Performance optimization method and device of intelligent manufacturing chip based on historical data

Abstract

The present invention relates to a performance optimization method and device for intelligent manufacturing chips based on historical data. The method: data cleaning and time-series sorting are performed on historical processing data collected by intelligent manufacturing chip manufacturing equipment to obtain a process parameter time-series data set, and a process semantic feature matrix is ​​generated according to the production process document; a process parameter optimization model is created based on the process parameter time-series data set and the process semantic feature matrix; parameter compensation is calculated for the real-time collected equipment status data and quality inspection data based on the process parameter optimization model to obtain an optimized process parameter combination; the optimal capacity allocation plan is solved based on the optimized process parameter combination, and a process parameter dynamic control instruction is output. The present invention realizes the optimal allocation of production capacity while ensuring the constraints of process parameters, thereby improving production efficiency.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a performance optimization method and device for an intelligent manufacturing chip based on historical data. Background Art

[0002] In the production process of smart manufacturing chips, the optimization and control of process parameters has an important impact on product quality and production efficiency. Traditional process parameter optimization methods mainly rely on expert experience and fixed process specifications, which are difficult to adapt to the complex and changing production environment and the differentiated needs of different batches of products.

[0003] With the development of intelligent manufacturing technology, a large amount of historical processing data and production process documents have been accumulated, which contain rich process knowledge and parameter optimization experience. However, due to the heterogeneity and complexity of the data, how to effectively integrate and utilize these multimodal data and build an intelligent process parameter optimization model has become an urgent problem to be solved. Summary of the invention

[0004] The main purpose of the present invention is to provide a performance optimization method and device for intelligent manufacturing chips based on historical data. The present invention achieves optimal allocation of production capacity while ensuring process parameter constraints, thereby improving production efficiency.

[0005] To achieve the above object, the present invention provides a method for optimizing the performance of an intelligent manufacturing chip based on historical data, comprising the following steps: Perform data cleaning and time series sorting on the historical processing data collected by smart manufacturing chip manufacturing equipment to obtain a process parameter time series data set, and generate a process semantic feature matrix based on the production process document; Creating a process parameter optimization model based on the process parameter time series data set and the process semantic feature matrix; Based on the process parameter optimization model, parameter compensation is calculated for the real-time collected equipment status data and quality inspection data to obtain an optimized process parameter combination; The optimal capacity allocation plan is solved based on the optimized process parameter combination, and a process parameter dynamic control instruction is output.

[0006] The present invention also provides a performance optimization device for intelligent manufacturing chips based on historical data, comprising: The preprocessing module is used to clean and sort the historical processing data collected by the intelligent manufacturing chip manufacturing equipment, obtain the process parameter time series data set, and generate the process semantic feature matrix according to the production process document; A creation module, used for creating a process parameter optimization model based on the process parameter time series data set and the process semantic feature matrix; A calculation module, used to calculate parameter compensation amounts for the equipment status data and quality inspection data collected in real time based on the process parameter optimization model, to obtain an optimized process parameter combination; The output module is used to solve the optimal capacity allocation plan based on the optimized process parameter combination and output the process parameter dynamic control instruction.

[0007] The present invention also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the above methods when executing the computer program.

[0008] The present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of any of the above-mentioned methods are implemented.

[0009] In summary, the technical solution provided by the present invention establishes a complete process knowledge representation system by cleaning and time-series sorting the historical processing data, combined with the semantic feature extraction of the production process documents, so that the process parameter optimization process can make full use of historical experience and expert knowledge. The cross-modal feature encoding network is used to realize the unified representation of process parameter time series data and process semantic features, and the expressive ability of the features is enhanced through the comparative learning strategy, which improves the adaptability of the parameter optimization model to different working conditions. Based on the feature decoupling mechanism of the self-encoder reconstruction network, the effective separation of process parameter features and quality indicator features is realized, and the accuracy of parameter optimization is improved through multi-task joint optimization. A dynamic parameter compensation mechanism is designed, which can quickly adjust the process parameters according to the real-time collected equipment status and quality inspection data, thereby enhancing the stability of the production process. The optimization solution method based on Lagrangian relaxation and KKT conditions is adopted to achieve the optimal allocation of production capacity while ensuring the constraints of process parameters, thereby improving production efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 This is a schematic diagram of the steps of a method for optimizing the performance of an intelligent manufacturing chip based on historical data in one embodiment of the present invention; Figure 2 It is a structural block diagram of a performance optimization device for intelligent manufacturing chips based on historical data in one embodiment of the present invention; Figure 3 It is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention.

[0011] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0012] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0013] Reference Figure 1 This embodiment provides a performance optimization method for an intelligent manufacturing chip based on historical data, comprising the following steps: S1, data cleaning and time series sorting of historical processing data collected by intelligent manufacturing chip manufacturing equipment to obtain a process parameter time series data set, and generate a process semantic feature matrix based on the production process document; Among them, the sampling cycle alignment processing is performed on the process parameters, equipment operation status and quality indicator data of intelligent manufacturing chip manufacturing equipment to unify the time scale of data from different sources to ensure the time series consistency of subsequent analysis. After sampling alignment, the original data set is obtained. The outlier detection method (such as the z-score method based on statistical rules or the isolation forest algorithm based on machine learning) is used to identify and eliminate the numerical data in the original data set to obtain a preliminary cleaned data set. The missing values ​​in the preliminary cleaned data set are interpolated based on time series correlation, such as using linear interpolation, spline interpolation or time series prediction model to fill the gaps in the data and convert it into a complete numerical sequence. After obtaining the complete numerical sequence, it is sorted in ascending order by timestamp to ensure logical continuity in the time dimension. On this basis, the sliding window technology is used to segment the data. The size of the sliding window is determined according to the process cycle or production batch. This segmentation method can extract process parameter information with a fixed time span to form a process parameter time series data set. At the same time, the production process documents are subjected to text analysis. The process documents are segmented and non-essential characters such as punctuation are eliminated to form the original text sequence. The original text sequence is segmented using rule-based methods (such as regular expressions) or natural language processing tools (such as NLTK or SpaCy). Through word segmentation, process terms and parameter descriptions in process documents are effectively extracted and converted into structured text data. Based on structured text data, a mapping table between process terms and parameters is established to clarify the correspondence between semantics and parameters in the process flow. On this basis, the process description is normalized to standardize the process information with the same or similar expressions into a consistent expression form to generate a standardized text collection. The standardized text collection is organized according to the order of the process flow, and an index structure is constructed to form a process knowledge corpus. A process label system is constructed using the process knowledge corpus. The system classifies the parameters and quality indicators in the process description, generates label templates, and integrates the relevant information in a templated manner. These templated labels are processed by using the SBERT (Sentence-BERT) semantic encoder to convert the text into a high-dimensional vector form. SBERT can capture the semantic information in the text, so that the semantic association between parameter descriptions and quality indicators can be reflected. The generated vector matrix is ​​the process semantic feature matrix.

