment module first calls the control constraint validator to perform legality verification on each label field one by one. If any field does not meet the constraints, it is considered to be out of bounds.

[0031] For all samples that exceed the limits, they are divided into partially exceeding the limits dataset and completely exceeding the limits dataset according to the degree of deviation: if only a few fields are out of bounds and can be corrected by fine-tuning, they are classified as partially exceeding the limits dataset; if multiple fields are seriously conflicting or the range of exceeding the limits is large, they are classified as completely exceeding the limits dataset.

[0032] This step introduces the definition of residual factor: taking the candidate sample as the center, calculate the average deviation between it and the upper and lower bounds of the control, and classify the sample categories based on the set threshold.

[0033] S6. Construct a precision completion model to optimize candidate data: A portion of the out-of-limit dataset was merged with the original dataset to construct a corrected dataset. This corrected dataset was used to train a second set of deep neural network models as a precision completion generator. Its network structure is similar to the basic generator, but a label weight penalty mechanism was added to the loss function to improve its ability to correct key fields.

[0034] After training, samples marked as "partially out of bounds" in the candidate data are corrected and predicted to generate corrected samples. To maintain the balance between the original generated features and the corrected features, the initial output and the completed results are linearly fused, with the fusion coefficients automatically adjusted based on the residual magnitude. This completion process has data hierarchical adjustment capabilities, effectively addressing the issue of bias accumulation caused by mildly violating samples during model generation.

[0035] S7. Generate residual completion data and synthesize the final data: For the data that completely exceeds the limits, perform label vector difference analysis. Difference the labels of the completely exceeding samples and the partially exceeding samples according to field dimensions to extract the set of missing label fields, denoted as the supplementary label set. Based on this set, add control indicator fields to the label fields, and construct a supplementary dataset based on the supplementary labels and the original data.

[0036] This supplementary dataset is used to train the residual completion generator, which uses a residual convolutional network to output residual vectors with a structure consistent with the candidate data. The generated residual data is then weighted and modified by applying weights to the candidate synthetic data to form the final synthetic data. This weighted fusion follows a joint scoring mechanism of field importance and residual credibility to ensure that the modified data is structurally continuous and semantically interpretable.

[0037] The final synthesized data not only meets all control constraints, but also has a reasonable distribution structure and cross-sample consistency, ensuring that it can be used for simulation and training tasks in complex scenarios.

[0038] S8. Output and Log Recording: Finally, all generated synthetic data samples are output to the target format file (such as CSV, Parquet, etc.), and a synthetic data generation log is constructed simultaneously. This log includes information such as model call records for each round, sample cluster identifiers, candidate / corrected / residual sample ratios, control constraint violation statistics, and field correction intensity distribution.

[0039] Log information is stored in structured JSON format for subsequent auditing, data quality analysis, and adjustments to generation strategies. It also supports association between logs and tag metadata tables, forming a complete backtracking chain of the generation process. This mechanism ensures the invention's strong applicability and regulatory transparency in industries with stringent data compliance requirements.

[0040] Example 2 This embodiment provides a synthetic data generation system that integrates AI modeling and a rule engine. It is suitable for high-quality synthesis of structured data in multi-label, multi-constraint scenarios, featuring strong label controllability, high constraint adaptability, good residual repair accuracy, and full-process traceability. The system can be deployed on cloud computing platforms, edge servers, private clusters, or local data centers integrated with business systems. It is compatible with various input / output data formats and business system interfaces, exhibiting good versatility and scalability. The following details the system's structural composition, functional module configuration, and interaction logic between modules in a specific implementation environment.

[0041] like Figure 2 As shown, this system includes the following functional modules: a first acquisition module, a second acquisition module, a clustering module, a training module, a judgment module, a correction module, a supplementary module, and a data output and log recording module. These modules are interconnected via an internal bus or a microservice framework, and can be called and interact with data according to the task flow sequence. The functions and specific implementations of each module are as follows: 1. First Acquisition Module: This module is used to acquire raw data and build a basic tag set. In implementation, this module connects to a database service or file import interface to obtain raw structured data from historical databases, business platforms, or external data acquisition systems, and performs filtering, extraction, and standardization processing on the data fields.

