A method and system for cross-linguistic semantic alignment and bilingual knowledge fusion for power grid standards

Abstract

This invention relates to the fields of computer data processing and power standardization management technology, specifically to a method and system for cross-language semantic alignment and bilingual knowledge fusion for power grid standards. The method includes: S1 collecting bilingual standard text and extracting clause context anchors; S2 performing contextual expansion on the power bilingual terminology database based on the clause context anchors to generate contextualized term vectors; S3 performing cross-language clause alignment and generating a translation ambiguity heatmap based on Shannon entropy; S4 fusing bilingual knowledge graphs and performing semantic drift inversion on terms to obtain term context drift vectors, which are coupled with the translation ambiguity heatmap to obtain a difference location probability distribution; S5 anchoring the difference clauses and difference types and feeding them back to S2 to form a closed loop. This invention inverts translation ambiguity and context drift from noise to be eliminated into difference location signals, improving the difference location accuracy by approximately 20%-30% compared to existing methods.

Description

Technical Field

[0001] This invention relates to the fields of computer data processing and power standardization management technology, specifically to a method and system for cross-language semantic alignment and bilingual knowledge fusion for power grid standards. Background Technology

[0002] Chinese invention patent application CN116303918A discloses a knowledge graph question-answering reasoning method and system based on the field of power and integration. The method uses topic entity recognition to obtain subgraphs and answer candidate sets, transforms dependency syntactic trees into hierarchical structure graphs, obtains hierarchical weights through bidirectional graph neural networks, and outputs the final answer after multi-step reasoning and scoring. The proposed solution is only for Chinese monolingual power knowledge graph question-answering scenarios and does not involve cross-language clause alignment and the construction of a bilingual power terminology database. It cannot support the semantic difference comparison between international standards and enterprise standards.

[0004] Chinese invention patent application CN114564950A discloses a method for recognizing Chinese named entities in the power sector by combining word sequences. This method mines professional terms from power texts using an unsupervised approach to construct a power vocabulary set. It co-encodes the word sequences and character sequences in an improved Transformer and then feeds them into a BERT-BiLSTM-CRF model to recognize power entities. However, this scheme only completes the named entity recognition task in the monolingual power sector and does not involve cross-language alignment or terminological extension. It also cannot establish a correspondence between international standards and enterprise standards at the clause level.

[0005] Chinese invention patent application CN106484686A discloses a patent intelligent translation system and its translation method. The method uses a sentence semantic analysis module to retrieve Chinese patents with semantic similarity to the target sentence from a Chinese patent database, and then retrieves the corresponding English patents from an English patent database and intelligently selects the corresponding English sentences to assemble them into an English patent document. The proposed solution is a sentence-level retrieval-based cross-language translation for patent documents. It does not model the semantic shift of contextualized terms in the power industry and lacks a mechanism for handling the semantic shift of terms under different voltage levels, equipment categories, and protection scenarios.

[0006] A comprehensive analysis of the existing technologies reveals that current solutions generally treat translation ambiguities and terminological context drift in cross-language comparisons of power grid standards as noise to be eliminated and smooth or discard them. They fail to recognize that translation ambiguities and terminological context drift themselves encode the location signals of the core differences between international standards and enterprise standards. This cognitive bias leads to the systematic limitation of the accuracy of cross-language standard difference location in existing methods on the upper limit of the serial architecture of translation followed by alignment. The real standard differences, such as differences in parameter thresholds, differences in technical requirement levels, and differences in the boundaries of applicable scenarios, are smoothed out by translation noise at the most significant positions of the differences, while insignificant literal differences are highlighted as differences. As a result, the difference list is dominated by noise, which cannot support the precise and intelligent needs of power industry standardization management. Summary of the Invention

[0007] To address the problem that existing cross-language comparison methods for power grid standards treat translation ambiguity and terminology context drift as noise to be eliminated, thus limiting the accuracy of difference localization to the upper limit of the sequential architecture of translation followed by alignment, this invention provides a cross-language semantic alignment and bilingual knowledge fusion method and system for power grid standards. By inverting translation ambiguity heatmaps and terminology context drift vectors from noise to be eliminated into standard difference localization signals and coupling them into difference localization probability distributions, and using the difference anchoring results to update the bilingual knowledge graph to form a closed loop, this invention achieves high-precision localization of cross-language power grid standard differences and deep fusion of bilingual knowledge at the signal inversion principle level, while maintaining the physical meaning constraints of clause-level context anchor points.

[0008] The technical solution of this invention is: a cross-lingual semantic alignment and bilingual knowledge fusion method for power grid standards, comprising the following steps: S1. Collect bilingual standard texts of international and corporate standards, and parse each standard clause to obtain clause context anchors, which include voltage level labels, equipment category labels, and protection scenario labels. S2. Based on the clause context anchors, perform contextual expansion on the power bilingual terminology database, generating contextualized term vectors for each term under different clause context anchors. S3. Based on the contextualized term vectors, perform cross-language semantic alignment at the clause level to obtain aligned clause pairs, and simultaneously generate a translation ambiguity heatmap, which includes the ambiguity entropy of each term. S4. Merge the aligned clause pairs into a bilingual knowledge graph, perform cross-contextual semantic drift inversion on each term to obtain a term context drift vector, and couple the translation ambiguity heatmap with the term context drift vector to obtain a difference positioning probability distribution. S5. Based on the difference positioning probability distribution, anchor the difference clauses and their difference types between the international standard and the corporate standard, and feed the difference anchoring results back to S2 to update the contextualized term vectors and the bilingual knowledge graph to form a closed loop.

[0009] This invention also provides a cross-language semantic alignment and bilingual knowledge fusion system for power grid standards, comprising: a bilingual text acquisition and contextual anchor extraction module, configured to acquire bilingual standard texts of international standards and enterprise standards, and parse each standard clause to obtain clause contextual anchors; a power bilingual terminology database contextualization module, configured to perform contextualization expansion on the power bilingual terminology database based on the clause contextual anchors to generate contextualized terminology vectors; and a cross-language clause alignment and ambiguity heatmap generation module, configured to perform cross-language semantic alignment at the clause level based on the contextualized terminology vectors to obtain aligned clause pairs and simultaneously generate a translation ambiguity heatmap; The bilingual knowledge graph fusion and drift inversion module is configured to fuse the alignment clause pairs into a bilingual knowledge graph, perform semantic drift inversion for each term across contexts to obtain a term context drift vector, and couple the translation ambiguity heatmap with the term context drift vector to obtain a difference positioning probability distribution; the standard difference anchoring and closed-loop feedback module is configured to anchor the difference clauses and their difference types between the international standard and the enterprise standard based on the difference positioning probability distribution, and feed the difference anchoring results back to the power bilingual terminology database contextualization module to update the contextualized term vector and form a closed loop with the bilingual knowledge graph.

[0010] The beneficial effects of this invention are as follows: First, in step S3, this invention generates a translation ambiguity heatmap by quantifying the probability uncertainty of candidate translations based on Shannon entropy and uses it as a standard difference location signal. This has significant advantages over existing technologies that treat translation ambiguity as a noise to be eliminated. The mechanism is that the degree of divergence between international power grid standards and enterprise standards at the terminology definition boundary can be encoded by the probability distribution entropy value of candidate translations. After inverting this entropy value into a filter for difference candidate positions, it can directly locate the clauses where there is divergence at the semantic boundary of the terminology, which improves the accuracy of difference location by about 20%-30%.