[0014] In this embodiment, the process terms and parameter descriptions in the process knowledge corpus are subjected to topic clustering analysis. The text content is analyzed by a clustering algorithm (such as hierarchical clustering or K-means clustering) to identify process parameter category information with similar semantics, reflecting the functional relevance and semantic logic between parameters. A parameter description label template is constructed based on the process parameter category information. The name, value range and adjustment rules of the process parameters are structured and encoded using information extraction technology and a rule engine. By parsing the semantic expression of the process parameters and combining with a specific coding standard, a formalized parameter description label is generated. At the same time, information extraction is performed on the quality description in the process knowledge corpus to extract key information such as target specifications, allowable deviations and detection methods. The parsing of the quality description is achieved by using named entity recognition (NER) technology or a method based on dependency syntax analysis to automatically identify key content describing product quality from the document. After the extracted information is standardized, a quality indicator label is generated. The quality indicator label contains constraints and detection basis for product quality, and together with the parameter description label, it constitutes the basis for process optimization. The parameter description label is merged with the quality indicator label to generate a standardized process text sequence. The merged text sequence can fully express the relationship between the parameter adjustment rules and the quality objectives in the process. On this basis, the SBERT (Sentence-BERT) model is used to encode the standardized process text sequence into word vectors. SBERT uses a pre-trained bidirectional Transformer architecture to perform deep semantic modeling on the input text and generate a high-dimensional initial semantic feature vector. These feature vectors can capture fine-grained semantic information in the text while retaining the contextual relationship between labels. In order to strengthen the correlation characteristics between parameters, an attention weight matrix is ​​constructed based on the initial semantic feature vector. By introducing the attention mechanism, important feature weights related to specific parameters are assigned to them, so that the correlation features between parameters are weighted and an enhanced semantic feature vector is generated. In order to improve computational efficiency and eliminate redundant information, the enhanced semantic feature vector is dimensionally reduced and feature selected, which is completed by PCA (principal component analysis) or embedding-based feature selection algorithm, and the semantic features that have the greatest impact on process optimization are extracted from it to obtain a process semantic feature matrix.

[0015] S2, creating a process parameter optimization model based on the process parameter time series data set and the process semantic feature matrix; Specifically, the process parameter time series data set is input into the time series encoding branch of the cross-modal feature encoding network, which uses time series modeling technology (such as LSTM, GRU or Transformer, etc.) to extract time series features to capture the dynamic changes of process parameters in the time dimension and generate a time series feature vector. At the same time, the process semantic feature matrix is ​​input into the semantic encoding branch of the cross-modal feature encoding network, and the process semantic information is transformed through the embedding layer or the Transformer-based semantic modeling module to obtain a semantic feature vector. The time series feature vector is decomposed by wavelet, and the time series features are divided into high-frequency, medium-frequency and low-frequency feature components using wavelet transform to capture the feature changes in different frequency ranges. The high-frequency component reflects the information of rapid parameter changes, the medium-frequency component represents the medium fluctuation feature, and the low-frequency component describes the overall trend feature. By combining these components into enhanced time series features, the modeling capability of complex process changes is improved. At the same time, for the semantic feature vector, a K nearest neighbor graph structure is constructed based on the semantic association between feature nodes. In this structure, nodes represent semantic features, edges represent the similarity between nodes, and the association strength between nodes is quantified by calculating the cosine similarity matrix. A two-layer graph convolutional network is used to perform message passing and feature aggregation on the K nearest neighbor graph. This process captures the contextual relationship of semantic features by updating node features to generate enhanced semantic features. The enhanced temporal features and enhanced semantic features are input into the feature alignment module, which calculates the attention weights between modalities through a cross-modal attention mechanism to align the semantic and temporal information between different modal features. The feature alignment module generates initial alignment features after dynamically adjusting the weights of temporal and semantic features, and these alignment features are optimized through contrastive learning. In contrastive learning, the learning objectives are defined by constructing positive and negative sample pairs. Positive sample pairs represent similar features from the same process or parameter, while negative sample pairs represent heterogeneous features from different process conditions. By minimizing the feature distance between positive sample pairs and maximizing the feature distance between negative sample pairs, the feature representation ability of the network is optimized to generate contrastive learning features. The optimized contrastive learning features are input into the feature fusion module, and multimodal features are fused into a unified feature vector through nonlinear transformation. On the basis of retaining temporal information and semantic information, the feature fusion module realizes nonlinear mapping of multimodal features through a deep learning network to generate a feature vector with stronger expression ability. In order to ensure the numerical stability and computational efficiency of the feature vector, the fused feature vector is L2 normalized to obtain a unified feature representation vector. The unified feature representation vector is input into the autoencoder reconstruction network, and the process parameter optimization model is finally constructed through the process of feature decoupling and parameter mapping. The autoencoder network uses its decoupling characteristics to decompose the complex feature vector into different latent space representations, and at the same time realizes the conversion of feature vectors to actual process parameters through the parameter mapping module.The entire training process adopts a multi-objective joint optimization strategy to balance the model's reconstruction error, feature consistency, and optimization performance. Through joint optimization training, a process parameter optimization model that can efficiently generate optimized process parameters is obtained.

[0016] In this embodiment, the unified feature representation vector is input into the encoder of the autoencoder reconstruction network for compression mapping. The encoder adopts three layers of fully connected layers, and each fully connected layer is followed by a LayerNorm normalization layer and a PReLU activation function. The introduction of LayerNorm can effectively stabilize the training process and improve the convergence speed of the network, while the PReLU activation function enhances the nonlinear modeling ability of the network by dynamically adjusting the activation coefficient of the negative part. After multi-layer mapping processing, the potential feature encoding is generated. The potential feature encoding is input into the decoupling module of the autoencoder reconstruction network, which consists of two parallel branches to extract process parameter features and quality indicator features respectively. The first branch extracts process parameter features through three layers of independent component analysis (ICA) layers, each with a dimension of 128. The ICA layer can effectively separate linear mixed features and highlight the independence between parameter features. The second branch extracts quality indicator features through two layers of variational inference layers, each with a dimension of 128. The variational inference layer uses a probabilistic inference method to perform distributed modeling on features to capture complex potential correlations. Through the processing of these two branches, the decoupling module decomposes the input potential feature encoding into more explanatory and structured decoupled features, which reflect the intrinsic characteristics of process parameters and quality indicators respectively. The decoupled features are input into the parameter reconstruction decoder, which consists of four layers of transposed fully connected layers, each of which is equipped with LeakyReLU activation function and residual connection. The LeakyReLU activation function can effectively alleviate the gradient vanishing problem, while the residual connection improves the gradient fluidity and representation ability of the network, so that the decoder can better restore the parameter information during the reconstruction process. Through multi-layer nonlinear mapping and inverse operation, the decoder restores the decoupled features to a reconstruction parameter vector, which represents the reconstruction result of the model for the process parameters. On this basis, a multi-task learning module is introduced into the autoencoder reconstruction network to simultaneously complete the parameter optimization task and the quality prediction task. The multi-task learning module designs two branches, in which the parameter optimization branch and the quality prediction branch respectively use three layers of fully connected layers with a dimension of 512 to perform feature transformation on the reconstruction parameter vector to generate the initial optimization result. Through this process, the model can simultaneously focus on the impact of parameter adjustment on the optimization goal and the constraint relationship on quality prediction, ensuring the comprehensiveness and accuracy of the optimization. Since the goals of different tasks may conflict during multi-task learning, a dynamic task weight calculation module is introduced. This module performs a weighted balance on the losses of parameter optimization tasks and quality prediction tasks based on the task losses calculated in real time during training. By dynamically adjusting the task weights, it is ensured that the model can achieve the optimal balance between optimization efficiency and quality constraints, and then define the joint optimization goal. This joint optimization goal is the core criterion for training. The model parameters are continuously iterated through the optimization algorithm based on gradient descent to gradually improve the performance and stability of the model. The optimized model parameters are input into the model integration module of the autoencoder reconstruction network for multi-model fusion.The model integration module improves the generalization ability and robustness of the overall model by combining the prediction results of multiple optimization models. At the same time, in order to meet the rationality requirements of process parameters in practical applications, a soft constraint layer is designed in the network to limit the range of parameter values. This constraint mechanism effectively prevents the model from generating parameter combinations that do not meet the actual process requirements, thereby ensuring that the optimization results are practical. Through the above steps, a process parameter optimization model is obtained. Based on historical data and semantic features, the model efficiently decouples the characteristic relationship between process parameters and quality indicators, and realizes dynamic adjustment and optimization of parameters.

