A method for constructing a performance influence factor conduction path for an intelligent production manufacturing system solution

Abstract

This invention relates to the field of intelligent manufacturing system technology, and particularly to a method for constructing a transmission path for the effectiveness of intelligent manufacturing system solutions. This method takes unstructured text such as various reports and operation logs as input, and through seven steps—word segmentation preprocessing, dual-channel effectiveness indicator extraction, dual-channel influencing factor extraction, automatic lifecycle stage mapping, business process node binding, construction of influencing factor transmission paths, and output of a structured effectiveness evidence chain—achieves automated transformation from raw text to a structured effectiveness evidence chain. This invention establishes an automated, traceable, and auditable mapping mechanism from abstract indicators in standard specifications to actual operational evidence of the manufacturing system, supporting a new paradigm for intelligent manufacturing effectiveness evaluation that moves from experience-based judgment to data-driven and standard-compliant approaches.

Description

Technical Field

[0001] This invention relates to the field of intelligent production system technology, and in particular to a method for constructing the transmission path of factors influencing the effectiveness of intelligent production and manufacturing system solutions. Background Technology

[0002] Currently, intelligent manufacturing system solutions are widely used in the industrial sector, but their implementation effectiveness evaluation still faces fundamental challenges: the influencing factors are unclear, and there is a lack of traceable mapping between indicators and the manufacturing process. Enterprises often equate "system launch" with "value realization," but fail to answer core questions such as why indicators change and which processes truly drive effectiveness. Particularly in production systems, the inability to accurately pinpoint the influencing factors in core business processes such as MRP (Material Requirements Planning) makes it difficult to verify the effectiveness of investments, renders project acceptance a mere formality, and lacks a basis for continuous improvement.

[0003] Chinese invention patent CN121328569B discloses a method and system for intelligent analysis of unstructured text data based on large models. Although it is also based on unstructured text data analysis, it only focuses on fault chain extraction at the text level and does not involve integration with business processes.

[0004] At its root, the existing assessment methods suffer from two major disconnects: First, there is a disconnect between the evaluation indicators and the operational logic of smart factories. Although national standards define enterprise performance evaluation indicators, these indicators are mostly static and isolated KPIs, not embedded in business process models and equipment operation mechanisms. For example, "improved order delivery on-time rate" may stem from "improved MPS planning accuracy," which in turn is affected by "MRP material availability rate" and "BOM process route optimization." Existing methods for manufacturing systems cannot automatically construct such cross-stage causal chains that span the main value chain from procurement to production to sales, resulting in evaluation results that are detached from the actual operational logic of the factory, reflecting only surface figures rather than the underlying mechanisms.

[0005] Second, there is a disconnect from unstructured evidence sources. The key factors that truly drive effectiveness are often hidden in unstructured texts such as acceptance reports, operation and maintenance logs, and expert review opinions. Traditional manual analysis relies heavily on expert experience, which has three major bottlenecks: 1) High subjectivity: for the same indicator, expert A and expert B may have different attributions, leading to inconsistent conclusions; 2) Lack of scalability: manual analysis reports, operation and maintenance logs, and expert review opinions take a long time, and the manpower required to support group-level assessments or large-scale policy applications is enormous; 3) Lack of auditability: conclusions are prone to lacking a structured chain of evidence, which may not meet the rigid requirements of "traceability and reproducibility" in compliance scenarios such as third-party evaluation and international mutual recognition. Summary of the Invention

[0006] This invention discloses a method for constructing the transmission path of factors influencing the effectiveness of intelligent manufacturing system solutions. It can establish an automated, traceable, and auditable mapping mechanism from abstract indicators in standard specifications to evidence of actual operation of the manufacturing system.

[0007] It is achieved through the following technical solution: Acquire unstructured text from intelligent manufacturing systems; Semantic segmentation of unstructured text is performed using an adaptive multi-scale perceptual word segmenter; Combining domain dictionary precise matching with BERT semantic classification, dual-channel recognition of performance indicators and non-standard expressions; The same mechanism was used to extract the influencing factors from both channels; Based on a pre-defined lifecycle-based semantic anchor word set and zero-sample semantic similarity calculation, performance indicators and influencing factors are automatically mapped and categorized. The mapping results are associated with specific task nodes in the intelligent manufacturing system to establish an explicit text-process mapping. Based on explicit text-process mapping, lifecycle stage relationships, and process topology, construct the transmission path of influencing factors; For intelligent manufacturing systems, output a structured chain of evidence for effectiveness.

[0008] Furthermore, the unstructured text includes, but is not limited to: system operation reports and system maintenance logs.

[0009] Furthermore, the word segmenter performs semantic segmentation on unstructured text, using the following specific methods: Suppose that the input text is segmented by WordPiece to obtain a token sequence. The data is encoded into a sequence of context vectors using a pre-trained Chinese BERT model. ; in, The hidden layer dimension of BERT; For each token Generate dictionary matching feature vectors : ; This prior knowledge vector is fused with the BERT output to generate a boundary-corrected context representation: ; in, , For learnable parameters, This indicates a splicing operation; LayerNorm represents layer normalization. Calculate the industrial semantic density for each location: ; in, , For learnable parameters, It is the sigmoid function; Several convolutional kernels with different dilation rates are applied in parallel to capture short abbreviations, medium-length terms, and ultra-long compound terms, respectively: ; in, This indicates that the kernel size is... One-dimensional convolutions are used to model single-word terms, medium-length terms, and long compound terms, respectively. Calculate the initial weights for features at each scale: ; in, , For learnable parameters, This represents element-wise multiplication; Modulation effect of semantic density mapping: ; in, For adjustment coefficients, It is a semantic density vector; Adaptive fusion generates a sequence of context-enhanced vectors: ; Context-enhanced vector sequences Emission fraction obtained by linear projection , Number of tags: ; in, , These are learnable parameters; The S sequence is then input into the CRF layer to decode the globally optimal label sequence. : ; in, This is the label transition matrix.

