Method for constructing manufacturing industry skill portrait based on knowledge graph

Abstract

The application discloses a manufacturing skill portrait construction method based on a knowledge graph and belongs to the technical field of knowledge graphs and data processing. The method comprises the following steps: after a skill description text set containing operation rules and regulations, equipment manuals and process flowcharts is acquired, skill meta-item analysis processing is performed, a meta-item set containing operation objects, action types and environment parameters is generated, application layer adaptation rules are determined for each meta-item, a skill knowledge graph containing skill nodes and skill correlation edges is constructed, and feature aggregation processing is performed on the graph to generate skill portraits of multiple work posts. The method realizes the automatic generation of manufacturing skill portraits by analyzing multiple source heterogeneous texts into structured meta-items and constructing a field knowledge graph.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph and data processing technology, specifically a method for constructing manufacturing skills profiles based on knowledge graphs. Background Technology

[0002] In the manufacturing sector, job skill descriptions are scattered across various unstructured texts such as operating regulations, equipment manuals, and process flowcharts. Existing methods for building skill profiles largely rely on manual sorting or automated extraction techniques based on simple keyword matching.

[0003] Manually compiling skill profiles is time-consuming and struggles to maintain consistency across profiles when dealing with a large number of devices and processes. Automated techniques based on keyword matching fail to effectively identify the semantic relationships between operational objects, action types, and environmental parameters within text, often isolating strongly related skill elements into isolated labels while ignoring the numerical environmental constraints provided by the textual context. This results in skill profiles that fail to accurately reflect the complex operational capabilities required for a specific job and the applicable working conditions.

[0004] The relationships between skill elements in manufacturing are characterized by deep hierarchies and numerous constraints. Current methods often employ flat lists or simple tree structures to organize these elements, which are insufficient to handle the complex network structures arising from co-occurrence coupling and sequential dependencies among skill components. Constructing skill profiles requires addressing how, based on the extraction of semantic elements, the meta-level parsing results can be reorganized into a graph system capable of expressing the complex relationships between skills, and how to aggregate and generate structured skill profiles with job-specific differentiation based on the topological characteristics of nodes in the graph. Summary of the Invention

[0005] This invention provides a method for constructing manufacturing skills profiles based on knowledge graphs. The purpose is to solve the problems that existing methods cannot automatically parse and associate skill elements from multi-source heterogeneous texts, and cannot construct complex semantic relationships between skill elements into a knowledge structured representation that can be aggregated and analyzed.

[0006] To achieve the above objectives, this invention provides the following technical solution: This invention provides a method for constructing manufacturing skill profiles based on knowledge graphs. This method acquires a set of skill description texts for the target manufacturing sector, which includes operating regulations texts, equipment manual texts, manufacturing-related professional qualification standards, and process flow diagram texts. By segmenting the collected operating regulations texts and equipment manual texts according to page layout, and then correcting the text line alignment of the segmented pages, the non-standardization barrier of multi-source heterogeneous texts in terms of structure is overcome, enabling subsequent semantic parsing to be performed on a unified page layout standard.

[0007] After obtaining the skill description text set, skill meta-item parsing is performed on the skill description text set to generate a skill meta-item set for the target manufacturing industry. This skill meta-item set includes operation object meta-items, action type meta-items, and environmental parameter meta-items. Preferably, the skill meta-item parsing process is based on paragraph boundaries to split entities, and the operation object meta-items and action type meta-items are extracted using the dependency relationship between verb and noun parts of speech. At the same time, environmental parameter meta-items are extracted based on numerical units and dimensional identifiers. By performing tuple combination processing on the extracted meta-items to generate skill meta-item groups, a precise transformation from unstructured text to structured skill elements is achieved, avoiding excessive reliance on manual annotation rules.

[0008] The application layer adaptation rule is determined based on the type attribute of each skill meta-item. This application layer adaptation rule includes the mapping relationship between meta-item attribute values ​​and knowledge graph node types. As a technical solution of this invention, for the operation object meta-item, a first mapping path is determined based on the physical attribute classification label, and a second mapping path is determined based on the functional attribute classification label. The weight values ​​of the two mapping paths are compared based on the frequency of occurrence of the operation object meta-item in the skill description text, and the mapping path with the higher weight value is set as the application layer adaptation rule. Through this dynamic competitive rule selection mechanism, the top-level model of the knowledge graph can adapt to the emphasis of actual terminology in the domain, improving the structural fit of knowledge modeling.

[0009] A skill knowledge graph for the target manufacturing sector is constructed based on defined application-layer adaptation rules. This skill knowledge graph includes a set of skill nodes and a set of skill-related edges. Each skill-related edge connects two skill nodes and is labeled with a correlation strength value. During construction, skill-related edges are established based on the co-occurrence relationships between skill elements, and the initial correlation strength values ​​are attenuated based on the distance between elements in the text to accurately reflect the spatiotemporal coupling degree between skill elements. Preferably, a depth-first traversal-based circular structure detection mechanism is introduced. When a circular structure is detected, the skill-related edge with the smallest correlation strength value is deleted, thus ensuring the acyclic hierarchical structure of the knowledge graph and making the dependencies and progression relationships between skills logically clear and without redundancy.

[0010] The constructed skill knowledge graph undergoes feature aggregation processing to generate skill profiles for multiple job positions within the target manufacturing sector. Preferably, core skill nodes are selected based on node degree values. Using each core skill node as the center, a breadth-first traversal is performed along the skill association edges in the skill knowledge graph within a specified number of hops to extract a set of adjacent skill nodes. The attribute values ​​of the nodes in this set are then vectorized to generate corresponding skill profile vectors. Through this feature aggregation method that combines topological centrality with vector space embedding, the extracted skill profiles can simultaneously represent the core hub characteristics of the skill structure and the surrounding associated skill graph, providing a high-dimensional, computable feature representation of the skill requirements for complex positions.

[0011] As a technical solution of this invention, the method also includes a dynamic feedback update mechanism. During a preset evaluation period, actual operation records from multiple job positions are collected. These records include operation time data and the frequency of operation errors. A first deviation value is generated by comparing the actual operation records with the skill profiles of the corresponding job positions based on the distribution range of operation time data. The association strength values ​​of skill association edges in the skill knowledge graph are adjusted according to the first deviation value, generating an updated set of skill association edges, which is then applied to the skill profile generation task in the next cycle. This feedback mechanism enables the static knowledge graph to perceive the dynamic evolution of front-line operational performance, achieving self-calibration and timeliness maintenance of the skill profile during continuous learning.

[0012] The technical effects and advantages provided by the present invention in the above technical solution are as follows: To address the issue of the inability to effectively correlate and parse the operation objects, actions, and environmental parameters in multi-source heterogeneous skill texts, this paper proposes a skill element parsing process for skill description text sets. Based on the dependency relationship between verbs and nouns, as well as numerical units and dimensional identifiers, operation object elements, action type elements, and environmental parameter elements are extracted separately and combined into skill element groups. This process achieves the structured decomposition and encapsulation of skill elements in unstructured texts from the manufacturing industry, avoiding the shortcomings of conventional keyword matching methods that sever the inherent connections between operation entities, actions, and working environments.

[0013] Based on the physical attribute classification labels and functional attribute classification labels of the operation object meta-items, mapping paths pointing to entity node types and concept node types are determined respectively. The two paths are compared using their frequency of occurrence in the skill description text as weight, and the optimal application-layer adaptation rule is selected. This adaptation method allows the same skill meta-item to be dynamically mapped to a more appropriate node dimension in the knowledge graph based on its distribution characteristics in the corpus. Compared to fixed template mapping, this preserves the polysemous adaptability of skill elements within the domain context.

[0014] In the construction of the skill knowledge graph, skill association edges are established based on the co-occurrence relationships between meta-items, and the initial value of the association strength is attenuated by the distance of the meta-items in the text. Distance attenuation correction makes the association strength in the graph structure closer to the actual distribution of semantic density in manufacturing texts, reducing the interference of noisy associations caused by accidental co-occurrence in long text paragraphs on the graph quality.

[0015] When generating skill profiles for each job position, core skill nodes are selected based on node degree values, and the hop count range is determined according to the average path length of the graph. Breadth-first traversal is performed along the associated edges to extract the set of adjacent nodes, and then the profile is generated through vectorized encoding. This mechanism automatically aggregates node groups belonging to the same skill cluster based on the topology, enabling the final skill profile to integrate the core skills required for a job position and the closely related operational capabilities, thus transcending the limitations of flat, list-style skill descriptions.

