A synthetic data generation method and system fusing AI modeling and rule engine

Abstract

The application discloses a kind of synthetic data generation method and system of fusion AI modeling and rule engine, it is related to data processing technical field, including following steps: S1, obtain sample data and business rule set and pre-process, construct sample state sequence and task condition sequence;S2, to sample state sequence is extended processing, obtain noise state sequence, and establish condition mapping relationship;S3, parse business rule set, construct rule constraint chain after being modeled and conflict resolution by dependence;S4, based on noise state sequence and condition mapping relationship executes reverse prediction, and constructs constraint correction state;S5, according to constraint correction state iterative generation synthetic data set;S6, check synthetic data set and adjust generation parameter and rule parameter.The application utilizes improved type condition diffusion model etc., has the advantages such as good data distribution retention, strong attribute association consistency and high rule compliance.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for generating synthetic data that integrates AI modeling and rule engines. Background Technology

[0002] With the increasing demand for AI model training, software testing, and data analysis, real-world data faces challenges such as insufficient sample size, uneven data distribution, and privacy leaks during acquisition and use. Synthetic data generation technology learns the data distribution patterns and attribute relationships of sample data to generate data for model training and business testing, reducing reliance on real-world data.

[0003] Existing synthetic data generation methods mainly include those based on statistical distributions, pre-defined rules, and deep learning models. Statistical distribution-based methods typically sample each data attribute separately, making it difficult to maintain the complex relationships between different attributes. Pre-defined rule-based methods rely on manually configured rules, making it difficult to learn the underlying distribution patterns in real data, and are prone to problems such as unclear rule dependencies and rule conflicts when there are many rules.

[0004] Data generation methods based on deep learning models can learn the distribution characteristics of sample data. Among them, diffusion models generate synthetic data through forward noise injection and reverse denoising. However, existing diffusion models usually adopt Markov diffusion, which lacks noise correlation between diffusion levels and easily weakens the long-term dependencies between sample states. At the same time, the lack of precise correspondence between task conditions and noise states affects the condition guidance effect in the reverse denoising process.

[0005] Furthermore, existing technologies typically filter synthesized data through a rule engine after generation. Rule constraints cannot directly participate in the reverse denoising process, resulting in data that does not conform to the rules being either deleted or regenerated. For business rules with dependencies and conflicts, existing rule engines also struggle to establish orderly constraint propagation relationships, thus affecting the accuracy of attribute associations and the degree of rule compliance in the synthesized data.

[0006] Therefore, how to provide a synthetic data generation method and system that integrates AI modeling and rule engine is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose a synthetic data generation method and system that integrates AI modeling and rule engine. This invention fully utilizes an improved conditional diffusion model and rule engine technology, learns the distribution patterns and state relationships of sample data through non-Markov diffusion processing, and embeds the rule constraint chain formed by dependency modeling and conflict resolution into the reverse denoising process. The denoising prediction results are subjected to step-by-step constraint matching and constraint-guided correction to achieve synthetic data generation that meets task conditions and business rules. It has the advantages of good data distribution preservation, strong attribute correlation consistency, high rule compliance, and controllable generation process.

[0008] A synthetic data generation method integrating AI modeling and rule engine according to an embodiment of the present invention includes the following steps: S1. Obtain the sample data and business rule set corresponding to the synthetic data generation task, and preprocess them to construct the sample state sequence and task condition sequence. S2. Construct an improved conditional diffusion model, perform non-Markov diffusion processing on the sample state sequence to obtain a noisy state sequence, determine the target state based on the task condition sequence, and establish a conditional mapping relationship between the noisy state sequence and each target state. S3. The business rule set is parsed through the rule engine, the rule constraint state corresponding to each business rule is extracted, and dependency modeling and conflict resolution operations are performed. Based on the task condition sequence, the rule constraint state corresponding to each target state is selected from the conflict resolution results to construct the rule constraint chain. S4. In the improved conditional diffusion model, an inverse denoising prediction operation is performed based on the mapping relationship between the noise state sequence and the condition. The denoising prediction result is subjected to step-by-step constraint matching along the rule constraint chain to construct the constraint correction state. S5. Perform constraint-guided correction on the denoising prediction results according to the constraint correction state, and iteratively perform inverse denoising prediction, step-by-step constraint matching and constraint-guided correction operations on the correction results until the preset termination condition is met, and generate a synthetic dataset. S6. Verify the synthesized dataset and adjust the parameters of the improved conditional diffusion model and rule engine based on the verification results.

[0009] Optionally, the sample data refers to the raw data obtained from real data sources for learning data distribution patterns and attribute relationships; the business rule set refers to a set of rules pre-configured for data attribute values, attribute relationships, state transition relationships, and data generation conditions; the preprocessing refers to performing data cleaning, missing state correction, abnormal state correction, data encoding, and numerical normalization on the sample data; the improved conditional diffusion model refers to a deep learning model used to perform conditional synthetic data generation; and the rule engine refers to a rule processing unit used to enforce rule constraints on the synthetic data generation process.

[0010] Optionally, S2 specifically includes: S21. Construct an improved conditional diffusion model, which includes a conditional diffusion unit and an inverse denoising unit; S22. In the conditional diffusion unit, non-Markov noise injection processing is performed on the sample state sequence according to the preset diffusion level to iteratively generate a noise state sequence. S23. Determine the target state corresponding to each task condition based on the task condition sequence, and perform state alignment processing on the noisy state sequence and each target state to establish a condition mapping relationship, specifically including: The generated objects, attribute constraints, and state ranges corresponding to each task condition in the task condition sequence are analyzed, and task conditions with the same generated objects are associated to form a set of condition states. The conditional state set is divided according to attribute constraints and state range, and the state differences between each conditional state are calculated to determine the target state corresponding to each task condition. The distribution features of each noise state and the condition features of each target state in the noise state sequence are extracted respectively. The state matching value between each noise state and each target state is calculated in turn, and a state matching matrix is ​​constructed. The Sinkhorn algorithm is used to perform alternating normalization on the state matching matrix so that the state matching values ​​corresponding to each noise state and the target state satisfy the preset allocation constraints, thus obtaining the state alignment matrix. The target state and alignment weight corresponding to each noise state are determined based on the state alignment matrix, and correlation processing is performed to establish a conditional mapping relationship.

