Enterprise business metadata intelligent management method based on multi-source heterogeneous data

Abstract

The invention relates to the field of data processing, in particular to an enterprise business metadata intelligent management method based on multi-source heterogeneous data, which comprises the following steps: performing event and blood relationship association analysis on enterprise multi-source heterogeneous data to obtain an enterprise business metadata dynamic monitoring data set; carrying out conjoint analysis on the time neighborhood of the metadata change event and the propagation characteristics of the influence domain to obtain a danger rate regulation factor; performing fusion analysis on the cross-window stability and the structure gradual change feature of the metadata change event to obtain a gradual change regulation factor; carrying out conjoint analysis on the danger rate regulation factor and the slow change regulation factor to obtain a dynamic danger rate; through joint distribution of the dynamic danger rate and the detection posteriori, comprehensive detection output is obtained, and therefore the problem that an existing Bayesian online change point detection algorithm cannot timely and accurately recognize real metadata changes in a multi-source heterogeneous environment is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to an enterprise business metadata intelligent management method based on multi-source heterogeneous data. BACKGROUND

[0002] In the multi-source heterogeneous information system environment of modern enterprises, business data and its corresponding metadata are widely distributed in enterprise resource planning (ERP), customer relationship management (CRM), production execution system (MES), data warehouse (DWH), and report analysis platform, etc. Enterprise business metadata not only includes the structure definition, value domain and verification rule of field and interface, but also covers index calculation caliber, data bloodline dependency relationship, version number, release window, gray release and rollback state, etc. With the continuous evolution of enterprise business logic and system architecture, these metadata show significant dynamic and multi-time characteristics, and their changes are often influenced by multiple factors such as business adjustment, system upgrade and cross-system task scheduling. The structure update or caliber adjustment of metadata is usually recorded in the form of release order, DDL audit, interface gateway policy update and task scheduling log, so it is of great significance to accurately identify and track the mutation and gradual change in the evolution process of metadata to ensure enterprise data governance and index consistency.

[0003] At present, the industry often uses Bayesian online change point detection (BOCPD) algorithm to identify the abnormality or mutation in the evolution process of metadata. This algorithm calculates the running length posterior and change point posterior by recursion, and dynamically judges the mutation point of time series distribution by combining the hazard rate and observation likelihood. It has the advantages of no manual annotation, online update and adaptation to streaming data, and has become a typical time series detection model. However, in actual enterprise application scenarios, metadata changes are often not random, but closely related to business release cycles, system dependency structures and bloodline relationships. The traditional Bayesian online change point detection algorithm only relies on sequence statistical features for posterior inference, lacks the recognition and use of business scenario features, and leads to detection delay, misjudgment or missed detection in multi-source heterogeneous environments. When there is asynchronous propagation or cross-system delay in metadata changes, the algorithm is difficult to capture the real changes in time; while in low-impact nodes or scenarios without substantial changes, it is easy to be misidentified as a mutation. Therefore, how to inject quantifiable business scenario features into the hazard rate modeling process under the framework of Bayesian online change point detection algorithm, so that the model has the ability of time and space perception and structure recognition, to realize the timely and accurate identification of real metadata changes, has become a technical problem to be solved in the field. SUMMARY

[0004] Therefore, the present application aims to propose an enterprise business metadata intelligent management method based on multi-source heterogeneous data to solve the problem that the existing Bayesian online change point detection algorithm cannot timely and accurately identify real metadata changes in a multi-source heterogeneous environment.

[0005] To achieve the above purpose, the technical scheme of the present application is as follows:

[0006] The enterprise business metadata intelligent management method based on multi-source heterogeneous data comprises the following steps:

[0007] Step S1: obtaining an enterprise business metadata dynamic monitoring data set by performing event and blood relationship analysis on enterprise multi-source heterogeneous data;

[0008] Step S2: obtaining a hazard rate adjustment factor by jointly analyzing the time proximity and influence domain propagation characteristics of metadata change events;

[0009] Step S3: obtaining a slow change adjustment factor by performing fusion analysis on the cross-window stability and structure slow change characteristics of metadata change events;

[0010] Step S4: obtaining a dynamic hazard rate by jointly analyzing the hazard rate adjustment factor and the slow change adjustment factor;

[0011] Step S5: obtaining a comprehensive detection output by jointly analyzing the dynamic hazard rate and the joint distribution of detection posteriori.

[0012] Further, the event and blood relationship analysis on enterprise multi-source heterogeneous data to obtain an enterprise business metadata dynamic monitoring data set comprises the following steps:

[0013] The log collection module and the audit module arranged in each business system of the enterprise are used to synchronously collect multi-source heterogeneous data from the enterprise resource planning system, the customer relationship management system, the production execution system, the data warehouse system and the report analysis platform;

[0014] The metadata change time is extracted from the structure change log and the audit record of the above system to form an event set with time stamp as the primary key and containing structure definition change content;

[0015] The upstream dependency set of the target field or index and the node set affected by the change event are extracted from the task scheduling system and the interface gateway to construct an upstream node set and an event node set, respectively;

[0016] The blood relationship graph containing the data flow relationship between each field and index is obtained from the enterprise data governance platform, and the betweenness centrality, average betweenness centrality and betweenness centrality standard deviation of each node in the blood relationship graph are calculated;

[0017] extracting a release event interval window data from an enterprise release system and a version control log, and calculating a time offset evaluation of each change event timestamp and a center of the release event window;

[0018] organizing and managing the collected event set, the upstream node set, the event node set, the blood relationship graph structure parameters, and the time offset evaluation result to form an enterprise business metadata dynamic monitoring data set containing event information, dependent node relationship, and time characteristics.

[0019] Further, the dangerous rate adjustment factor is obtained by jointly analyzing the time proximity of the metadata change event and the influence domain propagation characteristics, including:

[0020] The event time matching degree evaluation result is obtained by performing proximity analysis on the release window time data and the change event time data.

