Chip defect detection and fault diagnosis method

Abstract

The application relates to the chip detection technical field, in particular to a chip defect detection and fault diagnosis method, which comprises the following steps: collecting historical diagnosis log information and machine processing records in a chip manufacturing process, and constructing a time sequence dynamic domain knowledge graph with a time stamp and a production batch attribute; performing graph representation learning on the time sequence dynamic domain knowledge graph with the time stamp and the production batch attribute, to generate an entity feature vector set and a relationship feature vector set with a confidence attenuation weight; in the application, current defect information and retrieved historical knowledge are integrated into a structured text prompt, the prompt can drive a complex reasoning process which can simultaneously focus on a time sequence and a spatial position, a detailed diagnosis report is generated, deep cause analysis and executable maintenance suggestions are provided, and clear expert-level guidance is directly provided for engineers, so that rapid problem positioning and solving are realized.

Description

Technical Field

[0001] This invention relates to the field of chip testing technology, and in particular to a method for chip defect detection and fault diagnosis. Background Technology

[0002] Chip defect detection and fault diagnosis methods refer to a series of technical means and processes used in the entire semiconductor manufacturing process to ensure product yield and reliability. The core objective is to efficiently identify physical or electrical defects that occur in chips during the production process and to deeply analyze the root causes of these defects.

[0003] Existing chip defect diagnosis methods suffer from inherent limitations due to fragmented data and a lack of in-depth correlation analysis. In practice, visual data on defects, equipment operating parameters, and production line scheduling information are typically stored in independent systems. This operational model forces engineers to manually compile and compare information from different sources when troubleshooting. For example, when a specific type of defect is found in a wafer inspection report, engineers need to independently query the manufacturing execution system to obtain the wafer's processing history and separately analyze time-series data from dozens of sensors on multiple process devices. This manual correlation process is not only time-consuming and error-prone, but its effectiveness is also fundamentally limited by the engineer's personal experience, often leading to the overlooking of crucial connections between subtle parameter drifts and the final physical defects. Therefore, improvements are needed. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a chip defect detection and fault diagnosis method.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a chip defect detection and fault diagnosis method, comprising the following steps:

[0006] Collect historical diagnostic logs and machine processing records during the chip manufacturing process to construct a time-series dynamic domain knowledge graph with timestamps and production batch attributes;

[0007] The time-series dynamic domain knowledge graph with timestamps and production batch attributes is subjected to graph representation learning to generate a set of entity feature vectors and a set of relation feature vectors with confidence decay weights.

[0008] Acquire the target defect image data and target machine context data of the chip to be inspected, and perform feature extraction to generate a multimodal query feature vector;

[0009] Based on the multimodal query feature vector, a correlation matching retrieval is performed in the entity feature vector set and the relation feature vector set with confidence decay weight to obtain fault knowledge retrieval results;

[0010] The target defect image data, the target machine context data, and the fault knowledge retrieval results are fused into an enhanced contextual prompt text;

[0011] The enhanced contextual prompt text is input into a preset large-scale visual language model for spatiotemporal dual attention chain reasoning, and the output is a fault diagnosis result containing cause analysis and maintenance suggestions.

[0012] Preferably, the steps of collecting historical diagnostic log information and machine processing records during the chip manufacturing process to construct a time-series dynamic domain knowledge graph with timestamps and production batch attributes specifically include:

[0013] Acquire chip design specification documents, historical machine daily maintenance records, underlying process formula iteration documents, and wafer fabrication quantity data for the corresponding batch during the operation of the chip manufacturing production line;

[0014] The chip design specification document data, the historical machine daily maintenance record data, and the underlying process recipe iteration file data are converted to text format and invalid character filtering and cleaning operations are performed to obtain a clean specification text content sequence.

[0015] The clean, canonical text content sequence is subjected to deep scanning extraction using a pre-trained named entity recognition algorithm network model to obtain a set of entity data for core components of semiconductor equipment and a set of entity data for typical characteristics of chip faults.

[0016] Using the established dependency parsing algorithm model, a syntax tree correlation analysis operation is performed on the clean and canonical text content sequence to extract the node dependency relationship between the core component entity data set of the semiconductor equipment and the typical feature entity data set of chip faults, thereby obtaining the initial fault evolution logical relationship sequence.

[0017] The core component entity data set of the semiconductor device, the typical characteristic entity data set of chip faults, and the initial fault evolution logical relationship sequence are fused and processed to construct a static semiconductor defect basic knowledge graph.

[0018] The historical diagnostic log information and the machine processing record are parsed using regular expression parsing rules. The actual occurrence time node data of each closed-loop maintenance event record in the historical diagnostic log information and the machine processing record, as well as the cumulative wafer processing batch data within the event period, are extracted.

[0019] The actual occurrence time node data of the closed loop and the cumulative wafer processing batch data are configured on the corresponding logical relationship vector edge in the initial fault evolution logical relationship sequence in the static semiconductor defect basic knowledge graph as absolute occurrence time label and batch period label, to obtain the time-series dynamic domain knowledge graph with timestamp and production batch attributes.

[0020] Preferably, the step of performing graph representation learning on the time-series dynamic domain knowledge graph with timestamps and production batch attributes to generate a set of entity feature vectors and a set of relation feature vectors with confidence decay weights specifically includes:

[0021] Completely extract the full network node element features and full network edge element features from the time-series dynamic domain knowledge graph with timestamp and production batch attributes;

[0022] The full set of network node element features and the full set of network edge element features are subjected to uniformly distributed random initialization encoding to obtain the initial network node vector parameter matrix and the initial network edge vector parameter matrix.

[0023] The initial network node vector parameter matrix and the initial network edge vector parameter matrix are synchronously input into the pre-constructed knowledge graph translation metric model. Based on the triplet translation space distance metric mechanism, a high-dimensional vector space mapping iterative calculation is performed to obtain the target steady-state entity feature vector sequence and the target steady-state relation feature vector sequence.

[0024] The formula is: ,in This represents the value of the loss function for the translational spatial distance metric of the triplet. The feature set of positive sample triples representing correct logical associations. This represents the positive sample head entity feature parameters, positive sample relation feature parameters, and positive sample tail entity feature parameters corresponding to the positive sample triplet feature set. The feature set of negative sample triples representing erroneous logical associations. This represents the negative sample head entity feature parameters, negative sample relation feature parameters, and negative sample tail entity feature parameters corresponding to the negative sample triplet feature set. The hyperparameter representing the margin between positive and negative samples is... This represents the spatial correlation penalty coefficient between adjacent processes extracted based on the machine tool process flow topology. Represents the Euclidean distance metric function. This represents the head entity encoding tensor corresponding to the positive sample head entity feature parameters. This represents the relation encoding tensor corresponding to the relation feature parameters of the positive samples. This represents the tail entity encoding tensor corresponding to the tail entity feature parameters of the positive sample. This represents the negative head entity encoding tensor corresponding to the negative sample head entity feature parameters. This represents the negative tail entity encoding tensor corresponding to the negative sample tail entity feature parameters;

[0025] Obtain the current system diagnostic time scale value, calculate the absolute offset of the time interval between the current system diagnostic time scale value and the actual occurrence time node data of the closed loop in the time-series dynamic domain knowledge graph with timestamp and production batch attributes, and obtain the cumulative number of actual processed wafers corresponding to the absolute offset of the time interval.

[0026] Based on the absolute offset of the time interval and the cumulative number of actual processed wafers, the dynamic confidence decay weight index of the full network edge element features is calculated using a preset exponential aging decay algorithm function.