[0011] Second, in step S4, this invention couples the term context drift vector and the translation ambiguity heatmap through tensor product to form a differential localization probability distribution. Compared with the existing single signal source differential localization mechanism, this constitutes a nonlinear synergistic effect of 1+1>2. The mechanism is that the translation ambiguity heatmap describes the uncertainty of terms at the static translation level, while the term context drift vector describes the dynamic semantic shift of terms in cross-clause contexts. The two quantify the standard difference from two complementary dimensions: static boundary and dynamic shift. The recall rate of the coupled differential localization probability distribution is about 15% higher than that of the single signal source.

[0012] Third, in step S2, this invention performs contextual expansion on the power bilingual terminology database through clause context anchors and generates contextualized term vectors using hierarchical clause contextual encoding. Compared with single-layer word sequence encoding, it has the advantage of dimensional leap. The mechanism is that the length of power standard clauses ranges from tens of words to thousands of words, which is highly heterogeneous. Single-layer encoding cannot take into account both the global and local aspects. Hierarchical encoding reconciles the two scales through a gating fusion mechanism, so that the F1 of contextualized term vectors in mixed long and short clause scenarios is improved by about 15%-22% compared with single-layer encoding.

[0013] Fourth, this invention performs graph propagation correction on the probability distribution of difference location through power physical constraint graphs, upgrading the data-driven black box model to a physical constraint gray box model. The mechanism is that voltage-insulation level constraints, current-current carrying capacity constraints, and frequency-protection setting constraints are inviolable physical constraint chains in the power field, which constitute strong priors for difference candidates. Difference candidates that violate the constraint chains are likely to be false positives in the model, while difference candidates that conform to the constraint chains are likely to be true standard differences. Graph propagation correction reduces false positives of physical violations by about 35%-45% and improves the difference location F1 by about 8%-12%.

[0014] Fifth, in step S5, the present invention feeds back the difference anchoring result to step S2 to update the contextualized term vector and the bilingual knowledge graph to form a closed loop. Compared with the open-loop knowledge graph, it has the advantage of continuous learning. The mechanism is that the difference anchoring result serves as a high-quality annotation signal to correct the contextualized term vector and the bilingual knowledge graph, so that the difference positioning accuracy of the method continues to improve after multiple iterations. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the steps of the cross-language semantic alignment and bilingual knowledge fusion method for power grid standards according to the present invention. Figure 2 This is a module architecture diagram of the cross-language semantic alignment and bilingual knowledge fusion system for power grid standards, as described in this invention. Detailed Implementation

[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment uses the IEC 61850 smart substation communication standard and the China Southern Power Grid enterprise standard Q / CSG 1207002 "Application Specification for Communication Network and System of Smart Substation" as a typical bilingual standard comparison scenario to elaborate on the method embodiments and system embodiments of the present invention. It should be understood that the embodiments are only used to illustrate the present invention and are not intended to limit the present invention.

[0017] like Figure 1As shown, the cross-language semantic alignment and bilingual knowledge fusion method for power grid standards provided in this embodiment includes steps S1 to S5. Steps S1 to S5 form a deeply coupled closed loop during execution, and finally, in step S5, the difference anchoring result is fed back to step S2 to constitute a continuous learning mechanism. The specific implementation methods of each step are described in detail below.

[0018] Step S1: Bilingual Standard Text Acquisition and Clause Context Anchor Point Extraction. Step S1 involves acquiring bilingual standard texts from official data sources of international standardization organizations and enterprise standard management systems using a standard text acquisition module. The bilingual standard texts consist of two categories: English international standards and Chinese enterprise standards. The English international standards include, but are not limited to, core international standards in the fields of smart grids and new energy grid integration, such as the IEC 61850 series, IEC 61400 series, IEC 61727 series, and IEEE 1547 series. The Chinese enterprise standards include, but are not limited to, the Q / CSG Southern Power Grid enterprise standard series, the Q / GDW State Grid enterprise standard series, and the DL / T power industry standard series.

[0019] The acquired bilingual standard text is parsed into standard clauses according to a three-level hierarchy of chapters, sections, and clauses. Each standard clause serves as the basic semantic unit for subsequent processing. This embodiment employs a rule-driven chapter parser to perform hierarchical parsing of the bilingual standard text. The rule-driven chapter parser identifies chapter numbers, chapter titles, and boundaries between clause texts based on standard format specifications, resulting in a structured set of standard clauses. The specific implementation of the chapter parser uses hierarchical pattern matching based on regular expressions. For IEC standards, it identifies English hierarchical markers such as Clause, Sub-clause, and Note; for Chinese enterprise standards, it identifies Chinese hierarchical markers such as chapters, sections, clauses, and sub-clauses, parsing the entire standard document into a hierarchically organized standard clause tree structure.

[0020] Each standard clause is analyzed to obtain a clause context anchor, which includes three members: voltage level label, equipment category label, and protection scenario label. The voltage level label is divided into five ranges according to power grid industry conventions: low voltage, medium voltage, high voltage, ultra-high voltage, and extra-high voltage. Low voltage corresponds to 1kV and below; medium voltage corresponds to 1kV and below; high voltage corresponds to 35kV and below; high voltage corresponds to 220kV and below; ultra-high voltage corresponds to 750kV and below; and extra-high voltage corresponds to above 750kV. The equipment category label is divided into seven categories according to the function of the power equipment: power generation equipment, transmission equipment, substation equipment, distribution equipment, protection equipment, communication equipment, and measurement equipment. The protection scenario label is divided into five scenarios according to the protection logic triggering scenario: normal operation scenario, fault protection scenario, overload protection scenario, insulation coordination scenario, and backup power switching scenario.

[0021] The extraction process of clause context anchors is achieved by combining clause keyword retrieval with domain named entity recognition: First, predefined voltage level keywords, equipment category keywords, and protection scenario keywords are retrieved from the text of the standard clauses. The tags corresponding to the matched keywords are used as preliminary candidates for the clause context anchors. Then, for clauses with multiple candidates or no candidates, power domain named entity recognition is performed. By loading a pre-trained named entity recognition model finely tuned for the power domain, voltage level entities, equipment entities, and scenario entities are identified from the clause text and mapped to the corresponding clause context anchor tags. Step S1 outputs a sequence of binary pairs consisting of the structured standard clause set and its corresponding clause context anchors, which serves as the input for step S2.

[0022] Step S2: Contextual Expansion and Contextualized Terminology Vector Generation of the Power Bilingual Terminology Database. Step S2 extends the power bilingual terminology database using a contextualization module based on the contextual anchors of the clauses, and generates contextualized terminology vectors through hierarchical clause contextual encoding. The power bilingual terminology database is a pre-built Chinese-English bilingual database covering core terms in the power field. In this embodiment, the power bilingual terminology database contains no fewer than 10,000 terms, covering all aspects of power business, including power generation, transmission, substation, distribution, protection, control, communication, and measurement. Each term includes the original Chinese term, the original English term, the term definition, the relevant business aspect, and the source of the reference standard.