[0017] S3, based on the process parameter optimization model, the parameter compensation amount is calculated for the equipment status data and quality inspection data collected in real time to obtain the optimized process parameter combination; It should be noted that the equipment status data collected in real time is standardized to eliminate the dimension differences of data of different dimensions and obtain a standardized state vector. At the same time, the quality inspection data collected in real time is normalized and mapped to the interval [0,1] to form a standardized quality vector. The standardized state vector and the standardized quality vector are spliced ​​and fused to form a fused vector. The fused vector is input into the encoder module of the process parameter optimization model. The encoder compresses and maps the input data through multi-layer nonlinear mapping to extract the feature representation of the current working condition. Based on the feature representation of the current working condition, the process parameter features are separated from the quality indicator features through the decoupling module of the process parameter optimization model to obtain a decoupled feature pair. The decoupling module uses the capabilities of independent component analysis and variational inference to effectively distinguish the features directly related to the process parameter adjustment and the features associated with the quality constraints, ensuring that the expression of each feature has a clear physical meaning and optimization goal. The decoupled feature pair is input into the parameter reconstruction decoder of the process parameter optimization model. The decoder generates the initial parameter compensation by reconstructing the decoupled features. Based on the initial parameter compensation, the model optimizes the compensation and predicts the quality through the multi-task learning module. The multi-task learning module optimizes the accuracy of the compensation through the parameter optimization branch, and evaluates the changes in quality indicators after compensation through the quality prediction branch. The dual optimization ensures that the compensation suggestions can not only help improve equipment performance but also meet quality requirements. After the parameter adjustment suggestions are generated, the dynamic task weight calculation module prioritizes and assigns weights to the compensation suggestions. The task weight calculation module dynamically adjusts the priority of each compensation amount according to the real-time working conditions and multi-task loss weights to generate a weighted compensation vector. The weighted compensation vector is compared with the value constraints of the process parameters, and the amplitude of the compensation amount is adjusted through the soft constraint layer. The role of the soft constraint layer is to automatically correct the compensation amount that does not meet the process requirements, limit the adjustment range of the compensation amount, and ensure that the final output compensation amount meets the actual operating conditions. The compensation amount after constraint adjustment is linearly superimposed with the current process parameters to generate an updated parameter combination. The updated parameter combination is input into the model integration module, and the model integration module performs multi-model verification and integrated prediction on the updated parameter combination. By integrating the prediction results of multiple optimization models, the robustness and accuracy of parameter adjustment are improved, and an optimized process parameter combination is generated.

[0018] S4, solves the optimal capacity allocation plan based on the optimized process parameter combination and outputs the process parameter dynamic control instructions.

[0019] Specifically, based on the optimized process parameter combination, a binary integer linear programming model is established, and the maximum capacity is set as the objective function. The design of the objective function aims to maximize the production capacity under given resource conditions. At the same time, the value range of the process parameters and the limitation of the equipment resources are used as the constraints of the model to form a complete capacity planning mathematical model. The model represents the allocation status of resources in the form of binary variables, such as whether the equipment is assigned to a specific task or whether a certain process parameter is in a specific range. Since the binary integer programming problem has a high solution complexity, in order to improve the computational efficiency, the capacity planning mathematical model is Lagrangian relaxed. By introducing Lagrangian multipliers, the complex constraints in the model are separated from the objective function, and the original binary integer programming problem is transformed into a continuous optimization problem to form a relaxed optimization problem. Based on the relaxed optimization problem, the KKT (Karush-Kuhn-Tucker) optimality condition is used to solve it. The KKT condition is a necessary optimality condition for the continuous optimization problem. By taking the derivative of the objective function and its constraints and setting them equal to zero, a set of optimality necessary condition equations is obtained. In order to solve this system of equations, the subgradient method is used to iteratively update the Lagrange multiplier. The subgradient method is an optimization method that gradually approaches the optimal solution of the objective function by continuously adjusting the value of the multiplier. After multiple iterative updates, the optimal solution on the continuous solution space is obtained. The optimal solution on the continuous solution space is integerized by the branch and bound method. The branch and bound method is used to discretize the solution, and the branch space of the solution is traversed by the depth-first search strategy to ensure that the generated initial capacity allocation plan meets the requirements of the binary integer programming model while retaining the characteristics of the optimal solution. The feasibility of the initial capacity allocation plan is verified to check whether the plan meets the process constraints, such as the maximum load capacity of the equipment and the continuity of the production cycle. If an infeasible solution is found, a penalty function is constructed based on the process constraints, and the infeasible solution is adjusted by a local search algorithm. While introducing the penalty function, this adjustment process uses the search algorithm to find the optimal feasible solution in the local solution space to generate a feasible capacity plan that meets all constraints. Production scheduling is performed based on the feasible capacity plan. By calculating the processing sequence and equipment allocation of each process, the critical path method is used to determine the shortest completion time of the entire production process, that is, the production cycle. The critical path method analyzes the dependencies between tasks and determines the key processes that need to be focused on, thereby optimizing the execution order of the production plan. On this basis, a dynamic scheduling plan is generated and converted into a process parameter control sequence. The key content in the dynamic scheduling plan is converted into a control instruction set according to the parameter mapping rules. The control instruction set specifically includes specific requirements for equipment operation, the time point and amplitude of parameter adjustment, etc. These instructions are sent to the execution layer to realize dynamic regulation of the actual production process.

