A power service data security classification method and system

Abstract

The application discloses a kind of electric power service data security classification grading method and system, belong to data security technical field, method includes: obtaining the original description information of the data to be classified in electric power service data and carries out optimization and semantic vector coding;Based on electric power industry knowledge graph, automatically generate electric power classification rule;Parallel computing semantic similarity, rule matching degree and time decay history frequency;The above three are weighted and fused using dynamic weight coefficient, and a comprehensive confidence score is obtained;According to the pre-set confidence interval to which the score belongs, the corresponding electric power decision diversion operation is executed to output the final classification grading result.The application improves the accuracy and dynamic adaptability of data classification by multi-model dynamic fusion and intelligent decision diversion, optimizes the system processing efficiency, and enhances the fusion and rapid deployment capability of domain knowledge.

Description

Technical Field

[0001] This invention belongs to the field of power business data security technology, and in particular relates to a method and system for classifying and grading power business data security. Background Technology

[0002] With the accelerated advancement of digital transformation, data security classification and grading has become a core element in protecting critical information infrastructure, such as the data security classification and grading of power business data. Data security classification and grading aims to accurately identify and categorize the sensitivity of power business data, thereby driving the precise implementation of security measures such as data anonymization, encryption, and access control, and improving data flow and power business operation efficiency while ensuring basic security.

[0003] Currently, the mainstream data security classification technology in power business data is mainly based on traditional rule engines. This technical architecture typically includes three parts: a rule base, an inference engine, and a fact base. The rule base stores predefined power classification rules expressed in "IF-THEN" format; the inference engine often uses efficient algorithms such as Rete to perform pattern matching between the data to be classified (i.e., facts) and the rule conditions; and the fact base stores the data state during system operation. Its typical workflow covers four stages: pattern matching, rule triggering, conflict resolution, and action execution. This type of classification method based on deterministic rules can achieve high processing efficiency and interpretability when processing structured data with clear characteristics in power business, and to some extent meets the basic needs of early data classification in power business.

[0004] However, existing technologies still have significant limitations. First, traditional rule engines heavily rely on pre-compiled, relatively static rule sets, making it difficult to effectively handle complex and ever-changing unstructured or semi-structured power business data, resulting in insufficient classification accuracy and a high misclassification rate. Second, although agent technologies, such as Large Language Models (LLMs), can improve classification accuracy at the semantic understanding level, their reasoning process is usually time-consuming, making it difficult to meet the real-time requirements of power business data in large-scale, high-concurrency scenarios. Furthermore, both traditional rule engines and simple LLM solutions face the problem of insufficient domain adaptability, meaning they struggle to quickly and cost-effectively integrate with the professional knowledge and compliance requirements of the power industry. Consequently, when facing power business scenarios, the system's adaptability, accuracy, and reliability cannot meet the protection needs of critical information infrastructure. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a method and system for classifying and grading power business data security, thereby resolving the issues present in the prior art.

[0006] To achieve the above objectives, in a first aspect, the present invention provides a method for classifying and grading the security of power business data, comprising: Obtain the original descriptive information of the data to be classified in the power business data; The original description information is input into the LLM business inference module for context optimization to obtain the optimized description information; The optimized description information is semantically encoded to obtain the corresponding semantic vector; Based on the knowledge graph of the power industry, power classification rules are automatically generated through an improved particle swarm optimization algorithm; The semantic vector and the pre-stored classification label vector are input into a semantic matching model that integrates an attention mechanism to calculate semantic similarity. The optimized description information is matched with the automatically generated power classification rules, and the rule matching degree is calculated. Based on the historical access records of the data to be classified, calculate the historical frequency of time decay; Based on the dynamic weight coefficient, the semantic similarity, the rule matching degree and the historical frequency are weighted and fused to calculate the comprehensive confidence score of the data to be classified. Based on the preset interval to which the comprehensive confidence score belongs, the corresponding power decision-making and diversion operation is performed to output the final classification and grading results.