[0023] The contextualization extension refers to generating differentiated contextualized term vectors for each term under different clause context anchors. For term t at clause context anchors... Contextualized term vectors The contextualized term vector is generated through hierarchical clause context encoding, which includes two components: outer clause-level encoding and inner term local window encoding. The outer clause-level encoding encodes the entire standard clause to obtain a clause-level contextual representation, and the inner term local window encoding encodes the local windows around each term to obtain a term local contextual representation. Then, the clause-level contextual representation and the term local contextual representation are fused into the contextualized term vector through a gating fusion mechanism.

[0024] The outer clause-level encoding performs encoding on the entire sequence of standard clauses using a Transformer encoder: in: The clause-level context representation matrix output by the outer clause-level encoding is a matrix with dimension 1. The range of values ​​is the real number space. Dimensionless, it is obtained by performing Transformer encoding on the standard clause sequence C using the outer clause-level encoding, representing the global contextual information of the standard clause; C is the input standard clause sequence, which is a vector sequence with dimension . The range of values ​​is the real number space. Dimensionless, obtained by projecting the corresponding clauses in the structured standard clause set output in step S1 through a word embedding layer, representing the word-level original representation of the standard clauses; L is the length of the standard clause sequence, a scalar whose value ranges from the set of positive integers. The value range is [8, 4096], dimensionless, and determined by the number of words obtained by segmenting the standard clause using a word segmenter, representing the number of words in the standard clause; d is the hidden layer dimension of the outer clause-level encoding, a scalar, and its value range is the set of positive integers. The value is 512 or 768, dimensionless, determined by the model hyperparameter configuration, and represents the feature dimension of the outer clause-level encoded output vector; Let be the output dimension of the word embedding layer, and let be a scalar whose value range is the set of positive integers. The value is 768, which is dimensionless and determined by the hyperparameter configuration of the word embedding layer, representing the feature dimension of the output vector of the word embedding layer. The outer clause-level Transformer encoding function consists of a multi-head self-attention sublayer, a feedforward network sublayer, and a layer normalization sublayer stacked together. In this embodiment, the number of stacked layers is 12, representing a non-linear mapping that performs global contextual encoding on the standard clause sequence. The left side of the formula... for The dimensionless matrix, the Transformer function on the right side of the formula, is... After performing a projection transformation on the 3D input, the output is... A dimensionless matrix has the same dimensions on both sides.

[0025] Inner term local window encoding is performed by a Transformer encoder on the sequence of local windows surrounding each term: ,in: The term local context representation matrix output by the inner term local window encoding is a matrix with dimension 1. The range of values ​​is the real number space. Dimensionless, encoded by the local window sequence of the inner terminology. The fine-grained contextual information representing the local context surrounding the term is obtained by performing Transformer encoding calculations. Let be the sequence of local windows surrounding the input term t, and be a vector sequence with dimension . The range of values ​​is the real number space. Dimensionless, it is obtained by projecting a window word sequence of w words to the left and right of the term t from the standard clause sequence C through a word embedding layer, representing the local context-level word-level original representation of the term t; w is the local window radius, a scalar, and its value range is the set of positive integers. The value range is [3, 16], dimensionless, determined by the model hyperparameter configuration, and in this embodiment, the value is 8, representing the number of context words taken to the left and right of the term t; d is defined as in the aforementioned formula; The Transformer encoding function for the inner terminology local window is used. In this embodiment, the stacking layer is set to 4 layers, representing a non-linear mapping that performs fine-grained context encoding on the local window sequence. The left side of the formula... for The dimensionless matrix, the Transformer function on the right side of the formula, is... After performing a projection transformation on the 3D input, the output is... A dimensionless matrix has the same dimensions on both sides.

[0026] The gating fusion mechanism fuses the clause-level contextual representation with the term local contextual representation into a final contextualized term vector. Let... The clause-level context representation matrix The row vector at the position of the term t in the Chinese text. The local context representation matrix of the term The row vector at the location of the term 't' (i.e., the center position) is then calculated using the following formula: , ,in: Let be the gating fusion weight vector, be a vector with dimension d, and take values ​​in the real number space. Dimensionless, determined by the gating fusion mechanism for the... With the The concatenated vector is obtained by linear transformation and sigmoid activation calculation, representing the fusion weight of the clause-level context representation and the term local context representation on each feature dimension; For the term 't', anchor the clause context. The contextualized terminology vector below is a vector with dimension d and values ​​in the real number space. Dimensionless, determined by the gating fusion mechanism for the... With the According to the description The weighted summation is performed to represent the anchor point of the term t in the context of the clause. Contextualized semantic representation; The clause-level context representation matrix The row vector at the position of the term t is a vector with dimension d and its values ​​are in the real number space. , dimensionless, is extracted by term position index after the outer clause-level encoding output, and represents the global clause context features of term t; The local context representation matrix of the term The row vector at the center of the term t is a vector with dimension d and its values ​​are in the real number space. , dimensionless, is extracted by indexing the center position after the local window encoding output of the inner term, and represents the local window contextual features of the term t; Let be the gating fusion weight matrix, and be a matrix with dimension . The range of values ​​is the real number space. , dimensionless, obtained by model training, represents the linear projection of the gated fusion mechanism onto the concatenated vector to the gate weights; Let be the gated fusion bias vector, which is a vector with dimension d and values ​​in the real number space. , dimensionless, obtained from model training, characterizing the linear projection bias of the gated fusion mechanism; The sigmoid activation function is defined as follows: Dimensionless, mapping real number inputs to The interval represents the nonlinear activation of the linear projection output; This is a vector concatenation operator that concatenates two d-dimensional vectors along their feature dimensions to obtain... dimensional vector; This is a vector element-wise multiplication operator that multiplies two d-dimensional vectors by their corresponding positions to obtain a d-dimensional vector. Let be a d-dimensional vector with all elements equal to 1, representing the subtraction of element-wise 1s. The operational benchmark; 't' is a clause context anchor point, a symbolic label, whose value range is a finite set of the Cartesian product of the voltage level label, equipment category label, and protection scenario label. It is dimensionless and is obtained by extracting the current clause execution clause context anchor point in step S1, representing the contextual constraint of the term 't' under the current clause. The left side of the gating fusion weight formula... A dimensionless d-dimensional vector; right-hand linear transformation We obtain a dimensionless d-dimensional vector, and then add a bias. The vector remains a dimensionless d-dimensional vector, and the sigmoid function preserves its dimensionless nature, ensuring consistency of dimensions on both sides. The left side of the contextualized terminology vector formula... The first vector is a dimensionless d-dimensional vector; the summation of the element-wise multiplications on the right side yields a dimensionless d-dimensional vector with consistent dimensions on both sides.

[0027] By traversing each term t and each clause context anchor point Step S2 outputs the contextualized expansion result of the power bilingual terminology database, where each term has a differentiated contextualized term vector under different clause contextual anchor points. The contextualized term vector serves as the input for step S3.

[0028] Step S3: Cross-language clause alignment and translation ambiguity heatmap generation. Step S3 uses the cross-language clause alignment and ambiguity heatmap generation module to perform cross-language semantic alignment at the clause level based on the contextualized term vectors, obtain aligned clause pairs, and simultaneously generate a translation ambiguity heatmap.