[0020] In one example, the historical processing data collected by the intelligent manufacturing chip manufacturing equipment is cleaned and time-series sorted to obtain a process parameter time series data set, and a process semantic feature matrix is ​​generated according to the production process document, including: Perform sampling period alignment processing on the process parameters, equipment operating status and quality indicator data of intelligent manufacturing chip manufacturing equipment to obtain the original data set, and perform outlier detection and elimination on the numerical data in the original data set to obtain a preliminary cleaned data set; Perform interpolation calculation based on time series correlation on the missing values ​​in the preliminary cleaned data set to obtain a complete numerical sequence, sort the complete numerical sequence in ascending order according to the timestamp, and perform sliding window segmentation to obtain a process parameter time series data set; Perform text segmentation and punctuation filtering on the production process documents to obtain the original text sequence, and then perform word segmentation on the original text sequence to extract process terms and parameter descriptions to obtain structured text data; Based on structured text data, a process term-parameter mapping table is established, the process description is normalized to obtain a standardized text set, and the standardized text set is organized and indexed according to the process sequence to obtain a process knowledge corpus; A process labeling system is constructed based on the process knowledge corpus. The parameter description labels and quality indicator labels in the process labeling system are templated and merged. After being processed by the SBERT semantic encoder, the process semantic feature matrix is ​​obtained.

[0021] In this example, the sampling period alignment processing is performed on the process parameters, equipment operation status and quality indicator data of the intelligent manufacturing chip manufacturing equipment, and the data from different sources are synchronized in the time dimension to ensure the time sequence consistency of the data. Set the sampling time interval to , assuming that the timestamp sets of different data streams are , resample all data to a unified time axis using linear interpolation ,in Indicates The interpolation calculation formula is as follows: ; in, It is Parameters in time The value of and is the original time point. By this method, the aligned original data set is generated . Detect and remove outliers from the numerical data in the original data set. Assume that a process parameter The set of observations is , through statistical analysis methods such as the Зσ principle, the range of abnormal values ​​is determined to be ,in is the mean, is the standard deviation. Values ​​outside this range are recorded as outliers and removed. After removal, a preliminary cleaned data set is obtained. For the missing values ​​in the preliminary cleaned data set, interpolation calculations are performed based on the time series correlation to fill the data gaps. Assume that the missing points of the time series are , and its adjacent points are and , using the linear interpolation formula: ; Or use spline interpolation and time series forecasting models (such as ARIMA) to generate a complete numerical sequence. After sorting the sequence in ascending order of timestamp, set the sliding window size , segment the sequence. The sliding window segmentation formula is as follows: ; The generated segment collection That is, the process parameter time series data set. At the same time, the processing of the production process document includes text segmentation and punctuation filtering, and the document is divided into multiple small text segments for subsequent analysis. The original text sequence is segmented, and key information such as "environment", "operation", "parameter adjustment range", "unit", etc. are extracted and combined with numerical descriptions to form structured text data. Based on the structured text data, a process term-parameter mapping table is established. By normalizing the process description in the mapping table, a standardized text set is generated. The standardized text set is organized according to the process flow sequence, and an index structure is constructed based on the process node to obtain a process knowledge corpus. For example, it is organized in sequence as "process node 1->process node 2->process node 3", where each node stores relevant parameter descriptions and semantic information. Based on the process knowledge corpus, a process label system is constructed, and parameter description labels (such as "temperature" and "pressure") are templated and merged with quality indicator labels (such as "defect rate" and "stability"). For example, combine "temperature range" with "defect rate relationship" to form a template "temperature affects defect rate". Use the SBERT (Sentence-BERT) model to semantically encode these template texts and generate semantic feature vectors. SBERT uses a bidirectional Transformer structure to calculate embedding representations based on input text: ; in Represents the embedding function of the SBERT model. All feature vectors are combined into a matrix: ; Get the process semantic feature matrix , each column of which corresponds to the semantic embedding of a template, which serves as the input of the subsequent process optimization model.

[0022] In one example, a process labeling system is constructed based on a process knowledge corpus. The parameter description labels and quality indicator labels in the process labeling system are templated and merged. After being processed by the SBERT semantic encoder, a process semantic feature matrix is ​​obtained, including: The process terms and parameter descriptions in the process knowledge corpus are subjected to topic cluster analysis to obtain process parameter category information; A parameter description label template is constructed based on the process parameter category information, and the name, value range and adjustment rule of the process parameter are structuredly encoded to obtain a parameter description label; Extract information from the quality description in the process knowledge corpus, extract the target specification, allowable deviation and detection method information, and obtain the quality indicator label; The parameter description labels and quality indicator labels are merged to obtain a standardized process text sequence, and the SBERT model is used to encode the standardized process text sequence into word vectors to obtain the initial semantic feature vector. An attention weight matrix is ​​constructed based on the initial semantic feature vector, and the correlation features between parameters are weighted to obtain an enhanced semantic feature vector. The enhanced semantic feature vector is then dimensionally reduced and feature selected to obtain a process semantic feature matrix.

[0023] In this example, process terms and parameter descriptions are extracted from the process knowledge corpus and subject clustering analysis is performed on them. Assume that the set of process terms is ,in is a specific process term, such as "temperature", "pressure" or "humidity". Latent Dirichlet Allocation (LDA) is used to Perform topic modeling and construct process parameter category information. The core of LDA is to learn the topic distribution matrix based on word frequency and document distribution. and word-topic matrix , its mathematical form is: ; in, Yes word In the documentation The probability of Yes word In Theme The probability of It is a document Medium Theme The probability of is the number of topics. The above distribution is optimized by Gibbs sampling method to obtain the topic category of each process term. After obtaining the process parameter category information, a parameter description label template is constructed based on this category information. The construction of the parameter description label template includes structured coding of the process parameter name, value range and adjustment rules. The structured coding is expressed as: ; in, is the parameter name, is the value range, is the regulation rule, such as "linear increase". These encoded parameters are combined into parameter description tags At the same time, the quality description in the process knowledge corpus is extracted to extract the information of target specifications, allowable deviations and detection methods. Assume that the set of quality descriptions is ,in Represents a specific quality requirement. Parse and extract entity tags. Encode the extracted information into a set of quality indicator tags . Label the parameter description and quality indicator labels Merge to get a standardized process text sequence. Use the SBERT (Sentence-BERT) model to encode these texts and generate an initial semantic feature vector. The core of SBERT is to deeply model the text through a bidirectional Transformer network, and its embedding formula is: ; in, is text The semantic vector of is the embedding function of SBERT. Assume that the input standardized text sequence is , the generated feature vector set is In order to capture the correlation features between parameters, based on the initial semantic feature vector Construct the attention weight matrix. Assume and are two feature vectors, and the attention weight is calculated by cosine similarity: ; Attention weight matrix Through weighted operation, the feature vectors are weighted and combined to generate enhanced semantic feature vectors. : ; The enhanced semantic feature vector is subjected to dimensionality reduction and feature selection, and the principal component analysis method is used to reduce the dimension to retain the dominant features. is the enhanced feature matrix, which is converted into a reduced dimension matrix through principal component analysis : ; in, is the projection matrix of principal component analysis. The feature matrix after dimensionality reduction is That is the process semantic feature matrix, which has higher semantic expression ability and computational efficiency.

