Enterprise data analysis method and system based on big data mining
By constructing a constraint diagram and generating constraint residual data, the problem of misjudgment in big data mining systems when statistical definitions change was solved, enabling the distinction between rule-level changes and real business anomalies, and improving the stability and traceability of data mining results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SICHUAN RUIFANG TECH CO LTD
- Filing Date
- 2026-02-28
- Publication Date
- 2026-06-05
Smart Images
Figure CN122152912A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of big data mining technology, and more specifically to a method and system for enterprise data analysis based on big data mining. Background Technology
[0002] With the continuous growth of enterprise data scale, big data mining technology has been widely applied in scenarios such as sales anomaly identification, cost fluctuation analysis, risk monitoring, and business forecasting. In actual engineering deployments, data mining models typically rely on historical business data for training and perform anomaly identification or trend prediction calculations on newly added data during daily operation. To ensure model stability, engineering practice usually requires that the input data have a relatively stable statistical distribution and consistent data caliber.
[0003] However, in real-world business environments, data definitions are not fixed indefinitely. Adjustments to financial statistical definitions, changes in unit conversion methods, modifications to field derivation rules, and adjustments to data granularity can all cause abrupt changes in data distribution without altering the underlying business operations. For example, a change in the amount field from "yuan" to "ten thousand yuan," a change in sales revenue statistics from "including tax" to "excluding tax," or a change in the daily reporting standard to a weekly cumulative standard. These rule-level changes typically occur during system configuration or statistical specification adjustments and are not explicitly reflected in the business data.
[0004] Existing big data mining systems typically perform only basic cleaning processes before data enters the model, such as missing value imputation, outlier truncation, and simple consistency checks. They cannot identify rule-level changes such as "changes in statistical definitions." When the rule level changes, the data exhibits abrupt changes in its numerical distribution. The model often misinterprets these changes as business anomalies or risk signals, leading to incorrect warnings or misleading decision-making results.
[0005] Furthermore, existing technologies lack a mechanism for linking and controlling the "parsing rule version" with the "model mining results." The parsing process and the mining process are usually designed separately. When the parsing rules change, the model does not perceive the change and continues to perform anomaly identification calculations according to the original distribution assumptions, resulting in incorrect updates of model parameters and even amplifying the bias caused by rule layer changes in subsequent training phases.
[0006] Therefore, in enterprise big data mining application scenarios, how to identify changes in statistical standards during the data parsing stage and distinguish between rule-level changes and real business anomalies during the mining stage has become a technical problem that urgently needs to be solved in engineering practice. Summary of the Invention
[0007] The purpose of this invention is to provide a method and system for enterprise data parsing based on big data mining, so as to at least solve the technical problem that data distribution deviations caused by changes in statistical standards are misidentified as business anomalies by the mining model.
[0008] To achieve the above objectives, the first aspect of the present invention provides a method for enterprise data parsing based on big data mining. The method includes: acquiring multi-source business data, field dictionary data, and current parsing version data of an enterprise, and constructing a constraint graph based on the field dictionary data; generating a constraint set based on the current parsing version data, performing constraint calculations on the multi-source business data of the enterprise using the constraint set to generate constraint residual data, and calculating a caliber drift judgment result based on the constraint residual data; updating the current parsing version data when the caliber drift judgment result meets preset conditions, and performing parsing mapping processing on the multi-source business data of the enterprise based on the current parsing version data to generate mining sample data; inputting the mining sample data, together with the current parsing version data and constraint residual features generated based on the constraint residual data, into a preset big data mining model to perform anomaly identification calculations and generate a business anomaly judgment result; performing result separation processing based on the business anomaly judgment result and the caliber drift judgment result, and outputting a business anomaly judgment result or a rule layer change judgment result.
[0009] Optionally, the rules for constructing a caliber constraint graph based on the field dictionary data are as follows: extract field identifier information, field unit identifier information, and field lineage information from the field dictionary data; establish a set of field nodes using the field identifier information as field nodes, and establish a set of unit nodes using the field unit identifier information as unit nodes; establish a set of rule edges between the field nodes based on the field lineage information, the set of rule edges being used to represent the derivation relationship between fields; establish a set of unit mapping edges between the field nodes and the unit nodes based on the field unit identifier information, the set of unit mapping edges being used to represent the conversion relationship between fields and units; combine the set of field nodes, the set of unit nodes, the set of rule edges, and the set of unit mapping edges to form the caliber constraint graph.
[0010] Optionally, a constraint set is generated based on the current parsed version data, and constraint calculations of the constraint set are performed on the enterprise's multi-source business data to generate constraint residual data. This includes: extracting constraint identification information, field mapping information, and constraint expression information from the current parsed version data; locating corresponding field nodes and rule edges in the caliber constraint graph based on the constraint identification information and the field mapping information, and constructing a constraint function set based on the constraint expression information; converting the enterprise's multi-source business data into a sequence of field values corresponding to the constraint function set according to the field mapping information; performing function calculations of the constraint function set on the field value sequence to generate calculation result values corresponding to each constraint function; and performing deviation calculations based on the calculation result values and preset constraint conditions to generate the constraint residual data.
[0011] Optionally, calculating the caliber drift determination result based on the constraint residual data includes: sorting the constraint residual data according to constraint identifiers by time to generate a constraint residual sequence; calculating a residual aggregate value sequence based on the constraint residual sequence, wherein the residual aggregate value sequence is obtained by performing aggregation calculation on the constraint residual sequence within a preset time window; calculating a residual baseline value and a residual fluctuation value based on the residual aggregate value sequence; comparing the residual aggregate value sequence with the residual baseline value and the residual fluctuation value to generate a drift determination identifier sequence; and generating the caliber drift determination result based on the drift determination identifier sequence.
[0012] Optionally, updating the current parsing version data when the caliber drift determination result meets preset conditions includes: calculating the residual contribution value of each constraint according to the constraint identifier based on the constraint residual data, and selecting constraints with residual contribution values greater than the average residual contribution value to form a high contribution constraint set; regenerating the constraint expression set based on the field attribute information corresponding to the high contribution constraint set in the field dictionary data, and combining the constraint expression set to form candidate parsing version data; performing replay calculation on the enterprise multi-source business data based on the candidate parsing version data to generate candidate constraint residual data, and calculating the candidate residual aggregation value; when the candidate residual aggregation value is less than the current residual aggregation value, replacing the current parsing version data with the candidate parsing version data.