[0011] Optionally, S22 specifically includes: S221. Calculate the state correlation value between each sample state in the sample state sequence, and combine it with the positional interval between each sample state to form a state correlation sequence. S222. Map the state association sequence to a preset diffusion level, determine the noise memory parameter using the state association value corresponding to each diffusion level, and calculate the noise covariance between different diffusion levels using the noise memory parameter to construct a covariance sequence. S223. Using the Davies-Harte algorithm, the covariance sequence is expanded into a cyclic embedding sequence. The cyclic embedding sequence is then subjected to a fast Fourier transform to obtain the eigenvalues ​​corresponding to each frequency position. S224. Initialize multiple sets of independent random number pairs that follow a standard normal distribution. Use each set of random number pairs as the real and imaginary parts to form complex random quantities, and make the complex random quantities located at symmetrical frequency positions conjugate to each other. S225. Multiply the square root of the characteristic value corresponding to each frequency position by the corresponding complex random quantity, and perform the inverse fast Fourier transform operation to generate the relevant noise corresponding to each preset diffusion level. S226. Take the sample state sequence as the diffusion state of the first preset diffusion level, and set the corresponding state retention coefficient and noise injection coefficient for each preset diffusion level. Multiply the current diffusion state by the state retention coefficient, and multiply the corresponding related noise by the noise injection coefficient. Add the two multiplication results to obtain the current noise state, and use the current noise state as the diffusion state of the next preset diffusion level. After completing the calculations for all preset diffusion levels, the non-Markov diffusion process ends, and the noise state sequence is obtained.

[0012] Optionally, S3 specifically includes: S31. The business rule set is parsed one by one through the rule engine to extract the rule conditions, constraint attributes, constraint status and execution order corresponding to each business rule, and the status is encoded to generate the rule constraint status corresponding to each business rule. S32. Based on the rule conditions, constraint attributes and constraint states, perform dependency modeling operations on each rule constraint state to extract the dependency relationships between each rule constraint state. S33. Identify the constraint states of rules with the same constraint attributes but inconsistent constraint states, and compare and determine them according to the corresponding dependencies and execution order. Prioritize retaining the constraint states of rules that are in the earlier dependency positions. When the dependency positions are the same, retain the constraint states of rules that are executed earlier, and complete the conflict resolution process. Wherein, dependency position represents the position of each rule constraint state in the directed execution order formed by the dependency relationship; S34. Based on the sequence of task conditions, match each target state with the rule constraint state after conflict resolution, and filter the rule constraint state corresponding to each target state. S35. Arrange the selected rule constraint states according to the dependencies and execution order between each rule constraint state, establish the constraint transmission relationship between adjacent rule constraint states, and construct the rule constraint chain.

[0013] Optionally, S4 specifically includes: S41. In the reverse denoising unit of the improved conditional diffusion model, the current noise state is extracted sequentially from the noise state sequence according to the reverse order of the preset diffusion level, and the target state and alignment weight corresponding to the current noise state are determined according to the conditional mapping relationship. S42. Perform conditional fusion processing on the current noise state and the target state according to the alignment weight; S43. Perform noise prediction and state restoration processing on the conditional fusion results to obtain denoised prediction results; S44. Extract the rule constraint state corresponding to the target state from the rule constraint chain, and perform step-by-step constraint matching processing on the denoising prediction result according to the constraint propagation relationship; S45. Based on the constraint matching results, locate the state positions in the denoising prediction results that do not meet the rule conditions, and determine the state deviation direction and state deviation value corresponding to each state position. S46. Based on the direction and degree of state deviation, calculate the correction direction and correction value of the corresponding state position, and construct the corresponding constraint correction state.

[0014] Optionally, S43 specifically includes: S431. In the reverse denoising unit, noise residual prediction processing is performed on the conditional fusion result to obtain the noise prediction state corresponding to the current diffusion level. S432. Based on the current noise state, the noise prediction state, and the state retention coefficient and noise injection coefficient corresponding to the current diffusion level, the current data prediction state is obtained by inversion. S433. Extract the predicted state of historical data from the historical diffusion level that has completed the reverse denoising process, and determine the reverse solution step size based on the state retention coefficient and noise injection coefficient corresponding to the current pre-diffusion level and the historical diffusion level. S434. The DPM-Solver++ algorithm is used to calculate the multi-step correction state based on the current data prediction state, the historical data prediction state, and the reverse solution step size. S435. Perform reverse state update based on the current noise state, the current data prediction state, and the multi-step correction state to obtain the denoising prediction result and record the corresponding data prediction state.

[0015] Optionally, S5 specifically includes: S51. Based on the constraint correction state, the state values ​​at the corresponding state positions in the denoising prediction results are constrained and guided to be corrected to obtain the correction results. S52. Use the correction result as the input state for the next inverse denoising level, perform the inverse denoising prediction operation, and obtain a new denoising prediction result. S53. Perform stepwise constraint matching along the rule constraint chain on the new denoising prediction result, perform constraint-guided correction based on the constraint matching result, and iteratively execute the reverse denoising prediction, stepwise constraint matching and constraint-guided correction operations. S54. When the reverse denoising process of all preset diffusion levels is completed, the current correction result satisfies all rule constraint states in the rule constraint chain, and the state change between two adjacent constraint-guided correction results is less than the preset convergence threshold, the preset termination condition is satisfied, the iteration is stopped, and a synthetic dataset is generated based on the current correction result.

[0016] Optionally, S6 specifically includes: S61. Verification personnel manually verify the data distribution, attribute relationships, and rule compliance of the synthetic dataset, and generate verification results. S62. Determine the data distribution deviation, attribute association deviation, and rule constraint deviation corresponding to the synthetic dataset based on the verification results; S63. Adjust the generation parameters of the improved conditional diffusion model based on the data distribution deviation and attribute association deviation, and adjust the rule execution parameters of the rule engine based on the rule constraint deviation.

