Dynamic inference method and system based on Bayesian inference evolution probability model

Abstract

The invention relates to the technical field of model construction, and discloses a dynamic inference method and system based on a Bayesian inference evolution probability model, and the method comprises the steps: constructing an initial network topology of an initial data segment, initializing the probability distribution parameters of nodes in the initial network topology, and generating a basic probability graph model; variational inference is carried out on the basic probability graph model and the new data fragments to obtain an updated probability model; performing fitting degree evaluation on the updated probability model and the new data fragments to obtain a model fitting abnormal signal; performing mismatch degree evaluation on the updated probability model, and adjusting a network structure of the updated probability model to obtain an evolved probability model; calculating similarity measurement, and performing component removal and combination on the evolved probability model to obtain a simplified probability model; optimizing parameters of the simplified probability model to obtain a robust probability model; according to the method, the accuracy of dynamic inference based on the Bayesian inference evolution probability model can be improved.

Description

Technical Field

[0001] This invention relates to the field of model building technology, and in particular to a dynamic reasoning method and system based on a Bayesian inference evolution probability model. Background Technology

[0002] Existing technologies have significant shortcomings in the model building and incremental update stages of dynamic inference. They fail to construct a standardized initial network topology based on the dependencies of random variables in the initial data fragments, simply setting the model structure without accurately characterizing the intrinsic relationships between variables. Furthermore, the initialization of node probability distribution parameters does not fully incorporate observational statistics and domain prior knowledge, resulting in low initial accuracy of the basic model. After acquiring new data fragments, efficient incremental updates are not performed through variational inference; instead, the entire model is retrained or parameters are simply adjusted, which is not only time-consuming and labor-intensive but also prone to losing valuable information from historical data, failing to adapt to the dynamic changes in the data flow.

[0003] Existing technologies have significant shortcomings in the model optimization and robustness assurance aspects of dynamic inference. They fail to evaluate the fit and detect anomalies in the updated model, making it impossible to promptly identify model mismatches with new data; they do not adaptively adjust the network structure based on mismatches, resulting in a rigid and fixed model structure that struggles to adapt to the dynamic evolution of data distribution; they do not streamline and merge components in the evolved model, leading to redundant model structures and low inference efficiency; and they fail to incorporate historical statistical features to construct regularization constraints to optimize parameters, focusing solely on the model's fit to new data, resulting in weak generalization ability, poor robustness, and an inability to meet the inference needs of complex dynamic scenarios. Summary of the Invention

[0004] This invention provides a dynamic reasoning method and system based on a Bayesian inference evolution probability model to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, this invention provides a dynamic reasoning method based on a Bayesian inference evolutionary probability model, comprising:

[0006] S1. Using random variables in the initial data segment as nodes and the dependencies between variables as directed edges, construct the initial network topology of the initial data segment, initialize the probability distribution parameters of the nodes in the initial network topology, and generate the basic probabilistic graphical model of the initial data segment.

[0007] S2. Obtain a new data segment from the same data stream, and perform variational inference on the basic probabilistic graphical model and the new data segment to obtain an updated probabilistic model for the initial data segment;

[0008] S3. Evaluate the fit of the updated probability model and the new data segment to obtain the model fit anomaly signal of the updated probability model;

[0009] S4. Based on the abnormal signal fitted by the model, evaluate the degree of mismatch of the updated probability model, so as to adaptively adjust the network structure of the updated probability model and obtain the evolved probability model of the updated probability model.

[0010] S5. Calculate the similarity measure of the components in the evolved probability model, and remove and merge the components in the evolved probability model based on the similarity measure to obtain a simplified probability model of the evolved probability model.

[0011] S6. Based on historical statistical characteristics, construct regularization constraint terms to optimize the parameters of the simplified probability model, thereby obtaining a robust probability model of the simplified probability model.

[0012] In a preferred embodiment, the step of constructing an initial network topology for the initial data segment using random variables in the initial data segment as nodes and the dependencies between variables as directed edges, and initializing the probability distribution parameters of the nodes in the initial network topology to generate a basic probabilistic graphical model for the initial data segment includes:

[0013] Parse the initial data segment, identify the random variables in the initial data segment, and define the random variables as nodes in the initial data segment;

[0014] Analyze the statistical associations between the random variables, identify and determine the variable pairs with significant dependencies, and define the dependencies between the variable pairs as directed edges connecting the corresponding nodes to generate the initial network topology of the initial data fragment.

[0015] Based on the observation statistics and domain prior knowledge of the node in the initial data segment, the conditional probability distribution parameters of the node are initialized;

[0016] The initial network topology is combined with the conditional probability distribution parameters to obtain the basic probabilistic graphical model of the initial data segment.

[0017] In a preferred embodiment, the step of acquiring a new data segment from the same data stream and performing variational inference on the basic probabilistic graphical model and the new data segment to obtain an updated probabilistic model for the initial data segment includes:

[0018] Select an exponential family distribution that is compatible with the conditional distribution of nodes in the basic probabilistic graphical model to generate the variational distribution family of the basic probabilistic graphical model;

[0019] Based on the observed state of the variables in the new data segment, determine the initial variational parameters of the variational distribution family;

[0020] Based on the prior distribution of the basic probabilistic graphical model and the new data segment, a divergence metric between the variational distribution family and the posterior distribution is defined, and the initial variational parameters are iteratively adjusted with the goal of minimizing the divergence metric.

[0021] When the change in the initial variational parameters is lower than the preset convergence threshold, the variational inference process is determined to be complete, and the expected parameters of the variational distribution family are assigned to the corresponding node of the update probability model to obtain the update probability model of the initial data segment.

[0022] In a preferred embodiment, the step of evaluating the fit between the updated probability model and the new data segment to obtain a model fit anomaly signal for the updated probability model includes:

[0023] Based on the updated probability model, probabilistic inference is performed on the variables in the new data segment to obtain the model prediction distribution of the new data segment;

[0024] The observed values ​​of variables in the new data segment are compared with the expected values ​​of the model's predicted distribution to construct the residual vector of the updated probability model;

[0025] Based on the degree of difference in the dimensions of the residual vector, the overall fit between the updated probability model and the new data segment is evaluated.

[0026] The overall fit is compared with a preset anomaly detection threshold. When the overall fit is less than the preset anomaly detection threshold, a model fit anomaly signal for the updated probability model is generated.

[0027] In a preferred embodiment, the step of assessing the mismatch degree of the updated probability model based on the model-fitted abnormal signal, and adaptively adjusting the network structure of the updated probability model to obtain the evolved probability model of the updated probability model, includes:

[0028] The abnormal signal of the model fitting is analyzed to obtain the specific nodes and specific directed edges in the updated probability model that are mismatched with the new data segment, and the specific nodes and specific directed edges are used as the structural elements to be evaluated in the updated probability model.

[0029] Assess the overall degree of mismatch between the structuring element to be evaluated and the dependencies between variables reflected by the new data fragment;

[0030] Based on the overall degree of mismatch, determine the type of adjustment operation for the structural element to be evaluated;

[0031] Based on the adjustment operation type, the network structure of the updated probability model is modified to obtain the evolved probability model of the updated probability model.

