A semiconductor production report routing decision method, system, device and medium based on cognitive computing

Abstract

The present application relates to the technical field of semiconductor manufacturing, in particular to a semiconductor production report routing decision method, system, device and medium based on cognitive computing; first, input the obtained natural language query request into the constructed pre-training language model in the field of semiconductors to obtain a semantic encoding matrix and calculate the sentence vector representation of the natural language query request; then analyze the sentence vector representation to obtain the entity, intent type and slot value of the natural language query request and construct a structured query intent; finally, according to the structured query intent, construct a query intent vector to obtain a report routing decision result; by constructing a semiconductor production report cognitive knowledge graph, the semantic understanding ability of a large language model and the reasoning ability of a knowledge graph are fused to realize natural language driven cross-report intelligent question answering and root cause analysis, and the effect of non-technical personnel availability, cross-system automatic association and intelligent root cause analysis is realized.

Description

Technical Field

[0001] This invention relates to the field of semiconductor manufacturing technology, and more specifically, to a method, system, device, and medium for semiconductor production report routing decision-making based on cognitive computing. Background Technology

[0002] Semiconductor manufacturing is one of the most complex industrial manufacturing processes today. An advanced process production line involves hundreds of steps and thousands of pieces of equipment. Each batch of products must undergo hundreds of process steps, including photolithography, etching, thin film deposition, ion implantation, and chemical mechanical polishing. During this process, production managers, process engineers, equipment engineers, and quality engineers need to frequently query and analyze various production reports, including equipment status reports, process parameter reports, yield reports, defect reports, work-in-process (WIP) reports, and OEE reports.

[0003] As process nodes evolve towards 3nm and 2nm, the amount of production data is growing exponentially, with over 50 types of reports and thousands of data dimensions. Traditional manual query and analysis methods can no longer meet the needs of rapid decision-making, creating an urgent need for an intelligent report and question-and-answer system. Summary of the Invention

[0004] This invention addresses the problems of scattered report data, high query thresholds, and lack of cognitive reasoning capabilities in existing management and control systems. It proposes a method, system, device, and medium for routing decision-making of semiconductor production reports based on cognitive computing. By constructing a cognitive knowledge graph of semiconductor production reports and integrating the semantic understanding capabilities of a large language model with the reasoning capabilities of the knowledge graph, it achieves natural language-driven intelligent question answering and root cause analysis across reports, enabling non-technical personnel to use the system, automatic cross-system association, and intelligent root cause analysis.

[0005] The specific implementation details of this invention are as follows: A semiconductor production report routing decision-making method based on cognitive computing specifically includes the following steps: Step S1: Input the acquired natural language query request into the constructed semiconductor domain pre-trained language model to obtain the semantic encoding matrix and calculate the sentence vector representation of the natural language query request; Step S2: Parse the sentence vector representation to obtain the entity, intent type, and slot value of the natural language query request, and construct a structured query intent; Step S3: Based on the structured query intent, construct the query intent vector to obtain the report routing decision result.

[0006] To better realize the present invention, step S1 further includes the following steps: Step S11: Construct a pre-trained corpus for the semiconductor manufacturing field based on the obtained process documents, equipment manuals, report metadata, and historical query logs. Step S12: Call the pre-trained language model based on the BERT architecture, and perform incremental pre-training based on the pre-trained language model in the semiconductor field, on the basis of pre-training in general Chinese corpus, to obtain the pre-trained language model in the semiconductor field. Step S13: Perform WordPiece word segmentation on the received natural language query request input by the user to obtain the word segmentation results; Step S14: Input the word segmentation results into the pre-trained language model in the semiconductor field to obtain the context-related vector representation of the token and construct the semantic encoding matrix; Step S15: Calculate the sentence vector representation of the natural language query request based on the semantic encoding matrix.

[0007] To better realize the present invention, step S2 further includes the following steps: Step S21: Input the semantic encoding matrix into the bidirectional LSTM layer of the BiLSTM-CRF model to obtain the bidirectional hidden state, and call the Viterbi algorithm to obtain the entity of the natural language query request; Step S22: Input the sentence vector representation into the constructed attention mechanism multi-task learning model, calculate the intent probability distribution, and obtain the intent type; Step S23: Input the semantic encoding matrix into the pointer network to obtain the slot values; Step S24: Obtain the structured query intent based on the entity, intent type, and slot value.