[0010] Furthermore, the dual-channel identification of performance indicators and non-standard expressions is carried out using the following specific methods: Define the token sequence after word segmentation Context-enhanced vector sequences Global optimal label sequence BIOES format, and a dictionary of performance indicators in the field of intelligent manufacturing. ; A channel pair of token sequences processed by the word segmenter The Aho-Corasick automaton is used for efficient multi-pattern matching across the entire text. During matching, consecutive token subsequences that meet the following conditions are selected as candidate units: ; That is, only consecutive combinations of tokens marked as inside or outside the term by CRF are considered; For text segments not covered by the dictionary, i.e., the unmatched portions of channel one, sentences containing numerical information such as percentages, times, and frequencies are selected as a potential performance description candidate set, and context-enhanced vector sequences are used. Generate its sentence vector representation: ; in, The non-special token, non-BERT control character in this sentence, and the tag The index set, that is, only aggregates token vectors that belong to the term unit; For all candidate sentence vectors The K-Means clustering algorithm is used for unsupervised grouping, and the number of clusters K is automatically optimized based on the silhouette coefficient. After clustering is completed, domain experts manually review the central samples of each cluster to confirm whether they represent effective performance indicators. All sentences within the confirmed cluster were included in the candidate set of semantically extended performance indicators; The candidate segments output from Channel 1 and Channel 2 are merged, and deduplication, standardization, and context binding are performed. The final output is a structured set of candidate performance metrics, each containing the original text, standardized metric name, and source metadata.

[0011] Furthermore, the dual-channel influencing factors include: a structured set of candidate fragments for influencing factors, each containing the original text, a standardized influencing factor name, and source metadata.

[0012] Furthermore, performance indicators and influencing factors are automatically mapped and categorized, using the following methods: The lifecycle is divided into five stages: design, production, logistics, sales, and service. For each candidate fragment s, use the context-enhanced vector sequence Generate its sentence vector representation: ; in, For fragment s, the non-special token and tag The set of indices Enhance the context of vector sequences; For each anchor word set Calculate its average anchor vector: ; in, To represent the semantic vector of anchor word 'a', 'a' is treated as an independent term. After being encoded by BERT and processed by the ATC module, its output vector is taken as... Ensure that the sentence vector They exist in the same semantic space; Calculate the cosine similarity between the candidate segment s and the anchor vectors at each stage: ; Assign candidate fragments s to the lifecycle stage with the highest similarity: ; The maximum similarity value is recorded as the stage attribution confidence score: ; like , If the threshold can be set, it is marked as "stage undetermined" and requires manual review; The final output is a structured mapping result with lifecycle stage labels. Each item includes: the original text; the standardized name; the type, i.e., performance indicator or influencing factor; and the lifecycle stage label. Stage attribution confidence ; Furthermore, an explicit text-procedure mapping is established, as follows: Let the intelligent manufacturing business process model be... , where the node set Includes core business nodes; For each node Mark its keyword tag set This tag set originates from node descriptions, standard operating procedures, and business problem definitions; For each candidate fragment Only in its respective stage node subset The matching process is performed, and the sentence vector is generated as follows: ; in, For fragments China and Africa special tokens and tags index set For each node Its semantics are determined by a pre-labeled set of keywords and tags. Calculate its node keyword vector, which is the same as the sentence vector. Each keyword is in the same semantic space. Treating each term as an independent term, extract its corresponding context enhancement vector, and take the mean as the semantic representation of the node: ; in, Keywords ; Calculation fragment With nodes Cosine similarity: ; The business nodes have a clear topological relationship, and different constraints are introduced for different stages; When calculating similarity, a weighting coefficient is added to node pairs that conform to the business flow direction, business problem relevance, and data flow constraints. : ; in, The constraint must satisfy the indicator function; fragment Bind to the business process node with the highest similarity: ; The maximum similarity value is recorded as the node binding confidence score: ; like , If the threshold can be set, it is marked as an undetermined node and requires manual review; The final output is a structured result with business process nodes bound to it. Each item includes: the original text; a standardized name; the type, i.e., metric / influencing factor; and a lifecycle stage label. Stage attribution confidence Bind business process nodes Node binding confidence .

[0013] Furthermore, the transmission path of influencing factors is constructed, and the specific methods are as follows: Suppose any two co-occurring terms , Their life cycle stages are respectively , The bound business nodes are respectively , Based on the operational logic of the manufacturing system, the causal direction inference rules are defined as follows: like Based on the stage partial order relation, the stage partial order relation is defined. for: like Then it can be inferred that As an upstream influencing factor, Establish directed edges as downstream performance indicators ; like Then, the direction is determined based on the node topology order in the business process model: if the node Prior to in the flowchart ,but ; like and If the components are not sequential or preceding in the topology, then a causal priority score is introduced to determine the direction: ; in, The time series offset is used to extract the average timestamp of the corresponding item from the unstructured data in the operation and maintenance logs. ,like In the time sequence Priority occurs when the symbolic function value is 1, while synchronization results in a value of 0. The priority for occurrence is -1. Data dependency strength is used to measure the data flow between business nodes. If a node... The input parameters depend on The output result is then Conversely, 0 indicates a positive value and 0 indicates no dependency. , For weighting coefficients; when At that time, establish the causal path , Conversely, otherwise it will be impossible to determine; In the absence of operation and maintenance log timestamps or when the aforementioned methods cannot determine the issue, for nodes in parallel branches... , Semantic logical weight analysis is introduced, and the scores are as follows: ; in, Indicates the strength of semantic relations; If it contains strong causal conjunctions, and If it is the subject or the head word of the causal clause, then If there is no correlation, the value is 0. The value is -1 if it is the subject or the head word of the causal clause; The inherent character strength of the element type, if Identified as an influencing factor and If identified as a performance indicator, then Conversely, -1 is given if both factors or indicators are the same; 0 is given if both factors or indicators are the same. , These are the weighting coefficients. Then establish ,on the contrary 0 will be judged as "insufficient evidence", marked as a collaborative related item, and transferred to manual audit review; For each performance indicator Tracing back along the directed edges to identify all reachable influencing factors, a transmission path for influencing factors is formed: ; The overall confidence score of this path is calculated by a weighted average of the confidence scores of each edge: ; like , If a threshold can be set, it will be marked as a low-confidence path and requires manual review. The final output is a structured set of influencing factor transmission paths. Each path includes: path ID; initial influencing factor; intermediate links; final performance indicators; and overall confidence level.