[0042] The construction of the tag set is achieved using a tag modeling engine, including field type identification, business semantic annotation, field normalization conversion, and establishment of logical relationships between tags. For example, in the intelligent transportation scenario, tag fields include vehicle type, road segment number, time period number, congestion level, traffic rule number, event response time, etc.

[0043] The output structure of this module includes: a tag field dictionary, a field type mapping table, a dependency graph between tags, and an initial tag metadata set, providing a unified data entry point for subsequent modules.

[0044] 2. Second Acquisition Module: This module is responsible for generating a set of control constraint indicators and a first standardized data table. Based on the label fields generated by the first acquisition module, this module automatically identifies controllable fields and configures control methods, including: numerical range constraints, enumeration set constraints, conditional expression constraints, and dynamic formula constraints.

[0045] Taking "Event Response Time" as an example, this field is set to a value range of [0, 600] seconds and is associated with "Access Rule Number" to form a composite condition constraint. For example, if the access rule number is "TGR02", the response time must not exceed 300 seconds.

[0046] The set of control constraint indicators is stored in a rule tree structure for easy retrieval later.

[0047] The first standardized data table is output in JSON Schema or Avro format, with field order, type, constraints, and label fields corresponding one-to-one. This standardized structure is used in the deployment environment to verify whether candidate data meets the expected structure and to provide a fixed data path for the neural network input layer.

[0048] 3. Clustering module: This module constructs an initial data subset based on label vectors. After the original data is processed by label vectorization, a sample distribution in a high-dimensional label space is formed. This module uses an improved density peak clustering algorithm to cluster the label space.

[0049] In actual operation, the clustering algorithm first calculates the local density of each sample point and its distance from other high-density points, identifies several cluster centers, and then assigns the remaining points to the nearest center.

[0050] Clustering results are used to divide the input samples into several initial data subsets with similar internal structures, and each subset serves as input for training the generator model. This process reduces the complexity of a single model processing highly heterogeneous data and improves the model's fitting accuracy and generalization ability.

[0051] 4. Training Module: This module is used to train the base generator model. A separate deep neural network generator is trained for each initial subset of data; the network architecture typically uses a multilayer perceptron (MLP) or a variant autoencoder (VAE). The input is a latent representation of the label vectors, and the output is a fully normalized data sample.

[0052] The training process employs cross-entropy loss function, regularization constraints, and label balancing strategies. After each iteration, constraint verification and sample balance checks are performed to ensure that the model output covers all label spaces.

[0053] After training, the basic generator model group is stored in the model management service for dynamic use in subsequent synthesis tasks. This module also provides a model scoring interface, supporting model fusion, version control, and accuracy tracking.

[0054] 5. Judgment Module: This module is responsible for constraint verification and anomaly identification of candidate synthetic data. The system uses this module to call the rule engine to perform field-level control verification on each candidate sample output by the basic generator model.

[0055] Constraint validation includes: whether the labels are within the acceptable range, whether the field combinations are valid, and whether the logical relationships conflict. The system marks candidate data that does not meet the constraints as "out of bounds" and classifies them into "partially out of bounds" and "completely out of bounds" based on the degree of out-of-bounds.

[0056] Partially out of bounds refers to labels that deviate but can be fine-tuned; completely out of bounds refers to those with severe conflicts or missing fields. The module outputs a set of qualified samples, a partially out of bounds dataset, and a completely out of bounds dataset, which are then sent to different modules for processing.

[0057] 6. Correction Module: This module processes some out-of-limit samples, constructs a precision completion generator, and performs corrections. Specific steps include: 1) Constructing a corrected dataset: Merging some out-of-limit data with qualified samples of similar labels as supplementary training samples; 2) Training the completion model: A second set of neural network models is trained based on the corrected dataset, and the offset is predicted using a residual learning mechanism; 3) Data fusion: The original candidate samples and the correction results are linearly weighted and fused. The fusion coefficient is set according to the residual size to obtain smoothed correction data. 4) Output corrected and qualified data, and record each correction path, field change amount and fusion weight to provide data support for subsequent traceability and optimization.

