Intelligent retrieval method and device of process knowledge, electronic equipment and storage medium

Abstract

This application provides an intelligent retrieval method, apparatus, electronic device, and storage medium for process knowledge. The method includes: converting user queries into query vectors; querying multiple similar knowledge units with the highest semantic similarity to the query vectors in a pre-constructed process knowledge base; extracting fabric entity information from the user queries; if fabric entity information is obtained, extracting multiple candidate knowledge units that match the fabric entity information from the multiple similar knowledge units; performing keyword matching between the user queries and the multiple candidate knowledge units to obtain the keyword matching degree corresponding to each candidate knowledge unit; and selecting target knowledge units from the multiple candidate knowledge units based on the keyword matching degree and semantic similarity corresponding to each candidate knowledge unit. This solution provides users with accurate and actionable parameter configuration suggestions, and the recall results are both semantically relevant and strictly limited to the target process scenario, avoiding recommendation bias caused by overgeneralization.

Description

Technical Field

[0001] This application relates to the field of intelligent manufacturing technology, and in particular to an intelligent retrieval method and device for process knowledge, electronic equipment, computer-readable storage medium, and computer program product. Background Technology

[0002] In the garment manufacturing industry, the correct parameter settings of sewing machines (such as needle type, thread tension, feed dog height, stitch length, and speed) are crucial to the quality of the finished product. Currently, the industry generally faces the following technical problems: The configuration of process parameters is highly dependent on manual experience: sewing parameter configuration mainly relies on the personal experience of experienced sewing masters. Newcomers need to undergo a long apprenticeship learning period before they can independently complete parameter settings, and the consistency of processes between different operators is poor, with a lack of standardized basis for parameter adjustments. When problems such as "wrinkling," "skipped stitches," and "broken needles" occur, operators often adjust parameters through trial and error, which is inefficient and easily introduces new problems.

[0003] Technological knowledge is scattered and lacks structured accumulation: sewing process parameters and related knowledge are scattered in equipment manuals, PDF documents, historical process sheets, and oral transmissions from experienced workers, lacking unified and structured knowledge base management.

[0004] General-purpose AI assistants are not suitable for specialized industrial scenarios: While existing large language models can answer general questions, they lack a systematic understanding of specialized parameters in the sewing industry and cannot provide accurate and actionable parameter suggestions. Traditional RAG (Retrieval Enhanced Generation) systems are mostly geared towards general text question answering and are not optimized for highly structured industrial parameters (such as numerical and enumerable fields), resulting in insufficient retrieval accuracy. Summary of the Invention

[0005] This application provides a method and apparatus for intelligent retrieval of process knowledge, electronic equipment, and storage medium to achieve high-precision and executable sewing parameter recommendations.

[0006] In a first aspect, embodiments of this application provide an intelligent retrieval method for process knowledge, including: Convert user questions into question vectors; In the established process knowledge base, query the multiple similar knowledge units with the highest semantic similarity to the question vector; Fabric entity information is extracted from the user's question content. If fabric entity information is obtained, multiple candidate knowledge units that match the fabric entity information are extracted from the multiple similar knowledge units. The user's question content is matched with the multiple candidate knowledge units by keywords to obtain the keyword matching degree corresponding to each candidate knowledge unit; The target knowledge unit is selected from the multiple candidate knowledge units based on the keyword matching degree and semantic similarity corresponding to each candidate knowledge unit.

[0007] In one embodiment, before querying multiple similar knowledge units with the highest semantic similarity to the query vector in the constructed process knowledge base, the method further includes: Construct a process knowledge base containing multiple knowledge units. Each knowledge unit includes structured process knowledge and process semantic vectors for specified product categories and subcategories of fabrics.

[0008] In one embodiment, the structured process knowledge includes text document fields; the construction of a process knowledge base containing multiple knowledge units includes: The text document fields are vectorized to generate the process semantic vector.

[0009] In one embodiment, the structured process knowledge further includes structured process parameters, and the construction of a process knowledge base containing multiple knowledge units includes: Based on predefined regular expression templates and domain rules, the input unstructured text is parsed in layers to obtain structured process parameters; The structured process parameters, process constraint rules, parameter coupling relationships, and equipment adaptability declarations are input into the large language model to generate process guidance content, which is then used as a field in the text document.

[0010] In one embodiment, the step of performing hierarchical parsing on the input unstructured text based on predefined regular expression templates and domain rules to obtain structured process parameters includes: Extract the compound keywords with colon prefixes from the unstructured text, and split the compound keywords into category fields and subcategory fields according to a preset delimiter; Thickness information is extracted from the unstructured text using named entity recognition rules, and the fabric pattern classification field is determined based on a preset thickness-pattern mapping table. The process parameters in the unstructured text are converted into structured key-value pairs, and the units are standardized.

[0011] In one embodiment, the step of performing hierarchical parsing on the input unstructured text based on predefined regular expression templates and domain rules to obtain structured process parameters further includes: If some fields are missing, the missing fields will be filled in according to the process constraint rules.