[0017] A synthetic data generation system integrating AI modeling and rule engine according to an embodiment of the present invention includes: The data acquisition module is used to acquire sample data and business rule sets corresponding to the synthetic data generation task, and to preprocess them to construct sample state sequences and task condition sequences. The conditional diffusion module is used to construct an improved conditional diffusion model. It performs non-Markov diffusion processing on the sample state sequence to obtain a noisy state sequence, determines the target state based on the task condition sequence, and establishes a conditional mapping relationship between the noisy state sequence and each target state. The rule constraint module is used to parse the business rule set through the rule engine, extract the rule constraint state corresponding to each business rule, and perform dependency modeling and conflict resolution operations. Based on the task condition sequence, it selects the rule constraint state corresponding to each target state from the conflict resolution results and constructs the rule constraint chain. The inverse prediction module is used in the improved conditional diffusion model to perform inverse denoising prediction based on the mapping relationship between the noise state sequence and the condition. It performs step-by-step constraint matching processing on the denoising prediction results along the rule constraint chain to construct the constraint correction state. The iterative synthesis module is used to perform constraint-guided correction on the denoised prediction results according to the constraint correction state, and iteratively perform inverse denoising prediction, step-by-step constraint matching and constraint-guided correction operations on the correction results until the preset termination condition is met, and generate a synthetic dataset. The verification and adjustment module is used to verify the synthetic dataset and adjust the parameters of the improved conditional diffusion model and rule engine based on the verification results.

[0018] The beneficial effects of this invention are: First, this invention performs non-Markov diffusion processing through an improved conditional diffusion model and establishes a conditional mapping relationship between noise states and target states, effectively preserving the distribution patterns and attribute association characteristics of sample data and improving the data distribution consistency of synthetic data.

[0019] Secondly, this invention uses a rule engine to perform dependency modeling and conflict resolution on business rules, constructing a rule constraint chain corresponding to the target state, so that different business rules can participate in data generation in a determined order, thereby improving the orderliness of rule execution and the rule compliance of the synthesized data.

[0020] Finally, this invention embeds the rule constraint chain into the reverse denoising process, performs step-by-step constraint matching and constraint-guided correction on the denoising prediction results, and adjusts the model parameters and rule execution parameters according to the verification results, thereby improving the generation quality, stability and controllability of the synthetic data. Attached Figure Description

[0021] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a synthetic data generation method that integrates AI modeling and rule engine proposed in this invention; Figure 2 This is a non-Markov diffusion flowchart of a synthetic data generation method that integrates AI modeling and rule engine proposed in this invention. Figure 3 This is a module structure diagram of a synthetic data generation system that integrates AI modeling and rule engine proposed in this invention. Detailed Implementation

[0022] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0023] refer to Figures 1-2 A synthetic data generation method integrating AI modeling and rule engine includes the following steps: S1. Obtain the sample data and business rule set corresponding to the synthetic data generation task, and preprocess them to construct the sample state sequence and task condition sequence. S2. Construct an improved conditional diffusion model, perform non-Markov diffusion processing on the sample state sequence to obtain a noisy state sequence, determine the target state based on the task condition sequence, and establish a conditional mapping relationship between the noisy state sequence and each target state. S3. The business rule set is parsed through the rule engine, the rule constraint state corresponding to each business rule is extracted, and dependency modeling and conflict resolution operations are performed. Based on the task condition sequence, the rule constraint state corresponding to each target state is selected from the conflict resolution results to construct the rule constraint chain. S4. In the improved conditional diffusion model, an inverse denoising prediction operation is performed based on the mapping relationship between the noise state sequence and the condition. The denoising prediction result is subjected to step-by-step constraint matching along the rule constraint chain to construct the constraint correction state. S5. Perform constraint-guided correction on the denoising prediction results according to the constraint correction state, and iteratively perform inverse denoising prediction, step-by-step constraint matching and constraint-guided correction operations on the correction results until the preset termination condition is met, and generate a synthetic dataset. S6. Verify the synthesized dataset and adjust the parameters of the improved conditional diffusion model and rule engine based on the verification results.

[0024] In this embodiment, sample data refers to the raw data obtained from real data sources for learning data distribution patterns and attribute relationships. The business rule set refers to the set of rules pre-configured for data attribute values, attribute relationships, state transition relationships, and data generation conditions. Preprocessing refers to performing data cleaning, missing state correction, abnormal state correction, data encoding, and numerical normalization on the sample data. The improved conditional diffusion model refers to the deep learning model used to perform conditional synthetic data generation. The rule engine refers to the rule processing unit used to enforce rule constraints on the synthetic data generation process.

[0025] In this embodiment, S2 specifically includes: S21. Construct an improved conditional diffusion model, which includes a conditional diffusion unit and an inverse denoising unit; S22. In the conditional diffusion unit, non-Markov noise injection processing is performed on the sample state sequence according to the preset diffusion level to iteratively generate a noise state sequence. S23. Determine the target state corresponding to each task condition based on the task condition sequence, and perform state alignment processing on the noisy state sequence and each target state to establish a condition mapping relationship, specifically including: The generation object, attribute constraints and state range corresponding to each task condition in the task condition sequence are analyzed. Task conditions with the same generation object are associated to form a set of condition states. The discrete values ​​in the attribute constraints are encoded as state categories, and the continuous values ​​in the state range are mapped to the standardized interval of 0 to 1. The condition state set is divided according to attribute constraints and state range, and the state difference between each condition state is calculated. When the standardized distance between two condition states is less than 0.15, the corresponding condition states are assigned to the same target state interval. When the standardized distance between two condition states is not less than 0.15, the corresponding condition states are assigned to different target state intervals to determine the target state corresponding to each task condition. The distribution features of each noise state and the condition features of each target state in the noise state sequence are extracted respectively. The distribution mean, state variance, attribute correlation strength and condition constraint distance are used as matching criteria to calculate the state matching value between each noise state and each target state in turn, and construct the state matching matrix. The Sinkhorn algorithm is used to perform alternating normalization on the state matching matrix. During the alternating iteration of row normalization and column normalization, the sum of the elements in each row of the state matching matrix is ​​made equal to 1, the sum of the elements in each column is made equal to 1, and the total matching weights corresponding to each noise state and the total allocation weights corresponding to each target state satisfy the preset balanced allocation constraints. At the same time, it is ensured that any state matching value is between 0 and 1. The normalization process stops when the matrix change between two adjacent normalization results is less than the preset change threshold, and the state alignment matrix is ​​obtained. Based on the state alignment matrix, determine the target state and alignment weight corresponding to each noise state. Establish a primary mapping relationship between noise states with alignment weights greater than 0.6 and their corresponding target states. Establish an auxiliary mapping relationship between noise states with alignment weights no greater than 0.6 and greater than 0.3 and their corresponding target states. Then, perform association processing on the primary mapping relationship and the auxiliary mapping relationship to establish a conditional mapping relationship.

