Method, system and equipment for analyzing abnormal phenomenon of semiconductor structure

Abstract

The invention provides a method, a system and equipment for analyzing an abnormal phenomenon of a semiconductor structure. The method comprises the steps of obtaining a query request in a natural language text form; performing named entity recognition on the query request to obtain entity words associated with wafer manufacturing and appearing in the query request; calling a semantic disambiguation model, and performing semantic disambiguation on the entity word according to the query request based on a context sensing mechanism to obtain a disambiguation feature corresponding to the entity word; calculating semantic similarity between the disambiguation characteristics and each pre-stored field, determining a target field according to the semantic similarity, and calling each piece of resume information associated with wafer manufacturing based on the target field; and inputting each piece of called resume information into the event evolution linked list learning model according to a time sequence, identifying resume information associated with the abnormal phenomenon, and generating an abnormal resume information sequence. According to the method, the semantic gap problem of a natural language and a database structure is effectively improved, and the fault tracing efficiency and accuracy during wafer manufacturing are greatly improved.

Description

Technical Field

[0001] This application relates to the field of semiconductor manufacturing technology, and in particular to a method, system, and apparatus for analyzing anomalies in semiconductor structures. Background Technology

[0002] In the semiconductor manufacturing field, diverse process data exists, and this critical process data is scattered across multiple heterogeneous platforms such as Manufacturing Execution Systems (MES), Statistical Process Control Systems (SPCS), and Advanced Process Control Systems (APSCS). Users often need to rely on complex database queries to perform cross-system queries. However, this approach demands a high level of mastery of database structure and syntax, making it difficult for frontline engineers to efficiently complete data extraction and analysis tasks. Furthermore, there is a significant semantic mismatch between users' natural language query requirements and database field naming. Traditional methods mostly rely on fixed templates and single databases, making them ill-suited for complex manufacturing scenarios. Moreover, current methods largely only support static data retrieval, lacking the ability to dynamically model and deeply analyze process history, thus failing to accomplish critical tasks such as equipment anomaly localization or process root cause tracing. Therefore, there is a need to provide a method, system, and device for analyzing anomalies in semiconductor structures. Summary of the Invention

[0003] This application provides a method, system, device, and medium for analyzing abnormal phenomena in semiconductor structures, in order to improve the technical problem in the existing technology that, due to the mismatch between natural language queries and database structures, it is impossible to accurately and quickly determine the relevant information causing wafer abnormalities during the wafer manufacturing process.

[0004] This application provides a method for analyzing abnormal phenomena in semiconductor structures. The method includes: acquiring a query request in the form of natural language text; wherein the query request describes an abnormal phenomenon generated during wafer manufacturing; performing named entity recognition on the query request to obtain entity words in the query request that are associated with wafer manufacturing; invoking a semantic disambiguation model, based on a context-aware mechanism, performing semantic disambiguation on the entity words according to the query request to obtain disambiguation features corresponding to the entity words; wherein the semantic disambiguation model is a multilayer perceptron model; calculating the semantic similarity between the disambiguation features and pre-stored fields, thereby determining the target field, and retrieving various historical information related to wafer manufacturing based on the target field; inputting the retrieved historical information into an event evolution linked list learning model in chronological order to identify the historical information associated with the abnormal phenomenon and generate an abnormal historical information sequence; wherein the event evolution linked list learning model is a time-series model.

[0005] In one embodiment of this application, the invocation of the semantic disambiguation model, based on a context-aware mechanism, performs semantic disambiguation on entity words according to the query request to obtain disambiguation features corresponding to the entity words; wherein, the semantic disambiguation model is a multilayer perceptron model, including: performing word segmentation on the query request to obtain multiple segmented texts; performing semantic encoding on each segmented text to obtain corresponding semantic encoding features, and arranging the semantic encoding features in sequence to form a semantic encoding matrix; for each entity word: inputting the semantic encoding matrix into the semantic disambiguation model, and based on the context-aware mechanism, determining the interaction features between the entity word and other segmented texts in the semantic encoding matrix, and using them as the disambiguation features corresponding to the entity word.

[0006] In one embodiment of this application, the step of semantically encoding each segmented text to obtain corresponding semantic encoding features, and arranging the semantic encoding features in sequence to form a semantic encoding matrix, includes: semantically encoding each segmented text to obtain semantic encoding features corresponding to each segmented text; arranging the semantic encoding features in sequence to form an initial semantic encoding matrix; and performing global average pooling on the initial semantic encoding matrix to obtain the final semantic encoding matrix.

[0007] In one embodiment of this application, the step of inputting the semantic encoding matrix into the semantic disambiguation model and determining the interaction features between the entity word and other segmented texts in the semantic encoding matrix based on a context-aware mechanism, and using these features as the disambiguation features corresponding to the entity word, includes: inputting the semantic encoding matrix into the vertical fusion module of the semantic disambiguation model; performing vertical fusion of the semantic encoding features of all segmented texts in each column of the semantic encoding matrix based on a multilayer perceptron to obtain a vertically fused semantic encoding matrix; for each entity word: extracting the semantic encoding features at the corresponding position of the entity word from the vertically fused semantic encoding matrix; and inputting the semantic encoding features into the horizontal fusion module of the semantic disambiguation model to fuse the elements of different columns of the semantic encoding features to obtain the disambiguation features of the entity word.

[0008] In one embodiment of this application, the step of calculating the semantic similarity between the disambiguation feature and each pre-stored field, determining the target field accordingly, and retrieving the various history information associated with the wafer manufacturing based on the target field includes: calculating the semantic similarity between the disambiguation feature and each pre-stored field, determining the target field and its corresponding database table name accordingly, and generating multiple database query statements; and retrieving the history information in the corresponding database based on each database query statement.

[0009] In one embodiment of this application, the step of calculating the semantic similarity between the disambiguation feature and each pre-stored field, thereby determining the target field and its corresponding database table name, and generating multiple database query statements includes: for each disambiguation feature: calculating the semantic similarity between the disambiguation feature and each field in the field library, selecting the field with a similarity greater than a preset similarity threshold as the target field, and determining the database table name corresponding to the target field; and generating the database query statement corresponding to the disambiguation feature based on a preset template engine according to the target field and its corresponding database table name.