[0012] In one embodiment, if some fields are missing, filling in the missing fields according to process constraint rules includes: If the fabric pattern classification field is missing in the unstructured text, the corresponding fabric pattern classification field is determined based on the subclass field. If the needle plate model field is missing in the unstructured text, the needle plate model field is determined based on the fabric pattern classification field and the already determined needle model field.

[0013] In one embodiment, extracting fabric entity information from the user's question content includes: Extract category and subcategory identifiers from the user's question content; If the subclass identifier is missing, a preset designated identifier will be used as the subclass identifier; If the user's question does not contain a category identifier or a subcategory identifier, but contains synonyms of the category identifier and / or synonyms of the subcategory identifier, the category identifier and / or subcategory identifier contained in the user's question is determined according to the synonym mapping rules.

[0014] In one embodiment, before converting the user's question content into a question vector, the method further includes: The user's question content is classified into intent categories to obtain intent classification results; If the intent classification result indicates the query type intent, a knowledge retrieval task is triggered.

[0015] If the intent classification result indicates a fault diagnosis intent, a fault diagnosis task is triggered.

[0016] In one embodiment, before converting the user's question content into a question vector, the method further includes: A sliding window mechanism is used to extract the content of the most recent multiple rounds of dialogue; When the length of the multi-turn dialogue content exceeds a threshold, the multi-turn dialogue content is compressed into a summary.

[0017] In one embodiment, after selecting the target knowledge unit from the plurality of candidate knowledge units, the method further includes: The retrieved target knowledge units are output in the formatted way according to the fields of the sewing machine control panel.

[0018] Secondly, embodiments of this application also provide an intelligent retrieval device for process knowledge, comprising: The vector conversion module is used to convert user questions into question vectors; The unit query module is used to query multiple similar knowledge units with the highest semantic similarity to the question vector in the constructed process knowledge base; The unit matching module is used to extract fabric entity information from the user's question content. If the fabric entity information is obtained, multiple candidate knowledge units that match the fabric entity information are extracted from the multiple similar knowledge units. The matching degree calculation module is used to perform keyword matching between the user's question content and the multiple candidate knowledge units to obtain the keyword matching degree corresponding to each candidate knowledge unit. The target filtering module is used to filter target knowledge units from the multiple candidate knowledge units based on the keyword matching degree and semantic similarity corresponding to each candidate knowledge unit.

[0019] Thirdly, embodiments of this application also provide an electronic device, including: Memory is used to store executable instructions for a computer; The processor, when executing computer-executable instructions stored in the memory, implements the intelligent retrieval method for the aforementioned process knowledge.

[0020] Fourthly, embodiments of this application also provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the aforementioned intelligent retrieval method for process knowledge.

[0021] Fifthly, embodiments of this application also provide a computer program product, which includes computer-executable instructions stored in a computer-readable storage medium; When the processor of the electronic device reads the computer-executable instructions from the computer-readable storage medium and executes the computer-executable instructions, the above-mentioned intelligent retrieval method for process knowledge is implemented.

[0022] The technical solution provided in this application, through a pre-built process knowledge base, queries the target knowledge units corresponding to user queries, enabling new employees to obtain accurate and actionable parameter configuration suggestions without long-term mentorship, significantly shortening the training cycle and reducing operational error rates. During the query process for target knowledge units, semantic matching is performed first, followed by fabric entity information filtering, and then a fusion screening of keyword matching and semantic matching. This ensures that the retrieved results are both semantically relevant and strictly limited to the target process scenario, avoiding the recommendation bias caused by overgeneralization in traditional RAG systems. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating an intelligent retrieval method for process knowledge provided in an embodiment of this application; Figure 2 This is a schematic diagram of the dialog window for fault diagnosis provided in an embodiment of this application; Figure 3 This is a schematic diagram of the parameter query dialog window provided in the embodiments of this application; Figure 4 This is a schematic diagram of a window formatted according to the fields of the sewing machine control panel provided in this embodiment of the application; Figure 5 This is a block diagram of an intelligent retrieval device for process knowledge provided in an embodiment of this application; Figure 6 This is a block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0024] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this specification as detailed in the appended claims.

[0025] The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of this specification. The singular forms “a,” “the,” and “the” as used in this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0026] It should be understood that although the terms first, second, third, etc., may be used in this specification to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this specification, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0027] Figure 1 This is a flowchart illustrating an intelligent retrieval method for process knowledge provided in an embodiment of this application. For example... Figure 1 As shown, the method includes: Step S110: Convert the user's question into a question vector.

[0028] In one embodiment, before step S110 above, the user's question content can be classified into intents to obtain intent classification results; if the intent classification results indicate a parameter query intent, a knowledge retrieval task is triggered. If the intent classification results indicate a fault diagnosis intent, a fault diagnosis task is triggered.