[0026] In this embodiment, S22 specifically includes: S221. Calculate the state correlation value between each sample state in the sample state sequence, and combine it with the positional interval between each sample state to form a state correlation sequence. The state differences between sample states are normalized to 0 to 1. When the state difference decreases, the corresponding state association value is increased, and when the state difference increases, the corresponding state association value is decreased. The state association values ​​at different position intervals are corrected according to the interval decay coefficient of 0.86, so that sample states that are close to each other and have similar states maintain a high association strength in the state association sequence. S222. Map the state association sequence to 64 preset diffusion levels, use the state association values ​​corresponding to each diffusion level to determine the noise memory parameters, and use the noise memory parameters to calculate the noise covariance between different diffusion levels to construct a covariance sequence. The noise memory parameter is set between 0.35 and 0.92 based on the state correlation value. When the state correlation value corresponding to the diffusion level increases, the noise memory parameter of the corresponding diffusion level is increased to maintain strong continuity of the noise covariance between adjacent diffusion levels. When the state correlation value corresponding to the diffusion level decreases, the noise memory parameter of the corresponding diffusion level is decreased to increase the noise disturbance intensity of the corresponding diffusion level. S223. Using the Davies-Harte algorithm, the covariance sequence is expanded into a cyclic embedding sequence. The first part of the cyclic embedding sequence is arranged according to the original order of the covariance sequence, and the second part of the cyclic embedding sequence is completed according to the reverse decay order of the covariance sequence. The cyclic embedding sequence is subjected to a fast Fourier transform to obtain the eigenvalues ​​corresponding to each frequency position, and the eigenvalues ​​less than 0 are corrected to 0. S224. Initialize multiple sets of independent random number pairs that follow a standard normal distribution. Use each set of random number pairs as the real and imaginary parts to form complex random quantities. Make the complex random quantities located at symmetrical frequency positions conjugate to each other. After the frequency domain random disturbance is passed through the subsequent inverse fast Fourier transform, it becomes real noise. S225. Multiply the square root of the characteristic value corresponding to each frequency position with the corresponding complex random quantity, and perform inverse fast Fourier transform operation to generate the relevant noise corresponding to each preset diffusion level, so that the noise change between adjacent diffusion levels satisfies the correlation relationship defined by the covariance sequence, and so that a continuous noise evolution process is formed between the 1st diffusion level and the 64th diffusion level. S226. The sample state sequence is taken as the diffusion state of the first preset diffusion level. The corresponding state retention coefficient and noise injection coefficient are set at each preset diffusion level. The state retention coefficient decreases from 0.98 to 0.12 as the diffusion level increases, and the noise injection coefficient increases from 0.02 to 0.88 as the diffusion level increases. The current diffusion state is multiplied by the state retention coefficient, and the corresponding related noise is multiplied by the noise injection coefficient. The two multiplication results are added together to obtain the current noise state. The current noise state is then used as the diffusion state of the next preset diffusion level. This allows the diffusion state to gradually weaken the proportion of the original sample state and gradually increase the proportion of related noise as the diffusion level increases. After completing the calculations for all preset diffusion levels, the non-Markov diffusion process ends, and the noise state sequence is obtained.

[0027] In this embodiment, S3 specifically includes: S31. The business rule set is parsed one by one through the rule engine to extract the rule conditions, constraint attributes, constraint status and execution order corresponding to each business rule, and the status is encoded to generate the rule constraint status corresponding to each business rule. During status encoding, rule conditions, constraint attributes, and constraint states are uniformly mapped to an encoding range of 0 to 1, and each rule constraint state is sequentially marked according to the execution order, so that each business rule maintains the corresponding rule conditions, constraint attributes, constraint states, and execution order in the rule constraint states; S32. Based on the rule conditions, constraint attributes and constraint states, perform dependency modeling operations on each rule constraint state to extract the dependency relationships between each rule constraint state. When the constraint state corresponding to one rule constraint state satisfies the rule condition corresponding to another rule constraint state, and the correlation degree between the constraint attributes corresponding to the two rule constraint states is not less than 0.72, the two rule constraint states are determined to have a dependency relationship. S33. Identify the constraint states of rules with the same constraint attributes but inconsistent constraint states, and compare and determine them according to the corresponding dependencies and execution order. Prioritize retaining the constraint states of rules that are in the earlier dependency positions. When the dependency positions are the same, retain the constraint states of rules that are executed earlier, and complete the conflict resolution process. Dependency position indicates the position of each rule constraint state in the directed execution order formed by the dependency relationship. When the difference between the constraint states of two rule constraint states with the same constraint attributes is not less than 0.18, the constraint states of the two rule constraint states are determined to be inconsistent. S34. Based on the sequence of task conditions, match each target state with the rule constraint state after conflict resolution, and filter the rule constraint state corresponding to each target state. During matching, the sequence of task conditions corresponding to the target state is compared with the rule conditions, constraint attributes and constraint states corresponding to the rule constraint states. When the matching degree is not less than 0.65, the corresponding rule constraint states are selected as the rule constraint states corresponding to the target state. S35. Arrange the selected rule constraint states according to the dependency relationship and execution order between each rule constraint state, establish the constraint transit relationship between adjacent rule constraint states, and construct the rule constraint chain. During the arrangement, the preceding dependency positions of the rule constraint states are first determined according to the dependency relationship. Then, the rule constraint states at the same dependency position are arranged according to the execution order. The constraint states between adjacent rule constraint states are associated with the rule conditions to form a constraint transit relationship, thus obtaining a rule constraint chain.

