Abnormality control method and system for semiconductor manufacturing event

Abstract

The invention discloses an exception control method and system for a semiconductor manufacturing event, and relates to the technical field of exception control methods, and the method comprises the steps: determining a query matching coefficient according to the detection of each query path, and determining a query system based on each query matching coefficient, the to-be-queried content of a semiconductor and a semiconductor processing database; the multiple query channels of different dimensions are determined according to recognition of the query system, and the accuracy of the query system is improved. Therefore, a plurality of sub-repair items are determined according to the detection of the content repair event, and the repair process table is determined according to the plurality of sub-repair items, the importance level of the to-be-queried content of the semiconductor and the working state of the query system; and based on the repair process table, the corresponding repair node and the closed-loop control logic of the content repair event, determining the exception management and control system of the semiconductor manufacturing event, thereby improving the accuracy of the exception management and control system of the semiconductor manufacturing event.

Description

Technical Field

[0001] This invention relates to the technical field of anomaly control methods, and more particularly to an anomaly control method and system for semiconductor manufacturing events. Background Technology

[0002] With the development of technology, semiconductors are gradually being applied to electronic devices. Semiconductors are manufactured by semiconductor manufacturing equipment, and wafers are also a type of semiconductor. In the current technology, semiconductor workers log into the semiconductor database and output the corresponding query statement. The query statement is collected, and the content to be queried in the semiconductor is determined based on the identification of the query statement. However, the consideration of the query matching coefficient is ignored, which affects the accuracy of the query system and leads to a low accuracy of the abnormal control system for semiconductor manufacturing events. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art. This invention provides a method and system for anomaly control of semiconductor manufacturing events.

[0004] This invention provides an anomaly control method for semiconductor manufacturing events, including: Collect query statements output by semiconductor staff, and determine the content to be queried in the semiconductor field based on the recognition of the query statement; The query results are determined based on the semiconductor content to be queried and the semiconductor processing database; if the query results contain error information characteristics, multiple query paths are determined based on the semiconductor processing database and the semiconductor content to be queried. The query matching coefficient is determined based on the detection of each query path. The query system is determined based on the query matching coefficient, the query content of semiconductors, and the semiconductor processing database. Multiple query channels of different dimensions are determined based on the identification of the query system. Based on the identification of multiple query channels, the corresponding query health values ​​are matched. Based on the anomaly detection of the content to be queried in the semiconductor, multiple content anomaly features are determined. Based on the query health values ​​corresponding to each query channel, the corresponding query strategy, and multiple content anomaly features, the content repair event is determined. Multiple sub-repair projects are identified based on the detection of content repair events. A repair schedule is determined based on the multiple sub-repair projects, the importance level of the semiconductor content to be queried, and the working status of the query system. An anomaly control system for semiconductor manufacturing events is determined based on the repair schedule, the corresponding repair nodes, and the closed-loop control logic of the content repair events.

[0005] This invention provides an anomaly control system for semiconductor manufacturing events, which is applied to the aforementioned anomaly control method for semiconductor manufacturing events. The anomaly control system for semiconductor manufacturing events includes: The query content module is used to collect query statements output by semiconductor staff and determine the query content of semiconductors based on the recognition of the query statement. The query path module is used to determine the corresponding query results based on the query content of semiconductors and the semiconductor processing database; if the query results contain error information characteristics, multiple query paths are determined based on the semiconductor processing database and the query content of semiconductors. The query system module is used to determine the query matching coefficient based on the detection of each query path, and to determine the query system based on each query matching coefficient, the query content of semiconductors and the semiconductor processing database; and to determine multiple query channels of different dimensions based on the identification of the query system. The content repair event module is used to match the corresponding query health value based on the identification of multiple query channels, determine multiple content anomaly features based on the anomaly detection of the query content to be queried by semiconductors, and determine the content repair event according to the query health value corresponding to each query channel, the corresponding query strategy and multiple content anomaly features. The anomaly control system module is used to determine multiple sub-repair projects based on the detection of content repair events. Based on the multiple sub-repair projects, the importance level of the semiconductor content to be queried, and the working status of the query system, a repair schedule is determined. Based on the repair schedule, the corresponding repair nodes, and the closed-loop control logic of the content repair events, the anomaly control system for semiconductor manufacturing events is determined.

[0006] Compared with the prior art, the beneficial effects of the present invention are: In this embodiment of the invention, the method involves collecting query statements output by semiconductor workers, identifying the query statements to determine the semiconductor content to be queried, determining the corresponding query results based on the semiconductor content to be queried and the semiconductor processing database, identifying multiple query paths based on the semiconductor processing database and the semiconductor content to be queried, determining query matching coefficients based on the detection of each query path, and establishing a query system based on the query matching coefficients, the semiconductor content to be queried, and the semiconductor processing database. The identification of the query system determines multiple query channels of different dimensions, introducing multiple query paths, and considering the overall factors of various query matching coefficients, the semiconductor content to be queried, and the semiconductor processing database, thereby improving the accuracy of the query system and further controlling its operation.

[0007] Therefore, based on the identification of multiple query channels, corresponding query health values ​​are matched; based on the anomaly detection of the semiconductor's query content, multiple content anomaly features are determined; and content repair events are determined according to the query health values ​​corresponding to each query channel, the corresponding query strategies, and multiple content anomaly features. Multiple sub-repair projects are determined based on the detection of content repair events; and a repair timeline is determined based on these sub-repair projects, the importance level of the semiconductor's query content, and the working status of the query system. Based on this repair timeline, the corresponding repair nodes, and the closed-loop control logic of content repair events, an anomaly management system for semiconductor manufacturing events is determined. The introduction of content repair events further manages the repair timeline, achieving a holistic consideration of the repair timeline, the corresponding repair nodes, and the closed-loop control logic of content repair events, thus improving the accuracy of the anomaly management system for semiconductor manufacturing events. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating the abnormal control method for semiconductor manufacturing events in an embodiment of the present invention; Figure 2 This is a flowchart illustrating step S11 in the abnormal control method for semiconductor manufacturing events in an embodiment of the present invention. Figure 3 This is a flowchart illustrating step S12 in the abnormal control method for semiconductor manufacturing events in this embodiment of the invention. Figure 4 This is a flowchart illustrating step S13 in the abnormal control method for semiconductor manufacturing events in an embodiment of the present invention. Figure 5 This is a flowchart illustrating step S14 in the abnormal control method for semiconductor manufacturing events in this embodiment of the invention. Figure 6 This is a flowchart illustrating step S15 in the abnormal control method for semiconductor manufacturing events in an embodiment of the present invention. Figure 7 This is a schematic diagram of the structural composition of the abnormal control system for semiconductor manufacturing events in an embodiment of the present invention. Detailed Implementation

[0009] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0010] Please see Figures 1 to 7 An anomaly control method for semiconductor manufacturing events, applied to anomaly control scenarios; the anomaly control method for semiconductor manufacturing events includes: Step S11: Collect the query statement output by the semiconductor staff, and determine the content to be queried in the semiconductor field based on the identification of the query statement; Step S12: Determine the corresponding query results based on the semiconductor query content and the semiconductor processing database; if the query results contain error information characteristics, determine multiple query paths based on the semiconductor processing database and the semiconductor query content. Step S13: Determine the query matching coefficient based on the detection of each query path, determine the query system based on each query matching coefficient, the query content of semiconductors and the semiconductor processing database; determine multiple query channels of different dimensions based on the identification of the query system; Step S14: Match the corresponding query health value based on the identification of multiple query channels, determine multiple content anomaly features based on the anomaly detection of the content to be queried in the semiconductor, and determine the content repair event according to the query health value corresponding to each query channel, the corresponding query strategy and multiple content anomaly features. Step S15: Based on the detection of content repair events, determine multiple sub-repair projects. Based on the multiple sub-repair projects, the importance level of the semiconductor content to be queried, and the working status of the query system, determine the repair schedule. Based on the repair schedule, the corresponding repair nodes, and the closed-loop control logic of the content repair events, determine the abnormal control system for semiconductor manufacturing events.