[0016] By collecting actual operation time and error frequency data for each position, the operation time data is compared with a standard time interval to generate a deviation value. This deviation value is then used to adjust the strength of the related edges in the skill profile, forming an updated profile for the next cycle. This closed-loop mechanism keeps the skill profile dynamically synchronized with the actual execution on the job site, avoiding the problem of static profiles gradually becoming distorted due to changes in the production environment. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0018] Figure 1 This is a flowchart of a knowledge graph-based method for constructing skill profiles in the manufacturing industry. Figure 2 This is a flowchart of the skill description text generation process; Figure 3 This is a flowchart of the skill profile generation and dynamic update process; Figure 4 It is a skill description text region segmentation affinity analysis diagram; Figure 5 This is a statistical chart showing the hierarchical relationship of skill description paragraphs; Figure 6 This is a chart analyzing the deviation in operation time at different work positions. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] See Figure 1 This invention provides a method for constructing skill profiles in the manufacturing industry based on knowledge graphs. The method includes: acquiring a set of skill description texts for a target manufacturing sector, the set including operational regulations, equipment manuals, relevant manufacturing professional qualification standards, and process flow diagrams; parsing the skill description texts to generate a set of skill elements for the target manufacturing sector, the set including operation object elements, action type elements, and environmental parameter elements; determining corresponding application layer adaptation rules based on the type attributes of each skill element, the application layer adaptation rules including the mapping relationship between element attribute values ​​and knowledge graph node types; constructing a skill knowledge graph for the target manufacturing sector based on the application layer adaptation rules, the skill knowledge graph including a set of skill nodes and a set of skill association edges, each association edge connecting two skill nodes and marking the association strength value; and performing feature aggregation processing on the skill knowledge graph to generate skill profiles for multiple job positions within the target manufacturing sector.

[0021] Example 1: See Figure 2 Collect operating procedure texts from the target manufacturing sector. These operating procedure texts include descriptions of process steps, flowcharts, and safety regulations. Collect equipment manual texts from the target manufacturing sector. These equipment manual texts include control panel labels and parameter adjustment ranges.

[0022] In practice, parsing the flowcharts in the operational regulations text allows for the integration of graphic information from drawings with textual regulations. The specific parsing steps are as follows: 1. Vectorization of drawings: Perform vectorization processing on PDF process drawings and paper scanned flowcharts, identify three types of graphic elements: process rectangles, flow arrows, and equipment icons, and record the planar coordinates and related connection relationships of each graphic element; 2. OCR Extraction of Drawing Annotations: Perform OCR text recognition on the text annotations and parameter annotations within each process drawing element to extract the equipment component names, operating actions, and temperature / pressure / duration parameters within the drawing element; 3. Spatial alignment and fusion of text and graphics: Based on the plane coordinates of the drawing, the equipment, actions, and parameters identified in the flowchart are spatially matched with the operating rules and equipment manual text blocks in the same document; if the flowchart process and the text paragraph description are the same process, the text content of the two are merged. 4. Pre-generation of flowchart elements: Based on the graphic annotation text extracted by OCR, the same dependency parsing method as in Example 2 is used to pre-generate the operation object elements, action type elements, and environment parameter elements corresponding to the flowchart. These elements will then be uniformly incorporated into the global skill element set to participate in subsequent graph construction.

[0023] In practice, the collection methods for operating procedures and equipment manuals include extracting electronic documents from the enterprise document management system or generating editable text from paper pages using optical character recognition. The operating procedures and equipment manuals are then segmented according to page layout, with the segmentation based on the relationship between text block spacing and font size.

[0024] The region segmentation process is implemented as follows: For each page of the operating regulations text and each page of the equipment manual text, all text blocks within the page are identified. A text block is a character sequence with consecutive characters and a consistent font format, obtained through optical character recognition. The boundary coordinates, average character height, and average character width of each text block are extracted. Based on the vertical and horizontal spacing, as well as the average character height and average character width, between two adjacent text blocks, the affinity between them is calculated. Affinity indicates the degree to which two adjacent text blocks tend to belong to the same layout region. The calculation formula is as follows: in, Indicates affinity, This represents the vertical spacing between two adjacent text blocks, in pixels. This represents the average character height of the uppermost text block among two adjacent text blocks, in pixels. This represents the horizontal spacing between two adjacent text blocks, in pixels. This represents the average character width of the leftmost text block among two adjacent text blocks, in pixels. and These are the weighting coefficients. The value is 0.5, and the weighting coefficient is... The value is set to 0.5 to ensure that the vertical and horizontal spacing contribute equally to affinity. The preset threshold is set to 1.0, determined by statistically analyzing the distribution of spacing between text blocks within and across regions on a sample page, and is used to distinguish between text blocks within and across regions. When affinity... When the affinity is less than the preset threshold of 1.0, two adjacent text blocks are divided into the same layout area; when the affinity is less than the preset threshold of 1.0, the two adjacent text blocks are divided into the same layout area. When the value is greater than or equal to a preset threshold of 1.0, the two adjacent text blocks are divided into different layout regions. The process involves iterating through all adjacent text blocks on each page of the operating regulations text and each page of the device manual text to complete the region segmentation process, resulting in the segmented page.

[0025] Text line alignment correction is performed on the page after region segmentation, based on the projection distribution calculation of page coordinates. In practice, the text line alignment correction process is as follows: For each layout area in the segmented page, the baseline ordinates of all text lines within that area are extracted. The baseline ordinates are the average of the bottom ordinates of all characters in the text line. Using the vertical direction of the page as the projection axis, the frequency of occurrence of the baseline ordinates of all text lines within the layout area is counted, generating a horizontal projection histogram. The peak position in the horizontal projection histogram is detected, and the baseline ordinate corresponding to the peak position is determined as the standard line baseline for that layout area. The offset between the baseline ordinate of each text line within the layout area and the standard line baseline is calculated. Based on the offset, each text line is vertically shifted until its baseline ordinate coincides with the standard line baseline. This text line alignment correction operation is performed on all layout areas in the segmented page to obtain the corrected page. All corrected pages constitute a set of skill description text.

[0026] See Figure 4 The figure shows the distribution of affinity characteristics between adjacent text blocks in operating regulations and equipment manuals in the manufacturing industry in the form of a scatter plot. The horizontal axis represents the ratio of the vertical spacing between two adjacent text blocks to the average character height of the taller text block. The vertical axis represents the ratio of the horizontal spacing between two adjacent text blocks to the average character width of the leftmost text block. Both coordinates are dimensionless ratios. In the graph, blue dots mark text block pairs within the same region, red crosses mark text block pairs across regions, and the dashed line represents the preset affinity threshold. The dividing line and threshold are determined based on the statistical distribution of text block spacing within the same and cross-regional areas of the operating regulations text and equipment manual text sample pages.

[0027] As can be seen from the figure, the vertical spacing ratio of text block pairs in the same area are mostly concentrated between 0 and 1.0, and the vast majority are distributed in the range of 0.0 to 0.8, the horizontal spacing ratio is concentrated between 0.0 and 1.0, and most are lower than 0.75. The overall distribution is characterized by dense distribution in the lower left corner, indicating that the spacing between text blocks in the same layout area is relatively small, and both vertical and horizontal spacings maintain a relatively compact layout structure.

[0028] The vertical spacing ratio of cross-region text block pairs is mostly concentrated between 1.0 and 2.2, the horizontal spacing ratio is mainly distributed from 0.8 to above 2.0, the whole is located in the upper right area of the figure, and the vast majority of data points are located at the affinity threshold line in the upper right, indicating that the vertical and horizontal spacings between cross-region text blocks are significantly larger, with obvious spatial separation.

[0029] The dashed threshold boundary clearly distinguishes the same-region text block pairs from the cross-region text block pairs, which verifies the affinity calculation formula and the validity of the threshold of 1.0. In the actual region segmentation process, this threshold can be used to accurately determine whether adjacent text blocks belong to the same layout area. The specific operation is that when the affinity the text blocks are classified into the same region, they are divided into different regions, so as to realize accurate region segmentation based on spacing features.