[0027] The formula is: ,in This indicates the dynamic confidence decay weight index, This represents the manually set initial full-load confidence scalar value. The base of the natural logarithm mathematical constant, Indicates the inherent aging degradation coefficient constant of the equipment. Indicates the absolute offset of the time interval, Represents the physical fatigue coefficient constant of machine tool processing operations. This indicates the cumulative number of wafers actually processed;

[0028] The target steady-state entity feature vector sequence is filtered by a linear fully connected layer feature mapping to obtain the entity feature vector set;

[0029] The target steady-state relationship feature vector sequence is fused and adjusted by performing matrix element-wise multiplication with the dynamic confidence decay weight index to obtain the relationship feature vector set with confidence decay weight.

[0030] Preferably, the steps of acquiring the target defect image data of the chip to be inspected and the target machine context data, and performing feature extraction to generate a multimodal query feature vector specifically include:

[0031] The surface microstructure image features of the chip under test are acquired by a high-resolution electron microscope scanning imaging device to obtain the target defect image data, and the two-dimensional physical coordinate data of the chip under test on its respective wafer are extracted.

[0032] Simultaneously read the external sensor operating status parameters and internal chamber environment monitoring indicators of the test equipment where the chip under test is located in real time, and combine them to generate the target equipment context data;

[0033] The target defect image data is input into a pre-trained deep visual residual coding network to perform multi-scale convolutional feature extraction operations, resulting in a high-dimensional visual feature distribution tensor matrix.

[0034] The high-dimensional visual feature distribution tensor matrix is ​​subjected to global average pooling spatial dimensionality reduction processing to obtain a densely distributed image feature representation vector;

[0035] Convert the target machine context data into a long text formatted string feature sequence;

[0036] The long text formatted string feature sequence is input into the self-attention computation layer of the natural language text encoder for deep semantic feature extraction to obtain the machine semantic feature representation vector;

[0037] The densely distributed image feature representation vector is spatially encoded and compensated using the two-dimensional physical coordinate data. Then, the compensated densely distributed image feature representation vector and the machine semantic feature representation vector are fused together with cross-modal cross-attention in the feature channel numerical dimension to obtain a multimodal cascaded associated feature representation vector.

[0038] The multimodal cascaded associated feature representation vector is input into a nonlinear fully connected feature fusion network for feature dimensionality reduction and feature semantic space alignment to obtain the multimodal query feature vector.

[0039] Preferably, the step of performing correlation matching retrieval in the entity feature vector set and the relation feature vector set with confidence decay weights based on the multimodal query feature vector to obtain fault knowledge retrieval results specifically includes:

[0040] Extract each local entity feature vector to be matched by iterating through the set of entity feature vectors one by one;

[0041] Calculate the cosine angle mapping distance between the multimodal query feature vector and the feature vector of each local entity to be matched, and obtain the basic entity semantic similarity measurement score;

[0042] Each local relation feature vector to be matched is extracted by iterating through the set of relation feature vectors with confidence decay weights one by one.

[0043] Calculate the mathematical metric of the inner product projection between the multimodal query feature vector and each local relation feature vector to be matched, and obtain the basic relation pattern matching metric score;

[0044] Obtain the graph theory center metric of the adjacency network feature vector corresponding to each local entity feature vector to be matched in the temporal dynamic domain knowledge graph with timestamp and production batch attributes;

[0045] The graph theory center metric is used to perform a joint weighted fusion calculation on the basic entity semantic similarity measurement score and the basic relation pattern matching metric score to obtain the comprehensive knowledge retrieval confidence score.

[0046] The formula is: ,in This represents the total confidence score of the comprehensive knowledge retrieval. The standard logarithmic function with base 2 The graph theory centrality metric index represents the graph theory centrality metric index. This indicates the weighting coefficient for entity matching confidence. This represents the semantic similarity score of the basic entities. Indicates the weighting coefficient of relationship matching confidence. This refers to the dynamic confidence decay weight index carried by each feature vector of the local relation to be matched. This represents the basic relation pattern matching metric score;

[0047] The total confidence scores of the comprehensive knowledge retrieval are arranged in descending order from high to low to obtain a descending confidence score ranking list.

[0048] Extract the target entity knowledge text fragments and target relationship knowledge text fragments corresponding to the top ten scores in the descending confidence score ranking list;

[0049] The target entity knowledge text fragment and the target relation knowledge text fragment are processed by concatenating sentences using natural language grammar rules to obtain the fault knowledge retrieval result.

[0050] Preferably, the step of fusing the target defect image data, the target machine context data, and the fault knowledge retrieval results into enhanced contextual prompt text specifically includes:

[0051] Configure a standard system-level prompt and guidance execution instruction template, which includes expert role setting text placeholders and reasoning task description text placeholders;

[0052] The target machine context data and the two-dimensional physical coordinate data of the wafer to which it belongs are completely filled into the mapping area corresponding to the inference task description text placeholder of the standard system-level prompt guidance execution instruction template to obtain the basic background task prompt condition string;

[0053] The target defect image data is subjected to pixel normalization and bilinear scaling, and then converted into text escape image tagging labels in the target format;

[0054] The text escape image tag is embedded into the reserved image placeholder specified in the basic background task prompt condition string to obtain the multimodal visual basic prompt string;

[0055] Extract key defect cause phrase fragments and key equipment maintenance guidance phrase fragments from the fault knowledge retrieval results;

[0056] The phrase fragments describing the causes of key defects and the phrase fragments describing the maintenance guidance for key equipment are rearranged and arranged in chronological order according to the actual occurrence time to generate a context for reasoning based on expert knowledge.

[0057] The reference expert knowledge reasoning context is directly appended to the end region of the multimodal visual basic prompt string as an auxiliary visual reasoning judgment condition to obtain the enhanced context prompt text.

[0058] Preferably, the step of inputting the enhanced contextual cue text into a preset large-scale visual language model for spatiotemporal dual attention chain reasoning, and outputting fault diagnosis results including cause analysis and maintenance suggestions, specifically includes:

[0059] The enhanced contextual prompt text is processed by word segmentation rule mapping using a sub-word segmenter component to obtain a discrete language tag unit sequence;

[0060] The discrete language tagging unit sequence is input into the word embedding encoding layer of the preset large-scale visual language model to perform high-dimensional continuous space projection mapping operation, thereby obtaining a multi-dimensional word embedding representation feature tensor.

[0061] Extract the auxiliary reference occurrence time node attribute data corresponding to the key defect cause phrase fragment in the fault knowledge retrieval results, and extract the two-dimensional physical coordinate data contained in the enhanced contextual prompt text;

[0062] A time-aware bias attention mask matrix for the target dimension is constructed based on the auxiliary reference occurrence time node attribute data, and a wafer spatial position bias attention mask matrix is ​​constructed based on the two-dimensional physical coordinate data.

[0063] The multi-dimensional word embedding representation feature tensor, the time-aware bias attention mask matrix, and the wafer spatial location bias attention mask matrix are jointly input into the multi-head spatiotemporal attention calculation module of the preset large-scale visual language model to calculate the global weight attention allocation probability and obtain the global spatiotemporal correlation feature latent variable.