[0021] The node structure influence strength evaluation result is obtained by performing propagation characteristic analysis on the event influence node data and the blood relationship structure data, and the event time matching degree evaluation result is weighted and optimized by the node structure influence strength evaluation result to obtain the dangerous rate adjustment factor.

[0022] Further, the event time matching degree evaluation result is obtained by performing proximity analysis on the release window time data and the change event time data, including:

[0023] A time kernel width parameter is set, for any target node at any target time, the time difference between the change event time of the target node at the target time and the center of the release window is divided by the time kernel width parameter, and the calculation result is taken as the time difference evaluation of the target node at the target time. The inverse number of the calculation result of the square of the time difference evaluation of the target node at the target time divided by the constant 2 is exponentially mapped with the natural constant as the base number, and the corresponding mapping result is taken as the event time matching degree evaluation result of the target node at the target time.

[0024] Further, the node structure influence strength evaluation result is obtained by performing propagation characteristic analysis on the event influence node data and the blood relationship structure data, and the event time matching degree evaluation result is weighted and optimized by the node structure influence strength evaluation result to obtain the dangerous rate adjustment factor, including:

[0025] For any target node at any target time, the number of nodes in the intersection of the upstream dependent node set corresponding to the target node or the field and the change event influence node set of the target node at the target time is taken as the numerator, and the number of nodes in the upstream dependent node set corresponding to the target node or the field is taken as the denominator. The fraction formed accordingly is the time hit degree of the target node at the target time.

[0026] Set a centrality standardization adjustment coefficient; use the result of subtracting the betweenness centrality assessment of the target node from the average betweenness centrality of all nodes in the kinship graph as the numerator, and the standard deviation of the betweenness centrality of all nodes in the kinship graph as the denominator, and use the corresponding fraction as the centrality standardization assessment of the target node; use the result of multiplying the centrality standardization adjustment coefficient by the centrality standardization assessment of the target node as the first centrality assessment of the target node; apply the Sigmoid function mapping to the first centrality assessment of the target node, and use the corresponding mapping result as the structural propagation ability assessment of the target node;

[0027] The result of multiplying the temporal hit rate of the target node at the target time by the structural propagation capability assessment of the target node is taken as the structural influence intensity of the target node at the target time.

[0028] The result of multiplying the structural influence intensity of the target node at the target time by the event time matching degree assessment result of the target node at the target time is used as the risk rate adjustment factor of the target node at the target time.

[0029] Furthermore, the method of obtaining a gradual adjustment factor by fusing analysis of cross-window stability and structural gradual change characteristics of metadata change events includes:

[0030] By performing directional consistency and trend analysis on the structural influence intensity sequence data within a continuous time window, the cross-window trend consistency assessment results are obtained.

[0031] By performing hysteresis correlation analysis on the structural influence intensity sequence data of the target node and its upstream nodes, the local structural inertia assessment results are obtained.

[0032] By fusing cross-window trend consistency assessment data and local structural inertia assessment data, a gradual adjustment factor is obtained.

[0033] Furthermore, the step of performing directional consistency and trend analysis on the structural influence intensity sequence data within a continuous time window to obtain cross-window trend consistency assessment results includes:

[0034] The calculation expression for the cross-window trend consistency assessment of any target node at any target time is as follows:

[0035] ;

[0036] in, Indicates the first The node at the th Consistency assessment of cross-window trends at each moment; Indicates the window length; Indicates the sign of a function, used to determine the sign of the calculation result within the function; Indicates the first The node at the th The sequence of structural influence intensity at each time point; Indicates the first The node at the th The slope estimate obtained by linear regression of the structural influence intensity sequence at each time point on the event index; Indicates the first The node at the th The standard deviation of the structural influence intensity sequence at each time point; It represents a very small positive number.

[0037] Furthermore, the step of obtaining local structural inertia assessment results by performing hysteresis correlation analysis on the structural influence intensity sequence data of the target node and its upstream nodes includes:

[0038] The calculation expression for the local structural inertial assessment result of any target node at any target time is:

[0039] ;

[0040] in, Indicates the first The node at the th Local structural inertial assessment at each moment; Indicates the lag length; Indicates the maximum lag length; Indicates the first The node at the th The length of each time period is The structural influence intensity sequence; Indicates the first The node at the th The length of each time period is The structural influence intensity sequence; Indicates the number in the kinship chart The node to the first The bloodline edge weight of each node; Indicates the number in the kinship chart The set of upstream dependent nodes of each node; This represents the function for calculating the Pearson correlation coefficient.

[0041] Furthermore, the step of fusing cross-window trend consistency assessment data and local structural inertia assessment data to obtain a gradually varying adjustment factor includes:

[0042] Set trend assessment weights and inertia assessment weights; for any change event of any target node at any target time, the cross-window trend consistency assessment and the local structural inertia assessment of the target node at the target time are weighted and summed by the trend assessment weights and inertia assessment weights, respectively, and the corresponding weighted summation results are used as the gradual change adjustment factor of the target node at the target time.

[0043] Furthermore, the method of obtaining the dynamic hazard rate through joint analysis of the hazard rate adjustment factor and the slow-change adjustment factor includes:

[0044] Set a basic hazard rate constant; perform an exponential mapping with the natural constant as the base of the calculation result of adding the hazard rate adjustment factor of the target node at the target time and the slowly changing adjustment factor of the target node at the target time; and multiply the corresponding mapping result by the basic hazard rate constant as the dynamic hazard rate of the target node at the target time.