[0008] To better implement the present invention, step S21 further includes the following steps: Step S211: Construct a set of entity types in the semiconductor manufacturing field based on time, process, machine, product, batch, indicator, anomaly type, and report name; Step S212: Input the semantic encoding matrix into the bidirectional LSTM layer of the BiLSTM-CRF model to obtain the bidirectional hidden state; Step S213: Input the bidirectional hidden state into the CRF layer, and calculate the globally optimal label sequence based on the state transition matrix and emission matrix; Step S214: Call the Viterbi algorithm to decode the globally optimal label sequence, identify the entity of the natural language query request, and normalize it.

[0009] To better realize the present invention, step S22 further includes the following steps: Step S221: Construct a set of intent types based on query, comparison, statistics, trend, root cause analysis, and summary; Step S222: Input the sentence vector representation into the constructed attention mechanism multi-task learning model, and calculate the intent probability distribution through a fully connected layer and a Softmax function; Step S223: If the intent probability distribution is greater than or equal to the set intent classification threshold, output the intent type; if the intent probability distribution is less than the set intent classification threshold, trigger a clarification dialogue and return the candidate intent list.

[0010] To better realize the present invention, step S23 further includes the following steps: Step S231: Construct a set of slots; the set of slots includes time slots, process slots, machine slots, product slots, batch slots, indicator slots, dimension slots, and comparison object slots; Step S232: Input the semantic encoding matrix into the pointer network and calculate the probability distribution of the starting position and the probability distribution of the ending position of the slot in the semantic encoding matrix; Step S233: Based on the slot set, use greedy decoding to obtain the optimal start and end positions of the slots, and extract the slot value text; Step S234: Map the slot value text to a standardized value, and provide a missing indication for unfilled slots based on the intent type and the required field rule.

[0011] To better realize the present invention, step S3 further includes the following steps: Step S31: Construct a query intent vector based on the structured query intent; Step S32: Construct a knowledge graph subgraph retrieval based on the query intent vector to obtain report nodes, and calculate the relevance score based on the set entity weights; Step S33: Match according to intent type and slot value domain rule base to obtain report set; Step S34: Perform availability verification on the report set, prioritize the reports that pass the verification, and obtain the report routing decision results.

[0012] Based on the aforementioned cognitive computing-based semiconductor production report routing decision method, and to better realize this invention, a cognitive computing-based semiconductor production report routing decision system is further proposed for executing the aforementioned cognitive computing-based semiconductor production report routing decision method; it includes an input unit, a graph construction unit, and an intent understanding unit; The input unit is used to input the acquired natural language query request into the constructed semiconductor domain pre-trained language model to obtain a semantic encoding matrix and calculate the sentence vector representation of the natural language query request. The graph construction unit is used to parse the sentence vector representation to obtain the entity, intent type, and slot value of the natural language query request, and to construct a structured query intent. The intent understanding unit is used to construct a query intent vector based on the structured query intent to obtain the report routing decision result.

[0013] Based on the aforementioned cognitive computing-based semiconductor production report routing decision method, to better realize this invention, an electronic device is further proposed, including a memory and a processor; the memory stores a computer program; when the computer program is executed on the processor, the aforementioned cognitive computing-based semiconductor production report routing decision method is implemented.

[0014] Based on the aforementioned cognitive computing-based semiconductor production report routing decision method, to better realize the present invention, a computer-readable storage medium is further proposed, wherein computer instructions are stored on the computer-readable storage medium; when the computer instructions are executed on the aforementioned electronic device, the aforementioned cognitive computing-based semiconductor production report routing decision method is realized.

[0015] The present invention has the following beneficial effects: This invention constructs a semiconductor cognitive knowledge graph covering device signal level, process parameter level, and production line yield level. The system can automatically associate report data scattered across different systems and realize multi-hop reasoning across reports.

[0016] This invention's natural language question-and-answer interface allows non-technical personnel such as production managers and quality engineers to complete complex report queries directly using everyday language without needing to learn SQL or BI tools.