[0014] Furthermore, a structured chain of evidence for effectiveness is output, as follows: Each chain of evidence for effectiveness Defined as a quintuple: ; Where ID is a unique path identifier; Outcome is an abstract performance target; Path is the transmission path of influencing factors; and Confidence is the overall confidence level of the path. Source refers to the original document metadata.

[0015] Due to the adoption of the above technical solutions, this application has the following beneficial effects: 1. More accurate terminology recognition, ensuring semantic integrity. By introducing an adaptive multi-scale perceptual word segmentation algorithm that integrates industrial semantic priors, the problem of incorrect segmentation of manufacturing professional terms with varying lengths by general word segmentation tools is effectively solved. This algorithm combines an industrial entity boundary correction layer, a semantic density mapping mechanism, and an adaptive dilated convolutional network to significantly improve the accuracy of boundary recognition for ultra-long compound terms while maintaining inference efficiency, providing high-quality semantic units for subsequent extraction of indicators and influencing factors.

[0016] 2. The indicators and influencing factors are more comprehensive, balancing standardization and generalization. A dual-channel collaborative extraction mechanism of "dictionary matching + semantic clustering" is adopted. This mechanism achieves full coverage of known and unknown expressions without relying on large-scale labeled data, effectively solving the problems of fragmented evidence and diverse expressions in unstructured text.

[0017] 3. More reliable transmission path construction supports scientific decision-making and continuous optimization. By integrating the triple constraints of text co-occurrence, partial order relationships of lifecycle stages, and process topology of key business models in production and manufacturing, the system automatically derives the transmission paths of influencing factors across stages and introduces a confidence assessment mechanism. This avoids the subjectivity and arbitrariness of manual attribution, providing enterprises with structured root cause analysis to guide precise investment and cross-departmental collaborative improvement.

[0018] 4. End-to-end automation, supporting large-scale applications. The entire process, from unstructured text input to structured evidence chain output, is automated without manual intervention (only initial dictionary and process model configuration are required). It can process hundreds or thousands of project documents in batches, and is suitable for large-scale scenarios such as group-level performance evaluation, policy subsidy application, and standard compliance verification, significantly reducing labor costs and time expenditure.

[0019] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0020] The accompanying drawings of this invention are described below.

[0021] Figure 1 This is a schematic diagram of the overall process of the present invention.

[0022] Figure 2 This is a schematic diagram of the business model for an intelligent manufacturing system. Detailed Implementation

[0023] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0024] A method for constructing the transmission path of factors influencing the effectiveness of intelligent manufacturing system solutions, such as... Figure 1 As shown, perform the following steps: S1 acquires unstructured text from intelligent manufacturing systems.

[0025] Step S1: Unstructured text includes various reports, operation and maintenance logs, etc.

[0026] S2. Domain-specific word segmentation preprocessing: Semantic segmentation of unstructured text is performed using an adaptive multi-scale perceptual word segmenter.

[0027] Step S2, to accurately identify specialized terms in the manufacturing system and address the semantic fragmentation of industrial text by general word segmentation tools, adopts an adaptive multi-scale perceptual word segmentation algorithm that integrates prior industrial semantics. This algorithm, considering the wide distribution of terminology lengths in the manufacturing field, introduces prior distributions of industrial terminology lengths, knowledge graph constraint mechanisms, and a multi-scale dilatational perceptual network to achieve accurate boundary identification of specialized manufacturing terms. The specific steps are as follows: S21, Industrial entity boundary correction.

[0028] Suppose that the input text is segmented by WordPiece to obtain a token sequence. First, the data is encoded into a sequence of context vectors using a pre-trained Chinese BERT model: ; in is the dimension of the hidden layer in BERT.

[0029] To overcome the lack of hard-boundary common sense in the industrial field in general models, a boundary correction mechanism guided by deterministic knowledge was used. First, a domain dictionary of performance indicators for intelligent manufacturing was constructed. This dictionary can be derived from sources such as national standards, patents, and academic papers. Based on this dictionary, for each token... Generate dictionary matching feature vectors : ; This prior knowledge vector is fused with the BERT output to generate a boundary-corrected context representation: ; in, , For learnable parameters, This indicates a concatenation operation, and LayerNorm represents layer normalization. This mechanism enables the algorithm to possess semantic boundary awareness guided by dictionary priors.

[0030] S22, Semantic density mapping.

[0031] In manufacturing and production texts, the distribution of technical terms exhibits a highly non-uniform characteristic. A semantic density mapping mechanism is introduced to dynamically assign multi-scale convolution weights to different regions. First, the industrial semantic density at each location is calculated: ; in, , For learnable parameters, This is a sigmoid function. The density value is initialized using prior knowledge of the term distribution in the corpus generated during the statistical pre-training phase, rather than being randomly initialized.

[0032] To capture long-range dependent term features in manufacturing texts, traditional convolution is upgraded to adaptive dilated convolution. This module applies three convolutional kernels with different dilation rates in parallel to capture short abbreviations, medium-length terms, and ultra-long compound terms, respectively. ; in This indicates that the kernel size is... One-dimensional convolutions are used to model single-word terms, medium-length terms, and long compound terms, respectively.