[0032] In a preferred embodiment, the step of calculating the similarity measure of components in the evolved probabilistic model, and removing and merging components in the evolved probabilistic model based on the similarity measure to obtain a simplified probabilistic model of the evolved probabilistic model, includes:

[0033] Based on the network structure of the evolved probability model, identify the structural components in the network structure;

[0034] Calculate the similarity metric between the structural components;

[0035] Based on the similarity metric, the redundancy and complementarity between the structural components are evaluated, and based on the evaluation results, the structural components are adjusted to obtain a simplified probability model of the evolved probability model.

[0036] In a preferred embodiment, the component adjustment includes:

[0037] Based on the similarity metric, if the similarity between the first structural component and the second structural component in the evolved probabilistic model is lower than the correlation benchmark required by the evolved probabilistic model, then the components representing secondary data patterns in the first structural component and the second structural component are identified as redundant structural components and removed.

[0038] When the similarity between the first structural component and the second structural component in the evolved probabilistic model meets the correlation benchmark required by the evolved probabilistic model, the first structural component and the second structural component are determined as a highly complementary component pair, and the node sets of the first structural component and the second structural component are integrated with the directed edges.

[0039] In a preferred embodiment, the similarity metric is calculated using the following formula:

[0040] ;

[0041] In the formula, For structural components and structural components Similarity measurement between For components The number of nodes included. For components The number of nodes included. For nodes With nodes The connection strength weighting factor between them Let be the probability distribution similarity function. For nodes The set of parent nodes, For nodes The set of parent nodes, For nodes Conditional probability distribution For nodes Conditional probability distribution.

[0042] In a preferred embodiment, the step of constructing regularization constraint terms based on historical statistical characteristics to optimize the parameters of the simplified probability model, thereby obtaining a robust probability model of the simplified probability model, includes:

[0043] Extract historical statistical features from the initial data fragment and the new data fragment;

[0044] Based on the historical statistical characteristics, construct the regularization constraint term of the simplified probability model;

[0045] The regularization constraint term is combined with the goodness-of-fit term of the simplified probabilistic model on the new data segment to obtain the comprehensive optimization guide of the simplified probabilistic model;

[0046] Based on the aforementioned comprehensive optimization approach, the conditional probability distribution parameters of the simplified probability model are iteratively adjusted to obtain a robust probability model of the simplified probability model.

[0047] To address the aforementioned problems, this invention also provides a dynamic inference system based on a Bayesian inference evolution probability model, the system comprising:

[0048] The initial model building module is used to construct the initial network topology of the initial data segment using random variables in the initial data segment as nodes and the dependencies between variables as directed edges, and to initialize the probability distribution parameters of the nodes in the initial network topology to generate the basic probabilistic graphical model of the initial data segment.

[0049] The incremental inference update module is used to acquire new data segments from the same data stream and perform variational inference on the basic probabilistic graphical model and the new data segments to obtain the updated probabilistic model of the initial data segments.

[0050] The fitting anomaly detection module is used to evaluate the fit between the updated probability model and the new data segment, and obtain the model fitting anomaly signal of the updated probability model.

[0051] The structural adaptive evolution module is used to assess the degree of mismatch of the updated probability model based on the abnormal signal fitted by the model, so as to adaptively adjust the network structure of the updated probability model and obtain the evolved probability model of the updated probability model.

[0052] The component simplification and fusion module is used to calculate the similarity measure of components in the evolved probability model, and remove and merge components in the evolved probability model based on the similarity measure to obtain a simplified probability model of the evolved probability model.

[0053] The regularization robust optimization module is used to construct regularization constraint terms based on historical statistical characteristics to optimize the parameters of the simplified probability model, thereby obtaining a robust probability model of the simplified probability model.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] 1. This invention provides an efficient and adaptable model foundation for dynamic inference through precise modeling and incremental updates. An initial network topology is constructed using variables and dependencies in the initial data fragment. Parameters are initialized using observational statistics and prior knowledge to generate a basic probabilistic graphical model. Upon acquiring new data fragments, parameters are iteratively optimized through variational inference to generate an updated probabilistic model. Simultaneously, model mismatch anomalies are promptly captured through goodness-of-fit evaluation, ensuring the model's adaptability to the dynamic changes in the data stream.

[0056] 2. This invention significantly improves the efficiency and robustness of dynamic inference by leveraging structural evolution and parameter optimization. Based on mismatch anomaly signals, the network structure is adaptively adjusted to obtain the evolved model. Redundant components are simplified and complementary components are fused using similarity metrics to reduce model complexity. Regularization constraints are constructed using historical statistical features to optimize parameters, generating a robust probabilistic model that ensures both inference speed and improves the model's generalization ability and the reliability of inference results. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating a dynamic reasoning method based on a Bayesian inference evolution probability model, provided in an embodiment of the present invention.

[0058] Figure 2 A functional block diagram of a dynamic inference system based on a Bayesian inference evolution probability model provided in an embodiment of the present invention;

[0059] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0060] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0061] This application provides a dynamic reasoning method based on a Bayesian inference evolutionary probability model. The execution subject of this dynamic reasoning method based on a Bayesian inference evolutionary probability model includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the dynamic reasoning method based on a Bayesian inference evolutionary probability model can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0062] Reference Figure 1 The diagram shown is a flowchart illustrating a dynamic reasoning method based on a Bayesian inference evolutionary probability model, according to an embodiment of the present invention. In this embodiment, the dynamic reasoning method based on a Bayesian inference evolutionary probability model includes:

[0063] S1. Using random variables in the initial data segment as nodes and the dependencies between variables as directed edges, construct the initial network topology of the initial data segment, initialize the probability distribution parameters of the nodes in the initial network topology, and generate the basic probabilistic graphical model of the initial data segment.

[0064] In this embodiment of the invention, the step of constructing an initial network topology for the initial data segment using random variables in the initial data segment as nodes and the dependencies between variables as directed edges, and initializing the probability distribution parameters of the nodes in the initial network topology to generate a basic probabilistic graphical model for the initial data segment includes:

[0065] Parse the initial data segment, identify the random variables in the initial data segment, and define the random variables as nodes in the initial data segment;

[0066] Analyze the statistical associations between the random variables, identify and determine the variable pairs with significant dependencies, and define the dependencies between the variable pairs as directed edges connecting the corresponding nodes to generate the initial network topology of the initial data fragment.

[0067] Based on the observation statistics and domain prior knowledge of the node in the initial data segment, the conditional probability distribution parameters of the node are initialized;

[0068] The initial network topology is combined with the conditional probability distribution parameters to obtain the basic probabilistic graphical model of the initial data segment.

[0069] The initial data fragments are fully analyzed, and each element that reflects the data characteristics and may change is extracted. These elements are the random variables in the initial data fragments. Each random variable is defined as an independent node in the initial network topology to ensure that each random variable can be uniquely corresponded and reflected in the network topology.