[0028] In this embodiment, S4 specifically includes: S41. In the reverse denoising unit of the improved conditional diffusion model, the current noise state is extracted from the noise state sequence in reverse order of the 64 preset diffusion levels, and the target state and alignment weight corresponding to the current noise state are determined according to the conditional mapping relationship, so that the 64th level noise state is reversed to the 1st level noise state in sequence. S42. Perform conditional fusion processing on the current noise state and the target state according to the alignment weight. When the alignment weight is greater than 0.6, increase the participation ratio of the target state in the conditional fusion processing. When the alignment weight is not greater than 0.6 but greater than 0.3, retain the dominant ratio of the current noise state in the conditional fusion processing, so that the conditional fusion result maintains both the distribution characteristics of the current noise state and the generation constraints of the target state. S43. Perform noise prediction and state restoration processing on the conditional fusion results to obtain denoised prediction results; S44. Extract the rule constraint state corresponding to the target state from the rule constraint chain, and perform step-by-step constraint matching processing on the denoising prediction result according to the constraint propagation relationship; In the stepwise constraint matching process, the denoising prediction result is judged in turn according to the order of the rule constraint chain to determine whether it meets the rule conditions, constraint attributes and constraint states in the corresponding rule constraint state, and the matching result of the previous rule constraint state is passed to the next rule constraint state. S45. Based on the constraint matching results, locate the state positions in the denoising prediction results that do not meet the rule conditions, and determine the state deviation direction and state deviation value corresponding to each state position. When the state value corresponding to the state position is lower than the range of values ​​limited by the rule constraint state, the state deviation direction is determined as the positive correction direction. When the state value corresponding to the state position is higher than the range of values ​​limited by the rule constraint state, the state deviation direction is determined as the negative correction direction. The state deviation value is determined based on the difference between the state value corresponding to the state position and the rule constraint state. S46. Based on the direction and degree of state deviation, calculate the correction direction and correction value of the corresponding state position, and construct the corresponding constraint correction state.

[0029] In this embodiment, S43 specifically includes: S431. In the reverse denoising unit, noise residual prediction processing is performed on the conditional fusion result to obtain the noise prediction state corresponding to the current diffusion level. When processing is performed in reverse order according to the 64 preset diffusion levels, the noise residual prediction processing starts from the conditional fusion result corresponding to the 64th diffusion level and progresses to the 1st diffusion level step by step with the reverse denoising process, so that each current diffusion level generates a corresponding noise prediction state. S432. Based on the current noise state, the noise prediction state, and the state retention coefficient and noise injection coefficient corresponding to the current diffusion level, the current data prediction state is obtained by inversion. During the inversion process, the retained state components in the current noise state are reduced according to the state retention coefficient corresponding to the current diffusion level, and the noise components in the noise prediction state are canceled according to the noise injection coefficient corresponding to the current diffusion level, so that the current data prediction state gradually approaches the data recovery state corresponding to the current noise state at that diffusion level. S433. Extract the predicted state of historical data from the historical diffusion level that has completed the reverse denoising process, and determine the reverse solution step size based on the state retention coefficient and noise injection coefficient corresponding to the current diffusion level and the historical diffusion level. When the number of historical diffusion levels that have completed reverse denoising is 1, the reverse solution step size is determined by the current data prediction state and one historical data prediction state. When the number of historical diffusion levels that have completed reverse denoising is greater than 1, the reverse solution step size is determined by combining the current data prediction state and multiple historical data prediction states, and the reverse solution step size is constrained based on the coefficient change amplitude between adjacent diffusion levels. S434. The DPM-Solver++ algorithm is used to calculate the multi-step correction state based on the current data prediction state, the historical data prediction state, and the reverse solution step size. During the calculation process, the update direction of the multi-step correction state is determined based on the changing trend between the current data prediction state and the historical data prediction state, and the update magnitude of the multi-step correction state is controlled based on the reverse solution step size, so that the multi-step correction state can simultaneously retain the data recovery results of the current diffusion level and the data recovery trend of the historical diffusion level. S435. Perform reverse state update based on the current noise state, the current data prediction state, and the multi-step correction state to obtain the denoising prediction result and record the corresponding data prediction state. During the reverse state update process, the current data prediction state is used as the basic recovery result, the multi-step correction state is used as the state update compensation, and the hierarchical continuity of the reverse denoising process is maintained by combining the current noise state. The denoising prediction result corresponding to the current diffusion level is generated, and the corresponding data prediction state is recorded as the historical data prediction state in the subsequent historical diffusion levels.

[0030] In this embodiment, S5 specifically includes: S51. Based on the constraint correction state, the state value at the corresponding state position in the denoising prediction result is constrained and guided to be corrected. The state value to be corrected is locked according to the state position recorded by the constraint correction state. The adjustment direction of the corresponding state value is determined according to the correction direction. The adjustment range of the corresponding state value is determined according to the correction value. The corrected state value is brought closer to the state range limited by the rule constraint state, and the correction result is obtained. S52. Use the correction result as the input state for the next inverse denoising level, perform the inverse denoising prediction operation, and obtain a new denoising prediction result. During the reverse processing of the 64-level preset diffusion layers, after each reverse denoising layer completes the constraint-guided correction, the correction result formed at that layer is passed to the next reverse denoising layer, so that the subsequent reverse denoising prediction continues to be executed based on the state of the corrected rule deviation. S53. Perform stepwise constraint matching along the rule constraint chain on the new denoising prediction result, perform constraint-guided correction based on the constraint matching result, and iteratively execute the reverse denoising prediction, stepwise constraint matching and constraint-guided correction operations. S54. When the reverse denoising process of all preset diffusion levels is completed, the current correction result satisfies all rule constraint states in the rule constraint chain, and the state change between two adjacent constraint-guided correction results is less than the preset convergence threshold, the preset termination condition is satisfied, the iteration is stopped, and a synthetic dataset is generated based on the current correction result.

[0031] In this embodiment, S6 specifically includes: S61. Verification personnel manually verify the data distribution, attribute relationships, and rule compliance of the synthetic dataset, and generate verification results. S62. Determine the data distribution deviation, attribute association deviation, and rule constraint deviation corresponding to the synthetic dataset based on the verification results; S63. Adjust the generation parameters of the improved conditional diffusion model based on the data distribution deviation and attribute association deviation, and adjust the rule execution parameters of the rule engine based on the rule constraint deviation.