[0064] The formula is: ,in Represents the global spatiotemporal correlation feature latent variables, Represents the normalized exponential activation mapping function, This represents the query feature parameter matrix corresponding to the multi-dimensional word embedding representation feature tensor after linear transformation. The transpose of the key feature parameter matrix corresponding to the multi-dimensional word embedding representation feature tensor after linear transformation is represented by the transpose operation matrix. This indicates that the feature hidden layer corresponding to the key feature parameter matrix has a fixed dimension. This represents the time-aware biased attention mask matrix. This represents the wafer spatial location bias attention mask matrix. This represents the value feature parameter matrix corresponding to the multi-dimensional word embedding representation feature tensor after linear transformation;

[0065] The global spatiotemporal correlation feature latent variables are input into the feedforward fully connected neural network module of the preset large-scale visual language model to perform nonlinear high-order feature transformation processing to obtain high-level abstract logical semantic fusion features.

[0066] The high-level abstract logical semantic fusion features are input into the autoregressive text decoding module of the preset large-scale visual language model to perform stepwise word distribution probability prediction sampling, so as to obtain the optimal diagnostic word prediction output sequence.

[0067] The optimized diagnostic vocabulary prediction output sequence is divided into two text blocks according to the natural language semantic paragraph structure: a defect core cause analysis paragraph and a device maintenance and upkeep guidance paragraph.

[0068] The text block analyzing the core causes of the defects is directly used as the cause analysis, and the text block providing equipment maintenance and repair guidance is directly used as the repair suggestion.

[0069] The cause analysis and the repair suggestions are merged and output according to the preset standard fault result layout format to obtain the fault diagnosis result containing the cause analysis and repair suggestions.

[0070] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0071] This invention collects and structures historical diagnostic logs and machine processing records to construct a dynamic knowledge network that explicitly models time and production batch dimensions, forming a dynamic knowledge base. The relational connections in this network have confidence weights that decay over time and with usage, ensuring that diagnostic reasoning prioritizes the latest and most relevant historical events, improving the timeliness and accuracy of the applied knowledge. When a new defect occurs, the visual image of the defect and the machine's contextual operating data are simultaneously acquired and transformed into a unified multi-source feature representation, thus capturing a complete snapshot of the fault. This feature representation enables deep semantic retrieval within the historical knowledge base, retrieving not isolated similar defects, but complete fault scenarios of the same type. Finally, the current defect information and the retrieved historical knowledge are integrated into a structured text prompt. This prompt drives a complex reasoning process that simultaneously considers time series and spatial location, generating a detailed diagnostic report that provides in-depth causal analysis and actionable maintenance suggestions. It directly provides engineers with clear, expert-level guidance for rapid problem localization and resolution. Attached Figure Description

[0072] Figure 1 This is a schematic diagram of the steps of the present invention. Detailed Implementation

[0073] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0074] Please see Figure 1 This invention provides a technical solution, a method for chip defect detection and fault diagnosis, comprising the following steps:

[0075] Collect historical diagnostic logs and machine processing records during the chip manufacturing process to construct a time-series dynamic domain knowledge graph with timestamps and production batch attributes;

[0076] Graph representation learning is performed on a time-series dynamic domain knowledge graph with timestamps and production batch attributes to generate a set of entity feature vectors and a set of relation feature vectors with confidence decay weights.

[0077] Acquire the target defect image data and target machine context data of the chip to be inspected, and perform feature extraction to generate a multimodal query feature vector;

[0078] Based on multimodal query feature vectors, correlation matching retrieval is performed in the entity feature vector set and the relation feature vector set with confidence decay weights to obtain fault knowledge retrieval results;

[0079] The target defect image data, target machine context data, and fault knowledge retrieval results are integrated into enhanced contextual prompt text;

[0080] The enhanced contextual cue text is input into a pre-defined large-scale visual language model for spatiotemporal dual attention chain reasoning, and the output is a fault diagnosis result containing cause analysis and maintenance suggestions.

[0081] In this embodiment, the steps of collecting historical diagnostic log information and machine processing records during the chip manufacturing process to construct a time-series dynamic domain knowledge graph with timestamps and production batch attributes specifically include: acquiring chip design specification document data, historical machine daily maintenance record data, underlying process recipe iteration file data, and wafer fabrication quantity data for the corresponding batch during the operation of the chip manufacturing production line; performing text format conversion and invalid character filtering and cleaning operations on the chip design specification document data, historical machine daily maintenance record data, and underlying process recipe iteration file data to obtain a clean and standardized text content sequence; using a pre-trained named entity recognition algorithm network model to perform deep scanning extraction operations on the clean and standardized text content sequence to obtain a set of semiconductor equipment core component entity data and a set of chip fault typical feature entity data; and using a set dependency parsing algorithm model to perform syntax tree correlation analysis operations on the clean and standardized text content sequence to extract semiconductor equipment core components. The node dependencies between the core component entity data set and the typical chip fault feature entity data set are used to obtain the initial fault evolution logical relationship sequence. The core component entity data set, the typical chip fault feature entity data set, and the initial fault evolution logical relationship sequence are then fused into a topological network to construct a static semiconductor defect fundamental knowledge graph. Regular expression parsing rules are used to analyze historical diagnostic log information and machine processing records, extracting the actual occurrence time node data corresponding to each closed-loop maintenance event record in the historical diagnostic log information and machine processing records, as well as the cumulative wafer processing batch data within the event period. The actual occurrence time node data and the cumulative wafer processing batch data are configured on the corresponding logical relationship vector edges in the initial fault evolution logical relationship sequence of the static semiconductor defect fundamental knowledge graph as absolute occurrence time labels and batch period labels, resulting in a time-series dynamic domain knowledge graph with timestamps and production batch attributes.

[0082] Specifically, the process involves acquiring chip design specification documents, historical machine maintenance records, underlying process recipe iteration files, and wafer fabrication quantity data from the chip manufacturing production line. The acquired chip design specification documents, historical machine maintenance records, and underlying process recipe iteration files are then converted to UTF-8 encoded plain text format and an invalid character filtering and cleaning operation is performed using a predefined set of regular expressions, such as [^\u4e00-\u9fa5a-zA-Z0-9].The code snippet `,:;?!()-]` removes unstructured text, special control characters, and garbled characters to obtain a clean, canonical text sequence. Subsequently, a pre-trained named entity recognition (NID) network model combining a bidirectional long short-term memory (LSTM) network and a conditional random field (CRF) is used to perform deep scanning extraction on the clean, canonical text sequence. This model structure includes a word embedding layer, a bidirectional LSTM layer, and a CRF decoding layer. The input during training is manually annotated semiconductor-related corpus, where terms such as "etching machine" and "chamber pressure" are labeled as equipment component entities, and "particulate contamination" and "bridging" are labeled as fault features. The model outputs a text sequence with entity type labels, resulting in a set of entity data for core semiconductor equipment components and a set of entity data for typical chip fault characteristics. Then, using an algorithm based on dependency parsing, such as the open-source Stanza toolkit, a dependency tree is generated for each sentence in the clean, canonical text sequence. Based on preset syntactic rules, such as extracting a subject-verb-object structure with the fault entity as the object and the equipment entity as the subject, the node dependencies between the set of entity data for core semiconductor equipment components and the set of entity data for typical chip fault characteristics are analyzed. The initial fault evolution logical relationship sequence is used as the nodes of a graph, along with the extracted core component entity dataset of semiconductor equipment, the typical characteristic entity dataset of chip faults, and the initial fault evolution logical relationship sequence as the edges. A topological network fusion process is then performed to construct a static semiconductor defect fundamental knowledge graph. Subsequently, regular expression parsing rules are used, such as EventTime:\s*(\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}) and ProcessedBatches:\s*(\d+). This process involves analyzing historical diagnostic logs and machine processing records to extract the actual occurrence time of each closed-loop maintenance event and the cumulative wafer processing batch data within the event period. Finally, iterates through each logical relationship edge in the static semiconductor defect fundamental knowledge graph. Based on the co-occurrence relationships of the associated entities in the historical diagnostic logs, the extracted actual occurrence time of the closed loop and the cumulative wafer processing batch data are configured on the corresponding logical relationship vector edges as absolute occurrence time labels and batch period labels, resulting in a time-series dynamic domain knowledge graph with timestamps and production batch attributes.