[0070] Statistical analysis methods are used to compare all random variables pairwise to calculate the strength of the association between them. Variable pairs with significant association strengths are selected using fixed criteria; these pairs are considered to have significant dependencies. Based on the direction of the dependency in each variable pair, a directed edge is established for each pair, pointing from the node corresponding to the source variable to the node corresponding to the target variable. Connecting all nodes and directed edges according to the above rules forms the initial network topology for the initial data segment.

[0071] The actual observation data of the random variables corresponding to each node in the initial data segment are collected, and the observation data are statistically summarized to obtain observation statistics such as the frequency of occurrence and distribution characteristics of each variable. Combining existing mature theories, practical experience, and publicly available research conclusions in this field, the occurrence probability of the random variables corresponding to each node under different dependency conditions is initially set, thus completing the initialization of the conditional probability distribution parameters of the nodes in the initial network topology.

[0072] Using the constructed initial network topology as the basic framework, the conditional probability distribution parameters of each node after initialization are matched one by one to the corresponding node, so that each node has both topological connectivity and clear probability quantification attributes. By integrating the topological structure and probability parameters, a complete model that can simultaneously describe the dependencies between random variables and the probability distribution laws is formed. This model is the basic probabilistic graphical model of the initial data fragment.

[0073] The beneficial effects are that by parsing the initial data fragments and identifying random variables as nodes, the core elements of the data can be accurately captured, ensuring that every variable that affects the inference results can be reflected in the model, providing comprehensive and complete basic elements for subsequent network topology construction.

[0074] By analyzing the statistical associations between random variables and identifying pairs of variables with significant dependencies, the dependencies are defined as directed edges that generate the initial network topology. This clearly characterizes the inherent logic of the relationships between variables, enabling the model to accurately reflect the structural characteristics of the data and avoiding inference biases caused by ambiguity in the relationships between variables.

[0075] The conditional probability distribution parameters are initialized based on the observation statistics of nodes and the prior knowledge of the domain. By combining actual data characteristics and mature domain experience, the initial values ​​of the parameters are made to better fit the actual distribution of the data, thereby improving the initial accuracy of the basic probabilistic graphical model and reducing the cost of subsequent model iteration and optimization.

[0076] By combining the initial network topology with conditional probability distribution parameters, a basic probabilistic graphical model is generated, achieving an organic unity between variable relationship structure and probability quantification attributes. This constructs an initial model that combines logical rationality and numerical accuracy, laying a solid foundation for subsequent incremental updates and dynamic reasoning based on new data.

[0077] S2. Obtain a new data segment from the same data stream, and perform variational inference on the basic probabilistic graphical model and the new data segment to obtain an updated probabilistic model for the initial data segment;

[0078] In this embodiment of the invention, the step of acquiring a new data segment from the same data stream and performing variational inference on the basic probabilistic graphical model and the new data segment to obtain an updated probabilistic model for the initial data segment includes:

[0079] Select an exponential family distribution that is compatible with the conditional distribution of nodes in the basic probabilistic graphical model to generate the variational distribution family of the basic probabilistic graphical model;

[0080] Based on the observed state of the variables in the new data segment, determine the initial variational parameters of the variational distribution family;

[0081] Based on the prior distribution of the basic probabilistic graphical model and the new data segment, a divergence metric between the variational distribution family and the posterior distribution is defined, and the initial variational parameters are iteratively adjusted with the goal of minimizing the divergence metric.

[0082] When the change in the initial variational parameters is lower than the preset convergence threshold, the variational inference process is determined to be complete, and the expected parameters of the variational distribution family are assigned to the corresponding node of the update probability model to obtain the update probability model of the initial data segment.

[0083] By deeply analyzing the conditional distribution characteristics of each node in the basic probabilistic graphical model, the distribution patterns and data forms of their probability values ​​are clarified. Based on this, an exponential family distribution type that can accurately match the conditional distribution is selected. It is ensured that the selected exponential family distribution can fully cover the value range and trend of the node conditional distribution. Using this exponential family distribution as the core, a variational distribution family is constructed for the basic probabilistic graphical model, enabling the variational distribution family to approximate the true posterior distribution.

[0084] Each variable in the new data segment is checked individually to clarify the actual observation results of each variable and determine its specific observation state. Based on the value characteristics and distribution of these observation states, and combined with the structural characteristics of the variational distribution family, initial values ​​are assigned to each parameter in the variational distribution family. These initial values ​​should be able to initially reflect the observation patterns of the variables in the new data segment, thereby generating the initial variational parameters of the variational distribution family.

[0085] The prior distribution information pre-defined in the basic probabilistic graphical model is extracted, and the observed data and correlations of variables in the new data fragment are comprehensively reviewed. These two factors are used as the core basis for defining a measurement standard. This standard quantifies the degree of difference between the variational distribution family and the true posterior distribution, i.e., the divergence measure. With the core objective of reducing this divergence measure, the initial variational parameters are continuously adjusted. After each adjustment, the divergence measure is recalculated, and its changing trend is observed. The parameters are continuously optimized to gradually narrow the distance between the variational distribution family and the true posterior distribution.

[0086] During the iterative adjustment of the initial variational parameters, the difference between the current parameter and the previous parameter is calculated after each adjustment to measure the change in the parameter. This change is continuously compared with a preset convergence threshold. When the change in the parameter is lower than the convergence threshold, it indicates that the variational distribution family has sufficiently approximated the true posterior distribution, and the variational inference process is considered complete. The expected values ​​of each parameter in the variational distribution family are extracted, and these expected parameters are used as the parameters of the corresponding nodes in the updated probability model. Through parameter replacement and model integration, the updated probability model of the initial data fragment is finally obtained.

[0087] The beneficial effect is that by selecting an exponential family distribution that is compatible with the conditional distribution of the nodes in the basic probabilistic graphical model to generate a family of variational distributions, the variational distribution can accurately approximate the shape and characteristics of the true posterior distribution, providing a highly adaptable distributional basis for subsequent efficient inference.

[0088] Determine the initial variational parameters based on the observed state of variables in the new data fragment, so that the initial parameter values ​​fit the actual observation patterns of the new data, reduce the number of subsequent iterations and adjustments and the associated costs, and improve the efficiency of variational inference.

[0089] Based on the prior distribution of the basic probabilistic graphical model and the new data fragment, a divergence metric is defined. This not only preserves the effective information of the historical model but also makes full use of the real-time characteristics of the new data, enabling the divergence metric to accurately quantify the difference between the variational distribution and the true posterior distribution.

[0090] The initial variational parameters are iteratively adjusted with the goal of minimizing the divergence metric. Through continuous optimization, the distance between the variational distribution and the true posterior distribution is gradually narrowed, ensuring the scientific validity and effectiveness of the parameter adjustment direction and improving the accuracy of the model parameters.