[0029] For example, a lightweight classifier (or LLM prompt word engineering) can be used to determine the question type. For instance, if a user's question contains words like "parameters," "how to set," or "recommendation," it's considered a parameter query intent. If the question contains words like "wrinkling," "skipped stitch," or "broken thread," it's considered a fault diagnosis task. The knowledge retrieval task queries the process knowledge base to return parameter setting strategies, while the fault diagnosis task queries the preset debugging strategy library to return the corresponding debugging strategy for the fault. Specifically, differentiated suggestions can be returned based on fabric type (A / M / H). Figure 2 This is a diagram of the fault diagnosis dialog window, such as... Figure 2 As shown, based on the fabric parameters in the current session, targeted adjustments are given (e.g., "Currently using shirt - standard parameters, it is recommended to loosen the top thread tension first").

[0030] In one embodiment, a sliding window mechanism can be used to retain only the most recent N rounds of dialogue; when the history is too long (i.e. when the length of the most recent N rounds of dialogue content is greater than a threshold), the most recent N rounds of dialogue content is compressed into a summary (such as "The user is inquiring about the parameters of the fleece lining of the hoodie") to avoid token overflow.

[0031] Specifically, user questions can be encoded into question vectors using LLM.

[0032] Step S120: In the constructed process knowledge base, query multiple similar knowledge units with the highest semantic similarity to the question vector.

[0033] Specifically, sewing parameter knowledge can be abstracted into a standardized JSON Schema (database schema), with each record containing: { "id": "Category_Subcategory", / / e.g., "Sweatshirt_Fleece-lined" "category": "main category", / / e.g., "hoodies" "sub_category": "subcategory identifier", / / e.g., "fleece-lined" "params": { "rotation_speed": Integer, / / Rotational speed (rpm) "stitch_length": Floating-point number, / / Stitch pitch (mm) "tension_param": Floating-point number, / / Surface tension (%) "feed_dog_height": Floating-point number, / / Feed dog height (mm) "timing": Integer, / / Timing (degrees) "mode": String, / / Fabric pattern classification (A / M / H) "classification": string / / Human-readable category description }, "recommend": { "throat_plate": String, / / Recommended throat plate type "feed_dog": string, / / Recommended feeder tooth type "needle_spec": string / / Recommended needle specification }, "text": "Natural Language Description" / / Used for vectorization } The process knowledge base can contain multiple records similar to those described above (structured parameters) to guide the sewing parameters for different categories and subcategories of garments. In one embodiment, a Large Language Model (LLM) can be used to generate a dense vector (dimension ≥ 1024) for the text field of each record; simultaneously, the original structured parameters are preserved, forming a "vector + structure" bimodal storage, thus constructing a Milvus vector library (i.e., the process knowledge base). The construction of the process knowledge base will be described in detail below and will not be repeated here.

[0034] Each record in the process knowledge base can be called a knowledge unit. Each knowledge unit includes a vector and the structured parameters mentioned above. The structured parameters include information such as id (VarChar variable-length string, primary key), category, subcategory, sewing parameters, and text field.

[0035] Specifically, by calculating the semantic similarity between the question vector and the vector of each knowledge unit, the top K knowledge units with the highest semantic similarity can be found in the process database. These are called similar knowledge units for differentiation.

[0036] Step S130: Extract fabric entity information from the user's question content. If fabric entity information is obtained, extract multiple candidate knowledge units that match the fabric entity information from the multiple similar knowledge units.

[0037] In one embodiment, the fabric entity information may include a category identifier (such as "hoodie") and a subcategory identifier (such as "fleece-lined"). In another embodiment, the fabric entity information may consist only of a category identifier or a subcategory identifier.

[0038] In one embodiment, step S130 may specifically include: extracting category identifiers and subcategory identifiers from the user's question content; if the subcategory identifier is missing, using a preset designated identifier as the subcategory identifier. For example, if a subcategory is missing (e.g., only "hoodie"), "regular" is used as the subcategory identifier by default. If the user's question content does not contain category identifiers and subcategory identifiers, but contains synonyms of the category identifier and / or synonyms of the subcategory identifier, the category identifier and / or subcategory identifier contained in the user's question content are determined according to the synonym mapping rules.

[0039] For example, if a user's question contains a specific category (such as "shirt"), then the knowledge units corresponding to the category can be further filtered from multiple similar knowledge units. These are called candidate knowledge units for differentiation.

[0040] In another embodiment, the fabric entity information may further include fabric composition information. For example, if a user asks "How do I adjust the parameters of a 60% cotton, 40% polyester sweatshirt?", the system uses preset fabric composition identification rules (such as regular expressions like "(\d+)%cotton.*(\d+)%polyester" or "(\d+)%cotton") to extract "60% cotton content" and "40% polyester content" as fabric composition information. Then, based on a mapping table between fabric composition and fabric pattern (e.g., cotton content ≥50% and polyester content ≥30% mapped to category M, cotton content ≥80% mapped to category A, and polyester content ≥70% mapped to category H), the extracted fabric composition information is converted into a fabric pattern classification field (e.g., "M"). This fabric pattern classification field is then used for filtering from similar knowledge units. By introducing the extraction and mapping of fabric composition information, the accuracy of parameter recommendations can be further improved, especially for blended fabric processing scenarios.