[0013] Optionally, performing parsing mapping processing on the enterprise's multi-source business data based on the current parsed version data to generate mining sample data includes: extracting field mapping relationships and constraint expression information from the current parsed version data, and performing field rearrangement processing on the enterprise's multi-source business data according to the field mapping relationships to generate mapped field data; performing unit conversion calculations and field derivation calculations on the mapped field data based on the constraint expression information to generate standardized field data; performing missing value imputation processing and outlier correction rule calculations on the standardized field data to generate corrected field data; and combining the corrected field data according to a preset feature structure to form the mining sample data.
[0014] Optionally, the mined sample data, together with the current parsed version data and the constraint residual features generated based on the constraint residual data, are input into a preset big data mining model to perform anomaly identification calculation and generate a business anomaly judgment result. This includes: constructing a sample feature vector based on the mined sample data and constructing a residual feature vector based on the constraint residual features; generating a version identifier vector based on the current parsed version data and concatenating the version identifier vector with the sample feature vector and the residual feature vector to generate a model input vector; inputting the model input vector into the preset big data mining model to perform anomaly scoring calculation and generate anomaly score values; and comparing the anomaly score values with anomaly judgment threshold to generate the business anomaly judgment result.
[0015] Optionally, the preset big data mining model includes: a feature encoding unit, used to perform encoding calculations on the sample feature vector to generate a sample encoding vector; a residual modulation unit, used to perform feature modulation calculations on the sample encoding vector based on the residual feature vector to generate a modulation encoding vector; a version normalization unit, used to select a normalization parameter corresponding to the currently parsed version data according to the version identifier vector and perform version normalization calculations on the modulation encoding vector to generate a normalized vector; and an anomaly scoring unit, used to perform scoring calculations based on the normalized vector to generate the anomaly score value.
[0016] Optionally, based on the business anomaly determination result and the caliber drift determination result, separate processing is performed to output a business anomaly determination result or a rule layer change determination result, including: performing time alignment processing on the business anomaly determination result and the caliber drift determination result to generate aligned determination data; constructing a determination combination identifier based on the aligned determination data, wherein the determination combination identifier is formed by combining a business anomaly identifier and a caliber drift identifier; performing classification calculation based on the determination combination identifier to generate a business anomaly output identifier or a rule layer change output identifier; and generating a final output result based on the business anomaly output identifier or the rule layer change output identifier.
[0017] A second aspect of the present invention provides an enterprise data parsing system based on big data mining, characterized in that the system comprises: a data acquisition unit, used to acquire multi-source business data, field dictionary data, and current parsing version data of an enterprise, and construct a caliber constraint graph based on the field dictionary data; a drift determination unit, used to generate a constraint set based on the current parsing version data, perform constraint calculations on the multi-source business data of the enterprise using the constraint set to generate constraint residual data, and calculate a caliber drift determination result based on the constraint residual data; a sample generation unit, used to update the current parsing version data when the caliber drift determination result meets preset conditions, and perform parsing mapping processing on the multi-source business data of the enterprise based on the current parsing version data to generate mining sample data; an anomaly determination unit, used to input the mining sample data together with the current parsing version data and constraint residual features generated based on the constraint residual data into a preset big data mining model to perform anomaly identification calculations and generate a business anomaly determination result; and a result output unit, used to perform result separation processing based on the business anomaly determination result and the caliber drift determination result, and output a business anomaly determination result or a rule layer change determination result.
[0018] Through the above technical solution, this invention constructs a constraint graph and performs constraint calculations on multi-source business data of an enterprise. This solution can quantitatively identify rule layer offsets caused by changes in statistical caliber and promptly update the current parsing version data when caliber drift is detected, ensuring that the data entering the big data mining model remains consistent with the parsing version. Based on this, parsing version information and constraint residual features are introduced into the anomaly identification calculation process, and the business anomaly judgment results and caliber drift judgment results are separated. This achieves the ability to distinguish between rule layer changes and actual business anomalies, avoiding interference from statistical caliber adjustments on anomaly identification results and improving the stability and traceability of enterprise mining results.
[0019] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description
[0020] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the steps of an enterprise data parsing method based on big data mining provided by one embodiment of the present invention; Figure 2 This is a system architecture diagram of an enterprise data parsing system based on big data mining provided by one embodiment of the present invention. Detailed Implementation
[0021] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for illustration and explanation of the embodiments of this application and are not intended to limit the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.
[0022] It should be noted that if the embodiments of this application involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.
[0023] Furthermore, if the embodiments of this application involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed in this application.
[0024] like Figure 1 As shown, this invention provides a method for enterprise data parsing based on big data mining, the method comprising: Step S10: Obtain enterprise multi-source business data, field dictionary data, and current parsing version data, and construct a caliber constraint diagram based on the field dictionary data.
[0025] Specifically, the rules for constructing a caliber constraint graph based on the field dictionary data are as follows: extract field identifier information, field unit identifier information, and field lineage information from the field dictionary data; establish a set of field nodes using the field identifier information as field nodes, and establish a set of unit nodes using the field unit identifier information as unit nodes; establish a set of rule edges between the field nodes based on the field lineage information, the set of rule edges being used to represent the derivation relationship between fields; establish a set of unit mapping edges between the field nodes and the unit nodes based on the field unit identifier information, the set of unit mapping edges being used to represent the conversion relationship between fields and units; combine the set of field nodes, the set of unit nodes, the set of rule edges, and the set of unit mapping edges to form the caliber constraint graph.
[0026] In this embodiment of the invention, enterprise multi-source business data typically originates from different business systems, such as sales, finance, inventory, or reporting systems. These systems differ in field naming conventions, units of measurement, and statistical definitions. To avoid abrupt distribution changes due to inconsistent definitions during subsequent data parsing and mining stages, it is necessary to formally model the structural relationships between fields before the data enters the parsing process. Therefore, this solution constructs a definition constraint graph based on field dictionary data, transforming the semantic relationships and unit rules between fields into a computable graph structure.