[0029] Cross-language clause alignment is achieved by combining clause-level two-stream similarity calculation with maximum weighted matching of bipartite graphs. This applies to each clause in the international standard clause set. With each clause in the company's standard terms and conditions set The clause representation vectors are calculated based on their respective contextualized terminology vector sets, and then the alignment similarity between the two clauses is calculated using cosine similarity. The clause representation vectors are obtained by performing weighted average pooling on the contextualized terminology vectors of all terms within the clause. The weights of the weighted average pooling are determined by the TF-IDF scores of the terms in the clause. The cross-language clause alignment uses a bipartite graph maximum weight matching algorithm to perform a globally optimal match between international standard clauses and corporate standard clauses. The matching similarity threshold is set between 0.75 and 0.92, and the matching output is a set of aligned clause pairs, each pair containing one international standard clause and one corporate standard clause.

[0030] The generation of the translation ambiguity heatmap includes the following steps: for each term t to be aligned in the alignment clause pair, a set of candidate translations is retrieved from the power bilingual terminology database. The candidate translation set contains all possible translation candidates for term t; the distribution probability of the candidate translation set in the context of the alignment clause is calculated, and the Shannon entropy of the distribution probability is used as the ambiguity entropy of term t. The specific calculation formula is as follows: , ,in: For the term t, the k-th candidate translation The probability distribution of is a scalar, and its value range is . Dimensionless, determined by the contextual similarity between the candidate translation and the alignment clause. Temperature coefficient The normalized exponential transformation after scaling is used to calculate the term t, which is then translated into... The probability of; Let be the entropy of the term t, and let be a scalar with a range of values. The unit is bits, which is obtained by performing Shannon entropy calculation on the probability distribution, and characterizes the uncertainty of the distribution of candidate translations of term t; For the term t and candidate translations The contextual similarity is a scalar, with a value range of 1. Dimensionless, consisting of the contextualized term vector of the term t. With the aforementioned candidate translations Contextualized term vector The cosine similarity is calculated to represent the anchor points of the two terms in the context of the clause. The degree of semantic closeness; Temperature coefficient, a scalar quantity, has a range of values ​​of 1000. Furthermore, the value range in this embodiment is [0.5, 2.0], which is dimensionless and determined by the model hyperparameter configuration, representing the sharpness adjustment of the contextual similarity; Let be the number of candidate translations for term t, and be a scalar whose value range is the set of positive integers. The value range is [2, 32], dimensionless, and determined by the number of candidate translations of term t retrieved from the power bilingual terminology database, representing the scale of candidate translations for term t; k is the index variable for candidate translations, with a value range of... j is the index variable for the inner summation, and its value range is... ; Let be a logarithmic function to base 2, such that the unit of the ambiguity entropy is bits. The left side of the probability distribution formula... Scalar is dimensionless and takes values ; Right-side molecule It is a scalar and has the same type as the denominator. After normalization, it is a scalar and takes the value... The dimensions on both sides are consistent. The left side of the ambiguous entropy formula... The unit is scalar (bit); the probability distribution on the right is... Dimensionless The units are scalar bits, and the summation is still in scalar bits, with consistent dimensions on both sides.

[0031] The ambiguity entropy is higher than a preset ambiguity threshold. The terms are labeled as difference-sensitive terms, and the positions of these difference-sensitive terms in the alignment clause pair are used as standard difference candidate positions. The preset ambiguity threshold... Dynamic adaptation through the anchor points of the aforementioned clause context: ,in: For the term 't', anchor the clause context. The preset ambiguity threshold is a scalar with a value range of [value range missing]. The unit is bit, which is dynamically calculated by this formula according to the anchor point of the clause context, and represents the lower bound of the ambiguity entropy of the term t as a difference-sensitive term in the clause context; Let be the basic ambiguity threshold, and be a scalar with a range of values. The unit is bit, which is determined by the model hyperparameter configuration and is 1.0 bit in this embodiment, representing the basic threshold when no clause context anchor correction is applied; Here, is the high-voltage level correction factor, and is a scalar with a value range of . In this embodiment, the value is 0.4 bits, and the unit is bits. It is determined by the model hyperparameter configuration and represents the magnitude of threshold adjustment when the voltage level label is high voltage level. The correction factor for protecting the equipment is a scalar quantity with a range of values ​​of [value range missing]. In this embodiment, the value is 0.3 bits, and the unit is bits. It is determined by the model hyperparameter configuration and represents the extent of threshold tightening when the device category label is a protected device. Here, is the correction coefficient for fault protection scenarios; is a scalar with a value range of . In this embodiment, the value is 0.3 bits, and the unit is bits. It is determined by the model hyperparameter configuration and represents the degree of threshold tightening when the protection scenario label is a fault protection scenario. Here, is the high-voltage level indication function, and is a scalar with a value range of . Dimensionless, when the anchor point of the stated clause context The value is 1 when the voltage level label is high voltage level, ultra-high voltage level or extra-high voltage level, otherwise the value is 0; The function is a protection device indication function, where is a scalar and its value range is . Dimensionless, when the anchor point of the stated clause context The value is 1 when the equipment category label is "protected equipment"; otherwise, the value is 0. Here is the fault protection scenario indication function, where is a scalar and its value range is . Dimensionless, when the anchor point of the stated clause context The value of the protection scenario label is 1 when it is a fault protection scenario, and 0 otherwise. The left side of the formula... The scalar unit is bit; the right side , , , All are scalars in units of bit. The three indicator functions are dimensionless. A scalar in units of bit multiplied by a dimensionless scalar is still in units of bit, and the summation is still in units of bit. The dimensions on both sides are consistent.

[0032] The dynamic adaptation of the thresholds allows for a more lenient threshold for high-voltage clauses to accommodate the high concentration of terminology in high-voltage equipment, while tightening thresholds for protection equipment clauses and fault protection scenario clauses to accommodate the low fault tolerance of protection scenarios. Combining the set of difference-sensitive terms and their ambiguity entropy values, step S3 outputs the set of aligned clause pairs and the translation ambiguity heatmap, which serves as the input for step S4.

[0033] Step S4: Bilingual Knowledge Graph Fusion and Terminology Context Drift Inversion. Step S4 uses the bilingual knowledge graph fusion and drift inversion module to fuse the aligned terms into a bilingual knowledge graph, performs semantic drift inversion for each term across contexts to obtain a term context drift vector, and couples the translation ambiguity heatmap with the term context drift vector to obtain the difference localization probability distribution.

[0034] The bilingual knowledge graph is constructed using triples to represent clause entities, terminology entities, and their relationships. Each aligned clause pair generates a set of bidirectional relational triples, where the subject is the international standard clause entity in the aligned clause pair, the object is the corresponding enterprise standard clause entity, and the predicate is the alignment-to relation. Terminology entities within each standard clause are connected to the clause entities through the inclusion-term relation, and each pair of Chinese and English terminology entities is interconnected through the translation-to relation. The bilingual knowledge graph is stored in a graph database according to a node-relation-node graph structure. Node types include clause nodes and terminology nodes, and relation types include alignment-to, inclusion-term, and translation-to.