[0041] In another embodiment, the fabric entity information may further include fabric weight information. For example, if a user asks "How do I adjust the weight of a 320g hoodie?", the weight can be determined using a preset weight recognition rule (such as the regular expression "(\d+)\sg"). "or "(\d+)\s Extract the weight value "320" from the "g"); then, according to the mapping table between weight and fabric pattern (e.g., weight ≥ 300g / m),... 2 Mapped to class H, weight ≥ 200g / m³ 2 And <300g / m 2 Mapped to class M, weight <200g / m³ 2The extracted weight information is mapped to class A and converted into a fabric pattern classification field (such as "H"). Then, based on this fabric pattern classification field, similar knowledge units are selected. By introducing the extraction and mapping of fabric weight information, the process parameters for fabrics of different thicknesses can be matched more accurately, avoiding deviations caused by subjective judgments of fabric thickness.

[0042] Step S140: Match the user's question with the multiple candidate knowledge units using keywords to obtain the keyword matching degree for each candidate knowledge unit.

[0043] Keyword matching degree can be a weighted sum of one or more of the following four dimensions.

[0044] (i) Category Level Matching Degree: This calculates the degree of match between the category identifier extracted from the user's query and the category field of the candidate knowledge unit. If there is a perfect match, the score for this dimension is 1.0; if it is a synonym match (e.g., "pullover" and "hoodie"), the score is 0.8; if it is a hierarchical match (e.g., "top" and "hoodie"), the score is 0.6; if there is no match, the score is 0. The formula for calculating the category level matching degree is: Score_category = match_level(category_query, category_knowledge), where the return value of the match_level function is determined according to the matching rules described above.

[0045] (ii) Numerical Range Overlap: For the parameter range implicit in the user's question (e.g., "rotation speed is approximately 5000"), calculate the degree of overlap between this parameter range and the corresponding parameter value in the candidate knowledge unit (e.g., rotation_speed: 4800). Specifically, first, extract the numerical range [a, b] from the user's question (if there is only one value x, the range is considered to be [x-Δ, x+Δ], where Δ is a preset tolerance threshold, such as Δ=200); then calculate the overlap between this range and the parameter value p in the candidate knowledge unit: Overlap = max(0, min(b, p+δ) - max(a, p-δ)) / (b - a), where δ is a preset parameter tolerance (e.g., δ=100); finally, map the overlap to a score in the [0,1] interval. If the user's question does not implicitly contain any parameter range, the score for this dimension is 0.5 (neutral value).

[0046] (III) Synonym / Near-synonym Matching Degree: Calculate the synonym / near-synonym matching between keywords in the user's query and the text and structured fields of the candidate knowledge unit. Specifically, a thesaurus of process domains is pre-built (e.g., "rotation speed" and "speed", "stitch pitch" and "line length", "tension" and "tightness", etc.). When a keyword in the user's query matches a synonym of a field in the candidate knowledge unit, the keyword's matching score is 80% of the original word matching score (i.e., Score_synonym = 0.8 × Score_exact). Score_exact refers to the original word matching score, which can be 1, indicating that the keyword in the user's query is the same as a field in the candidate knowledge unit. Specifically, an exact match has a score of 1.0; a synonym / near-synonym match has a score of 0.8; and a no match has a score of 0.

[0047] The formula for calculating the matching degree of synonyms / near-synonyms is: Score_synonym_total = Σ (weight of keyword i × matching score of keyword i) / (weight of all keywords), where the weight of each keyword is pre-set according to its importance in the process field (e.g., the weight of core parameters such as "speed" and "needle pitch" is 2.0, and the weight of auxiliary words such as "equipment" and "model" is 1.0).

[0048] (iv) Entity Density Matching Degree: This calculates the proportion of entities extracted from the user's query content to the total number of entities in the candidate knowledge unit. Specifically, firstly, all process-related entities (including categories, subcategories, parameter names, values, equipment models, etc.) are extracted from the user's query content to obtain the entity set E_query; then, all entities are extracted from the structured fields of the candidate knowledge unit to obtain the entity set E_knowledge; the formula for calculating the entity density matching degree is: Score_density = |E_query ∩ E_knowledge| / |E_knowledge|, where |·| represents the number of elements in the set. The entity density matching degree reflects the consistency between the user's query content and the knowledge unit in terms of information richness; a higher density indicates a better match between the information coverage of the query content and the knowledge unit.