[0007] Preferably, the dynamic weight coefficients include a first weight coefficient, a second weight coefficient, and a third weight coefficient; the method further includes a dynamic weight adjustment step: based on the error rate statistics of the classification results within a preset period, the first weight coefficient, the second weight coefficient, and the third weight coefficient are adaptively adjusted, and the adjustment range is limited in a single instance.

[0008] Preferably, the semantic vector encoding step is implemented using an encoding model optimized based on a contrastive learning framework. The contrastive learning framework is trained using a triplet loss function so that the semantic vector distance between synonymous expressions is less than a first threshold, while the semantic vector distance between heteronyms is greater than a second threshold.

[0009] Preferably, the power industry knowledge graph includes multiple predefined core business entity types and various semantic relationship types; the power classification rules are based on the search space constituted by the core business entity types and the semantic relationship types.

[0010] Preferably, the semantic matching model with the fusion attention mechanism includes a multi-layer Transformer encoder and multiple attention heads; the semantic similarity is obtained by aggregating the output vectors of each attention head.

[0011] Preferably, the model for calculating the historical frequency incorporates a time decay factor, with access records closer to the current time having a higher weight.

[0012] Preferably, the weighted fusion calculation formula is: Comprehensive confidence score = First weight coefficient × Semantic similarity + Second weight coefficient × Rule matching degree + Third weight coefficient × Historical frequency + Correction factor; wherein, the correction factor is determined based on historical classification deviation cases.

[0013] Preferably, the power decision-making and diversion operation includes: when the comprehensive confidence score belongs to the high confidence interval, directly outputting the current TOP-1 classification result; when the comprehensive confidence score belongs to the medium confidence interval, initiating a secondary voting mechanism based on multiple candidate results; and when the comprehensive confidence score belongs to the low confidence interval, triggering a manual review process.

[0014] Preferably, the secondary voting mechanism uses multiple candidate classification labels output by the semantic matching model and their corresponding comprehensive confidence scores as weights for weighted voting to determine the final classification label.

[0015] Secondly, the present invention also provides a power business data security classification and grading system for implementing the method described in the first aspect, the system comprising: The information optimization module is used to obtain the original descriptive information of the data to be classified in the power business data; the original descriptive information is input into the LLM business inference module for context optimization to obtain the optimized descriptive information; The semantic vector encoding module is used to perform semantic vector encoding on the optimized description information to obtain the corresponding semantic vector; The classification rule generation module is used to automatically generate power classification rules based on the power industry knowledge graph and through an improved particle swarm algorithm. The first calculation module is used to input the semantic vector and the pre-stored classification label vector into a semantic matching model that integrates an attention mechanism to calculate semantic similarity; The second calculation module is used to match the optimized description information with the automatically generated power classification rules and calculate the rule matching degree. The third calculation module is used to calculate the historical frequency of time decay based on the historical access records of the data to be classified. The fourth calculation module is used to perform weighted fusion of the semantic similarity, the rule matching degree and the historical frequency according to the dynamic weight coefficient, and calculate the comprehensive confidence score of the data to be classified. The result output module is used to perform the corresponding power decision diversion operation based on the preset interval to which the comprehensive confidence score belongs, so as to output the final classification and grading result.

[0016] Compared with the prior art, the present invention has the following advantages and technical effects: This invention overcomes the limitations of single models or static rules by employing a multi-model dynamic fusion decision architecture that calculates semantic similarity, rule matching degree, and historical frequency in parallel, and uses dynamic weight coefficients for weighted fusion. This technical feature enables the system to integrate information advantages from different dimensions and adaptively adjust decision priorities based on data characteristics, thereby achieving more accurate and robust classification when facing complex and ever-changing business data.

[0017] This invention relates to a power decision-making and diversion operation based on comprehensive confidence scores, which can adopt differentiated processing strategies for classification results with different confidence levels. This technical feature avoids performing in-depth, time-consuming uniform analysis on all data, initiating secondary verification or manual intervention only for low-confidence results. Therefore, while ensuring classification quality, it significantly reduces average processing time and improves system throughput.