[0083] In this embodiment, the steps of performing graph representation learning on a time-series dynamic domain knowledge graph with timestamps and production batch attributes to generate a set of entity feature vectors and a set of relation feature vectors with confidence decay weights specifically include: completely extracting the full set of network node element features and the full set of network edge element features from the time-series dynamic domain knowledge graph with timestamps and production batch attributes; performing uniformly distributed random initialization encoding on the full set of network node element features and the full set of network edge element features to obtain an initial network node vector parameter matrix and an initial network edge vector parameter matrix; synchronously inputting the initial network node vector parameter matrix and the initial network edge vector parameter matrix into a pre-constructed knowledge graph translation metric model, and performing high-dimensional vector space mapping iterative calculation based on the triplet translation space distance metric mechanism to obtain the target steady-state entity feature vector sequence and the target steady-state relation feature vector sequence; the formula is: ,in This represents the value of the loss function for the translational spatial distance metric of the triplet. The feature set of positive sample triples representing correct logical associations. This represents the positive sample head entity feature parameters, positive sample relation feature parameters, and positive sample tail entity feature parameters corresponding to the positive sample triplet feature set. The feature set of negative sample triples representing erroneous logical associations. This represents the negative sample head entity feature parameters, negative sample relation feature parameters, and negative sample tail entity feature parameters corresponding to the negative sample triplet feature set. The hyperparameter representing the margin between positive and negative samples is... This represents the spatial correlation penalty coefficient between adjacent processes extracted based on the machine tool process flow topology. Represents the Euclidean distance metric function. Represents the head entity encoding tensor corresponding to the head entity feature parameters of the positive sample. This represents the relation encoding tensor corresponding to the relation feature parameters of positive samples. Represents the tail entity encoding tensor corresponding to the tail entity feature parameters of the positive sample. This represents the negative head entity encoding tensor corresponding to the negative sample head entity feature parameters. This represents the negative tail entity encoding tensor corresponding to the negative sample tail entity feature parameters; it obtains the current system diagnosis time scale value, calculates the absolute offset of the time interval between the current system diagnosis time scale value and the actual occurrence time node data of the closed loop in the time-series dynamic domain knowledge graph with timestamp and production batch attributes, and obtains the cumulative number of actual processed wafers corresponding to the absolute offset of the time interval; based on the absolute offset of the time interval and the cumulative number of actual processed wafers, it calculates the dynamic confidence decay weight index of the full network edge element features using a preset exponential aging decay algorithm function; the formula is: ,in Indicates the dynamic confidence decay weight index, This represents the manually set initial full-load confidence scalar value. The base of the natural logarithm mathematical constant, Indicates the inherent aging degradation coefficient constant of the equipment. Indicates the absolute offset of the time interval. Represents the physical fatigue coefficient constant of machine tool processing operations. This represents the cumulative number of wafers actually processed; the target steady-state entity feature vector sequence is filtered by linear fully connected layer feature mapping to obtain the entity feature vector set; the target steady-state relationship feature vector sequence is fused and adjusted by matrix element-wise multiplication with the dynamic confidence decay weight index to obtain the relationship feature vector set with confidence decay weight.

[0084] Specifically, the full set of network node element features and full set of network edge element features with timestamps and production batch attributes are extracted from the time-series dynamic domain knowledge graph. For each element in the full set of network node element features and full set of network edge element features, random initialization encoding is performed using a Xavier uniform distribution to generate a vector of dimension 256, resulting in the initial network node vector parameter matrix and the initial network edge vector parameter matrix. The initial network node vector parameter matrix and the initial network edge vector parameter matrix are synchronously input into a pre-built knowledge graph translation metric model. This model is based on the translation embedding idea, and its training process is as follows: In each training iteration, a set of positive sample triplet features with correct logical associations of a batch is randomly extracted from the knowledge graph. For sets Each triple in By randomly replacing the head entity or tail entity Construct a negative sample triplet with incorrect logical association in this way. The feature set of negative sample triples is formed. Then, optimization calculations are performed based on the following loss function: ,in This represents the value of the loss function that measures the spatial distance of the triplet translation. This represents the feature set of positive sample triples in this batch. To represent a positive sample triple, This represents the feature set of the corresponding negative sample triplet. To represent a negative sample triple, The hyperparameter for the positive and negative sample boundary margin is set to 1.0 based on experience. This refers to the spatial association penalty coefficient between adjacent processes extracted based on the machine tool process flow topology. For example, if two entities are adjacent in the process flow, Set it to 1.2, otherwise set it to 1.0. This represents the Euclidean distance metric function. , , , , Let the encoded tensors of the head entity, relation, tail entity, negative sample head entity, and negative sample tail entity be represented respectively, and the loss function be minimized by the Adam optimizer. This process updates the vector representations of all entities and relations until the model converges, yielding the target steady-state entity feature vector sequence and the target steady-state relation feature vector sequence. Subsequently, the current system diagnostic time scale value is obtained, and the absolute offset of the time interval between it and the actual occurrence time node data of the closed loop on each relation edge in the knowledge graph is calculated. At the same time, obtain the cumulative number of actual wafers processed within that time period. Based on this, the dynamic confidence decay weight index of the features of all network edge elements is calculated using the exponential aging decay algorithm function. The formula is as follows: ,in This represents the dynamic confidence decay weighting index. The initial full-load confidence scalar value is set to 1.0. The base of the natural logarithm, The inherent aging degradation coefficient constant of the equipment is set to 0.01 based on historical equipment data analysis. The physical fatigue coefficient constant for machine processing is set to 0.0005 based on the fit between wafer processing volume and equipment failure rate. The target steady-state entity feature vector sequence is directly filtered by feature mapping through a linear fully connected layer to obtain the entity feature vector set. The target steady-state relationship feature vector sequence is then multiplied element-wise with the calculated dynamic confidence decay weight index to obtain the relationship feature vector set with confidence decay weight.

[0085] In this embodiment, the steps of acquiring target defect image data and target equipment context data of the chip under test, and performing feature extraction to generate a multimodal query feature vector specifically include: acquiring surface micro-morphology image features of the chip under test using a high-resolution electron microscope scanning imaging device to obtain target defect image data, and extracting the two-dimensional physical coordinate data of the chip under test on its respective wafer; synchronously and in real-time reading the external sensor operating status parameters and internal chamber environmental monitoring indicators of the testing equipment where the chip under test is located, and combining them to generate target equipment context data; inputting the target defect image data into a pre-trained deep visual residual coding network for multi-scale convolutional feature extraction operations to obtain a high-dimensional visual feature distribution tensor matrix; and performing global average pooling spatial dimensionality reduction processing on the high-dimensional visual feature distribution tensor matrix to obtain... The densely distributed image feature representation vector is generated; the target machine context data is converted into a long text formatted string feature sequence; the long text formatted string feature sequence is input into the self-attention computation layer of the natural language text encoder for deep semantic feature extraction, resulting in a machine semantic feature representation vector; spatial position encoding compensation processing is performed on the densely distributed image feature representation vector using two-dimensional physical coordinate data, and then cross-modal cross-attention fusion processing is performed on the compensated densely distributed image feature representation vector and the machine semantic feature representation vector in the feature channel numerical dimension to obtain a multimodal cascaded associated feature representation vector; the multimodal cascaded associated feature representation vector is input into a nonlinear fully connected feature fusion network for feature dimensionality reduction and feature semantic space alignment processing to obtain a multimodal query feature vector.