[0035] The inversion of term context drift vectors is performed on the set of contextualized term vectors for each term t under multiple clause context anchors. This is done for term t under M different clause context anchors. The contextualized term vector set below First, calculate its contextual mean vector. With cross-context covariance matrix Then, principal component analysis is used to extract the main drift direction, and finally, the deviation in each context is projected onto the main direction to obtain the drift vector. The specific calculation formula is as follows: , , , , in: Let be the contextual mean vector of term t, and let be a vector of dimension d, taking values ​​in the real number space. , dimensionless, is obtained by performing an arithmetic mean of the contextualized term vectors of the term t under the contextual anchor points of M clauses, and represents the cross-contextual semantic center of the term t; Let be the cross-contextual covariance matrix of term t, and let be a matrix with dimension . The value range is a set of symmetric positive semi-definite matrices, dimensionless, and is composed of the contextualized term vector and the context mean vector of the term t in each context. The deviation vectors between are summed by performing an outer product and divided by . We obtain the covariant structure that characterizes the cross-contextual semantic deviation of the term t; Let be the unit vector of the principal direction of drift for the term t, and let be a vector of dimension d, with values ​​ranging from a d-dimensional unit sphere. Dimensionless, derived from the cross-contextual covariance matrix The eigenvector corresponding to the largest eigenvalue is obtained, which represents the direction of the most significant cross-contextual semantic drift of the term t. For the term 't', anchor the clause context. The terminology context shift vector is a vector with dimension d and values ​​in the real number space. Dimensionless, defined by the term t in the context of the clause. The deviation vector below Projected onto the main drift direction Then multiply by the main drift direction to obtain the term t, which represents the anchor point of the clause context. The drift component relative to its contextual mean; M is the number of different clause contextual anchors covered by the term t, which is a scalar and takes the value of a set of positive integers. The value range is [2, 64], dimensionless, and determined by the contextual expansion results of the power bilingual terminology database, representing the cross-contextual sample size of term t; m is the index variable of the clause context anchor point, with a value range of [missing value]. ; The definition is the same as the formula mentioned above; The definition is the same as the formula mentioned above; This represents the matrix transpose operation; Represents the L2 norm; This represents selecting parameters on the constraint set that maximize the objective function. The left side of the contextual mean vector formula is a dimensionless d-dimensional vector; the right side is a sum of d-dimensional vectors multiplied by... It remains a dimensionless d-dimensional vector, with consistent dimensions on both sides. The left side of the covariance matrix formula is... The matrix is ​​dimensionless; the outer product of the d-dimensional vector on the right and its transpose yields... The matrix is ​​dimensionless; sum and multiply by . Still The matrix is ​​dimensionless, and the dimensions on both sides are consistent. The left side of the drift principal direction formula is a dimensionless d-dimensional unit vector; the right side constraint optimization result is also a dimensionless d-dimensional unit vector, with the dimensions on both sides consistent. Similarly, the left side of the terminology context drift vector formula is a dimensionless d-dimensional vector; the right side scalar multiplied by the d-dimensional vector yields a dimensionless d-dimensional vector, with the dimensions on both sides consistent.

[0036] For the term semantic role recognition, the inner term local window encoding outputs an attention weight matrix. The attention weight matrix is ​​subjected to term semantic role recognition to obtain term semantic roles, which include subject terms, constraint terms, and parameter terms. The specific calculation formula is as follows: , , in: Let be the attention weight matrix for term t, and be a matrix with dimension . The range of values ​​is Furthermore, the sum of each row equals 1, is dimensionless, and is determined by the multi-head self-attention sub-layer in the local window encoding of the inner terminology for the query matrix. Bond matrix The value is calculated by performing a scaled dot product and a normalized exponential transformation, representing the attention intensity between the term t and each pair of positions within the local window; Let be the semantic role probability vector of term t, be a vector with dimension 3, and take values ​​ranging from 1 to 2. Furthermore, the sum of the components equals 1, which is dimensionless, as stated in the attention weight matrix. The flattened text is obtained by linear transformation and normalized exponential transformation, representing the probability of the term t corresponding to the three semantic roles of subject term, constraint term, and parameter term, respectively. The query matrix is ​​a matrix with dimensions of . The range of values ​​is the real number space. Dimensionless, the local window sequence is determined by the query projection matrix of the inner term local window encoding layer. The projection is performed to obtain; Let be the key matrix, and be a matrix with dimension . The range of values ​​is the real number space. Dimensionless, the local window sequence is determined by the key projection matrix of the inner term local window encoding layer. The projection is performed to obtain; Let be the key projection dimension, and be a scalar whose value range is the set of positive integers. Furthermore, in this embodiment, the value is 64, which is dimensionless and determined by the model hyperparameter configuration; Let be a semantic role classification weight matrix for terms, and let be a matrix with dimension . The range of values ​​is the real number space. , dimensionless, obtained from model training; Let be a semantic role classification bias vector for terms, denoted as a vector with dimension 3 and values ​​in the real number space. , dimensionless, obtained from model training; To normalize the exponential function, map the real vector to the probability simplex; For the matrix flattening operator, Matrix concatenated by row and flattened A dimensional vector. The left side of the formula for the attention weight matrix is... Dimensionless matrix; right side for Matrix is ​​dimensionless, divide by scalar The resulting matrix retains the same dimension, and softmax preserves both the dimension and dimensionless properties, ensuring consistency between the left and right sides. The left side of the terminology semantic role formula is a dimensionless 3-dimensional probability vector; the right side, after linear transformation, is also a dimensionless 3-dimensional vector. Softmax preserves both the dimension and dimensionless properties, ensuring consistency between the left and right sides.

[0037] The translation ambiguity heatmap is coupled with the terminology context drift vector to obtain the difference localization probability distribution. The coupling process is achieved through tensor product fusion and the application of terminology semantic role-differentiated coupling weights. The specific calculation formula is as follows: , , , in: Let t be the differential coupling weight, and let t be a scalar with a value range of t. Dimensionless, derived from the semantic role probability vector of the term. Coupled weight vectors corresponding to the three types of semantic roles The dot product calculation is performed to obtain the weighted weight representing the role of the term t in the differential localization; Let be the coupling difference vector of term t, and let be a vector of dimension d, taking values ​​in the real number space. Dimensionless, derived from the ambiguous entropy The terminology context drift vector With the differentiated coupling weights The three are calculated by tensor coupling according to this formula, which characterizes the coupling difference strength of the term t in both static translation ambiguity and dynamic context drift. Let t be the probability of differential localization for the term t, and let be a scalar with a range of values. , dimensionless, is calculated from the L2 norm of the coupled difference vector through a normalized exponential transformation, and represents the probability that the position corresponding to the term t is a standard difference point; The terminology coupling weight is the main term, and the value range is 1. In this embodiment, the value is 0.2, which is dimensionless and determined by the model hyperparameter configuration, representing the lower weight of the subject term in the differential localization. The coupling weights are used to constrain the terminology; is a scalar with a range of values. Dimensionless, determined by the constraint strength of the constraint term. The constraint strength is obtained through dynamic calculation. The larger the value, the greater the weight, representing the moderate dynamic weight of the constraint term in differential positioning; The parameter term is coupling weight, which is a scalar with a range of values. Furthermore, in this embodiment, the value is 0.7, which is dimensionless and determined by the model hyperparameter configuration, representing the higher weight of parameter terms in differential localization; The constraint strength is the constraint term, and is a scalar with a value range of . Dimensionless, determined by the scope of the binding term in the clause and the frequency of the binding term; The definition is the same as the formula mentioned above; The definition is the same as the formula mentioned above; The set of all terms in the current alignment clause pair; This is the L2 norm operator for vectors. The left side of the differential coupling weight formula is a dimensionless scalar; the right side, the dot product of the 3D probability vector and the 3D weight vector, is also a dimensionless scalar, and the dimensions of the left and right sides are consistent. The left side of the coupling difference vector formula is a dimensionless d-dimensional vector; the right side is a scalar... scalar With d-dimensional vectors Multiplying them yields a dimensionless d-dimensional vector, with both sides having the same dimensions. The left side of the differential localization probability formula represents scalar values. Dimensionless; the L2 norm on the right is a scalar dimensionless value, and the normalized exponent gives the scalar value. Dimensionless, with consistent dimensions on both sides.