[0058] 7. Supplementary Module: This module performs supplementary generation processing for samples that completely exceed the limits. The main process is as follows: 1) Difference analysis: Perform field-level difference calculations on samples that are completely out of bounds and partially out of bounds, identify missing or conflicting fields, and construct a supplementary label set; 2) Supplementing label modeling: Add missing fields to the original label set, construct a second standardized data table, and form an extended label structure; 3) Construct a supplementary dataset: Cross-merge the supplementary label set with the original data to generate training samples for modeling the residual completion generator; 4) Training the residual completion model: The residual prediction model is trained using a residual convolutional neural network (ResCNN) to output the residual data stream; 5) Fusion to generate final data: The residual data and the original candidate data are fused together at the field level to obtain the final synthetic data with complete structure and satisfying constraints.

[0059] The weighted fusion strategy supports weighting by field importance, weighting by residual stability, and adaptive balancing by model confidence, ensuring that the fused data has business consistency and reasonable distribution.

[0060] 8. Data Output and Log Recording Module: The final generated synthetic data is output by this module to the target directory or data platform interface, supporting formats such as CSV, Parquet, ORC, and JSON. Output samples are stored by batch number, and each sample retains fields such as "generation source identifier," "field correction path," and "fusion weight record."

[0061] The log recording section includes: 1) Records of model calls in each round (including model ID and version number); 2) Data flow path (basic generation → judgment module → correction or supplementation module → output); 3) Control constraint hit rate and statistics of violation fields; 4) Each sample's correction actions, field changes, residual size, and fusion weights, etc.

[0062] Log files are stored in a structured manner, support external audit interface calls, and can be used for operations such as trajectory backtracking, generation process optimization, and model accuracy diagnosis through a visualization platform.

[0063] System operation instructions: In actual operation, this system can be deployed as an independent service, receiving configuration files or API task instructions to complete batch or incremental data synthesis generation processes as needed. The system supports concurrent execution, multi-task scheduling, fault recovery, and log traceability mechanisms, and has good engineering stability and system adaptability.

[0064] For example, in financial risk control scenarios, users can upload customer profile tagging systems and initial compliance data. The system then uses the aforementioned modules to generate simulated transaction records that meet regulatory constraints and multi-tag control requirements, which can be used for AI model training and strategy testing.

[0065] In summary, the synthetic data generation system provided by this invention has the ability to dynamically model labels, a hierarchical management mechanism for control constraints, a collaborative completion mechanism for accuracy and residuals, and a complete log and audit system. It solves the problems of uncontrollable, low-accuracy, and unexplainable anomalies in existing technologies. It can be widely applied to synthetic data generation tasks in multiple industries such as finance, power, medical care, and transportation, and has significant engineering practical value and industrial promotion prospects.

[0066] As can be seen from the above implementation process, the method of this invention not only covers all key aspects of synthetic data generation, but also constructs a structurally sound, functionally closed-loop synthetic data generation system with precision control and multi-source adaptability through a nested standardized label system, control constraint modeling, completion mechanism coordination, and log tracking system. This method demonstrates high versatility, controllability, and interpretability when applied to scenarios involving complex label control and dynamic sample evolution.