[0045] Compared with the prior art, the present invention has the following advantages:

[0046] The intelligent management method for enterprise business metadata based on multi-source heterogeneous data described in this invention introduces a dynamic hazard rate modeling mechanism into a Bayesian online change point detection framework, enabling the detection process to have business scenario awareness capabilities. By comprehensively analyzing the release window, dependency structure, and lineage relationships in the enterprise's multi-source heterogeneous systems, the model can identify the correlation between metadata changes and actual business operations. This allows for timely responses to real changes at high-propagation, high-impact nodes, while automatically suppressing low-propagation, non-substantive modifications. This mechanism effectively eliminates the detection lag and false alarm problems caused by relying solely on time-series statistical features in traditional methods, transforming change identification from a statically statistically driven process to a context-aware, dynamic identification process, significantly improving the real-time performance and accuracy of the metadata governance system. Furthermore, this invention considers both abrupt and gradual change characteristics in hazard rate modeling, achieving unified identification of sudden events and gradual evolution through dynamic factor adjustment. In complex enterprise environments with asynchronous releases and cross-system dependencies, the algorithm can adaptively adjust the response intensity to hazard signals, enabling immediate alerts for rapid changes and continuous tracking of slow evolutions. This method not only improves the capture rate and robustness of metadata monitoring systems for real changes, but also maintains detection stability in scenarios involving multi-node link propagation and rolling version updates, providing enterprises with a highly reliable and scalable intelligent detection method for data governance and indicator consistency management. Attached Figure Description

[0047] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0048] Figure 1 This is a flowchart illustrating the intelligent management method for enterprise business metadata based on multi-source heterogeneous data, as described in an embodiment of the present invention. Detailed Implementation

[0049] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0050] See Figure 1 This is a flowchart of the intelligent management method for enterprise business metadata based on multi-source heterogeneous data provided in Embodiment 1 of the present invention. Figure 1 As shown, an intelligent management method for enterprise business metadata based on multi-source heterogeneous data may include:

[0051] Step S1: By performing event and lineage correlation analysis on the enterprise's multi-source heterogeneous data, a dynamic monitoring dataset of enterprise business metadata is obtained.

[0052] In scenarios where multiple systems operate in parallel within an enterprise, the dynamic evolution information of metadata does not originate from a single system but is distributed across multiple heterogeneous data sources, including structure change logs, interface release logs, task orchestration dependency tables, and data lineage diagrams. In order to accurately depict the changes in metadata in subsequent analysis processes, this step aims to obtain a dynamic monitoring dataset of metadata that includes events, influencing nodes, and lineage relationships through correlation analysis of multi-source data.

[0053] Specifically, firstly, metadata change event records are extracted from the logs and audit modules of each system. All collected events are denoted as the event set, where each event uses a timestamp as its primary key and includes changes to the event's structural definition, i.e., the structural change record. Simultaneously, corresponding dependency information is extracted from the task scheduling system and interface gateway, forming the upstream dependency set (upstream node set) of the target field or metric and the set of nodes affected by the event (event node set).

[0054] Secondly, in order to describe the data transmission path between systems, a lineage graph is obtained from the enterprise data governance platform. Each node in the lineage graph corresponds to a field or indicator, and each edge represents the dependency relationship of data flow. The betweenness centrality of each node is calculated based on the lineage graph to measure the propagation ability of the node in the overall data link. At the same time, the standard deviation of the average betweenness centrality of the lineage graph and the betweenness centrality of all nodes is calculated.

[0055] In the time dimension, the internal release time interval window is extracted by the release system and version control logs to define the time range within which adjustments may be made. For each event, the time difference between it and the center of the release time interval window is calculated, i.e., the time offset assessment.

[0056] Ultimately, the acquired data and evaluation results are combined to form a dynamic monitoring dataset of enterprise business metadata.

[0057] This completes the dynamic monitoring dataset of enterprise business metadata obtained by performing event and lineage correlation analysis on multi-source heterogeneous data of the enterprise.

[0058] Step S2: Obtain the risk rate adjustment factor by jointly analyzing the temporal proximity and influence domain propagation characteristics of metadata change events.

[0059] In the dynamic management of enterprise business metadata, structural changes in metadata are often not isolated events, but rather occur synchronously with enterprise business activities, system releases, and task dependencies. Changes in metadata definitions exhibit clear business temporal and structural dependency characteristics. However, existing Bayesian online change point detection algorithms rely solely on posterior inference based on time-series statistical features when identifying metadata mutations. They fail to incorporate contextual awareness of the relationship between events and structures during the risk modeling stage. This results in Bayesian online change point detection algorithms being able to detect mutation points, but unable to respond promptly within the change's time window, and unable to distinguish between global structural propagation and pseudo-changes caused by local noise. This modeling approach, detached from contextual characteristics, limits the algorithm's detection accuracy in multi-source, heterogeneous enterprise environments.

[0060] To address this issue, this step starts from a real-world scenario, incorporating temporal proximity and influence domain propagation characteristics as quantifiable scenario features into the risk rate modeling process of the Bayesian online change point detection algorithm. By jointly modeling the event offset and lineage propagation tendency within the publishing window, the algorithm gains scenario recognition capabilities during the detection phase. When an event highly overlaps with the publishing window and hits a lineage node with a high propagation tendency, the change point risk rate is increased, enabling the Bayesian online change point detection algorithm to quickly identify real business changes. Conversely, when an event is far from the publishing window or only affects low-propagation nodes, the risk rate remains at the baseline level, thus suppressing false alarms.

[0061] In summary, this invention first obtains the event time matching degree evaluation result by performing nearest neighbor analysis on the release window time data and change event time data. Specifically, a time kernel width parameter is set. In this embodiment of the invention, the time kernel width parameter is set to 3. The time kernel width parameter can be adjusted according to the actual scenario and is not required. For any change event of any target node at any target time, the time difference between the change event time of the target node at the target time and the time of the release window center is divided by the time kernel width parameter to obtain the time difference evaluation of the target node at the target time. The square of the time difference evaluation of the target node at the target time is divided by the constant 2, and the negative number of the result is used for exponential mapping with the natural constant as the base. The corresponding mapping result is used as the event time matching degree evaluation result of the target node at the target time.