[0032] refer to Figure 3 A synthetic data generation system integrating AI modeling and rule engine, comprising: The data acquisition module is used to acquire sample data and business rule sets corresponding to the synthetic data generation task, and to preprocess them to construct sample state sequences and task condition sequences. The conditional diffusion module is used to construct an improved conditional diffusion model. It performs non-Markov diffusion processing on the sample state sequence to obtain a noisy state sequence, determines the target state based on the task condition sequence, and establishes a conditional mapping relationship between the noisy state sequence and each target state. The rule constraint module is used to parse the business rule set through the rule engine, extract the rule constraint state corresponding to each business rule, and perform dependency modeling and conflict resolution operations. Based on the task condition sequence, it selects the rule constraint state corresponding to each target state from the conflict resolution results and constructs the rule constraint chain. The inverse prediction module is used in the improved conditional diffusion model to perform inverse denoising prediction based on the mapping relationship between the noise state sequence and the condition. It performs step-by-step constraint matching processing on the denoising prediction results along the rule constraint chain to construct the constraint correction state. The iterative synthesis module is used to perform constraint-guided correction on the denoised prediction results according to the constraint correction state, and iteratively perform inverse denoising prediction, step-by-step constraint matching and constraint-guided correction operations on the correction results until the preset termination condition is met, and generate a synthetic dataset. The verification and adjustment module is used to verify the synthetic dataset and adjust the parameters of the improved conditional diffusion model and rule engine based on the verification results.

[0033] Example 1: To verify the feasibility of this invention in implementation, it was applied to a comprehensive business data management scenario. This scenario includes 38 business terminals, 11 business processing nodes, and 6 types of data objects. During daily operation, records are generated for account status, terminal status, business applications, processing progress, cost changes, and anomaly handling. Since low-frequency anomalies occur infrequently, existing real data cannot cover all test conditions. Furthermore, real data contains sensitive attributes and cannot be directly used for model training and business function verification. Therefore, it is necessary to generate synthetic data that combines data distribution realism, attribute correlation consistency, and business rule compliance.

[0034] This embodiment selects 24,680 historical business data entries as sample data, which contains 64 data attributes. After preprocessing, 312 missing status corrections and 87 abnormal status corrections are completed. Category attributes are encoded, and continuous attributes are numerically normalized to form a sample status sequence. Based on data usage requirements, five types of task conditions are formed, and 16 target statuses are further determined, including normal acceptance, status suspension, processing timeout, terminal offline, fee abnormality, duplicate application, and cross-status linkage.

[0035] The business rule set contains 176 business rules, of which 41 rules have dependencies and 13 sets of rules have constraint conflicts. Business rules are mainly used to define the range of attribute values, the order of state transitions, and the relationship between terminal operating states and business processing states. When using a common random rule generation method, although some data can satisfy individual rules, it is prone to problems such as distorted attribute combinations and insufficient coverage of low-frequency states. When using a conventional conditional diffusion model, although the data distribution is close to the sample data, the generated results still contain errors in state relationships and rule conflicts.

[0036] In this invention, the sample state sequence is input into an improved conditional diffusion model, and 48 diffusion levels are set. First, the state correlation values ​​between different sample states are calculated, and a state correlation sequence is constructed based on the state position intervals. Then, the noise memory parameter corresponding to each diffusion level is determined based on the state correlation sequence. In this embodiment, the noise memory parameter is distributed between 0.42 and 0.78. The noise covariance between diffusion levels is calculated based on the noise memory parameter, and the Davies-Harte algorithm is used to generate noise with hierarchical correlation, ensuring that the noise injected at different diffusion levels maintains a certain correlation and reducing the loss of sample attribute relationships during continuous diffusion.

[0037] Based on the analysis of the attribute constraints and state ranges corresponding to each generated object according to the task condition sequence, 16 target states are determined. The distribution features of the noise states and the conditional features of the target states are extracted respectively, and the state matching values ​​between the noise states and the target states are calculated to form a state matching matrix. The Sinkhorn algorithm is used to perform alternating normalization on the state matching matrix. After 35 iterations, the preset allocation constraints are satisfied, thereby determining the target state and alignment weight corresponding to each noise state and establishing a conditional mapping relationship.

[0038] The rule engine parses 176 business rules, extracting rule conditions, constraint attributes, constraint states, and execution order, and generates rule constraint states. A directed execution order is established based on the dependencies between these constraint states. For rules with the same constraint attributes but inconsistent constraint states, the rule with the earlier dependency position is retained first; if the dependency positions are the same, the rule with the earlier execution order is retained. After conflict resolution, 148 valid rule constraint states are obtained, and rule constraint chains are constructed based on 16 target states, with each chain containing 6 to 14 rule constraint states.

[0039] In the reverse denoising process, the improved conditional diffusion model determines the target state and alignment weight corresponding to the current noise state based on the conditional mapping relationship, performs conditional fusion between the current noise state and the target state, and then uses the DPM-Solver++ algorithm to perform noise prediction and state recovery to obtain the denoising prediction result. The rule engine performs step-by-step constraint matching on the denoising prediction result along the corresponding rule constraint chain. When it finds that the state position does not meet the rule conditions, it determines the state deviation direction and state deviation value, and forms a constraint correction state.

[0040] The denoising prediction result is constrained and guided to be corrected based on the constraint correction state. The corrected result is then used as the input state for the next inverse denoising level, and inverse denoising prediction, step-by-step constraint matching, and constraint-guided correction are continued. The iteration stops when all 48 diffusion levels have completed inverse denoising processing, the current correction result satisfies all rule constraint states in the rule constraint chain, and the state change between two adjacent correction results is less than 0.006.

[0041] To verify the technical effect of the present invention, the same data generation task was completed using a rule-based random generation method, a conventional conditional diffusion method, and the method of the present invention, respectively, wherein: Data distribution bias represents the difference in data distribution between synthetic data and sample data; the smaller the value, the closer the synthetic data is to the sample data. Attribute association bias indicates the degree of deviation in the association between different attributes in the composite data. The smaller the value, the better the attribute association is preserved. The rule compliance level indicates the degree to which the synthesized data meets the business rule set; the closer the value is to 1, the better the rule compliance effect. The target state coverage number represents the number of target states actually covered by the generated data; The number of invalid data indicates the number of data that does not meet the generation requirements after verification.

[0042] Table 1 Performance comparison of different synthetic data generation methods

[0043] Table 1 shows that the rule-based random generation method achieves a rule compliance rate of 0.984 in regular state data, but the data distribution bias is 0.143, the attribute association bias is 0.116, and the number of target states covered is only 9. This indicates that although the method can meet some rule requirements, it is difficult to maintain the true distribution and attribute association of the sample data. In the generation of low-frequency state data, the data distribution bias of the rule-based random generation method increases to 0.211, the attribute association bias increases to 0.174, and the number of target states covered is only 7, indicating that fixed rules have a weak ability to generate low-frequency states and complex associated states.