[0049] Finally, the comprehensive formula for calculating keyword matching degree is: Score_keyword = w1 × Score_category + w2 × Score_interval + w3 × Score_synonym_total + w4 × Score_density Where w1, w2, w3, and w4 are the weight coefficients for each dimension, and w1 + w2 + w3 + w4 = 1.0. In one embodiment, the default values ​​for each weight coefficient can be set as follows: w1 = 0.35, w2 = 0.25, w3 = 0.20, and w4 = 0.20. In another embodiment, the weight coefficients can be dynamically adjusted according to the type of user question content. For example, when the user question content contains explicit numerical parameters, the value of w2 can be appropriately increased; when the user question content contains many synonyms, the value of w3 can be appropriately increased. The dynamic adjustment of the weight coefficients can be achieved through a preset rule engine or through a lightweight classifier to predict based on the features of the question content.

[0050] Step S150: Select the target knowledge unit from the multiple candidate knowledge units based on the keyword matching degree and semantic similarity corresponding to each candidate knowledge unit.

[0051] In one embodiment, the Reciprocal Rank Fusion (RRF) algorithm can be used to fuse and rank the semantic similarity ranking and the keyword matching ranking. Specifically, for each candidate knowledge unit i, its ranking in the semantic similarity ranking is R_semantic(i), and its ranking in the keyword matching ranking is R_keyword(i). Then, its RRF composite score is: Score_RRF(i) = 1 / (k + R_semantic(i)) + 1 / (k + R_keyword(i)) Where k is a smoothing constant used to prevent candidate units with excessively high rankings from receiving excessively large score differences; in one embodiment, k is set to 60. The core idea of ​​the RRF algorithm is that candidate knowledge units that perform well in both high semantic similarity ranking and high keyword matching ranking will receive the highest comprehensive score. By integrating the ranking using the RRF algorithm, the algorithm fully utilizes both the ability of semantic similarity to understand vague user expressions and the role of keyword matching in defining precise process scenarios.

[0052] In another embodiment, a weighted summation method can also be used to calculate the comprehensive score. Specifically, the semantic similarity and keyword matching degree are first normalized, mapping their numerical ranges to the [0,1] interval. For semantic similarity, the cosine similarity value (already located in the [0,1] interval) can be directly used; for keyword matching degree, since its value range may vary depending on the calculation method, a minimum-maximum normalization method can be used: Score_keyword_norm(i) = (Score_keyword(i) - min(Score_keyword)) / (max(Score_keyword) - min(Score_keyword)), where min(Score_keyword) and max(Score_keyword) are the minimum and maximum values ​​of keyword matching degree among all candidate knowledge units, respectively. Then, the comprehensive score is calculated: Score_total(i) = α × Score_semantic(i) + β × Score_keyword_norm(i) Here, α and β are weighting coefficients, and α + β = 1.0. In one embodiment, α = 0.5, β = 0.5, meaning semantic similarity and keyword matching are equally important. In another embodiment, the values ​​of α and β can be dynamically adjusted according to the application scenario: when the user's question is relatively vague (e.g., "how to adjust thick fabric"), the value of α is appropriately increased (e.g., α = 0.7, β = 0.3) to emphasize semantic understanding; when the user's question is relatively precise (e.g., "fleece-lined hoodie, 4800 RPM, 3.5 stitch length"), the value of β is appropriately increased (e.g., α = 0.3, β = 0.7) to emphasize precise keyword matching. The dynamic adjustment of the weighting coefficients can be achieved through preset rules, such as judging the precision of the question based on the number of entities extracted from the user's question: the more entities, the higher the value of β.

[0053] The target knowledge unit can be the knowledge unit with the highest comprehensive score in keyword matching and semantic similarity among multiple candidate knowledge units.

[0054] The above search strategy supports both fuzzy semantics ("How to adjust elastic fabric" → matching "Category A") and ensures category accuracy ("Suit slips" → only returns the suit subcategory). For example... Figure 3 As shown, regarding the user question "How to set the parameters for a thin shirt?", as follows: Figure 4As shown, the retrieved target knowledge units can be output in a formatted manner according to the fields on the sewing machine control panel. This allows users to quickly obtain the sewing parameters for thin shirts without relying on random settings based on experience. Furthermore, the target knowledge units can be generated as QR codes, which the sewing machine can scan to import the information.

[0055] The following describes the process of building the process knowledge base.

[0056] Prior to step S120, the method provided in this application embodiment further includes: constructing a process knowledge base containing multiple knowledge units, each knowledge unit including structured process knowledge of a specified category and a specified subcategory of fabric, as well as a process semantic vector.

[0057] The `category` field represents the product category, and the `sub_category` field represents the subcategory. Structured process knowledge, as shown above, includes the `id`, `category`, `sub_category`, `params` (process parameter settings), `recommend` (hardware selection recommendations), and `text` (text document) fields. Process semantic vectors can be generated by vectorizing the text document fields. For distinction, all fields except the `text` field can be referred to as structured process parameters.

