Data governance method based on industrial big data platform

Abstract

The application relates to the technical field of data governance, and discloses a data governance method based on an industrial big data platform, which comprises the following steps: processing organizational structure documents and business process definitions through a text semantic analysis technology to generate a responsibility allocation matrix; analyzing user historical data access behaviors to characterize user data use characteristics; according to a user model and the responsibility matrix, combining data sensitivity scoring, scene characteristic extraction and multi-dimensional risk calculation to realize dynamic permission decision; aiming at the decision mechanism, verifying the safety influence of the authorization strategy through simulation testing; according to the test results, adaptively adjusting the model parameters and dynamically updating the initial responsibility matrix to form a closed-loop management; through organizational structure intelligent analysis, user behavior modeling, risk perception-based permission decision, digital twin verification and adaptive optimization, the accurate division and dynamic adjustment of data governance responsibilities are realized, and the intelligentization and dynamic adaptation capability of data permission management are improved.

Description

Technical Field

[0001] This invention relates to the field of data governance technology, and more specifically, to a data governance method based on an industrial big data platform. Background Technology

[0002] With the rapid development of industry and intelligent manufacturing, industrial big data platforms have become an important infrastructure for enterprise digital transformation, containing a large amount of sensitive data such as production process parameters and R&D experimental data. Existing data governance methods mainly rely on role-based access control (RBAC) or attribute-based access control (ABAC) models, which manage permissions through predefined static rules.

[0003] However, these traditional methods have several technical problems: the responsibility allocation mechanism relies on manual definition, which makes it difficult to reflect changes in organizational structure in a timely manner, resulting in unclear division of responsibilities; static permission rules cannot cope with dynamically changing access scenarios and risk conditions, either over-restricting and affecting data usage efficiency, or providing insufficient protection and causing security risks; and there is a lack of a pre-assessment mechanism for the impact of authorization decisions, which may lead to unexpected security risks or efficiency losses after permission adjustments.

[0004] The aforementioned technical issues severely restrict the secure and efficient use of data on industrial big data platforms, necessitating a new data governance method capable of intelligently analyzing organizational structures, dynamically sensing access risks, pre-assessing security impacts, and adaptively optimizing. Summary of the Invention

[0005] This invention provides a data governance method based on an industrial big data platform, which solves the technical problems of unclear responsibility allocation, lack of adaptability of static permission management, and lack of security impact assessment of authorization decisions in related technologies.

[0006] This invention provides a data governance method based on an industrial big data platform, including:

[0007] Organizational structure documents and business process definitions are processed using text semantic analysis technology to generate a responsibility assignment matrix;

[0008] Analyze users' historical data access behavior, construct a time-series model of user operation patterns, and characterize user data usage features;

[0009] Based on the user model and responsibility matrix, combined with data sensitivity scoring, scenario feature extraction and multi-dimensional risk calculation, dynamic permission decision-making is achieved;

[0010] Based on dynamic permission decisions, a virtual environment for data interaction is constructed, and the security impact of the authorization policy is verified through simulation testing.

[0011] Based on the simulation test results, the model parameters are adaptively adjusted, and the initial responsibility matrix is ​​dynamically updated to form a closed-loop management system.

[0012] Furthermore, the steps for generating a responsibility assignment matrix by processing organizational structure documents and business process definitions using text semantic analysis technology include:

[0013] Process unstructured text to extract role entities, responsibility entities, and relationships; unstructured text includes organizational structure documents, job descriptions, and business process documents;

[0014] Construct an organizational structure knowledge graph and calculate node centrality and influence metrics;

[0015] By combining the results of the graph analysis with the business process definition, a RACI responsibility matrix is ​​generated.

[0016] Furthermore, the steps for analyzing users' historical data access behavior and constructing a time-series model of user operation patterns include:

[0017] Collect user data access logs, and generate a structured user behavior dataset after cleaning and preprocessing;

[0018] The preprocessed behavioral data is organized into a time series and modeled using a deep learning model that combines long short-term memory networks and attention mechanisms.

[0019] Cluster analysis is performed on users with similar behavioral patterns to identify typical behavioral types and patterns.

[0020] Furthermore, based on data sensitivity scoring, scene feature extraction, and multi-dimensional risk calculation, the steps to achieve dynamic permission decision-making include:

[0021] Classify and assess the sensitivity of data assets, and construct a sensitivity scoring model that comprehensively considers the type, value, and scope of impact of data content;

[0022] Real-time extraction of access scenario features;

[0023] Multidimensional risk calculation and decision-making: Based on user behavior models, data sensitivity and scenario characteristics, it calculates access risk scores and determines authorization based on the comparison results of scores and thresholds.

[0024] Furthermore, the steps of constructing a virtual environment for data interaction and verifying the security impact of the authorization policy through simulation testing include:

[0025] Create virtual copies of data assets, user simulation agents, and interactive behavior simulators to form a virtual environment;

[0026] The virtual environment simulates user access behavior to data of varying sensitivity.

[0027] We use causal inference to analyze the impact of different permission policies on data security and business efficiency.

[0028] Furthermore, the steps for adaptively adjusting model parameters based on governance effectiveness monitoring results include:

[0029] Construct a data governance effectiveness evaluation model that includes basic indicators, process indicators, and outcome indicators;

[0030] The parameters of the risk assessment model are adaptively adjusted using a reinforcement learning algorithm, and the adjustment process is modeled as a Markov decision process.

[0031] Furthermore, the steps for dynamically updating the responsibility matrix include:

[0032] Forced updates based on organizational structure changes ensure that the responsibility matrix remains consistent with the current organizational structure.

[0033] The system is optimized and updated based on the results of the governance effectiveness evaluation. The responsibility allocation scheme is optimized by analyzing the relationship between responsibility allocation and governance effectiveness.

[0034] Furthermore, the computation process of the Long Short-Term Memory network includes:

[0035]

[0036] f t =σ(W f ·(h t-1 x t )+b f );

[0037]

[0038]

[0039] o t =σ(W o ·(h t-1 x t )+b o );

[0040] h t =o t ⊙tanh(c t );

[0041] in f t and o t These represent the values ​​of the input gate, forget gate, and output gate, respectively. t-1 and h t Let X represent the hidden states at time t-1 and time t, respectively. tc represents the input at time t. t and c t-1 These represent the cell states at time t and time t-1, respectively. These are candidate unit states, ⊙ represents element-wise multiplication, σ is the sigmoid activation function, and W... i W f W c W o Let b represent the weight matrices for the input gate, forget gate, cell state, and output gate, respectively. i b f b c b o represents the bias terms of the input gate, forget gate, cell state, and output gate, respectively, and tanh represents the hyperbolic tangent activation function.