[0033] S23, Adaptive fusion driven by prior knowledge of industrial terminology length.

[0034] To address the wide distribution characteristics of manufacturing terminology lengths, an adaptive fusion mechanism driven by prior knowledge of industrial terminology lengths was adopted. First, the initial weights of features at each scale were calculated: ; in , For learnable parameters, This represents element-wise multiplication. Subsequently, the modulation effect of the industrial semantic density mapping is introduced: ; in For adjustment coefficients, For the semantic density vector calculated in step S22, this mechanism automatically enhances it in high-density regions. Weighting enhances the ability to recognize long, complex terms; in low-density areas, it strengthens... Weighting optimizes the recognition of abbreviations.

[0035] Finally, adaptive fusion generates a sequence of context-enhanced vectors: ; Finally, the context-enhanced vector sequence Emission fraction obtained by linear projection ( (Number of tags) ; in, , These are learnable parameters.

[0036] The S sequence is then input into the CRF layer to decode the globally optimal label sequence. : ; in, This is the label transition matrix.

[0037] This architecture will ensure semantically complete word segmentation results, providing high-quality input for subsequent extraction of metrics and influencing factors.

[0038] S3. Dual-channel performance indicator extraction: Combining domain dictionary precise matching and BERT semantic classification, dual-channel recognition of performance indicators and non-standard expressions is used.

[0039] Step S3 involves comprehensively and accurately identifying indicators related to the effectiveness of smart manufacturing implementation from the preprocessed text. A dual-channel collaborative extraction mechanism is adopted to balance the precise recall of known standard indicators with the semantic generalization ability of unknown expressions.

[0040] The input is the triplet output from step S2, "Domain-Specific Token Preprocessing": the token sequence after tokenization. Context-enhanced vector sequences Global optimal label sequence (BIOES format) and a dictionary of performance indicators in the field of intelligent manufacturing The specific processing steps are as follows: S31, Channel 1: Precise matching based on domain dictionary.

[0041] The token sequence after word segmentation The Aho-Corasick Algorithm is employed for efficient multi-pattern matching across the entire text. During matching, consecutive token subsequences satisfying the following conditions are selected as candidate units: ; That is, only consecutive token combinations marked as "within or outside the term" by CRF are considered, avoiding matching non-term fragments.

[0042] If a candidate unit is completely identical to any indicator term in the dictionary or satisfies the preset synonym mapping rules, it is marked as a candidate fragment of high-confidence performance indicator, and its standardized subject, original expression, and context window are recorded.

[0043] S32, Channel Two: Discovery of Unknown Indicators Based on Semantic Clustering

[0044] For text segments not covered by the dictionary (i.e., unmatched portions in channel one), a semantically driven metric discovery mechanism is activated. Specifically, sentences containing numerical information such as percentages, time, and frequency are selected as a candidate set of potential performance descriptions. The context-enhanced vector sequence output from S2 is then utilized. Generate its sentence vector representation: ; in, The non-special token (non-BERT control character) and tag in this sentence The index set, which aggregates only token vectors belonging to "term units", further enhances semantic focus.

[0045] Subsequently, for all candidate sentence vectors Unsupervised grouping was performed using the K-Means clustering algorithm, with the number of clusters K automatically optimized based on the silhouette coefficient. After clustering, domain experts manually reviewed the center samples of each cluster to confirm whether they represented effective performance metrics. All sentences within the confirmed clusters were included in the semantically expanded performance metric candidate set.

[0046] S33, Candidate set fusion and deduplication.

[0047] The candidate segments output from Channel 1 and Channel 2 are merged, and the following post-processing is performed: S331. Deduplication. If the same original segment is captured by two channels simultaneously, retain the result from channel one. S332. Standardization. Mapping synonyms to dictionary terms; S333, Context Binding. Retains the original sentence, source document, and paragraph position information for subsequent source tracing.

[0048] The final output is a structured set of candidate performance metrics, each containing the original text, standardized metric name, and source metadata.

[0049] This dual-channel mechanism ensures high recall of standard metrics while effectively expanding coverage of non-standard expressions, providing high-quality input for subsequent lifecycle mapping and transmission path construction.

[0050] S4. Dual-channel influencing factor extraction: Extract dual-channel influencing factors using the same mechanism.

[0051] To comprehensively identify the key constraints or promoting factors driving the effectiveness of intelligent manufacturing implementation, a dual-channel influencing factor extraction mechanism is constructed and executed in parallel with the S3 indicator extraction. This mechanism aims to identify implicit influencing factors from unstructured text, providing causal evidence for subsequently constructing the transmission path from influencing factors to effectiveness indicators.

[0052] Similar to the extraction of performance indicators, the technical approach is the same. The final output is a structured set of candidate fragments for influencing factors, each containing the original text, the standardized name of the influencing factor, and source metadata.

[0053] S5. Automatic lifecycle stage mapping: Based on the preset lifecycle, a set of semantic anchor words is generated, and zero-sample semantic similarity calculation is performed to automatically map and classify performance indicators and influencing factors.

[0054] To accurately categorize the extracted performance indicators and influencing factors to the corresponding stages of the intelligent manufacturing lifecycle, a zero-sample automatic mapping method based on semantic anchors is adopted. This method eliminates the need for manual annotation of each new indicator; instead, it calculates the semantic similarity between a pre-defined set of domain semantic anchor terms and candidate segments, thus achieving automatic stage assignment. The specific steps are as follows: S51. Preset semantic anchor word set for five major life cycle stages.