[0024] In one example, a process parameter optimization model is created based on a process parameter time series data set and a process semantic feature matrix, including: The process parameter time series data set is input into the time series encoding branch of the cross-modal feature encoding network to extract time series features and obtain a time series feature vector; The process semantic feature matrix is ​​input into the semantic encoding branch of the cross-modal feature encoding network for feature transformation to obtain a semantic feature vector; Perform wavelet decomposition on the time series feature vector to extract high-frequency, medium-frequency and low-frequency feature components respectively to obtain enhanced time series features; Based on the semantic feature vector, a K-nearest neighbor graph structure is constructed, the cosine similarity matrix between feature nodes is calculated, and message passing and feature aggregation are performed through a two-layer graph convolutional network to obtain enhanced semantic features; The enhanced temporal features and enhanced semantic features are input into the feature alignment module, the attention weights between the modalities are calculated, and the initial alignment features are obtained; Perform contrastive learning on the initial alignment features, construct positive and negative sample pairs, minimize the feature distance between the positive sample pairs, and maximize the feature distance between the negative sample pairs to obtain contrastive learning features; The contrastive learning features are input into the feature fusion module for nonlinear transformation to obtain a fused feature vector, and the fused feature vector is L2 normalized to obtain a unified feature representation vector; The unified feature representation vector is input into the autoencoder reconstruction network for feature decoupling and parameter mapping, and the process parameter optimization model is obtained through multi-objective joint optimization training.

[0025] In this example, the process parameter time series data set is input into the time series encoding branch of the cross-modal feature encoding network, and the time series features are extracted through time series encoding. Assume that the input time series data set is ,in Indicates The time series encoding branch uses LSTM (Long Short-Term Memory Network) to model the time series data, and its state update formula is: ; in, Indicates The hidden state of time steps, and are the weight matrices for hidden states and inputs, respectively, is the bias term, is the activation function. Through the recursive modeling of LSTM, the time series feature vector is generated , contains the dynamic change information of process parameters in the time dimension. At the same time, the process semantic feature matrix Input the semantic encoding branch of the cross-modal feature encoding network for feature transformation. Assume is a The matrix of is the number of features, is the dimension of each feature. The semantic encoding branch adopts a Transformer-based architecture and performs feature modeling through a multi-head attention mechanism. The attention weight calculation formula is: ; in, are query, key, and value matrices respectively, is the dimension of the key vector. Through the stacking structure of Transformer, a semantic feature vector is generated , these feature vectors capture the complex relationship between process semantic features. Wavelet decomposition is performed to extract high-frequency, medium-frequency and low-frequency characteristic components and enhance their time series characteristics. The mathematical expression of wavelet decomposition is: ; in, is the high frequency wavelet basis, is the low frequency wavelet basis, and The corresponding coefficients are obtained through decomposition, namely, high-frequency components (capturing fast-changing information), medium-frequency components (capturing periodic change information) and low-frequency components (capturing overall trend information). These components are combined to form enhanced time series features. At the same time, based on the semantic feature vector Construct a K-nearest neighbor graph structure by calculating the cosine similarity matrix between feature nodes Describe the association between nodes: ; Using graph convolutional networks For message passing and feature aggregation, the update rule of graph convolution is: ; in, is the updated node feature matrix, is the weight matrix, is the activation function. After two layers of graph convolutional networks, enhanced semantic features are generated. . Will enhance the timing characteristics and enhanced semantic features The input feature alignment module uses the cross-modal attention mechanism to calculate the attention weights between modalities. The formula for attention weights is: ; The different modal features are weighted and combined through attention weights to generate initial alignment features . For the initial alignment feature Perform contrastive learning and construct positive and negative sample pairs. Assume that the positive sample pair is , the negative sample pair is , define the contrast loss as: ; in, represents the cosine similarity, is the temperature parameter. By minimizing the distance between positive samples and maximizing the distance between negative samples, a contrastive learning feature is generated. . Compare the learning features Input feature fusion module to generate fused feature vector through nonlinear transformation The formula for the fusion process is: ; in, is the weight matrix, is the bias term. Perform L2 normalization: ; Generate a unified feature representation vector .Will Input the autoencoder reconstruction network to complete feature decoupling and parameter mapping, and generate the final process parameter optimization model through multi-objective joint optimization training. The objective function of joint optimization is: ; in, To rebuild the losses, is the regression loss, is the weight factor. The optimized model can effectively generate process parameters for performance optimization.

[0026] In one example, the unified feature representation vector is input into the autoencoder reconstruction network for feature decoupling and parameter mapping, and a process parameter optimization model is obtained through multi-objective joint optimization training, including: The unified feature representation vector is input into the encoder of the autoencoder reconstruction network for compression mapping. The encoder contains three fully connected layers, each of which is followed by a LayerNorm normalization layer and a PReLU activation function to obtain the potential feature encoding. The latent feature encoding is input into the decoupling module of the autoencoder reconstruction network. The decoupling module contains two parallel branches. The first branch extracts the process parameter features through three layers of independent component analysis with a dimension of 128, and the second branch extracts the quality indicator features through two layers of variational inference with a dimension of 128 to obtain the decoupled features. The decoupled features are input into the autoencoder to reconstruct the parameters of the network. The decoder contains four transposed fully connected layers. Each transposed fully connected layer uses the LeakyReLU activation function and residual connection to obtain the reconstruction parameter vector. A multi-task learning module of the autoencoder reconstruction network is constructed based on the reconstruction parameter vector. The multi-task learning module includes a parameter optimization branch and a quality prediction branch. Three fully connected layers with a dimension of 512 are used for feature conversion to obtain the initial optimization result. The initial optimization results are input into the dynamic task weight calculation module of the autoencoder reconstruction network to perform a weighted balance on the losses of the parameter optimization task and the quality prediction task to obtain the joint optimization target. Gradient descent optimization is performed based on the joint optimization objective to obtain the optimized model parameters, and the optimized model parameters are input into the model integration module of the autoencoder reconstruction network for multi-model fusion. At the same time, the parameter value range is restricted through the soft constraint layer to obtain the process parameter optimization model.