[0042] Furthermore, the multidimensional risk calculation and decision-making are implemented using a gradient boosting decision tree model. Authorization is determined based on a comparison between the risk score and a preset threshold. The implementation of the gradient boosting decision tree model includes:

[0043] Construct a combined feature vector, combining user behavior features, data sensitivity features, scenario features, and corresponding values ​​from the responsibility matrix into a unified feature vector:

[0044] X = (B u S(d), C scene M RACI (u, d));

[0045] Where X is the combined feature vector, B u S(d) is the behavioral feature vector of user u, and S(d) is the sensitivity feature vector of data d. scene It is a scene feature vector, M BACI (u, d) is the responsibility value of user u for data d in the responsibility matrix;

[0046]

[0047] in This is the output value of the m1-th tree. It is the number of leaf nodes in tree m1. This represents the feature space region corresponding to the j1st leaf node in the m1th tree. It is the output value of the j1-th leaf node in the m1-th tree, I ind (·) is an indicator function. This represents all trees in tree m1. Summing is performed on each leaf node;

[0048] The model's final prediction is the sum of all the trees;

[0049] The structure of the new tree is determined by minimizing the loss function.

[0050] A data governance system based on an industrial big data platform is used to execute the data governance methods described above, including:

[0051] The intelligent organizational structure analysis module is used to process organizational structure documents and business process definitions through text semantic analysis to generate a responsibility allocation matrix.

[0052] The user behavior modeling module is used to analyze users' historical data access behavior and build a time-series model of user operation patterns.

[0053] The risk perception and decision-making module is used to make dynamic permission decisions based on data sensitivity scores, scenario characteristics, and multi-dimensional risk calculations.

[0054] The digital twin verification module is used to construct a virtual environment for data interaction and verify the security impact of authorization policies.

[0055] The adaptive optimization module is used to adjust model parameters and update the responsibility matrix based on the monitoring results of governance effectiveness.

[0056] The processor and memory are used to perform the calculations and data storage of the above modules.

[0057] The beneficial effects of this invention are as follows: through intelligent analysis of organizational structure, user behavior modeling, risk-aware permission decision-making, digital twin verification, and adaptive optimization, an intelligent data governance method based on an industrial big data platform is realized, which solves the key technical problems of unclear responsibility allocation, static and rigid permission management, and lack of prior security assessment in traditional data governance.

[0058] It enables precise division and dynamic adjustment of data governance responsibilities, reducing the error rate of responsibility allocation by 42% and the response time for responsibility updates by 65%.

[0059] It improves the intelligence and dynamic adaptability of data access control, reducing unnecessary access restrictions by 38% while maintaining the same security level, thus improving data utilization efficiency.

[0060] It achieves pre-emptive security verification for access control decisions, which can identify 85% of potential access risk points in advance, reducing the incidence of risk events by 56% compared to traditional post-event monitoring methods. Attached Figure Description

[0061] Figure 1 This is a flowchart of the data governance method based on an industrial big data platform in this invention;

[0062] Figure 2 This is a flowchart of the intelligent analysis steps of the organizational structure in this invention;

[0063] Figure 3 This is a flowchart of the user behavior modeling steps of the present invention;

[0064] Figure 4 This is a flowchart of the authorization decision-making steps for risk perception in this invention;

[0065] Figure 5 This is a flowchart of the digital twin verification steps of the present invention;

[0066] Figure 6 This is a flowchart of the adaptive optimization steps of the present invention. Detailed Implementation

[0067] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0068] At least one embodiment of the present invention discloses a data governance method based on an industrial big data platform, such as... Figures 1 to 6 As shown, it includes the following steps:

[0069] Step 1: Process organizational structure documents and business process definitions using text semantic analysis technology to generate a responsibility assignment matrix;

[0070] This step utilizes text semantic analysis algorithms and graph structure analysis models to process enterprise organizational structure documents and business process definition files, generating a structured responsibility allocation matrix; the specific implementation includes the following sub-steps:

[0071] Sub-step 1.1, text semantic analysis;

[0072] Natural language processing techniques are applied to process unstructured text to extract key entities and relationships; unstructured text includes organizational structure documents, job descriptions, and business process documents.

[0073] In the steps of this document:

[0074] First, the text is preprocessed, including word segmentation, stop word removal, and part-of-speech tagging.

[0075] Then, a named entity recognition algorithm is used to identify role entities (such as departments, positions, personnel, etc.) and responsibility entities in the text;

[0076] Finally, a relation extraction algorithm is applied to extract management relationships and business collaboration relationships between entities. The relation extraction uses a dependency parsing-based model, represented as follows:

[0077] R e =f relation (E1, E2, C) org );

[0078] Where R e Indicates the type of relationship between entity E1 and entity E2, C org f represents the context feature vector in organizational structure analysis. relation Represents a relational classification function.

[0079] Sub-step 1.2: Organization map construction;

[0080] Based on the entities and relationships extracted in the previous step, an organizational structure knowledge graph is constructed, which is formally represented as a directed graph:

[0081] G o = (V, E, T) rel );

[0082] Among them G o This represents an organizational graph, where V represents the set of nodes (including entities such as departments, roles, and personnel), E represents the set of edges (representing the relationships between entities), and T... rel Represents a set of relation types.

[0083] Graph structure analysis algorithms are applied to the constructed organizational graph to calculate key metrics such as centrality and influence of each node.

[0084]

[0085] Where C a Let w represent the centrality of node a, N(a) represent the set of all nodes connected to node a, and w represent the centrality of node a. ab A represents the weight of the relationship between node a and node b. inf This represents the influence index of a node, and ∑ represents the summation symbol.

[0086] Sub-step 1.3: Generation of the RACI responsibility matrix;

[0087] Based on the organizational chart analysis results and business process definitions, a RACI (Responsible, Approving, Consulting, Informed) responsibility matrix is ​​generated. This matrix describes the specific responsibility types of each role in data governance activities, and is formally represented as follows:

[0088] M RACI ={m ij}, m ij ∈{R, A, C, I, -};

[0089] Where M RACI Denotes the RACI responsibility matrix, m ij This indicates the type of responsibility of role i for activity j, where R means responsible, A means approved, C means consulted, I means informed, and - means unrelated;

[0090] The matrix generation process uses a rule-based and graph pattern matching algorithm to map organizational relationships, business processes, and data governance activities.

[0091] In this formula, if role i has a direct relationship with activity j, then m ij Let m be one of R, A, C, I; if role i has no direct relationship with activity j, then m ij for-.

[0092] The RACI responsibility matrix M generated in this sub-step RACI As a key output, M RACI It is a two-dimensional matrix. The rows of the matrix represent various roles in the organization (such as departments, positions, and individuals); the columns of the matrix represent various activities or decisions related to data governance.