[0044] The conventional conditional diffusion method outperforms the rule-based random generation method in preserving data distribution, reducing the data distribution bias of regular state data to 0.061 and high-frequency state data to 0.054. However, because business rules are not continuously involved in the reverse denoising process, the rule compliance rate is relatively low. Particularly in conflict rule data, the rule compliance rate of the conventional conditional diffusion method drops to 0.846, with 326 invalid data points, indicating that relying solely on the model to learn the data distribution is insufficient to guarantee that the generated results consistently meet business rules.

[0045] The method of this invention demonstrates good overall performance across four types of data. Specifically, for the regular state data, the data distribution bias is 0.038, the attribute association bias is 0.031, the rule compliance rate reaches 0.994, and the number of invalid data is only 17. The low-frequency state data covers all 16 target states, with a data distribution bias of 0.047, an attribute association bias of 0.039, and the number of invalid data is 24. The rule compliance rate for the conflict rule data remains at 0.982, with 46 invalid data, significantly outperforming conventional conditional diffusion methods.

[0046] Therefore, this invention preserves the distribution patterns and attribute association features in the sample data through non-Markov diffusion processing, and embeds the rule constraint chain into the inverse denoising process. This allows business rules to continuously exert a constraining effect during the generation of synthetic data, rather than simply being filtered after data generation. Thus, this invention improves the consistency of synthetic data distribution and the accuracy of attribute associations while enhancing the rule compliance, target state coverage, and controllability of the generation process, verifying its feasibility and effectiveness in real-world data generation scenarios.

[0047] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for generating synthetic data that integrates AI modeling and rule engines, characterized in that, Includes the following steps: S1. Obtain the sample data and business rule set corresponding to the synthetic data generation task, and preprocess them to construct the sample state sequence and task condition sequence. S2. Construct an improved conditional diffusion model, perform non-Markov diffusion processing on the sample state sequence to obtain a noisy state sequence, determine the target state based on the task condition sequence, and establish a conditional mapping relationship between the noisy state sequence and each target state. S3. The business rule set is parsed through the rule engine, the rule constraint state corresponding to each business rule is extracted, and dependency modeling and conflict resolution operations are performed. Based on the task condition sequence, the rule constraint state corresponding to each target state is selected from the conflict resolution results to construct the rule constraint chain. S4. In the improved conditional diffusion model, an inverse denoising prediction operation is performed based on the mapping relationship between the noise state sequence and the condition. The denoising prediction result is subjected to step-by-step constraint matching along the rule constraint chain to construct the constraint correction state. S5. Perform constraint-guided correction on the denoising prediction results according to the constraint correction state, and iteratively perform inverse denoising prediction, step-by-step constraint matching and constraint-guided correction operations on the correction results until the preset termination condition is met, and generate a synthetic dataset. S6. Verify the synthesized dataset and adjust the parameters of the improved conditional diffusion model and rule engine based on the verification results.

2. The synthetic data generation method integrating AI modeling and rule engine according to claim 1, characterized in that, The sample data refers to the raw data obtained from real data sources for learning data distribution patterns and attribute relationships. The business rule set refers to a set of rules pre-configured for data attribute values, attribute relationships, state transition relationships, and data generation conditions. The preprocessing refers to performing data cleaning, missing state correction, abnormal state correction, data encoding, and numerical normalization on the sample data. The improved conditional diffusion model refers to a deep learning model used to perform conditional synthetic data generation. The rule engine refers to a rule processing unit used to enforce rule constraints on the synthetic data generation process.

3. The synthetic data generation method integrating AI modeling and rule engine according to claim 1, characterized in that, S2 specifically includes: S21. Construct an improved conditional diffusion model, which includes a conditional diffusion unit and an inverse denoising unit; S22. In the conditional diffusion unit, non-Markov noise injection processing is performed on the sample state sequence according to the preset diffusion level to iteratively generate a noise state sequence. S23. Determine the target state corresponding to each task condition based on the task condition sequence, and perform state alignment processing on the noisy state sequence and each target state to establish a condition mapping relationship, specifically including: The generated objects, attribute constraints, and state ranges corresponding to each task condition in the task condition sequence are analyzed, and task conditions with the same generated objects are associated to form a set of condition states. The conditional state set is divided according to attribute constraints and state range, and the state differences between each conditional state are calculated to determine the target state corresponding to each task condition. The distribution features of each noise state and the condition features of each target state in the noise state sequence are extracted respectively. The state matching value between each noise state and each target state is calculated in turn, and a state matching matrix is ​​constructed. The Sinkhorn algorithm is used to perform alternating normalization on the state matching matrix so that the state matching values ​​corresponding to each noise state and the target state satisfy the preset allocation constraints, thus obtaining the state alignment matrix. The target state and alignment weight corresponding to each noise state are determined based on the state alignment matrix, and correlation processing is performed to establish a conditional mapping relationship.

4. The synthetic data generation method integrating AI modeling and rule engine according to claim 3, characterized in that, S22 specifically includes: S221. Calculate the state correlation value between each sample state in the sample state sequence, and combine it with the positional interval between each sample state to form a state correlation sequence. S222. Map the state association sequence to a preset diffusion level, determine the noise memory parameter using the state association value corresponding to each diffusion level, and calculate the noise covariance between different diffusion levels using the noise memory parameter to construct a covariance sequence. S223. Using the Davies-Harte algorithm, the covariance sequence is expanded into a cyclic embedding sequence. The cyclic embedding sequence is then subjected to a fast Fourier transform to obtain the eigenvalues ​​corresponding to each frequency position. S224. Initialize multiple sets of independent random number pairs that follow a standard normal distribution. Use each set of random number pairs as the real and imaginary parts to form complex random quantities, and make the complex random quantities located at symmetrical frequency positions conjugate to each other. S225. Multiply the square root of the characteristic value corresponding to each frequency position by the corresponding complex random quantity, and perform the inverse fast Fourier transform operation to generate the relevant noise corresponding to each preset diffusion level. S226. Take the sample state sequence as the diffusion state of the first preset diffusion level, and set the corresponding state retention coefficient and noise injection coefficient for each preset diffusion level. Multiply the current diffusion state by the state retention coefficient, and multiply the corresponding related noise by the noise injection coefficient. Add the two multiplication results to obtain the current noise state, and use the current noise state as the diffusion state of the next preset diffusion level. After completing the calculations for all preset diffusion levels, the non-Markov diffusion process ends, and the noise state sequence is obtained.