[0011] refer to Figure 2 In step S11, the specific steps are as follows: S111: Real-time monitoring of semiconductor manufacturing events and collection of feedback statements from semiconductor workers on semiconductor manufacturing events. Based on the identification of the feedback statements and the semiconductor workers, the query statements output by the semiconductor workers are determined. These query statements cover the corresponding semiconductor manufacturing data in the semiconductor manufacturing process. S112: Based on the detection of the query statement, multiple sub-query segments are determined. Based on the identification of the multiple sub-query segments, multiple key manufacturing contents are determined. Based on the multiple key manufacturing contents, the corresponding content importance level, and the semiconductor process corresponding to the semiconductor manufacturing event, the semiconductor query content is determined.

[0012] In the embodiments of this application, semiconductor manufacturing events are monitored in real time, and feedback statements from semiconductor workers on these events are collected. Based on the identification of these feedback statements and the semiconductor workers' identification, a query statement output by the semiconductor workers is determined. This query statement covers the corresponding semiconductor manufacturing data during the semiconductor manufacturing process, and is compatible with the identification of feedback statements and the overall considerations of the semiconductor workers, ensuring the accuracy of the query statement output by the semiconductor workers.

[0013] At this point, by deeply integrating into the Manufacturing Execution System (MES), the system can capture various events occurring on the production line in real time, including every change in equipment status, the completion status of process steps, and the immediate recording of product defects. To ensure the timeliness and reliability of data transmission, event bus or message queue technology is used as the data transmission channel. In this way, even in a high-concurrency production environment, event data can be delivered to the analysis system in a timely and accurate manner, providing a solid data foundation for subsequent anomaly control.

[0014] After acquiring the raw event data, the system needs to collect the professional judgments and observations of semiconductor workers. To this end, the system provides a variety of convenient feedback channels, including a concise text input box for quick description, a voice recognition function to facilitate verbal feedback from engineers when they are busy, and an image annotation tool for visually marking the location of defects. The key is that each piece of feedback will be automatically associated with the corresponding event data, such as binding the engineer's comments with process parameters of specific equipment and specific time points to form a complete event context.

[0015] The system uses Natural Language Processing (NLP) technology to deeply analyze the feedback from staff, accurately extracting key information such as wafer ID, equipment number, anomaly type, and time range. More importantly, the system combines the identity information of the person providing the feedback (such as process engineer or equipment maintenance worker) and their work area to intelligently infer their potential query intent. Based on these analyses, the system automatically generates structured query statements and directly accesses relevant databases to obtain the required information.

[0016] Furthermore, multiple sub-query segments are determined based on the detection of the query statement, and multiple key manufacturing contents are determined based on the identification of the multiple sub-query segments. Based on the multiple key manufacturing contents, the corresponding content importance level, and the semiconductor process corresponding to the semiconductor manufacturing event, the semiconductor query content is determined. This approach takes into account the overall consideration of multiple key manufacturing contents, the corresponding content importance level, and the semiconductor process corresponding to the semiconductor manufacturing event, ensuring the accuracy of the semiconductor query content.

[0017] At this point, the system uses syntactic analysis technology to decompose the complete query statement into multiple clauses or phrases with independent functions. For example, it can clearly distinguish the SELECT clause representing the query target, the WHERE clause setting the filtering conditions, and the time phrases that limit the time range. Through this refined decomposition, the system can accurately identify the key information contained in each segment, such as what the user wants to query (wafer ID), under what conditions to query (specific device), and the time range of the query (last week), laying the foundation for subsequent semantic understanding.

[0018] After completing the structural decomposition, the system enters the semantic understanding stage. It maps each subquery segment to a specific semiconductor manufacturing entity based on the key information in each subquery segment. For example, it identifies "Etch_Tool_03" as a device ID and "wafer ID" as a product identifier. Furthermore, the system uses semantic analysis technology to understand the logical relationships between these manufacturing contents. For example, it can accurately understand that the time range "last week" is used to modify the action of "processing on Etch_Tool_03" rather than an isolated condition. This deep semantic association ensures that the system's understanding of the query intent is accurate.

[0019] After identifying all key manufacturing content, the system determines the query target based on a set of built-in priority rules. These rules comprehensively consider the importance level of the content (for example, when tracing issues, the wafer ID is usually more critical than the device ID) and the semiconductor process type to which the event belongs (e.g., different processes such as etching and deposition have different query focuses). The system can use a preset rule base or a trained machine learning model, combined with historical query data and expert knowledge, to intelligently optimize and sort the query content, ensuring that the generated query task best meets the user's actual needs.

[0020] Specifically, a process engineer's query request is initially transformed into a query statement by the system: "Query the wafer IDs processed on Etch_Tool_03 last week". Through syntactic analysis, the statement is broken down into four core sub-query segments: "query" (query target), "last week" (time range), "Etch_Tool_03" (equipment), and "wafer ID" (query target). Next, the system maps these segments to key manufacturing information: wafer ID, equipment ID (Etch_Tool_03), and time range (last week). When determining the query content, the system judges based on rules that the user's ultimate goal is to obtain a specific list, namely, "all wafer IDs processed using Etch_Tool_03 last week". Based on this clear query content, the system generates and executes the corresponding SQL query, quickly returning the data needed by the engineer.

[0021] refer to Figure 3 In step S12, the specific steps are as follows: S121: Based on the tracing of semiconductor manufacturing events, determine the corresponding semiconductor processing database, determine the corresponding content query space based on the traversal of the semiconductor processing database, and determine the corresponding query result based on the autonomous matching of the query content of the semiconductor and the corresponding content query space. S122: Based on the detection of the query result, determine the query error statement; based on the identification of the query error statement, determine the characteristics of each error message; based on the characteristics of each error message and the corresponding semiconductor manufacturing event, determine multiple query methods for the semiconductor processing database; based on the matching of multiple query methods and the semiconductor query content, determine multiple query paths.

[0022] In the embodiments of this application, the corresponding semiconductor processing database is determined based on the tracing of semiconductor manufacturing events, the corresponding content query space is determined based on the traversal of the semiconductor processing database, and the corresponding query result is determined based on the autonomous matching of the semiconductor's query content and the corresponding content query space. This approach takes into account the overall consideration of the autonomous matching of the semiconductor's query content and the corresponding content query space, ensuring the accuracy of the corresponding query results.

[0023] At this point, the system does not blindly search through a vast ocean of data, but relies on a pre-built global data asset catalog and event-data source mapping model. When a specific manufacturing event is received (such as "processing event on Etch_Tool_03"), a dedicated traceability engine is immediately activated. This engine parses the type of event—whether it is equipment processing, quality inspection, or material flow—and quickly locates the physical or logical database storing the core data of the event based on the mapping model. For example, the data for equipment processing events is mainly stored in the Manufacturing Execution System (MES) database, while measurement data is stored in the Quality Management System (QMS) database. This mechanism ensures that the query finds the right place from the beginning, greatly improving the efficiency of subsequent steps.

[0024] After identifying the target database, it connects to the database's metadata store, automatically parses and loads the database's table structure, field definitions, primary and foreign key constraints, and the lineage relationships between tables. Based on this metadata and the previously defined "query content" (such as wafer ID and device ID), the system dynamically constructs a logical query space. This space is not a physical entity, but a virtual data view and relationship graph tailored to this query. It precisely outlines the data tables, key fields, and their relationships involved in this query, clearly defining the boundaries for the next step of precise querying.

[0025] The system utilizes a semantic mapping and query generation engine to perform high-precision matching between natural language or structured "query content" and the newly constructed "content query space." For example, the engine converts "last week" into a specific timestamp range using a built-in date parser; it precisely maps "Etch_Tool_03" to the tool_id field of the process_log table; and it maps "wafer ID" to the wafer_id field of the wafer_info table. After completing all mappings, the engine automatically generates an optimized structured query statement (such as SQL), submits it to the database for execution, and retrieves the query results.