[0093] Step 2: Analyze users' historical data access behavior, construct a time-series model of user operation patterns, and characterize user data usage features;

[0094] This step involves analyzing users' historical data access and operation records to construct a user behavior pattern model, thereby achieving a precise characterization of user data usage characteristics. Specifically, it includes the following sub-steps:

[0095] Sub-step 2.1: Data access log collection and preprocessing;

[0096] Data access logs from users are collected from industrial big data platforms, including access time, data type, operation type, and results. The collected raw logs are cleaned, standardized, and feature extracted to generate a structured user behavior dataset. The feature extraction process includes multi-dimensional information such as time features (e.g., access time period, frequency), operation features (e.g., query, modification, deletion operation types and proportions), and data type features (e.g., accessed data categories and sensitivity).

[0097] Sub-step 2.2, constructing the behavior sequence model;

[0098] The preprocessed user behavior data is organized into a time series, and a sequence model is used to model user operation patterns. Specifically, this is implemented as a time-series deep learning model, which combines a Long Short-Term Memory (LSTM) network and an attention mechanism, and is represented as follows:

[0099] h t =LSMT(xt h t-1 c t-1 );

[0100]

[0101] Where x t h represents the user behavior feature vector at time t. t and c t These are the hidden state and cell state of the LSTM, respectively, h t-1 and c t-1 Let α represent the hidden state and the cell state at time t-1, respectively. att It is attention weight, C ctx B represents organizational context information in user behavior modeling. user Let f represent the user behavior pattern vector, and let f represent the user behavior pattern vector B. user The generating function is defined by: `Attention` (representing the attention mechanism), `LSTM` (representing the Long Short-Term Memory network model), and `f` (representing the generating function for user behavior pattern vectors); `h1`, `h2`, ... These represent the LSTM network at time step 1, time step 2, and the last time step T, respectively. seq The hidden state vector, T seq The length of the temporal sequence is represented by these elements, which together constitute the temporal representation of the user behavior sequence and are used for contextual information extraction in the attention mechanism computation.

[0102] The LSTM part of this model consists of input gates, forget gates, output gates, and cell states. The detailed calculation process is as follows:

[0103]

[0104] f t =σ(W f ·(h t-1 x t )+b f );

[0105]

[0106] o t =σ(W o ·(h t-1 x t )+b o );

[0107] h t =o t ⊙tanh(c t );

[0108] in ft and o t These represent the values ​​of the input gate, forget gate, and output gate, respectively. t-1 and h t Let x represent the hidden states at time t-1 and time t, respectively. t c represents the input at time t. t and c t-1 These represent the cell states at time t and time t-1, respectively. These are candidate unit states, ⊙ represents element-wise multiplication, σ is the sigmoid activation function, and W... i W f W c W o Let b represent the weight matrices for the input gate, forget gate, cell state, and output gate, respectively. o b f b c b o represents the bias terms of the input gate, forget gate, cell state, and output gate, respectively, and tanh represents the hyperbolic tangent activation function.

[0109] The attention mechanism is calculated as follows:

[0110] e t =v·tanh(W) h ·h t +W s ·s t-1 );

[0111]

[0112] Where e t and e n h represents the attention scores at time steps t and n, respectively. t Let s represent the hidden state vector at time t. t-1 α represents the decoder state at time t-1. att Let W represent the attention weights, where v represents the weight vector of the attention mechanism. h W represents the weight matrix of the hidden state. s The weight matrix represents the decoder state, tanh represents the hyperbolic tangent activation function, exp represents the natural exponential function, and T... seq Indicates the length of the time sequence.

[0113] The following is a specific application example of this model in an industrial big data platform environment: Analyzing the data access behavior of a steel plant, a time-series model of user behavior is constructed by extracting user access sequences for different types of data, such as production parameters, quality data, and equipment status. The model can identify normal access patterns, such as the fixed pattern of quality inspectors regularly viewing quality data, as well as abnormal behavior, such as production personnel suddenly exporting large amounts of commercially sensitive data. Through this model, the time-series patterns, preferences, and abnormal behavioral characteristics of user data access can be captured.

[0114] Sub-step 2.3: Behavioral pattern clustering and profile generation;

[0115] Based on the constructed user behavior model, cluster analysis is performed on users with similar behavior patterns to identify typical user behavior types and patterns; an improved K-means clustering algorithm is used to group user behavior vectors.

[0116]

[0117] Where S clust S represents the clustering result. cl，ic Let μ represent the i-th cluster. ic X is the center of the ic-th cluster. jc It is a user behavior vector, R reg (S clust ) is a regularization term that considers organizational structure information, λ clust It is the balance parameter in clustering, where k represents the total number of clusters, ||x jc -μ ic || 2 Represents vector x jc With cluster center μ ic The square of the Euclidean distance between them This represents summing over all k clusters (from ic=1 to ic=k). Indicates the cluster S cl，ic All vectors x jc Summation, This represents the clustering result S that minimizes the objective function. clust ;

[0118] Based on the clustering results, a behavioral profile is generated for each user or user group, including information on typical operating patterns, data access preferences, risk tendencies, and other dimensions.

[0119] In this method, the generic user behavior pattern vector B generated in step 2.2 user B is used to represent the overall behavior model, while for a specific user u, its personalized behavior model is denoted as B. u It is based on B userAn individualized model derived from the user's specific behavioral characteristics and clustering results.

[0120] The user behavior model generated in this step (including the general model B) user And user individual model B u () is an important input for subsequent steps.

[0121] Step 3: Based on the user model and responsibility matrix, combined with data sensitivity scoring, scenario feature extraction and multi-dimensional risk calculation, dynamic permission decisions are made.

[0122] This step, based on the responsibility matrix and user behavior model obtained in the previous two steps, combines data sensitivity and access scenario information to construct a multi-dimensional risk assessment model and achieve dynamic permission decisions; specifically, it includes the following sub-steps:

[0123] Sub-step 3.1, Data asset sensitivity assessment;

[0124] The data assets in the industrial big data platform are classified and their sensitivity assessed, and a data sensitivity scoring model is constructed. This model comprehensively considers factors such as data content type (e.g., personal information, trade secrets, production parameters), data value, and the scope of data impact to calculate the sensitivity score for each type of data. The sensitivity scoring function is expressed as follows:

[0125] S(d) = w1·S type (d)+w2·S value (d)+w3·S impact (d);

[0126] Where S(d) represents the sensitivity score of data d, S type (d), S value (d) and S impact (d) represents the scoring functions for the data type, value, and scope of influence of data d, respectively, where w1, w2, and w3 are the weight parameters for the data type, value, and scope of influence.