[0086] Specifically, a high-resolution electron microscope scanning imaging device is used to acquire surface microstructure images of the chip under test, obtaining grayscale target defect image data with a resolution of 1024×1024 pixels. The two-dimensional physical coordinates of the chip under test on its wafer are extracted, for example, coordinates (15.3, 28.1). Simultaneously, external sensor operating status parameters of the testing equipment where the chip is located are read in real time, such as chamber temperature, gas flow rate, and RF power, as well as internal chamber environmental monitoring indicators, such as residual gas analyzer readings. These are combined into a key-value pair JSON object as target equipment context data. The target defect image data is input into a deep visual residual coding network, such as ResNet-101, pre-trained on the ImageNet dataset and fine-tuned on a semiconductor defect image dataset, to perform multi-scale convolutional feature extraction. The feature map output by the last convolutional block of the network is extracted, resulting in a 7×7×2048-dimensional high-dimensional visual feature distribution tensor matrix. This high-dimensional visual feature distribution tensor matrix is ​​then used to perform multi-scale convolutional feature extraction. The feature distribution tensor matrix is ​​subjected to global average pooling to compress it spatially, resulting in a 2048-dimensional densely distributed image feature representation vector. The target machine context data, i.e., the JSON object, is serialized into a long text formatted string feature sequence in the format "temperature: 35.2; gas flow rate: 50.1; RF power: 300.5". This long text formatted string feature sequence is input into a pre-trained natural language text encoder, such as the self-attention layer of the BERT-base model, for deep semantic feature extraction. The output vector corresponding to its [CLS] label is then taken to obtain a 768-dimensional machine semantic feature representation vector. Using two-dimensional physical coordinate data (15.3, 28.1) Generate a 2048-dimensional location encoding vector. This encoding vector encodes the coordinate values ​​by alternately using sine and cosine functions. Then, it is added element-wise to the densely distributed image feature representation vector to complete the spatial location encoding compensation process. Subsequently, the 2048-dimensional densely distributed image feature representation vector after compensation is used as the query, and the 768-dimensional machine semantic feature representation vector is used as the key and value. Cross-modal cross-attention fusion processing is performed on the numerical dimension of the feature channels to obtain a 2048-dimensional fused feature vector. This vector is then concatenated with the original machine semantic feature representation vector to obtain a 2816-dimensional multimodal cascaded associated feature representation vector. Finally, this multimodal cascaded associated feature representation vector is input into a nonlinear fully connected feature fusion network containing two fully connected layers. The first layer reduces the dimension from 2816 to 1024, and the second layer further reduces it to 256. The ReLU activation function is used between layers to perform feature dimensionality reduction and feature semantic space alignment operations to obtain the multimodal query feature vector.

[0087] In this embodiment, the steps of performing correlation matching retrieval in the entity feature vector set and the relation feature vector set with confidence decay weights based on multimodal query feature vectors to obtain fault knowledge retrieval results specifically include: extracting each local entity feature vector to be matched from the entity feature vector set one by one; calculating the cosine angle mapping distance between the multimodal query feature vector and each local entity feature vector to be matched to obtain the basic entity semantic similarity measurement score; extracting each local relation feature vector to be matched from the relation feature vector set with confidence decay weights one by one; calculating the inner product projection mathematical metric between the multimodal query feature vector and each local relation feature vector to be matched to obtain the basic relation pattern matching metric score; obtaining the graph theory centrality metric of the adjacency relation network feature vector corresponding to each local entity feature vector to be matched in the time-series dynamic domain knowledge graph with timestamp and production batch attributes; and using the graph theory centrality metric to perform joint weighted fusion calculation of the basic entity semantic similarity measurement score and the basic relation pattern matching metric score to obtain the comprehensive knowledge retrieval confidence score; the formula is: ,in This represents the total confidence score for comprehensive knowledge retrieval. The standard logarithmic function with base 2 Represents the graph theory centrality metric. This indicates the weighting coefficient for entity matching confidence. Indicates the basic entity semantic similarity measurement score, Indicates the weighting coefficient of relationship matching confidence. This represents the dynamic confidence decay weight index carried by each feature vector of the local relation to be matched. The basic relation pattern matching metric score is represented; the total confidence score of the comprehensive knowledge retrieval is arranged in descending order from high to low, resulting in a descending confidence score ranking list; the target entity knowledge text fragments and target relation knowledge text fragments corresponding to the top ten scores in the descending confidence score ranking list are extracted; the target entity knowledge text fragments and target relation knowledge text fragments are processed by natural language grammar rules to generate sentences, resulting in the fault knowledge retrieval results.

[0088] Specifically, each local entity feature vector to be matched is extracted by iteratively traversing the entity feature vector set obtained in the previous steps. The cosine similarity between the 256-dimensional multimodal query feature vector and each 256-dimensional local entity feature vector to be matched is calculated and used as the basic entity semantic similarity measurement score. Its value range is [-1, 1]. Then, each local relation feature vector to be matched is extracted by iteratively traversing the set of relation feature vectors with confidence decay weights. The dot product between the multimodal query feature vector and each local relation feature vector to be matched is calculated as the basic relation pattern matching metric score. Obtain the graph theory center metric of each local entity feature vector to be matched in the temporal dynamic domain knowledge graph. Specifically, this involves calculating the degree centrality of each entity node, which is the number of edges directly connected to that node, and then using the calculated graph theory centrality metric. Basic entity semantic similarity measurement score Matching metric score with underlying relation patterns A joint weighted fusion calculation is performed to obtain the total confidence score of the comprehensive knowledge retrieval. The calculation formula is: ,in This represents the total confidence score for comprehensive knowledge retrieval. It is a logarithmic function with base 2. The degree centrality index of a node, and The confidence weights for entity and relationship matching are assigned coefficients, which are then optimized on the validation set using a grid search. Set it to 0.6. Set it to 0.4. This represents the dynamic confidence decay weight index carried by each feature vector of the local relation to be matched. This represents the basic entity semantic similarity measurement score. This represents the underlying relational pattern matching metric score; for example, for an entity, its... It is 0.85. The value is 15, and one of the associated relationships is... It is 0.9. If the value is 1.2, then its fraction is The calculated total confidence scores for comprehensive knowledge retrieval are arranged in descending order to obtain a descending confidence score ranking list. The original text descriptions of the entities and relations corresponding to the top 10 scores in this descending confidence score ranking list are extracted, namely, target entity knowledge text fragments and target relation knowledge text fragments. Finally, these extracted target entity knowledge text fragments and target relation knowledge text fragments are processed according to a preset template, such as "the defect '{entity A}' may be caused by '{entity B}' in the process of '{relation}'", and then processed by natural language grammar rules to generate the fault knowledge retrieval results.