[0091] Setting a convergence threshold to determine when variational inference is complete avoids resource waste caused by excessive iteration, while ensuring that the parameters reach a stable state. The expected parameters of the variational distribution family at this point are then assigned to the corresponding nodes of the update probability model. The generated update probability model can take into account both historical data patterns and new data features, achieving efficient incremental updates of the model.

[0092] S3. Evaluate the fit of the updated probability model and the new data segment to obtain the model fit anomaly signal of the updated probability model;

[0093] In this embodiment of the invention, the step of evaluating the fit between the updated probability model and the new data segment to obtain a model fit anomaly signal for the updated probability model includes:

[0094] Based on the updated probability model, probabilistic inference is performed on the variables in the new data segment to obtain the model prediction distribution of the new data segment;

[0095] The observed values ​​of variables in the new data segment are compared with the expected values ​​of the model's predicted distribution to construct the residual vector of the updated probability model;

[0096] Based on the degree of difference in the dimensions of the residual vector, the overall fit between the updated probability model and the new data segment is evaluated.

[0097] The overall fit is compared with a preset anomaly detection threshold. When the overall fit is less than the preset anomaly detection threshold, a model fit anomaly signal for the updated probability model is generated.

[0098] Based on the conditional probability distribution of each node and the dependencies between variables in the updated probability model, the possible values ​​of each variable in the new data segment and their corresponding probabilities are derived and calculated. Following the correlation logic of the variables in the new data segment, the probability derivation results of all variables are integrated to form a probability distribution that fully reflects the variable value patterns in the new data segment. This distribution is the model prediction distribution for the new data segment.

[0099] The actual observed values ​​of each variable are extracted from the new data segment one by one. Simultaneously, the expected value of the corresponding variable in the model's predicted distribution is calculated. The expected value is obtained by combining all possible values ​​of the variable in the predicted distribution and their corresponding probabilities. The actual observed value of each variable is compared with its corresponding expected predicted value. The expected value is subtracted from the observed value of each variable to obtain the residual for each variable. All residuals are arranged according to the order of the variables in the new data segment to construct the residual vector of the updated probability model.

[0100] The residual values ​​for each dimension in the residual vector are analyzed, and the differences between dimensions are measured by calculating the overall dispersion of all residual values. The calculation of the overall dispersion must consider the magnitude and distribution characteristics of each residual value to avoid the influence of extreme cases of a single residual value on the overall judgment. Based on the results of the overall dispersion, the degree of fit between the prediction results of the updated probability model and the actual observed data of the new data segment is determined; this degree of fit is the overall goodness of fit between the updated probability model and the new data segment.

[0101] Based on the characteristics of the data domain, the accuracy requirements of the model application scenario, and experience with historical fitting data, a fixed preset anomaly detection threshold is set in advance. This threshold is used to define the normal and abnormal ranges of the model's fit. The previously obtained overall fit is directly compared with the preset anomaly detection threshold. When the overall fit value is less than the preset anomaly detection threshold, it indicates that the prediction result of the updated probability model deviates from the actual situation of the new data segment beyond the acceptable range. At this time, a signal is immediately generated to indicate the model fit anomaly; this signal is the updated probability model's model fit anomaly signal.

[0102] The beneficial effects are that the model predicts the distribution of variables in new data segments by performing probabilistic inference based on the updated probability model, making full use of the model's parameters and structural information, accurately characterizing the predicted value patterns of variables, and providing a clear prediction benchmark for subsequent fit evaluation.

[0103] By comparing the observed values ​​of variables with the expected values ​​of the model's predicted distribution to construct residual vectors, the deviation between the actual and predicted values ​​of each variable is directly quantified, making the differences in the model's fit with the new data explicit and providing accurate basic data for evaluating the overall fit.

[0104] The overall fit is assessed based on the degree of difference in the dimensions of the residual vector, taking into account the deviation of all variables, avoiding the one-sidedness of the assessment caused by the deviation of a single variable, and comprehensively reflecting the overall adaptability of the model to the new data.

[0105] By comparing the model with a preset anomaly detection threshold, abnormal signals are generated for model fitting, clearly defining the normal and abnormal boundaries of model fitting, and timely capturing the mismatch between the model and new data. This provides a precise trigger for subsequent model structure adjustments and parameter optimization, ensuring that the model can dynamically adapt to data changes.

[0106] S4. Based on the abnormal signal fitted by the model, evaluate the degree of mismatch of the updated probability model, so as to adaptively adjust the network structure of the updated probability model and obtain the evolved probability model of the updated probability model.

[0107] In this embodiment of the invention, the step of assessing the mismatch degree of the updated probability model based on the model-fitted abnormal signal, and adaptively adjusting the network structure of the updated probability model to obtain the evolved probability model of the updated probability model, includes:

[0108] The abnormal signal of the model fitting is analyzed to obtain the specific nodes and specific directed edges in the updated probability model that are mismatched with the new data segment, and the specific nodes and specific directed edges are used as the structural elements to be evaluated in the updated probability model.

[0109] Assess the overall degree of mismatch between the structuring element to be evaluated and the dependencies between variables reflected by the new data fragment;

[0110] Based on the overall degree of mismatch, determine the type of adjustment operation for the structural element to be evaluated;

[0111] Based on the adjustment operation type, the network structure of the updated probability model is modified to obtain the evolved probability model of the updated probability model.

[0112] A comprehensive analysis of the abnormal signals in the model fit is performed. Combining the network structure of the updated probabilistic model with the variable observation information of the new data segment, nodes in the model whose predicted results deviate significantly from the actual observed data are located. These nodes are the specific nodes that are mismatched with the new data segment. At the same time, directed edges connecting specific nodes and other related nodes are examined to identify directed edges that cannot accurately reflect the true dependencies between variables. These directed edges are the specific directed edges that are mismatched with the new data segment. All specific nodes and specific directed edges are uniformly collected and determined as the structural elements to be evaluated in the updated probabilistic model.

[0113] Collect the actual observed values ​​of all variables in the new data segment and the true relationships between variables to clarify the actual dependencies between variables. For each structural element to be evaluated, analyze the fit between the variable relationships represented by it in the updated probability model and the true variable dependencies reflected in the new data segment, and calculate the degree of deviation between the two. Combine the degree of deviation of all structural elements to be evaluated with the importance ratio of each element in the model to obtain a comprehensive value that can fully reflect the degree of mismatch between the structural elements to be evaluated and the true variable dependencies. This value is the comprehensive mismatch degree.

[0114] A fixed set of rules is established to correspond to the overall mismatch degree range and the adjustment operation type. Different mismatch degree ranges correspond to specific adjustment operation directions. The obtained overall mismatch degree is compared with the preset range. If the overall mismatch degree is in the slight mismatch range, the adjustment operation type is determined to be parameter fine-tuning; if it is in the moderate mismatch range, the adjustment operation type is determined to be directed edge modification, including the addition or deletion of directed edges; if it is in the severe mismatch range, the adjustment operation type is determined to be a combination adjustment of nodes and directed edges, ensuring that the adjustment operation type and the overall fit degree are accurately matched.