[0127] Sub-step 3.2: Access scene feature extraction;

[0128] Real-time extraction and analysis of data access scenario features, including access time, access location, access device, network environment, and other factors; through feature engineering techniques, these scenario features are converted into structured feature vectors, which serve as input for risk assessment; the scenario feature vector is represented as follows:

[0129]

[0130] Where C scene Representing scene feature vectors, c1, c2, They represent the 1st, 2nd, and nth respectively.feat The values ​​of scene features, such as time anomaly, location anomaly, device reliability, etc., n feat This represents the scene feature dimension, and T represents the transpose operator.

[0131] Sub-step 3.3, Multidimensional Risk Calculation and Decision-Making;

[0132] Based on user behavior models, data sensitivity, and scenario characteristics, a risk assessment model is constructed. The input to this step includes the RACI responsibility matrix M generated in sub-step 1. RACI Step 2 generates user behavior model B u The behavioral profile, the sensitivity score S(d) generated in step 3.1, and the scene feature vector C generated in step 3.2 are used. scene ;

[0133] Real-time calculation of risk scores for users accessing specific data:

[0134] R(u, d, c) = f risk (B user S(d), C scene M RACI );

[0135] Where R(u, d, c) represents the risk score of user u accessing data d in scenario c, and B u Here is the behavioral model of user u, S(d) is the data sensitivity of data d, and C... scene It is a scene feature vector; M RACI It is a responsibility matrix, f risk It is a risk scoring function, implemented as a gradient boosting decision tree model.

[0136] The specific implementation of this gradient boosting decision tree model is as follows:

[0137] Construct a combined feature vector, combining user behavior features, data sensitivity features, scenario features, and corresponding values ​​from the responsibility matrix into a unified feature vector:

[0138] X = (B u S(d), C scene M RACI (u, d));

[0139] Where X is the combined feature vector, B u S(d) is the behavioral feature vector of user u, and S(d) is the sensitivity feature vector of data d. scene It is a scene feature vector, M RACI (u, d) is the responsibility value of user u for data d in the responsibility matrix.

[0140] Ensemble learning is performed using multiple decision trees, with each tree outputting an estimate of the risk score;

[0141] The m1th tree can be represented as:

[0142]

[0143] in This is the output value of the m1-th tree. It is the number of leaf nodes in tree m1. This represents the feature space region corresponding to the j1st leaf node in the m1th tree. It is the output value of the j1-th leaf node in the m1-th tree, I ind The · sign is an indicator function (it takes the value 1 when the condition is true, and 0 otherwise). This represents all trees in tree m1. Summing is performed on the leaf nodes (from j1=1 to ...). ).

[0144] The model's final prediction is the sum of all the trees:

[0145]

[0146] Where f risk (X) is the model's final predicted value, m1 is the index of the current tree, and M... trees It is the total number of decision trees. This indicates that for all M trees Summing the outputs of the trees (from m1=1 to m1=M) trees ).

[0147] The model is trained using a forward successive addition modeling approach.

[0148]

[0149] in This is the predicted value for the m1-th tree. These are the predicted values ​​for the first m1-1 trees. It is the predicted value for the m1th tree.

[0150] For each step, the structure of the new tree is determined by minimizing the loss function:

[0151]

[0152] in Let L be the predicted value of the m1-th tree, and L be the loss function. In this implementation, the mean squared error is used as the loss function. sample It is the number of training samples. It is the predicted value of the i1th sample. It is the true label of the i1th sample. It is the feature vector of the i1th sample. This means for all n sample Summing is performed on each training sample (from i1=1 to i1=n). sample ), argmin h Let h represent the function that minimizes the loss function.

[0153] In practical applications, the model assigns different weights to access behaviors with different risk levels. For example, it assigns higher weights to high-risk operations that may lead to data breaches, ensuring that the model is more sensitive to high-risk behaviors.

[0154] For example, in the industrial big data platform of a chemical company, this risk assessment model analyzes access requests for production formula data. When R&D personnel access the formula data they are responsible for using the company's secure terminals during normal working hours, the model gives a low-risk score; however, when the same R&D personnel attempt to download large amounts of sensitive formula data, including data outside their responsibility, outside of working hours using non-company devices, the model gives a high-risk score, triggering access denial or multi-factor authentication requirements. During the company's trial operation, this model successfully identified and prevented 12 potential data breaches, while keeping the false interception rate in daily operations below 5%.

[0155] Based on the calculated risk score and the preset risk threshold θ, dynamic permission decisions are made.

[0156]

[0157] Where R(u, d, c) is the risk score, A(u, d, c) = 1 indicates granting permission, and A(u, d, c) = 0 indicates denying permission; the risk threshold θ can be adjusted according to the organization's security policy and the importance of the data.

[0158] The authorization decision result A(u, d, c) and risk score R(u, d, c) of this step are important outputs. This step also receives the responsibility matrix M provided in step 1 (Intelligent Analysis of Organizational Structure). RACI And the user behavior model B provided in step 2 (user behavior modeling) u As input, upstream and downstream dependencies are formed in the data flow.

[0159] Step 4: Based on dynamic permission decisions, construct a virtual environment for data interaction and verify the security impact of the authorization policy through simulation testing;

[0160] This step constructs a virtual environment for data interaction, and uses simulation technology to verify the effectiveness and security impact of the authorization policy, thereby achieving pre-verification of authorization decisions; specifically, it includes the following sub-steps:

[0161] Sub-step 4.1: Construction of the virtual environment for data interaction;

[0162] Based on the real data structure and access patterns of the industrial big data platform, a virtual environment for data interaction is constructed.

[0163] Input: The RACI responsibility matrix M generated in step 1 RACI This is used to construct the permission mapping relationship of users in the virtual environment; the user behavior model generated in step 2 is used to generate the virtual user agent U. virtual Provide behavioral pattern B class ;

[0164] This environment includes components such as virtual copies of data assets, user simulation agents, and interaction behavior simulators, enabling permission policy testing without affecting the actual production system; the formal representation of the virtual environment is as follows:

[0165] E virtual ={D virtual U virtual I virtual P test};

[0166] Where E virtual D represents a virtual environment. virtual U represents a collection of virtual data assets. virtual I represents the set of virtual user agents. virtual P represents the set of virtual interactive behaviors. test This represents the set of permission policies to be tested; the virtual environment is constructed using a domain-specific language-based representation method, mapping key elements from the real system to the virtual environment.

[0167] Specifically, the implementation of a virtual environment includes the following core components:

[0168] The data asset virtualization module is responsible for creating virtual copies of real data assets, preserving their structural characteristics and sensitive attributes, but replacing or masking the actual content to protect privacy. This is achieved by employing data anonymization techniques, transforming the original data while maintaining its distribution characteristics.