[0026] Specifically, in response to the query "Wafer IDs processed on Etch_Tool_03 last week", the system identified this as a "device processing event". After the tracing engine queries the mapping model, it quickly pinpointed the target to the MES database. The system connected to the metadata of the MES database and found that the process_log table recorded all processing events, including key fields such as tool_id, wafer_id, and process_timestamp. Based on this, the system constructed a logical query space with the process_log table as its core.

[0027] The semantic mapping engine begins to work: "last week" is parsed as the specific date range of the process_timestamp field; "Etch_Tool_03" is mapped as the filter condition tool_id='Etch_Tool_03'; "wafer ID" is determined as the wafer_id field that needs to be returned; the query generation engine constructs and executes an SQL query based on this, and uses the list of wafer IDs returned from the database as the final result of step S121.

[0028] Furthermore, based on the detection of the query results, query error statements are identified, and based on the identification of the query error statements, the characteristics of each error message are determined. Based on the characteristics of each error message and the corresponding semiconductor manufacturing events, multiple query methods for the semiconductor processing database are determined. Based on the matching of multiple query methods and the semiconductor query content, multiple query paths are determined, which takes into account the overall consideration of matching multiple query methods and the semiconductor query content, and ensures the accuracy of multiple query paths.

[0029] At this point, the system will perform a comprehensive "check-up" on the query results returned by S121. This goes far beyond simply checking whether the results are empty; it covers multiple dimensions, including execution status (such as database error codes and connection timeouts), result set analysis (whether the number of rows is zero, whether it is far beyond expectations, and the percentage of NULL values), and performance metrics (query time). If any metric is abnormal, the system will immediately mark the executed query statement along with its context (original problem, database structure, etc.) as a "query error statement" and activate the subsequent in-depth diagnostic process.

[0030] After marking errors, the system performs in-depth analysis, extracting quantified error characteristics and proceeding to root cause analysis. This includes using an Abstract Syntax Tree (AST) to analyze the syntax and structural features of the SQL, detecting whether there are field reference errors; comparing the values ​​in the WHERE condition with the actual data distribution of the target field to find mismatches between the value range and the encoding features (e.g., using the natural language word "complete" to match a field storing numeric codes); analyzing the logical association features of multi-table JOINs to determine whether there are missing join conditions; and comparing the SQL logic with the original natural language intent to identify deviations between semantic and intent features.

[0031] Based on the extracted detailed error features, the system will match multiple feasible "query methods" from a dynamic query strategy knowledge base. This knowledge base encapsulates a large amount of expert experience and best practices for dealing with the complexity of semiconductor data. For example, if the error feature is a value range mismatch, the system will suggest query methods such as "enabling domain terminology-encoding conversion table" or "performing fuzzy matching". If the error is a missing logical association, it will suggest "introducing an intermediate entity table for multi-table join" or "using a subquery instead of JOIN". If it is a data timing problem, it will recommend "window query based on event flow".

[0032] The system combines these abstract "query methods" with the explicit "query content" in S11, instantiating them into one or more executable query paths. Each path represents a data-driven potential solution to bypass the current error. The system generates an independent and complete query execution plan for each path, evaluating and selecting subsequent steps, thereby achieving an intelligent transformation from "passive error reporting" to "actively exploring multiple possibilities".

[0033] Specifically, the system detected that the number of rows in the result set was 0, which was a clear anomaly signal. It packaged the original query, the generated SQL, and the empty result set and marked it as a "query error statement". The system began to identify the error characteristics. Through step-by-step conditional analysis, it found that querying the device or time individually yielded results, but the result was empty when both conditions were present. After further investigating the process_log table, the system found that in the record of Etch_Tool_03 last week, the value of the process_status field was the code 'C' (Complete), while the system's default query logic implicitly included a filter condition of process_status='FINISHED'. Therefore, the error characteristics were precisely located as: "low value range health" (status code mismatch) and "implicit logical conflict".

[0034] Based on these characteristics, the system matches two query methods from the strategy library: the first is "value range correction", which ignores the status code and directly queries based on the device ID and timestamp; the second is "logical extension", which not only queries the process log table, but also cross-validates with the device status log table.

[0035] These two query methods are instantiated into two specific query paths: Path 1 generates a new SQL statement that removes all filter conditions related to process_status; Path 2 generates a more complex SQL statement that uses the JOIN of the equipment_status_log table to filter out the wafer IDs within the time period when both logs are confirmed as "processing".

[0036] refer to Figure 4 In step S13, the specific steps are as follows: S131: In multiple query paths, multiple query nodes are determined based on the identification of each query path, and the node position and node result of each query node are marked. The query matching coefficient of the query path is determined according to the node position, node result and the semiconductor query content, so as to collect multiple query matching coefficients. S132: Determine the content deviation range based on each query matching coefficient and the content to be queried in the semiconductor. Determine multiple query deviation paths based on the content deviation range, the corresponding deviation level, and the semiconductor processing database. Construct a query system based on the multiple query deviation paths, the content to be queried in the semiconductor, and the semiconductor processing database. This query system serves as the optimized query part and is adapted to the query of the content to be queried in the semiconductor. S133: Determine the optimized query path based on the detection of the query system, determine the main query path and multiple secondary query paths based on the identification of the optimized query path, and determine multiple query channels of different dimensions based on the main query path, multiple secondary query paths and semiconductor processing database.

[0037] In the embodiments of this application, multiple query nodes are determined based on the identification of each query path in multiple query paths, and the node position and node result of each query node are marked. The query matching coefficient of the query path is determined according to the node position, node result and the content to be queried of the semiconductor, so as to collect multiple query matching coefficients, which takes into account the overall consideration of the node position, node result and the content to be queried of the semiconductor, and ensures the accuracy of the query matching coefficient of the query path.

[0038] At this point, the system will break down each complex query path into an execution sequence consisting of multiple "query nodes". Each node is the smallest atomic operation unit in the path, which can be an independent SQL query, a data cleaning function, or an API call. The system will mark the logical position of each node in the path, such as "starting node", "intermediate association node" or "final aggregation node", and this position determines its importance. Subsequently, the system will pre-execute or simulate execution for each node to obtain its predicted node results, including key indicators such as the number of returned data rows, field integrity, execution time, and resource consumption.

[0039] After completing the structured decomposition, the system uses a multi-dimensional weighted evaluation model to calculate the comprehensive score of each path—the query matching coefficient. This model integrates information from multiple dimensions into a precise value between 0 and 1. The evaluation dimensions include: semantic relevance score, which compares the matching degree between the path output and the "query content"; data quality score, which evaluates the completeness and accuracy of the data based on the node results; execution efficiency score, which evaluates the total estimated time and resource consumption of the path; and logical complexity score, which evaluates the stability and maintenance difficulty of the path.

[0040] The weight of the results of nodes at different locations on the final score also varies. By weighting and summing the scores of these dimensions, the system generates a comprehensive query matching coefficient for each query path. The closer the coefficient is to 1, the better the overall performance of the path.

[0041] Specifically, S122 generates two paths for "querying wafer IDs processed on Etch_Tool_03 last week"; Path 1: Ignore status codes and directly query the process_log table; Path 2: Join the process_log table and the equipment_status_log table to cross-validate the equipment processing time period; For Path 1, it is decomposed into two nodes: Node 1 (starting node) queries the process_log table, returns 500 rows of data, and takes 50ms; Node 2 (final aggregation node) filters out 500 wafer IDs, which takes 10ms; For Path 2, it is decomposed into three nodes: Node 1 and Node 2 query the two log tables respectively, and Node 3 (final aggregation node) performs a JOIN operation on them, finally returning 450 wafer IDs, with a total time of 280ms.