[0010] In one embodiment of this application, the time series model is a Mamba model. The step of inputting each retrieved history information into the event evolution linked list learning model according to the time order, identifying the history information associated with the abnormal phenomenon, and generating an abnormal history information sequence includes: inputting each history information into the Mamba model according to the time order, detecting whether the input history information is abnormal, and marking the history information as an abnormal history node when it is abnormal; after all history information has been detected, counting all abnormal history nodes, and arranging the corresponding abnormal history information according to the time order to form an abnormal history information sequence.

[0011] In one embodiment of this application, for each piece of resume information, the steps of inputting the resume information into the Mamba model, detecting whether the input resume information is abnormal, and marking the resume information as an abnormal resume node when it is abnormal include: inputting the current resume information into the backbone network of the Mamba model, updating the state of the current resume information according to the pre-stored historical hidden states, obtaining and saving the hidden state corresponding to the current resume information; wherein, if there is no historical hidden state, the historical hidden state is preset to the initial value; inputting the hidden state corresponding to the current resume information into the classifier of the Mamba model, detecting whether the current resume information is abnormal, and marking the resume information as an abnormal resume node when it is abnormal.

[0012] This application also provides an analysis system for anomalies in semiconductor structures. The system includes: a query request acquisition module for acquiring query requests in natural language text form, wherein the query request describes anomalies occurring during wafer fabrication; an entity word recognition module for performing named entity recognition on the query request to obtain entity words in the query request that are associated with wafer fabrication; and a semantic disambiguation module for invoking a semantic disambiguation model, based on a context-aware mechanism, performing semantic disambiguation on entity words according to the query request to obtain disambiguation features corresponding to the entity words; wherein the semantic disambiguation model is multi-layered. The system includes: a perceptron model; a statement generation module for calculating the semantic similarity between disambiguation features and pre-stored fields, determining the target field and its corresponding database table name, and generating multiple database query statements; a history information extraction module for retrieving history information from the corresponding database based on each database query statement; and an event chain generation module for inputting all retrieved history information into an event evolution linked list learning model in chronological order, identifying history information associated with the abnormal phenomenon based on the temporal relationship between each history information, and generating an abnormal history information sequence; wherein the event evolution linked list learning model is a temporal model.

[0013] This application also provides an electronic device, including: one or more processors; and a storage device for storing one or more programs, which, when executed by one or more processors, cause the electronic device to perform an analysis method for abnormal phenomena of the semiconductor structure described above.

[0014] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a computer's processor, causes the computer to perform an analysis method for abnormal phenomena of any of the aforementioned semiconductor structures.

[0015] The beneficial effects of this application are as follows: The method, system, and device for analyzing anomalies in semiconductor structures proposed in this application have the following unexpected effects: By acquiring query requests in natural language form, named entity recognition is used to extract entity words related to wafer manufacturing. A semantic disambiguation model is then invoked, and context-aware mechanisms are used to perform semantic disambiguation on the entity words, obtaining the disambiguation features corresponding to the entity words. By calculating the semantic similarity between the disambiguation features and pre-stored fields, automatic alignment between entity words and database fields is achieved, improving the semantic gap between natural language and structured data fields. This allows users to achieve efficient and accurate information retrieval without needing to master the underlying database language. After matching the target field through semantic similarity, a database query statement can be generated accordingly, enabling rapid and accurate retrieval of wafer manufacturing-related historical information. An event evolution linked list learning model with temporal modeling capabilities is used to perform temporal analysis on the retrieved historical information, identifying historical information related to the anomalies requested by the user, thereby constructing an anomaly historical information sequence. Through the above process, this application can not only effectively improve the semantic gap between natural language and database structure in the prior art, but also greatly improve the efficiency and accuracy of fault tracing during wafer manufacturing. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0017] In the attached diagram: Figure 1 A diagram of an existing data-driven decision-making process provided for one embodiment of this application; Figure 2 A flowchart illustrating a method for analyzing abnormal phenomena in a semiconductor structure according to an embodiment of this application; Figure 3 A schematic diagram of a semantically driven process for identifying abnormal history information provided in an embodiment of this application; Figure 4 An overall system architecture diagram of a wafer-specific process quality traceability analysis system for incomplete information provided in an embodiment of this application; Figure 5 A schematic diagram of the overall system architecture of the MCP protocol provided in an embodiment of this application; Figure 6 This is a structural block diagram of a system for analyzing abnormal phenomena in a semiconductor structure provided in one embodiment of this application. Figure 7This is a schematic diagram of the structure of an electronic device provided in one embodiment of this application. Detailed Implementation

[0018] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0019] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. The drawings only show the components related to this application and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0020] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the present application. However, it will be apparent to those skilled in the art that embodiments of the present application may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the present application.

[0021] like Figure 1As shown, research has revealed that data-driven decision-making processes in the semiconductor manufacturing field face significant challenges, resulting in substantial obstacles for non-technical users when acquiring and analyzing manufacturing data. This is because current wafer manufacturing processes primarily rely on frontline process engineers submitting query requests in the form of work orders based on actual analytical needs. The heads of various heterogeneous systems then assist in retrieving relevant data based on the request. These systems include, but are not limited to, Manufacturing Execution Systems (MES), Statistical Process Control (SPC), and Advanced Process Control (APC). IT personnel, based on feedback from the system heads, extract the required data from these multiple heterogeneous systems and manually import it, then use Data Expert General (DEG) tools for correlation analysis. It is clear that this traditional query process is cumbersome, lacks automation, and severely limits query efficiency. Since key data is often scattered across multiple manufacturing systems such as MES, SPC, and APC, cross-system data correlation analysis takes an average of over three hours. Furthermore, this process heavily relies on traditional SQL query methods, requiring users to be proficient in database structures and complex syntax, making it difficult for most frontline process engineers to independently complete data extraction tasks. Moreover, a significant semantic gap exists between the natural language expressions used by engineers and the underlying physical structure of the database, further raising the barrier to entry for database use and increasing communication and maintenance costs in multi-role collaborations.

[0022] To address the aforementioned issues, this application provides a method for analyzing anomalies in semiconductor structures. Based on a wafer history analysis project, it aims to construct a new dialogue-driven root cause query paradigm, enabling engineers to conveniently obtain the necessary wafer history data through natural language interaction. This application integrates a Large Language Model (LLM), a dynamic relational database, and a real-time processing engine to construct a highly efficient closed-loop process of "natural language input—real-time multi-source retrieval—decision suggestion generation." Compared to the insufficient real-time performance of traditional retrieval-augmented generation (RAG) methods, this application significantly improves the overall response speed and supports production anomaly tracing and precise location of suspicious machines, thereby comprehensively enhancing the speed and accuracy of anomaly analysis during wafer manufacturing.