[0038] To further suppress discrepancy candidates that do not conform to the physical constraints of the power sector and enhance discrepancy candidates that do conform to the physical constraints of the power sector, the discrepancy localization probability distribution is reverse-corrected through the power physical constraint map. The power physical constraint map... A modeling electrical quantity constraint chain is established, comprising three types of members: voltage-insulation level constraints, current-current carrying capacity constraints, and frequency-protection setting constraints. The power physics constraint graph serves as a priori to perform graph propagation correction on the differential location probability distribution. , , in: Let be the symmetric normalized Laplace matrix of the power physics constraint graph, and let be a matrix with dimension . The value range is a set of symmetric positive semi-definite matrices, dimensionless, and derived from the degree matrix of the power physics constraint diagram. Adjacency Matrix The frequency domain structure of the power physics constraint diagram, calculated according to this formula, represents the structure of the power physics constraint diagram. For the first The probability distribution vector for differential localization in the next iteration is a vector with dimension N and a value range of [value]. Furthermore, the summation of the components is approximately 1, dimensionless, and calculated by this formula according to the graph propagation iteration rule, representing the differential location probability distribution after the correction of the power physics constraint graph; Let be the initial differential location probability distribution vector, be a vector with dimension N, and take values ​​ranging from . Furthermore, the sum of the components equals 1, which is dimensionless, and is determined by the aforementioned difference location probability. The probability distribution of differential location before correction of the power physics constraint map is obtained by aggregating and normalizing all terms. To define the probability distribution vector of the difference location in the k-th iteration, define the same... ; Let be the graph propagation coefficient, and be a scalar with a range of values ​​of . In this embodiment, the value is 0.5, which is dimensionless and determined by the model hyperparameter configuration. It represents the correction strength of the prior of the power physics constraint map to the initial differential location probability distribution; I is an N-dimensional identity matrix. Let be the degree matrix of the power physics constraint graph, and be a diagonal matrix with dimension . The adjacency matrix of the power physics constraint graph The degree of each node in the power physics constraint graph is obtained by summing the rows and constructing a diagonal matrix. Let be the adjacency matrix of the power physics constraint graph, and let be a matrix with dimension . The range of values ​​is Dimensionless, calculated from the existence and constraint strength of the electrical quantity constraint chains for each node pair in the electrical physical constraint graph, characterizing the topological structure of the electrical quantity constraint chains; N is the number of nodes in the electrical physical constraint graph, a scalar whose value ranges from the set of positive integers. , dimensionless, determined by the number of centering terms in the alignment clause; k is the index variable for graph propagation iteration, taking values ​​from the set of non-negative integers. The iterative convergence condition is: ,in To achieve convergence tolerance, and in this embodiment the value is taken as follows: The left side of the formula for the symmetric normalized Laplace matrix is... The matrix is ​​dimensionless; right side , , All Matrix is ​​dimensionless, and matrix multiplication preserves the matrix's dimension and dimensionless property, with consistent dimensions on both sides. The left side of the graph propagation iteration formula is an N-dimensional dimensionless vector; the right side... and All are dimensionless N-dimensional vectors. for A matrix is ​​dimensionless. Multiplying a matrix by a vector results in an N-dimensional dimensionless vector. The weighted summation also results in an N-dimensional dimensionless vector, with the dimensions of the left and right sides being consistent.

[0039] After the graph propagation iteration converges, the final differential location probability distribution, corrected by the power physics constraint graph, is obtained. This differential location probability distribution serves as the input for step S5.

[0040] Step S5: Standard Difference Anchoring and Closed-Loop Feedback. Step S5 uses the standard difference anchoring and closed-loop feedback module to anchor the difference clauses and their types between the international standard and the enterprise standard based on the difference positioning probability distribution, and feeds back the difference anchoring results to step S2 to form a closed loop. The anchoring of difference clauses is performed by sorting the terms in the aligned clause pairs in descending order according to the difference positioning probability distribution, and the clauses containing terms with a difference positioning probability higher than the difference anchoring threshold are marked as difference clauses. The difference anchoring threshold is between 0.6 and 0.85, and in this embodiment, it is 0.75.

[0041] The discrimination of difference types is further refined based on the semantic role of the term and the ambiguity entropy. The difference types include three categories: parameter threshold difference, term definition difference, and scenario application difference. When the difference-sensitive term is a parameter term, it is determined to be a parameter threshold difference; when the difference-sensitive term is a subject term and the ambiguity entropy is greater than 2.0 bits, it is determined to be a term definition difference; when the difference-sensitive term is a constraint term and there is a discrepancy in the contextual anchor point of the aligned clause pair, it is determined to be a scenario application difference. The difference anchoring result is fed back to step S2 as a high-quality annotation signal. The feedback content includes a list of difference clauses, the difference type classification result, and the ambiguity entropy record of the difference-sensitive term. After receiving the feedback content, the contextualization module of the power bilingual terminology database performs an exponential moving average update on the contextualized term vectors of the differing terms under the corresponding clause contextual anchor point; after receiving the feedback content, the bilingual knowledge graph fusion and drift inversion module supplements the difference type attribute edges to the aligned clause pairs with differences in the bilingual knowledge graph, so that the bilingual knowledge graph continuously accumulates domain knowledge of cross-language comparison of power grid standards in multiple iterations. The closed-loop feedback mechanism in step S5 enables the method of the present invention to continuously improve the difference localization accuracy during multiple rounds of execution, rather than remaining at the initial training accuracy.

[0042] To further illustrate the practical application effect of the method in this embodiment, the typical example of the alignment clause pair is illustrated by Section 6-2 of the IEC 61850 Smart Substation Communication Standard for GOOSE Application Layer Services and Clause 7.3 of the China Southern Power Grid Enterprise Standard Q / CSG 1207002 for GOOSE Communication Application Specification. In step S3 of this embodiment, the ambiguity entropy of the transmission delay under the two clauses is identified as 2.4 bits, which is higher than the preset ambiguity threshold of 1.0 bits for the corresponding clause, and it is marked as a difference-sensitive term. In step S4, the term context drift vector inversion is performed on the transmission delay, and it is found that its context anchor point in the IEC 61850 clause is a general substation scenario, while its context anchor point in the Q / CSG 1207002 clause is a 220kV high-voltage substation fault protection scenario. After correction by the frequency-protection setting constraint of the power physics constraint diagram, the difference location probability is 0.86, which is higher than the difference anchoring threshold of 0.75, and the semantic role of the term is identified as a parametric term. In step S5, the difference is anchored as a parameter threshold difference. IEC 61850 specifies a GOOSE transmission delay of 3ms, while the corresponding clause of Q / CSG 1207002 specifies a GOOSE transmission delay of no more than 2ms. The difference type is verified to be a parameter threshold difference. The working example verifies the accurate location capability of the method of this embodiment for substantial standard differences.