[0017] (3) The present invention improves the analysis efficiency by using root cause path association reasoning based on knowledge graph. Attached Figure Description

[0018] Figure 1 A schematic flowchart of the semiconductor production report routing decision method based on cognitive computing provided by the present invention. Detailed Implementation

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments, and therefore should not be regarded as a limitation on the scope of protection. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set up," "connected," and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0021] Example 1: This embodiment proposes a semiconductor production report routing decision-making method based on cognitive computing, which specifically includes the following steps: Step S1: Input the acquired natural language query request into the constructed semiconductor domain pre-trained language model to obtain the semantic encoding matrix and calculate the sentence vector representation of the natural language query request; Step S1 specifically includes the following steps: Step S11: Construct a pre-trained corpus for the semiconductor manufacturing field based on the obtained process documents, equipment manuals, report metadata, and historical query logs. Step S12: Call the pre-trained language model based on the BERT architecture, and perform incremental pre-training based on the pre-trained language model in the semiconductor field, on the basis of pre-training in general Chinese corpus, to obtain the pre-trained language model in the semiconductor field. Step S13: Perform WordPiece word segmentation on the received natural language query request input by the user to obtain the word segmentation results; Step S14: Input the word segmentation results into the pre-trained language model in the semiconductor field to obtain the context-related vector representation of the token and construct the semantic encoding matrix; Step S15: Calculate the sentence vector representation of the natural language query request based on the semantic encoding matrix.

[0022] Step S2: Parse the sentence vector representation to obtain the entity, intent type, and slot value of the natural language query request, and construct a structured query intent; Step S2 specifically includes the following steps: Step S21: Input the semantic encoding matrix into the bidirectional LSTM layer of the BiLSTM-CRF model to obtain the bidirectional hidden state, and call the Viterbi algorithm to obtain the entity of the natural language query request; Step S21 specifically includes the following steps: Step S211: Construct a set of entity types in the semiconductor manufacturing field based on time, process, machine, product, batch, indicator, anomaly type, and report name; Step S212: Input the semantic encoding matrix into the bidirectional LSTM layer of the BiLSTM-CRF model to obtain the bidirectional hidden state; Step S213: Input the bidirectional hidden state into the CRF layer, and calculate the globally optimal label sequence based on the state transition matrix and emission matrix; Step S214: Call the Viterbi algorithm to decode the globally optimal label sequence, identify the entity of the natural language query request, and normalize it.

[0023] Step S22: Input the sentence vector representation into the constructed attention mechanism multi-task learning model, calculate the intent probability distribution, and obtain the intent type; Step S22 specifically includes the following steps: Step S221: Construct a set of intent types based on query, comparison, statistics, trend, root cause analysis, and summary; Step S222: Input the sentence vector representation into the constructed attention mechanism multi-task learning model, and calculate the intent probability distribution through a fully connected layer and a Softmax function; Step S223: If the intent probability distribution is greater than or equal to the set intent classification threshold, output the intent type; if the intent probability distribution is less than the set intent classification threshold, trigger a clarification dialogue and return the candidate intent list.

[0024] Step S23: Input the semantic encoding matrix into the pointer network to obtain the slot values; Step S23 specifically includes the following steps: Step S231: Construct a set of slots; the set of slots includes time slots, process slots, machine slots, product slots, batch slots, indicator slots, dimension slots, and comparison object slots; Step S232: Input the semantic encoding matrix into the pointer network and calculate the probability distribution of the starting position and the probability distribution of the ending position of the slot in the semantic encoding matrix; Step S233: Based on the slot set, use greedy decoding to obtain the optimal start and end positions of the slots, and extract the slot value text; Step S234: Map the slot value text to a standardized value, and provide a missing indication for unfilled slots based on the intent type and the required field rule.

[0025] Step S24: Obtain the structured query intent based on the entity, intent type, and slot value.

[0026] Step S3: Based on the structured query intent, construct the query intent vector to obtain the report routing decision result.

[0027] Step S3 specifically includes the following steps: Step S31: Construct a query intent vector based on the structured query intent; Step S32: Construct a knowledge graph subgraph retrieval based on the query intent vector to obtain report nodes, and calculate the relevance score based on the set entity weights; Step S33: Match according to intent type and slot value domain rule base to obtain report set; Step S34: Perform availability verification on the report set, prioritize the reports that pass the verification, and obtain the report routing decision results.

[0028] Working principle: This embodiment constructs a knowledge graph of semiconductor production reports, integrates the semantic understanding ability of a large language model with the reasoning ability of the knowledge graph, and realizes natural language-driven cross-report intelligent question answering and root cause analysis, achieving the effects of being usable by non-technical personnel, automatic cross-system association, and intelligent root cause analysis.