[0062] After obtaining the event time matching degree assessment result, the propagation characteristic analysis is further performed on the event-affected node data and kinship structure data to obtain the node structure influence intensity assessment result. The event time matching degree assessment result is then weighted and optimized using the node structure influence intensity assessment result to obtain the risk rate adjustment factor. Specifically, for any change event of any target node at any target time, the number of nodes at the intersection of the upstream dependent node set of the indicator or field corresponding to the target node and the node set affected by the change event of the target node at the target time is used as the numerator, and the number of nodes in the upstream dependent node set of the indicator or field corresponding to the target node is used as the denominator. The resulting fraction is used as the time hit degree of the target node at the target time. A centrality standardization adjustment coefficient is set. In this embodiment of the invention, the centrality standardization adjustment coefficient is set to 4.5. The centrality standardization adjustment coefficient is used to control the slope of the Sigmoid function in the propagation tendency mapping, so that the influence of betweenness centrality on node propagation is adjustable. A larger centrality normalization adjustment coefficient enhances the distinction between high-centrality nodes and ordinary nodes, suppressing false triggering by low-centrality nodes. The result of subtracting the average betweenness centrality of all nodes in the kinship graph from the target node's betweenness centrality assessment is used as the numerator, and the standard deviation of the betweenness centrality of all nodes in the kinship graph is used as the denominator. This fraction is then used as the centrality normalization assessment of the target node. The result of multiplying the centrality normalization adjustment coefficient by the target node's centrality normalization assessment is used as the target node's first centrality assessment. The first centrality assessment of the target node is then mapped using a Sigmoid function, and the resulting mapping is used as the target node's structural propagation capability assessment. The result of multiplying the target node's temporal hit rate at the target time by its structural propagation capability assessment is used as the target node's structural influence strength at the target time. The result of multiplying the target node's structural influence strength at the target time by its event time matching degree assessment at the target time is used as the target node's hazard rate adjustment factor at the target time.

[0063] In one implementation, assume the first The set of upstream dependent nodes for each node's corresponding metric or field is: ;No. The node at the th The set of nodes affected by the change event at each moment is: ;No. The betweenness centrality of the nodes is The average betweenness centrality of all nodes in the kinship graph is: The standard deviation of the betweenness centrality of all nodes in the kinship graph is: The centrality-standardized adjustment coefficient is The difference between the event time and the time in the publishing window center is: The time kernel width parameter is Then the first The node at the th The formula for calculating the risk rate adjustment factor at each time point is:

[0064]

[0065] in, Indicates the first The node at the th Risk rate adjustment factor at each time point; Indicates the first The set of upstream dependent nodes for each node's corresponding metric or field; Indicates the first The node at the th A change event at a given moment affects the set of nodes; Indicates the first The node at the th The event hit rate at a given moment reflects the degree of overlap between the scope of the event's impact and the target field's dependency domain. Indicates the first The betweenness centrality of a node is used to measure the node's ability to propagate data flow in the kinship graph; This represents the average betweenness centrality of all nodes in the kinship graph; The standard deviation of the betweenness centrality of all nodes in the kinship graph; This represents the centrality standardization adjustment coefficient, used to control the slope of the sigmoid function, making the influence of propagation tendency adjustable and sensitive. This indicates the difference between the event time and the time displayed in the publishing window center. The time kernel width parameter represents the scale of the time range of an event's influence and is used to control the rate at which the event decays.

[0066] It should be noted that actual changes in metadata definitions or structures typically occur within a company's release window. Therefore, risk adjustment can be performed using time proximity. When an event's time is close to the center of the release window, i.e., the time difference between the event's time and the center of the release window is close to zero, When the value is close to 1, the risk rate should be increased; when the event time is far from the release window, The risk level rapidly decreases, returning to the baseline. However, in real-world scenarios, non-substantial noise from the same window—such as minor changes to document wording or low-impact edge nodes performing maintenance actions within the release window—can also inflate the risk level due to temporal proximity assessment errors. Furthermore, in cross-system asynchronous propagation, where real structural changes often first hit critical nodes in the chain and propagate along lineage, nodes not hit within the same time window should not be synchronously inflated. Relying solely on time weights amplifies prior information, leading to over-detection and false alarms. Therefore, this invention further optimizes the structural constraints of event proximity through the characteristics of influence domain propagation. It requires that the time frame must not only match the event but also hit the structural location that should be affected, and be located on a critical node that is more easily propagated, in order to allow a significant increase in the risk level. For the propagation characteristics of the influence domain, a comprehensive evaluation is conducted using hit rate and propagation tendency. Hit rate measures whether the event falls within the upstream dependency set of the node at the target time, i.e., whether it actually affects the node or indicator. Propagation tendency reflects the node's diffusion ability in the lineage graph through a sigmoid mapping standardized to betweenness centrality. The higher the centrality, the more likely it is to trigger a chain reaction. Thus, the risk rate is increased only when both hit rate and propagation capability are met. Specifically, the first part of the risk rate adjustment factor... No. The node at the th The structural influence strength at time t is used to describe the propagation characteristics of metadata at the structural level. This part consists of two elements: temporal hit rate and node propagation tendency. Its core logic lies in determining the strength of the structural influence at time t. Does the change event that occurs at a given time hit the target node? The upstream data domain on which it relies, and the propagation impact of nodes is weighted according to their centrality in the kinship graph. Among them, This indicates the time hit rate, reflected in the node. The proportion of the node directly or indirectly affected by this event within its dependency domain; if the event does not hit the target dependency domain, the ratio approaches 0, indicating that the node has not been structurally affected. The algorithm standardizes the betweenness centrality of nodes using the Sigmoid function, reflecting the node's propagation ability within the data lineage structure. Nodes with high centrality play a crucial bridging role between systems, and changes to them are more likely to cause interconnected fluctuations in the global data distribution. Nodes with low centrality are mostly local terminals, and their changes have a priority impact on the overall structure. By multiplying these two factors, nodes that are hit by events and are in high propagation positions receive higher response values, while nodes that are not hit or have weak propagation capabilities are automatically suppressed. This ensures that the optimization factor responds only to real and propagable changes at the structural level. Through the product of these two parts, a joint regulatory relationship between structural propagation characteristics and temporal proximity characteristics is formed. This ensures that the risk rate adjustment factor only significantly increases when both the key nodes are hit in time and have high propagation ability, and the event is within the publication window. This gives the algorithm a strong spatiotemporal awareness when detecting real business changes, enabling it to focus on the real change window in the temporal dimension and filter nodes with high propagation risk in the structural dimension. Thus, in risk rate adjustment, it achieves rapid response to real changes and effective suppression of noise interference.