[0027] The field dictionary data contains at least field identifier information, field unit identifier information, and field lineage information. Field identifier information uniquely identifies a business field, such as "sales amount," "return amount," and "tax-inclusive sales amount." Field unit identifier information identifies the unit of measurement corresponding to the field, such as "yuan," "ten thousand yuan," "piece," and "ton." Field lineage information describes the derivation rules or calculation relationships between fields, such as "net sales amount = sales amount - return amount" and "tax amount = net sales amount × tax rate." All of the above information originates from the field dictionary data itself and does not rely on external manual input, thus ensuring the repeatability of the diagram construction process.
[0028] When constructing the caliber constraint graph, a set of field nodes is established based on field identification information, with each field node corresponding to a specific field instance in the enterprise data. A set of unit nodes is established based on field unit identification information, with each unit node corresponding to a specific unit of measurement. A set of rule edges is established between field nodes based on field lineage information, used to represent the derivation or combination relationships between fields; a set of unit mapping edges is established between field nodes and unit nodes based on field unit identification information, used to represent the mapping or conversion relationship between fields and units of measurement. The above set of field node nodes, unit node nodes, rule edge sets, and unit mapping edge sets combine to form the caliber constraint graph, which can fully express the computational dependencies and unit conversion paths between fields at the structural level.
[0029] For example, when the "Sales Amount" field in the sales system uses "ten thousand yuan" as the unit, while the "Sales Revenue" field in the financial system uses "yuan" as the unit, the unit mapping edge can explicitly express the conversion relationship between "ten thousand yuan" and "yuan". When the "Net Sales Amount" field in the report is calculated from "Sales Amount" and "Return Amount", the rule edge can express this derivation path. By structurally expressing the above relationships as a graph model, a clear structural foundation can be provided for the subsequent generation of constraint sets. The caliber constraint graph not only limits the scope of computational dependencies between fields, but also limits the effective path of unit conversion. The protection scope covers the three types of nodes and their edge structures—fields, units, and rules—built based on field dictionary data in any enterprise's multi-source business data environment, and is not limited to a specific industry or specific field name.
[0030] Step S20: Generate a constraint set based on the current parsed version data, perform constraint calculations on the enterprise multi-source business data to generate constraint residual data, and calculate the caliber drift judgment result based on the constraint residual data.
[0031] Specifically, the process involves generating a constraint set based on the current parsed version data, and performing constraint calculations on the enterprise's multi-source business data to generate constraint residual data. This includes: extracting constraint identification information, field mapping information, and constraint expression information from the current parsed version data; locating corresponding field nodes and rule edges in the caliber constraint graph based on the constraint identification information and the field mapping information, and constructing a constraint function set based on the constraint expression information; converting the enterprise's multi-source business data into a sequence of field values corresponding to the constraint function set according to the field mapping information; performing function calculations on the constraint function set on the field value sequence to generate calculation result values for each constraint function; and performing deviation calculations based on the calculation result values and preset constraint conditions to generate the constraint residual data.
[0032] Furthermore, the calculation of the caliber drift determination result based on the constraint residual data includes: sorting the constraint residual data according to constraint identifiers by time to generate a constraint residual sequence; calculating a residual aggregate value sequence based on the constraint residual sequence, wherein the residual aggregate value sequence is obtained by performing aggregation calculation on the constraint residual sequence within a preset time window; calculating a residual baseline value and a residual fluctuation value based on the residual aggregate value sequence; comparing the residual aggregate value sequence with the residual baseline value and the residual fluctuation value to generate a drift determination identifier sequence; and generating the caliber drift determination result based on the drift determination identifier sequence.
[0033] In this embodiment of the invention, the process of generating a constraint set based on the current parsed version data and performing constraint calculations on the enterprise's multi-source business data to form constraint residual data, and further calculating the caliber drift determination result is as follows.
[0034] At the point of time Below, enterprise multi-source business data is transformed into a field value vector after being mapped by field relationships: ; in, This indicates the number of fields involved in the constraint calculation. Indicates the first The fields at the time point The value of . The current parsed version of the data contains constraint identifier information and corresponding constraint expression information. For the . Each constraint is defined, and its constraint function is constructed as follows: ; in, , This represents the total number of constraints in the constraint set. (Constraint function) The function is determined by the field derivation relationships or unit conversion relationships defined in the current parsed version of the data. For example, when there is a derivation relationship between fields, "Net Sales Amount = Sales Amount - Return Amount", its function form can be expressed as: ; For example, when there is a unit conversion relationship "Sales amount (yuan) = Sales amount (ten thousand yuan)\timesk", its function form is: ; in, This indicates the unit conversion factor.
[0035] field value vector Substituting each constraint function in turn, we obtain the function calculation results: ; To quantify the degree of deviation, define the constraint residuals: ; in, Indicates the first The reference field value used as the baseline in the constraint. This serves as a stabilizing factor. This forms the time point. The constrained residual vector is as follows: ; The constrained residual data were then sorted chronologically within a preset time window. Perform sliding aggregation calculations on each constraint residual: ; At the same time point The overall aggregated residual value is obtained by averaging the aggregated residuals of all constraints: ; During the baseline period of stable system operation Within, calculate the residual baseline value: ; And calculate the residual fluctuation value: ; in, This indicates the number of time points within the base time period.
[0036] During the real-time operation phase, the current time point When compared with the baseline statistic, if the following conditions are met: ; At that time, a drift determination identifier is generated. ;otherwise ,in The determination coefficient is used. The final caliber drift determination result is generated based on the determination identifier sequence within a continuous time period.
[0037] Through the above continuous execution process, the constraint expression is determined by the current parsed version data, the constraint residual is calculated by the enterprise's multi-source business data, and the drift judgment result is derived from the residual statistics. The calculation path is coherent, the parameter source is clear, and it can quantitatively identify the rule layer offset caused by changes in statistical caliber.