[0027] In this example, the unified feature representation vector The encoder part of the input autoencoder reconstruction network is compressed and mapped. The encoder consists of three fully connected layers, each of which is followed by a LayerNorm normalization layer and a PReLU activation function. Layer, the mapping formula is: ; in Indicates The output features of the layer, is the weight matrix, is the bias vector, LayerNorm normalizes the output of each layer to stabilize the training process, and PReLU improves the nonlinear expression ability of the model through a learnable activation function. The encoder generates the potential feature encoding , the encoding is the input vector The latent features are encoded The input is sent to the decoupling module of the autoencoder reconstruction network, which contains two parallel branches. The first branch uses three layers of independent component analysis (ICA) to extract process parameter features. The ICA layer finds the independent representation of the process parameters by decomposing the linear mixed features. Its mathematical formula is: ; in is the independent component matrix, is the separation matrix of ICA. The second branch extracts quality indicator features through two layers of variational inference. The goal of variational inference is to use parameterized distribution To approximate the true posterior distribution . Its optimization goal is to minimize the evidence lower bound (ELBO): ; in, is the variational feature, and KL is the Kullback-Leibler divergence. Through these two branches, the decoupling module generates process parameter features and quality indicator features, collectively referred to as decoupled features. The decoupled features are input into the parameter reconstruction decoder of the autoencoder reconstruction network. The decoder contains four layers of transposed fully connected layers, each of which uses the LeakyReLU activation function and residual connection. Assume that The input of the layer is , the output is: ; in and are weight matrices and bias vectors respectively. The decoder generates the reconstruction parameter vector , whose size is consistent with the process parameter space. Based on the reconstruction parameter vector , a multi-task learning module is introduced into the autoencoder reconstruction network. This module contains a parameter optimization branch and a quality prediction branch, each of which uses three fully connected layers to transform features. Assume that the reconstruction parameters are , the output of the parameter optimization branch is: ; The output of the quality prediction branch is: ; in and is the corresponding conversion function, and the initial optimization result is finally generated. The initial optimization result is input into the dynamic task weight calculation module, which performs a weighted balance on the losses of the parameter optimization task and the quality prediction task. Assume that the losses of the two tasks are and , the total loss is defined as: ; in and It is a dynamic weight that is automatically adjusted during the training process. Based on the joint optimization objective, gradient descent optimization is performed to update the model parameters : ; in is the learning rate. The optimized model parameters are input into the model integration module, and multi-model fusion is performed by combining the outputs of multiple sub-models. At the same time, in order to ensure that the generated parameters meet the actual process requirements, the parameter value range is restricted through the soft constraint layer. For example, if a parameter The constraint is , then adjust it as follows: ; The generated process parameter optimization model can achieve the best performance while meeting quality and process constraints.

[0028] In one example, the parameter compensation amount is calculated based on the real-time collected equipment status data and quality inspection data based on the process parameter optimization model to obtain an optimized process parameter combination, including: The device status data collected in real time is normalized to obtain a standardized state vector, and the quality detection data collected in real time is normalized to obtain a standardized quality vector; The standardized state vector and the standardized quality vector are concatenated and fused to obtain a fused vector, and the fused vector is input into the encoder of the process parameter optimization model to obtain a feature representation of the current working condition; Based on the feature representation of the current working conditions, the process parameter features and quality index features are separated through the decoupling module of the process parameter optimization model to obtain the decoupled feature pair; The decoupled features are input into the parameter reconstruction decoder of the process parameter optimization model to generate the initial parameter compensation. The multi-task learning module is used to perform parameter optimization and quality prediction to obtain parameter adjustment suggestions. Based on the parameter adjustment suggestions, the dynamic task weight calculation module is used to prioritize and weight parameter compensation to obtain a weighted compensation vector. The weighted compensation vector is compared with the process parameter value constraint, and the amplitude of the compensation amount is adjusted through the soft constraint layer to obtain the compensation amount that meets the constraint; The compensation amount that meets the constraints and the current process parameters are linearly superimposed to obtain an updated parameter combination, and the updated parameter combination is verified by multiple models and integrated prediction through the model integration module to obtain the optimized process parameter combination.

[0029] In this example, the status data collected from the device in real time is recorded as ,in It is The equipment status parameters are such as temperature, pressure or speed. In order to eliminate the differences in the scales of different parameters, the status data is standardized. The standardization formula is: ; in is a parameter The mean of is its standard deviation. After standardization, we get the standardized state vector , the distribution of each component has zero mean and unit variance, which is convenient for subsequent modeling. Normalize it and map it to the [0, 1] interval to unify the scale. The normalization formula is: ; in and They are The minimum and maximum values ​​of . After normalization, the standardized mass vector is obtained . The normalized state vector and the normalized mass vector Perform splicing and fusion to obtain the fused vector . Assume that the state vector dimension is , the quality vector dimension is , then the fusion vector dimension is .Will Input into the encoder of the process parameter optimization model. The encoder consists of a multi-layer fully connected network, and its feature mapping formula is: ; in It is The output of the layer, is the weight matrix, is the bias vector, is the activation function. After multiple layers of mapping, the encoder outputs the feature representation of the current working condition. Based on the current working condition characteristics , the decoupling module of the process parameter optimization model is used to separate the process parameter characteristics and quality index characteristics to form a decoupled feature pair Among them, the process parameter characteristics Extracted by the independent component analysis (ICA) layer, the formula is: ; is the separation matrix. Quality indicator characteristics Extracted through a variational inference layer, its goal is to minimize the following evidence lower bound: ; Where KL is the Kullback-Leibler divergence. The decoupled feature pair is input into the parameter reconstruction decoder to generate the initial parameter compensation The decoder uses a multi-layer transposed fully connected network, and its mapping formula is: ; in and are weight matrices and biases respectively. Through the multi-task learning module The multi-task learning module contains a parameter optimization branch and a quality prediction branch, which minimize the following losses respectively: ; in and is the target value. Based on the optimized parameter compensation, the dynamic task weight calculation module is used to prioritize and assign weights to the parameter compensation. The dynamic weight is calculated using the following formula: ; in It is The weight of the task, is its loss value. Constrain the weighted compensation vector and the process parameter value and Compare and adjust the amplitude of the compensation through the soft constraint layer. The adjustment formula is: ; The compensation amount that will satisfy the constraint With the current process parameters Perform linear superposition: ; in is the adjustment coefficient. The updated parameter combination is adjusted by the model integration module Perform multi-model verification and integrated prediction to ultimately generate an optimized process parameter combination , to achieve optimal performance.

[0030] In one example, the optimal capacity allocation solution is solved based on the optimized process parameter combination, and the process parameter dynamic control instructions are output, including: Based on the optimized process parameter combination, a binary integer linear programming model is constructed, and the capacity maximization is set as the objective function. The process parameter value range and equipment resource limit are used as constraints to obtain the capacity planning mathematical model. The capacity planning mathematical model is Lagrangian relaxed, Lagrangian multipliers are introduced, and the binary integer programming problem is transformed into a continuous optimization problem to obtain a relaxed optimization problem. Based on the relaxed optimization problem, the KKT optimality condition is constructed, and the optimality necessary condition equations are obtained by taking the derivative of the objective function and setting it equal to zero. Solve the optimality necessary condition equations, use the subgradient method to iteratively update the Lagrange multiplier, and obtain the optimal solution in the continuous solution space; The optimal solution in the continuous solution space is integerized by the branch-and-bound method, and the depth-first search strategy is used to perform branch traversal to obtain the initial capacity allocation plan; Verify the feasibility of the initial capacity allocation plan, build a penalty function based on process constraints, adjust the infeasible solution through the local search algorithm, and obtain a feasible capacity plan; The processing timing and equipment allocation of each process are calculated based on the feasible capacity plan, and the critical path method is used to determine the production rhythm to obtain a dynamic scheduling plan. The dynamic scheduling plan is converted into a process parameter control sequence, and a control instruction set is generated through parameter mapping rules to output dynamic control instructions for process parameters.