[0055] Based on national standards and industry practices, and building upon the five stages of design, production, logistics, sales, and service, a semantic anchor term set is constructed to address the characteristics of the manufacturing system. In particular, it enriches the fine-grained semantic representation of the "production" stage, with each anchor word... All have been standardized and mapped to unique lifecycle stage tags. .

[0056] S52, Semantic Vector Generation and Similarity Calculation.

[0057] For each candidate segment s, i.e., the original text field in the structured candidate segments output in steps S3 and S4, the context-enhanced vector sequence output in S2 is utilized. Generate its sentence vector representation: ; in For fragment s, the non-special token and tag The set of indices Enhance the context vector sequence.

[0058] For each anchor word set Calculate its average anchor vector: ; in, This is the semantic vector representation of the anchor word 'a' under the same model as S2. Specifically, 'a' is treated as an independent term, encoded using BERT, processed by the ATC module, and its output vector is taken as... Ensure that the sentence vector They exist in the same semantic space.

[0059] Subsequently, the cosine similarity between the candidate segment s and the anchor vectors of each stage is calculated: ; S53, Automatic Stage Assignment and Confidence Assessment.

[0060] Assign candidate fragments s to the lifecycle stage with the highest similarity: ; At the same time, the maximum similarity value is recorded as the stage attribution confidence score: ; like (threshold) If it can be set, it will be marked as "Stage Undetermined" and requires manual review.

[0061] S54. Output the structured stage mapping results.

[0062] The final output is a structured mapping result with lifecycle stage labels. Each item includes: original text; standardized name; type (performance indicator / influencing factor); lifecycle stage label. Stage attribution confidence Source metadata.

[0063] This method achieves automatic lifecycle stage mapping through semantic anchors, providing a structured foundation for subsequent business process node binding and transmission path construction. The stage mapping results provide constraints for subsequent business process node binding, ensuring that fine-grained matching is performed within a reasonable semantic domain.

[0064] S6. Business process node binding: Associate the mapping results with specific task nodes in the intelligent manufacturing system to establish an explicit text-process mapping.

[0065] To further link the performance indicators and influencing factors of S5 mapped to lifecycle stages to the core business model of the manufacturing system, this paper uses a node binding mechanism based on a business process model. This mechanism achieves an explicit mapping from abstract indicators to specific business process tasks by matching semantic terms with process node keywords, thereby supporting the construction of a traceable transmission path of "indicator-process-root cause". The manufacturing system business model is as follows: Figure 2 As shown, the specific operation steps are as follows: S61. Business process model preprocessing and node annotation.

[0066] Let the intelligent manufacturing business process model be... , where the node set It includes the following core business nodes: 1) Procurement node: Processes supplier purchase orders (POs); 2) Raw material inventory node: Manages raw material inventory; 3) MRP (Material Requirements Planning) node: Calculates material requirements based on the Bill of Materials (BOM); 4) Production node: Executes production and processing tasks; 5) MPS (Master Production Schedule) node: Develops the master production schedule; 6) Finished goods inventory node: Manages finished goods inventory; 7) Sales node: Processes customer order of goods (COs); 8) BOM / BOP (Bill of Materials / Process Specifications) node: Defines product composition and process flow; 9) Supplier node: Provides raw materials; 10) Customer node: Receives products; 11) Accounts payable node: Processes purchase payments; 12) Accounts receivable node: Processes sales receipts; 13) Material node: Manages material information; 14) Product node: Manages product information.

[0067] For each node Mark its keyword tag set The tag set is derived from node descriptions, standard operating procedures, and business problem definitions, as shown in the table below: node Keyword tag set Procurement Node {"Supplier", "PO", "Purchase Order", "Purchase Cycle", "Payment Terms", "On-Time Purchase Rate"} Raw material inventory nodes {"Raw Material Inventory", "Inventory Turnover", "Material Shortage", "Inventory Level", "Inventory Accuracy"} MRP node {"MRP", "Material Requirements Planning", "BOM Deployment", "Material Shortages", "Demand Calculation"} Production Node {"Production Execution", "Production Schedule", "Process Route", "Changeover Time", "OEE", "Work-in-Process"} MPS Node {"MPS", "Master Production Schedule", "Capacity Balance", "Order Delivery", "Plan Achievement Rate"} Finished Goods Inventory Node {"Finished Goods Inventory", "Inventory Turnover Rate", "Shipping Delay", "Inventory Accuracy Rate"} Sales Node {"Sales Orders", "CO", "Delivery Time", "Customer Satisfaction", "Order Fulfillment Rate"} BOM / BOP process list nodes {"BOM", "Bill of Materials", "Bill of Processes", "Product Structure", "Process Routing Design"} Supplier Node {"Suppliers", "Supplier Management", "Supplier Performance", "Supplier Delivery Rate", "Supplier Quality"} Customer Node {"Customer Orders", "Customer Satisfaction", "Customer Complaint Rate", "Customer On-Time Delivery Rate", "Customer Needs"} Material nodes {"Material Information", "Material Code", "Material Attributes", "Material Category", "Bill of Materials"} Product Node {"Product Information", "Product Specifications", "Product Structure", "Product Version", "Product Lifecycle"} Nodes to be dealt with {"Accounts Payable", "Purchase Payments", "Payment Terms", "Payment Cycle", "On-Time Payment Rate"} Accounts receivable nodes {"Accounts Receivable", "Sales Receivables", "Payment Terms", "Payment Cycle", "On-Time Payment Rate"} S62, Semantic similarity-driven node matching.

[0068] For each candidate fragment That is, S5 output, only in its respective stage. node subset The matching is performed within the range.

[0069] like For the design phase, then For {BOM / BOP process list node, product node} like For the production stage, then For {MRP Material Requirements Planning node, Production node, MPS Master Production Schedule node, Material node} like For the logistics stage, then For {procurement node, raw material inventory node, finished goods inventory node} like If it is the sales stage, then For {sales node, customer node} like For the service phase, then For {Accounts Receivable Node, Accounts Payable Node, Supplier Node} To maintain semantic space consistency, the same sentence vector generation method as S5 will still be used: ; in For fragments China and Africa special tokens and tags The set of indices.