[0038] In the above execution process, the construction of constraint functions and the calculation of residuals are not limited to linear relationships. Although the embodiments use addition and subtraction operations or unit conversion relationships as examples, the constraint expression information can include any computable algebraic expression, proportional relationship, conditional piecewise function, or matrix transformation relationship, as long as the expression can be expressed in functional form. Furthermore, the evaluation can be performed by mapping the fields in the enterprise's multi-source business data with the field mapping information in the current parsed version of the data, all of which fall within the protection scope of this invention. For example, when there are multiplication, division, weighted summation, or exponential relationships between fields, their constraint function can be expressed as follows: ,in To parse the function expressions defined in the version data.
[0039] For the residual calculation method, this implementation method describes it in the form of normalized absolute deviation, that is: ; However, the definition of residuals is not limited to this form. Without changing the core idea of quantifying the degree of deviation using constraint functions, the squared error form can also be used. The residuals can be calculated in logarithmic or relative error form. The aggregation of residuals can also employ the average, median, weighted summation, or exponential weighting methods. As long as the aggregation method can statistically summarize the residual data within the time window and form comparable statistics, it falls within the technical scope of this invention.
[0040] During the drift determination phase, the baseline statistic... With volatility The calculation can also be implemented based on different statistical methods, such as using the median and absolute deviation, quantile intervals, or adaptive threshold models. As long as a stable period statistical benchmark is constructed based on the residual aggregate value sequence, and the current residual level is compared and judged accordingly, it is an equivalent replacement for the core judgment mechanism of this invention.
[0041] Therefore, the scope of protection of this invention is not limited to a specific mathematical formula or a fixed statistical model, but covers the complete technical path of determining constraint functions based on analytical versions, calculating constraint residual data using multi-source business data of enterprises, and generating caliber drift determination results by performing statistical determination on the residual time series. Any implementation method that makes equivalent transformations to the form of constraint functions, the definition of residuals, or the statistical determination method within this technical path framework should fall within the scope of protection of this invention.
[0042] Step S30: When the caliber drift determination result meets the preset conditions, update the current parsing version data, and perform parsing mapping processing on the enterprise multi-source business data based on the current parsing version data to generate mining sample data.
[0043] Specifically, updating the current parsing version data when the caliber drift determination result meets preset conditions includes: calculating the residual contribution value of each constraint according to the constraint identifier based on the constraint residual data, and selecting constraints with residual contribution values greater than the average residual contribution value to form a high contribution constraint set; regenerating the constraint expression set based on the field attribute information corresponding to the high contribution constraint set in the field dictionary data, and combining the constraint expression set to form candidate parsing version data; performing replay calculation on the enterprise multi-source business data based on the candidate parsing version data to generate candidate constraint residual data, and calculating the candidate residual aggregation value; when the candidate residual aggregation value is less than the current residual aggregation value, replacing the current parsing version data with the candidate parsing version data.
[0044] Furthermore, based on the current parsed version data, parsing and mapping processing is performed on the enterprise's multi-source business data to generate mining sample data. This includes: extracting field mapping relationships and constraint expression information from the current parsed version data, and performing field rearrangement processing on the enterprise's multi-source business data according to the field mapping relationships to generate mapped field data; performing unit conversion calculations and field derivation calculations on the mapped field data based on the constraint expression information to generate standardized field data; performing missing value imputation processing and outlier correction rule calculations on the standardized field data to generate corrected field data; and combining the corrected field data according to a preset feature structure to form the mining sample data.
[0045] In this embodiment of the invention, when the caliber drift determination result calculated based on the constraint residual data meets the preset conditions, the system enters the parsing version update process. This update process is not a simple replacement of the rule file, but is executed in three stages: residual contribution analysis, candidate version generation, and replay verification, to ensure that the update of the parsing version data has a calculable basis and a verifiable path.
[0046] At the point of time Below, the constrained residual vector has been obtained. First, the constraint residuals are aggregated within a preset window to obtain the constraint aggregated residuals. Definition of the first The residual contribution value of the constraint is: ; in, Indicates the first The relative contribution of each constraint at the current time. Calculate the mean residual contribution: ; Select the one that satisfies The constraints form a set of high-contribution constraints. This set is used to locate the field ranges where rules may change. Subsequently, data related to the set is extracted from the field dictionary data. The constraint's corresponding field attribute information includes field identifier, unit identifier, and lineage information. Based on the above information, the constraint expression set is reconstructed to form a candidate constraint expression set. And based on this, candidate parsing version data is generated.
[0047] To verify the validity of the candidate parsing version data, replay calculations were performed on the enterprise's multi-source business data within a historical time interval. Let the replay time interval be... Candidate constraint functions are constructed using candidate parsing version data. Calculate the candidate constraint residuals: ; Further calculation of candidate residual aggregate values: ; in This is the window aggregation value of the candidate residuals. When the following conditions are met... If a candidate parsing version is determined to be statistically superior to the current parsing version, the system will replace the current parsing version with the candidate parsing version and record the version switch time to ensure that the version evolution path is traceable.
[0048] After the parsed version data update is complete, the parsing mapping process begins. The current parsed version data contains field mapping relationships and constraint expression information. Based on the field mapping relationships, the fields of the enterprise's multi-source business data are rearranged to generate mapped field data. The field rearrangement process ensures that fields from different source systems are aligned according to a unified field identifier.
[0049] Then, based on the constraint expression information, unit conversion calculations and field derivation calculations are performed on the mapped field data. For example, when the field unit changes, a unit conversion factor is used. implement: ; When fields have a derivation relationship, field calculations are performed based on the constraint function to generate standardized field data. .
[0050] Based on standardized field data, missing value imputation and outlier correction rules are calculated. Missing value imputation can be based on historical mean or nearby values: ; in These are replacement values calculated based on preset imputation rules. Outlier correction can be adjusted through threshold truncation or distribution constraints to ensure that field values fall within a reasonable range.
[0051] The corrected field data will be combined according to a preset feature structure to form the mining sample data. The preset feature structure defines the feature arrangement order and dimensional structure, ensuring that the input dimensions of the mining model remain stable.
[0052] In other implementations, a field dependency path sensitivity analysis mechanism is introduced during the parsing version update and parsing mapping process to address the cascading impact of statistical caliber changes on cross-system field propagation paths. In real-world engineering environments, changes to the calculation rules of a field often affect not only the field itself but also multiple downstream derived fields along the field lineage path. For example, when the unit for "sales amount" is changed from "yuan" to "ten thousand yuan," it not only affects "net sales amount" but also multiple levels of derived indicators such as "gross profit margin" and "sales per employee."