[0058] Specifically, constructing a process knowledge base containing multiple knowledge units may include the following steps: performing hierarchical parsing on the input unstructured text based on predefined regular expression templates and domain rules to obtain structured process parameters; inputting the structured process parameters into a large language model to generate process guidance content, and using the process guidance content as the text document field.

[0059] Regular expression templates are text matching patterns based on regular expressions, used to accurately extract information with specific formats from unstructured text. Domain rules are logical judgment rules based on sewing technology expertise, used to interpret semantics, infer missing values, and verify parameter validity.

[0060] Unstructured text refers to unstructured process records. Unstructured text undergoes hierarchical parsing to obtain structured process parameters. These parameters can include category fields, subcategory fields, params fields, and recommend fields stored as key-value pairs. The params field may contain parameters such as rotation speed, stitch length, thread tension, feed dog height, timing, fabric pattern classification, and human-readable classification descriptions. The recommend field may contain recommended needle plate type, recommended feed dog type, and recommended needle specifications.

[0061] Layered parsing can include: category-subcategory extraction, parameter slot filling, and process parameter standardization. Category-subcategory extraction: Extract compound keywords with colon prefixes (e.g., "fleece-lined hoodie") from the unstructured text, and split these keywords into a category field (category, hoodie) and a subcategory field (sub_category, fleece-lined) based on preset separators (e.g., spaces / underscores). Parameter slot filling: Extract thickness information (e.g., "extra-thick fabric") from the unstructured text using named entity recognition rules, and determine the fabric pattern classification field (e.g., "H") based on a preset thickness-pattern mapping table (extra-thick → H, medium-thick → M, thin → L). Process parameter standardization: Convert process parameters in the unstructured text into structured key-value pairs and standardize the units. For example, convert expressions like "needle: 14#" into structured key-value pairs (needle_type: "14#"), and standardize the units to international standards (e.g., feed dog height is forcibly converted to mm).

[0062] If some fields are missing, they can be filled in according to the process constraint rules. Specifically, if the unstructured text is missing a fabric pattern classification field, the corresponding fabric pattern classification field is determined based on the sub-class field (e.g., "fleece-lined" → H category, "thin" → L category); if the unstructured text is missing a needle plate model field, the needle plate model field is determined based on the fabric pattern classification field and the determined needle model field (e.g., H category + 14# needle → xx machine 2.6 needle plate).

[0063] A globally unique ID (e.g., down jacket_thick) is generated using the composite naming rule {category}_{sub_category}, which serves as the index key for the knowledge base, ensuring that parameter configurations are strictly aligned with the process scenario during subsequent recalls.

[0064] The `text` field of the structured process parameters is used to record process guidance content. Specifically, structured process parameters, process constraint rules, parameter coupling relationships, and equipment adaptability declarations can be input into the Large Language Model (LLM) to generate process guidance content, which is then used as the `text` field. Specifically, a context template (Prompt) is first injected, embedding process constraint rules (e.g., "For extra-thick fabrics, increase the feed dog height to avoid skipped stitches") within the Prompt. The Prompt template can also include role definitions (e.g., "You are a garment sewing process expert with 20 years of experience"), structured process parameters, and task requirements (please output the process guidance content). Multi-parameter coupling relationships and equipment adaptability declarations can also be added to the Prompt. Multi-parameter coupling relationships guide the physical relationships between parameters in the LLM (e.g., stitch length ↑ corresponds to thread tension ↓, speed ↑ corresponds to timing ↓); equipment adaptability declarations explicitly associate recommended components (e.g., "xx machine 21-tooth fine tooth") with process objectives ("reduce thin material slippage").

[0065] The method provided in the above embodiments of this application: 1) Effectively solves the problem that sewing parameter configuration is highly dependent on human experience. By constructing a structured fabric-parameter knowledge graph, the implicit process knowledge scattered in manuals, PDFs, or oral instruction is transformed into standardized, machine-readable JSON schema data. This allows new employees to obtain accurate and actionable parameter configuration suggestions without long-term apprenticeship, significantly shortening the training cycle and reducing the rate of operational errors.

[0066] 2) Significantly improved the accuracy and scenario adaptability of parameter retrieval and recommendation. Innovatively, a dual-modal hybrid retrieval mechanism of "vector semantics + structured metadata" is adopted: on the one hand, a large language model is used to understand the user's fuzzy expression (such as "how to adjust a thick sweatshirt"), and on the other hand, hard filtering is performed through the category (category / sub_category) field, and combined with the RRF algorithm (Reciprocal Rank Fusion) to merge and sort, so as to ensure that the recall results are both semantically relevant and strictly limited to the target process scenario, avoiding the recommendation bias caused by overgeneralization in traditional RAG systems.

[0067] 3) It achieves highly operable and context-aware intelligent fault diagnosis. For typical faults such as "wrinkling" and "skipped stitches", a layered debugging rule library based on fabric mode (A / M / H) is pre-built, and special strategies for elastic fabrics are supported for activation. At the same time, combined with the fabric type and equipment parameters confirmed in the user's current session (such as "shirt - regular, 11# needle"), priority-based adjustment suggestions are dynamically generated (such as "it is recommended to loosen the top thread tension first"), so as to achieve a precise mapping from "problem description" to "specific operation instructions", which greatly improves the efficiency of troubleshooting.