[0031] In this example, a binary integer linear programming model is constructed based on the optimized process parameter combination. Assuming that the range of process parameters and equipment resource constraints have been determined, define the variables Indicates whether the Device assigned to process, including The goal is to maximize the total production capacity, denoted by The objective function is expressed as: ; in, Yes Equipment In process The unit capacity contribution on and are the total number of equipment and processes respectively. The constraints include process parameter range constraints and equipment resource limitations. The process parameter range constraints are: ; in, Yes Equipment Process The contribution of process parameters, and The processes are The minimum and maximum values ​​of the parameters are as follows. The device resource limits are: ; Indicates that each device is assigned to at most one process. The above binary integer programming problem is Lagrangian relaxed. By introducing the Lagrangian multiplier and , adding parameter range constraints and device resource restrictions to the objective function to form a relaxed optimization problem: By taking the derivative of the objective function and setting it equal to zero, the KKT (Karush-Kuhn-Tucker) optimality condition is constructed to describe the necessary conditions for the optimal solution. Taking the derivative, we get: ; This set of equations describes the optimality under relaxed conditions and is further solved in combination with constraints. In order to solve the optimality condition equations, the subgradient method is used to iteratively update the Lagrange multiplier. The update formula is: ; ; in is the learning rate. Through multiple iterations, the optimal solution in the continuous solution space is obtained The continuous solution is integerized by the branch-and-bound method, and the branch space is traversed using a depth-first search strategy. The core of the branch-and-bound method is to divide the problem into multiple sub-problems and reduce the search space by pruning the upper and lower bounds. The final solution is is the initial capacity allocation plan. The feasibility of the initial capacity allocation plan is verified. If there is an allocation plan that does not meet the process constraints, a penalty function is constructed based on the process constraints, for example: The infeasible solution is adjusted through the local search algorithm, and the allocation scheme is gradually optimized to meet all constraints and generate a feasible capacity plan. Based on the feasible capacity plan, the processing sequence and equipment allocation of each process are calculated. The critical path method is used to determine the production cycle. Assume that the processing time of each process is , the completion time of the path in the processing network is: ; The critical path method identifies the path in the production process that determines the total completion time, thereby optimizing equipment utilization. The final dynamic scheduling plan is converted into a process parameter control sequence, and a control instruction set is generated through parameter mapping rules. For example, if a certain equipment In process Parameters that need to be adjusted , the generated control instruction is: "Device In time Start, set the parameters to Through this process, dynamic process parameter control instructions are output to ensure efficient operation of equipment and processes.

[0032] Reference Figure 2 This embodiment provides a performance optimization device for an intelligent manufacturing chip based on historical data, including: Preprocessing module 1 is used to clean and sort the historical processing data collected by the intelligent manufacturing chip manufacturing equipment, obtain the process parameter time series data set, and generate the process semantic feature matrix according to the production process document; Creating module 2, for creating a process parameter optimization model based on a process parameter time series data set and a process semantic feature matrix; Calculation module 3, used to calculate parameter compensation amount for the equipment status data and quality inspection data collected in real time based on the process parameter optimization model, and obtain an optimized process parameter combination; The output module 4 is used to solve the optimal capacity allocation plan based on the optimized process parameter combination and output the process parameter dynamic control instruction.

[0033] In this embodiment, for the specific implementation of each unit in the above device embodiment, please refer to the above method embodiment, which will not be repeated here.

[0034] Reference Figure 3 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3 As shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0035] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0036] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0037] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided by the present invention and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM.

[0038] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0039] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A performance optimization method for intelligent manufacturing chips based on historical data, characterized in that: The following steps are involved: Perform data cleaning and time series sorting on the historical processing data collected by smart manufacturing chip manufacturing equipment to obtain a process parameter time series data set, and generate a process semantic feature matrix based on the production process document; Creating a process parameter optimization model based on the process parameter time series data set and the process semantic feature matrix; Based on the process parameter optimization model, parameter compensation is calculated for the real-time collected equipment status data and quality inspection data to obtain an optimized process parameter combination; The optimal capacity allocation plan is solved based on the optimized process parameter combination, and a process parameter dynamic control instruction is output.

2. The performance optimization method of intelligent manufacturing chip based on historical data according to claim 1 is characterized in that: The historical processing data collected by the intelligent manufacturing chip manufacturing equipment is cleaned and sorted in time series to obtain a process parameter time series data set, and a process semantic feature matrix is ​​generated according to the production process document, including: Perform sampling period alignment processing on the process parameters, equipment operating status and quality indicator data of the intelligent manufacturing chip manufacturing equipment to obtain the original data set, and perform outlier detection and elimination on the numerical data in the original data set to obtain a preliminary cleaned data set; Performing interpolation calculation based on time series correlation on the missing values ​​in the preliminary cleaned data set to obtain a complete numerical sequence, sorting the complete numerical sequence in ascending order according to timestamps, and performing sliding window segmentation to obtain a process parameter time series data set; Performing text segmentation and punctuation filtering on the production process document to obtain an original text sequence, and performing word segmentation processing on the original text sequence to extract process terms and parameter descriptions to obtain structured text data; A process term-parameter mapping table is established based on the structured text data, the process description is normalized to obtain a standardized text set, and the standardized text set is organized and indexed according to the process flow sequence to obtain a process knowledge corpus; A process labeling system is constructed based on the process knowledge corpus, and the parameter description labels and quality indicator labels in the process labeling system are templated and merged, and processed by the SBERT semantic encoder to obtain a process semantic feature matrix.

3. The performance optimization method of intelligent manufacturing chip based on historical data according to claim 2 is characterized in that: The process label system is constructed based on the process knowledge corpus, and the parameter description labels and quality indicator labels in the process label system are templated and merged, and processed by the SBERT semantic encoder to obtain a process semantic feature matrix, including: Performing a topic cluster analysis on the process terms and parameter descriptions in the process knowledge corpus to obtain process parameter category information; Constructing a parameter description label template based on the process parameter category information, performing structured coding on the name, value range and adjustment rule of the process parameter to obtain a parameter description label; Extracting information from the quality description in the process knowledge corpus, extracting target specifications, allowable deviations and detection method information, and obtaining quality indicator labels; The parameter description label and the quality indicator label are merged to obtain a standardized process text sequence, and the SBERT model is used to perform word vector encoding on the standardized process text sequence to obtain an initial semantic feature vector; An attention weight matrix is ​​constructed based on the initial semantic feature vector, and the correlation features between parameters are weighted to obtain an enhanced semantic feature vector. The enhanced semantic feature vector is then dimensionally reduced and feature selected to obtain a process semantic feature matrix.