[0029] Example 2: This embodiment is based on the above embodiment 1, such as... Figure 1 As shown, a specific embodiment will be described in detail, which includes the following steps.

[0030] Step 1: Receive the user's natural language query request. Example: "Why did the yield of the lithography machine EE-01 drop yesterday?"; Step 2: Intent and entity resolution; Step 21: Semantic Understanding – Word Segmentation and Semantic Encoding Based on Domain Pre-trained Language Models; ① Construct a pre-trained corpus in the semiconductor field, collect process documents, equipment manuals, report metadata, and historical query logs in the semiconductor manufacturing field to form a corpus in the semiconductor field; ② A pre-trained language model based on the BERT architecture is adopted (the language model (such as Semiconductor-BERT) is pre-trained on general Chinese corpus (such as Wiki Chinese, Baidu Encyclopedia) and incrementally pre-trained using semiconductor domain corpus to learn the distributed representation of domain terms, thereby obtaining a semiconductor domain pre-trained language model as the basis for subsequent word segmentation and semantic encoding). ③ Perform WordPiece word segmentation on the natural language query request input by the user, and segment the word "Why did the yield of the lithography machine EE-01 decrease yesterday" into: yesterday / lithography process / machine / EE-01 / of / yield / why / decline; ④ Input the word segmentation results into the semiconductor domain pre-trained language model obtained in step ②, obtain the context-related vector representation of each token (i.e., the sub-word unit obtained by WordPiece segmentation in step ③), and form a semantic encoding matrix H∈R^(n×d), where n is the number of tokens and d is the hidden layer dimension (e.g., 768 dimensions). ⑤ Based on the semantic encoding matrix, calculate the sentence vector representation of the query request v_query=MeanPooling(H), which is used for subsequent intent classification and similarity calculation.

[0031] Step 22: Named Entity Recognition (NER) – Sequence Labeling Based on BiLSTM-CRF; ① Define the set of entity types in the semiconductor manufacturing field as E = {time, process, machine, product, batch, indicator, exception type, report name}, which consists of 8 types of entities; ② Sequence labeling using the BiLSTM-CRF model: Input the semantic encoding matrix H obtained in step 21.④ into the bidirectional LSTM layer. The forward LSTM captures context information from left to right, and the backward LSTM captures context information from right to left. The two layers are concatenated to obtain the bidirectional hidden state h_i=[h→_i;h←_i]; ③ Input the bidirectional hidden state into the CRF layer. The CRF layer models the dependency relationship between adjacent labels using the state transition matrix A (this matrix is ​​the model parameter learned by the BiLSTM-CRF model during the training phase using the backpropagation algorithm on the semiconductor domain labeled corpus, used to model the transition probability between adjacent labels) and calculates the globally optimal label sequence: P(y∣x)=Z(x)1exp(∑i=1n(Wyi+hi+Ayi−1,yi)); Where W is the emission matrix (obtained by linear transformation of the hidden state of the bidirectional LSTM layer, which is the parameter matrix learned during model training), A is the state transition matrix (i.e., the state transition matrix A in step ③, which is the parameter learned during model training), and Z(x) is the normalization factor. ④ The optimal label sequence is decoded using the Viterbi algorithm to identify the entities in the query request:

[0032] ⑤ Normalize the identified entities: normalize "yesterday" to the day before the current date (e.g., 2026-05-17), map "lithography process" to the standard process node ID in the knowledge graph (e.g., PROCESS_LITHO_001), and map "EE-01" to the standard machine code in the equipment master data table.

[0033] Step S2.3: Intent Classification – Joint Intent-Slot Model Based on Multi-Task Learning ① Define the intent type set I = {Query, Comparison, Statistics, Trend, Root Cause Analysis, Summary}, which includes 6 intent types; ② Construct a multi-task learning model based on the attention mechanism (this model includes an attention layer, a fully connected layer, and a softmax output layer, which is obtained through supervised training on annotated corpora in the semiconductor field using the cross-entropy loss function); ③ Introduce domain feature enhancement: Combine the entity types identified in step 22.④ as auxiliary feature inputs (such as feature codes of [time, process, machine, index]), concatenate them with sentence vectors, and perform intent classification together to improve the recognition accuracy of domain-specific intents; ④ Set the intent classification threshold θ=0.85: Input the augmented vector concatenated in step ③ into the multi-task learning model in step ②. The updated intent probability distribution is calculated through the attention layer, fully connected layer and Softmax function. If the highest intent probability P_max≥θ is taken, the intent type is output; if P_max<θ, a clarification dialogue is triggered, and a list of candidate intents is returned to the user for confirmation. ⑤ For the example query "Why did the yield of the lithography machine EE-01 decrease yesterday?", the model output intent type is "root cause analysis" and the confidence level is 0.94.