[0018] This invention utilizes a knowledge graph of the power industry and automatically generates power classification rules, structurally integrating domain-specific entities, relationships, and business logic into the core of the classification. This invention reduces reliance on manually written rules, automates the construction and iteration of the rule base, significantly shortens the system adaptation cycle for new scenarios, and improves the versatility and portability of the technology. Attached Figure Description

[0019] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a power business data security classification and grading method according to an embodiment of the present invention; Figure 2 This is a flowchart of the model voting process according to an embodiment of the present invention; Figure 3 This is a flowchart illustrating the dynamic parameter tuning process according to an embodiment of the present invention. Figure 4 This is a schematic diagram of a power business data security classification and grading system according to an embodiment of the present invention. Detailed Implementation

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0022] Example 1 like Figure 1 As shown in the figure, this embodiment provides a method for classifying and grading the security of power business data, including: S1. Obtain the original descriptive information of the data to be classified in the power business data; Specifically, based on the types of business in the power industry, they can be roughly divided into the following categories: The power generation side includes: thermal power generation, hydropower generation, nuclear power generation, and new energy power generation, etc. The power grid side includes: power transmission, substation, distribution, and integrated operation (such as provincial / municipal power supply companies); Electricity consumption and sales include: grid electricity sales, independent electricity sales, load aggregators, and major users; Market transactions include: power trading centers and green certificate trading; Technical and service support includes: planning and design, research institutes, equipment suppliers, and third-party technical service providers.

[0023] This embodiment takes the business data on the power grid side as an example. The business data is the data generated during the operation of the power grid. The data table and data fields of the power business data are shown in Table 1.

[0024] Table 1

[0025] S2. Input the original description information into the LLM business inference module for context optimization to obtain the optimized description information; Specifically, LLM stands for Large Language Model. It refers to a foundational artificial intelligence model based on deep learning technology, which is trained on a large amount of text data to gradually acquire the ability to understand and generate natural language.

[0026] In the LLM business inference module, if the original description is d and the optimized description is d′, then the optimization process can be represented as a function: d′= (d, c); Here, 'c' represents context information, which includes entity name, list of field names, system name, etc.

[0027] S3. Perform semantic vector encoding on the optimized description information to obtain the corresponding semantic vector; Furthermore, the semantic vector encoding step is implemented using an encoding model optimized based on a contrastive learning framework. The contrastive learning framework is trained using a triplet loss function so that the semantic vector distance of synonymous expressions is less than a first threshold, while the semantic vector distance of heteronymous expressions is greater than a second threshold.

[0028] Specifically, semantic vector encoding is implemented using an improved BERT embedding algorithm, and the specific steps are as follows: ① Text preprocessing: The input text t (such as data field names and business rule descriptions) is segmented and entity-labeled (based on entity types in the power industry knowledge graph). ② Pre-trained model encoding: Initial word vectors are generated using a domain-fine-tuned BERT model, with dimension d set to 768; ③ Vector optimization: Leveraging the LLM business inference module function d′= (d, c), incorporating the context information c, the formula is: vec(t) = BERT(t) + Atten; The Attention layer is used to capture the semantic association weights between the context and the text.

[0029] S4. Based on the knowledge graph of the power industry, power classification rules are automatically generated through an improved particle swarm algorithm; Furthermore, the power industry knowledge graph includes multiple predefined core business entity types and various semantic relationship types; the automatically generated power classification rules are based on the search space constituted by the core business entity types and the semantic relationship types.