[0043] This embodiment also provides a cross-language semantic alignment and bilingual knowledge fusion system for power grid standards, corresponding to the above method embodiments. The modular architecture of the system is as follows: Figure 2As shown, the system comprises five modules: a bilingual text acquisition and contextual anchor extraction module, a power bilingual terminology database contextualization module, a cross-language clause alignment and ambiguity heatmap generation module, a bilingual knowledge graph fusion and drift inversion module, and a standard difference anchoring and closed-loop feedback module. These five modules correspond to steps S1 to S5 of the method embodiment, and the modules are deeply coupled and collaboratively linked through well-defined data interfaces. The following provides a detailed description of each module of the system.

[0044] The bilingual text acquisition and contextual anchor extraction module is configured to acquire bilingual standard texts of international and enterprise standards, and parse each standard clause to obtain clause contextual anchors, which include voltage level labels, equipment category labels, and protection scenario labels. The module's inputs are the official data source from the International Organization for Standardization (ISO) and the data source from the enterprise standard management system. The output is the structured standard clause set and its corresponding clause contextual anchor binary sequence. This module consists of three sub-units: a standard text acquisition sub-unit, a standard clause hierarchical parsing sub-unit, and a clause contextual anchor extraction sub-unit. The standard text acquisition sub-unit is responsible for pulling bilingual standard texts and performing format standardization processing. The standard clause hierarchical parsing sub-unit uses a rule-driven chapter parser to parse the entire standard document into a hierarchically organized standard clause tree structure. The clause contextual anchor extraction sub-unit extracts the clause contextual anchors for each standard clause using a combination of keyword retrieval and power industry named entity recognition. The output binary sequence of this module is passed to the power bilingual terminology database contextualization module through an inter-module data interface.

[0045] The contextualization module for the power bilingual terminology database is configured to perform contextualization expansion on the database based on the clause contextual anchors, generating a contextualized term vector for each term under different clause contextual anchors. The module's input includes the structured standard clause set, the clause contextual anchor sequence, and the pre-built power bilingual terminology database. The output is a set of contextualized term vectors organized by the Cartesian product of term and clause contextual anchor. The core component of this module is a hierarchical clause contextual encoder, which includes two sub-encoders: an outer clause-level Transformer and an inner term local window Transformer, as well as a gating fusion mechanism unit. The outer clause-level Transformer stacks 12 layers of multi-head self-attention sublayers, feedforward network sublayers, and layer normalization sublayers to encode the entire standard clause, obtaining the clause-level contextual representation. The inner term local window Transformer stacks 4 layers of multi-head self-attention sublayers, feedforward network sublayers, and layer normalization sublayers to encode a local window with a radius of 8 around each term, obtaining the term local contextual representation. The gated fusion mechanism unit executes the gated fusion formula described in step S2 of the reference method embodiment to fuse the clause-level contextual representation and the term local contextual representation into the contextualized term vector. The output contextualized term vector set of this module is simultaneously passed to the cross-language clause alignment and ambiguity heatmap generation module and the bilingual knowledge graph fusion and drift inversion module through the inter-module data interface.

[0046] The cross-language clause alignment and ambiguity heatmap generation module is configured to perform cross-language semantic alignment at the clause level based on the contextualized term vectors to obtain aligned clause pairs, and simultaneously generate a translation ambiguity heatmap, which includes the ambiguity entropy of each term. The module's input is the set of contextualized term vectors, and its output is the set of aligned clause pairs and the translation ambiguity heatmap. This module consists of six sub-units: a clause representation pooling sub-unit, a bipartite graph maximum weight matching sub-unit, a candidate translation retrieval sub-unit, a distribution probability calculation sub-unit, a Shannon entropy calculation sub-unit, and a dynamic threshold adaptive sub-unit. The clause representation pooling subunit performs weighted average pooling on the contextualized term vectors of all terms within each clause to obtain a clause representation vector; the bipartite graph maximum weight matching subunit performs bipartite graph maximum weight matching based on the cosine similarity between clause representation vectors to obtain the aligned clause pair set; the candidate translation retrieval subunit retrieves the candidate translation set for each term to be aligned in the power bilingual terminology database; the distribution probability calculation subunit, the Shannon entropy calculation subunit, and the dynamic threshold adaptive subunit calculate the candidate translation distribution probability, term ambiguity entropy, and preset ambiguity threshold according to the formula in step S3 of the method embodiment, and mark difference-sensitive terms. The output of this module, the aligned clause pair set and the translation ambiguity heatmap, are passed to the bilingual knowledge graph fusion and drift inversion module and the standard difference anchoring and closed-loop feedback module.

[0047] The bilingual knowledge graph fusion and drift inversion module is configured to fuse the aligned clause pairs into a bilingual knowledge graph, perform semantic drift inversion for each term across contexts to obtain a term context drift vector, and couple the translation ambiguity heatmap with the term context drift vector to obtain a difference localization probability distribution. The module's inputs include the set of aligned clause pairs, the translation ambiguity heatmap, and the set of contextualized term vectors; the outputs are the bilingual knowledge graph and the difference localization probability distribution. This module consists of six sub-units: a bilingual knowledge graph construction sub-unit, a term context drift inversion sub-unit, a term semantic role recognition sub-unit, a coupling difference vector calculation sub-unit, a difference localization probability distribution generation sub-unit, and a power physics constraint graph correction sub-unit. The bilingual knowledge graph construction subunit uses triples to represent clause entities, terminology entities, and their relationships, and stores them in the graph database. The terminology context drift inversion subunit, terminology semantic role recognition subunit, coupling difference vector calculation subunit, difference location probability distribution generation subunit, and power physics constraint graph correction subunit sequentially perform terminology drift inversion, terminology semantic role recognition, coupling difference vector calculation, difference location probability distribution generation, and power physics constraint graph correction operations according to the formula described in step S4 of the method embodiment. The final difference location probability distribution output by this module is passed to the standard difference anchoring and closed-loop feedback module.