[0030] Embodiment 2: The skill description text set is subjected to entity splitting according to paragraph boundaries to generate a plurality of skill description paragraphs, and the entity splitting is realized based on the matching relationship between the title level identifier and the paragraph indentation value. In a specific implementation, for each page in the skill description text set, the title lines in the page are identified. The identification method of the title line is: extracting the character content of the text line, matching the character content with a predefined title serial number regular expression, the predefined title serial number regular expression covers patterns such as "Chapter [One Two Three Four Five Six Seven Eight Nine Ten]" and "\d+(\.\d+) "; at the same time, the font attribute of the text line is detected. When the font bold attribute of the text line is true or the font size is larger than the preset font size threshold of 14 points, the text line is determined to be a title line. For the identified title line, the title level identifier is extracted, and the title level identifier is the level depth value of the number sequence in the title serial number. For example, the title level depth value of the title serial number "2.1.3" is 3. The paragraph indentation value of the title line is recorded, and the paragraph indentation value is the difference between the upper left abscissa of the first character of the title line and the abscissa of the left margin of the page, and the unit is millimeter.

[0031] For each paragraph following the heading line, extract the paragraph indentation value. The paragraph indentation value is the difference between the top-left x-coordinate of the first character of the paragraph and the x-coordinate of the left edge of the page. The hierarchical relationship between the heading line and subsequent paragraphs is established by calculating the association score. Implementation, related scores The expression is: in, Indicates the associated score, This indicates the paragraph indentation value for the heading line, in millimeters. This indicates the paragraph indentation value, in millimeters. This indicates the heading level depth value. The value of is an integer greater than or equal to 1. This represents the indentation difference weight, with a value of 0.55. This represents the hierarchy depth weight, with a value of 0.45. Indentation difference weight. and hierarchy depth weight The selection criteria are as follows: Based on a manually indexed set of 300 manufacturing document layout templates, the weight combinations were iterated and the paragraph attribution prediction accuracy was calculated; the weight values ​​that resulted in the highest accuracy were then selected. When the correlation score... If the score is less than or equal to the preset attribution threshold of 0.35, the paragraph is assigned to the area where the current heading line is located; when the correlation score is less than or equal to the threshold of 0.35, the paragraph is assigned to the area where the current heading line is located. When the value exceeds a preset threshold of 0.35, paragraph collection for the current title row's corresponding area is stopped, and that paragraph is used as the starting paragraph for the next title row area. This process is repeated across the entire page, packaging each title row and its corresponding paragraph into a single skill description paragraph. When no title row exists on the page, the entire page is divided into natural paragraphs and treated as independent skill description paragraphs.

[0032] For each skill description paragraph, the following processing is performed: Based on the dependency relationship between verb and noun parts of speech in the skill description paragraph, at least one operation object meta-item and at least one action type meta-item are extracted. The operation object meta-item corresponds to the device component identifier, and the action type meta-item corresponds to the operation action category. In specific implementation, the skill description paragraph is input into a pre-trained dependency parsing model. The pre-trained dependency parsing model adopts a bilinear attention dependency parsing architecture based on a Transformer encoder, containing 6 Transformer encoding blocks, each with 8 multi-head attention heads, a hidden layer dimension of 512, and a feedforward network dimension of 2048. The encoder output is connected to two bilinear classifiers, used to predict the dependency arc direction and dependency relation label, respectively. The input of the dependency parsing model is the word sequence of the skill description paragraph after being processed by a word segmenter. The word segmenter adopts a manufacturing domain sub-word segmenter based on byte-pair encoding, with a vocabulary size of 32,000. The output of the dependency parsing model includes part-of-speech tagging for each word, the word index pointed to by the headword of each word, and the dependency relation type. The training process of the dependency parsing model is as follows: pre-training is performed using a general domain dependency treebase for 100 epochs, with the AdamW optimizer used, an initial learning rate of 0.0005, and a batch size of 64. After pre-training, fine-tuning is performed using a corpus labeled with dependency relations of manufacturing operation instructions. The fine-tuning corpus contains 12,000 operation instruction sentences, and fine-tuning is performed for 30 epochs. The learning rate is adjusted to 0.00003, and the loss function is the sum of the log-likelihood loss for dependency arc prediction and the cross-entropy loss for dependency relation classification.

[0033] To ensure model reproducibility, specific labeling standards, dependency labels, and loss weighting rules are defined for the manufacturing industry: 1. Manufacturing-specific dependency label list: Only three types of domain-specific dependency labels are used, namely equipment-operation, operation-parameter, and component-working condition. Redundant dependency labels in general domains are discarded, and relationship labels are only made for manufacturing process text. 2. Weighting of loss function: Total loss = 0.6 × Dependency arc log-likelihood loss + 0.4 × Dependency relationship cross-entropy classification loss. Arc prediction task has higher weight to prioritize the accuracy of entity action matching. 3. Criteria for determining whether to include a corpus in the database: All manufacturing operation instruction corpora must be independently annotated by two senior process engineers. Only when the dependency arcs and dependency labels annotated by the two engineers are completely consistent can the sample be included in the fine-tuning training corpus. If there is a discrepancy between the two annotations, a third senior process engineer must be organized to review and adjudicate. Only after the adjudication is approved can the sample be included in the database. Discrepancies will be directly removed and will not be used for model training.

[0034] For each sentence's dependency parsing result, all dependency arcs are traversed. When the dependency relation type is "dobj" (direct object) and the part-of-speech tag of the head word belongs to the verb category, the head word is used as the source for extracting the action type meta-item, and the direct object is used as the source for extracting the operation object meta-item. The original verb form of the head word undergoes lemmatization processing, restoring the inflectional changes to the verb infinitive, and mapping it to a pre-constructed operation action classification vocabulary. The pre-constructed operation action classification vocabulary contains standardized verbs representing manufacturing operation actions, such as "start," "shut down," "adjust," "switch," "assemble," "disassemble," "clean," "calibrate," "monitor," and "record." The extraction method for operation object meta-items is as follows: extract the core noun of the direct object noun phrase, remove quantifier prefixes and modifier suffixes, retain the textual description of the equipment component identification, and unify the English abbreviations to uppercase, for example, "hydraulic oil pump" is directly used as the operation object meta-item.

[0035] Based on the numerical units and dimensional identifiers in the skill description paragraph, extract at least one environmental parameter element, corresponding to a temperature range or pressure range value. Combine the object element, action type element, and environmental parameter element into tuples to generate a skill element group. In practice, a single skill description paragraph is used as a processing unit, and the total number of object elements, action type elements, and environmental parameter elements extracted within that unit is counted. The sets of object elements, action type elements, and environmental parameter elements are input into a combiner, which generates all possible ordered triples. Each ordered triple has the structure (object element, action type element, environmental parameter element). When no environmental parameter element is extracted within a processing unit, the position of the environmental parameter element in the triple is filled with the placeholder "NULL". When a processing unit extracts multiple operation object meta-items and multiple action type meta-items, but the environment parameter meta-item is missing or only has a single environment parameter meta-item, the combiner iterates through the Cartesian product of the operation object meta-items and action type meta-items, forming multiple triples with the same environment parameter meta-item or placeholder. The set of triples generated by all processing units constitutes the skill meta-item group. The filtering rules for NULL parameter meta-items are: NULL triples are not included in co-occurrence calculations; NULL parameter associated edges are not generated during graph construction; NULL meta-item nodes are removed during vector encoding.

[0036] See Figure 5The bar chart shows the frequency distribution of the association score B between the title line and subsequent paragraphs in the skill description text. The horizontal axis represents the association score B, ranging from 0 to 1, and the vertical axis represents the frequency of occurrence of the corresponding association score. Green bars indicate the frequency of paragraphs belonging to the title line area with an association score B less than or equal to a preset attribution threshold of 0.35, while purple bars indicate the frequency of paragraphs with an association score B greater than the threshold of 0.35 where data collection stops. The preset attribution threshold of 0.35 is marked by a black dashed line, clearly dividing the overall association score range into two parts.