[0089] In this embodiment, the step of fusing target defect image data, target machine context data, and fault knowledge retrieval results into enhanced contextual prompt text specifically includes: configuring a standard system-level prompt guidance execution instruction template, which includes expert role setting text placeholders and inference task description text placeholders; completely filling the target machine context data and the two-dimensional physical coordinate data of the wafer to which it belongs into the mapping area corresponding to the inference task description text placeholders of the standard system-level prompt guidance execution instruction template to obtain the basic background task prompt condition string; and performing pixel normalization bilinear scaling processing on the target defect image data and converting it into a target... The text is escaped and labeled with image tags; these tags are then embedded into the reserved image placeholders in the basic background task prompt string to obtain the multimodal visual basic prompt string; key defect cause phrase fragments and key equipment maintenance guidance phrase fragments are extracted from the fault knowledge retrieval results; these fragments are then rearranged according to their actual occurrence time sequence to generate a reference expert knowledge reasoning context; this reference expert knowledge reasoning context is directly appended to the end region of the multimodal visual basic prompt string as an auxiliary visual reasoning judgment condition to obtain the enhanced context prompt text.

[0090] Specifically, a standard system-level prompt instruction template is configured. This template is a string of text containing specific placeholders, such as: "You are a senior semiconductor fault diagnosis expert. Please analyze the root cause of the fault and provide repair suggestions based on the following background information, defect image, and related knowledge. Background information: {context}, Defect image: {image}, Related knowledge: {knowledge}", where {context}, {image}, and {knowledge} are placeholders. The target machine context data and the two-dimensional physical coordinate data of the wafer obtained in the previous steps are integrated into a descriptive text, such as "Machine ID: ETCH-03, Process recipe: P-101, Chamber pressure: 1.2Pa, Wafer coordinates: " (15.3,28.1)”, and fill this text completely into the mapping area corresponding to the {context} placeholder of the standard system-level prompt execution instruction template to obtain the basic background task prompt condition string. Normalize the pixel values ​​of the target defect image data, then scale it to 224×224 pixels using bilinear interpolation, and encode it into a Base64 format string. Add specific image tagging labels before and after this string, for example...<image_start> and<image_end> This process generates text-to-image tags, which are then precisely embedded into the specified positions of the {image} placeholders in the basic background task prompt string to obtain the multimodal visual basic prompt string. Key defect cause phrases, such as "uneven etching rate," and key equipment maintenance guidance phrases, such as "check the gas spray head," are extracted from the fault knowledge retrieval results obtained in the previous steps. These phrases are then rearranged according to their time tags associated with them in the knowledge graph, following the chronological order of their actual occurrence, to generate a coherent text that serves as the reference expert knowledge reasoning context. For example, "Historical records show that uneven etching rate occurred 3 days ago, and the associated operation was checking the gas spray head." Finally, this reference expert knowledge reasoning context is directly appended to the end of the multimodal visual basic prompt string, specifically filling the {knowledge} placeholder, as an auxiliary visual reasoning judgment condition, resulting in the enhanced context prompt text.

[0091] In this embodiment, the step of inputting enhanced contextual prompt text into a preset large-scale visual language model for spatiotemporal dual attention chain reasoning, and outputting fault diagnosis results including cause analysis and maintenance suggestions, specifically includes: using a word segmenter component to perform word segmentation rule mapping processing on the enhanced contextual prompt text to obtain a discrete language tag unit sequence; inputting the discrete language tag unit sequence into the word embedding encoding layer of the preset large-scale visual language model for high-dimensional continuous spatial projection mapping to obtain a multi-dimensional word embedding representation feature tensor; and extracting the key defect cause phrase fragments corresponding to the fault knowledge retrieval results. The system uses auxiliary reference time node attribute data and extracts 2D physical coordinate data from the enhanced contextual cue text. Based on the auxiliary reference time node attribute data, it constructs a target-dimensional time-aware bias attention mask matrix and a wafer spatial location bias attention mask matrix based on the 2D physical coordinate data. The multi-dimensional word embedding representation feature tensor, the time-aware bias attention mask matrix, and the wafer spatial location bias attention mask matrix are jointly input into the multi-head spatiotemporal attention calculation module of a pre-defined large-scale visual language model to calculate the global weight attention allocation probability, thus obtaining the global spatiotemporal correlation feature latent variables. The formula is: ,in Representing global spatiotemporal correlation features latent variables, Represents the normalized exponential activation mapping function, This represents the query feature parameter matrix corresponding to the multi-dimensional word embedding representation feature tensor after linear transformation. The transpose of the key feature parameter matrix after linear transformation of the multi-dimensional word embedding representation feature tensor is represented by the matrix. The key feature parameter matrix represents the fixed dimension of the hidden feature layer. Represents the time-aware biased attention mask matrix, This represents the wafer spatial location bias attention mask matrix. This involves a feature parameter matrix representing the value of a multi-dimensional word embedding representation feature tensor after linear transformation; inputting global spatiotemporal correlation feature latent variables into a pre-defined large-scale visual language model's feedforward fully connected neural network module for nonlinear high-order feature transformation processing to obtain high-level abstract logical semantic fusion features; inputting these high-level abstract logical semantic fusion features into a pre-defined large-scale visual language model's autoregressive text decoding module for stepwise word distribution probability prediction sampling to obtain an optimized diagnostic word prediction output sequence; dividing the optimized diagnostic word prediction output sequence into a defect core cause analysis paragraph text block and an equipment maintenance guidance paragraph text block according to the natural language semantic paragraph structure; directly using the defect core cause analysis paragraph text block as the cause analysis and the equipment maintenance guidance paragraph text block as the maintenance suggestion; and merging the cause analysis and maintenance suggestion according to the pre-defined standard fault result formatting requirements to obtain a fault diagnosis result containing both cause analysis and maintenance suggestion.

[0092] Specifically, a word segmenter component based on a byte-pair encoding algorithm is used to perform word segmentation rule mapping on the enhanced contextual prompt text, resulting in a discrete sequence of language tag units composed of integer IDs. This discrete sequence of language tag units is input into the word embedding encoding layer of a pre-defined large-scale visual language model, such as the LLaVA-1.5 model. By looking up the embedding matrix, each tag ID is mapped to a 1024-dimensional high-dimensional vector. A high-dimensional continuous space projection mapping operation is then performed to obtain a multi-dimensional word embedding representation feature tensor. Auxiliary reference occurrence time node attribute data corresponding to key defect cause phrase fragments in the fault knowledge retrieval results is extracted, and two-dimensional physical coordinate data contained in the enhanced contextual prompt text is extracted. Based on the auxiliary reference occurrence time node attribute data, the relative time difference between each word and the current diagnosis time is calculated and mapped to a 1024-dimensional time encoding vector. The time encoding vectors of all words form a time-aware bias attention mask matrix of a specific dimension. Similarly, based on the two-dimensional physical coordinate data, a 1024-dimensional spatial location encoding vector is also generated for each word, forming a wafer spatial location bias attention mask matrix. Multi-dimensional word embedding representation feature tensors and time-aware biased attention mask matrices are used to embed words into the representation feature tensor. And wafer spatial location bias attention mask matrix The calculation process in the multi-head spatiotemporal attention module of the large-scale visual language model, which is jointly input into it, follows the formula below: ,in This represents the latent variables representing the global spatiotemporal correlation features of the output. , , These represent the query, key, and value feature parameter matrices obtained by transforming the multi-dimensional word embedding representation feature tensor through different linear transformations. It is the transpose of the key feature parameter matrix. This refers to the dimension of the hidden feature layer in the key feature parameter matrix, which is 64 in this case. For normalized exponential activation function, and The bias matrix, directly added to the attention score, is used to input the obtained global spatiotemporal correlation feature latent variables into the model's feedforward fully connected neural network module for nonlinear high-order feature transformation, resulting in high-level abstract logical semantic fusion features. These features are then input into the model's autoregressive text decoding module, where a beam search decoding strategy is employed to predict probability distributions word by word and select the optimal sequence, yielding the optimal diagnostic vocabulary prediction output sequence. This sequence is then divided into paragraphs based on preset keywords, such as "Cause Analysis:" and "Maintenance Suggestions:", resulting in a core defect cause analysis paragraph and an equipment maintenance guidance paragraph. The core defect cause analysis paragraph is directly used as the cause analysis, and the equipment maintenance guidance paragraph is directly used as the maintenance suggestion. Finally, the cause analysis and maintenance suggestions are merged and output according to a preset report template to obtain a fault diagnosis result containing both cause analysis and maintenance suggestions.