[0023] like Figure 2 As shown, the analysis method for anomalies in semiconductor structures includes the following steps: S21. Obtain a query request in the form of natural language text; wherein the query request is used to describe anomalies that occur during wafer manufacturing.

[0024] During wafer manufacturing, when users need to perform anomaly analysis on the wafers, they will submit a query request. This query request characterizes the user's analytical needs regarding anomalies occurring during wafer manufacturing. Since this query request is often issued by frontline process engineers, it is typically expressed in natural language text. For example, a query request might be: "The accuracy of the AA layer is relatively low; is there a problem with the exposure machine?"

[0025] S22. Perform named entity recognition on the query request to obtain entity words in the query request that are associated with the wafer manufacturing.

[0026] After receiving a user's query request, named entity recognition (NAME) is performed on the query request to extract entity words related to wafer manufacturing. In one example, the query request can be input into a pre-trained named entity recognition model to identify and extract the entity words and their corresponding entity types contained in the query request. It is understood that the named entity recognition model can be any model using a natural language processing framework, including but not limited to BiLSTM, BERT, or Transformer, as long as it can effectively identify entity words in the query request. Entity words refer to words in the query request that have a clear semantic correspondence with structured data fields actually existing in the manufacturing system. For example, if the query request is "The accuracy of the AA layer is relatively low; is there a problem with the exposure machine?", after named entity recognition, the entity words (such as AA, layer, exposure machine) and their corresponding entity types (such as layer, device) are identified.

[0027] S23. Invoke the semantic disambiguation model, based on the context-aware mechanism, and perform semantic disambiguation on entity words according to the query request to obtain the disambiguation features corresponding to the entity words; wherein, the semantic disambiguation model is a multilayer perceptron model.

[0028] For each entity word, context-aware processing is performed using a semantic disambiguation model. Specifically, a multilayer perceptron is used to progressively aggregate the context information of the sentence containing the entity word, and by analyzing the semantic relationships between the entity word and other words in the query request, the actual meaning of the entity word in the current context is obtained, generating disambiguation features that reflect the true semantics of the entity word.

[0029] In an optional embodiment of this application, step S23 includes steps S231 to S233: S231. Perform word segmentation on the query request to obtain multiple segmented texts.

[0030] Considering that query requests may contain invalid components such as interjections and stop words, the query requests can first be cleaned to remove irrelevant words. The cleaned query requests are then input into a word segmenter for word segmentation, thereby dividing the query requests into multiple segmented texts. For each segmented text, based on the aforementioned entity recognition results, it is labeled as an entity word or a non-entity word for subsequent semantic disambiguation.

[0031] S232. Semantically encode each segmented text to obtain the corresponding semantic coding features, and arrange the semantic coding features in sequence to form a semantic coding matrix.

[0032] After the query request is segmented, all segmented texts are semantically encoded. For example, all segmented texts can be input into a pre-trained language model, and each segmented text is encoded based on a context mechanism to obtain its semantic encoding features. These semantic encoding features characterize the semantic information of the corresponding segmented text within the entire query request context. It is understood that the language model can be any text encoding model such as BERT, Word2Vec, or the Transformer encoder; no specific limitation is made. The semantic encoding features of all segmented texts are arranged sequentially according to their order in the query request to obtain a semantic encoding matrix. Each row of the semantic encoding matrix corresponds to the semantic encoding feature of one segmented text, and each column corresponds to the numerical value of the semantic encoding feature in different semantic dimensions.

[0033] In an optional embodiment of this application, step S232 includes the following processing steps: semantically encoding each segmented text to obtain semantic encoding features corresponding to each segmented text; arranging each semantic encoding feature in sequence to form an initial semantic encoding matrix; and performing global average pooling on the initial semantic encoding matrix to obtain the final semantic encoding matrix.

[0034] Specifically, all segmented texts are sequentially input into a pre-trained language model. Based on the context of each segmented text, it is encoded to obtain its semantic encoding features. After all segmented texts are encoded, the semantic encoding features are arranged in the original segmentation order to form an initial semantic encoding matrix. To reduce data dimensionality, enhance the universality of the semantic disambiguation model, and improve its efficiency in handling complex queries, this embodiment also performs global average pooling on the initial semantic encoding matrix to obtain the final semantic encoding matrix.

[0035] S233. For each entity word: input the semantic encoding matrix into the semantic disambiguation model, and based on the context-aware mechanism, determine the interaction features between the entity word and other segmented texts in the semantic encoding matrix, and use them as the disambiguation features corresponding to the entity word.

[0036] For each entity word extracted through named entity recognition, the following processing is performed: The semantic encoding matrix corresponding to the query request is input into the semantic disambiguation model. Since the semantic encoding matrix contains the contextual semantic features of all segmented texts in the query request, the semantic disambiguation model can extract the interaction features between the entity word and the remaining text words in the query request. Since the interaction features comprehensively reflect the semantic tendency of the entity word in the entire query request context, they can be used as the disambiguation features of the entity word for subsequent field matching and database query statement generation, thereby ensuring that the target field in the structured database can be accurately mapped according to the disambiguation features and the required resume information can be obtained.

[0037] In an optional embodiment of this application, step S233 includes the following process: inputting the semantic encoding matrix into the vertical fusion module of the semantic disambiguation model, and based on a multilayer perceptron, vertically fusing the semantic encoding features of all word segmented texts in each column of the semantic encoding matrix to obtain a vertically fused semantic encoding matrix; for each entity word, the following processing steps are performed: extracting the semantic encoding features of the corresponding position of the entity word from the vertically fused semantic encoding matrix; inputting the semantic encoding features into the horizontal fusion module of the semantic disambiguation model, and fusing the elements of different columns of the semantic encoding features to obtain the disambiguation features of the entity word.