[0067] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A synthetic data generation method fusing AI modeling and rule engine, characterized in that, The method comprises the following steps: S1, constructing a label set: collecting original data and constructing a label set corresponding to the target application scenario, the label set comprising a plurality of label fields for characterizing key elements; S2, establishing a standardized data table and control constraints: determining control constraint indicators according to the label fields and their associated relationships, and constructing a first standardized data table based on the label set, the fields of the first standardized data table corresponding one by one to the label fields; S3, constructing a label vector set and clustering division: establishing a first association relationship based on the original data and the first standardized data table to construct a label vector set, and performing a clustering operation on the label vector set to divide it into a plurality of initial data sets; S4, training a base generator model to generate candidate synthetic data: training a deep neural network model based on the initial data set as a base generator for generating candidate synthetic data, each field of the candidate synthetic data corresponding to a field of the first standardized data table; S5, performing control judgment and abnormal division: judging the candidate synthetic data according to the control constraint indicators, if it does not meet the constraint indicators, then through a judgment module, the candidate synthetic data is classified into a partial out-of-limit data set or a complete out-of-limit data set; S6, constructing a precision completion model to optimize the candidate data: combining the partial out-of-limit data set with the original data to construct a revised data set, and training a second deep neural network model based on the revised data set as a precision completion generator to revise and optimize the candidate synthetic data; S7, generating residual completion data and synthesizing final data: constructing a supplementary label data set based on the difference value of the out-of-limit data set, training a residual completion model to generate residual data, and weighting and fusing the candidate data to obtain final synthetic data; S8, output and log recording: outputting the final synthetic data and the second standardized data table, and constructing a synthetic data generation log. 2.The synthetic data generation method of fusing AI modeling and rule engine according to claim 1, wherein, The label fields in step S1 include at least geographic information labels, time labels, device information labels, user behavior labels, and time limit control labels for measuring the timeliness of synthetic data generation. 3.The synthetic data generation method of fusing AI modeling and rule engine according to claim 1, characterized in that, In step S3, the clustering operation of the label vector set comprises: calculating the distribution density of each label vector in the original data, and dividing the label vector set into high-density and low-density subsets to form the initial data set. 4.The synthetic data generation method of fusing AI modeling and rule engine according to claim 1, wherein, In step S5: The control constraint indicators include a first-level control constraint constructed based on the label fields and a second-level control constraint generated based on the supplementary labels; The first-level control constraint includes parameters for specifying the upper limit, lower limit or interval range of the label field values; The second-level control constraint is determined by comparing the dispersion of the residual data and the candidate synthetic data in each label field. 5.The synthetic data generation method of fusing AI modeling and rule engine according to claim 1, characterized in that, In step S6, the precision completion generator is composed of an initial neural network model and a completion neural network model constructed based on the partial out-of-limit data set in parallel, and outputs the revised data after fusion.

6. The synthetic data generation method of fusing AI modeling and rule engine according to claim 1, characterized in that, Step S7 specifically includes: Performing label vector difference operation on the complete out-of-limit data set and the partial out-of-limit data set to obtain a supplementary label set; A second standardized data table is constructed by adding control constraint indicators to the label field according to the supplementary label set, and a supplementary data set is constructed by using the supplementary label set and the original data; A third deep neural network model is trained by using the supplementary data set, as a residual complement generator to generate residual data; The residual data is used to supplement the part of the synthetic data, and the residual data and the candidate synthetic data are weighted and averaged to obtain corrected synthetic data as the final synthetic data.

7. A synthetic data generation system fusing AI modeling with a rules engine for implementing a synthetic data generation method fusing AI modeling with a rules engine according to any one of claims 1 - 6, characterized in that, Comprise: The first acquisition module is configured to collect original data and construct a label set corresponding to a target application scenario, the label set including a label field used to describe key elements; The second acquisition module is configured to determine control constraint indicators according to the label field and construct a first standardized data table, the fields of the first standardized data table corresponding to the label field one by one; The clustering module is configured to divide the original data into a plurality of data subsets based on a label vector set; The training module is configured to train a deep neural network model based on the data subsets, as a base generator to generate candidate synthetic data; The judgment module is configured to judge the candidate synthetic data according to the control constraint indicators, and if the candidate synthetic data does not meet the constraint indicators, the candidate synthetic data is classified into a part of the out-of-limit data set or a complete out-of-limit data set; The correction module is configured to train a precision complement generator based on the part of the out-of-limit data set and correct the bias in the candidate synthetic data; The supplement module is configured to train a residual complement generator based on the complete out-of-limit data set and generate residual data used to supplement the part of the synthetic data, and the residual data and the candidate synthetic data are weighted and averaged to obtain final synthetic data.

8. The synthetic data generation system that fuses AI modeling with a rules engine of claim 7, wherein, The first acquisition module is further configured to extract original data from a database and perform a data cleaning operation to remove missing items and outliers.

9. The synthetic data generation system that fuses AI modeling with a rules engine of claim 7, wherein, The supplement module is further configured to perform a label difference operation between the complete out-of-limit data set and the part of the out-of-limit data set to obtain a missing label field difference of the label vector set in some indicators, generate a residual label for the difference label field, and obtain a supplementary label set by correcting the part of the error.

10. The synthetic data generation system that fuses AI modeling with rules engine of claim 7, wherein, The correction module generates lower bound complement data, upper bound complement data, and correction complement data calculated based on the original data for each original data sample when correcting the candidate synthetic data, and generates corrected synthetic data by weighted averaging the three.

Images