[0042] The system calculates the query matching coefficient: For path one, its semantic relevance is high, the data quality is moderate (analysis shows it contains 50 test wafers), the execution efficiency is high, and the logical complexity is low; after comprehensive calculation, the query matching coefficient is 0.85; For path two, its semantic relevance is high, the data quality is high (all 450 IDs are valid production wafers), but the execution efficiency is moderate and the logical complexity is high; after comprehensive calculation, the query matching coefficient is 0.88.

[0043] Through the quantitative evaluation of S131, the system obtained precise scores for the two paths. Although path one is faster, path two obtained a higher overall score due to its superior data quality. This precise coefficient provides a solid decision-making basis for subsequent steps to determine how to optimize and combine paths, avoiding the blindness of making choices based solely on intuition.

[0044] Furthermore, based on the matching coefficients of each query and the content to be queried in the semiconductor, the content deviation range is determined. Based on the content deviation range, the corresponding deviation level, and the semiconductor processing database, multiple query deviation paths are determined. Based on the multiple query deviation paths, the content to be queried in the semiconductor, and the semiconductor processing database, a query system is constructed. This query system serves as the optimized query component and is adapted to the query of the content to be queried in the semiconductor. It takes into account the overall consideration of the content deviation range, the corresponding deviation level, and the semiconductor processing database, ensuring the accuracy of the multiple query deviation paths.

[0045] At this point, the system performs a distribution analysis on all query matching coefficients generated by S131. If the highest coefficient is lower than the preset "excellent" threshold (e.g., 0.9), or if the coefficients of each path differ significantly, the system determines that no existing path can perfectly satisfy the "query content" and there is a significant content deviation. In order to accurately locate the deviation, the system performs difference and intersection analysis on the "node results" of high and low coefficient paths to quantify the specific link where the deviation occurs. For example, the deviation range is precisely described as: "Path A contains 50 more wafer IDs with a status of 'TEST' than path B".

[0046] After identifying the deviation range, the system will activate a deviation compensation engine to generate a correction plan. The engine assesses the severity of the deviation, i.e., the deviation level (for example, "containing test data" is a medium deviation, while "missing critical production data" is a high deviation). Based on the deviation level and specific content, a new "query deviation path" specifically for correcting or compensating for the deviation is generated from the policy library. For example, for the deviation "containing test data", the system generates a new query path: "Query the wafer attribute table, obtain all wafer types, and generate a 'list of valid production wafer IDs' for filtering."

[0047] The system logically combines and arranges the original query path with the newly generated query deviation path to construct a complete query system. This system is no longer linear, but a directed acyclic graph (DAG) or decision tree structure. It clearly defines the logical relationships and calling order between the main query, backup query, verification query, and compensation query. For example, the system can be defined as: "Execute deviation path A to obtain the filter list, pass the list as input to the original path B for querying, and output the results." This constructed system is saved as an optimized and reusable query module, which can dynamically adapt to similar query needs in the future.

[0048] Specifically, S131 calculated the matching coefficients for the two paths as follows: Path 1 (ignoring status codes) 0.85, Path 2 (associated with device logs) 0.88; the system found that the highest coefficient was only 0.88, which is lower than the excellent threshold of 0.9, and determined that there was a content deviation; by comparing the results of the two paths, the system accurately located the deviation as: "Path 1 contains 50 wafer IDs that are not found in Path 2, and the common attribute of these IDs is that the wafer_type field is 'TEST'"; the deviation level was rated as "medium".

[0049] To address this discrepancy, the deviation compensation engine is activated and generates a path to query the deviation: "Query the wafer_attribute table, filter out all wafer IDs with wafer_type 'PROD' (production), and generate a temporary dataset of 'valid production wafer IDs'."

[0050] The system combines the original path 2 with the newly generated deviation path to construct a query system. The logic of this system is defined as follows: "Execute 'Query Deviation Path' to obtain a temporary dataset of 'Valid Production Wafer IDs'; use this dataset as input and pass it to the original path 2, add a filter condition to the JOIN operation to ensure that the wafer_id in the process_log table must exist in this temporary dataset; execute the optimized query and return the result."

[0051] Therefore, the optimized query path is determined based on the detection of the query system. The main query path and multiple secondary query paths are determined based on the identification of the optimized query path. Based on the main query path, multiple secondary query paths, and the semiconductor processing database, multiple query channels of different dimensions are determined. This approach takes into account the overall considerations of the main query path, multiple secondary query paths, and the semiconductor processing database, ensuring the accuracy of multiple query channels of different dimensions. At the same time, the introduction of multiple query paths takes into account the overall considerations of various query matching coefficients, the semiconductor query content, and the semiconductor processing database, improving the accuracy of the query system and further controlling the query system.

[0052] At this point, the system performs static analysis on the "query system" built by S132 to check for logical deadlocks or performance bottlenecks in the data flow and dependencies. Subsequently, dynamic simulation is performed in a sandbox environment to verify the logical correctness and collaborative working ability of the system using small sample data. After this series of rigorous tests, the system will solidify one or more complete and optimal "optimized query paths" from the system. These paths are specific instantiations of the system and are reliable solutions that can be directly deployed to the production environment.

[0053] After obtaining all optimized paths, the system assigns them different roles based on a primary and secondary path decision-making algorithm. Typically, the path with the highest overall matching coefficient, the best data quality, and the most direct logic is designated as the "primary query path," which bears the primary responsibility of obtaining the core and most reliable query results. Other paths are defined as "secondary query paths," which play important auxiliary roles, such as: acting as "degradation paths," automatically taking over when the primary path fails due to failure or excessive load to ensure service availability; acting as "validation paths," cross-validating the results of the primary path; or acting as "supplementary paths," providing additional contextual information.

[0054] The system encapsulates primary and secondary query paths into channels based on their functions, data sources, and execution timing. Each query channel is an independent, schedulable, and monitorable execution unit, encapsulating its own query logic, data source connection, and exception handling mechanism. Channels are divided according to different business dimensions, such as: a "core data channel" based on the primary path, responsible for acquiring highly reliable production data; a "quality verification channel" based on the verification path, acquiring data from systems such as QMS for cross-validation; and a "high availability channel" based on the degradation path, ensuring that basic services can still be provided when the primary data source is unavailable. These channels can be designed to execute in parallel, with a result aggregation engine integrating the data from each channel to form a comprehensive and highly reliable query output.

[0055] Specifically, the system performs static analysis and dynamic simulation on the architecture to confirm that its logic is correct and its performance is acceptable, and then solidifies it into a complete optimized query path. The system designates this optimized path as the primary query path because it can provide the most accurate and purest results. At the same time, the system retains the original "Path 1 (ignore status code, query directly)" in S122 as a secondary query path, whose role is defined as a "degradation path" for use in extreme cases such as timeouts caused by multi-table JOINs in the primary path. The system encapsulates these two paths into channels to construct two parallel query channels.

[0056] The primary data channel (high-precision channel) executes the optimized primary query path, retrieving a verified, high-precision list of wafer IDs from multiple tables in the MES database. This is the default and preferred execution channel. The secondary query channel (high-availability channel) executes the secondary query path, querying only the `process_log` table in the MES database. It monitors the execution status of the primary data channel; if the primary channel times out or fails, the secondary channel is immediately triggered, returning a noisy but basically usable result to ensure uninterrupted service. Ultimately, the system deploys an intelligent query system as an industrial-grade query execution solution with primary and secondary criteria, fault tolerance, and a multi-dimensional perspective. This not only guarantees the accuracy of the query results but also ensures the high availability and robustness of the entire query service.