[0034] Step S2.4: Slot filling - Slot value extraction based on pointer network. The probability distribution of the position of the slot in the input sequence is calculated through the following steps ① to ⑤, the slot value text is decoded and mapped to the standardized value. ① Define a slot set S = {time slot, process slot, machine slot, product slot, batch slot, indicator slot, dimension slot, comparison object slot}, with a total of 8 slots; ② Use a pointer network for slot filling: Use the semantic encoding matrix H from step S2.1.④ as the input of the pointer network encoder. For each slot s_j, calculate its start position probability distribution α_start^j and end position probability distribution α_end^j in the input sequence. This probability distribution is used to determine the optimal start and end positions of the slot in step ③. ③ Based on the probability distribution of the starting position and the probability distribution of the ending position calculated in step ②, determine the optimal starting position and ending position of each slot through greedy decoding or beam search, and extract the slot value text; ④ Map the slot value text to a standardized value (this standardized value is used for slot resolution and inheritance in the subsequent step S2.5 dialogue state tracking, as the basis for entity matching in step S3.2 knowledge graph subgraph retrieval, and for slot value domain verification in step S3.3 intent-report mapping rule matching): Time slot "Yesterday" → "2026-05-17", Process slot "Lithography Process" → "PROCESS_LITHO_001", Machine slot "EE-01" → "EQP_LITHO_001", Index slot "Yield" → "YIELD_RATE"; ⑤ For unfilled slots, provide missing information based on intent type and required field rules: If the intent is "root cause analysis" and no time slot is provided, automatically fill in the default time range (e.g., the last 7 days) and indicate in the response "No time range specified, default query for the last 7 days of data".

[0035] Step 2.5: Context Management – ​​Multi-turn Dialogue Based on Dialogue State Tracking; ① Construct a dialogue state representation DS={dialogue ID, round number, historical intent sequence, historical slot value, currently focused entity, slot to be clarified}; ② Determine if the current query is the first query: Check if the dialogue ID exists. If it does not exist, create a new dialogue state DS_new and initialize the round number to 1. If it exists, read the existing dialogue state DS_existing and increment the round number by 1. ③ Perform pronoun resolution: If the current query contains pronouns (such as "it", "this batch", "this machine"), resolution is performed based on the currently focused entity in the historical dialogue state. The resolution rules are as follows: "It" → is dissolved into the indicator entity in the previous query; "This batch" → is resolved to the batch entity from the previous query; "This machine" → is resolved to the machine entity from the previous query; ④ Execution Intent Inheritance: If the current query is an abbreviated follow-up question (such as "Why", "What is the trend", "What about other machines"), then the historical intent type of the previous round is inherited, and only the changed slots are updated; ⑤ Slot resolution: If the current query and the historical query have the same slot type but different values, the value of the current query will be used first, and the historical values ​​will be stored in the historical slot value list for comparison and analysis. ⑥ Update the dialogue state: Write the current parsing results (intent type, slot value, focused entity) into the dialogue state and persist them to the dialogue state database (such as Redis) for use in the next round of queries.

[0036] Step S3: Cognitive Reasoning: Determine the Source of the Report Step S3.1: Construct the query intent vector – integrate the structured information output from step S2 ① Receive the structured query intent QI={intent type, entity set, slot value set} output from step S2, where: Intent type="root cause analysis" Entity set = {Time: 2026-05-17, Process: PROCESS_LITHO_001, Machine: EQP_LITHO_001, Index: YIELD_RATE} Slot value set = {Time slot: 2026-05-17, Process slot: PROCESS_LITHO_001, Machine slot: EQP_LITHO_001, Index slot: YIELD_RATE} ② Encode the intent type as an intent vector v_intent: Use One-Hot encoding or intent embedding matrix to map "root cause analysis" into a 128-dimensional dense vector; ③ Encode the entity set into an entity feature vector v_entity: For each entity type, obtain its embedding representation (such as TransE embedding) from the knowledge graph and concatenate them to form a joint entity feature vector; ④ Encode the set of slot values ​​into a slot feature vector v_slot: For each slot value, obtain its semantic vector through a domain pre-trained language model, and form a joint slot feature vector after MeanPooling; ⑤ The query intent vector v_query_fused = Concat(v_intent, v_entity, v_slot) is formed by merging the elements, which has a dimension of (128+256+256) = 640.