[0037] As shown in the figure, the frequency of correlation score B in the range of 0 to 0.35 is relatively high and exhibits a right-skewed distribution. The frequency peaks are concentrated between approximately 0.05 and 0.20, with the highest frequency exceeding 70 times. This indicates that the indentation differences of most paragraphs match the heading level depth quite well, resulting in a large number of paragraphs belonging to the corresponding heading lines, which conforms to the rule in Example 2 for determining paragraph affiliation based on correlation score B. The purple area where the correlation score B exceeds 0.35 is relatively evenly distributed, with a frequency concentrated between 30 and 45 times. The correlation scores are mostly in the range of 0.45 to 0.75, indicating that when the correlation score exceeds the threshold, paragraph affiliation collection stops, and the paragraphs are divided into different heading regions, ensuring the accuracy and clear hierarchy of the skill description paragraph division.

[0038] This graph visually reflects the rationality and effectiveness of using the association score B as a paragraph attribution criterion in Example 2 through frequency statistics. It verifies that when the association score B is less than or equal to 0.35, the paragraph should belong to the title line area; conversely, when the association score B is greater than 0.35, the collection of the currently assigned paragraphs stops, and a new title area division is initiated, thus achieving entity splitting of the skill description text. The clear segmentation of the two frequency distributions in the graph also demonstrates the discriminative power of the preset attribution threshold of 0.35, which is helpful for the accurate extraction of subsequent skill elements and the construction of the knowledge graph.

[0039] Example 3: In practical implementation, a physical attribute thesaurus and a functional attribute thesaurus are pre-built. The physical attribute thesaurus includes terms representing physical components, parts, materials, and tooling of manufacturing equipment, such as "motor," "spindle," "hydraulic cylinder," "conveyor belt," "clamp," "tool holder," "pneumatic valve," and "sensor probe." Each term in the physical attribute thesaurus has a name and a physical attribute category tag, with the value of the physical attribute category tag being "physical entity." The functional attribute thesaurus includes terms representing equipment functions, control loops, process effects, and signal interfaces, such as "drive module," "cooling loop," "pressure control unit," "signal feedback path," "position closed loop," and "temperature compensation device." Each term in the functional attribute thesaurus has a name and a functional attribute category tag, with the value of the functional attribute category tag being "functional concept." The physical attribute lexicon and the functional attribute lexicon are constructed as follows: 3,000 noun phrases are manually sampled from the skill description text set of the target manufacturing field. At least two manufacturing process engineers independently assign each noun phrase a category of "physical entity" or "functional concept". Words with consistent labeling are assigned to the physical attribute lexicon and the functional attribute lexicon respectively. Words with inconsistent labeling are assigned to the corresponding lexicon after discussion and decision.

[0040] For the operation object meta-items to be processed, the text string of the operation object meta-items is subjected to exact matching and vector semantic matching in the physical attribute lexicon and the functional attribute lexicon, respectively. For exact matching, a match is successful if the edit distance between the text of the operation object meta-item and the name of a term in a certain lexicon is zero. For vector semantic matching, the operation object meta-items and lexicon terms are encoded into 768-dimensional sentence embedding vectors using a pre-trained sentence vector model. The sentence vector model adopts the all-mpnet-base-v2 model based on the SBERT architecture, with the term string as input and a 768-dimensional vector as output. The cosine similarity between the sentence vector of the operation object meta-item and the sentence vector of the term in the lexicon is calculated. When the cosine similarity is greater than or equal to a preset similarity threshold of 0.82, the vector semantic matching is considered successful. If the operation object meta-item matches any term in the physical attribute lexicon, the physical attribute classification label of the operation object meta-item is "physical entity"; if it matches any term in the functional attribute lexicon, the functional attribute classification label is "functional concept". If the operation object meta item fails to match either of the two thesauruses, the physical attribute category label and the functional attribute category label will be set to "Undetermined" by default, and the operation object meta item will not participate in the subsequent mapping path weight value calculation.

[0041] The first mapping path is determined based on the physical attribute classification label, which maps the operation object meta-item to the entity node type in the knowledge graph; the second mapping path is determined based on the functional attribute classification label, which maps the operation object meta-item to the concept node type in the knowledge graph.

[0042] In practical implementation, the knowledge graph's node type system defines two top-level node types: entity node types and concept node types. Entity node types are used to instantiate and represent specific identifiable components in the physical configuration of equipment, while concept node types are used to represent abstract categories of equipment functions and process logic. When the physical attribute classification label of the operation object meta-item is "physical entity," the target node type of the first mapping path is set to the entity node type; when the functional attribute classification label of the operation object meta-item is "functional concept," the target node type of the second mapping path is set to the concept node type. When the mapping operation of the first mapping path is executed, a node with the corresponding name is created or referenced under the entity node type level of the knowledge graph, and the node identifier is set to the standardized code of the operation object meta-item; when the mapping operation of the second mapping path is executed, a node with the corresponding name is created or referenced under the concept node type level of the knowledge graph.

[0043] The path weight values ​​of the first mapping path and the second mapping path are compared. The path weight value is calculated based on the frequency of occurrence of the operation object meta-item in the skill description text. The mapping path with the higher path weight value is set as the application layer adaptation rule.

[0044] In practical implementation, the formula for calculating the path weight value is as follows: in, Indicates the path weight value. This indicates the frequency of occurrence of the operand meta-item within a specific text subset, that is, the total number of times the operand meta-item appears in the specified text type. This represents the total number of words in a specific text subset, i.e., the total number of all words in the specified text type. The specific text subset is divided into a first text subset and a second text subset. The first text subset consists of all operating regulations texts and all equipment manual texts from the skill description text set, while the second text subset consists of all process flow diagram texts from the skill description text set.

[0045] Calculate the first path weight value corresponding to the first mapping path respectively. The second path weight value corresponding to the second mapping path First path weight value In the calculation, Take the frequency of occurrence of the operation object's meta-item in the first text subset. The total word count of the first text subset. Second path weight value. In the calculation, Take the frequency of occurrence of the operation object's meta-item in the second text subset. Calculate the total word count of the second text subset. When the physical attribute category label of the operation object meta-item is "physical entity" and the functional attribute category label is "functional concept", calculate the first path weight value respectively. Second path weight value Compare the weight values ​​of the first path. Second path weight value The size, if the first path weight value Greater than the second path weight value If the weight value of the second path is less than 1, then the first mapping path is set as the application layer adaptation rule; if ... Greater than the first path weight value Then the second mapping path is set as the application layer adaptation rule; if the weight value of the first path is... Equal to the second path weight value If the first mapping path is set as the default application layer adaptation rule, then when the only physical attribute category label value of the operation object meta-item is "physical entity", the first mapping path is directly selected as the application layer adaptation rule, without comparing path weight values. When the only functional attribute category label value of the operation object meta-item is "functional concept", the second mapping path is directly selected as the application layer adaptation rule. The application layer adaptation rule contains the mapping relationship between meta-item attribute values ​​and knowledge graph node types. This mapping relationship records the allocation scheme from operation object meta-items to entity node types or concept node types. When constructing the skill knowledge graph subsequently, node type selection and node instantiation are performed based on this mapping relationship.

[0046] Example 4: In the specific implementation, an empty initial knowledge graph is pre-constructed. The initial knowledge graph adopts an attribute graph model, and node storage structure and edge storage structure are defined. The node storage structure includes a node identifier field, a node type field, and an attribute value dictionary field; the edge storage structure includes an edge identifier field, a start node identifier field, a stop node identifier field, an association strength value field, and an edge type field.

[0047] According to the application layer adaptation rules, at least one skill node is inserted into the pre-built initial knowledge graph. Each skill node contains a node identifier and node attribute values. For each skill meta-item in the skill meta-item set, the meta-item name is extracted. The mapping relationship between the meta-item attribute values ​​recorded in the application layer adaptation rules and the knowledge graph node types is queried to determine the target node type corresponding to the skill meta-item. The target node type is either an entity node type or a concept node type. The skill meta-item name is standardized by taking the lowercase form of the original string, removing punctuation and spaces, and prepending it with the abbreviation string of the target node type to generate a node identifier. A search is performed in the initial knowledge graph using the node identifier as an index. If the search result is non-existent, a new skill node is created, and the node identifier and node type are written to the new skill node. The attribute value set of the skill meta-item is also written to the attribute value dictionary field of the new skill node. If the search result indicates that the skill meta-item already exists, the attribute value set of the skill meta-item is merged and updated to the attribute value dictionary field of the existing skill node.