[0030] Specifically, a knowledge graph for the power industry, including entity types and relationship definitions, is constructed. Entity types include core business entities such as "user information," "transaction records," and "account data"; relationship definitions include 12 semantic relationship types such as "contains," "associates," and "derives." An improved particle swarm optimization algorithm is used to automatically generate rules, with the algorithm's objective function being: fitness(ω) = α·Accuracy(ω)+β·Coverage(ω)-γ·Complexity(ω); Where ω is the particle swarm position vector, with a dimension equal to the number of rule features (default 32); α, β, and γ are weight coefficients, which are 0.6, 0.3, and 0.1, respectively; Accuracy represents the rule accuracy; Coverage represents the rule coverage; and Complexity represents the rule complexity (calculated based on the number of rule conditions).

[0031] S5. Input the semantic vector and the pre-stored classification label vector into the semantic matching model with the attention mechanism to calculate the semantic similarity. Furthermore, the semantic matching model that integrates the attention mechanism includes a multi-layer Transformer encoder and multiple attention heads; the semantic similarity is obtained by aggregating the output vectors of each attention head.

[0032] Specifically, an attention mechanism is introduced on top of semantic vector encoding. The model structure includes 12 attention heads and a 6-layer Transformer encoder, achieving accurate capture of key semantics through dynamic weight allocation. The semantic matching calculation formula is optimized as follows: ; in, Let be the output vector of the i-th attention head, representing the output vectors of text a and text b at the i-th attention head, respectively, generated by a 6-layer Transformer encoder.

[0033] S6. Match the optimized description information with the automatically generated power classification rules and calculate the rule matching degree; S7. Calculate the historical frequency of time decay based on the historical access records of the data to be classified; S8. Based on the dynamic weight coefficient, the semantic similarity, the rule matching degree, and the historical frequency are weighted and fused to calculate the comprehensive confidence score of the data to be classified. Furthermore, the dynamic weight coefficients include a first weight coefficient, a second weight coefficient, and a third weight coefficient; the method also includes a dynamic weight adjustment step: based on the error rate statistics of the classification results within a preset period, the first weight coefficient, the second weight coefficient, and the third weight coefficient are adaptively adjusted, and the adjustment range is limited in a single instance.

[0034] Furthermore, the model for calculating the historical frequency introduces a time decay factor, with access records closer to the current time having a higher weight.

[0035] Furthermore, the weighted fusion calculation formula is: Comprehensive confidence score = First weight coefficient × Semantic similarity + Second weight coefficient × Rule matching degree + Third weight coefficient × Historical frequency + Correction factor; wherein, the correction factor is determined based on historical classification deviation cases.

[0036] Specifically, an original confidence algorithm is introduced to achieve multi-dimensional result fusion, and a comprehensive score formula is defined: S = ; Where S is the overall score, and α, β, and γ are weighting coefficients. For semantic similarity, For rule matching degree, For historical frequencies, This is a correction factor.

[0037] Sim (Semantic Similarity): Based on an improved cosine similarity algorithm, it is calculated by fusing word vector cosine distance (weight 0.4), semantic role matching degree (weight 0.3), and contextual relevance degree (weight 0.3), with a value range of [0, 1]. This calculation method breaks through the limitation of traditional vector similarity, which only considers literal matching. By introducing semantic role labeling technology, it identifies deep semantic relationships between entities, improving the accuracy of similarity calculation for variant expressions such as "ID number" and "citizen identity code" by 35%.

[0038] Rule (Rule Matching Degree): A layered rule matching mechanism is adopted. The basic rule layer (such as keyword matching) has a weight of 0.5, the semantic rule layer (such as entity relationship verification) has a weight of 0.3, and the business rule layer (such as industry compliance requirements) has a weight of 0.2. The overall matching degree is obtained by weighted summation, with a value range of [0, 1]. Its innovation lies in realizing dynamic sorting of rule priorities and automatically adjusting the weight of each level according to the matching success rate, solving the technical problem that traditional rule engines cannot handle rule conflicts.