[0037] Construct a semiconductor cognitive knowledge graph (aligned with entity normalization in step S2.2): Extract entities such as equipment, processes, machines, indicators, and reports, as well as relationships such as attribution, association, and time sequence, from equipment manuals, process specifications, production report schemas, and historical work orders in the semiconductor manufacturing field. Construct a heterogeneous graph containing entity nodes, relationship edges, and attributes. Use the TransE graph embedding model to jointly train entities and relationships to obtain entity embedding vectors and report node vectors, and align them with the standard node IDs of the knowledge graph obtained from entity normalization in step S2.2. This knowledge graph serves as the underlying data for multi-hop neighbor retrieval in step 3.2, enabling semantic association across reports.

[0038] Step 3.2: Knowledge Graph Subgraph Retrieval – Relational Reasoning Based on Entity Matching; ① Using the entity in step S3.1.③ as the seed node, perform multi-hop neighbor retrieval in the semiconductor cognitive knowledge graph: First-hop retrieval: Starting from the seed node "YIELD_RATE" (yield rate indicator), retrieve directly related report nodes along the "association" edges: Yield metrics → (related) → Daily Yield Report (REPORT_YIELD_DAILY) Yield metrics → (related) → Yield trend report (REPORT_YIELD_TREND) Yield metrics → (Related) → Yield Comparison Report (REPORT_YIELD_COMPARE) Second hop search: Starting from "PROCESS_LITHO_001" (lithography process), search along the "containment" relationship: Photolithography process → (includes) → Photolithography equipment status report (REPORT_EAP_LITHO) Photolithography process → (includes) → Photolithography process parameter report (REPORT_MES_LITHO_PARAM) Third-hop search: Starting from "EQP_LITHO_001" (EE-01 machine), search along the "influence" relationship: EE-01 Machine → (Impact) → Equipment Overall Efficiency Report (REPORT_OEE_EQUIP) EE-01 Machine → (Impact) → Fault Detection and Classification Report (REPORT_FDC_ALARM) ② Obtain entity text from step S2.2 (named entity recognition), such as "yield". Search for the standard node ID corresponding to the entity in the knowledge graph entity database, such as "YIELD_RATE". Train the TransE knowledge graph embedding model to obtain an entity embedding vector ve corresponding to the entity.

[0039] ③ Define a set of report type nodes R = {REPORT_YIELD_DAILY, REPORT_EAP_LITHO, REPORT_MES_LITHO_PARAM,...} in the knowledge graph. Use the same TransE model and entity embedding for joint training. Report nodes participate in training as ordinary nodes in the knowledge graph to obtain the vector vr corresponding to each report node.

[0040] ④ Define entity weights we, with the highest entity weight we being 1.0, followed by process entities at 0.8, and machine entities at 0.6.

[0041] ⑤ Deduplicate and prioritize the retrieved report nodes: Sort them in descending order of relevance score to the query intent. The relevance score is calculated as follows: Score(r)=∑e∈seed_entitiessim(ve,vr)×we Step S3.3: Intent-Report Mapping Rule Matching – Report Routing Decision Based on Rule Engine ① Load predefined intents – report mapping rule base R, with rule format IF-THEN rule:

[0042] ② Match the intent type from step 3.1.② and the standardized slot value obtained from step S2.4.④ with the rule base: The current query intent type is "root cause analysis", the indicator type is "yield", and the matching rule is R001; ③ Extract the set of mandatory reports specified by rule R001: {Daily Yield Report, Equipment Status Report, Process Parameter Report, Anomaly Alarm Report}; ④ Take the intersection of the set of mandatory reports and the subgraph results of the knowledge graph retrieved in step 3.2.② to ensure that the source of the reports conforms to both the rule constraints and the semantic association of the knowledge graph.