[0067] Thus, the risk rate adjustment factor was obtained by jointly analyzing the temporal proximity and impact domain propagation characteristics of metadata change events.

[0068] Step S3: Obtain the gradual change adjustment factor by performing a fusion analysis on the cross-window stability and structural gradual change characteristics of metadata change events.

[0069] In real-world enterprise business systems, changes to metadata are often not instantaneous but rather spread slowly as multiple system modules are deployed asynchronously and tasks are scheduled. For example, when an enterprise restructures its indicator system, the definitions of core indicator fields may be updated first in the first system, while downstream dependent systems experience a synchronization delay. Alternatively, in large, group-based enterprise environments, the release cycles of different subsystems are not entirely consistent, resulting in a gradual propagation of adjustments from upstream to downstream. These changes manifest as a slow drift across time windows at the data layer and as delayed updates to local dependency chains at the structural layer.

[0070] In this scenario, the hazard rate adjustment factor constructed in step S2, while rapidly amplifying temporal proximity and structural propagation, can capture abrupt changes within a window. However, it may be overly sensitive to gradual, delayed-propagation structural changes. When an event fluctuates at a low amplitude across multiple adjacent release windows, the exponential amplification effect of the hazard rate adjustment factor leads to frequent fluctuations in the hazard rate curve, causing periodic overshooting of the posterior probability of the Bayesian online change point detection algorithm, thus triggering duplicate alarms or false positives. Simultaneously, the cumulative impact of some real changes is underestimated because they do not exhibit significant characteristics within a single window, resulting in detection lag.

[0071] To address this issue, this step employs a joint analysis of trend consistency and local structural inertia across time windows to smooth and correct the dynamic changes in risk rates. By introducing a gradual adjustment mechanism into the algorithm, transient high-risk signals caused by short-term noise can be suppressed, and the risk rate response can be gradually amplified in continuously evolving scenarios. This extends the Bayesian online change point detection algorithm from instantaneous mutation detection to dynamic evolution detection, achieving dual-domain collaborative identification of mutations and gradual changes.

[0072] In one embodiment, for the first The node at the th The calculation expression for the cross-window trend consistency assessment at each time point is:

[0073]

[0074] in, Indicates the first The node at the th Consistency assessment of cross-window trends at each moment; This indicates the window length; in this embodiment of the invention, the window length is set to 5 publishing windows. Indicates the sign of a function, used to determine the sign of the calculation result within the function; Indicates the first The node at the th The sequence of structural influence intensity at each time step, i.e. ; Indicates the first The node at the th The slope estimate obtained by linear regression of the structural influence intensity sequence at each time point on the event index; Indicates the first The node at the th The standard deviation of the structural influence intensity sequence at each time point; Representing extremely small positive numbers, as set in the embodiments of the present invention. .

[0075] After obtaining the number The node at the th After assessing the cross-window trend consistency at time step 1, continue to obtain the 2nd time step 2. The node at the th Local structural inertial assessment at time 1 :

[0076]

[0077] in, Indicates the first The node at the th Local structural inertial assessment at each moment; Indicates the lag length; This represents the maximum hysteresis length, which is set to 3 in this embodiment of the invention. Indicates the first The node at the th The length of each time period is The structural influence intensity sequence; Indicates the first The node at the th The length of each time period is The structural influence intensity sequence; Indicates the number in the kinship chart The node to the first The bloodline edge weight of each node; Indicates the number in the kinship chart The set of upstream dependent nodes of each node; This represents the function for calculating the Pearson correlation coefficient.

[0078] After obtaining the cross-window trend consistency assessment and the local structure inertia assessment, a comprehensive evaluation is performed on the cross-window trend consistency assessment and the local structure inertia assessment to obtain the gradual change adjustment factor. Specifically, trend assessment weight and inertia assessment weight are set. In this embodiment of the invention, both trend assessment weight and inertia assessment weight are set to 0.5. For any change event of any target node at any target time, the cross-window trend consistency assessment and the local structure inertia assessment of the target node at the target time are weighted and summed using the trend assessment weight and the inertia assessment weight, respectively. The corresponding weighted summation result is used as the gradual change adjustment factor of the target node at the target time.

[0079] For the The node at the th The formula for calculating the slowly varying adjustment factor at time point is:

[0080]

[0081] in, Indicates the first The node at the th The slowly varying regulatory factor at each moment; Indicates the weight of trend assessment; The inertia evaluation weight is set in the embodiments of the present invention. ; Indicates the first The node at the th Consistency assessment of cross-window trends at each moment; Indicates the first The node at the th Local structural inertial assessment at each moment.