[0068] 4) It has established a closed loop for industrial implementation, connecting "knowledge → decision-making → equipment execution". The parameter recommendations are strictly organized according to the fields on the sewing machine control panel (such as speed, stitch length, tension, etc.), and support the generation of QR codes for one-click import by the equipment. Fault suggestions are presented in a point-by-point natural language, retaining specific values ​​(such as "top thread tension reduced by 5%–10%)" and the names of special parts (such as "synchronous presser foot" and "2.0 needle plate"), to ensure that frontline workers can "understand, adjust accurately, and use effectively".

[0069] 5) Ensures reliability and security in industrial application scenarios. All responses are generated based on a predefined rule base, structured knowledge units, and process constraint logic, without relying on large language models to make their own decisions. This effectively avoids the risks of "illusions" or unreasonable suggestions common in general AI assistants, and meets the stringent requirements of garment manufacturing for production stability and process consistency.

[0070] 6) Enabled automated knowledge extraction and continuous evolution of unstructured process documents. Through multi-granularity rule-based parsing, missing value reasoning, and LLM-assisted semantic expansion, historical process records can be efficiently and automatically transformed into high-quality knowledge units, supporting subsequent manual verification and iterative updates, thus laying the foundation for building enterprise-level digital assets for sewing processes.

[0071] The above are merely preferred embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

[0072] Figure 5 This is a block diagram of an intelligent retrieval device for process knowledge provided in an embodiment of this application. Figure 5 As shown, the device includes: a vector conversion module 610, a cell query module 620, a cell matching module 630, a matching degree calculation module 640, and a target filtering module 650.

[0073] Vector conversion module 610 is used to convert user questions into question vectors; The unit query module 620 is used to query multiple similar knowledge units with the highest semantic similarity to the question vector in the constructed process knowledge base; The unit matching module 630 is used to extract fabric entity information from the user's question content. If the fabric entity information is obtained, multiple candidate knowledge units that match the fabric entity information are extracted from the multiple similar knowledge units. The matching degree calculation module 640 is used to perform keyword matching between the user's question content and the multiple candidate knowledge units to obtain the keyword matching degree corresponding to each candidate knowledge unit. The target filtering module 650 is used to filter target knowledge units from the multiple candidate knowledge units based on the keyword matching degree and semantic similarity corresponding to each candidate knowledge unit.

[0074] The specific implementation process of the functions and roles of each module in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0075] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. Each method step is implemented by a corresponding module, which will not be described in detail here. The device embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules, that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of the solution in this specification according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0076] Figure 6 This specification illustrates a hardware structure diagram of an electronic device containing a process knowledge intelligent retrieval device according to an exemplary embodiment, such as... Figure 6 As shown, the device may include: a processor 301, a memory 302, an input / output interface 303, a communication interface 304, and a bus 305. The processor 301, memory 302, input / output interface 303, and communication interface 304 are interconnected within the device via the bus 305.

[0077] The processor 301 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the steps of the intelligent retrieval method for process knowledge provided in the embodiments of this specification.

[0078] The memory 302 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 302 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 302 and is called and executed by the processor 301.

[0079] Input / output interface 303 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components in the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touch screens, microphones, various sensors, etc., and output devices may include displays, speakers, vibrators, indicator lights, etc.

[0080] Communication interface 304 is used to connect a communication module (not shown in the figure) to enable communication and interaction between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0081] Bus 305 includes a pathway for transmitting information between various components of the device (e.g., processor 301, memory 302, input / output interface 303, and communication interface 304).

[0082] It should be noted that although the above-described device only shows the processor 301, memory 302, input / output interface 303, communication interface 304, and bus 305, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0083] This specification also provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the intelligent retrieval method for process knowledge as provided in this specification.

[0084] This specification also provides a computer program product, which includes computer-executable instructions stored in a computer-readable storage medium; wherein, when the processor of an electronic device reads the computer-executable instructions from the computer-readable storage medium and executes the computer-executable instructions, it implements the steps of the intelligent retrieval method for process knowledge described in the above embodiments of this application.

[0085] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0086] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0087] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

Claims

1. An intelligent retrieval method for process knowledge, characterized in that, include: Convert user questions into question vectors; In the established process knowledge base, query the multiple similar knowledge units with the highest semantic similarity to the question vector; Fabric entity information is extracted from the user's question content. If fabric entity information is obtained, multiple candidate knowledge units that match the fabric entity information are extracted from the multiple similar knowledge units. The user's question content is matched with the multiple candidate knowledge units by keywords to obtain the keyword matching degree corresponding to each candidate knowledge unit; The target knowledge unit is selected from the multiple candidate knowledge units based on the keyword matching degree and semantic similarity corresponding to each candidate knowledge unit.