[0169]

[0170] Where D virtual It is a collection of virtual data assets. It refers to virtual data assets, i2 is the index of the virtual data asset, Anonymize() is the data anonymization function, and α dsm It is the desensitization parameter for the corresponding data type, D real It is a real data set;

[0171] Different anonymization strategies are used for different types of data. For example, noise is added to numerical data, value mapping is performed to replace categorical data, and synonyms are used to replace text data.

[0172] The user behavior simulator, based on the previously constructed user behavior model, generates virtual user agents that conform to statistical characteristics, simulating the data access behavior of different types of users:

[0173]

[0174] U virtual This represents a collection of virtual user agents. `GenerateAgent()` is the agent generation function. class It is a behavioral model of user categories, P cfg This refers to the permission configuration for this type of user, P. set It is a set of permission configurations.

[0175] The simulator uses a policy-based behavior generation algorithm to generate access sequences based on the historical behavior patterns of user types, while adding random mutations to simulate the uncertainty in the real environment.

[0176] The interaction event engine is responsible for simulating interactions between users and data assets, including operations such as querying, modifying, deleting, and exporting, and recording the interaction process and results.

[0177]

[0178] Among them I virtual Represents a set of virtual interactive events, u vir Indicates virtual user, d vir This represents virtual data, op represents the operation type, and t represents virtual data. op Indicates the operation time, r op The operation result is represented by OP, the set of operation types is represented by T1, and R is represented by R. result Represents the set of operation results;

[0179] The engine uses a discrete event simulation method to process events in a preset timeline order and trigger corresponding permission checks and operation executions.

[0180] For example, in a power industry data governance system, a virtual environment reflecting grid operation data, customer electricity consumption data, and electricity transaction data was constructed. The system simulates user access scenarios for different sensitive data levels by creating virtual user agents with different roles such as dispatchers, analysts, and marketing personnel. For instance, the system generated behavioral patterns of dispatchers frequently accessing grid operation parameters during peak load periods, and analysts regularly exporting summary reports but not frequently accessing detailed customer information. In this way, new access control policies, such as dynamic policies like "temporarily elevating dispatchers' access permissions to specific system parameters during grid emergencies," can be securely tested without exposing actual data. The virtual environment successfully simulated 95% of typical user behavior patterns and 87% of extreme access scenarios over a year, providing strong support for optimizing access control policies.

[0181] Sub-step 4.2, permission policy simulation test;

[0182] This sub-step simulates user access behavior to data of varying sensitivity in a virtual environment to test the response of the access control decision model. It directly receives the access control decision result A(u, d, c) and risk score R(u, d, c) from sub-step 3.3 as input, and evaluates the model's effectiveness by reproducing its operating conditions and parameters. Test scenarios include three categories: normal access scenarios, boundary condition scenarios, and extreme scenarios. Large-scale simulation tests are executed using an automated testing framework to collect decision results and key indicators. The testing process records logs for each interaction, including access request parameters, decision results, and decision basis, providing data support for subsequent analysis.

[0183] Sub-step 4.3, safety impact assessment;

[0184] Based on simulation test results, a security impact assessment model is constructed to analyze the impact of different access control policies on data security and business efficiency. This model employs a causal inference-based approach to characterize the causal relationship between access control policies and security risks and business efficiency.

[0185] I impact (P)=I security (P), I efficiency (P);

[0186] I security (P)=E(Y security |do(P));

[0187] I efficiency (P)=E(Y efficiency |do(P));

[0188] Among them I impact (P) represents the overall impact of the permission policy P, I security(P) and L efficiency (P) represents the impact on safety and efficiency, respectively, Y security and Y efficiency The corresponding outcome variables are do(P), which represents the intervention strategy variable P in the causal model (i.e., setting variable P to a specific value without considering its cause), and E(·) represents the expected value operator (calculating the average of random variables).

[0189] The specific implementation of this causal inference model is as follows:

[0190] Construct a structural causal model (SCM) to represent the causal relationships between variables:

[0191] SCM = V scm U scm F scm ,Pr(U scm );

[0192] Where SCM represents the structural causal model, V scm It is a set of observed variables (including policy variables, safety indicators, efficiency indicators, etc.), U scm It is the set of latent variables (representing unobserved confounding factors), F scm It is a set of functions (describing deterministic relationships between variables), Pr(U scm ) is the joint distribution of the latent variables.

[0193] Cause-effect graphs are used to represent causal relationships between variables. Nodes in the graph represent variables, and directed edges represent direct causal relationships.

[0194] For evaluating the effect of strategy P on security Y security The impact is calculated using do-calculus:

[0195]

[0196] Where Pr(Y) security |do(P)) represents the safety Y after intervention strategy P. security The probability of z is Z sec The value of Z sec Representing policy P and security Y security The set of moderating variables, such as user behavior characteristics and data sensitivity, are represented by Pr(·), which indicates probability.

[0197] Similarly, the computational strategy P affects the efficiency Y. efficiency Impact:

[0198]

[0199] Where Pr(Y) efficiency|do(P)) represents the efficiency Y after intervention strategy P. efficiency The probability of z′ is Z. eff The value of Z eff Representing strategy P and efficiency Y efficiency The set of adjustment variables between them, where Pr(·) represents probability.

[0200] Calculate the combined impact index of different strategies:

[0201]

[0202] in This represents the partial derivative of the overall effect with respect to the strategy (measuring the degree to which a small change in strategy affects the overall effect). and W represents the partial derivatives of the safety and efficiency impacts with respect to the policy, respectively. security and W efficiency These are weighting coefficients used to balance the importance of security and efficiency.

[0203] For example, in the data governance system of a smart manufacturing enterprise, this model was applied to evaluate the impact of a "time-based sensitive recipe data access strategy." The system simulated the impact of different strategy variables (such as access time restrictions and approval process levels) on security indicators (such as data leakage risk rate and unauthorized access frequency) and efficiency indicators (such as data access waiting time and R&D process completion cycle) in a virtual environment. Causal analysis revealed that restricting access to sensitive recipe data to weekday working hours, coupled with two-factor authentication, reduced the risk of data leakage by 78%, while only extending the process time by 8% in terms of R&D efficiency. This model helped the enterprise determine the optimal balance and optimize its data governance strategy. This evaluation model quantifies the risk-benefit ratio of different strategies, providing a basis for access control decisions.

[0204] Safety impact assessment results for this step: I impact (P) serves as the key output, flowing in two directions:

[0205] As an emergency feedback, it flows directly back to sub-step 3.3 in step 3, and is used to adjust the risk threshold θ and decision rules in real time to prevent immediate security risks;

[0206] As a long-term optimization basis, step 5 is used to evaluate the governance effect. gov Provide reference metrics and define the reward function R for the reinforcement learning model. t，rl Provides a rating component for security and efficiency.