[0043] Step S3.4: Report Availability Validation and Prioritization ① Perform availability verification on the report set determined in step S3.3.④: Data timeliness verification: Check whether the report data has been updated to the query time range (e.g., whether the data from 2026-05-17 has been synchronized). Permission verification: Checks whether the current user role has the right to access the report (via the permissions and security management module). Data integrity verification: Check whether there is any missing or abnormal data in the report within the query time range; ② Prioritize the reports that have passed the validation: P0 level (highest): Reports directly related to the query intent, such as the daily yield report (because the indicator is "yield"); P1 level (high): Reports strongly correlated with root cause analysis, such as equipment status reports and process parameter reports; P2 Level (Medium): Auxiliary verification reports, such as abnormal alarm reports; P3 Level (Low): Extends related reports, such as the historical case library; ③ Output report routing decision result RD={report ID, report name, data source system, priority, preload flag}:

[0044] Working Principle: This embodiment constructs a semiconductor cognitive knowledge graph covering equipment signal level, process parameter level, and production line yield level. The system can automatically associate report data scattered across different systems, achieving multi-hop reasoning across reports. Compared with existing single-report query methods, the coverage of cross-report association queries is increased from 35% to 95%, and the data association integrity is significantly improved. The natural language question-and-answer interface allows non-technical personnel such as production managers and quality engineers to complete complex report queries directly using everyday language without needing to learn SQL or BI tools. User availability reaches 100%, and query response time is reduced from an average of 15 minutes (manual operation) to ≤3 seconds. Root cause path association reasoning based on the knowledge graph compresses the traditional 4-6 hour manual investigation work into less than 1 minute. In the verification of actual production data, the root cause localization accuracy reaches 91%, which is more than 300 times more efficient than traditional manual analysis methods.

[0045] The other parts of this embodiment are the same as those in Embodiment 1 above, so they will not be described again.

[0046] Example 3: Based on any one of Embodiments 1-2 above, this embodiment proposes a semiconductor production report routing decision system based on cognitive computing, used to execute the above-mentioned semiconductor production report routing decision method based on cognitive computing; it includes an input unit, a graph construction unit, and an intent understanding unit; The input unit is used to input the acquired natural language query request into the constructed semiconductor domain pre-trained language model to obtain a semantic encoding matrix and calculate the sentence vector representation of the natural language query request. The graph construction unit is used to parse the sentence vector representation to obtain the entity, intent type, and slot value of the natural language query request, and to construct a structured query intent. The intent understanding unit is used to construct a query intent vector based on the structured query intent to obtain the report routing decision result.

[0047] This embodiment also proposes an electronic device, including a memory and a processor; the memory stores a computer program; when the computer program is executed on the processor, it implements the above-described semiconductor production report routing decision method based on cognitive computing.

[0048] This embodiment also proposes a computer-readable storage medium storing computer instructions; when the computer instructions are executed on the aforementioned electronic device, the aforementioned semiconductor production report routing decision method based on cognitive computing is implemented.

[0049] The other parts of this embodiment are the same as any one of the above embodiments 1-2, so they will not be described again.

[0050] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications or equivalent changes made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A semiconductor production report routing decision-making method based on cognitive computing, characterized in that, Specifically, the following steps are included: Step S1: Input the acquired natural language query request into the constructed semiconductor domain pre-trained language model to obtain the semantic encoding matrix and calculate the sentence vector representation of the natural language query request; Step S2: Parse the sentence vector representation to obtain the entity, intent type, and slot value of the natural language query request, and construct a structured query intent; Step S3: Based on the structured query intent, construct the query intent vector to obtain the report routing decision result.

2. The semiconductor production report routing decision method based on cognitive computing according to claim 1, characterized in that, Step S1 specifically includes the following steps: Step S11: Construct a pre-trained corpus for the semiconductor manufacturing field based on the obtained process documents, equipment manuals, report metadata, and historical query logs. Step S12: Call the pre-trained language model based on the BERT architecture, and perform incremental pre-training based on the pre-trained language model in the semiconductor field, on the basis of pre-training in general Chinese corpus, to obtain the pre-trained language model in the semiconductor field. Step S13: Perform WordPiece word segmentation on the received natural language query request input by the user to obtain the word segmentation results; Step S14: Input the word segmentation results into the constructed semiconductor domain pre-trained language model to obtain the context-related vector representation of the word segmentation results and construct the semantic encoding matrix; Step S15: Calculate the sentence vector representation of the natural language query request based on the semantic encoding matrix.