[0093] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for chip defect detection and fault diagnosis, characterized in that, Includes the following steps: Collect historical diagnostic logs and machine processing records during the chip manufacturing process to construct a time-series dynamic domain knowledge graph with timestamps and production batch attributes; The time-series dynamic domain knowledge graph with timestamps and production batch attributes is subjected to graph representation learning to generate a set of entity feature vectors and a set of relation feature vectors with confidence decay weights. Acquire the target defect image data and target machine context data of the chip to be inspected, and perform feature extraction to generate a multimodal query feature vector; Based on the multimodal query feature vector, a correlation degree matching retrieval is performed in the entity feature vector set and the relation feature vector set with confidence decay weight to obtain fault knowledge retrieval results; The target defect image data, the target machine context data, and the fault knowledge retrieval results are fused into an enhanced contextual prompt text; The enhanced contextual prompt text is input into a preset large-scale visual language model for spatiotemporal dual attention chain reasoning, and the output is a fault diagnosis result containing cause analysis and maintenance suggestions.

2. The chip defect detection and fault diagnosis method according to claim 1, characterized in that, The steps involved in collecting historical diagnostic logs and machine processing records during chip manufacturing to construct a time-series dynamic domain knowledge graph with timestamps and production batch attributes include: Acquire chip design specification documents, historical machine daily maintenance records, underlying process formula iteration documents, and wafer fabrication quantity data for the corresponding batch during the operation of the chip manufacturing production line; The chip design specification document data, the historical machine daily maintenance record data, and the underlying process recipe iteration file data are converted to text format and invalid character filtering and cleaning operations are performed to obtain a clean specification text content sequence. The clean, canonical text content sequence is subjected to deep scanning extraction using a pre-trained named entity recognition algorithm network model to obtain a set of entity data for core components of semiconductor equipment and a set of entity data for typical characteristics of chip faults. Using the established dependency parsing algorithm model, a syntax tree correlation analysis operation is performed on the clean and canonical text content sequence to extract the node dependency relationship between the core component entity data set of the semiconductor equipment and the typical feature entity data set of chip faults, thereby obtaining the initial fault evolution logical relationship sequence. The core component entity data set of the semiconductor device, the typical characteristic entity data set of chip faults, and the initial fault evolution logical relationship sequence are fused and processed to construct a static semiconductor defect basic knowledge graph. The historical diagnostic log information and the machine processing record are parsed using regular expression parsing rules. The actual occurrence time node data of each closed-loop maintenance event record in the historical diagnostic log information and the machine processing record, as well as the cumulative wafer processing batch data within the event period, are extracted. The actual occurrence time node data of the closed loop and the cumulative wafer processing batch data are configured on the corresponding logical relationship vector edge in the initial fault evolution logical relationship sequence in the static semiconductor defect basic knowledge graph as absolute occurrence time label and batch period label, to obtain the time-series dynamic domain knowledge graph with timestamp and production batch attributes.

3. The chip defect detection and fault diagnosis method according to claim 2, characterized in that, The steps of performing graph representation learning on the time-series dynamic domain knowledge graph with timestamps and production batch attributes to generate a set of entity feature vectors and a set of relation feature vectors with confidence decay weights specifically include: Completely extract the full network node element features and full network edge element features from the time-series dynamic domain knowledge graph with timestamp and production batch attributes; The full set of network node element features and the full set of network edge element features are subjected to uniformly distributed random initialization encoding to obtain the initial network node vector parameter matrix and the initial network edge vector parameter matrix. The initial network node vector parameter matrix and the initial network edge vector parameter matrix are synchronously input into the pre-constructed knowledge graph translation metric model. Based on the triplet translation space distance metric mechanism, a high-dimensional vector space mapping iterative calculation is performed to obtain the target steady-state entity feature vector sequence and the target steady-state relation feature vector sequence. The formula is: ,in This represents the value of the loss function for the translational spatial distance metric of the triplet. The feature set of positive sample triples representing correct logical associations. This represents the positive sample head entity feature parameters, positive sample relation feature parameters, and positive sample tail entity feature parameters corresponding to the positive sample triplet feature set. The feature set of negative sample triples representing erroneous logical associations. This represents the negative sample head entity feature parameters, negative sample relation feature parameters, and negative sample tail entity feature parameters corresponding to the negative sample triplet feature set. The hyperparameter representing the margin between positive and negative samples is... This represents the spatial correlation penalty coefficient between adjacent processes extracted based on the machine tool process flow topology. Represents the Euclidean distance metric function. This represents the head entity encoding tensor corresponding to the positive sample head entity feature parameters. This represents the relation encoding tensor corresponding to the relation feature parameters of the positive samples. This represents the tail entity encoding tensor corresponding to the tail entity feature parameters of the positive sample. This represents the negative head entity encoding tensor corresponding to the negative sample head entity feature parameters. This represents the negative tail entity encoding tensor corresponding to the negative sample tail entity feature parameters; Obtain the current system diagnostic time scale value, calculate the absolute offset of the time interval between the current system diagnostic time scale value and the actual occurrence time node data of the closed loop in the time-series dynamic domain knowledge graph with timestamp and production batch attributes, and obtain the cumulative number of actual processed wafers corresponding to the absolute offset of the time interval. Based on the absolute offset of the time interval and the cumulative number of actual processed wafers, the dynamic confidence decay weight index of the full network edge element features is calculated using a preset exponential aging decay algorithm function. The formula is: ,in This indicates the dynamic confidence decay weight index, This represents the manually set initial full-load confidence scalar value. The base of the natural logarithm mathematical constant, Indicates the inherent aging degradation coefficient constant of the equipment. Indicates the absolute offset of the time interval, Represents the physical fatigue coefficient constant of machine tool processing operations. This indicates the cumulative number of wafers actually processed; The target steady-state entity feature vector sequence is filtered by a linear fully connected layer feature mapping to obtain the entity feature vector set; The target steady-state relationship feature vector sequence is fused and adjusted by performing matrix element-wise multiplication with the dynamic confidence decay weight index to obtain the relationship feature vector set with confidence decay weight.