[0207] This step also receives the responsibility matrix M from step 1. RACI Step 2 User Behavior Model B userUsing the risk assessment model R(u, d, c) from step 3 as input, we combine this information to construct a virtual environment and simulated test scenarios.

[0208] Step 5: Based on the simulation test results, adaptively adjust the model parameters and dynamically update the initial responsibility matrix to form a closed-loop management system;

[0209] This step establishes a feedback mechanism for the effectiveness of data governance strategies, continuously optimizes the responsibility allocation and permission decision-making model, and achieves closed-loop management of data governance. Specifically, it includes the following sub-steps:

[0210] Sub-step 5.1: Monitoring and evaluating the governance effect;

[0211] A multi-level data governance quality monitoring model is constructed to collect and analyze key data governance indicators. The monitoring indicator system comprises three levels: basic indicators (such as the number of authorized operations and rejection rate), process indicators (such as decision response time and user feedback satisfaction), and outcome indicators (such as the number of data breaches and compliance scores). Based on these indicators, a data governance effectiveness evaluation model is constructed.

[0212] E gov =f evaluate (I base I process I result W ind );

[0213] Where E gov Indicates the governance effectiveness score, I base I process and I result W represents the sets of basic indicators, process indicators, and outcome indicators, respectively. ind f represents the set of indicator weights; evaluate This represents the evaluation function, implemented as a weighted scoring model.

[0214] Sub-step 5.2: Adaptive adjustment of model parameters;

[0215] Based on the governance effectiveness evaluation results and user feedback, a reinforcement learning algorithm is used to adaptively adjust the parameters of the permission decision-making model. By observing the difference between actual user behavior and predicted behavior through reinforcement learning, the behavioral model parameters are dynamically adjusted to improve the model's predictive accuracy. Specifically, the optimization targets include the weight parameters of the risk scoring function and the risk threshold θ in step three.

[0216] The adjustment process is modeled as a Markov decision process:

[0217]

[0218] Where π *argmax represents the optimal policy function. π This represents the search for the optimal policy in the policy space, where T represents the time step and γ represents the time step. disc It is a discount factor. It is the reward obtained at time t (based on governance effectiveness score). The adjusted parameters include the weights in the risk assessment model, the risk threshold, and the update frequency of the responsibility matrix. E(·) represents the expected value operator.

[0219] The specific implementation of this reinforcement learning model is as follows:

[0220] State space definition, state vector s t Includes key data governance metrics, user behavior pattern distribution, current permission denial rate, security incident incidence rate, and other information.

[0221]

[0222] Where s t Represents the state vector, I base，t I process，t and I result，t U represents the values ​​of the basic indicator, process indicator, and result indicator at time t, respectively. pattern，t Represents the distribution of user behavior patterns, Reject r ate t This indicates the current permission denial rate, Security. e vent r ate t ... indicates the incidence rate of security incidents, ... indicates other indicators.

[0223] Action space definition, action vector a t Includes adjustable parameters, such as the weighting coefficients of each feature in the risk assessment model and risk thresholds:

[0224]

[0225] Where a t Represents the action vectors, w1, w2, They represent the 1st, 2nd, and nth respectively. weight There are several weighting parameters, where θ is the risk threshold, and f is the weighting parameter. update It is the update frequency of the responsibility matrix.

[0226] The reward function design takes into account both security and efficiency considerations.

[0227]

[0228] in Represents the reward at time t, Security score t This indicates the security score, Efficiency. s core t Indicates efficiency rating, (Change) r ate t ) represents the rate of change of the parameter (used to prevent excessive parameter fluctuations), and λ1, λ2 and λ3 are trade-off parameters.

[0229] Implement reinforcement learning algorithms using Deep Q-Networks (DQN):

[0230]

[0231] Where Q(s) t a t ) indicates that in state s t Take action a t The expected cumulative reward, E(·) denotes the expected value operator, γ disc It is a discount factor. Let s represent the rewards at time t, t+1, and t+2 respectively. t and a t Indicates state and action.

[0232] Using neural network approximation functions Function representing the Q value:

[0233]

[0234] Where Q(s) t a t ) is in state s t Take action a t The expected cumulative reward, The parameter is θ Q The neural network approximation function.

[0235] Balancing exploration and exploitation using an ε-greedy strategy:

[0236]

[0237] Where a t Represents the action vector, ∈ exp It's the exploration rate, argmax a Q(s t a) indicates that in state s t The action with the largest Q value is selected, and the random action means randomly selecting an action, which gradually decreases over time to achieve the transition from exploration to exploitation.

[0238] For example, in a financial institution's data governance system, this reinforcement learning model was applied to dynamically adjust risk decision parameters. The system set different initial risk weights and access thresholds for different types of data, such as credit data, customer information, and transaction records. As business operations progressed, the system observed a high false rejection rate (25%) for customer information data, while authorization for transaction record data was too lenient (8% of security incidents). Through the reinforcement learning model, the system automatically increased the sensitivity weight of transaction record data and lowered the risk scoring threshold for customer information access in specific scenarios. After three months of adaptive adjustment, the false rejection rate for customer information dropped to 7%, security incidents for transaction records dropped to 2%, and overall data governance efficiency improved by 31%. This model is particularly suitable for the ever-changing financial environment and can automatically optimize data governance strategies based on actual operational results.

[0239] Sub-step 5.3: Dynamically update the responsibility matrix;

[0240] The RACI responsibility matrix is ​​updated periodically based on organizational changes and governance effectiveness assessments; the updating process includes two parts:

[0241] Forced updates based on organizational structure changes ensure that the responsibility matrix remains consistent with the current organizational structure;

[0242] Based on the optimization and update of governance effectiveness evaluation, the responsibility allocation scheme is optimized by analyzing the relationship between the current responsibility allocation and governance effectiveness.

[0243] The responsibility matrix is ​​dynamically updated using the Update function to ensure that responsibility allocation remains consistent with organizational changes and governance effectiveness. The responsibility matrix update function is expressed as follows:

[0244]

[0245] in and Let G represent the responsibility matrices before and after the update, respectively. Let Update represent the update function. o E indicates a change in organizational structure. gov This indicates the score for the effectiveness of the governance.

[0246] This step is the optimization hub of the entire data governance methodology. It receives feedback information from all other steps and outputs optimized parameters and models to each step.

[0247] Sub-step 5.1 Receive the decision record from step 3 and the safety impact assessment results from step 4. impact (P), comprehensively generating a governance effectiveness score E gov ;

[0248] Sub-step 5.2 Based on E govThe behavioral model parameters and risk assessment model parameters in step 2 are optimized using reinforcement learning algorithms.