3. The semiconductor production report routing decision method based on cognitive computing according to claim 2, characterized in that, Step S2 specifically includes the following steps: Step S21: Input the semantic encoding matrix into the bidirectional LSTM layer of the BiLSTM-CRF model to obtain the bidirectional hidden state, and call the Viterbi algorithm to obtain the entity of the natural language query request; Step S22: Input the sentence vector representation into the constructed attention mechanism multi-task learning model, calculate the intent probability distribution, and obtain the intent type; Step S23: Input the semantic encoding matrix into the pointer network to obtain the slot values; Step S24: Obtain the structured query intent based on the entity, intent type, and slot value.

4. The semiconductor production report routing decision method based on cognitive computing according to claim 3, characterized in that, Step S21 specifically includes the following steps: Step S211: Construct a set of entity types in the semiconductor manufacturing field based on time, process, machine, product, batch, indicator, anomaly type, and report name; Step S212: Input the semantic encoding matrix into the bidirectional LSTM layer of the BiLSTM-CRF model to obtain the bidirectional hidden state; Step S213: Input the bidirectional hidden state into the CRF layer of the BiLSTM-CRF model, and calculate the globally optimal label sequence based on the state transition matrix and emission matrix; Step S214: Call the Viterbi algorithm to decode the globally optimal label sequence, identify the entity of the natural language query request, and normalize it.

5. A semiconductor production report routing decision method based on cognitive computing according to claim 3, characterized in that, Step S22 specifically includes the following steps: Step S221: Construct a set of intent types based on query, comparison, statistics, trend, root cause analysis, and summary; Step S222: Concatenate the entity and sentence vector representation of the normalized natural language query request to obtain the augmented vector; Step S223: Enhance the vector input to the constructed attention mechanism multi-task learning model, and calculate the intent probability distribution through fully connected layers and the Softmax function; Step S224: If the intent probability distribution is greater than or equal to the set intent classification threshold, output the intent type; if the intent probability distribution is less than the set intent classification threshold, trigger a clarification dialogue and return the candidate intent list.

6. The semiconductor production report routing decision method based on cognitive computing according to claim 5, characterized in that, Step S23 specifically includes the following steps: Step S231: Construct a set of slots; the set of slots includes time slots, process slots, machine slots, product slots, batch slots, indicator slots, dimension slots, and comparison object slots; Step S232: Input the semantic coding matrix into the pointer network encoder and calculate the probability distribution of the starting position and the probability distribution of the ending position of the slot in the semantic coding matrix; Step S233: Based on the probability distribution of the starting position and the probability distribution of the ending position, call the greedy decoding algorithm to obtain the optimal starting position and the optimal ending position of the slot, and extract the slot value text; Step S234: Map the slot value text to a standardized slot value, and provide a missing indication for unfilled slots based on the intent type and the required fill rule.

7. A semiconductor production report routing decision method based on cognitive computing according to claim 6, characterized in that, Step S3 specifically includes the following steps: Step S31: Construct a query intent vector based on the structured query intent; Step S32: Construct a semiconductor cognitive graph subgraph retrieval based on the query intent vector, obtain report nodes, and calculate the relevance score based on the set entity weights; Step S33: Match according to intent type and slot value domain rule base to obtain report set; Step S34: Perform availability verification on the report set, prioritize the reports that pass the verification, and obtain the report routing decision results.

8. A semiconductor production report routing decision system based on cognitive computing, used to execute the semiconductor production report routing decision method based on cognitive computing as described in claim 1; characterized in that, It includes an input unit, a graph construction unit, and an intent understanding unit; The input unit is used to input the acquired natural language query request into the constructed semiconductor domain pre-trained language model to obtain a semantic encoding matrix and calculate the sentence vector representation of the natural language query request. The graph construction unit is used to parse the sentence vector representation to obtain the entity, intent type, and slot value of the natural language query request, and to construct a structured query intent. The intent understanding unit is used to construct a query intent vector based on the structured query intent to obtain the report routing decision result.

9. An electronic device, characterized in that, It includes a memory and a processor; the memory stores a computer program; when the computer program is executed on the processor, it implements a semiconductor production report routing decision method based on cognitive computing as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions; when the computer instructions are executed on the electronic device as described in claim 9, they implement a semiconductor production report routing decision method based on cognitive computing as described in any one of claims 1-7.

Images