[0053] In this implementation, a field dependency matrix is constructed based on the caliber constraint graph. ,in Representation field Dependency fields Otherwise, the value is 0. For fields in the high-contribution constraint set, the set of reachable paths in the dependency matrix is calculated, and the affected field set is constructed. During the parsing version update, not only are the constraint expressions corresponding to the high-contribution constraint set regenerated, but the constraint expression reconstruction is also performed on all derived paths corresponding to the affected field set. Subsequently, in the replay calculation phase, candidate residual verification is performed only on the data subspace corresponding to the affected field set, without performing replay on all fields, thereby reducing computational complexity.
[0054] Step S40: Input the mined sample data, the current parsed version data, and the constraint residual features generated based on the constraint residual data into a preset big data mining model to perform anomaly identification calculation and generate a business anomaly judgment result.
[0055] Specifically, a sample feature vector is constructed based on the mined sample data, and a residual feature vector is constructed based on the constrained residual features; a version identifier vector is generated based on the current parsed version data, and the version identifier vector is concatenated with the sample feature vector and the residual feature vector to generate a model input vector; the model input vector is input into the preset big data mining model to perform anomaly scoring calculation and generate anomaly score values; the anomaly score values are compared with anomaly judgment thresholds to generate the business anomaly judgment result.
[0056] Furthermore, the preset big data mining model includes: a feature encoding unit, used to perform encoding calculations on the sample feature vector to generate a sample encoding vector; a residual modulation unit, used to perform feature modulation calculations on the sample encoding vector based on the residual feature vector to generate a modulation encoding vector; a version normalization unit, used to select a normalization parameter corresponding to the currently parsed version data according to the version identifier vector and perform version normalization calculations on the modulation encoding vector to generate a normalized vector; and an anomaly scoring unit, used to perform scoring calculations based on the normalized vector to generate the anomaly score value.
[0057] In this embodiment of the invention, the mined sample data is not directly input into a conventional anomaly detection model, but rather participates in the model calculation together with the currently parsed version data and constraint residual features, enabling the anomaly identification process to have version awareness and rule residual modulation capabilities. The entire anomaly identification execution process is as follows.
[0058] At the point of time Next, the parsing mapping process generates the mining sample data vector: ; in, To extract the number of features, Indicates the first These features at a given time point The value of is determined. A sample feature vector is constructed based on the mined sample data. .
[0059] Simultaneously, residual feature vectors are constructed based on the constrained residual data: ; in, Indicates the first Constraints at a point in time The residual value, To constrain the quantity, the currently parsed version data contains version identifier information, which is encoded as a version identifier vector: ; in, This is the version encoding dimension. The version identifier vector can be generated using one-hot encoding or numerical mapping, and is used to represent the parsed version corresponding to the current data.
[0060] The sample feature vector, residual feature vector, and version identifier vector are concatenated to form the model input vector: ; The symbols This indicates a vector concatenation operation.
[0061] The preset big data mining model is a version-aware anomaly identification model, whose internal structure includes a feature encoding unit, a residual modulation unit, a version normalization unit, and an anomaly scoring unit.
[0062] The feature encoding unit performs encoding calculations on the sample feature vector: ; in, The feature encoding weight matrix, For bias vectors, It is a non-linear activation function. This is the sample encoding vector.
[0063] The residual modulation unit performs modulation calculations on the sample encoding vector using the residual feature vector: ; in, The residual mapping matrix is... For modulation function, This represents element-wise multiplication. This is the modulation and coding vector. This process causes the constraint residuals to have a dynamic weighting effect on the principal features.
[0064] The version normalization unit selects the corresponding normalization parameter based on the version identifier vector. and And perform normalization calculation on the modulation and coding vector: ; in, and This is the set of normalized parameters corresponding to the current parsed version of the data.
[0065] The anomaly scoring unit performs scoring calculations based on normalized vectors: ; in, For the rating weight vector, For the rating bias, This is an abnormal score.
[0066] Compare the anomaly score with the anomaly determination threshold Comparison: ; in, This indicates the generation of business anomaly determination results. This indicates that the business has not been identified as abnormal.
[0067] Through the above calculation process, sample features, constraint residual features, and parsing version information are structurally coupled within the model. The sample encoding vector originates from the mined sample data, the residual modulation vector originates from the constraint residual data, the version normalization parameter originates from the current parsing version data, and the anomaly score is calculated from the normalized vector, ultimately generating the business anomaly determination result. The entire execution path has clear correspondences at the data source, variable definition, and function operation levels, avoiding uncontrollable deviations in the model when the rule layer changes, and maintaining the stability and scalability of the mining calculation structure.
[0068] Furthermore, regarding the pre-defined big data mining model proposed in this invention, the pre-defined big data mining model is not limited to a specific network structure, but refers to a version-aware anomaly identification model that can receive sample feature vectors, residual feature vectors, and version identifier vectors and perform joint calculations. The core of this model lies in introducing parsed version information and constraint residual information in a structured manner, enabling the model to distinguish between rule-layer fluctuations and business-layer anomalies during parameter update and inference phases.
[0069] In practical implementation, the preset big data mining model can adopt a multi-layer feedforward neural network structure, a recurrent neural network structure, or a graph neural network structure. Regardless of the structure used, its input layer includes three types of feature channels: sample feature channel, residual feature channel, and version identifier channel. The sample feature channel is used to express the statistical attributes of the business data itself; the residual feature channel is used to express the degree of deviation from the rule layer constraints; and the version identifier channel is used to express the parsing version to which the current data belongs.
[0070] During the model training phase, a loss function can be constructed based on historical labeled data: ; in, This represents the anomaly detection loss function. For real labels, This represents the model's prediction results. During training, the version normalization unit dynamically selects normalization parameters based on the version identifier vector, ensuring that data from different parsing versions maintain a stable distribution in the feature space. The residual modulation unit dynamically adjusts the weights of the feature representations, causing samples with significant deviations from the constraints to be explicitly labeled within the model.