[0249] Sub-step 5.3 then updates the responsibility matrix. Feedback is sent to step 1 to form a closed-loop optimization of responsibility allocation.

[0250] This multi-directional feedback and parameter adjustment mechanism ensures that the entire data governance system can continuously optimize itself and adapt to the dynamic evolution of organizational changes and security needs.

[0251] Through the above five steps, this implementation method realizes an intelligent data governance method based on an industrial big data platform, which solves key technical problems in traditional data governance such as unclear responsibility allocation, static and rigid access control, and lack of prior security assessment, thereby improving the intelligence level and adaptability of data governance.

[0252] A data governance system based on an industrial big data platform is used to execute the data governance methods described above, including:

[0253] The intelligent organizational structure analysis module is used to process organizational structure documents and business process definitions through text semantic analysis to generate a responsibility allocation matrix.

[0254] The user behavior modeling module is used to analyze users' historical data access behavior and build a time-series model of user operation patterns.

[0255] The risk perception and decision-making module is used to make dynamic permission decisions based on data sensitivity scores, scenario characteristics, and multi-dimensional risk calculations.

[0256] The digital twin verification module is used to construct a virtual environment for data interaction and verify the security impact of authorization policies.

[0257] The adaptive optimization module is used to adjust model parameters and update the responsibility matrix based on the monitoring results of governance effectiveness.

[0258] The processor and memory are used to perform the calculations and data storage of the above modules.

[0259] Here, the present invention provides an implementation example:

[0260] This implementation method was applied to the industrial big data platform of a large petrochemical group, achieving intelligent management of various data types, including production control data, process formulations, and equipment management data. The application scenarios, implementation process, and effect verification will be detailed below.

[0261] This petrochemical group has 10 production bases, 3 R&D centers, and 5 sales branches. Its data governance faces the following challenges:

[0262] The organizational structure is complex: the group has about 1,200 employees, covering multiple departments such as production, R&D, sales, and operations and maintenance. The unclear division of responsibilities has led to a lack of accountability for a large amount of data governance.

[0263] The data types are diverse and their sensitivity varies: The platform contains 17 types of core business data, as shown in Table 1. Among them, highly sensitive data such as formula data and process parameters account for 33%, requiring stricter access control.

[0264] Access scenarios vary greatly: employees’ access behaviors differ significantly at different times, locations and devices, making it difficult for static permission models to cope with dynamically changing access needs and risks;

[0265] Data sharing and security conflicts: Production bases and R&D centers need to share data to improve efficiency, but at the same time they face the risk of data leakage. Under the existing mechanism, the authorization cycle is long and there are problems of excessive restrictions or blind opening.

[0266] The core data types and sensitivities of the petrochemical group's industrial big data platform are shown in Table 1:

[0267] Table 1: Core Data Types and Sensitivity of Petrochemical Group's Industrial Big Data Platform

[0268]

[0269] In this environment, traditional static permission management and role-based access control (RBAC) models struggle to balance data security and efficient sharing, necessitating a data governance solution that dynamically senses organizational changes, user behavior, and access risks.

[0270] First, we collected 158 unstructured documents, including organizational structure documents, job descriptions, and business process documents from the petrochemical group. Through text semantic analysis, we extracted 245 role entities, 187 responsibility entities, and 423 relationships, constructing an organizational structure knowledge graph.

[0271] Based on organizational chart analysis and business process definition, a RACI responsibility matrix is ​​generated to clarify the specific responsibilities of each role in data governance activities. The responsibility allocation for some key roles for different data types is shown in Table 2.

[0272] Table 2: Example of a RACI Responsibility Matrix (Partial)

[0273]

[0274] Where: R means responsible, A means approved, C means consulted, I means informed, and - means unrelated.

[0275] User access logs from the group over two years were collected, containing approximately 28 million records. After cleaning and preprocessing, a structured user behavior dataset was obtained. A deep learning model combining LSTM and attention mechanisms, as described in this implementation, was used to model the operational patterns of 1324 active users.

[0276] The key parameters of the model were set as follows: 2 LSTM layers, 128 hidden units, batch size of 64, learning rate of 0.001, and 50 training epochs. The model performance was validated through a sequence prediction task, achieving an F1 score of 87.3% on the test set.

[0277] Based on the constructed user behavior model, the improved K-means clustering algorithm (k=12) was used to group the user behavior vectors and identify 12 typical user behavior types.

[0278] First, the data assets in the group's industrial big data platform were classified and their sensitivity assessed. A sensitivity scoring model was constructed, and the sensitivity scoring results of the main data types were displayed, as shown in Table 3.

[0279] Table 3: Data Sensitivity Scoring Results

[0280]

[0281] Based on user behavior models, data sensitivity, and real-time extracted scene features, a risk assessment model is constructed using the gradient boosting decision tree model in this implementation. The number of decision trees is set to 500, the maximum tree depth is 6, and the learning rate is 0.05. The model is trained using 100,000 risk assessment samples labeled by experts.

[0282] A two-level risk threshold system was finally constructed, with different risk decision thresholds set according to data sensitivity, as shown in Table 4:

[0283] Table 4: Risk Threshold Settings

[0284]

[0285] A virtual data interaction environment was constructed, with key process formulations and R&D experimental data being modeled. Typical user access behavior over a year was simulated in the virtual environment, totaling approximately 5 million interaction events. Based on the simulation test results, the causal inference model presented in this implementation was used to analyze the security and efficiency impacts of different permission policies.

[0286] A data governance effectiveness monitoring system was established, comprising 12 sets of core indicators, covering basic indicators, process indicators, and outcome indicators. The parameters of the risk assessment model were adaptively adjusted using reinforcement learning algorithms. The actual deployment employed a deep Q-network with a state space dimension of 28 and an action space dimension of 16.

[0287] The adaptive optimization results after 6 months of system deployment are shown in Table 5:

[0288] Table 5: Comparison of Adaptive Optimization Effects

[0289]

[0290] After this implementation method was carried out in the petrochemical group, a comprehensive evaluation was conducted through two key technical performance indicators: 1) the effect of accurate division and dynamic adjustment of data governance responsibilities; 2) the effect of intelligent and dynamic adaptation of data access control.

[0291] Following the implementation of this methodology, a 12-month follow-up evaluation was conducted on the group's data governance responsibility allocation. The quality and responsiveness indicators of the responsibility matrix are presented in Table 6.

[0292] Table 6: Quality and responsiveness assessment of the responsibility matrix

[0293]

[0294] The responsibility coverage rate increased from an initial 78.4% to 99.7%, the responsibility allocation conflict rate decreased from 14.6% to 1.2%, the number of missing responsibility items decreased from 325 to only 5, the average update response time shortened from 24.5 days to 4.2 days, and the responsibility allocation satisfaction score increased from 62.3 to 91.4. These indicators collectively demonstrate the advantages of this method in clarifying responsibility allocation and dynamically responding to organizational changes.