[0048] Based on the co-occurrence relationships between each skill element in the skill element set, establish at least one skill association edge. Each skill association edge includes a start node identifier, an end node identifier, and an association strength value. Obtain the number of times the first and second skill elements in the skill element set co-occur in the skill description text set. The co-occurrence count is calculated as follows: divide the skill description text set into multiple skill description paragraphs, traverse each skill description paragraph, and check whether the first and second skill elements appear simultaneously in the same skill description paragraph. If they appear simultaneously, it is counted as one co-occurrence. Accumulate the co-occurrence detection counts for all skill description paragraphs during the traversal to obtain the total co-occurrence count.

[0049] The initial value of the association strength between the first skill meta-item and the second skill meta-item is determined based on the ratio of the co-occurrence frequency to the total number of paragraphs in the skill description text set. The initial value of the association strength is calculated by dividing the co-occurrence frequency by the total number of paragraphs in the skill description text set; the quotient is the initial value of the association strength.

[0050] The initial value of the association strength is subjected to attenuation correction, which is based on the positional distance between the first skill element and the second skill element in the text. The positional distance is calculated as follows: for each skill description paragraph where the first and second skill elements co-occur, the sentence index of the first occurrence of the first skill element in the skill description paragraph is located, the sentence index of the first occurrence of the second skill element in the skill description paragraph is located, and the absolute difference between the two sentence indices is calculated as the single paragraph distance in that co-occurring paragraph; the single paragraph distances of all co-occurring paragraphs are arithmetically averaged to obtain the average sentence distance between paragraphs, which is recorded as the positional distance value.

[0051] The correlation strength value after attenuation correction is determined by the attenuation correction formula, which is: in, This represents the correlation strength value after attenuation correction. This represents the initial value of the association strength. This represents the positional distance value, specifically the average sentence distance between the first and second skill elements across all co-occurring paragraphs, expressed in sentences. This represents the attenuation adjustment coefficient. The value is 0.15. Attenuation adjustment coefficient. The value of 0.15 is determined based on the selected location distance value. Attenuation multiplier when there are 10 sentence intervals Approximately 0.22, this compresses the association strength value to about one-fifth of its initial value in scenarios with large-interval co-occurrence, preventing excessive contribution of association strength from long-distance co-occurrence; simultaneously, in terms of location distance value... When the interval is 2 sentences, the attenuation multiplier is approximately 0.74, which is sufficient to preserve the main association strength of close co-occurrence.

[0052] Assign the attenuation-corrected association strength value to the skill association edge. Create a new skill association edge in the initial knowledge graph, setting the starting node identifier of the skill association edge to the node identifier of the skill node corresponding to the first skill element, and setting the ending node identifier of the skill association edge to the node identifier of the skill node corresponding to the second skill element. Set the association strength value field of the skill association edge to... The value is set to "Skill Co-occurrence Relationship" in the edge type field. If a skill association edge already exists between the start node identifier and the end node identifier, then use... The value is updated in the association strength value field of the existing skill association edge.

[0053] A deduplication rule for unidirectional storage of associated edges has been added. Before creating a skill-related edge, the two node identifiers are compared in size, and only the smaller node identifier is used as the starting node and the larger node identifier is used as the ending node to generate a unidirectional edge. In all operation logic such as graph traversal, node degree value statistics, and cycle detection, this unidirectional edge is treated as a bidirectional relationship, and the reverse edge is no longer generated repeatedly, thus eliminating the problem of graph storage redundancy and the doubling distortion of node degree value statistics.

[0054] Cycle structure detection is performed on skill-related edges, based on path marking using depth-first traversal. In practice, a list of all skill nodes and all skill-related edges in the current knowledge graph is obtained. An adjacency list is constructed to represent the knowledge graph structure. The key of the adjacency list is the node identifier, and the value is a list of all skill-related edges originating from that node. Each skill-related edge records the terminating node identifier and the association strength value.

[0055] Initialize a global access state set to record whether each node has been fully explored in this global detection. Iterate through each node in the list of all skill nodes. If the current node is not in the global access state set, perform a depth-first traversal starting from the current node. During the depth-first traversal, maintain a path node boolean tag set and a path edge list. The path node boolean tag set records the identifiers of all nodes along the path from the starting node of this traversal to the current recursion level. When visiting each adjacent edge of the current node, obtain the terminal node identifier of the adjacent edge. If the terminal node identifier exists in the path node boolean tag set, a loop structure is detected. Backtrack along the path edge list to locate the sequence of consecutive skill-related edges contained in the loop structure, and add all skill-related edges in this sequence to the candidate deletion set. If the terminal node identifier is not in the path node boolean tag set and is not marked by the global access state set, recursively enter the terminal node, push the current adjacent edge into the path edge list, and add the terminal node identifier to both the path node boolean tag set and the global access state set. Upon recursion, the termination node identifier is removed from the path node boolean flag set, and the current adjacent edge is popped from the path edge list. After all adjacent edges have been processed, the current node is added to the global access state set.

[0056] When a circular structure is detected, the skill association edge with the lowest association strength value within the circular structure is deleted. Specifically, all skill association edges covered by detected circular structures are extracted from the candidate deletion set. These edges are then sorted in ascending order by their association strength value. The edge with the lowest association strength value is selected and deleted from the knowledge graph. If multiple circular structures share the same edge with the lowest association strength value, that edge is deleted globally, not just within the current circular structure. After deletion, the knowledge graph is re-detected for circular structures until no new circular structures are found after traversing all skill nodes.

[0057] Example 5: See Figure 3The process involves obtaining the degree value of each skill node in the skill knowledge graph. The degree value represents the number of skill-related edges connected to that skill node. In practice, a degree value statistics dictionary is constructed. The keys of the degree value statistics dictionary are the node identifiers of the skill nodes, and the values ​​are integer counters, all initialized to 0. Each skill-related edge in the skill knowledge graph's set of skill-related edges is traversed. The start and end node identifiers of the skill-related edge are read, and the integer counters in the degree value statistics dictionary with the start node identifier as the key and the end node identifier as the key are incremented by 1. After completing the full traversal of the skill-related edge set, the value of each key-value pair in the degree value statistics dictionary is the degree value of the corresponding skill node.

[0058] Skill nodes are sorted based on their degree values, and a predetermined number of these nodes are selected as core skill nodes. Specifically, all key-value pairs in the degree value dictionary are sorted in descending order of their values, resulting in an ordered list. The predetermined number is denoted as [number to be filled in]. , The value of is the smaller of one-fifth of the total number of skill nodes in the skill knowledge graph and the value 50. If the total number of skill nodes is... ,but If the total number of nodes is less than 3, all nodes are considered core nodes. Based on this set number, the top-ranked nodes are extracted from the ordered list. Each skill node will... The node identifier of each skill node is stored in the core skill node list.

[0059] Centered on each core skill node, extract a set of adjacent skill nodes within a specified number of hops from the skill knowledge graph. Obtain the node identifier of the core skill node and the specified number of hops, which is determined based on the average path length of the skill knowledge graph. The average path length is calculated as follows: calculate the average shortest path between all nodes in the skill knowledge graph, i.e., exhaustively enumerate the shortest path length between any two different skill nodes in the skill node set, sum these shortest path lengths, and divide by the total number of skill node pairs to obtain the average path length value. Specify the number of hops. The value is set to , This indicates rounding up to the nearest integer, resulting in a specified number of hops. It can cover the topological substructure of the central node within most of the connected regions of the knowledge graph. For example, if there are 100 skill nodes in the current skill knowledge graph, the average shortest path length between any two skill nodes is calculated to be... Specify the number of hops This means that when performing a breadth-first traversal starting from the core skill node and following the skill association edges, the traversal will extend to all adjacent skill nodes within a maximum of 3 hops. Furthermore, to avoid setting the hop count too high when the total number of nodes is very small, causing the traversal range to exceed the actual size of the knowledge graph, or setting the hop count too low when the total number of nodes is very large, resulting in insufficient capture of associated skill information, the following boundary condition is set: when the total number of skill nodes in the skill knowledge graph is less than 20, a hop count is forcibly specified. The value is 2; when the total number of skill nodes is greater than 500, the specified number of jumps is used. The upper limit is fixed at 5. This boundary condition ensures that, in skill knowledge graphs of different scales, the set of adjacent skill nodes extracted by feature aggregation processing can reflect the local topological structure of core skills while also taking into account computational efficiency and the representativeness of the profile.