[0071] During the inference phase, when the parsing version is updated, only the version identifier vector and the normalization parameter set need to be updated, without retraining all model parameters. This design enables the model to adapt to changes in the parsing version. Any anomaly identification model based on a multi-channel input structure that uses version identifiers and residual features to participate in internal encoding, normalization, or scoring calculations falls under the implementation of the pre-defined big data mining model of this invention, and its protection scope is not limited to a specific number of neural network layers or a specific activation function form.
[0072] Step S50: Based on the business anomaly determination result and the caliber drift determination result, perform result separation processing and output the business anomaly determination result or the rule layer change determination result.
[0073] Specifically, the business anomaly determination result and the caliber drift determination result are time-aligned to generate aligned determination data; a determination combination identifier is constructed based on the aligned determination data, wherein the determination combination identifier is formed by combining the business anomaly identifier and the caliber drift identifier; a classification calculation is performed based on the determination combination identifier to generate a business anomaly output identifier or a rule layer change output identifier; and the final output result is generated based on the business anomaly output identifier or the rule layer change output identifier.
[0074] In this embodiment of the invention, after the anomaly identification calculation yields the business anomaly determination result, it is not directly used as the final output. Instead, it is combined with the caliber drift determination result to perform result separation processing. This separation processing is used to distinguish between anomalies caused by actual business changes and rule-level changes caused by changes in statistical calibers or field rules, thereby avoiding misjudging rule adjustments as business anomalies.
[0075] At the point of time Let the result of the business anomaly determination be: ,in: ; Meanwhile, the caliber drift determination result is recorded as ,in: ; To ensure consistency between the two judgment results in the time dimension, time alignment processing is first performed on the business anomaly judgment result sequence and the caliber drift judgment result sequence. Specifically, based on the timestamp... The two sequences are matched to form an alignment determination dataset: ; This alignment determination dataset is used to construct the determination combination identifier. Determination combination identifier Defined as an ordered combination of business anomaly indicators and caliber drift indicators: ; in, Different values correspond to different combinations of judgment scenarios. For example, when When, it means and This corresponds to a business anomaly and no misrepresentation of the scope of the report; when When, it means and This corresponds to a situation where there are no business anomalies but there is a shift in the scope of the report; when This indicates that both business anomalies and policy drift exist simultaneously.
[0076] Based on the classification calculation of the result of the combined identifier, define the output function: ; When both business anomalies and caliber shifts are detected simultaneously, the result is categorized as a rule-layer change output identifier to avoid interference with the business anomaly determination during the rule adjustment phase.
[0077] The final output result is based on The system generates and records the current parsed version data identifier for the corresponding time point. This processing path ensures the logical independence and distinguishability of the business anomaly judgment result and the caliber drift judgment result. All variables originate from the aforementioned anomaly identification calculation and drift judgment calculation results. The time alignment, combined identifier generation, and classification calculation steps form a complete data processing chain, achieving a structured separation of anomalies between the rule layer and the business layer.
[0078] In one specific implementation, we will illustrate the scenario of sales anomaly monitoring in a chain retail enterprise. The enterprise performs daily anomaly identification calculations on the sales figures of each store to detect stores with abnormal fluctuations. During a certain period, the finance department changes the statistical scope of "sales amount" from "amount including tax" to "amount excluding tax," and simultaneously changes the unit of the field from "yuan" to "ten thousand yuan." This adjustment is not explicitly marked in the business system; it is only reflected in changes to the field dictionary data and parsed version data.
[0079] On the day of the adjustment, store A's sales revenue changed from 125,000 yuan to 125,000 yuan. After parsing and mapping, if the logic for separating the caliber drift identification from the result is not executed, the anomaly identification model will treat this numerical change as a sudden drop in sales, thus generating a business anomaly judgment result. Meanwhile, the caliber drift determination result is obtained based on the constraint residual calculation. This indicates a change in the statistical methodology.
[0080] In the result separation process, time points The two judgment results are aligned to form aligned judgment data. According to the rules for determining combined identifiers, we obtain... This addresses the situation where both business anomalies and statistical misalignments coexist. It outputs a rule-level change determination result based on the classification rules, but not a business anomaly determination result. This output is intended to indicate to operations personnel that the current anomaly primarily stems from changes in statistical rules, rather than store operational anomalies.
[0081] After the change in scope is completed and the parsing version is updated stably, subsequent time points The system is reset to 0, and the model re-performs normal anomaly identification for sales fluctuations. Through this implementation, the system can avoid numerous false alarms caused by adjustments to units or statistical rules in real-world engineering environments, while retaining its ability to identify genuine operational anomalies, achieving separate output of rule-layer changes and business-layer anomalies.
[0082] In another possible implementation, this is applied to multi-currency settlement data analysis scenarios for cross-border e-commerce enterprises. In this scenario, the enterprise simultaneously receives sales data denominated in different currencies such as RMB, USD, and EUR, and converts it into the local currency according to the exchange rate of the day for report statistics in the settlement system. In actual operation, the exchange rate source may be switched from a "real-time exchange rate interface" to a "monthly average exchange rate interface," or the exchange rate precision may be adjusted from four decimal places to two decimal places. Such changes do not alter the sales quantity or order structure, but they will cause an overall shift in the amount field.
[0083] In this implementation, the field dictionary data includes currency identifiers and exchange rate lineage information, and a rule edge is established in the constraint graph that states "base currency amount = original currency amount × exchange rate". When the source of the exchange rate is adjusted, the constraint residuals will show a concentrated increase in the amount-related constraints. The system identifies the set of high-contribution constraints through residual contribution analysis and automatically generates candidate parsing version data, updating the source identifier and conversion expression of the exchange rate field to the new rule version.
[0084] During the data mining phase, the residual feature vector contains exchange rate constraint residual information, enabling the model to identify whether current fluctuations are related to changes in exchange rate rules during anomaly detection. In the results separation and processing phase, when both business anomaly determination results and caliber drift determination results exist simultaneously, the system prioritizes outputting the rule-layer change determination result, indicating that the current amount fluctuation is primarily caused by adjustments to exchange rate rules.