[0038] To improve the accuracy of semantic disambiguation, this embodiment uses a cross-fusion mechanism to mine semantic information related to entity words by leveraging the contextual information of entity words, thereby achieving more accurate semantic disambiguation. Specifically, since the semantic encoding matrix is ​​a two-dimensional real-valued matrix composed of the semantic encoding features of multiple segmented texts, it is denoted as M. The semantic encoding matrix M corresponding to the query request is input into the vertical fusion module in the semantic disambiguation model, and the semantic encoding values ​​of all segmented texts in each column of the semantic encoding matrix, i.e., each feature dimension, are vertically fused. Specifically, the original semantic encoding matrix M is subjected to layer normalization, and then the t-th column of the normalized semantic encoding matrix is ​​selected. The data in the t-th column is multiplied by the preset first weight matrix, and after being processed by the activation function, it is multiplied by the preset second weight matrix. The result of the multiplication is added to the t-th column of the original semantic encoding matrix to obtain the data in the t-th column of the vertically fused semantic encoding matrix. The above process is shown in formula (1): (1) in, This represents the data in column t of the semantic coding matrix after vertical fusion. The data in column t of the original semantic encoding matrix. , The first and second weight matrices are preset. (*) is the activation function. This refers to the data in column t after layer normalization of the original semantic coding matrix.

[0039] Obtain the semantic encoding matrix after vertical fusion Then, using the horizontal integration module, targeting For each row, the features of the segmented text corresponding to that row are horizontally fused across all dimensions to further explore the deep connections between different semantic dimensions within the same event, thereby obtaining the disambiguation features of the entity word. Specifically, the semantic encoding matrix after vertical fusion is subjected to layer normalization. The k-th row of the normalized semantic encoding matrix is ​​selected, and the data of the k-th row is multiplied by the preset third weight matrix. After being processed by the activation function, it is multiplied by the preset fourth weight matrix. The result of the multiplication is added to the data of the corresponding row of the original semantic encoding matrix to obtain the data of the k-th row of the horizontally fused semantic encoding matrix. The above process is shown in formula (2): (2) in, This represents the data in the k-th row of the semantic coding matrix after horizontal fusion. This represents the data in the k-th row of the original semantic encoding matrix. , These are the preset third and fourth weight matrices, respectively. This represents the k-th row of data after layer normalization of the vertically fused semantic encoding matrix. The row vector corresponding to the position of the entity word is extracted from the horizontally fused semantic encoding matrix; this is the disambiguation feature of the entity word. The disambiguation feature obtained in this way can more accurately reflect its true meaning within the entire query context.

[0040] S24. Calculate the semantic similarity between the disambiguation feature and each pre-stored field, determine the target field accordingly, and retrieve each history information associated with the manufacturing of the wafer based on the target field.

[0041] After semantic disambiguation of entity words is completed, each disambiguation feature is processed as follows: Semantic similarity is calculated between the disambiguation feature and the semantic vector of a pre-stored database field. The semantic similarity can be calculated using cosine similarity or vector dot product, etc., and is not limited here. The field with the highest semantic similarity to the disambiguation feature can be selected as the target field corresponding to that entity word. After all entity words are matched in the above manner, based on the determined target fields and their respective database tables, a database query is constructed to retrieve the historical information associated with the wafer manufacturing process. This provides a data foundation for subsequent identification and analysis of abnormal events. The historical information refers to the structured process record data formed during wafer manufacturing, including the wafer's processing trajectory at different process stages, the equipment used, key parameters, and test results. Each piece of historical information records the specific state of the wafer at the corresponding step.

[0042] In an optional embodiment of this application, step S24 includes steps S241 and S242: S241. Calculate the semantic similarity between the disambiguation features and the pre-stored fields, determine the target field and its corresponding database table name accordingly, and generate multiple database query statements.

[0043] After obtaining the disambiguation feature corresponding to each entity word, the following processing is performed on each disambiguation feature: The semantic similarity between the disambiguation feature and the pre-stored semantic vector of each field is calculated, where semantic similarity characterizes the degree of semantic matching between the disambiguation feature and each structured field. The field with the highest semantic similarity is selected as the target field for matching the disambiguation feature, and the database table name to which the target field belongs can be determined based on the target field. It is understood that this application can perform semantic search on one or more databases. For a single database, the encoded fields in that database can be directly matched to determine the target field corresponding to the entity word, and the database query statement can be determined based on the target field and its corresponding database table name to retrieve the content stored in the target field. For multiple databases, semantic vectors can be pre-encoded for each available field in all databases, and all field semantic vectors can be uniformly stored in a field semantic vector library to achieve unified semantic retrieval and field location across databases.

[0044] In an optional embodiment of this application, step S241 includes the following processing procedures: For each disambiguation feature: calculate the semantic similarity between the disambiguation feature and each field in the field library, select the field with a similarity greater than a preset similarity threshold as the target field, and determine the database table name corresponding to the target field; generate the database query statement corresponding to the disambiguation feature based on the target field and its corresponding database table name, using a preset template engine.

[0045] For each entity word, the semantic disambiguation feature is processed as follows: The semantic similarity between this disambiguation feature and the semantic vectors of all fields in the pre-stored field semantic vector library is calculated. The field with the highest semantic similarity is then selected as the target field. Based on a pre-established field-database table name mapping, the corresponding database table name for the target field can be automatically determined. A preset template generation engine can be invoked to dynamically generate SQL query statements conforming to structured database specifications based on the target field and the database table name. S242. Retrieve resume information from the corresponding database based on the query statements of each database.

[0046] By invoking the generated database queries, database resources from different manufacturing systems can be retrieved to extract relevant historical information related to the user's query request. For example, specific database tables and fields are retrieved through the database queries, allowing the extraction of necessary process parameters, equipment status, and other relevant data. By accessing these heterogeneous databases, the required historical data can be obtained and integrated to form a historical table composed of related events, laying the data foundation for subsequent event chain analysis.

[0047] This application achieves dynamic semantic mapping from user natural language queries (such as "AALOOP accuracy") to cross-system SQL statements (such as "select acc from accuracy where layer='AA'") by introducing a semantic disambiguation mechanism and a context-aware weight allocation strategy. Simultaneously, a multi-layer context-aware fusion mechanism is used as a bridge to connect user-input query requests with complex database structures. Through an event evolution linked list learning model, pre-constructed historical data is modeled and analyzed, enabling automatic data extraction and query statement generation across systems, thereby significantly improving the processing efficiency and accuracy of anomaly analysis during wafer manufacturing.

[0048] S25. Input the retrieved resume information into the event evolution linked list learning model in chronological order, identify the resume information associated with the abnormal phenomenon, and generate an abnormal resume information sequence; wherein, the event evolution linked list learning model is a time series model.