[0057] refer to Figure 5 In step S14, the specific steps are as follows: S141: In multiple query channels, multiple query events are determined based on the traversal of each query channel, and the corresponding result accuracy coefficient is determined based on the identification of each query event. The corresponding query health value is matched according to the event type of each query event, the corresponding result accuracy coefficient and the query strategy of the query channel, so as to collect the query health value corresponding to each query channel. S142: Based on the identification of the query content of semiconductors, multiple sub-query contents are determined, and the corresponding abnormal statements are determined based on the detection of multiple sub-query contents. Multiple content abnormal features are determined based on each abnormal statement and the corresponding sub-query contents. S143: Collect the query strategies corresponding to each query channel, determine the first-level repair coefficient based on the query strategies corresponding to each query channel and multiple content anomaly features, determine the second-level repair coefficient based on the query health values ​​corresponding to each query channel and multiple content anomaly features, and determine the content repair event based on the mapping relationship between the first-level repair coefficient, the second-level repair coefficient and the content repair event.

[0058] In the embodiments of this application, multiple query events are determined based on the traversal of each query channel in multiple query channels, and the corresponding result accuracy coefficient is determined based on the identification of each query event. The corresponding query health value is matched according to the event type of each query event, the corresponding result accuracy coefficient and the query strategy of the query channel, so as to collect the query health value corresponding to each query channel. This takes into account the overall consideration of the identification of each query event and ensures the accuracy of the corresponding result accuracy coefficient.

[0059] At this point, the system deploys an "event probe" during the execution of each query channel. This probe decomposes the entire query lifecycle into a series of discrete, observable query events, such as join events, parsing events, data scan events, aggregation events, and result return events.

[0060] For each event, the system calculates a "result accuracy coefficient" based on preset business rules and performance baselines. This coefficient, between 0 and 1, measures the degree to which the output of the event matches the "ideal state." For example, for a data scanning event, if the actual number of rows scanned is within 10% of the estimated error, its accuracy coefficient is 1.0; if the error exceeds 50%, it drops to 0.5.

[0061] After obtaining the accuracy coefficients of all events, the system uses a multi-dimensional weighted evaluation model to calculate the "query health value". The unique feature of this model is that it is dynamic and adjusts the weight of different event types according to the "query strategy" (such as "high accuracy", "high availability" and "low latency") defined in S133 for each query channel.

[0062] For high-precision channels, events related to data integrity and logical correctness (such as logical consistency of multi-table joins) are given the highest weight; for high-availability channels, events related to connection stability and execution speed are more critical; the system multiplies the "result accuracy coefficient" of each event by its corresponding "strategy weight", adds up all the weighted scores, and finally obtains a comprehensive health value between 0 and 1, which serves as the final quantitative representation of the channel's health status.

[0063] Specifically, the system uses event probes to identify each query event and calculates its accuracy coefficient: For the primary data channel, the data scan event was estimated to scan 1000 rows, but actually scanned 1050 rows, with an error of 5% and an accuracy coefficient of 0.95; the multi-table join event was successful but was discarded due to some data mismatch, with an accuracy coefficient of 0.90; the result return event met expectations, with an accuracy coefficient of 1.0; for the degraded channel, both the data scan and result return events were executed perfectly, with an accuracy coefficient of 1.0 for both.

[0064] The system matches weights and calculates health values ​​according to their respective strategies: For the main data channel (high-precision strategy), the weight of multi-table association events is the highest (0.5) because it is directly related to the accuracy of the results; its health value is calculated as: (0.95×0.3)+(0.90×0.5)+(1.0×0.2)=0.935; For the degradation channel (high-availability strategy), faster return is more important, so the weight of result return events is higher (0.6); its health value is calculated as: (1.0×0.4)+(1.0×0.6)=1.0.

[0065] The system obtained health values ​​for two channels; although the results for the main data channel were more accurate, its health value was slightly lower than 1.0 due to minor data loss during the association operation; while the degradation channel performed perfectly and had a full score in health. These precise values ​​provided crucial quantitative input for subsequent steps to determine whether there was "content anomaly" and whether "repair" was needed.

[0066] Furthermore, multiple sub-query contents are determined based on the identification of the query content of semiconductors, corresponding abnormal statements are determined based on the detection of multiple sub-query contents, and multiple content abnormal features are determined based on each abnormal statement and its corresponding sub-query content. This approach takes into account the overall consideration of each abnormal statement and its corresponding sub-query content, ensuring the accuracy of multiple content abnormal features.

[0067] At this point, the system uses a built-in semantic parsing and knowledge graph to break down the "query content" determined in S11 into multiple finer-grained, independently detectable sub-query contents. This process is not just about splitting fields, but also about a deep understanding of their inherent business meaning. For example, the requirement to "query the wafer IDs processed on Etch_Tool_03 last week" can be intelligently decomposed by the system into: a list of wafer IDs as the core entity, a sequence of processing timestamps as the time dimension, device IDs as the spatial dimension, and the relationship between preceding and subsequent processes as the process flow dimension.

[0068] After completing the content decomposition, the system equips each subquery with a business rule engine and a statistical anomaly detection model built on knowledge of the semiconductor manufacturing field and historical data. The business rule engine includes a series of hard constraints, such as "wafer IDs must be unique" and "processing time must be within the shift". The statistical anomaly detection model is responsible for identifying data that does not conform to the usual pattern, such as judging whether the number of wafers being processed is abnormal by learning from historical data. When a subquery violates the rules or is judged as abnormal by the model, the system will generate a human-readable, descriptive "abnormal statement" that clearly points out the problem.

[0069] The system abstracts and characterizes the generated "abnormal statements," extracting their core concepts and mapping them to a standardized content abnormality feature tag library. The purpose of this step is to transform specific, descriptive abnormalities into features that machines can understand, process, and reuse. For example, the abnormal statement "Duplicate items exist in the wafer ID list" will be abstracted into "Entity uniqueness violation feature," while "Processing timestamp exceeds the planned maintenance window" will be abstracted into "Temporal logic conflict feature." These standardized features are key inputs for subsequent repair decisions.

[0070] Specifically, the main data channel successfully returned 450 wafer IDs, and S141 calculated their health value to be 0.935. The system decomposed the query results into multiple sub-contents for inspection, focusing on the "list of 450 wafer IDs" and "their corresponding processing timestamp sequences".

[0071] The system begins detecting and identifying anomalous statements: Inspection of the wafer ID list reveals a jump in the sequence, skipping directly from WAFER_100 to WAFER_102, omitting WAFER_101; the system generates anomalous statement 1: "A non-continuous gap was detected in the wafer ID sequence between WAFER_100 and WAFER_102, with WAFER_101 suspected to be missing." Analysis of the timestamp sequence reveals that most wafer processing time is concentrated during the day shift, but one wafer, WAFER_200, has a processing timestamp of 3:00 AM on Sunday; after querying the equipment calendar, the system finds that Etch_Tool_03 is under planned maintenance during that time period; the system generates anomalous statement 2: "The processing timestamp of wafer WAFER_200 conflicts with the equipment's planned maintenance time window, indicating a timing logic contradiction."

[0072] The system abstracts the above abnormal statements and determines the content anomaly characteristics: Abnormal statement 1 is mapped to content anomaly characteristic 1: "Discontinuous key entity sequence characteristic"; Abnormal statement 2 is mapped to content anomaly characteristic 2: "Entity-device timing logic conflict characteristic". Although the system confirms that the query is successful at the execution level, it has found two potentially serious problems at the business content level. These two standardized "content anomaly characteristics" will be passed to subsequent steps as the core basis for deciding whether to start the repair process and how to repair it. This perfectly reflects the system's deep insight from "technical correctness" to "business rationality".

[0073] Therefore, the query strategies corresponding to each query channel are collected, and the first-level repair coefficient is determined based on the query strategies corresponding to each query channel and multiple content anomaly characteristics. The second-level repair coefficient is determined based on the query health values ​​corresponding to each query channel and multiple content anomaly characteristics. The content repair event is determined based on the mapping relationship between the first-level repair coefficient, the second-level repair coefficient and the content repair event. This approach takes into account the overall consideration of the mapping relationship between the first-level repair coefficient, the second-level repair coefficient and the content repair event, ensuring the accuracy of the content repair event.