[0082] It should be noted that the cross-window trend consistency assessment is constructed through the product of two sub-quantities during the calculation process. The first is directional consistency, which is used to statistically assess the trend consistency across a length of [missing information]. Within a continuous window, the direction of individual changes in the intensity sequence of node structure influence is consistent with the overall direction of change. Secondly, the monotonic trend strength measures the significance of the overall upward or downward slope of the sequence within the window against a fluctuating background. Directional consistency is given by the matching result of the sign function, such that when most individual changes are consistent with the composite direction, the directional consistency approaches 1; when the direction changes repeatedly, the directional consistency decreases rapidly. The monotonic trend strength is obtained by mapping the slope estimate obtained from linear regression of the intensity sequence of node structure influence to the sample standard deviation using a hyperbolic tangent function. The scope, preserving trend significance while suppressing the impact of extreme values, requires that cross-window trend consistency assessment achieves a high value only when both directional stability and trend significance are simultaneously established after multiplying directional consistency and monotonic trend strength. Cross-window trend consistency assessment addresses the issues of short-term noise and discontinuous changes in the time domain. The slow evolution of enterprise metadata typically manifests as unidirectional accumulation across windows. Even if the change in a single window is not significant, the continuity of direction and the stability of the overall trend constitute key evidence. Directional consistency is equivalent to a majority vote on the direction of change, excluding occasional small fluctuations in the opposite direction; trend strength uses the ratio of slope to fluctuation to evaluate the significance of slow but continuous changes. Combining the two can accumulate small progress over multiple consecutive windows into robust evidence, avoiding missed detections due to insufficient evidence from a single window, and also avoiding mistaking noise with fluctuating directions as valid signals. When the indicator field is in the most recent When continuous fine-tuning, consistent direction, and clear overall trend occur within a release cycle, the cross-window trend consistency assessment improves, and the gradual change adjustment factor increases accordingly. This allows the risk rate to rise smoothly and continuously within these windows until it reaches the cross-domain trigger threshold, thus correctly capturing slow evolution. Conversely, during gray-scale trials, frequent rollbacks, or operational jitters causing repeated direction changes, the cross-window trend consistency assessment automatically decreases, reducing the instantaneous rise in risk rate and suppressing repeated warnings. In the calculation process, the local structural inertia assessment is evaluated using the maximum post-correlation between the target node time series and its upstream weighted aggregated time series. First, a weighted average of the upstream node series is calculated using lineage edge weights to obtain the upstream aggregated time series. This reflects the synthetic influence of the upstream on the target node. Subsequently, the Pearson correlation coefficient between the target sequence and the upstream sequence is calculated within the subsequent set, and the maximum value is obtained. This coefficient is then linearly normalized to... Scope. The entire process allows for small-scale time delays to align with the actual situation in enterprise metadata chains where upstream adjustments precede downstream adjustments. Local structural inertia assessment addresses the issues of delayed propagation and chain consistency within the structural domain. Slow changes often propagate gradually along dependency chains. If the change in the target node is significantly correlated with the change in the upstream over a lag time, it indicates that the change in the target node is not noise but a delayed response driven by stable structural inertia. Using maximum lag correlation rather than zero lag correlation as the basis can distinguish between genuine chain propagation and irrelevant disturbances, avoiding misjudging occasional fluctuations lacking upstream consistency as slow changes. Local structural inertia assessment provides robust weighting for hazard rate updates. When the upstream system has been producing according to the new caliber for several consecutive cycles and the downstream is gradually following the scheduling cycle, the correlation between the target node and the upstream increases over a lag time, and the value of the local structural inertia assessment increases, keeping the slow change adjustment factor at a high level. This drives the hazard rate to rise smoothly in subsequent windows, achieving continuous tracking of delayed consistency. If changes at the target node are unrelated to or have low correlation with the upstream, the local structural inertia assessment value decreases, thereby slowing down the risk rate update and avoiding the misinterpretation of local fluctuations as systemic evolution. Cross-window trend consistency assessment and local structural inertia assessment are numerically integrated through a weighted summation, forming complementary constraints in the time and structural dimensions. When a stage exhibits stable direction and a clear trend over time, and shows a significant lag correlation with the upstream in terms of structure, the gradual change adjustment factor increases, causing the risk rate to steadily rise according to the gradual change curve. This reliably captures gradual evolution without relying on evidence of sudden window events. When evidence in either dimension is weak, the gradual change adjustment factor automatically decreases, suppressing short-term noise and occasional anomalies.

[0083] Thus, the analysis of cross-window stability and gradual structural change characteristics of metadata change events has been completed, and the gradual change adjustment factor has been obtained.

[0084] Step S4: Obtain the dynamic hazard rate by jointly analyzing the hazard rate adjustment factor and the slow change adjustment factor.

[0085] In steps S2 and S3, the hazard rate regulation characteristics are analyzed and evaluated from the perspectives of mutation detection and gradual change detection, respectively. After obtaining the hazard rate regulation factor and the gradual change regulation factor, the hazard rate can be jointly regulated using the hazard rate regulation factor and the gradual change regulation factor. Specifically, a base hazard rate constant is set, which is set to 0.02 in this embodiment of the invention. The calculation result of adding the hazard rate regulation factor of the target node at the target time and the gradual change regulation factor of the target node at the target time is subjected to an exponential mapping with the natural constant as the base. The result of multiplying the corresponding mapping result by the base hazard rate constant is used as the dynamic hazard rate of the target node at the target time.

[0086] No. The node at the th The formula for calculating the dynamic hazard rate at a given time is:

[0087]

[0088] in, Indicates the first The node at the th Dynamic risk rate at any given moment; Indicates the first The node at the th Risk rate adjustment factor at each time point; Indicates the first The node at the th The slowly varying regulatory factor at each moment; This represents the basic hazard rate constant.

[0089] This completes the process of obtaining dynamic hazard rates through joint analysis of hazard rate adjustment factors and slow-change adjustment factors.