[0049] The retrieved historical information is sorted according to its corresponding timestamp and sequentially input into the event evolution linked list learning model. The event evolution linked list learning model can be of various types, including but not limited to time-series models such as the Mamba model and the Transformer architecture model, as long as it possesses long-term modeling capabilities and can dynamically learn the changing relationships between various historical information. Specifically, the event evolution linked list learning model combines the currently input historical information with the context states retained in historical inputs to identify key events that are causally related to the anomalies described in the user's query request. These key events are further labeled, generating an anomaly historical information sequence composed of multiple anomaly events in sequence. Based on this anomaly historical information sequence, the user can systematically understand the evolution path and potential root causes of a particular anomaly in the manufacturing process, thus providing a theoretical basis for subsequent process optimization or equipment intervention.

[0050] To achieve efficient identification of abnormal events in the wafer manufacturing process, in an optional embodiment of this application, the timing model is the Mamba model, and step S25 includes steps S251 and S252: S251. Based on the time sequence, each resume information is input into the Mamba model, and the input resume information is checked for abnormality. If the resume information is abnormal, it is marked as an abnormal resume node.

[0051] Based on chronological order, each piece of historical information is sequentially input into the Mamba model according to the time step sequence. As a novel state-space modeling framework, the Mamba model possesses a linear recursive structure and efficient long-sequence modeling capabilities, capturing long-range temporal dependencies while maintaining computational efficiency. At each time step, the currently input historical information is fused and updated with previously stored state information, and the temporal features of that time step are extracted. The model performs anomaly detection on the current historical information, obtaining an anomaly score, and compares the anomaly score with a preset anomaly score threshold. If the anomaly score is greater than or equal to the anomaly score threshold, it indicates that the historical information has a strong correlation with the current anomaly query, and therefore the historical information is marked as an anomaly node. Conversely, if the anomaly score is less than the anomaly score threshold, it indicates that the historical information has a weak correlation with the current anomaly, and the historical information can be marked as a normal node or left unmarked.

[0052] In an optional embodiment of this application, step S251 includes the following processing procedures: inputting the current resume information into the backbone network of the Mamba model, updating the current resume information according to the pre-stored historical hidden states, obtaining and saving the hidden state corresponding to the current resume information; wherein, if there is no historical hidden state, the historical hidden state is preset to the initial value; inputting the hidden state corresponding to the current resume information into the classifier of the Mamba model, detecting whether the current resume information is abnormal, and marking the resume information as an abnormal resume node when the resume information is abnormal.

[0053] The historical information of the current time step is input into the backbone network of the Mamba model, and participates in the state update operation together with the pre-stored historical state information to generate the hidden state of the current time step. If the current time step is the first time step, a preset initial state is used as the historical input. Based on its internally constructed state space mechanism, the Mamba model's backbone network can fully capture the temporal dependencies of historical information during the manufacturing process by fusing the currently input historical information with historical hidden states. The generated current hidden state is input into the classifier (such as a multilayer perceptron) module of the Mamba model to determine whether the currently input historical information is abnormal. If the historical information is abnormal, it is marked as an abnormal historical node. Through the above process, dynamic identification and anomaly labeling of time-series historical data in the manufacturing process can be achieved, providing a data foundation for the subsequent construction and analysis of abnormal historical sequences.

[0054] It is understandable that, in addition to the aforementioned Mamba model, the temporal model in this application can also be a Transformer architecture model. Specifically, multiple retrieved historical information are arranged sequentially according to time to form a historical information sequence, and then input into the backbone network of the event evolution linked list learning model to generate anomalous features for each historical information. This backbone network includes multiple parallel detection units, each of which includes a multi-head attention module (MHA), a feedforward neural network (FF), layer normalization (LN), and residual connections (RC). The multi-head attention mechanism is used to capture the dependency relationship between current and historical historical information. The feedforward neural network enhances the nonlinear expressive power of the event evolution linked list learning model through two layers of linear transformation and the GELU nonlinear activation function. Layer normalization and residual connections improve the integrity of information transmission. Through the above processing, the global dependency relationship between the current input resume information and historical resume information can be captured more accurately, thereby providing more accurate feature support for subsequent anomaly event identification. This processing can be represented by the following formula (3): (3) Where x is the input resume information sequence, For having weight parameters Multi-head attention module, For having weight parameters feedforward neural networks For having weight parameters Layer normalized network, These are the anomalous features output by the backbone network. `r` is a composite function operator used to indicate that the output of the previous network is used as the input of the next network, and `r` represents the residual connection. The generated anomaly features are input into the classifier of the event evolution linked list learning model, which can determine whether the current input history information is an anomalous event and mark it if it is. Through this method, the event evolution linked list learning model can not only improve its ability to perceive the relationship between time and events, but also more accurately distinguish the subtle differences between similar events at different time points, thus enabling it to process knowledge graphs with complex temporal evolution features.

[0055] It is understandable that the temporal model can also be other models with temporal modeling capabilities, including but not limited to recurrent neural networks (RNN), long short-term memory networks (LSTM), temporal convolutional networks (TCN), gated recurrent units (GRU), etc., but due to space limitations, they will not be discussed in detail here.

[0056] S252. After all the resume information has been detected, all abnormal resume nodes are counted, and the corresponding abnormal resume information is arranged in chronological order to form an abnormal resume information sequence.

[0057] After all historical data has been detected, the historical data of all nodes marked as abnormal are collected and statistically analyzed. By extracting the timestamp field corresponding to each abnormal node, these abnormal historical data are sorted according to their occurrence time, thereby generating an abnormal historical data sequence reflecting the evolution of the abnormality. This sequence retains the key attributes of each abnormal event, facilitating subsequent root cause analysis of the cause of the abnormality. Furthermore, after forming the abnormal historical data sequence, a root cause report and faulty machines can be generated based on the sequence, where faulty machines refer to machines related to the current abnormality. Experiments show that using the method of this application, the accuracy of the top 30 causes in the root cause ranking exceeds 70%, root cause reports are generated quickly, and root cause analysis results for specific wafers can be accurately generated, thus significantly improving the efficiency and accuracy of problem localization.