[0074] At this point, the system employs a strategy-feature impact assessment matrix, which defines the severity level of each "content anomaly feature" under different "query strategies" (such as "high precision" and "high availability"). For example, "discontinuous data sequence" is fatal for a "high precision" strategy that strives for perfection, but may be tolerable for a "high availability" strategy that only requires "rough availability". The system queries this matrix, assigns a severity score to each anomaly feature under the current strategy, and calculates a weighted average of all scores to obtain the first-level repair coefficient. The lower this coefficient, the more severe the conflict between the anomaly and the strategy, and the higher the necessity for repair.

[0075] The system determines a second-level repair coefficient based on the query health values ​​of each query channel and multiple content anomaly characteristics. It employs a health-anomaly correlation model to assess the correlation between "content anomaly" and "channel execution health status." If a channel's query health value is already low, and a content anomaly is added, the problem is more systemic, and the urgency of repair is higher. Conversely, if the channel health is high, the content anomaly is an isolated, accidental data issue. Based on this correlation, the system adjusts the first-level repair coefficient to derive a second-level repair coefficient, which more accurately reflects the priority of repair.

[0076] Based on the mapping relationship between the first-level repair coefficient, the second-level repair coefficient, and the content repair event, the system determines the content repair event. At this point, the system merges the two repair coefficients to generate a comprehensive repair assessment value. By querying a predefined "assessment value-repair event" mapping knowledge base, the system determines the final "content repair event" to be triggered. This knowledge base defines the repair actions corresponding to different assessment value ranges. For example: a high assessment value triggers the "no repair needed" event; a medium assessment value triggers the "data marking" event, adding a warning to the result; a low assessment value triggers the "data compensation" event, attempting to automatically correct the data; and an extremely low assessment value triggers the "query reconstruction and alarm" event, abandoning the current result and escalating the problem.

[0077] Specifically, the health value of the main data channel (high-precision strategy) was 0.935, and two content anomaly features were found: "discontinuous key entity sequence feature" and "entity-device timing logic conflict feature". The system determined the first-level repair coefficient. For the "high-precision strategy", the severity score of the sequence discontinuity was 0.2, while the severity score of the timing logic conflict was 0.1. After weighted averaging, the first-level repair coefficient was 0.15. This extremely low value indicates that from a strategy perspective, these anomalies are fatal.

[0078] The system determines the second-level repair coefficient. The system analysis believes that "temporal logic conflict" is one of the reasons why the health value is not the full score of 1.0, and there is a certain correlation. Therefore, the first-level coefficient is fine-tuned to slightly reduce the urgency of repair. Assuming that after the adjustment, the second-level repair coefficient is 0.25, this value is still very low, indicating that the priority of repair is very high.

[0079] The system makes a fusion decision, assuming equal weights, and the overall repair assessment value is 0.20. The system queries the "assessment value - repair event" mapping knowledge base and finds that 0.20 falls in the range of ≤0.30. Therefore, the system finally determines to trigger the "query reconstruction and alarm" content repair event. This means that the system will abandon the current result with serious content defects and escalate the problem to S15, start a more complex repair process, and notify relevant personnel that the data queried from the main data channel has serious business logic problems.

[0080] refer to Figure 6 In step S15, the specific steps are as follows: S151: Collect content repair events, determine multiple repair contents of different repair levels based on the detection of content repair events, determine the corresponding sub-repair items according to the identification of each repair content, and collect multiple sub-repair items. S152: Collect the query content of the semiconductor and mark the importance level of the query content of the semiconductor. Determine the first repair combination based on multiple sub-repair items and the importance level of the query content of the semiconductor. At the same time, collect the working status of the query system and determine the repair schedule based on multiple sub-repair items, the working status of the query system and the first repair combination. S153: Based on the identification of the repair history table, multiple repair nodes are identified. The first level of anomaly control combination is determined according to the node position and corresponding node content of the multiple repair nodes. At the same time, the closed-loop control logic of the content repair event is collected. The second level of anomaly control combination is determined based on the closed-loop control logic of the content repair event and the node content of the multiple repair nodes. The anomaly control system for semiconductor manufacturing events is constructed based on the first level of anomaly control combination and the second level of anomaly control combination.

[0081] In the embodiments of this application, content repair events are collected, and multiple repair contents of different repair levels are determined based on the detection of content repair events. The corresponding sub-repair items are determined according to the identification of each repair content, so as to collect multiple sub-repair items, which takes into account the overall consideration of the identification of each repair content and ensures the accuracy of the corresponding sub-repair items.

[0082] At this point, the system receives a "content repair event" (such as "query reconstruction and alarm") from S143. Based on the severity of the event and its potential impact on the business, it assigns a repair level (such as P0-urgent, P1-high, P2-medium, P3-low) through an event severity assessment model. Subsequently, the system queries a repair event decomposition knowledge base to map the event into a set of high-level, function-oriented "repair content". This content describes several key aspects that need to be repaired, forming the top-level architecture of the repair work, such as "data tracing and cross-validation", "query logic root cause analysis and reconstruction", and "anomaly reporting and knowledge accumulation".

[0083] After determining the top-level architecture of the repair content, the system will further refine it into a series of specific, atomic, and independently executable "sub-repair projects". Each sub-project clearly defines its inputs, outputs, execution standards, and acceptance conditions. This is a deepening process from "what to do" to "how to do it". For example, the repair content of "data tracing and cross-validation" can be refined into several specific sub-repair projects such as "starting the degradation channel query", "performing a traceability query on the original logs of the device", and "comparing third-party data and generating a difference analysis report".

[0084] Specifically, the system collects the event; since this query involves critical production data, the system uses an evaluation model to classify its repair level as P1 (high); the system queries the knowledge base and decomposes this P1 level event into three repair contents: A. Data tracing and cross-validation; B. Query logic root cause analysis and reconstruction; C. Anomaly reporting and knowledge accumulation.

[0085] The system refines each repair task, generating a task list containing multiple sub-repair projects: For repair task A, the system generates sub-projects A1 (parallel execution of degradation channel query), A2 (initiating a traceability query to the device log database), and A3 (performing set operations to identify differences in three-party data); For repair task B, the system generates sub-projects B1 (analyzing and locating SQL logic fragments with timing conflicts), B2 (analyzing and locating SQL logic fragments with discontinuous entity sequences), and B3 (rewriting SQL based on the analysis results to generate new query statements); For repair task C, the system generates sub-projects C1 (encapsulating standardized anomaly case reports) and C2 (calling the alarm service API to push reports to relevant personnel).

[0086] Furthermore, the query content of the semiconductor is collected and its importance level is marked. The first repair combination is determined based on the importance levels of multiple sub-repair projects and the query content of the semiconductor. At the same time, the working status of the query system is collected. The repair schedule is determined based on the multiple sub-repair projects, the working status of the query system, and the first repair combination. This comprehensive consideration of multiple sub-repair projects, the working status of the query system, and the first repair combination ensures the accuracy of the repair schedule.

[0087] At this point, the system obtains the original "query content" and assigns it an importance level (e.g., Critical, High, Medium) using a business impact assessment model. This level directly determines the depth and breadth of the remediation strategy. For example, for "Critical" level queries, the system will select the most comprehensive and thorough combination of sub-remediation projects, sparing no effort to ensure the final accuracy of the data. For "Medium" level queries, the system will select a lightweight combination, prioritizing service availability and skipping some time-consuming in-depth analysis projects. This process forms a targeted "first remediation combination".

[0088] Once the scope of the repair is determined, the system acts like an intelligent job scheduler, collecting real-time data on the operational status of the "query system," including system resources such as CPU, memory, and I / O, as well as business load information such as the number of active database connections and query queues. Based on this status information and the "first repair combination," the system generates a detailed and dynamic repair timeline. This timeline is far more than just a task list; it is an execution plan that includes a timeline, resource allocation, and dependencies. It defines the execution order and dependencies of sub-projects using a Directed Acyclic Graph (DAG), employs predictive scheduling algorithms to schedule resource-intensive tasks during periods of low system load, and allocates resource quotas and fault tolerance strategies to each sub-project to ensure that the repair process itself does not affect normal business queries.