[0070] For each node Its semantics are determined by a pre-labeled set of keywords and tags. Calculate its node keyword vector, which is the same as the sentence vector. Each keyword is in the same semantic space. Treating each term as an independent term, inputting it into the S2 algorithm model, extracting its corresponding context enhancement vector, and taking the mean as the semantic representation of the node: ; in Keywords .

[0071] Subsequently, the calculation fragment With nodes Cosine similarity: ; S63. Node matching optimization for business model topology constraints.

[0072] In a manufacturing system, business nodes have a clear topological relationship, and the following constraints are introduced: 1) Business flow direction constraints: Procurement → Raw material inventory → MRP → Production → MPS → Finished goods inventory → Sales.

[0073] 2) Business problem-related constraints, such as: "When to buy what / how much" → Procurement Node "How much raw material is missing?" → Raw material inventory node "When to produce what / how much" → Production node "How to manufacture" → BOM / BOP process list node "How many finished products are missing?" → MPS node "When and how much to deliver" → Sales milestones "What materials does it consist of?" → BOM / BOP process list node.

[0074] 3) Data flow constraints: MRP node output → MO (Production Order) → Production node MPS node output → CO (Customer Order) → Sales node When calculating similarity, a weighting coefficient is added to node pairs that conform to the business flow direction, business problem relevance, and data flow constraints. ; in The constraint must satisfy the indicator function (0 or 1).

[0075] S64, Optimal Node Allocation and Confidence Filtering.

[0076] fragment Bind to the business process node with the highest similarity: ; At the same time, the maximum similarity value is recorded as the node binding confidence score: ; like (threshold) If the node is undefined (can be set), it will be marked as "node undetermined" and requires manual review.

[0077] S65, Output the structured node binding results.

[0078] The final output is a structured result with business process nodes bound to it. Each item includes: original text; standardized name; type (metric / influencing factor); and lifecycle stage label. Stage attribution confidence Bind business process nodes Node binding confidence Source metadata.

[0079] This mechanism achieves explicit association between indicators / influencing factors and business process nodes through semantic matching, providing a structured foundation for "constructing the transmission path of influencing factors" and truly opening up a traceable chain from "textual evidence to business process".

[0080] S7. Constructing the transmission path of influencing factors: Based on text-process explicit mapping, life cycle stage relationships and process topology, construct the transmission path of influencing factors.

[0081] To achieve traceable attribution from abstract results to specific business processes, a cross-lifecycle transmission path of influencing factors is constructed based on the structured indicators and influencing factors output from previous steps. This path uses "influencing factors → results indicators → business process nodes" as its basic unit, and automatically derives the causal chain driving changes in results through co-occurrence analysis and stage sequence constraints. The specific steps are as follows: S71, Co-occurrence Relationship Extraction.

[0082] Construct a co-occurrence graph for all structured items within the same document paragraph or semantic context window (default 3 sentences before and after). , where the node set For all indicators and influencing factors; edge set This indicates that two items co-occur in the text, meaning they appear in the same context window.

[0083] For each co-occurring edge Calculate its co-occurrence intensity: ; in, For text and distance, Enhance the cosine similarity of two context vectors.

[0084] S72, Causal Direction Inference under Stage Sequence Constraints.

[0085] Suppose any two co-occurring terms , Their life cycle stages are respectively , The bound business nodes are respectively , Based on the operational logic of the manufacturing system, the causal direction inference rules are defined as follows: like Based on the stage partial order relation, the stage partial order relation is defined. as follows ; like Then it can be inferred that As an upstream influencing factor, Establish directed edges as downstream performance indicators .

[0086] When the node topology order in the business process model is used to determine the direction, if the node... Prior to in the flowchart ,but Its topology constraints are detailed in S63, Service Flow Direction Constraints.

[0087] like and If the components are not sequential or preceding in the topology, then a causal priority score is introduced to determine the direction: ; in, The time series offset is used to extract the average timestamp of the corresponding item from the unstructured data in the operation and maintenance logs. ,like In the time sequence Priority occurs when the symbolic function value is 1, while synchronization results in a value of 0. The priority for occurrence is -1. Data dependency strength is used to measure the data flow between business nodes. If a node... The input parameters depend on The output result is then Conversely, 0 indicates a positive value and 0 indicates no dependency. , For weighting coefficients. When At that time, establish the causal path , Conversely, otherwise it will be impossible to determine.

[0088] In the absence of operation and maintenance log timestamps or when the aforementioned methods cannot determine the issue, for nodes in parallel branches... , Semantic logical weight analysis can be introduced, with the following scores: ; in, To determine the strength of semantic relationships, examine the text fragments extracted in the above steps for explicit causal indicators or syntactic dependency relationships. If the text contains strong causal conjunctions such as "cause," "due to," "trigger," or "drive," and... If it is the subject or the head word of the causal clause, then If there is no correlation, the value is 0. The value is -1 if it is the subject or the head word of the causal clause; The inherent character strength of the element type, if Identified as an influencing factor and If identified as a performance indicator, then Conversely, -1 is given if both factors or indicators are the same; 0 is given if both factors or indicators are the same. , These are the weighting coefficients. Then establish ,on the contrary 0 will be judged as "insufficient evidence", marked as a collaborative related item, and transferred to manual audit review.

[0089] S73, Transmission Path Generation and Confidence Calculation.

[0090] For each performance indicator Tracing back along the directed edges to identify all reachable influencing factors, a transmission path for influencing factors is formed: ; The overall confidence score of this path is calculated by a weighted average of the confidence scores of each edge: ; like (threshold) If it can be set, it will be marked as a "low confidence path" and requires manual review.