[0060] Starting from the core skill node, perform a breadth-first traversal along the skill-related edges, recording the hop count for each traversal. The specific traversal method is as follows: initialize a set of visit nodes and a set of boundary nodes; add the node identifier of the core skill node to both the boundary node set and the visit node set, and record the current hop count corresponding to the core skill node as 0; when the boundary node set is not empty and the current hop count is less than a specified hop count... At this point, all node identifiers at the current hop count level are retrieved from the boundary node set to form the current layer node list. Each node identifier in the current layer node list is traversed, and its adjacent node identifiers directly connected to it via skill association edges in the skill knowledge graph are obtained. Each adjacent node identifier is checked to see if it already exists in the visitor node set. If not, the adjacent node identifier is added to the visitor node set and temporarily stored in the next layer node list. After completing the traversal of the current layer node list, the entire next layer node list is assigned to the boundary node set, and the hop count is incremented by 1. When the hop count reaches the specified number of hops... When the current traversal path is terminated, the identifiers of all skill nodes in the visited node set, excluding the core skill node itself, are aggregated to generate an adjacent skill node set. It should be noted that hop count 0 (the core node itself) is not included; all adjacent nodes with hop counts 1 to H are stored in the adjacent skill node set.

[0061] The attribute values ​​of skill nodes in the adjacent skill node set are vectorized and encoded to generate the skill profile vector corresponding to the core skill node. The steps of attribute value vectorization encoding are as follows: For each skill node in the adjacent skill node set, all attribute value texts are extracted from the node attribute value dictionary of the skill node, and the attribute value texts are concatenated into a long text string. The long text string is input into a pre-trained text embedding model. The pre-trained text embedding model adopts the Sentence-BERT model based on the Transformer encoder. The Sentence-BERT model contains 12 Transformer encoder blocks, each with 12 multi-head attention heads, a hidden layer dimension of 768, a feedforward network dimension of 3072, and uses an average pooling strategy. After the input text is segmented into token sequences, the model outputs a 768-dimensional real number vector as the node embedding vector of the skill node. After calculating the node embedding vectors of all skill nodes in the adjacent skill node set one by one, the arithmetic mean of these node embedding vectors is calculated dimension by dimension to generate a 768-dimensional average vector. This average vector is used as the skill profile vector corresponding to the core skill node. All core skill nodes' skill profile vectors are associated with their node identifiers and stored to form a skill profile vector set. Each core skill node corresponds to a job position or is directly associated with the skill requirement description of a job position, thus the skill profile vector set constitutes the skill profiles of multiple job positions within the target manufacturing sector. To achieve precise binding between core skill nodes and specific job positions, the following job matching rules are set: First, obtain the process requirement description text of a certain job position in the target manufacturing sector. This text usually comes from the job operation procedures or job description, and uses the same skill element parsing processing method as in claim 3 to parse the text, extracting the operation object element, action type element, and environmental parameter element corresponding to the job position, forming the job element set. Second, concatenate all element name texts in the job element set to form a job description string, and encode the string using the same pre-trained text embedding model (Sentence-BERT model) as in embodiment 5 to generate a 768-dimensional job requirement vector. Subsequently, calculate the cosine similarity between the job requirement vector and the skill profile vector of each core skill node. Third, set a similarity threshold. When the cosine similarity between the job requirement vector and the skill profile vector of a certain core skill node is greater than... When determining that the skill cluster represented by the core skill node is highly matched with the skill requirements of the job position, the core skill node is bound as the core skill of the job position. By traversing all job positions, each core skill node in the skill profile vector set can be associated with one or more specific job positions in a one-to-one or one-to-many relationship, ultimately forming a complete and applicable job skill profile system.

[0062] Within a preset evaluation period, actual operation records for multiple job positions within the target manufacturing sector are collected. These records include operation time data and the frequency of operation errors. The preset evaluation period is set to thirty calendar days. Data collection is performed by accessing the operation log database of the Manufacturing Execution System (MES). All operation event records within the preset evaluation period are extracted using the job position identifier as the query condition. Each operation event record includes the operation step name, start timestamp, end timestamp, and an error flag field. A value of 0 indicates successful completion, while a value of 1 indicates an error occurred. The operation time for each operation event record is calculated as the time difference between the end timestamp and the start timestamp, in seconds. The operation event records are aggregated by operation step name and job position identifier to form actual operation record entries for each job position at each operation step. Each actual operation record entry includes an operation time data sequence and the frequency of operation errors, where the error frequency is the number of records with an error flag field value of 1.

[0063] The actual operation records are compared with the skill profiles of the corresponding job positions to generate a first deviation value. This deviation analysis is based on comparing the distribution intervals of operation time data. Operation time data is extracted from the actual operation records, which include the start and end timestamps of each operation step. The actual operation records are organized into a time-consuming list indexed by the operation step identifier. A standard time interval for the corresponding operation step is extracted from the skill profile. This standard time interval is calculated based on the statistical distribution of historical operation data. The calculation method is as follows: extract all operation time data for the operation step within the past twelve months from the historical operation record database of the manufacturing execution system, sort these historical operation time data in ascending order, and take the 5th percentile as the lower limit of the standard time interval and the 95th percentile as the upper limit. Precision machining processes use the 10 / 90 quantile, and emergency repairs use the 1 / 99 quantile; furthermore, the quantile switching condition is based on process type label matching.

[0064] The frequency of operation time data falling outside the standard time interval is calculated to obtain the first frequency statistic. Specifically, all operation time data in the actual operation record for this operation step are iterated through, and each operation time data value is checked to see if it is less than the lower limit of the standard time interval or greater than the upper limit of the standard time interval. If the above conditions are met, the first frequency statistic is incremented by 1. Based on the first frequency statistic and the total number of operations in the actual operation record, the first deviation value is calculated. The first deviation value is calculated by dividing the first frequency statistic by the total number of operations in the actual operation record, i.e.: in, This represents the first deviation value. This represents the first frequency statistic, i.e., the frequency with which the operation time falls outside the standard time range. This indicates the total number of operations recorded in the actual operation log.

[0065] Based on the first deviation value, the association strength values ​​of skill-related edges in the skill knowledge graph are adjusted to generate an updated set of skill-related edges. Specifically, for job positions where the first deviation value is greater than a preset deviation threshold of 0.25, the core skill node corresponding to that job position is located. All skill-related edges within the neighborhood of this core skill node are searched in the skill knowledge graph, and the association strength value of each edge is updated. The update operation involves multiplying the original association strength value by an adjustment coefficient. Adjustment coefficient The calculation formula is ,in This is the attenuation factor, with a value of 0.3. The parameter values ​​are selected based on a step size of 0.05, traversing the range from 0.1 to 0.5 in a simulated bias feedback experiment, to achieve the highest bias prediction accuracy of the adjusted knowledge graph in the next cycle. The updated association strength values ​​are written back to the skill association edges, generating an updated set of skill association edges. This updated set of skill association edges is then applied to the skill profile generation task in the next cycle. The application method is as follows: before the start of the next preset evaluation cycle, the updated set of skill association edges replaces the original set of skill association edges in the skill knowledge graph, and feature aggregation processing is re-executed to generate a new skill profile.

[0066] The incremental update logic within the cycle is as follows: collect operation records every 7 days, calculate short-term deviations, incrementally fine-tune the strength of related edges and locally redraw the corresponding job profile vectors, and perform a full regeneration every 30 days.

[0067] See Figure 6In the graph, the horizontal axis represents the operation time in seconds, and the vertical axis represents the frequency. The two vertical black dashed lines mark the lower limit of the standard operation time interval (18.4 seconds) and the upper limit (58.6 seconds), respectively. The green bars represent the frequency distribution of operation times falling within the standard operation time interval, and the red bars represent the frequency distribution of operation times falling outside the standard operation time interval.

[0068] As shown in the graph, the frequency of operations within the specified time range is mainly concentrated between 18.4 seconds and 58.6 seconds, exhibiting a trend of rapid increase followed by slow decrease. A significant peak frequency is observed, exceeding 120 times, indicating that most operations fall within this standard time range. The frequency outside this range is primarily distributed between 18.4 seconds and 58.6 seconds, with an overall significantly lower frequency than within the range. However, a considerable number of operations occur below the lower limit of 18.4 seconds, peaking at approximately 80 times, reflecting that some operations are significantly shorter than the standard range. Furthermore, the frequency of operations exceeding the upper limit of 58.6 seconds exhibits a long-tailed distribution, gradually decreasing in frequency, indicating fewer instances of abnormally long operations. This frequency distribution graph reflects the match between the actual operation time of the target manufacturing sector's job positions and the standard time range in the skill profile within the preset assessment period. The high frequency within the green range indicates that most operations meet the expected standard time, while the presence outside the red range reveals abnormal operation times, providing a quantitative basis for deviation analysis.