2. The method according to claim 1, characterized in that, Before querying the multiple similar knowledge units with the highest semantic similarity to the question vector in the constructed process knowledge base, the method further includes: Construct a process knowledge base containing multiple knowledge units. Each knowledge unit includes structured process knowledge and process semantic vectors for specified product categories and subcategories of fabrics.

3. The method according to claim 2, characterized in that, The structured process knowledge includes text document fields; The construction of a process knowledge base comprising multiple knowledge units includes: The text document fields are vectorized to generate the process semantic vector.

4. The method according to claim 3, characterized in that, The structured process knowledge also includes structured process parameters, and the construction of a process knowledge base containing multiple knowledge units includes: Based on predefined regular expression templates and domain rules, the input unstructured text is parsed in layers to obtain structured process parameters; The structured process parameters, process constraint rules, parameter coupling relationships, and equipment adaptability declarations are input into the large language model to generate process guidance content, which is then used as a field in the text document.

5. The method according to claim 4, characterized in that, The method of performing hierarchical parsing on the input unstructured text based on predefined regular expression templates and domain rules to obtain structured process parameters includes: Extract the compound keywords with colon prefixes from the unstructured text, and split the compound keywords into category fields and subcategory fields according to a preset delimiter; Thickness information is extracted from the unstructured text using named entity recognition rules, and the fabric pattern classification field is determined based on a preset thickness-pattern mapping table. The process parameters in the unstructured text are converted into structured key-value pairs, and the units are standardized.

6. The method according to claim 5, characterized in that, The method of performing hierarchical parsing on the input unstructured text based on predefined regular expression templates and domain rules to obtain structured process parameters also includes: If some fields are missing, the missing fields will be filled in according to the process constraint rules.

7. The method according to claim 6, characterized in that, If some fields are missing, the missing fields will be filled in according to the process constraint rules, including: If the fabric pattern classification field is missing in the unstructured text, the corresponding fabric pattern classification field is determined based on the subclass field. If the needle plate model field is missing in the unstructured text, the needle plate model field is determined based on the fabric pattern classification field and the already determined needle model field.

8. The method according to claim 1, characterized in that, Extracting fabric entity information from the user's question content includes: Extract category and subcategory identifiers from the user's question content; If the subclass identifier is missing, a preset designated identifier will be used as the subclass identifier; If the user's question does not contain a category identifier or a subcategory identifier, but contains synonyms of the category identifier and / or synonyms of the subcategory identifier, the category identifier and / or subcategory identifier contained in the user's question is determined according to the synonym mapping rules.

9. The method according to claim 1, characterized in that, Before converting the user's question content into a question vector, the method further includes: The user's question content is classified into intent categories to obtain intent classification results; If the intent classification result indicates the query class intent of the parameter, a knowledge retrieval task is triggered. If the intent classification result indicates a fault diagnosis intent, a fault diagnosis task is triggered.

10. The method according to claim 1, characterized in that, Before converting the user's question content into a question vector, the method further includes: A sliding window mechanism is used to extract the content of the most recent multiple rounds of dialogue; When the length of the multi-turn dialogue content exceeds a threshold, the multi-turn dialogue content is compressed into a summary.

11. The method according to claim 1, characterized in that, After selecting the target knowledge unit from the plurality of candidate knowledge units, the method further includes: The retrieved target knowledge units are output in the formatted way according to the fields of the sewing machine control panel.

12. An intelligent retrieval device for process knowledge, characterized in that, include: The vector conversion module is used to convert user questions into question vectors; The unit query module is used to query multiple similar knowledge units with the highest semantic similarity to the question vector in the constructed process knowledge base; The unit matching module is used to extract fabric entity information from the user's question content. If the fabric entity information is obtained, multiple candidate knowledge units that match the fabric entity information are extracted from the multiple similar knowledge units. The matching degree calculation module is used to perform keyword matching between the user's question content and the multiple candidate knowledge units to obtain the keyword matching degree corresponding to each candidate knowledge unit. The target filtering module is used to filter target knowledge units from the multiple candidate knowledge units based on the keyword matching degree and semantic similarity corresponding to each candidate knowledge unit.

13. An electronic device, characterized in that, include: Memory is used to store executable instructions for a computer; A processor, when executing computer-executable instructions stored in the memory, implements the intelligent retrieval method for process knowledge as described in any one of claims 1 to 11.

14. A computer-readable storage medium, characterized in that, The device stores computer-executable instructions, which, when executed by a processor, implement the intelligent retrieval method for process knowledge as described in any one of claims 1 to 11.

15. A computer program product, characterized in that, The computer program product includes computer-executable instructions stored in a computer-readable storage medium; When the processor of the electronic device reads the computer-executable instructions from the computer-readable storage medium and executes the computer-executable instructions, it implements the intelligent retrieval method of process knowledge as described in any one of claims 1 to 11.

Images