[0089] Specifically, the system has generated 7 sub-repair projects, while the original query "Query wafer IDs processed on Etch_Tool_03 last week" has been marked as "Critical". The system has determined the first repair combination. Since the query level is "Critical", the system has included all 7 sub-repair projects into the first repair combination according to the strategy to ensure the thoroughness of the repair. The system collects the current working status of the query system and finds that it is currently the peak production period of the day shift, and the system load is high. Based on this status, the system has formulated the following repair schedule: At time T0 (immediately): Start sub-projects A1 (degradation channel query) and A2 (device log trace) in parallel, because these two queries are relatively independent and put less pressure on the main database; Time T1 (after A1 and A2 are completed): Immediately start sub-project A3 (data difference analysis), which is an in-memory computing task that puts no pressure on the database; Time T2 (after A3 is completed): Immediately start sub-projects B1 and B2 (root cause analysis), which are CPU-intensive tasks, but the preceding tasks have already released resources by this time; At time T3 (after B1 and B2 are completed): immediately start sub-projects B3 (generate new SQL) and C1 (generate report), which can be executed in parallel; At time T4 (after B3 and C1 are completed): immediately start sub-project C2 (send alarm); Meanwhile, the system has set up a fault tolerance strategy such as a 5-minute timeout for sub-project A2 to prevent a single point of failure from blocking the entire repair process.

[0090] Therefore, multiple repair nodes are identified based on the repair timeline. A first-level anomaly control combination is determined based on the node positions and corresponding node contents of these nodes. Simultaneously, the closed-loop control logic for content repair events is collected. Based on this closed-loop control logic and the node contents of the multiple repair nodes, a second-level anomaly control combination is determined. An anomaly control system for semiconductor manufacturing events is constructed based on both the first and second-level anomaly control combinations. This system incorporates the closed-loop control logic of content repair events and the overall consideration of the node contents of multiple repair nodes, ensuring the accuracy of the second-level anomaly control combination. Furthermore, the introduction of content repair events further controls the repair timeline, achieving an overall consideration of the repair timeline, corresponding repair nodes, and the closed-loop control logic of content repair events, thus improving the accuracy of the anomaly control system for semiconductor manufacturing events.

[0091] At this point, the system creates each sub-repair project in the "Repair Process Schedule" generated by S152 as a "repair node" that can be managed by a state machine. Each node has a clear logical order, execution content, and status (such as pending execution, running, success, failure). Based on the position and content of all nodes, the system constructs a "first-level exception control combination," which is essentially a workflow orchestration engine configuration. It defines the flow rules between nodes, including sequential flow, parallel flow, conditional branch flow, and exception sub-processes triggered when a node fails, ensuring the orderliness and robustness of the repair process.

[0092] The system extracts its closed-loop control logic from the definition of "content repair event" in S143, such as "update the knowledge base if the repair is successful, and escalate the alarm if the repair fails". The system binds this high-level logic with the specific content of each repair node to form a "second-level anomaly control combination". This is essentially a configuration of a feedback control loop. It defines how to automatically update the knowledge base when a key node succeeds, how to trigger different compensation measures according to the error type when a node fails, and how the entire repair process should be stored as a case in the case library for future learning, regardless of success or failure.

[0093] The system deeply integrates the first layer of control, representing "workflow orchestration," with the second layer of control, representing "feedback control loop," to build a complete and automated semiconductor manufacturing event anomaly control system. This system is a living, autonomous intelligent agent that can not only automatically execute and monitor the workflow in real time, but also dynamically adjust the execution path according to the feedback loop. Through closed-loop control logic, it transforms the experience of each repair into system knowledge, enabling continuous learning and self-evolution.

[0094] Specifically, the system has established a repair schedule comprising 7 sub-projects; the system has determined the first layer of anomaly control combination; it instantiates the 7 sub-projects into 7 repair nodes (N1 to N7), and constructs a workflow based on the schedule: parallel startup (N1, N2) > N1, N2 successful startup N3 > N3 successful startup (N4, N5) > N4, N5 successful startup (N6, N7); at the same time, the system is configured with anomaly sub-processes, for example, if N2 (device log trace) times out and fails, its status is marked as "failed", but the workflow is allowed to continue.

[0095] The system determines the second layer of anomaly control combination; it extracts the closed-loop logic of the "query reconstruction and alarm" event and binds the logic of "optimizing the query system if the repair is successful" to node N6 (generating new SQL). This means that when node N6 is successfully executed, the system will automatically trigger a feedback operation, writing the newly generated SQL statement, its applicable query mode, and the anomaly characteristics of this repair into the "query system knowledge base" of S13.

[0096] The system integrates workflow and feedback loop, and the anomaly control system is formally built and launched. It begins to automatically execute the repair process. When N6 successfully generates new SQL, the system not only completes the repair, but also ensures that the S13 query system will directly use this verified and optimized new path the next time any user initiates a similar query, thereby fundamentally preventing the recurrence of similar anomalies.

[0097] Please see Figure 7 , Figure 7This is a schematic diagram of the structural composition of the abnormal control system for semiconductor manufacturing events in an embodiment of the present invention; the abnormal control system for semiconductor manufacturing events includes: The query content module 21 is used to collect the query statements output by semiconductor staff and determine the query content of semiconductor based on the recognition of the query statements. The query path module 22 is used to determine the corresponding query results based on the query content of semiconductors and the semiconductor processing database; if the query results have error information characteristics, multiple query paths are determined based on the semiconductor processing database and the query content of semiconductors. The query system module 23 is used to determine the query matching coefficient based on the detection of each query path, and to determine the query system based on each query matching coefficient, the query content of semiconductors and the semiconductor processing database; and to determine multiple query channels of different dimensions based on the identification of the query system. The content repair event module 24 is used to match the corresponding query health value based on the identification of multiple query channels, determine multiple content anomaly features based on the anomaly detection of the query content to be queried by semiconductors, and determine the content repair event according to the query health value corresponding to each query channel, the corresponding query strategy and multiple content anomaly features. The anomaly control system module 25 is used to determine multiple sub-repair projects based on the detection of content repair events, determine a repair schedule based on multiple sub-repair projects, the importance level of the semiconductor content to be queried, and the working status of the query system, and determine the anomaly control system for semiconductor manufacturing events based on the repair schedule, the corresponding repair nodes, and the closed-loop control logic of the content repair events.

[0098] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for anomaly control in semiconductor manufacturing events, characterized in that, include: Collect query statements output by semiconductor staff, and determine the content to be queried in the semiconductor field based on the recognition of the query statement; The query results are determined based on the content to be queried about semiconductors and the semiconductor processing database. If the query results contain error information, multiple query paths are determined based on the semiconductor processing database and the semiconductor content to be queried. The query matching coefficient is determined based on the detection of each query path, and the query system is determined based on each query matching coefficient, the query content of semiconductors, and the semiconductor processing database. Based on the identification of the query system, multiple query channels of different dimensions are determined; Based on the identification of multiple query channels, the corresponding query health values ​​are matched. Based on the anomaly detection of the content to be queried in the semiconductor, multiple content anomaly features are determined. Based on the query health values ​​corresponding to each query channel, the corresponding query strategy, and multiple content anomaly features, the content repair event is determined. Multiple sub-repair projects are identified based on the detection of content repair events. A repair schedule is determined based on the multiple sub-repair projects, the importance level of the semiconductor content to be queried, and the working status of the query system. An anomaly control system for semiconductor manufacturing events is determined based on the repair schedule, the corresponding repair nodes, and the closed-loop control logic of the content repair events.