[0090] Step S5: Obtain the comprehensive detection output by combining the dynamic hazard rate and the post-detection distribution.

[0091] In steps S2 and S3, a hazard rate adjustment factor to characterize the mutation response and a gradual change adjustment factor to characterize the progressive evolution were constructed, respectively. In step S4, the dynamic hazard rate of the node in both the time and structural domains was obtained through joint analysis of the hazard rate adjustment factor and the gradual change adjustment factor. Based on this, this step further introduces the dynamic hazard rate into the posterior evaluation process of the Bayesian online change point detection algorithm to form a comprehensive detection output, thereby achieving adaptive identification of real changes in the dynamic evolution environment of enterprise business metadata.

[0092] In the implementation of the standard Bayesian online change point detection algorithm, the detection logic is recursively updated based on the runtime distribution. Its core is the calculation of the posterior probability of whether a change point has occurred at the current time. Combined with the dynamic risk rate obtained in step S4 of this invention, the comprehensive detection posterior of the node at the target time can be expressed as:

[0093]

[0094] in, Indicates the first The node at the th The overall posterior probability of detection at each moment; Indicates the first The node at the th Dynamic risk rate at any given moment; Indicates the first The observational likelihood at a given moment when a change point is assumed to occur; Indicates the first The observational likelihood at each time step assuming a continuous state; Indicates the first The node at the th Metadata observation index sequence at each time point; Indicates the first The node at the th The length of the operation at each moment.

[0095] The core of the comprehensive detection posterior calculation formula lies in introducing a dynamic hazard rate, which makes the update of the posterior probability more accurate. In the traditional Bayesian online change point detection process, the hazard rate is fixed as a constant. When faced with different types of changes in a multi-source heterogeneous system, the algorithm cannot automatically adjust its response according to the context risk. This invention, through the dynamic replacement of the dynamic hazard rate, allows the detection process to be jointly regulated by a hazard rate adjustment factor and a gradual change adjustment factor. When the hazard rate adjustment factor is high, it indicates that a sudden event has hit a critical node and has high propagation potential. The hazard rate increases, the comprehensive detection posterior probability increases, and the model prioritizes its identification as a new change point. When the gradual change adjustment factor is high, it indicates that the system is in a slow evolution process. Although the algorithm does not understand the triggering of high-intensity alarms, the continuous maintenance of the hazard rate makes the comprehensive detection posterior probability show a steady upward trend, thereby accumulating trend signals. When both the hazard rate adjustment factor and the gradual change adjustment factor are high, the system has significant changes in both the time and structural domains. The comprehensive detection posterior probability quickly crosses the domain alarm threshold, achieving real-time identification.

[0096] Through the comprehensive detection output mechanism in this step, the Bayesian online change point detection algorithm no longer relies on a fixed risk rate assumption, but rather dynamically and adaptively updates under the joint drive of the risk rate adjustment factor and the gradual change adjustment factor. This approach not only maintains the real-time performance and mathematical rigor of traditional Bayesian online change point detection algorithms, but also significantly enhances the model's ability to perceive scenarios where abrupt changes and gradual changes coexist in multi-source heterogeneous enterprise environments, achieving an intelligent detection transition from statistical driving to scenario-driven approaches. The final comprehensive detection posterior accurately reflects the structural state changes of each metadata node, providing a reliable basis for automated change identification in enterprise metadata management systems.

[0097] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent management of enterprise business metadata based on multi-source heterogeneous data, characterized in that: The method includes: Step S1: By performing event and lineage correlation analysis on the enterprise's multi-source heterogeneous data, a dynamic monitoring dataset of enterprise business metadata is obtained; Step S2: Obtain the risk rate adjustment factor by jointly analyzing the temporal proximity and impact domain propagation characteristics of metadata change events; Step S3: Obtain the gradual change adjustment factor by performing a fusion analysis on the cross-window stability and structural gradual change characteristics of metadata change events; Step S4: Obtain the dynamic hazard rate by jointly analyzing the hazard rate adjustment factor and the slow-change adjustment factor; Step S5: Obtain the comprehensive detection output by combining the dynamic hazard rate and the post-detection distribution.

2. The intelligent management method for enterprise business metadata based on multi-source heterogeneous data according to claim 1, characterized in that, The process involves performing event and lineage correlation analysis on multi-source heterogeneous data of the enterprise to obtain a dynamic monitoring dataset of enterprise business metadata, including: By setting up log collection and audit modules in various business systems of the enterprise, multi-source heterogeneous data from enterprise resource planning system, customer relationship management system, production execution system, data warehouse system and report analysis platform are collected synchronously. Extract metadata change times from the structure change logs and audit records of the above systems to form an event set containing the structure definition changes, with timestamps as the primary key. Extract the upstream dependency set of the target field or metric and the set of nodes affected by the change event from the task scheduling system and interface gateway, and construct the upstream node set and event node set respectively; Obtain a lineage graph containing the data flow relationships between various fields and indicators from the enterprise data governance platform, and calculate the betweenness centrality, mean betweenness centrality, and standard deviation of betweenness centrality for each node in the lineage graph; Extract release event interval window data from the enterprise release system and version control logs, and calculate and evaluate the time offset between the timestamp of each change event and the center of the release event window. The collected event sets, upstream node sets, event node sets, lineage graph structure parameters, and time offset evaluation results are organized and managed to form a dynamic monitoring dataset of enterprise business metadata containing event information, dependent node relationships, and time characteristics.

3. The intelligent management method for enterprise business metadata based on multi-source heterogeneous data according to claim 1, characterized in that, The method involves jointly analyzing the temporal proximity and impact domain propagation characteristics of metadata change events to obtain a risk rate adjustment factor, including: By performing proximity analysis on the release window time data and change event time data, the event time matching degree evaluation results are obtained; By analyzing the propagation characteristics of event-affected node data and kinship structure data, the impact intensity assessment results of node structure are obtained. The event time matching degree assessment results are then weighted and optimized using the impact intensity assessment results of node structure to obtain the risk rate adjustment factor.