[0091] S74, Output structured transmission path.

[0092] The final output is a structured set of influencing factor transmission paths. Each path includes: path ID; starting influencing factor (standardized name, original text, stage, node); intermediate links (if any); endpoint performance indicators (standardized name, original text, stage, node); overall confidence level; and source document metadata.

[0093] This mechanism constructs an interpretable, auditable, and traceable transmission path of influencing factors that conforms to the operating logic of a smart factory through triple constraints of text co-occurrence, stage sequence, and process topology, providing a structured basis for performance attribution and continuous optimization.

[0094] S8. Output a structured chain of evidence for effectiveness: For intelligent manufacturing systems, output a structured chain of evidence for effectiveness.

[0095] To support third-party evaluation, policy application, and continuous optimization of intelligent manufacturing system solutions, the outputs of previous steps are integrated into a structured chain of effectiveness evidence, achieving end-to-end transformation from unstructured text to standardized evaluation criteria that are auditable, reproducible, and traceable. The specific steps are as follows: S81, Each chain of evidence for effectiveness Defined as a quintuple: ; In this context, ID represents a unique path identifier; Outcome represents an abstract outcome goal (e.g., "improving flexible manufacturing capabilities"); Path represents the transmission path of influencing factors (including initial influencing factors, intermediate stages, and endpoint indicators); and Confidence represents the overall confidence level of the path. (From step S7); Source is the original document metadata (filename, page number, paragraph).

[0096] S82. The final output is structured JSON or database records.

[0097] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for constructing the transmission path of factors influencing the effectiveness of intelligent manufacturing system solutions, characterized in that, The specific method is as follows: Acquire unstructured text from intelligent manufacturing systems; Semantic segmentation of unstructured text is performed using an adaptive multi-scale perceptual word segmenter; Combining domain dictionary precise matching with BERT semantic classification, dual-channel recognition of performance indicators and non-standard expressions; The same mechanism was used to extract the influencing factors from both channels; Based on a pre-defined lifecycle-based semantic anchor word set and zero-sample semantic similarity calculation, performance indicators and influencing factors are automatically mapped and categorized. The mapping results are associated with specific task nodes in the intelligent manufacturing system to establish an explicit text-process mapping. Based on explicit text-process mapping, lifecycle stage relationships, and process topology, construct the transmission path of influencing factors; For intelligent manufacturing systems, output a structured chain of evidence for effectiveness.

2. The method for constructing the transmission path of factors influencing the effectiveness of intelligent manufacturing system solutions as described in claim 1, characterized in that, The unstructured text includes, but is not limited to: system operation reports and system maintenance logs.

3. The method for constructing the transmission path of factors influencing the effectiveness of intelligent manufacturing system solutions as described in claim 1, characterized in that, The word segmenter performs semantic segmentation on unstructured text, and the specific method is as follows: Suppose that the input text is segmented by WordPiece to obtain a token sequence. The data is encoded into a sequence of context vectors using a pre-trained Chinese BERT model. ； in, The hidden layer dimension of BERT; For each token Generate dictionary matching feature vectors : ； This prior knowledge vector is fused with the BERT output to generate a boundary-corrected context representation: ； in, , For learnable parameters, This indicates a splicing operation; LayerNorm represents layer normalization. Calculate the industrial semantic density for each location: ； in, , For learnable parameters, It is the sigmoid function; Several convolutional kernels with different dilation rates are applied in parallel to capture short abbreviations, medium-length terms, and ultra-long compound terms, respectively: ； in, This indicates that the kernel size is... One-dimensional convolutions are used to model single-word terms, medium-length terms, and long compound terms, respectively. Calculate the initial weights for features at each scale: ； in, , For learnable parameters, This represents element-wise multiplication; Modulation effect of semantic density mapping: ； in, For adjustment coefficients, It is a semantic density vector; Adaptive fusion generates a sequence of context-enhanced vectors: ； Context-enhanced vector sequences Emission fraction obtained by linear projection , Number of tags: ； in, , These are learnable parameters; The S sequence is then input into the CRF layer to decode the globally optimal label sequence. : ； in, This is the label transition matrix.

4. The method for constructing the transmission path of factors influencing the effectiveness of intelligent manufacturing system solutions as described in claim 1, characterized in that, The dual-channel identification of effectiveness indicators and non-standard expressions is presented in the following specific methods: Define the token sequence after word segmentation Context-enhanced vector sequences Global optimal label sequence BIOES format, and a dictionary of performance indicators in the field of intelligent manufacturing. ; A channel pair of token sequences processed by the word segmenter The Aho-Corasick Algorithm is used for efficient multi-pattern matching across the entire text. During matching, consecutive token subsequences that meet the following conditions are selected as candidate units: ； That is, only consecutive combinations of tokens marked as inside or outside the term by CRF are considered; For text segments not covered by the dictionary, i.e., the unmatched portions of channel one, sentences containing numerical information such as percentages, times, and frequencies are selected as a potential performance description candidate set, and context-enhanced vector sequences are used. Generate its sentence vector representation: ； in, The non-special token, non-BERT control character in this sentence, and the tag The index set, that is, only aggregates token vectors that belong to the term unit; For all candidate sentence vectors The K-Means clustering algorithm is used for unsupervised grouping, and the number of clusters K is automatically optimized based on the silhouette coefficient. After clustering is completed, domain experts manually review the central samples of each cluster to confirm whether they represent effective performance indicators. All sentences within the confirmed cluster were included in the candidate set of semantically extended performance indicators; The candidate segments output from Channel 1 and Channel 2 are merged, and deduplication, standardization, and context binding are performed. The final output is a structured set of candidate performance metrics, each containing the original text, standardized metric name, and source metadata.