[0069] To verify this invention, three sets of mainstream existing technologies in the industry were set as control schemes. Parallel tests were conducted on the same batch of multi-source test data from the manufacturing industry (including 200 operating regulations, 150 equipment manuals, and 80 process flow diagrams, covering two major manufacturing scenarios: machining and assembly). Two manufacturing process engineers jointly and manually annotated the data to form a truth benchmark library, which consisted of 12,600 sets of standard skill triples, 4,820 real process association edges, and 320 job standard skill vectors. Unified evaluation indicators were used: skill element extraction accuracy, skill association noise error, job skill profile matching accuracy, and profile iteration update cycle. The quantitative data are shown in Table 1 below.

[0070] Definition of three control schemes: Compared to Solution 1 (traditional keyword matching method): Skill tags are extracted by using existing keyword segmentation and matching, without syntactic dependency parsing or knowledge graph modeling, and the job skill tag list is directly generated in a flattened manner; Comparison Scheme 2 (Fixed Mapping Knowledge Graph Method): The skill graph is constructed using fixed template mapping rules, without setting a dual-path weight selection and adaptation mechanism, and without text position distance attenuation correction of association strength; Comparison with Scheme 3 (Non-closed-loop static profiling method): The map construction and topology aggregation process of this invention is adopted, but the MES operation record deviation feedback update mechanism is cancelled, and the skill map and profiling remain permanently static after generation.

[0071] This test uses four quantitative evaluation indicators, and the complete definitions of the statistical scope and calculation method of each indicator are as follows: Skill element extraction accuracy: Based on the total number of manually annotated complete triplet samples (operation object element, action type element, and environmental parameter element) in the test corpus, the number of valid triplets obtained by automatic parsing of each scheme without mismatch, omission, or redundancy is counted. The calculation formula is: Extraction accuracy = Number of correctly parsed triplets ÷ Total number of manually annotated triplets × 100%. Judgment criteria: The semantics of operation object, action, and environmental parameter must match the original text of the procedure. Only when there is no component confusion, action misclassification, or parameter range loss can it be considered a valid triplet.

[0072] Skill association noise error: Using all skill association edges in the graph as the statistical base, two manufacturing process engineers manually determine whether each association edge has a real process coupling relationship; false association edges that have no process logic and are only generated by accidental co-occurrence of paragraphs are defined as noise edges. The calculation formula is: Association noise error = Number of false association edges ÷ Total number of association edges in the graph × 100%. The lower this value, the more closely the skill association relationship in the graph matches the logic of the real production process.

[0073] Job Skill Profile Matching Accuracy: Using the standard skill requirement vectors for each job manually calibrated by process engineers as the ground truth, the cosine similarity between the job skill profile vectors output by each solution and the ground truth vectors is calculated; the similarity threshold is set at 0.75, and a similarity ≥ 0.75 is considered a successful match. Profile Matching Accuracy = Number of Successfully Matched Job Samples ÷ Total Number of Job Samples in the Test Set × 100%. The higher the value, the higher the consistency between the automatically generated skill profile and the actual skill requirements of the job.

[0074] Effective update cycle for job profiles: Continuously collect production line MES operation logs and calculate the first deviation value D between the job profile and the actual operation record daily. dev The effective update cycle is the time from when the portrait is generated until it first appears above the deviation threshold; unit: natural days. The longer the cycle, the longer the skill portrait can stably adapt to on-site production conditions.

[0075] Table 1: Data Results of the Comparison Scheme The indicator values ​​in the table are all derived from unified testing and statistics based on the same standardized manufacturing test dataset. The complete description of the test environment, dataset, and calculation process is as follows: 1. Standardized testing basic conditions (1) Test dataset The test corpus includes: 200 operating regulations, 150 equipment manuals, and 80 process flow diagrams, covering two major manufacturing scenarios: machining and automated assembly; manual annotation benchmark: jointly annotated by two senior process engineers, producing a manual standard truth library. A total of 12,600 valid skill triples (operation object + action + environmental parameters) were labeled; 4820 actual process coupling skill-related edges were labeled; Standard skill requirement vectors for 320 job positions were labeled.

[0076] (2) Unified testing rules The three control schemes and the input of the method of this invention are completely consistent, with no additional human intervention; all output results are compared with the human truth database, and the indicators are calculated in batches according to fixed formulas without any subjective human correction.

[0077] 2. Derivation process for numerical calculation of each indicator (1) Accuracy of skill element extraction Calculation formula: Extraction accuracy = Number of correctly identified triples ÷ Total number of manually labeled triples (12600) × 100%; Comparison 1 (traditional keywords): 7850 groups were correctly extracted, 7850 ÷ 12600 × 100% ≈ 62.3%; Comparison 2 (fixed mapping map): 9891 groups were correctly extracted, 9891 ÷ 12600 × 100% = 78.5%; Comparison 3 (static profile without closed loop): 9941 groups were correctly extracted, 9941 ÷ 12600 × 100% ≈ 78.9%; This invention correctly extracted 11,932 groups, with a success rate of 11,932 ÷ 12,600 × 100% ≈ 94.7%. (2) Skill-related noise error Calculation formula: Noise error = Number of false related edges ÷ Total number of related edges generated by the solution × 100% Manual judgment: Only edges that appear simultaneously in the same process and have sequential / cooperative constraints are considered true related edges; edges that appear occasionally in the same paragraph are judged as noise edges.

[0078] Comparison 1: Total number of edges: 3875; Noise edge: 1426; 1426 ÷ 3875 × 100% = 36.8%; Comparison 2: Total number of edges: 3962; Noise edge: 721; 721 ÷ 3962 × 100% ≈ 18.2%; Comparison 3: Total number of edges: 3943; Noise edge: 694; 694 ÷ 3943 × 100% ≈ 17.6%; This invention has a total of 4139 edges, with 178 noisy edges. 178 ÷ 4139 × 100% ≈ 4.3%; (3) Job skill profile matching accuracy Calculation formula: Matching accuracy = Number of positions with cosine similarity ≥ 0.75 ÷ Total number of tested positions (320) × 100%; Comparison 1: 182 qualified positions 182 ÷ 320 × 100% ≈ 57.1%; Comparison 2: 231 qualified positions 231 ÷ 320 × 100% ≈ 72.4%; Comparison 3: 234 qualified positions 234 ÷ 320 × 100% ≈ 73.0%; This invention: 293 qualified positions; 293 ÷ 320 × 100% ≈ 91.5%; (4) Effective portrait update cycle Calculation method: Simulate production line process fluctuations and personnel operation deviations for 60 consecutive days, calculate Ddev daily, and record the number of natural days when the profile first exceeds the 0.25 threshold (average of multiple experiments): Traditional keyword matching: average deviation exceeds the standard over 7 days; Fixed mapping pattern: Average deviation exceeds standard over 16 days; Static profiling without closed-loop control: average deviation exceeds standard every 22 days; The method of this invention has a full cycle of 30 days and supports incremental fine-tuning every 7 days, with controllable deviations within the cycle.

[0079] The following conclusions can be drawn from the quantitative data in the table above: 1. At the level of skill element extraction, traditional keyword matching can only identify surface nouns and verbs, severing the semantic relationship between operations, actions, and working condition parameters, resulting in the lowest extraction accuracy. This invention relies on layout normalization preprocessing and dependency syntax triple parsing, which improves the element extraction accuracy by 32.4% compared to the control scheme 1 and by 16.2% compared to the fixed mapping graph scheme.

[0080] 2. In terms of graph association quality, compared with schemes 1 and 2, which do not introduce a text position distance attenuation correction mechanism, a large number of paragraphs co-occur accidentally to form false skill associations, resulting in significantly higher noise errors. This invention optimizes the graph by using co-occurrence intensity distance attenuation and cyclic weak edge pruning, resulting in a false association edge ratio of only 4.3%, which greatly reduces graph redundancy noise.