4. The intelligent management method for enterprise business metadata based on multi-source heterogeneous data according to claim 3, characterized in that, The step of performing nearest neighbor analysis on the release window time data and change event time data to obtain the event time matching degree evaluation result includes: Set the time kernel width parameter. For any change event of any target node at any target time, divide the time difference between the change event time of the target node at the target time and the time at the center of the publishing window by the time kernel width parameter as the time difference assessment of the target node at the target time. Divide the square of the time difference assessment of the target node at the target time by the negative of the result of the constant 2 and perform an exponential mapping with the natural constant as the base. Use the corresponding mapping result as the event time matching degree assessment result of the target node at the target time.

5. The intelligent management method for enterprise business metadata based on multi-source heterogeneous data according to claim 3, characterized in that, The process involves analyzing the propagation characteristics of event-affected node data and kinship structure data to obtain an assessment result of the node structure's influence intensity. This assessment result is then used to weight and optimize the event time-matching degree assessment result, resulting in a risk rate adjustment factor, including: For any change event of a target node at any target time, the number of nodes in the intersection of the set of upstream dependent nodes of the target node's corresponding indicator or field and the set of nodes affected by the change event of the target node at the target time is used as the numerator, and the number of nodes in the set of upstream dependent nodes of the target node's corresponding indicator or field is used as the denominator. The resulting fraction is used as the time hit rate of the target node at the target time. Set a centrality standardization adjustment coefficient; use the result of subtracting the betweenness centrality assessment of the target node from the average betweenness centrality of all nodes in the kinship graph as the numerator, and the standard deviation of the betweenness centrality of all nodes in the kinship graph as the denominator, and use the corresponding fraction as the centrality standardization assessment of the target node; use the result of multiplying the centrality standardization adjustment coefficient by the centrality standardization assessment of the target node as the first centrality assessment of the target node; apply the Sigmoid function mapping to the first centrality assessment of the target node, and use the corresponding mapping result as the structural propagation ability assessment of the target node; The result of multiplying the temporal hit rate of the target node at the target time by the structural propagation capability assessment of the target node is taken as the structural influence intensity of the target node at the target time. The result of multiplying the structural influence intensity of the target node at the target time by the event time matching degree assessment result of the target node at the target time is used as the risk rate adjustment factor of the target node at the target time.

6. The intelligent management method for enterprise business metadata based on multi-source heterogeneous data according to claim 1, characterized in that, The method involves fusing analysis of cross-window stability and gradual structural change characteristics of metadata change events to obtain gradual adjustment factors, including: By performing directional consistency and trend analysis on the structural influence intensity sequence data within a continuous time window, the cross-window trend consistency assessment results are obtained. By performing hysteresis correlation analysis on the structural influence intensity sequence data of the target node and its upstream nodes, the local structural inertia assessment results are obtained. By fusing cross-window trend consistency assessment data and local structural inertia assessment data, a gradual adjustment factor is obtained.

7. The intelligent management method for enterprise business metadata based on multi-source heterogeneous data according to claim 1, characterized in that, The step involves performing directional consistency and trend analysis on the structural influence intensity sequence data within a continuous time window to obtain cross-window trend consistency assessment results, including: The calculation expression for the cross-window trend consistency assessment of any target node at any target time is as follows: ； in, Indicates the first The node at the th Consistency assessment of cross-window trends at each moment; Indicates the window length; Indicates the sign of a function, used to determine the sign of the calculation result within the function; Indicates the first The node at the th The sequence of structural influence intensity at each time point; Indicates the first The node at the th The slope estimate obtained by linear regression of the structural influence intensity sequence at each time point on the event index; Indicates the first The node at the th The standard deviation of the structural influence intensity sequence at each time point; It represents a very small positive number.

8. The intelligent management method for enterprise business metadata based on multi-source heterogeneous data according to claim 1, characterized in that, The process of obtaining local structural inertia assessment results by performing hysteresis correlation analysis on the structural influence intensity sequence data of the target node and its upstream nodes includes: The calculation expression for the local structural inertial assessment result of any target node at any target time is: ； in, Indicates the first The node at the th Local structural inertial assessment at each moment; Indicates the lag length; Indicates the maximum lag length; Indicates the first The node at the th The length of each time period is The structural influence intensity sequence; Indicates the first The node at the th The length of each time period is The structural influence intensity sequence; Indicates the number in the kinship chart The node to the first The bloodline edge weight of each node; Indicates the number in the kinship chart The set of upstream dependent nodes of each node; This represents the function for calculating the Pearson correlation coefficient.

9. The intelligent management method for enterprise business metadata based on multi-source heterogeneous data according to claim 1, characterized in that, The step of fusing cross-window trend consistency assessment data and local structural inertia assessment data to obtain a gradual adjustment factor includes: Set trend assessment weights and inertia assessment weights; for any change event of any target node at any target time, the cross-window trend consistency assessment and the local structural inertia assessment of the target node at the target time are weighted and summed by the trend assessment weights and inertia assessment weights, respectively, and the corresponding weighted summation results are used as the gradual change adjustment factor of the target node at the target time.

10. The intelligent management method for enterprise business metadata based on multi-source heterogeneous data according to claim 1, characterized in that, The process of obtaining dynamic hazard rates through joint analysis of hazard rate adjustment factors and slow-change adjustment factors includes: Set a basic hazard rate constant; perform an exponential mapping with the natural constant as the base of the calculation result of adding the hazard rate adjustment factor of the target node at the target time and the slowly changing adjustment factor of the target node at the target time; and multiply the corresponding mapping result by the basic hazard rate constant as the dynamic hazard rate of the target node at the target time.

Images