[0058] like Figure 3 As shown, the user-input query request, after word segmentation and entity recognition, forms a semantic encoding feature matrix, which is then input into the semantic disambiguation model. After global mean pooling, the matrix is ​​fed into multiple parallel modules based on a multilayer perceptron (MLP). The initial disambiguation features generated by each module are fused to obtain the final disambiguation feature corresponding to each entity word. For each module: vertical feature fusion and horizontal feature fusion are performed sequentially to obtain the disambiguation feature of each entity word in the query request, denoted as . , … In the vertical feature fusion stage, the semantically encoded feature matrix is ​​input into the Layer Normalization (LN) network of the semantic disambiguation model. The feature values ​​of each column are normalized. Through the State Space Duality (SSD) mechanism, the long-distance dependencies between different entity words in the context are effectively captured, thereby improving the semantic disambiguation model's ability to understand contextual semantics. Furthermore, a Gated Multilayer Perceptron (Gated MLP) is used to dynamically adjust the input features by introducing a gating mechanism. This allows the semantic disambiguation model to adaptively emphasize key information based on the semantic context, enhancing its feature selection ability. Based on this, the semantic context information of all words is aggregated within the feature dimension to complete the vertical feature fusion. The fused output is then normalized again and then subjected to state space duality and the Gated MLP to achieve horizontal feature fusion, obtaining the initial disambiguation features corresponding to each entity word.

[0059] Continue as Figure 3As shown, after obtaining all the historical information corresponding to the user's query request, this historical information is sequentially input into the event evolution linked list learning model in chronological order. The event evolution linked list learning model consists of three stage-specific sub-modules: a front-end process module, a mid-end process module, and a back-end process module, corresponding to different stages of the wafer manufacturing process. Specifically, the input historical information is divided into stages, and historical data belonging to the front-end, mid-end, and back-end process stages are sent to the corresponding modules to identify whether there is any abnormal historical information related to the user's query anomaly in the current stage. All abnormal historical information is summarized and arranged in chronological order to obtain an abnormal historical information sequence. It should be further noted that before each piece of historical information is input into the model, it needs to undergo feature mapping, mapping it to a unified feature embedding space. Specifically, each piece of historical information includes two parts: a process parameter vector... , , , One part consists of basic process data used to characterize the event, and the other part consists of semantic embedding features related to the process stage to which it belongs.

[0060] Understandably, depending on the different process stages, the event evolution linked list learning model of this application adopts differentiated feature embedding strategies: in the front-end process module, features such as process classification (E_cls), the location or node of the event (E_bush), the frequency of event visits (E_visit), and the timestamp of the event in the front-end process (E_time1) are extracted; in the middle-end process module, features such as process classification (E_cls), usage status (E_USA), setting features (E_pres), and the timestamp of the event in the middle-end process (E_time2) are extracted; in the back-end process module, features such as process classification (E_cls), fault observation (E_oba), success marker (E_suc), and the timestamp of the event in the back-end process (E_time3) are extracted. After combining the semantic embedding features with the corresponding process parameter vectors, the corresponding process module is used to determine whether the history information is abnormal history information. When the event evolution linked list learning model is a Transformer model, its structure is as follows: Figure 3 As shown on the left, it includes network layers such as multi-head attention, feedforward neural network, and layer normalization (Add&Norm). By globally modeling the input history information sequence, it captures the dependency features between various history information, thereby obtaining abnormal history information related to the current anomaly. The abnormal history information in each stage module is arranged sequentially according to the corresponding timestamp to form a complete abnormal history information sequence, so as to perform root cause analysis of the anomaly and identify the faulty machine.

[0061] It should be further explained that, Figure 3 The diagram below shows the process history of wafer NBY182.010 during its manufacturing process. It arranges the various process steps the wafer underwent in chronological order, revealing that the wafer passed through multiple photolithography (PHOTO PROC) processes, interspersed with an ash process. Different colored nodes in the diagram indicate whether an anomaly occurred at that step; for example, red nodes represent an anomaly detected in that process step, green nodes represent normal processes, and gray nodes represent nodes unrelated to the current anomaly.

[0062] Understandably, this application can, on the one hand, extract entity words from user-input queries to form an entity selection set, and then use database query statements to map these entity words to historical data in the database, obtaining a unified representation of events. It also utilizes an event evolution linked list learning model to identify abnormal historical information and construct an abnormal historical information sequence. On the other hand, to further enhance the model's time-series awareness, this application can also introduce a time window through a time candidate mechanism, focusing only on historical information within the time period most relevant to the current query intent, thereby significantly improving the accuracy of event recognition and query efficiency.

[0063] like Figure 4The diagram illustrates the overall system architecture of a wafer-specific process quality traceability analysis system for incomplete information. The system comprises four layers: application interaction layer, business logic layer, data service layer, and data resource layer. In the data resource layer, the system integrates structured data (such as process parameters and machine sensor data), semi-structured data (such as R&D process information and machine sensor data), and unstructured data (such as wafer defect diagrams and GDS layouts), providing more comprehensive data for subsequent analysis. The data service layer processes this multi-source heterogeneous data through a knowledge extraction module. For example, it employs named entity recognition, cross-modal alignment, and zero-shot learning to unify the data representation of the aforementioned multi-modal data, obtaining knowledge embedding representations of these data and saving them to a database (DB) via the MCP protocol. Based on this, the business logic layer semantically encodes user-input natural language queries to obtain corresponding query embedding representations. The similarity between the query embedding and the stored knowledge embeddings is calculated. The history information corresponding to the knowledge embedding with the highest similarity is input into a large language model, which includes a semantic disambiguation model and an event evolution linked list learning model. The semantic disambiguation model identifies and parses entity words in the user's query request through hybrid contextual knowledge reasoning. The event evolution linked list learning model identifies history information related to the anomaly based on time-series data, constructing an anomaly history information sequence. Through this process, accurate tracing of process anomalies and precise location of machine faults are achieved in multi-source heterogeneous and incomplete information scenarios, thereby significantly improving the accuracy and speed of root cause analysis.