[0115] According to the determined adjustment operation type, targeted modifications are made to the network structure of the updated probabilistic model. For parameter fine-tuning, only the conditional probability distribution attributes of specific nodes are adjusted, without changing the network topology. For directed edge modification, invalid specific directed edges are deleted based on the variable dependencies reflected in the new data fragments, or directed edges are added between nodes corresponding to variables with real dependencies. For combined adjustments, specific nodes are added, removed, or their attributes are corrected simultaneously, and the connection relationships of related directed edges are optimized concurrently. After all modifications are completed, the model structure is integrated and verified to ensure that the dependencies between variables are accurate and logically coherent, resulting in the evolved probabilistic model of the updated probabilistic model.

[0116] The beneficial effects are that the analytical model can locate specific nodes and directed edges that are mismatched by fitting abnormal signals, accurately pinpoint the core structural location where the model does not match the new data, avoid blindly adjusting the entire network structure, and provide a clear target for subsequent targeted optimization.

[0117] Assess the overall degree of mismatch between the structural elements to be evaluated and the variable dependencies reflected by the new data, quantify the severity and scope of the mismatch, provide quantitative support for adjustment decisions, and ensure that the adjustment intensity is precisely matched with the degree of mismatch.

[0118] The type of adjustment operation is determined based on the overall degree of mismatch, and differentiated adjustment plans are formulated for different levels of mismatch to avoid insufficient optimization or over-adjustment caused by a single adjustment method, thereby improving the scientific and rational nature of network structure adjustment.

[0119] The network structure of the updated probability model is modified according to the type of adjustment operation, so as to realize the adaptive evolution of the model structure, break the rigidity of the fixed structure, and enable the model to dynamically adapt to the changes in variable dependencies of new data, significantly improving the model's ability to adapt to the evolution of data distribution.

[0120] The entire process forms a closed-loop optimization workflow of "anomaly localization - mismatch assessment - type determination - structural modification", ensuring that the evolved probability model is structurally more in line with the actual patterns of the new data, laying a high-quality structural foundation for subsequent component simplification and parameter optimization, and improving the accuracy of dynamic inference.

[0121] S5. Calculate the similarity measure of the components in the evolved probability model, and remove and merge the components in the evolved probability model based on the similarity measure to obtain a simplified probability model of the evolved probability model.

[0122] In this embodiment of the invention, the step of calculating the similarity measure of components in the evolved probabilistic model, and removing and merging components in the evolved probabilistic model based on the similarity measure to obtain a simplified probabilistic model of the evolved probabilistic model, includes:

[0123] Based on the network structure of the evolved probability model, identify the structural components in the network structure;

[0124] Calculate the similarity metric between the structural components;

[0125] Based on the similarity metric, the redundancy and complementarity between the structural components are evaluated, and based on the evaluation results, the structural components are adjusted to obtain a simplified probability model of the evolved probability model.

[0126] The component adjustments include:

[0127] Based on the similarity metric, if the similarity between the first structural component and the second structural component in the evolved probabilistic model is lower than the correlation benchmark required by the evolved probabilistic model, then the components representing secondary data patterns in the first structural component and the second structural component are identified as redundant structural components and removed.

[0128] When the similarity between the first structural component and the second structural component in the evolved probabilistic model meets the correlation benchmark required by the evolved probabilistic model, the first structural component and the second structural component are determined as a highly complementary component pair, and the node sets of the first structural component and the second structural component are integrated with the directed edges.

[0129] The formula for calculating the similarity measure is:

[0130] ;

[0131] In the formula, For structural components and structural components Similarity measurement between For components The number of nodes included. For components The number of nodes included. For nodes With nodes The connection strength weighting factor between them Let be the probability distribution similarity function. For nodes The set of parent nodes, For nodes The set of parent nodes, For nodes Conditional probability distribution For nodes Conditional probability distribution.

[0132] A comprehensive review of the network structure of the evolved probabilistic model is conducted to identify the interconnected clusters of nodes and combinations of directed edges. These units, which are formed by specific logical connections and can independently perform local probability inference functions, are the structural components. This ensures that each structural component can fully reflect the dependencies and probability propagation paths between a set of variables, and identifies all structural components without omission or duplication.

[0133] For each pair of identified structural components, a comparative analysis is conducted from three dimensions: node composition, directed edge connection pattern, and conditional probability distribution characteristics. The types and number of nodes in each component are compared, the connection direction and association strength of directed edges are determined to be consistent, and the similarity of the conditional probability distribution patterns of nodes within the components is analyzed. By comprehensively considering the degree of fit among these three dimensions, a quantitative result measuring the similarity between two structural components is obtained; this result is the similarity measure between structural components.

[0134] Based on the similarity measurement results between structural components, the relationship attributes between components are determined. If the similarity measurement reaches the preset redundancy judgment standard, it indicates that the two components highly overlap in variable dependency representation and probability inference functions, and are judged as redundant. If the similarity measurement is at a low level, but the variable relationships represented by each component can complement each other and jointly improve the model's characterization of the overall data patterns, they are judged as complementary. For redundant components, the core component with more comprehensive functions and more accurate probability inference is retained, and the remaining redundant components are removed. For complementary components, they are kept independent without merging or deletion. Through such component adjustments, a simplified probability model of the evolved probability model is obtained.

[0135] First, define the correlation benchmark required by the evolved probabilistic model. This benchmark, based on the data patterns of the model's application scenario, the accuracy requirements of probabilistic inference, and the standards and specifications within the field, defines the minimum degree of correlation that structural components must meet. Calculate the similarity metric between the first and second structural components in the evolved probabilistic model. Directly compare this metric with the correlation benchmark. If the similarity metric is lower than the benchmark, it indicates insufficient correlation between the two components in core dimensions such as variable dependency and probability distribution characteristics. Further analyze the data patterns represented by the two components. Combining the core patterns of the new data fragments with the model's core inference objectives, determine which component represents a secondary data pattern. Identify the component representing the secondary data pattern as a redundant structural component. Remove the node set and directed edges of this component from the network structure of the evolved probabilistic model to eliminate the redundant structural component.

[0136] Using the correlation benchmark required by the evolved probabilistic model as the criterion, the similarity measure between the first and second structural components is calculated. When the measure reaches or exceeds the correlation benchmark, it indicates that the two components have a high degree of compatibility in terms of core functions and variable correlation logic, and can complement each other to improve the model's characterization of data patterns. These components are thus identified as highly complementary component pairs. The node sets of the two components in the highly complementary pair are analyzed, and duplicate or functionally overlapping nodes are merged, while retaining their unique core nodes to form a unified merged node set. The original directed edge connections of the two components are analyzed, and conflicting or redundant directed edges are deleted. Based on the complementary logic between the components, necessary directed edges are added to improve the dependency propagation path between variables, achieving a comprehensive integration of the node sets and directed edges of the first and second structural components. This ensures that the integrated components can perform probabilistic inference more efficiently and accurately.