[0039] Freq (historical frequency): A frequency calculation model based on the time decay factor, the formula is: ; in, Let i be the frequency of the i-th access. The number of days since the current time is given, and λ is the attenuation coefficient (default 0.1). This method solves the problem of interference from "long-tailed high-frequency" data in traditional frequency statistics, giving higher weight to recent high-frequency access data and dynamically reducing the influence of historical low-frequency data.

[0040] α, β, γ (Dynamic Weight Coefficients): Employing a reinforcement learning-based dynamic adjustment mechanism, the initial configuration is α = 0.5, β = 0.3, γ = 0.2 (α + β + γ = 1). Every 24 hours, the system optimizes the weights based on the previous day's classification accuracy. When the classification error rate for a certain class exceeds a threshold (default 15%), the corresponding parameter weight is automatically increased: a high Sim error rate increases α, a high Rule error rate increases β, and a high Freq error rate increases γ. The weight adjustment range is limited to ±0.1 to ensure system stability. Figure 3 As shown.

[0041] In this embodiment, the reinforcement learning parameter tuning is based on the previous day's classification results. The error rate is calculated using the error rate statistics module: if Sim errors > 15% → α + 0.1; if Rule errors > 15% → β + 0.1; if Freq errors > 15% → γ + 0.1; the adjustment range is ±0.1 and α + β + γ = 1; the correction factor is... The value range is (-0.05, 0.05), which is used to correct the classification bias of low-frequency special fields. By analyzing historical classification error cases, a bias correction matrix is ​​constructed, and compensation values ​​are automatically applied to special format data such as "encrypted mobile phone number".

[0042] In this embodiment, semantic vector encoding and the automatic rule generation mechanism based on the power industry knowledge graph form a fundamental supporting relationship: entities / relationships in the knowledge graph need to be converted into vector form through semantic vector encoding, serving as input features for improving the particle swarm optimization algorithm; simultaneously, the generated rules need to calculate similarity through semantic vectors to avoid synonym conflicts and trigger dynamic adjustments to the rule base when the knowledge graph is updated. This forms a progressively reinforcing relationship with the semantic matching model that integrates an attention mechanism: semantic vector encoding provides the initial vector, and the attention mechanism optimizes the weight allocation through 12 attention heads, ensuring that the output vector of the i-th attention head... and By focusing more on core features, the two work together to improve the accuracy of semantic similarity calculation to 96%. The relationship with the original confidence algorithm is as follows: semantic vector encoding is the core calculation basis of "semantic similarity (Sim)" in the confidence algorithm. The "word vector cosine distance" (weight 0.4) in the Sim value comes from the vector similarity of semantic vector encoding, so that Sim can be weighted and fused with features such as rule matching degree (Rule) and historical frequency (Freq) in the formula S=α·Sim+β·Rule+γ·Freq +ε.

[0043] S9. Based on the preset interval to which the comprehensive confidence score belongs, perform the corresponding power decision diversion operation to output the final classification and grading result.

[0044] Furthermore, the power decision-making and diversion operation includes: when the comprehensive confidence score belongs to the high confidence interval, directly outputting the current TOP-1 classification result; when the comprehensive confidence score belongs to the medium confidence interval, initiating a secondary voting mechanism based on multiple candidate results; and when the comprehensive confidence score belongs to the low confidence interval, triggering a manual review process.

[0045] Furthermore, the secondary voting mechanism uses multiple candidate classification labels output by the semantic matching model and their corresponding comprehensive confidence scores as weights for weighted voting to determine the final classification label.

[0046] Specifically, such as Figure 2As shown, the final classification conclusion is generated by fusing results from multiple sources. The generated top-2 candidate names / field names are combined with the top-3 classification results to perform weighted voting. The classification result S value corresponding to each candidate name is used as the weight, and the candidate with the highest weight and its classification label are determined as the final output, for example, "ID number → Personal sensitive information (S = 0.89)".

[0047] This embodiment also includes a degradation processing step: when the LLM business inference module is unavailable, it switches to a similarity retrieval mode based on the semantic vector and the pre-stored vector library, and uses the classification result corresponding to the highest similarity as the output.