[0048] The standard difference anchoring and closed-loop feedback module is configured to anchor the difference clauses and their types between the international standard and the enterprise standard based on the difference positioning probability distribution, and feed the difference anchoring results back to the power bilingual terminology database contextualization module to update the contextualized term vectors and form a closed loop with the bilingual knowledge graph. The module's inputs are the final difference positioning probability distribution, the set of aligned clause pairs, the translation ambiguity heatmap, and the terminology semantic role recognition results; the output is the difference anchoring results and feedback update instructions. This module consists of four sub-units: a difference clause anchoring sub-unit, a difference type discrimination sub-unit, a visualization output sub-unit, and a closed-loop feedback scheduling sub-unit. The difference clause anchoring subunit marks clauses containing terms with difference positioning probabilities higher than the difference anchoring threshold as difference clauses; the difference type discrimination subunit determines the difference type based on the semantic role and ambiguity entropy of the difference-sensitive terms as parameter threshold difference, term definition difference, or scenario applicability difference, respectively; the visualization output subunit generates bilingual annotations of difference clauses, difference type classification statistics charts, and difference positioning probability heatmaps for user review; the closed-loop feedback scheduling subunit sends the difference anchoring results to the power bilingual terminology database contextualization module in the form of feedback update instructions to update the contextualized term vectors using exponential moving average, and sends them to the bilingual knowledge graph fusion and drift inversion module to supplement the difference type attribute edges of the bilingual knowledge graph. This closed-loop feedback mechanism continuously improves the difference positioning accuracy of the system during multiple execution rounds.

[0049] The above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for cross-linguistic semantic alignment and bilingual knowledge fusion oriented towards power grid standards, characterized in that: Includes the following steps: S1: Collect bilingual standard texts of international standards and enterprise standards, and analyze each standard clause to obtain the clause context anchor points. The clause context anchor points include voltage level labels, equipment category labels, and protection scenario labels. S2: Based on the aforementioned clause context anchors, perform contextual expansion on the power bilingual terminology database, and generate a contextualized term vector for each term under different clause context anchors; S3: Based on the contextualized term vector, perform cross-language semantic alignment at the clause level to obtain aligned clause pairs, and simultaneously generate a translation ambiguity heatmap, which includes the ambiguity entropy of each term; S4: The alignment terms are fused into a bilingual knowledge graph. Semantic drift inversion is performed across contexts for each term to obtain a term context drift vector. The translation ambiguity heatmap is coupled with the term context drift vector to obtain a difference location probability distribution. S5: Based on the difference positioning probability distribution, anchor the difference clauses and difference types between the international standard and the enterprise standard, and feed the difference anchoring results back to S2 to update the contextualized terminology vector and the bilingual knowledge graph to form a closed loop.

2. The cross-lingual semantic alignment and bilingual knowledge fusion method for power grid standards according to claim 1, characterized in that: The generation of the translation ambiguity heatmap in step S3 includes retrieving a set of candidate translations for each term to be aligned in the power bilingual terminology database, calculating the distribution probability of the candidate translations in the context of the alignment terms, and using the Shannon entropy of the distribution probability as the ambiguity entropy. Terms whose ambiguity entropy is higher than a preset ambiguity threshold are marked as difference-sensitive terms, and the positions of the difference-sensitive terms in the alignment clause pair are used as standard difference candidate positions.

3. The cross-lingual semantic alignment and bilingual knowledge fusion method for power grid standards according to claim 2, characterized in that: The preset ambiguity threshold is dynamically adaptive through the clause context anchor point: a first threshold is used for clauses with voltage level label as high voltage level, and a second threshold is used for clauses with voltage level label as low voltage level, wherein the first threshold is higher than the second threshold. Threshold tightening corrections are applied to clauses labeled as "protected equipment" and to clauses labeled as "fault protection" for protection scenarios.

4. The cross-lingual semantic alignment and bilingual knowledge fusion method for power grid standards according to claim 3, characterized in that: The contextualized term vector in step S2 is generated through hierarchical clause contextual encoding. The hierarchical clause contextual encoding includes outer clause-level encoding and inner term local window encoding. The outer clause-level encoding encodes the entire standard clause to obtain a clause-level contextual representation. The inner term local window encoding encodes the local windows around each term to obtain a term local contextual representation. Then, the clause-level contextual representation and the term local contextual representation are fused into the contextualized term vector through a gating fusion mechanism.

5. The cross-lingual semantic alignment and bilingual knowledge fusion method for power grid standards according to claim 4, characterized in that: The inner term local window encoding outputs an attention weight matrix, and term semantic role recognition is performed on the attention weight matrix to obtain term semantic roles. The term semantic roles include subject terms, constraint terms, and parameter terms. In step S4, the coupling of the differential localization probability distribution applies differential coupling weights to the term semantic roles, making the differential coupling weight of the parameter terms higher than that of the subject terms, and making the differential coupling weight of the constraint terms dynamically adjusted according to the constraint strength of the constraint terms.

6. The cross-lingual semantic alignment and bilingual knowledge fusion method for power grid standards according to claim 5, characterized in that: The differential location probability distribution is reverse-corrected through an electrical physical constraint graph, which models electrical quantity constraint chains, including voltage-insulation level constraints, current-current carrying capacity constraints, and frequency-protection setting constraints. The electrical physical constraint graph serves as a priori to perform graph propagation correction on the differential location probability distribution, thereby suppressing differential candidates that violate the electrical quantity constraint chains and enhancing differential candidates that conform to the electrical quantity constraint chains.

7. The cross-lingual semantic alignment and bilingual knowledge fusion method for power grid standards according to claim 1, characterized in that: The bilingual terminology database for power has no fewer than 10,000 terms, the contextualized term vector has a dimension of 512 or 768, and the semantic similarity threshold of the aligned terms is between 0.75 and 0.

92.

8. The cross-lingual semantic alignment and bilingual knowledge fusion method for power grid standards according to claim 1, characterized in that: The international standards include at least one of IEC standards and IEEE standards, and the enterprise standards include at least one of the Southern Power Grid enterprise standards Q / CSG series and the power industry standards DL / T series.

9. The cross-lingual semantic alignment and bilingual knowledge fusion method for power grid standards according to claim 1, characterized in that: It also includes performing visualization output on the difference clauses and the difference types, the visualization output including bilingual annotations of difference clauses, statistical charts of difference type classification, and heat maps of difference location probability.

10. A cross-lingual semantic alignment and bilingual knowledge fusion system for power grid standards, used to implement the cross-lingual semantic alignment and bilingual knowledge fusion method for power grid standards as described in any one of claims 1-9, characterized in that, include: The bilingual text acquisition and context anchor extraction module is configured to acquire bilingual standard texts of international standards and enterprise standards, and parse each standard clause to obtain clause context anchors, which include voltage level labels, equipment category labels and protection scenario labels. The contextualization module of the power bilingual terminology database is configured to perform contextualization expansion on the power bilingual terminology database based on the contextual anchor points of the clauses, and generate a contextualized terminology vector for each term under different clause contextual anchor points. The cross-language clause alignment and ambiguity heatmap generation module is configured to perform cross-language semantic alignment at the clause level based on the contextualized term vector to obtain aligned clause pairs, and simultaneously generate a translation ambiguity heatmap, wherein the translation ambiguity heatmap includes the ambiguity entropy of each term. The bilingual knowledge graph fusion and drift inversion module is configured to fuse the alignment terms into a bilingual knowledge graph, perform semantic drift inversion across context for each term to obtain a term context drift vector, and couple the translation ambiguity heat map with the term context drift vector to obtain a difference location probability distribution. The standard difference anchoring and closed-loop feedback module is configured to anchor the difference clauses and difference types between the international standard and the enterprise standard based on the difference positioning probability distribution, and feed the difference anchoring results back to the power bilingual terminology database contextualization module to update the contextualized term vector and form a closed loop with the bilingual knowledge graph.

Images