[0137] Similarity measurement is derived from the number of nodes in the structural components, the connection strength weight factor between nodes, the conditional probability distribution of nodes, and the result of the probability distribution similarity function. Through the integrated calculation of multi-dimensional parameters, the similarity between two structural components is quantified.

[0138] The number of nodes in a structural component is directly derived from the network structure of the evolved probabilistic model. It is obtained by counting the number of independent nodes contained in each structural component, and is a basic parameter reflecting the scale of the structural component.

[0139] The connection strength weighting factor between nodes is derived from the association analysis between nodes in the evolved probabilistic model, based on the nodes. With nodes The actual connection between the directed edges, the frequency of their co-occurrence in historical data, and domain prior knowledge are used to determine a numerical value that can characterize the degree of their association. The closer the association, the larger the weight factor.

[0140] The set of parent nodes of a node is derived from the network topology of the evolved probabilistic model. All predecessor nodes pointing to that node in the topology are directly extracted to form the set of parent nodes of that node, thus clarifying the basis of the node probability distribution.

[0141] The conditional probability distribution of a node is derived from the parameter settings of the evolved probability model. Based on the node's parent node set, the statistical results of historical observation data, and prior knowledge of the domain, the probability of each node occurring under different combinations of states in its parent node set is determined, thus forming the node's conditional probability distribution.

[0142] The probability distribution similarity function is a pre-defined fixed operation rule. By comparing the shape, value range and distribution trend of the conditional probability distributions of two nodes, it calculates a result that can quantify the degree of similarity between the two. The closer the distributions are, the larger the result value.

[0143] The significance of the formula lies in its ability to comprehensively and accurately quantify the similarity between two structural components by taking into account the scale of the structural components, the strength of the association between nodes, and the similarity of the conditional probability distribution of nodes.

[0144] The operation logic first calculates the distribution similarity between nodes in different components using a probability distribution similarity function, then amplifies the influence of key related nodes by combining the connection strength weight factor between nodes, and finally normalizes the results by multiplying the number of nodes in the components to eliminate interference caused by differences in component size.

[0145] This multi-dimensional integrated computational approach takes into account both the topological characteristics of structural components and their probability distribution characteristics, avoiding similarity assessment biases caused by single-dimensional judgments. It provides a scientific and reliable quantitative basis for the subsequent removal of redundancy and complementary integration of structural components, ensuring that the simplified probability model retains core functions while possessing efficient inference performance.

[0146] The beneficial effects are that, based on the network structure of the evolved probability model, structural components can be identified, units with independent probability inference functions in the model can be accurately separated, the core role and coverage of each component can be clarified, and clear operational objects can be provided for subsequent similarity assessment and component adjustment.

[0147] The similarity metric between structural components is calculated, and the degree of fit between components is quantified from multiple dimensions. This avoids relying on subjective judgment to evaluate component relationships and provides an objective and accurate quantitative basis for judging redundancy and complementarity.

[0148] The redundancy and complementarity of structural components are evaluated based on similarity metrics, clearly distinguishing between redundant components with overlapping functions and core components with complementary functions, ensuring that the direction of component adjustment is clear and avoiding ineffective adjustments.

[0149] Based on the evaluation results, the structural components were adjusted accordingly. Removing redundant components can simplify the model structure, reduce the consumption of computing resources, and improve inference efficiency. Merging complementary components can integrate core functions and improve the characterization of data patterns. The final simplified probabilistic model balances lightweight and functionality, laying an efficient and high-quality model foundation for subsequent parameter optimization.

[0150] Using similarity measurement and correlation benchmark as dual judgment criteria, the redundant and complementary attributes of structural components are clearly distinguished, avoiding component adjustment deviations caused by subjective judgment, and ensuring the objectivity and scientific nature of the adjustment logic.

[0151] For cases where the similarity is lower than the correlation benchmark, redundant structural components representing secondary data patterns are accurately identified and removed. Overlapping and non-core contributing parts in the model are eliminated, effectively simplifying the model structure, reducing computational resource consumption, and improving inference efficiency.

[0152] For components whose similarity meets the correlation benchmark, they are identified as highly complementary component pairs and their node sets and directed edges are integrated. This combines the core functions and data characterization capabilities of both components, improves the transmission path of dependencies between variables, and allows the model to capture data patterns more comprehensively and accurately.

[0153] Two targeted adjustment methods form an optimization scheme covering different component relationships. This avoids model bloat caused by redundant components and strengthens the core capabilities of the model through complementary component integration. Ultimately, the simplified probabilistic model maintains its lightweight nature while possessing better probabilistic inference performance, laying a high-quality model foundation for subsequent robust optimization.

[0154] The formula integrates key information from multiple dimensions, such as the number of nodes, connection strength weighting factors, and probability distribution similarity, to comprehensively consider the topological characteristics and probability distribution properties of structural components. This avoids similarity judgment bias caused by single-dimensional evaluation, making the measurement results more comprehensive and objective.

[0155] Introducing a connection strength weighting factor between nodes can highlight the impact of core related nodes on component similarity, weaken the interference of secondary nodes, make the similarity measurement more in line with the actual functional association logic of the components, and improve the measurement accuracy.

[0156] By quantifying the degree of fit of the conditional probability distribution of nodes through the probability distribution similarity function, the similarity of the probability transmission rules within components can be accurately captured, so that the measurement results not only reflect the structural-level correlation, but also reflect the functional-level consistency, thus enhancing the depth of similarity judgment.

[0157] Using the product of the number of nodes in two components as a normalization factor effectively eliminates the interference of component size differences on similarity measurement, ensuring fair and reasonable similarity comparison between structural components of different sizes and improving the universality of measurement results.

[0158] The formula generates accurate similarity measurement results through the organic combination and quantitative calculation of multi-dimensional parameters, providing a scientific and reliable quantitative basis for the subsequent redundancy identification, complementarity judgment and component adjustment of structural components, ensuring that the simplified probabilistic model retains its core functions and has efficient inference performance.

[0159] S6. Based on historical statistical characteristics, construct regularization constraint terms to optimize the parameters of the simplified probability model, thereby obtaining a robust probability model of the simplified probability model.

[0160] In this embodiment of the invention, the step of constructing regularization constraint terms based on historical statistical characteristics to optimize the parameters of the simplified probability model, thereby obtaining a robust probability model of the simplified probability model, includes:

[0161] Extract historical statistical features from the initial data fragment and the new data fragment;

[0162] Based on the historical statistical characteristics, construct the regularization constraint term of the simplified probability model;

[0163] The regularization constraint term is combined with the goodness-of-fit term of the simplified probabilistic model on the new data segment to obtain the comprehensive optimization guide of the simplified probabilistic model;

[0164] Based on the aforementioned comprehensive optimization approach, the conditional probability distribution parameters of the simplified probability model are iteratively adjusted to obtain a robust probability model of the simplified probability model.