[0048] Specifically, when a large model is unavailable, the system automatically reverts to a similarity-based selection: ) .

[0049] Regarding the confidence guarantee mechanism, let the confidence sequence of candidate results be C = { , ,..., }, then the reliability index of the result is: When Reliability < At that time, a manual review process is triggered.

[0050] The beneficial effects of this embodiment: This embodiment proposes a data security classification method and system based on multi-model dynamic fusion and rule engine for power business data, which realizes automated and accurate classification of power industry business data and effectively improves classification efficiency and accuracy.

[0051] This embodiment adopts a dynamic rule weight adjustment mechanism, which automatically fine-tunes α / β / γ (±0.1) every 24 hours based on the accuracy of the previous day, thereby improving the rule matching success rate and solving the problem of poor adaptability of static rules.

[0052] This embodiment uses "triple loss + differentiable hard threshold" semantic vector encoding to force the vector distance of synonym variants to ≤0.3 and the distance of heteronyms to ≥0.7, thereby improving the Top-1 accuracy of classification and overcoming the insufficient classification accuracy.

[0053] This embodiment employs a particle swarm optimization-knowledge graph joint rule factory, which automatically generates rules using the entity-relationship structure of the graph as the search space. This achieves zero manual rule writing, shortens the industry cold start time, and fills in the gaps in industry knowledge.

[0054] This embodiment employs a three-level distribution based on Reliability adjacency levels. In the high-confidence region, classification and grading are directly output, while in the low-confidence region, similarity retrieval is performed first, followed by manual review. This significantly reduces the daily manual review workload while maintaining a reduction in grading errors.

[0055] Ranked by confidence score (S= This implements a three-level decision-making process: results are directly output in the high confidence region (S≥0.9), and in the low confidence region (S... 0.6) Triggers manual review, medium confidence interval (0.6≤S) 0.9) Initiate a second round of voting. This mechanism reduces the daily manual review volume from 20,000 to 11,600, lowering the cost of manual intervention by 42%; in the power industry classification scenario, the error rate drops from 20% to below 5%.

[0056] Example 2 like Figure 4 As shown, this embodiment provides a power business data security classification and grading system for implementing the method described in Embodiment 1. The system includes: The information optimization module is used to obtain the original descriptive information of the data to be classified in the power business data; the original descriptive information is input into the LLM business inference module for context optimization to obtain the optimized descriptive information; The semantic vector encoding module is used to perform semantic vector encoding on the optimized description information to obtain the corresponding semantic vector; The classification rule generation module is used to automatically generate power classification rules based on the power industry knowledge graph and through an improved particle swarm algorithm. The first calculation module is used to input the semantic vector and the pre-stored classification label vector into a semantic matching model that integrates an attention mechanism to calculate semantic similarity; The second calculation module is used to match the optimized description information with the automatically generated power classification rules and calculate the rule matching degree. The third calculation module is used to calculate the historical frequency of time decay based on the historical access records of the data to be classified. The fourth calculation module is used to perform weighted fusion of the semantic similarity, the rule matching degree and the historical frequency according to the dynamic weight coefficient, and calculate the comprehensive confidence score of the data to be classified. The result output module is used to perform the corresponding power decision diversion operation based on the preset interval to which the comprehensive confidence score belongs, so as to output the final classification and grading result.

[0057] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for classifying and grading the security of power business data, characterized in that, Includes the following steps: Obtain the original descriptive information of the data to be classified in the power business data; The original description information is input into the LLM business inference module for context optimization to obtain the optimized description information; The optimized description information is semantically encoded to obtain the corresponding semantic vector; Based on the knowledge graph of the power industry, power classification rules are automatically generated through an improved particle swarm optimization algorithm; The semantic vector and the pre-stored classification label vector are input into a semantic matching model that integrates an attention mechanism to calculate semantic similarity. The optimized description information is matched with the automatically generated power classification rules, and the rule matching degree is calculated. Based on the historical access records of the data to be classified, calculate the historical frequency of time decay; Based on the dynamic weight coefficient, the semantic similarity, the rule matching degree and the historical frequency are weighted and fused to calculate the comprehensive confidence score of the data to be classified. Based on the preset interval to which the comprehensive confidence score belongs, the corresponding power decision-making and diversion operation is performed to output the final classification and grading results.