5. The method for constructing the transmission path of factors influencing the effectiveness of intelligent manufacturing system solutions as described in claim 4, characterized in that, The dual-channel influencing factors include: a structured set of candidate fragments for influencing factors, each containing the original text, a standardized influencing factor name, and source metadata.

6. The method for constructing the transmission path of factors influencing the effectiveness of intelligent manufacturing system solutions as described in claim 1, characterized in that, The performance indicators and influencing factors are automatically mapped and categorized, using the following method: The lifecycle is divided into five stages: design, production, logistics, sales, and service. For each candidate fragment s, use the context-enhanced vector sequence Generate its sentence vector representation: ； in, For fragment s, the non-special token and tag The set of indices Enhance the context of vector sequences; For each anchor word set Calculate its average anchor vector: ； in, To represent the semantic vector of anchor word 'a', 'a' is treated as an independent term. After being encoded by BERT and processed by the ATC module, its output vector is taken as... Ensure that the sentence vector They exist in the same semantic space; Calculate the cosine similarity between the candidate segment s and the anchor vectors at each stage: ； Assign candidate fragments s to the lifecycle stage with the highest similarity: ； The maximum similarity value is recorded as the stage attribution confidence score: ； like , If the threshold can be set, it is marked as "stage undetermined" and requires manual review; The final output is a structured mapping result with lifecycle stage labels. Each item includes: the original text; the standardized name; the type, i.e., performance indicator or influencing factor; and the lifecycle stage label. Stage attribution confidence .

7. The method for constructing the transmission path of factors influencing the effectiveness of intelligent manufacturing system solutions as described in claim 6, characterized in that, To establish an explicit text-procedure mapping, follow these steps: Let the intelligent manufacturing business process model be... , where the node set Includes core business nodes; For each node Mark its keyword tag set This tag set originates from node descriptions, standard operating procedures, and business problem definitions; For each candidate fragment Only in its respective stage node subset The matching process is performed, and the sentence vector is generated as follows: ； in, For fragments China and Africa special tokens and tags index set For each node Its semantics are determined by a pre-labeled set of keywords and tags. Calculate its node keyword vector, which is the same as the sentence vector. Each keyword is in the same semantic space. Treating each term as an independent term, extract its corresponding context enhancement vector, and take the mean as the semantic representation of the node: ； in, Keywords ; Calculation fragment With nodes Cosine similarity: ； The business nodes have a clear topological relationship, and different constraints are introduced for different stages; When calculating similarity, a weighting coefficient is added to node pairs that conform to the business flow direction, business problem relevance, and data flow constraints. : ； in, The constraint must satisfy the indicator function; fragment Bind to the business process node with the highest similarity: ； The maximum similarity value is recorded as the node binding confidence score: ； like , If the threshold can be set, it is marked as an undetermined node and requires manual review; The final output is a structured result with business process nodes bound to it. Each item includes: the original text; a standardized name; the type, i.e., metric / influencing factor; and a lifecycle stage label. Stage attribution confidence Bind business process nodes Node binding confidence .

8. The method for constructing the transmission path of factors influencing the effectiveness of intelligent manufacturing system solutions as described in claim 1, characterized in that, The specific methods for constructing the transmission path of influencing factors are as follows: Suppose any two co-occurring terms , Their life cycle stages are respectively , The bound business nodes are respectively , Based on the operational logic of the manufacturing system, the causal direction inference rules are defined as follows: like Based on the stage partial order relation, the stage partial order relation is defined. for: ； like Then it can be inferred that As an upstream influencing factor, Establish directed edges as downstream performance indicators ; like Then, the direction is determined based on the node topology order in the business process model: if the node Prior to in the flowchart ,but ; like and If the components are not sequential or preceding in the topology, then a causal priority score is introduced to determine the direction: ； in, The time series offset is used to extract the average timestamp of the corresponding item from the unstructured data in the operation and maintenance logs. ,like In the time sequence Priority occurs when the symbolic function value is 1, while synchronization results in a value of 0. The priority for occurrence is -1. Data dependency strength is used to measure the data flow between business nodes. If a node... The input parameters depend on The output result is then Conversely, 0 indicates a positive value and 0 indicates no dependency. , For weighting coefficients; when At that time, establish the causal path , Conversely, otherwise it will be impossible to determine; In the absence of operation and maintenance log timestamps or when the aforementioned methods cannot determine the issue, for nodes in parallel branches... , Semantic logical weight analysis is introduced, and the scores are as follows: ； in, Indicates the strength of semantic relations; If it contains strong causal conjunctions, and If it is the subject or the head word of the causal clause, then If there is no correlation, the value is 0. The value is -1 if it is the subject or the head word of the causal clause; The inherent character strength of the element type, if Identified as an influencing factor and If identified as a performance indicator, then Conversely, -1 is given if both factors or indicators are the same; 0 is given if both factors or indicators are the same. , These are the weighting coefficients. Then establish ,on the contrary 0 will be judged as "insufficient evidence", marked as a collaborative related item, and transferred to manual audit review; For each performance indicator Tracing back along the directed edges to identify all reachable influencing factors, a transmission path for influencing factors is formed: ； The overall confidence score of this path is calculated by a weighted average of the confidence scores of each edge: ； like , If a threshold can be set, it will be marked as a low-confidence path and requires manual review. The final output is a structured set of influencing factor transmission paths. Each path includes: path ID; initial influencing factor; intermediate links; final performance indicators; and overall confidence level.

9. The method for constructing the transmission path of factors influencing the effectiveness of intelligent manufacturing system solutions as described in claim 1, characterized in that, Output a structured chain of evidence for effectiveness, as follows: Each chain of evidence for effectiveness Defined as a quintuple: ； Where ID is a unique path identifier; Outcome is an abstract performance target; Path is the transmission path of influencing factors; and Confidence is the overall confidence level of the path. Source refers to the original document metadata.

Images