4. The performance optimization method of intelligent manufacturing chip based on historical data according to claim 3 is characterized in that: The creating a process parameter optimization model based on the process parameter time series data set and the process semantic feature matrix includes: Inputting the process parameter time series data set into the time series encoding branch of the cross-modal feature encoding network to extract time series features to obtain a time series feature vector; Inputting the process semantic feature matrix into the semantic encoding branch of the cross-modal feature encoding network for feature transformation to obtain a semantic feature vector; Performing wavelet decomposition on the time series feature vector to extract high-frequency, medium-frequency and low-frequency feature components respectively to obtain enhanced time series features; Based on the semantic feature vector, a K-nearest neighbor graph structure is constructed, a cosine similarity matrix between feature nodes is calculated, and message transmission and feature aggregation are performed through a two-layer graph convolutional network to obtain enhanced semantic features; Inputting the enhanced temporal features and the enhanced semantic features into a feature alignment module, calculating the attention weights between the modalities, and obtaining initial alignment features; Performing contrastive learning on the initial alignment features, constructing positive and negative sample pairs, minimizing the feature distance between the positive sample pairs, and maximizing the feature distance between the negative sample pairs, to obtain contrastive learning features; Inputting the contrastive learning features into a feature fusion module for nonlinear transformation to obtain a fused feature vector, and performing L2 normalization on the fused feature vector to obtain a unified feature representation vector; The unified feature representation vector is input into an autoencoder reconstruction network for feature decoupling and parameter mapping, and a process parameter optimization model is obtained through multi-objective joint optimization training.

5. The performance optimization method of intelligent manufacturing chip based on historical data according to claim 4 is characterized in that: The unified feature representation vector is input into the autoencoder reconstruction network for feature decoupling and parameter mapping, and a process parameter optimization model is obtained through multi-objective joint optimization training, including: Inputting the unified feature representation vector into an encoder of an autoencoder reconstruction network for compression mapping, the encoder comprises three fully connected layers, each fully connected layer is followed by a LayerNorm normalization layer and a PReLU activation function, to obtain a potential feature code; The potential feature encoding is input into the decoupling module of the autoencoder reconstruction network, wherein the decoupling module comprises two parallel branches, wherein the first branch extracts process parameter features through three independent component analysis layers with a dimension of 128, and the second branch extracts quality indicator features through two variational inference layers with a dimension of 128 to obtain decoupled features; The decoupled features are input into a parameter reconstruction decoder of the autoencoder reconstruction network, wherein the decoder comprises four transposed fully connected layers, each transposed fully connected layer uses a LeakyReLU activation function and a residual connection to obtain a reconstruction parameter vector; A multi-task learning module of the autoencoder reconstruction network is constructed based on the reconstruction parameter vector, wherein the multi-task learning module includes a parameter optimization branch and a quality prediction branch, and three fully connected layers with a dimension of 512 are respectively used for feature conversion to obtain an initial optimization result; Inputting the initial optimization result into the dynamic task weight calculation module of the autoencoder reconstruction network, performing weighted balance on the losses of the parameter optimization task and the quality prediction task, and obtaining a joint optimization target; Gradient descent optimization is performed based on the joint optimization objective to obtain optimized model parameters, and the optimized model parameters are input into the model integration module of the autoencoder reconstruction network for multi-model fusion. At the same time, the parameter value range is restricted through the soft constraint layer to obtain a process parameter optimization model.

6. The performance optimization method of intelligent manufacturing chip based on historical data according to claim 5 is characterized in that: The process parameter optimization model is used to calculate the parameter compensation amount of the equipment status data and quality detection data collected in real time to obtain the optimized process parameter combination, including: The device status data collected in real time is normalized to obtain a standardized state vector, and the quality detection data collected in real time is normalized to obtain a standardized quality vector; Concatenating and fusing the standardized state vector and the standardized quality vector to obtain a fused vector, and inputting the fused vector into an encoder of the process parameter optimization model to obtain a feature representation of the current working condition; Based on the characteristic representation of the current working condition, the process parameter characteristics and the quality index characteristics are separated by a decoupling module of the process parameter optimization model to obtain a decoupled characteristic pair; Inputting the decoupled feature pairs into the parameter reconstruction decoder of the process parameter optimization model to generate initial parameter compensation, and performing parameter optimization and quality prediction through the multi-task learning module to obtain parameter adjustment suggestions; Based on the parameter adjustment suggestion, the dynamic task weight calculation module is used to prioritize and weight parameter compensation to obtain a weighted compensation vector; Comparing the weighted compensation vector with the process parameter value constraint, adjusting the amplitude of the compensation amount through the soft constraint layer, and obtaining the compensation amount that satisfies the constraint; A linear superposition operation is performed on the compensation amount that satisfies the constraints and the current process parameters to obtain an updated parameter combination, and the updated parameter combination is subjected to multi-model verification and integrated prediction through the model integration module to obtain an optimized process parameter combination.

7. The performance optimization method of intelligent manufacturing chip based on historical data according to claim 6 is characterized in that: The step of solving the optimal capacity allocation scheme based on the optimized process parameter combination and outputting a process parameter dynamic control instruction includes: Based on the optimized process parameter combination, a binary integer linear programming model is constructed, production capacity maximization is set as the objective function, and the process parameter value range and equipment resource limit are used as constraints to obtain a capacity planning mathematical model; Performing Lagrangian relaxation processing on the capacity planning mathematical model, introducing Lagrangian multipliers, converting the binary integer programming problem into a continuous optimization problem, and obtaining a relaxed optimization problem; Constructing the KKT optimality condition based on the relaxed optimization problem, and obtaining the optimality necessary condition equation set by taking the derivative of the objective function and setting it equal to zero; Solving the optimality necessary condition equation group, iteratively updating the Lagrange multiplier using the subgradient method, and obtaining the optimal solution in the continuous solution space; The optimal solution on the continuous solution space is integerized by using the branch and bound method, and the branch traversal is performed using a depth-first search strategy to obtain an initial capacity allocation plan; The feasibility of the initial capacity allocation plan is verified, a penalty function is constructed based on process constraints, and an infeasible solution is adjusted through a local search algorithm to obtain a feasible capacity plan; The processing timing and equipment allocation of each process are calculated based on the feasible capacity plan, the production rhythm is determined by the critical path method, a dynamic scheduling plan is obtained, and the dynamic scheduling plan is converted into a process parameter control sequence, a control instruction set is generated through parameter mapping rules, and a process parameter dynamic control instruction is output.

8. A performance optimization device for intelligent manufacturing chips based on historical data, characterized in that: For implementing the steps of the method according to any one of claims 1 to 7, the device comprises: The preprocessing module is used to clean and sort the historical processing data collected by the intelligent manufacturing chip manufacturing equipment, obtain the process parameter time series data set, and generate the process semantic feature matrix according to the production process document; A creation module, used for creating a process parameter optimization model based on the process parameter time series data set and the process semantic feature matrix; A calculation module, used to calculate parameter compensation amounts for the equipment status data and quality inspection data collected in real time based on the process parameter optimization model, to obtain an optimized process parameter combination; The output module is used to solve the optimal capacity allocation plan based on the optimized process parameter combination and output the process parameter dynamic control instruction.

9. A computer device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images