[0085] This implementation extends rule-layer changes to multi-currency conversion scenarios, making it suitable for enterprise data environments with exchange rate mapping relationships or cross-currency field derivation relationships. Any implementation that identifies currency conversion rule changes based on field lineage and combines this with anomaly detection and result separation falls within the technical scope of this invention.
[0086] Example: In existing solutions of this invention, parsed version data is essentially a collection of internal system parameter structures used to define field mapping relationships, constraint expression information, and unit conversion rules. When the application scenario expands from data mining within a single enterprise to a collaborative scenario involving multiple enterprises across the upstream and downstream of the industry chain, the differences in organizational structure, business processes, field naming rules, and statistical calibers among different enterprises can essentially be regarded as differences in system structural parameters. Therefore, in the industry chain docking scenario, the parsed version data in this invention no longer represents only the statistical rule version of a single enterprise, but is expanded to a set of structural parameters for different enterprise structures, used to describe the differences between supplier structural parameters, manufacturing enterprise structural parameters, distribution enterprise structural parameters, and service enterprise structural parameters.
[0087] The platform provides data integration services for a specific equipment manufacturing industry chain. Upstream suppliers use "batch shipment amount" and "settlement amount including tax" as the main monetary fields, midstream manufacturers use "material procurement cost" and "cost amount excluding tax" fields, and downstream distributors use "outbound sales amount" and "receipt amount" fields. Although there are semantic correspondences between the fields of each enterprise, there are differences in field identification information, field unit identification information, and field lineage information. When the system connects to enterprises, it generates a corresponding enterprise structure parameter set for each enterprise and writes this enterprise structure parameter set into the parsed version data. The parsed version data thus becomes a combination of "enterprise structure parameters + constraint expression set".
[0088] In actual operation, when the system needs to compare the matching relationship between upstream procurement amount, midstream production cost, and downstream sales amount, a unified constraint diagram is constructed based on the enterprise structural parameter set corresponding to each enterprise, and a cross-enterprise constraint function set is generated, such as "Sales amount = Production cost + Reasonable profit" and "Production cost = Procurement amount + Processing fee". After the enterprise's multi-source business data completes field rearrangement and unit conversion according to its respective structural parameter set, it enters the unified constraint calculation path. If a distribution enterprise adjusts its internal system and changes the "Sales amount" from statistics based on outbound time to statistics based on receipt time, the field lineage information in the corresponding parsed version data of that enterprise will change. The system detects a concentrated increase in cross-enterprise matching constraint residuals during the constraint residual calculation stage and generates a caliber drift judgment result.
[0089] Furthermore, based on the high-contribution constraint set, the structural parameters of affected enterprises are located, and candidate parsing version data is generated. Historical cross-enterprise data is replayed to verify whether the candidate structural parameters can reduce the residual aggregation value. Once the candidate parsing version data passes verification, it replaces the original enterprise structural parameter set. The updated parsing version data continues to participate in the anomaly identification model calculation, and the results are separated to distinguish between "matching offset caused by structural parameter adjustment" and "real supply-demand imbalance anomaly." Through this implementation, the present invention achieves system-level parameter adaptive adjustment for differences in the structural structure of different enterprises in the industrial chain docking scenario, ensuring the stability of cross-enterprise docking calculations when structural parameters change, and avoiding false alarms in industrial chain collaborative early warning caused by adjustments to internal enterprise statistical rules.
[0090] like Figure 2 As shown, this invention provides an enterprise data parsing system based on big data mining. The system includes: a data acquisition unit, used to acquire multi-source business data, field dictionary data, and current parsing version data of an enterprise, and construct a caliber constraint graph based on the field dictionary data; a drift determination unit, used to generate a constraint set based on the current parsing version data, perform constraint calculations on the multi-source business data of the enterprise to generate constraint residual data, and calculate a caliber drift determination result based on the constraint residual data; a sample generation unit, used to update the current parsing version data when the caliber drift determination result meets preset conditions, and perform parsing mapping processing on the multi-source business data of the enterprise based on the current parsing version data to generate mining sample data; an anomaly determination unit, used to input the mining sample data, together with the current parsing version data and the constraint residual features generated based on the constraint residual data, into a preset big data mining model to perform anomaly identification calculations and generate a business anomaly determination result; and a result output unit, used to perform result separation processing based on the business anomaly determination result and the caliber drift determination result, and output a business anomaly determination result or a rule layer change determination result.
[0091] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a microcontroller, chip, or processor to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.
[0092] The optional embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details described above. Within the scope of the technical concept of the embodiments of the present invention, various simple modifications can be made to the technical solutions of the embodiments of the present invention, and these simple modifications all fall within the protection scope of the embodiments of the present invention. It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the embodiments of the present invention will not further describe the various possible combinations.
[0093] Furthermore, various different embodiments of the present invention can be combined in any way, as long as they do not violate the spirit of the embodiments of the present invention, they should also be regarded as the content disclosed by the embodiments of the present invention.
Claims
1. A method for analyzing enterprise data based on big data mining, characterized in that, The method includes: Acquire enterprise multi-source business data, field dictionary data, and current parsing version data, and construct a caliber constraint diagram based on the field dictionary data; Based on the current parsed version data, a constraint set is generated. The constraint calculation of the constraint set is performed on the enterprise's multi-source business data to generate constraint residual data. The caliber drift judgment result is calculated based on the constraint residual data. When the caliber drift determination result meets the preset conditions, the current parsing version data is updated, and parsing mapping processing is performed on the enterprise multi-source business data based on the current parsing version data to generate mining sample data; The mined sample data, together with the current parsed version data and the constraint residual features generated based on the constraint residual data, are input into a preset big data mining model to perform anomaly identification calculations and generate business anomaly judgment results. Based on the separation processing of the business anomaly determination result and the caliber drift determination result, the business anomaly determination result or the rule layer change determination result is output.
2. The enterprise data parsing method based on big data mining according to claim 1, characterized in that, The rules for constructing a caliber constraint diagram based on the aforementioned field dictionary data are as follows: Extract field identifier information, field unit identifier information, and field lineage information from the field dictionary data; A set of field nodes is established using the field identifier information as field nodes, and a set of unit nodes is established using the field unit identifier information as unit nodes. Based on the lineage information of the fields, a set of rule edges is established between the field nodes, and the set of rule edges is used to represent the derivation relationship between the fields; Based on the unit identification information of the field, a set of unit mapping edges is established between the field node and the unit node. The set of unit mapping edges is used to represent the conversion relationship between the field and the unit. The set of field nodes, the set of unit nodes, the set of rule edges, and the set of unit mapping edges are combined to form a caliber constraint graph.