2. The data security classification and grading method according to claim 1, characterized in that, The dynamic weight coefficients include a first weight coefficient, a second weight coefficient, and a third weight coefficient; the method further includes a dynamic weight adjustment step: based on the error rate statistics of the classification results within a preset period, the first weight coefficient, the second weight coefficient, and the third weight coefficient are adaptively adjusted, and the adjustment range is limited in a single instance.

3. The data security classification and grading method according to claim 1, characterized in that, The semantic vector encoding step is implemented using an encoding model optimized based on a contrastive learning framework. The contrastive learning framework is trained using a triplet loss function so that the semantic vector distance between synonymous expressions is less than a first threshold, while the semantic vector distance between heteronymous expressions is greater than a second threshold.

4. The data security classification and grading method according to claim 1, characterized in that, The power industry knowledge graph contains multiple predefined core business entity types and various semantic relationship types; the power classification rules are based on the search space formed by the core business entity types and the semantic relationship types.

5. The data security classification and grading method according to claim 1, characterized in that, The semantic matching model with the fusion attention mechanism includes a multi-layer Transformer encoder and multiple attention heads; the semantic similarity is obtained by aggregating the output vectors of each attention head.

6. The data security classification and grading method according to claim 1, characterized in that, The model for calculating the historical frequency introduces a time decay factor, with access records closer to the current time having a higher weight.

7. The data security classification and grading method according to claim 1, characterized in that, The weighted fusion calculation formula is: Comprehensive confidence score = First weight coefficient × Semantic similarity + Second weight coefficient × Rule matching degree + Third weight coefficient × Historical frequency + Correction factor; wherein, the correction factor is determined based on historical classification deviation cases.

8. The data security classification and grading method according to claim 1, characterized in that, The power decision-making and diversion operation includes: when the comprehensive confidence score belongs to the high confidence interval, directly outputting the current TOP-1 classification result; when the comprehensive confidence score belongs to the medium confidence interval, initiating a secondary voting mechanism based on multiple candidate results; and when the comprehensive confidence score belongs to the low confidence interval, triggering a manual review process.

9. The data security classification and grading method according to claim 8, characterized in that, The secondary voting mechanism uses multiple candidate classification labels output by the semantic matching model and their corresponding comprehensive confidence scores as weights to perform weighted voting, so as to determine the final classification label.

10. A power business data security classification and grading system, characterized in that, The system for implementing the method according to any one of claims 1-9 comprises: The information optimization module is used to obtain the original descriptive information of the data to be classified in the power business data; the original descriptive information is input into the LLM business inference module for context optimization to obtain the optimized descriptive information; The semantic vector encoding module is used to perform semantic vector encoding on the optimized description information to obtain the corresponding semantic vector; The classification rule generation module is used to automatically generate power classification rules based on the power industry knowledge graph and through an improved particle swarm algorithm. The first calculation module is used to input the semantic vector and the pre-stored classification label vector into a semantic matching model that integrates an attention mechanism to calculate semantic similarity; The second calculation module is used to match the optimized description information with the automatically generated power classification rules and calculate the rule matching degree. The third calculation module is used to calculate the historical frequency of time decay based on the historical access records of the data to be classified. The fourth calculation module is used to perform weighted fusion of the semantic similarity, the rule matching degree and the historical frequency according to the dynamic weight coefficient, and calculate the comprehensive confidence score of the data to be classified. The result output module is used to perform the corresponding power decision diversion operation based on the preset interval to which the comprehensive confidence score belongs, so as to output the final classification and grading result.

Images