[0165] A comprehensive review of all variables in the initial and new data segments is conducted, including statistical analysis of key attributes such as data distribution patterns, the strength of correlations between variables, and core value ranges. Information reflecting the long-term characteristics of the data is extracted, including the frequency distribution of variables, the synergistic occurrence patterns of different variable combinations, and the overall fluctuation trend of the data. These core features extracted from historical data constitute historical statistical features, ensuring a comprehensive coverage of the essential attributes of the data.

[0166] Based on extracted historical statistical features and combined with the network structure and parameter characteristics of the simplified probabilistic model, a regularization constraint term is constructed to constrain the model parameter values. The construction of this constraint term requires clearly defining the reasonable boundaries of parameter values ​​to ensure that parameter adjustments do not deviate from the patterns reflected in historical data. Simultaneously, it suppresses overfitting of parameters to local anomalies in new data segments, enabling the constraint term to effectively guide model parameters towards optimization in a direction consistent with historical statistical patterns.

[0167] The goodness-of-fit term of the simplified probabilistic model on the new data segment is calculated. This goodness-of-fit term quantifies the degree of fit between the model's predictions and the actual observations of the new data segment; a higher goodness-of-fit indicates better performance. The constructed regularization constraint term is organically combined with this goodness-of-fit term to balance the constraints of historical statistical patterns with the accuracy requirements of fitting the new data. This forms a comprehensive optimization guide that ensures the model closely matches the actual situation of the new data while not deviating from the overall patterns of historical data, providing clear direction for parameter adjustment.

[0168] With a focus on comprehensive optimization, an iterative adjustment process is initiated to refine the conditional probability distribution parameters of the simplified probabilistic model. Each adjustment first calculates the model's overall performance under the current parameters, including the degree of fit to new data and the degree of satisfaction with regularization constraints. Based on this overall performance, the direction and magnitude of parameter adjustment are determined. Parameter values ​​are gradually corrected, ensuring that the model improves its adaptability to new data while consistently adhering to historical statistical constraints, thus preventing parameter drift. Adjustment stops when the model's overall performance no longer significantly improves and the parameter values ​​stabilize. The resulting model is the robust probabilistic model of the simplified probabilistic model.

[0169] The beneficial effects are that historical statistical features are extracted from initial and new data fragments, comprehensively covering core attributes such as long-term data distribution patterns and variable correlation logic, providing a basis for constructing regularization constraints that fits the essence of the data, and ensuring that the direction of constraints conforms to the overall evolution trend of the data.

[0170] Regularization constraints are constructed based on historical statistical characteristics to clarify the reasonable value boundaries of parameters. This not only suppresses the overfitting of parameters to local anomalies in new data, but also prevents parameters from deviating from the patterns of historical data, thus providing a scientific constraint framework for parameter optimization.

[0171] By combining regularization constraints with goodness-of-fit terms for new data, a comprehensive optimization approach is formed, balancing the inheritance of historical patterns with the need to adapt to new data. This avoids insufficient model generalization ability caused by solely pursuing good fit, and ensures that the optimization direction takes into account both stability and accuracy.

[0172] Based on the comprehensive optimization-oriented iterative adjustment of conditional probability distribution parameters, the parameter values ​​are gradually corrected so that the model can improve the prediction accuracy of new data while always following the constraints of historical statistical characteristics, effectively avoiding parameter drift.

[0173] The resulting robust probabilistic model retains the lightweight structural advantages of the simplified model while enhancing its generalization and anti-interference capabilities through parameter optimization. It can stably adapt to changes in dynamic data streams and significantly improve the reliability and consistency of dynamic inference results.

[0174] like Figure 2The diagram shown is a functional block diagram of a dynamic reasoning system based on a Bayesian inference evolution probability model provided in an embodiment of the present invention.

[0175] The dynamic inference system 100 based on a Bayesian inference evolutionary probability model described in this invention can be installed in an electronic device. Depending on the functions implemented, the dynamic inference system 100 may include an initial model construction module 101, an incremental inference update module 102, a fitting anomaly detection module 103, a structural adaptive evolution module 104, a component simplification and fusion module 105, and a regularization robust optimization module 106. The module described in this invention can also be referred to as a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, stored in the memory of the electronic device.

[0176] In this embodiment, the functions of each module / unit are as follows:

[0177] The initial model construction module 101 is used to construct the initial network topology of the initial data segment using random variables in the initial data segment as nodes and the dependencies between variables as directed edges, and to initialize the probability distribution parameters of the nodes in the initial network topology to generate the basic probabilistic graphical model of the initial data segment.

[0178] The incremental inference update module 102 is used to acquire new data segments from the same data stream and perform variational inference on the basic probabilistic graphical model and the new data segments to obtain the updated probabilistic model of the initial data segments.

[0179] The fitting anomaly detection module 103 is used to evaluate the fitting degree of the updated probability model and the new data segment, and obtain the model fitting anomaly signal of the updated probability model.

[0180] The structure adaptive evolution module 104 is used to evaluate the degree of mismatch of the updated probability model based on the abnormal signal fitted by the model, so as to adaptively adjust the network structure of the updated probability model and obtain the evolved probability model of the updated probability model.

[0181] The component simplification and fusion module 105 is used to calculate the similarity measure of components in the evolved probability model, and remove and merge components in the evolved probability model based on the similarity measure to obtain a simplified probability model of the evolved probability model.

[0182] The regularization robust optimization module 106 is used to construct regularization constraint terms based on historical statistical characteristics to optimize the parameters of the simplified probability model, thereby obtaining a robust probability model of the simplified probability model.

[0183] In the several embodiments provided by this invention, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0184] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0185] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0186] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0187] This application embodiment can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0188] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A dynamic reasoning method based on a Bayesian inference evolutionary probability model, characterized in that, The method includes: S1. Using random variables in the initial data segment as nodes and the dependencies between variables as directed edges, construct the initial network topology of the initial data segment, initialize the probability distribution parameters of the nodes in the initial network topology, and generate the basic probabilistic graphical model of the initial data segment. S2. Obtain a new data segment from the same data stream, and perform variational inference on the basic probabilistic graphical model and the new data segment to obtain an updated probabilistic model for the initial data segment; S3. Evaluate the fit of the updated probability model and the new data segment to obtain the model fit anomaly signal of the updated probability model; S4. Based on the abnormal signal fitted by the model, evaluate the degree of mismatch of the updated probability model, so as to adaptively adjust the network structure of the updated probability model and obtain the evolved probability model of the updated probability model. S5. Calculate the similarity measure of the components in the evolved probability model, and remove and merge the components in the evolved probability model based on the similarity measure to obtain a simplified probability model of the evolved probability model. S6. Based on historical statistical characteristics, construct regularization constraint terms to optimize the parameters of the simplified probability model, thereby obtaining a robust probability model of the simplified probability model.