[0064] like Figure 5As shown, this diagram illustrates the overall system architecture of the MCP protocol, comprising four layers: a user interaction layer, a client core layer, a service layer, and external systems. Users input query requests in natural language via a command-line interface. The MCP client module (MCPClient) in the client core layer establishes a connection with the service layer using the connectToServer() function. The processQuery() function parses the input query request, and the chatLoop() function sends the parsed result to the service layer, establishing a multi-turn dialogue between the user and the system. Upon receiving the user's query request, the service layer retrieves the corresponding tool using the getTools() function in the tool server (ToolService()), executes the relevant tool operations using the callTool() function, and returns the result to the client core layer. Here, the tool in this application refers to a functional unit that interacts with a database. For example, the tool can access resume information in the corresponding database, thereby enabling data retrieval from multiple heterogeneous systems. The user's query request is submitted to the OpenAI SDK port of the Large Language Model (LLM) via the `sendMessage()` function in the Large Language Model Server (LLMService). The LLM performs semantic understanding and reasoning, and returns a sequence of generated exception history information to the client core layer. This sequence is then fed back to the command-line interface for user review. After feedback, the client core layer releases relevant data resources using `cleanup()` to reduce memory usage. Furthermore, the service layer uses the `addLogs()` and `clearLog()` functions in the Log Server (LogService()) to record and clean up runtime logs, ensuring traceability. Further, the service layer can call external systems, such as the MCP server (MCPServer) or other Large Language Model APIs, via API interfaces to achieve richer tool calls and query collaboration capabilities.

[0065] like Figure 6As shown, the semiconductor structure anomaly analysis system 600 includes: a query request acquisition module 610, an entity word recognition module 620, a semantic disambiguation module 630, a statement generation module 640, a history information extraction module 650, and an event chain generation module 660. The query request acquisition module 610 acquires query requests in natural language text form; wherein, the query request describes anomalies occurring during wafer manufacturing. The entity word recognition module 620 performs named entity recognition on the query request to obtain entity words associated with wafer manufacturing appearing in the query request. The semantic disambiguation module 630 invokes a semantic disambiguation model, based on a context-aware mechanism, to perform semantic disambiguation on entity words according to the query request, obtaining disambiguation features corresponding to the entity words; wherein, the semantic disambiguation model is a multilayer perceptron model. The statement generation module 640 calculates the semantic similarity between the disambiguation features and pre-stored fields, thereby determining the target field and its corresponding database table name, and generating multiple database query statements. The resume information extraction module 650 is used to retrieve resume information from the corresponding database based on various database query statements. The event chain generation module 660 is used to input all retrieved resume information into the event evolution linked list learning model in chronological order, and based on the temporal relationship between each resume information, identify the resume information associated with the abnormal phenomenon and generate an abnormal resume information sequence; wherein, the event evolution linked list learning model is a temporal model.

[0066] Specific limitations regarding the analysis system for anomalies in semiconductor structures can be found in the limitations of the analysis methods for anomalies in semiconductor structures described above, and will not be repeated here. Each module in the aforementioned analysis system for anomalies in semiconductor structures can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware format within or independently of the processor in a computer device, or stored in software format in the memory of a computer device, so that the processor can call the corresponding operations of each module.

[0067] It should be noted that, in order to highlight the innovative aspects of this application, this embodiment does not include modules that are not closely related to solving the technical problems proposed in this application, but this does not mean that there are no other modules in this embodiment.

[0068] like Figure 7 As shown, the electronic device 7 may include a memory 71, a processor 72, and a bus, and may also include a computer program stored in the memory 71 and executable on the processor 72, such as an analysis program for anomalies in semiconductor structures.

[0069] The memory 71 includes at least one type of readable storage medium, including flash memory, portable hard drive, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 71 can be an internal storage unit of the electronic device 7, such as a portable hard drive. In other embodiments, the memory 71 can be an external storage device of the electronic device 7, such as a plug-in portable hard drive, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the electronic device 7. Furthermore, the memory 71 can include both internal and external storage units of the electronic device 7. The memory 71 can be used not only to store application software and various types of data installed on the electronic device 7, such as code for analyzing anomalies in semiconductor structures, but also to temporarily store data that has been output or will be output.

[0070] In some embodiments, the processor 72 may be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits with the same or different functions, including combinations of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips. The processor 72 is the control unit of the electronic device 7, connecting various components of the entire electronic device 7 via various interfaces and lines. It executes programs or modules stored in the memory 71 (such as programs for analyzing anomalies in semiconductor structures) and calls data stored in the memory 71 to perform various functions and process data in the electronic device 7.

[0071] The processor 72 executes the operating system of the electronic device 7 and various installed application programs. The processor 72 executes the application programs to implement the steps in the above-described method for analyzing abnormal phenomena in semiconductor structures.

[0072] For example, a computer program can be divided into one or more modules, one or more of which are stored in memory 71 and executed by processor 72 to complete this application. One or more modules can be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in electronic device 7. For example, the computer program can be divided into a query request retrieval module 610, an entity recognition module 620, a semantic disambiguation module 630, a sentence generation module 640, a resume information extraction module 650, and an event chain generation module 660.

[0073] The integrated unit implemented as a software functional module described above can be stored in a computer-readable storage medium, which can be non-volatile or volatile. The software functional module, stored in the storage medium, includes several instructions to cause a computer device (which may be a personal computer, computer equipment, or network device, etc.) or processor to execute part of the functions of the anomaly analysis method for the semiconductor structure of the various embodiments of this application.

[0074] In summary, the method, system, device, and medium for analyzing anomalies in semiconductor structures proposed in this application have the following unexpected effects: By acquiring query requests in natural language form, named entity recognition is used to extract entity words related to wafer manufacturing. A semantic disambiguation model is then invoked, and context-aware mechanisms are used to perform semantic disambiguation on the entity words, obtaining disambiguation features corresponding to the entity words. By calculating the semantic similarity between the disambiguation features and pre-stored fields, automatic alignment between entity words and database fields is achieved, improving the semantic gap between natural language and structured data fields. This allows users to achieve efficient and accurate information retrieval without needing to master the underlying database language. After matching the target field through semantic similarity, a database query statement can be generated accordingly, enabling rapid and accurate retrieval of wafer manufacturing-related historical information. An event evolution linked list learning model with temporal modeling capabilities is used to perform temporal analysis on the retrieved historical information, identifying historical information related to the anomalies requested by the user, thereby constructing an anomaly historical information sequence. Through the above process, this application can not only effectively improve the semantic gap between natural language and database structure in the prior art, but also greatly improve the efficiency and accuracy of fault tracing during wafer manufacturing.