[0295] When significant changes occur in the organizational structure (the group underwent departmental integration in the 8th month), the system can automatically adjust and stabilize the allocation of responsibilities within 2 weeks, while the traditional manual adjustment method takes 2 months and has a higher error rate.

[0296] A comparative experiment was conducted by setting up a control group to evaluate the performance differences between the proposed method and the traditional RBAC permission model on various key indicators. The control group used the traditional RBAC model, while the experimental group used the proposed method. The experiment was run for 12 months under the same data environment, and the results are shown in Table 7.

[0297] Table 7: Comparative Experiment Results of Access Management

[0298] Evaluation indicators Traditional RBAC model This method Relative increase (%) Average authorization decision time (seconds) 47.56 0.38 -99.2％ Data access rejection rate (%) 22.73 8.21 -63.9％ Abnormal authorization rate (%) 8.47 1.36 -83.9％ Excessive permissions issue (item) 685 45 -93.4％ Insufficient permissions (item) 524 67 -87.2％ Data breach incidents (times) 7 0 -100％ Authorization satisfaction score (points) 67.8 92.6 +36.6％

[0299] Compared to the traditional RBAC model, this method reduces the average authorization decision time from 47.56 seconds to 0.38 seconds, the data access denial rate from 22.73% to 8.21%, and the abnormal authorization rate from 8.47% to 1.36%. Excessive and insufficient permissions issues are reduced by 93.4% and 87.2%, respectively. Notably, the traditional RBAC model experienced 7 data breaches during implementation, while this method successfully avoided all breach risks, and the authorization satisfaction score increased from 67.8 to 92.6.

[0300] To further verify the dynamic adaptability of this method, various special scenario tests were designed, including emergency response, remote collaboration, and personnel relocation scenarios. Test results show that this method reduces the average response time from 38 minutes to 45 seconds in emergency scenarios, increases the authorization rate from 37% to 92% in cross-site collaboration scenarios, and shortens the permission adaptation time from 2-3 weeks to within 3 days in organizational restructuring scenarios, enhancing the adaptability of data governance in complex scenarios.

[0301] In summary, the application of this implementation method in the petrochemical group's industrial big data platform has fully verified its technical effectiveness in two aspects: precise division and dynamic adjustment of data governance responsibilities, and intelligent and dynamic adaptation of access control. It has solved the core technical problems faced by traditional data governance, effectively balanced the contradiction between data security and efficient use, and has broad application value and promotion significance.

[0302] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A data governance method based on an industrial big data platform, characterized in that, include: Organizational structure documents and business process definitions are processed using text semantic analysis technology to generate a responsibility assignment matrix; The unstructured text is processed to extract role entities, responsibility entities, and relationships; based on the entities and relationships extracted in the previous step, an organizational structure knowledge graph is constructed, and the RACI responsibility matrix is ​​generated by combining the graph analysis results and business process definitions. Analyze users' historical data access behavior, construct a time-series model of user operation patterns, and characterize user data usage features; Based on the constructed user behavior model, cluster analysis is performed on users with similar behavior patterns to identify typical user behavior types and patterns; Based on user behavior models and responsibility allocation matrices, combined with data sensitivity scoring, scenario feature extraction, and multi-dimensional risk calculation, dynamic permission decisions are achieved. Multidimensional risk calculation and decision-making: Based on user behavior models, data sensitivity and scenario characteristics, it calculates access risk scores and determines authorization based on the comparison results of scores and thresholds; The multidimensional risk calculation and decision-making are implemented using a gradient boosting decision tree model. Authorization is determined based on a comparison between the risk score and a preset threshold. The implementation of the gradient boosting decision tree model includes: Construct a combined feature vector, combining user behavior features, data sensitivity features, scenario features, and corresponding values ​​from the responsibility allocation matrix into a unified feature vector: Based on the calculated risk score and the preset risk threshold This enables dynamic permission decisions. Based on dynamic permission decisions, a virtual environment for data interaction is constructed, and the security impact of the authorization policy is verified through simulation testing. Based on simulation test results, a security impact assessment model is constructed to analyze the impact of different permission policies on data security and business efficiency. The model uses a causal inference-based approach to characterize the causal relationship between permission policies and security risks and business efficiency. Safety impact assessment results As key outputs, they flow in two directions: As an emergency feedback mechanism, it is used to adjust risk thresholds in real time. Decision-making rules to prevent immediate security risks; As a basis for long-term optimization and for evaluating governance effectiveness. Provide reference metrics and the reward function for reinforcement learning algorithms. Provides a rating component for security and efficiency; Based on the simulation test results, the parameters of the risk assessment model are adaptively adjusted, and the initial responsibility allocation matrix is ​​dynamically updated to form a closed-loop management. The steps for adaptively adjusting the parameters of the risk assessment model include: Construct a data governance effectiveness evaluation model that includes basic indicators, process indicators, and outcome indicators; The parameters of the risk assessment model are adaptively adjusted using a reinforcement learning algorithm, and the adjustment process is modeled as a Markov decision process. Based on the governance effectiveness evaluation results and user feedback, a reinforcement learning algorithm is used to adaptively adjust the parameters of permission decision-making. The difference between the user's actual behavior and the predicted behavior is observed through the reinforcement learning algorithm, and the behavior model parameters are dynamically adjusted. The specific implementation of this reinforcement learning algorithm is as follows: State space definition, state vector It includes key data governance metrics, user behavior pattern distribution, current permission denial rate, and security incident occurrence rate information; Action space definition, action vectors Includes adjustable parameters, such as the weighting coefficients and risk thresholds of each feature in the risk assessment model; The reward function design takes into account both security and efficiency considerations. ； in This represents the reward at time t. Indicates the safety score. Indicates efficiency score. Indicates the rate of change of the parameter. , and It is a trade-off parameter.

2. A data governance system based on an industrial big data platform, characterized in that: The data governance method based on an industrial big data platform as described in claim 1 includes: The intelligent organizational structure analysis module is used to process organizational structure documents and business process definitions through text semantic analysis to generate a responsibility allocation matrix. The user behavior modeling module is used to analyze users' historical data access behavior and build a time-series model of user operation patterns. The risk perception and decision-making module is used to make dynamic permission decisions based on data sensitivity scores, scenario characteristics, and multi-dimensional risk calculations. The digital twin verification module is used to construct a virtual environment for data interaction and verify the security impact of authorization policies. The adaptive optimization module is used to adjust model parameters and update the responsibility allocation matrix based on the monitoring results of governance effectiveness. The processor and memory are used to perform the calculations and data storage of the above modules.

Images