[0081] 3. Regarding the accuracy of job profile matching, the two existing solutions cannot aggregate core skill clusters based on the graph topology structure, but only achieve simple label listing, with a matching degree of less than 75% with the composite skill requirements of real jobs; the present invention filters core skills based on node degree, limits the number of jumps, performs breadth-first traversal of local subgraphs and vector average encoding, improving the profile matching accuracy to 91.5%, and can accurately distinguish the differentiated skill requirements of different jobs.

[0082] 4. Regarding the timeliness of the profile, compared to Solution 3, which lacks a closed-loop feedback mechanism for on-site operational data, the profile deviation quickly exceeds the standard as production line processes are fine-tuned and personnel operating habits change, with an effective update cycle of only 22 days. This invention uses the deviation of MES operation time to dynamically correct the correlation strength of the map, supports monthly full updates and weekly incremental fine-tuning, and extends the effective usage period of the profile to 30 days, solving the defect of static profiles becoming distorted quickly with production iterations.

[0083] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes 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.

Claims

1. A method for constructing manufacturing skills profiles based on knowledge graphs, characterized in that, The method includes: Obtain a set of skill description texts for the target manufacturing sector, including operating regulations texts, equipment manual texts, manufacturing-related professional qualification standards, and process flow diagram texts. The skill description text set is parsed to generate a skill meta-item set for the target manufacturing industry. The skill meta-item set includes operation object meta-items, action type meta-items, and environmental parameter meta-items. The corresponding application layer adaptation rule is determined based on the type attribute of each skill meta item, and the application layer adaptation rule includes the mapping relationship between meta item attribute values ​​and knowledge graph node types; Based on the application layer adaptation rules, a skill knowledge graph for the target manufacturing field is constructed. The skill knowledge graph includes a set of skill nodes and a set of skill association edges. Each association edge in the skill association edge set connects two skill nodes and is marked with an association strength value. The skill knowledge graph is subjected to feature aggregation processing to generate skill profiles for multiple job positions within the target manufacturing sector.

2. The method for constructing manufacturing skills profiles based on knowledge graphs according to claim 1, characterized in that, The acquisition of the skill description text set for the target manufacturing sector includes: Collect operational regulations texts within the target manufacturing sector, the operational regulations texts including process step descriptions and safety specification clauses; Collect equipment manual texts from the target manufacturing sector, the equipment manual texts including operation panel labels and parameter adjustment ranges; The operating rules text and the equipment manual text are divided into regions according to the page layout. The region division is based on the relationship between the text block spacing and the font size. After the page is segmented, text line alignment correction is performed to generate the skill description text set. The text line alignment correction is based on the projection distribution calculation of the page coordinates.

3. The method for constructing manufacturing skills profiles based on knowledge graphs according to claim 1, characterized in that, The step of parsing the skill description text set to generate the skill meta-item set for the target manufacturing sector includes: The skill description text set is split into entities according to paragraph boundaries to generate multiple skill description paragraphs. The entity splitting is based on the matching relationship between the title level identifier and the paragraph indentation value. For each of the skill description paragraphs, perform the following processing: Based on the dependency relationship between verb and noun parts of speech in the skill description paragraph, at least one operation object element and at least one action type element are extracted. The operation object element corresponds to the device component identifier, and the action type element corresponds to the operation action category. Based on the numerical units and dimensional identifiers in the skill description paragraph, at least one environmental parameter element is extracted, which corresponds to a temperature range value or a pressure range value. The operation object element, the action type element, and the environment parameter element are combined into a tuple to generate a skill element group.

4. The method for constructing manufacturing skills profiles based on knowledge graphs according to claim 1, characterized in that, The step of determining the corresponding application layer adaptation rule based on the type attribute of each skill element includes: Obtain the type attribute of the operation object meta item, the type attribute including physical attribute category label and functional attribute category label; A first mapping path is determined based on the physical attribute classification label, and the first mapping path maps the operation object meta item to the entity node type in the knowledge graph. A second mapping path is determined based on the functional attribute classification label, and the second mapping path maps the operation object meta item to the concept node type in the knowledge graph. The path weight values ​​of the first mapping path and the second mapping path are compared. The path weight values ​​are calculated based on the frequency of occurrence of the operation object meta-item in the skill description text. The mapping path with the higher path weight value is set as the application layer adaptation rule.

5. The method for constructing manufacturing skills profiles based on knowledge graphs according to claim 1, characterized in that, The construction of the skill knowledge graph for the target manufacturing sector based on the application layer adaptation rules includes: According to the application layer adaptation rules, at least one skill node is inserted into the pre-constructed initial knowledge graph, and the skill node includes a node identifier and a node attribute value. Based on the co-occurrence relationship between each skill element in the skill element set, at least one skill association edge is established, and the skill association edge includes a start node identifier, an end node identifier, and an association strength value. The skill-related edges are subjected to loop structure detection, which is implemented based on path marking in depth-first traversal. When multiple cycles are detected sharing the same skill association edge with the lowest association strength value, the edge is deleted globally. After deletion, a full-graph cycle detection is performed again until no cycles are found.

6. The method for constructing manufacturing skills profiles based on knowledge graphs according to claim 5, characterized in that, The step of establishing at least one skill association edge based on the co-occurrence relationship between each skill element in the skill element set includes: Obtain the number of times the first skill element and the second skill element in the skill description text set are co-occurring in the skill element set; The initial value of the association strength between the first skill meta-item and the second skill meta-item is determined based on the ratio of the co-occurrence frequency to the total number of paragraphs in the skill description text set; The initial value of the association strength is subjected to attenuation correction processing, which is calculated based on the positional distance between the first skill element and the second skill element in the text. The attenuation-corrected correlation strength value is assigned to the skill correlation edge.

7. The method for constructing manufacturing skills profiles based on knowledge graphs according to claim 1, characterized in that, The step of performing feature aggregation processing on the skill knowledge graph to generate skill profiles for multiple job positions within the target manufacturing sector includes: Obtain the node degree value of each skill node in the skill knowledge graph, where the node degree value represents the number of skill association edges connected to that skill node; The skill nodes are sorted in descending order according to their node degree values, and a set number of skill nodes with the highest node degree values ​​are selected as core skill nodes. Centered on each core skill node, extract a set of adjacent skill nodes within a specified number of jumps from the skill knowledge graph; The skill nodes in the set of adjacent skill nodes are vectorized with attribute values ​​to generate the skill profile vector corresponding to the core skill node.

8. The method for constructing manufacturing skills profiles based on knowledge graphs according to claim 7, characterized in that, The step of extracting a set of adjacent skill nodes within a specified number of jumps from the skill knowledge graph, centered on each core skill node, includes: Obtain the node identifier of the core skill node and the value of the specified number of jumps, wherein the specified number of jumps is determined based on the average path length of the skill knowledge graph; Starting from the core skill node, a breadth-first traversal is performed along the skill-related edges, and the breadth-first traversal records the jump count for each traversal; When the hop count reaches the specified hop count, the current traversal path is terminated; The identifiers of all skill nodes visited during the traversal are aggregated to generate the set of adjacent skill nodes.

9. The method for constructing manufacturing skills profiles based on knowledge graphs according to claim 1, characterized in that, The method further includes: Within a preset evaluation period, actual operation records of multiple job positions in the target manufacturing sector are collected. The actual operation records include operation time data and the frequency of operation errors. The actual operation records are compared with the skill profiles of the corresponding job positions to generate a first deviation value. The deviation analysis is based on the comparison of the distribution range of the operation time data. Adjust the association strength value of the skill association edges in the skill knowledge graph according to the first deviation value, and generate an updated set of skill association edges; The updated set of skill-related edges will be applied to the skill profile generation task in the next cycle.

10. The method for constructing manufacturing skills profiles based on knowledge graphs according to claim 9, characterized in that, The step of performing deviation analysis on the actual operation records and the corresponding job skill profiles to generate a first deviation value includes: Extract the operation time data from the actual operation record. The operation time data includes the start timestamp and end timestamp of each operation step. Extract the standard time interval of the corresponding operation steps from the skill profile, the standard time interval being calculated based on the statistical distribution of historical operation data; Calculate the frequency of the operation time data falling outside the standard time interval to obtain a first frequency statistic; The first deviation value is calculated based on the first frequency statistics and the total number of operations recorded in the actual operation records.

Images