[0075] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A method of analyzing an abnormality of a semiconductor structure, characterized by, The method comprises: acquiring a query request in the form of natural language text; wherein the query request is used to describe an abnormal phenomenon generated during wafer manufacturing; performing named entity recognition on the query request to obtain an entity word associated with the wafer manufacturing appearing in the query request; calling a semantic disambiguation model to perform semantic disambiguation on the entity word based on a context perception mechanism according to the query request to obtain a disambiguation feature corresponding to the entity word; wherein the semantic disambiguation model is a multilayer perception machine model; calculating semantic similarity between the disambiguation feature and a pre-stored semantic of each field to determine a target field, and calling each history information associated with the wafer manufacturing based on the target field; inputting each history information called to an event evolution chain table learning model in time sequence to identify history information associated with the abnormal phenomenon and generate an abnormal history information sequence; wherein the event evolution chain table learning model is a time sequence model.

2. The analysis method of abnormality of a semiconductor structure according to claim 1, characterized by, The calling of the semantic disambiguation model to perform semantic disambiguation on the entity word based on the context perception mechanism according to the query request to obtain the disambiguation feature corresponding to the entity word, wherein the semantic disambiguation model is a multilayer perception machine model, comprises: performing word segmentation processing on the query request to obtain a plurality of segmented texts; performing semantic coding on each segmented text to obtain a corresponding semantic coding feature, and sequentially arranging each semantic coding feature to form a semantic coding matrix; for each entity word: inputting the semantic coding matrix to the semantic disambiguation model to determine an interaction feature between the entity word and other segmented texts in the semantic coding matrix based on the context perception mechanism, and taking the interaction feature as the disambiguation feature corresponding to the entity word.

3. The analysis method of abnormality of a semiconductor structure according to claim 2, characterized by, The performing of semantic coding on each segmented text to obtain a corresponding semantic coding feature, and the sequentially arranging of each semantic coding feature to form a semantic coding matrix, comprises: performing semantic coding on each segmented text to obtain a corresponding semantic coding feature; sequentially arranging each semantic coding feature to form an initial semantic coding matrix; performing global average pooling processing on the initial semantic coding matrix to obtain a final semantic coding matrix.

4. The method of analyzing abnormality of a semiconductor structure according to claim 2, wherein The inputting of the semantic coding matrix to the semantic disambiguation model to determine an interaction feature between the entity word and other segmented texts in the semantic coding matrix based on the context perception mechanism, and the taking of the interaction feature as the disambiguation feature corresponding to the entity word, comprises: inputting the semantic coding matrix to a longitudinal fusion module of the semantic disambiguation model to perform longitudinal fusion on semantic coding features of all segmented texts in each column of the semantic coding matrix based on a multilayer perception machine to obtain a longitudinally fused semantic coding matrix; for each entity word: extracting a semantic coding feature at a position corresponding to the entity word from the longitudinally fused semantic coding matrix; inputting the semantic coding feature to a transverse fusion module of the semantic disambiguation model to fuse elements in different columns of the semantic coding feature to obtain the disambiguation feature of the entity word.

5. The method of analyzing anomalies of a semiconductor structure according to claim 1, wherein The computing of the disambiguation feature and the semantic similarity of each pre-stored field determines the target field, and each history information associated with the wafer manufacturing is called based on the target field, including: The computing of the disambiguation feature and the semantic similarity of each pre-stored field determines the target field and its corresponding database table name, and generates a plurality of database query statements; Each database query statement is used to call the history information in the corresponding database.

6. The analysis method of abnormality of a semiconductor structure according to claim 5, characterized by, The computing of the disambiguation feature and the semantic similarity of each pre-stored field determines the target field and its corresponding database table name, and generates a plurality of database query statements, including: For each disambiguation feature: The semantic similarity of the disambiguation feature and each field in the field library is calculated, the field with a similarity greater than a pre-set similarity threshold is selected as the target field, and the database table name corresponding to the target field is determined; According to the target field and its corresponding database table name, the disambiguation feature corresponding database query statement is generated based on the pre-set template engine.

7. The method of analyzing anomalies of a semiconductor structure according to claim 1, wherein The timing model is a Mamba model, and each history information is input into the event evolution chain table learning model according to the time sequence, the history information associated with the abnormal phenomenon is identified, and an abnormal history information sequence is generated, including: According to the time sequence, each history information is input into the Mamba model, and whether the input history information is abnormal is detected, and when the history information is abnormal, it is marked as an abnormal history node; After monitoring that all history information is detected, all abnormal history nodes are counted, and the corresponding abnormal history information is arranged in time sequence to form an abnormal history information sequence.

8. The analysis method of abnormality of a semiconductor structure according to claim 7, characterized by, For each history information, the history information is input into the Mamba model, and whether the input history information is abnormal is detected, and when the history information is abnormal, it is marked as an abnormal history node, including: The current history information is input into the backbone network of the Mamba model, the state of the current history information is updated according to the pre-stored historical hidden state, and the hidden state corresponding to the current history information is obtained and saved; If there is no historical hidden state, the historical hidden state is pre-set to an initial value; The hidden state corresponding to the current history information is input into the classifier of the Mamba model, whether the current history information is abnormal is detected, and when the history information is abnormal, the history information is marked as an abnormal history node.

9. An abnormality analysis system of a semiconductor structure, characterized by comprising: The system comprises: A query request acquisition module is configured to acquire a query request in the form of natural language text; wherein the query request is used to describe an abnormal phenomenon generated during wafer manufacturing; An entity word recognition module is configured to perform named entity recognition on the query request to obtain an entity word associated with the wafer manufacturing appearing in the query request; A semantic disambiguation module is configured to call a semantic disambiguation model, perform semantic disambiguation on the entity word based on a context perception mechanism according to the query request, and obtain a disambiguation feature corresponding to the entity word; wherein the semantic disambiguation model is a multi-layer perception machine model; The sentence generation module is configured to calculate the semantic similarity between the disambiguation feature and each field, determine the target field and the corresponding database table name, and generate a plurality of database query statements; The resume information extraction module is configured to call the resume information in the corresponding database based on each database query statement; The event chain generation module is configured to input all the called resume information into an event evolution chain table learning model according to the time sequence, identify the resume information associated with the abnormal phenomenon based on the time sequence relationship between each resume information, and generate an abnormal resume information sequence; wherein the event evolution chain table learning model is a time sequence model.

10. An electronic device, comprising: The electronic device includes: one or more processors; a storage device configured to store one or more programs that, when executed by the one or more processors, cause the electronic device to implement the analysis method for abnormal phenomena of the semiconductor structure according to any one of claims 1 to 8.

Images