5. The synthetic data generation method integrating AI modeling and rule engine according to claim 1, characterized in that, S3 specifically includes: S31. The business rule set is parsed one by one through the rule engine to extract the rule conditions, constraint attributes, constraint status and execution order corresponding to each business rule, and the status is encoded to generate the rule constraint status corresponding to each business rule. S32. Based on the rule conditions, constraint attributes and constraint states, perform dependency modeling operations on each rule constraint state to extract the dependency relationships between each rule constraint state. S33. Identify the constraint states of rules with the same constraint attributes but inconsistent constraint states, and compare and determine them according to the corresponding dependencies and execution order. Prioritize retaining the constraint states of rules that are in the earlier dependency positions. When the dependency positions are the same, retain the constraint states of rules that are executed earlier, and complete the conflict resolution process. Wherein, dependency position represents the position of each rule constraint state in the directed execution order formed by the dependency relationship; S34. Based on the sequence of task conditions, match each target state with the rule constraint state after conflict resolution, and filter the rule constraint state corresponding to each target state. S35. Arrange the selected rule constraint states according to the dependencies and execution order between each rule constraint state, establish the constraint transit relationship between adjacent rule constraint states, and construct the rule constraint chain.

6. The synthetic data generation method integrating AI modeling and rule engine according to claim 1, characterized in that, S4 specifically includes: S41. In the reverse denoising unit of the improved conditional diffusion model, the current noise state is extracted sequentially from the noise state sequence according to the reverse order of the preset diffusion level, and the target state and alignment weight corresponding to the current noise state are determined according to the conditional mapping relationship. S42. Perform conditional fusion processing on the current noise state and the target state according to the alignment weight; S43. Perform noise prediction and state restoration processing on the conditional fusion results to obtain denoised prediction results; S44. Extract the rule constraint state corresponding to the target state from the rule constraint chain, and perform step-by-step constraint matching processing on the denoising prediction result according to the constraint propagation relationship; S45. Based on the constraint matching results, locate the state positions in the denoising prediction results that do not meet the rule conditions, and determine the state deviation direction and state deviation value corresponding to each state position. S46. Based on the direction and degree of state deviation, calculate the correction direction and correction value of the corresponding state position, and construct the corresponding constraint correction state.

7. The synthetic data generation method integrating AI modeling and rule engine according to claim 6, characterized in that, Specifically, S43 includes: S431. In the reverse denoising unit, noise residual prediction processing is performed on the conditional fusion result to obtain the noise prediction state corresponding to the current diffusion level. S432. Based on the current noise state, the noise prediction state, and the state retention coefficient and noise injection coefficient corresponding to the current diffusion level, the current data prediction state is obtained by inversion. S433. Extract the predicted state of historical data from the historical diffusion level that has completed the reverse denoising process, and determine the reverse solution step size based on the state retention coefficient and noise injection coefficient corresponding to the current pre-diffusion level and the historical diffusion level. S434. The DPM-Solver++ algorithm is used to calculate the multi-step correction state based on the current data prediction state, the historical data prediction state, and the reverse solution step size. S435. Perform reverse state update based on the current noise state, the current data prediction state, and the multi-step correction state to obtain the denoising prediction result and record the corresponding data prediction state.

8. The synthetic data generation method integrating AI modeling and rule engine according to claim 1, characterized in that, S5 specifically includes: S51. Based on the constraint correction state, the state values ​​at the corresponding state positions in the denoising prediction results are constrained and guided to be corrected to obtain the correction results. S52. Use the correction result as the input state for the next inverse denoising level, perform the inverse denoising prediction operation, and obtain a new denoising prediction result. S53. Perform stepwise constraint matching along the rule constraint chain on the new denoising prediction result, perform constraint-guided correction based on the constraint matching result, and iteratively execute the reverse denoising prediction, stepwise constraint matching and constraint-guided correction operations. S54. When the reverse denoising process of all preset diffusion levels is completed, the current correction result satisfies all rule constraint states in the rule constraint chain, and the state change between two adjacent constraint-guided correction results is less than the preset convergence threshold, the preset termination condition is satisfied, the iteration is stopped, and a synthetic dataset is generated based on the current correction result.

9. The synthetic data generation method integrating AI modeling and rule engine according to claim 1, characterized in that, S6 specifically includes: S61. Verification personnel manually verify the data distribution, attribute relationships, and rule compliance of the synthetic dataset, and generate verification results. S62. Determine the data distribution deviation, attribute association deviation, and rule constraint deviation corresponding to the synthetic dataset based on the verification results; S63. Adjust the generation parameters of the improved conditional diffusion model based on the data distribution deviation and attribute association deviation, and adjust the rule execution parameters of the rule engine based on the rule constraint deviation.

10. A synthetic data generation system integrating AI modeling and rule engine, comprising executing the synthetic data generation method integrating AI modeling and rule engine as described in any one of claims 1 to 9, characterized in that, include: The data acquisition module is used to acquire sample data and business rule sets corresponding to the synthetic data generation task, and to preprocess them to construct sample state sequences and task condition sequences. The conditional diffusion module is used to construct an improved conditional diffusion model. It performs non-Markov diffusion processing on the sample state sequence to obtain a noisy state sequence, determines the target state based on the task condition sequence, and establishes a conditional mapping relationship between the noisy state sequence and each target state. The rule constraint module is used to parse the business rule set through the rule engine, extract the rule constraint state corresponding to each business rule, and perform dependency modeling and conflict resolution operations. Based on the task condition sequence, it selects the rule constraint state corresponding to each target state from the conflict resolution results and constructs the rule constraint chain. The inverse prediction module is used in the improved conditional diffusion model to perform inverse denoising prediction based on the mapping relationship between the noise state sequence and the condition. It performs step-by-step constraint matching processing on the denoising prediction results along the rule constraint chain to construct the constraint correction state. The iterative synthesis module is used to perform constraint-guided correction on the denoised prediction results according to the constraint correction state, and iteratively perform inverse denoising prediction, step-by-step constraint matching and constraint-guided correction operations on the correction results until the preset termination condition is met, and generate a synthetic dataset. The verification and adjustment module is used to verify the synthetic dataset and adjust the parameters of the improved conditional diffusion model and rule engine based on the verification results.

Images