4. The chip defect detection and fault diagnosis method according to claim 1, characterized in that, The steps of acquiring the target defect image data of the chip to be inspected and the target machine context data, and performing feature extraction to generate a multimodal query feature vector specifically include: The surface microstructure image features of the chip under test are acquired by a high-resolution electron microscope scanning imaging device to obtain the target defect image data, and the two-dimensional physical coordinate data of the chip under test on its respective wafer are extracted. Simultaneously read the external sensor operating status parameters and internal chamber environment monitoring indicators of the test equipment where the chip under test is located in real time, and combine them to generate the target equipment context data; The target defect image data is input into a pre-trained deep visual residual coding network to perform multi-scale convolutional feature extraction operations, resulting in a high-dimensional visual feature distribution tensor matrix. The high-dimensional visual feature distribution tensor matrix is ​​subjected to global average pooling spatial dimensionality reduction processing to obtain a densely distributed image feature representation vector; Convert the target machine context data into a long text formatted string feature sequence; The long text formatted string feature sequence is input into the self-attention computation layer of the natural language text encoder for deep semantic feature extraction to obtain the machine semantic feature representation vector; The densely distributed image feature representation vector is spatially encoded and compensated using the two-dimensional physical coordinate data. Then, the compensated densely distributed image feature representation vector and the machine semantic feature representation vector are fused together with cross-modal cross-attention in the feature channel numerical dimension to obtain a multimodal cascaded associated feature representation vector. The multimodal cascaded associated feature representation vector is input into a nonlinear fully connected feature fusion network for feature dimensionality reduction and feature semantic space alignment to obtain the multimodal query feature vector.

5. The chip defect detection and fault diagnosis method according to claim 3, characterized in that, The steps for obtaining fault knowledge retrieval results by performing correlation matching retrieval in the entity feature vector set and the relation feature vector set with confidence decay weights based on the multimodal query feature vectors specifically include: Extract each local entity feature vector to be matched by iterating through the set of entity feature vectors one by one; Calculate the cosine angle mapping distance between the multimodal query feature vector and the feature vector of each local entity to be matched, and obtain the basic entity semantic similarity measurement score; Each local relation feature vector to be matched is extracted by iterating through the set of relation feature vectors with confidence decay weights one by one. Calculate the mathematical metric of the inner product projection between the multimodal query feature vector and each local relation feature vector to be matched, and obtain the basic relation pattern matching metric score; Obtain the graph theory center metric of the adjacency network feature vector corresponding to each local entity feature vector to be matched in the temporal dynamic domain knowledge graph with timestamp and production batch attributes; The graph theory center metric is used to perform a joint weighted fusion calculation on the basic entity semantic similarity measurement score and the basic relation pattern matching metric score to obtain the comprehensive knowledge retrieval confidence score. The formula is: ,in This represents the total confidence score of the comprehensive knowledge retrieval. The standard logarithmic function with base 2 This represents the graph theory centrality metric index. This indicates the weighting coefficient for entity matching confidence. This represents the semantic similarity score of the basic entities. Indicates the weighting coefficient of relationship matching confidence. This refers to the dynamic confidence decay weight index carried by each feature vector of the local relation to be matched. This represents the basic relation pattern matching metric score; The total confidence scores of the comprehensive knowledge retrieval are arranged in descending order from high to low to obtain a descending confidence score ranking list. Extract the target entity knowledge text fragments and target relationship knowledge text fragments corresponding to the top ten scores in the descending confidence score ranking list; The target entity knowledge text fragment and the target relation knowledge text fragment are processed by concatenating sentences using natural language grammar rules to obtain the fault knowledge retrieval result.

6. The chip defect detection and fault diagnosis method according to claim 4, characterized in that, The step of fusing the target defect image data, the target machine context data, and the fault knowledge retrieval results into enhanced contextual prompt text specifically includes: Configure a standard system-level prompt and guidance execution instruction template, which includes expert role setting text placeholders and reasoning task description text placeholders; The target machine context data and the two-dimensional physical coordinate data of the wafer to which it belongs are completely filled into the mapping area corresponding to the inference task description text placeholder of the standard system-level prompt guidance execution instruction template to obtain the basic background task prompt condition string; The target defect image data is subjected to pixel normalization and bilinear scaling, and then converted into text escape image tagging labels in the target format; The text escape image tag is embedded into the reserved image placeholder specified in the basic background task prompt condition string to obtain the multimodal visual basic prompt string; Extract key defect cause phrase fragments and key equipment maintenance guidance phrase fragments from the fault knowledge retrieval results; The phrase fragments describing the causes of key defects and the phrase fragments describing the maintenance guidance for key equipment are rearranged and arranged in chronological order according to the actual occurrence time to generate a context for reasoning based on expert knowledge. The reference expert knowledge reasoning context is directly appended to the end region of the multimodal visual basic prompt string as an auxiliary visual reasoning judgment condition to obtain the enhanced context prompt text.

7. The chip defect detection and fault diagnosis method according to claim 6, characterized in that, The steps of inputting the enhanced contextual cue text into a preset large-scale visual language model for spatiotemporal dual attention chain reasoning, and outputting fault diagnosis results including cause analysis and maintenance suggestions, specifically include: The enhanced contextual prompt text is processed by word segmentation rule mapping using a sub-word segmenter component to obtain a discrete language tag unit sequence; The discrete language tagging unit sequence is input into the word embedding encoding layer of the preset large-scale visual language model to perform high-dimensional continuous space projection mapping operation, thereby obtaining a multi-dimensional word embedding representation feature tensor. Extract the auxiliary reference occurrence time node attribute data corresponding to the key defect cause phrase fragment in the fault knowledge retrieval results, and extract the two-dimensional physical coordinate data contained in the enhanced contextual prompt text; A time-aware bias attention mask matrix for the target dimension is constructed based on the auxiliary reference occurrence time node attribute data, and a wafer spatial position bias attention mask matrix is ​​constructed based on the two-dimensional physical coordinate data. The multi-dimensional word embedding representation feature tensor, the time-aware bias attention mask matrix, and the wafer spatial location bias attention mask matrix are jointly input into the multi-head spatiotemporal attention calculation module of the preset large-scale visual language model to calculate the global weight attention allocation probability and obtain the global spatiotemporal correlation feature latent variable. The formula is: ,in Represents the global spatiotemporal correlation feature latent variables, Represents the normalized exponential activation mapping function, This represents the query feature parameter matrix corresponding to the multi-dimensional word embedding representation feature tensor after linear transformation. The transpose of the key feature parameter matrix corresponding to the multi-dimensional word embedding representation feature tensor after linear transformation is represented by the transpose operation matrix. This indicates that the feature hidden layer corresponding to the key feature parameter matrix has a fixed dimension. This represents the time-aware biased attention mask matrix. This represents the wafer spatial location bias attention mask matrix. This represents the value feature parameter matrix corresponding to the multi-dimensional word embedding representation feature tensor after linear transformation; The global spatiotemporal correlation feature latent variables are input into the feedforward fully connected neural network module of the preset large-scale visual language model to perform nonlinear high-order feature transformation processing to obtain high-level abstract logical semantic fusion features. The high-level abstract logical semantic fusion features are input into the autoregressive text decoding module of the preset large-scale visual language model to perform stepwise word distribution probability prediction sampling, so as to obtain the optimal diagnostic word prediction output sequence. The optimized diagnostic vocabulary prediction output sequence is divided into two text blocks according to the natural language semantic paragraph structure: a defect core cause analysis paragraph and a device maintenance and upkeep guidance paragraph. The text block analyzing the core causes of the defects is directly used as the cause analysis, and the text block providing equipment maintenance and repair guidance is directly used as the repair suggestion. The cause analysis and the repair suggestions are merged and output according to the preset standard fault result layout format to obtain the fault diagnosis result containing the cause analysis and repair suggestions.

Images