3. The enterprise data parsing method based on big data mining according to claim 2, characterized in that, Based on the current parsed version of the data, a constraint set is generated. Constraint calculations are then performed on the enterprise's multi-source business data using the constraint set to generate constraint residual data, including: Extract constraint identification information, field mapping information, and constraint expression information from the currently parsed version data; Based on the constraint identification information and the field mapping information, the corresponding field nodes and rule edges are located in the caliber constraint diagram, and a set of constraint functions is constructed according to the constraint expression information; The enterprise's multi-source business data is converted into a sequence of field values corresponding to the constraint function set according to the field mapping information. Perform function calculations on the set of constraint functions on the sequence of field values to generate the calculation result values of each constraint function. Based on the calculated result value and the preset constraint conditions, deviation calculation is performed to generate the constraint residual data.
4. The enterprise data parsing method based on big data mining according to claim 3, characterized in that, The caliber drift determination result is calculated based on the constrained residual data, including: The constraint residual data is sorted according to the constraint identifier and execution time to generate a constraint residual sequence; The residual aggregated value sequence is calculated based on the constrained residual sequence, and the residual aggregated value sequence is obtained by performing aggregate calculation on the constrained residual sequence within a preset time window; Calculate the residual baseline value and residual fluctuation value based on the residual aggregated value sequence; The residual aggregated value sequence is compared with the residual baseline value and the residual fluctuation value to generate a drift determination identifier sequence; The aperture drift determination result is generated based on the drift determination identifier sequence.
5. The enterprise data parsing method based on big data mining according to claim 1, characterized in that, When the caliber drift determination result meets the preset conditions, the current parsing version data is updated, including: Based on the constraint residual data, the residual contribution value of each constraint is calculated according to the constraint identifier, and constraints with residual contribution values greater than the average residual contribution value are selected to form a set of high contribution constraints. Based on the field attribute information corresponding to the high contribution constraint set in the field dictionary data, a constraint expression set is regenerated, and the constraint expression set is combined to form candidate parsing version data. Based on the candidate parsing version data, replay calculations are performed on the enterprise's multi-source business data to generate candidate constraint residual data, and candidate residual aggregate values are calculated. When the candidate residual aggregate value is less than the current residual aggregate value, the candidate parsing version data replaces the current parsing version data.
6. The enterprise data parsing method based on big data mining according to claim 5, characterized in that, Based on the current parsed version data, parsing and mapping processing is performed on the enterprise's multi-source business data to generate mining sample data, including: Extract field mapping relationships and constraint expression information from the current parsed version data, and perform field rearrangement processing on the enterprise multi-source business data according to the field mapping relationships to generate mapped field data; Based on the constraint expression information, unit conversion calculations and field derivation calculations are performed on the mapped field data to generate standardized field data; The standardized field data is processed to perform missing value imputation and outlier correction rule calculation to generate corrected field data; The modified field data is combined according to a preset feature structure to form the mining sample data.
7. The enterprise data parsing method based on big data mining according to claim 1, characterized in that, The mined sample data, along with the current parsed version data and the constraint residual features generated based on the constraint residual data, are input into a preset big data mining model to perform anomaly identification calculations, generating business anomaly determination results, including: Based on the mined sample data, a sample feature vector is constructed, and based on the constrained residual features, a residual feature vector is constructed. A version identifier vector is generated based on the current parsed version data, and the version identifier vector is concatenated with the sample feature vector and the residual feature vector to generate the model input vector. The model input vector is input into the preset big data mining model to perform anomaly scoring calculation and generate anomaly scoring values. The anomaly score is compared with the anomaly determination threshold to generate the business anomaly determination result.
8. The enterprise data parsing method based on big data mining according to claim 7, characterized in that, The preset big data mining model includes: A feature encoding unit is used to perform encoding calculations on the sample feature vector to generate a sample encoding vector; A residual modulation unit is used to perform feature modulation calculation on the sample coding vector based on the residual feature vector to generate a modulation coding vector; The version normalization unit is used to select the normalization parameter corresponding to the current parsed version data based on the version identifier vector and perform version normalization calculation on the modulation and coding vector to generate a normalized vector; An anomaly scoring unit is used to perform scoring calculations based on the normalized vector to generate the anomaly score value.
9. The enterprise data parsing method based on big data mining according to claim 8, characterized in that, Based on the separation processing of the business anomaly determination result and the caliber drift determination result, the business anomaly determination result or rule layer change determination result is output, including: The business anomaly determination result and the caliber drift determination result are time aligned to generate aligned determination data; A judgment combination identifier is constructed based on the alignment judgment data, wherein the judgment combination identifier is formed by combining the business anomaly identifier and the caliber drift identifier; Based on the classification calculation of the judgment combination identifier, a business anomaly output identifier or a rule layer change output identifier is generated; The final output result is generated based on the business anomaly output identifier or the rule layer change output identifier.
10. An enterprise data analysis system based on big data mining, characterized in that, The system includes: The data acquisition unit is used to acquire enterprise multi-source business data, field dictionary data, and current parsing version data, and to construct a caliber constraint diagram based on the field dictionary data; The drift determination unit is used to generate a constraint set based on the current parsed version data, perform constraint calculations on the enterprise multi-source business data to generate constraint residual data, and calculate the caliber drift determination result based on the constraint residual data. The sample generation unit is used to update the current parsing version data when the caliber drift determination result meets the preset conditions, and to perform parsing mapping processing on the enterprise multi-source business data based on the current parsing version data to generate mining sample data; An anomaly determination unit is used to input the mined sample data, together with the current parsed version data and the constraint residual features generated based on the constraint residual data, into a preset big data mining model to perform anomaly identification calculation and generate a business anomaly determination result. The result output unit is used to perform result separation processing based on the business anomaly determination result and the caliber drift determination result, and output the business anomaly determination result or the rule layer change determination result.