2. The method for anomaly control of semiconductor manufacturing events according to claim 1, characterized in that, The process of collecting query statements output by semiconductor personnel and determining the target query content for semiconductors based on the identification of these query statements includes: Real-time monitoring of semiconductor manufacturing events and collection of feedback statements from semiconductor workers on these events; identification of these feedback statements and determination of the query statements output by the semiconductor workers, which cover the corresponding semiconductor manufacturing data during the semiconductor manufacturing process. Based on the detection of the query statement, multiple sub-query segments are determined. Based on the identification of multiple sub-query segments, multiple key manufacturing contents are determined. Based on multiple key manufacturing contents, their corresponding content importance levels, and the semiconductor process corresponding to the semiconductor manufacturing event, the semiconductor query content is determined.

3. The method for anomaly control of semiconductor manufacturing events according to claim 1, characterized in that, The query results are determined based on the semiconductor-based query content and the semiconductor processing database. If the query results contain error message characteristics, multiple query paths are determined based on the semiconductor processing database and the semiconductor content to be queried, including: The corresponding semiconductor processing database is determined by tracing semiconductor manufacturing events, the corresponding content query space is determined by traversing the semiconductor processing database, and the corresponding query result is determined by the autonomous matching of the query content of the semiconductor and the corresponding content query space. Based on the detection of the query results, query error statements are identified. Based on the identification of the query error statements, the characteristics of each error message are determined. Based on the characteristics of each error message and the corresponding semiconductor manufacturing events, multiple query methods for the semiconductor processing database are determined. Based on the matching of multiple query methods and the semiconductor query content, multiple query paths are determined.

4. The method for anomaly control of semiconductor manufacturing events according to claim 1, characterized in that, The query matching coefficient is determined based on the detection of each query path, and the query system is determined based on each query matching coefficient, the query content of the semiconductor, and the semiconductor processing database; Based on the identification of the query system, multiple query channels of different dimensions are determined, including: In multiple query paths, multiple query nodes are identified based on the identification of each query path, and the node position and node result of each query node are marked. The query matching coefficient of the query path is determined based on the node position, node result and the semiconductor query content, so as to collect multiple query matching coefficients. Based on the matching coefficients of each query and the content to be queried in the semiconductor field, the content deviation range is determined. Based on the content deviation range, the corresponding deviation level, and the semiconductor processing database, multiple query deviation paths are determined. Based on the multiple query deviation paths, the content to be queried in the semiconductor field, and the semiconductor processing database, a query system is constructed. This query system serves as the optimized query part and is adapted to the query of the content to be queried in the semiconductor field.

5. The method for anomaly control of semiconductor manufacturing events according to claim 4, characterized in that, The query matching coefficient is determined based on the detection of each query path, and the query system is determined based on each query matching coefficient, the query content of the semiconductor, and the semiconductor processing database; Based on the identification of the query system, multiple query channels of different dimensions are determined, including: The optimized query path is determined based on the detection of the query system. The main query path and multiple secondary query paths are determined based on the identification of the optimized query path. Multiple query channels of different dimensions are determined based on the main query path, multiple secondary query paths and semiconductor processing database.

6. The method for anomaly control of semiconductor manufacturing events according to claim 1, characterized in that, The process involves identifying and matching corresponding query health values ​​based on multiple query channels, determining multiple content anomaly features based on anomaly detection of the query content in the semiconductor, and determining content repair events based on the query health values ​​corresponding to each query channel, the corresponding query strategy, and the multiple content anomaly features, including: In multiple query channels, multiple query events are determined based on the traversal of each query channel, and the corresponding result accuracy coefficient is determined based on the identification of each query event. The corresponding query health value is matched according to the event type of each query event, the corresponding result accuracy coefficient and the query strategy of the query channel, so as to collect the query health value corresponding to each query channel. Based on the identification of the query content in semiconductors, multiple sub-query contents are determined. Based on the detection of multiple sub-query contents, corresponding abnormal statements are determined. Based on each abnormal statement and its corresponding sub-query content, multiple content abnormal features are determined.

7. The method for anomaly control of semiconductor manufacturing events according to claim 6, characterized in that, The process of matching corresponding query health values ​​based on the identification of multiple query channels, determining multiple content anomaly features based on anomaly detection of the query content in the semiconductor, and determining content repair events based on the query health values ​​corresponding to each query channel, the corresponding query strategy, and the multiple content anomaly features also includes: Collect the query strategies corresponding to each query channel, determine the first-level repair coefficient based on the query strategies corresponding to each query channel and multiple content anomaly features, determine the second-level repair coefficient based on the query health values ​​corresponding to each query channel and multiple content anomaly features, and determine the content repair event based on the mapping relationship between the first-level repair coefficient, the second-level repair coefficient and the content repair event.

8. The method for anomaly control of semiconductor manufacturing events according to claim 1, characterized in that, The process involves identifying multiple sub-repair projects based on the detection of content repair events, determining a repair timeline based on these sub-repair projects, the importance level of the semiconductor's queried content, and the operational status of the query system, and establishing an anomaly control system for semiconductor manufacturing events based on this repair timeline, corresponding repair nodes, and closed-loop control logic for content repair events. This system includes: Collect content repair events, determine multiple repair contents of different repair levels based on the detection of content repair events, and determine the corresponding sub-repair items according to the identification of each repair content, so as to collect multiple sub-repair items; Collect the query content of the semiconductor and mark the importance level of the query content of the semiconductor. Determine the first repair combination based on multiple sub-repair items and the importance level of the query content of the semiconductor. At the same time, collect the working status of the query system and determine the repair schedule based on multiple sub-repair items, the working status of the query system and the first repair combination.

9. The method for anomaly control of semiconductor manufacturing events according to claim 8, characterized in that, The process of determining multiple sub-repair projects based on the detection of content repair events, determining a repair timeline based on these sub-repair projects, the importance level of the semiconductor's queried content, and the operational status of the query system, and determining an anomaly control system for semiconductor manufacturing events based on this repair timeline, corresponding repair nodes, and closed-loop control logic for content repair events, further includes: Multiple repair nodes are identified based on the repair timeline. A first-level anomaly control combination is determined based on the node positions and corresponding node contents of the multiple repair nodes. At the same time, the closed-loop control logic of the content repair event is collected. A second-level anomaly control combination is determined based on the closed-loop control logic of the content repair event and the node contents of the multiple repair nodes. An anomaly control system for semiconductor manufacturing events is constructed based on the first-level and second-level anomaly control combinations.

10. An abnormal control system for semiconductor manufacturing events, characterized in that, The abnormal control system for semiconductor manufacturing events is applied to the abnormal control method for semiconductor manufacturing events as described in any one of claims 1-9, wherein the abnormal control system for semiconductor manufacturing events includes: The query content module is used to collect query statements output by semiconductor staff and determine the query content of semiconductors based on the recognition of the query statement. The query path module is used to determine the corresponding query results based on the query content of semiconductors and the semiconductor processing database; if the query results contain error information characteristics, multiple query paths are determined based on the semiconductor processing database and the query content of semiconductors. The query system module is used to determine the query matching coefficient based on the detection of each query path, and to determine the query system based on each query matching coefficient, the query content of semiconductors and the semiconductor processing database; and to determine multiple query channels of different dimensions based on the identification of the query system. The content repair event module is used to match the corresponding query health value based on the identification of multiple query channels, determine multiple content anomaly features based on the anomaly detection of the query content to be queried by semiconductors, and determine the content repair event according to the query health value corresponding to each query channel, the corresponding query strategy and multiple content anomaly features. The anomaly control system module is used to determine multiple sub-repair projects based on the detection of content repair events. Based on the multiple sub-repair projects, the importance level of the semiconductor content to be queried, and the working status of the query system, a repair schedule is determined. Based on the repair schedule, the corresponding repair nodes, and the closed-loop control logic of the content repair events, the anomaly control system for semiconductor manufacturing events is determined.

Images