2. The dynamic reasoning method based on a Bayesian inference evolutionary probability model as described in claim 1, characterized in that, The process involves constructing an initial network topology for the initial data segment using random variables as nodes and directed edges representing the dependencies between variables, and initializing the probability distribution parameters of the nodes in the initial network topology to generate a basic probabilistic graphical model for the initial data segment, including: Parse the initial data segment, identify the random variables in the initial data segment, and define the random variables as nodes in the initial data segment; Analyze the statistical associations between the random variables, identify and determine the pairs of variables with significant dependencies, and define the dependencies between the pairs of variables as directed edges connecting the corresponding nodes to generate the initial network topology of the initial data fragment. Based on the observation statistics and domain prior knowledge of the node in the initial data segment, the conditional probability distribution parameters of the node are initialized; The initial network topology is combined with the conditional probability distribution parameters to obtain the basic probabilistic graphical model of the initial data segment.

3. The dynamic reasoning method based on a Bayesian inference evolutionary probability model as described in claim 1, characterized in that, The process of acquiring a new data segment from the same data stream and performing variational inference on the basic probabilistic graphical model and the new data segment to obtain an updated probabilistic model for the initial data segment includes: Select an exponential family distribution that is compatible with the conditional distribution of nodes in the basic probabilistic graphical model to generate the variational distribution family of the basic probabilistic graphical model; Based on the observed state of the variables in the new data segment, determine the initial variational parameters of the variational distribution family; Based on the prior distribution of the basic probabilistic graphical model and the new data segment, a divergence metric between the variational distribution family and the posterior distribution is defined, and the initial variational parameters are iteratively adjusted with the goal of minimizing the divergence metric. When the change in the initial variational parameters is lower than the preset convergence threshold, the variational inference process is determined to be complete, and the expected parameters of the variational distribution family are assigned to the corresponding node of the update probability model to obtain the update probability model of the initial data segment.

4. The dynamic reasoning method based on a Bayesian inference evolutionary probability model as described in claim 1, characterized in that, The step of evaluating the fit between the updated probability model and the new data segment to obtain anomaly signals in the model fit of the updated probability model includes: Based on the updated probability model, probabilistic inference is performed on the variables in the new data segment to obtain the model prediction distribution of the new data segment; The observed values ​​of variables in the new data segment are compared with the expected values ​​of the model's predicted distribution to construct the residual vector of the updated probability model; Based on the degree of difference in the dimensions of the residual vector, the overall fit between the updated probability model and the new data segment is evaluated. The overall fit is compared with a preset anomaly detection threshold. When the overall fit is less than the preset anomaly detection threshold, a model fit anomaly signal for the updated probability model is generated.

5. The dynamic reasoning method based on a Bayesian inference evolutionary probability model as described in claim 1, characterized in that, The step of assessing the mismatch degree of the updated probability model based on the fitted abnormal signal to adaptively adjust the network structure of the updated probability model, thereby obtaining the evolved probability model of the updated probability model, includes: The abnormal signal of the model fitting is analyzed to obtain the specific nodes and specific directed edges in the updated probability model that are mismatched with the new data segment, and the specific nodes and specific directed edges are used as the structural elements to be evaluated in the updated probability model. Assess the overall degree of mismatch between the structuring element to be evaluated and the dependencies between variables reflected by the new data fragment; Based on the overall degree of mismatch, determine the type of adjustment operation for the structural element to be evaluated; Based on the adjustment operation type, the network structure of the updated probability model is modified to obtain the evolved probability model of the updated probability model.

6. The dynamic reasoning method based on a Bayesian inference evolutionary probability model as described in claim 1, characterized in that, The process of calculating the similarity measure of components in the evolved probabilistic model, and removing and merging components based on the similarity measure to obtain a simplified probabilistic model of the evolved probabilistic model, includes: Based on the network structure of the evolved probability model, identify the structural components in the network structure; Calculate the similarity measure between the structural components; Based on the similarity metric, the redundancy and complementarity between the structural components are evaluated, and based on the evaluation results, the structural components are adjusted to obtain a simplified probability model of the evolved probability model.

7. The dynamic reasoning method based on a Bayesian inference evolutionary probability model as described in claim 6, characterized in that, The component adjustments include: Based on the similarity metric, if the similarity between the first structural component and the second structural component in the evolved probabilistic model is lower than the correlation benchmark required by the evolved probabilistic model, then the components representing secondary data patterns in the first structural component and the second structural component are identified as redundant structural components and removed. When the similarity between the first structural component and the second structural component in the evolved probabilistic model meets the correlation benchmark required by the evolved probabilistic model, the first structural component and the second structural component are determined as a highly complementary component pair, and the node sets of the first structural component and the second structural component are integrated with the directed edges.

8. The dynamic reasoning method based on a Bayesian inference evolutionary probability model as described in claim 6, characterized in that, The formula for calculating the similarity measure is as follows: ； In the formula, For structural components and structural components Similarity measurement between For components The number of nodes included. For components The number of nodes included. For nodes With nodes Connection strength weighting factor between Let be the probability distribution similarity function. For nodes The set of parent nodes, For nodes The set of parent nodes, For nodes Conditional probability distribution For nodes Conditional probability distribution.

9. The dynamic reasoning method based on a Bayesian inference evolutionary probability model as described in claim 1, characterized in that, The step of constructing regularization constraint terms based on historical statistical characteristics to optimize the parameters of the simplified probabilistic model, thereby obtaining a robust probabilistic model of the simplified probabilistic model, includes: Extract historical statistical features from the initial data fragment and the new data fragment; Based on the historical statistical characteristics, a regularization constraint term is constructed for the simplified probability model; By combining the regularization constraint term with the goodness-of-fit term of the simplified probabilistic model on the new data segment, a comprehensive optimization guide for the simplified probabilistic model is obtained. Based on the aforementioned comprehensive optimization approach, the conditional probability distribution parameters of the simplified probability model are iteratively adjusted to obtain a robust probability model of the simplified probability model.

10. A dynamic reasoning system based on a Bayesian inference evolution probability model, characterized in that, The system is used to implement the dynamic reasoning method based on a Bayesian inference evolutionary probability model as described in claim 1, the system comprising: The initial model building module is used to construct the initial network topology of the initial data segment using random variables in the initial data segment as nodes and the dependencies between variables as directed edges, and to initialize the probability distribution parameters of the nodes in the initial network topology to generate the basic probabilistic graphical model of the initial data segment. The incremental inference update module is used to acquire new data segments from the same data stream and perform variational inference on the basic probabilistic graphical model and the new data segments to obtain the updated probabilistic model of the initial data segments. The fitting anomaly detection module is used to evaluate the fit between the updated probability model and the new data segment, and obtain the model fitting anomaly signal of the updated probability model. The structural adaptive evolution module is used to assess the degree of mismatch of the updated probability model based on the abnormal signal fitted by the model, so as to adaptively adjust the network structure of the updated probability model and obtain the evolved probability model of the updated probability model. The component simplification and fusion module is used to calculate the similarity measure of components in the evolved probability model, and remove and merge components in the evolved probability model based on the similarity measure to obtain a simplified probability model of the evolved probability model. The regularization robust optimization module is used to construct regularization constraint terms based on historical statistical characteristics to optimize the parameters of the simplified probability model, thereby obtaining a robust probability model of the simplified probability model.

Images