A line edge instant artificial intelligence system

Abstract

The present application relates to a real-time artificial intelligence system for production line edge, which comprises a data collection system, a statistical analysis system, an artificial intelligence modeling system and an automatic tuning system. The data collection system is used to preprocess and clean the data. The statistical analysis system analyzes the preprocessed and cleaned data to generate a dataset. The artificial intelligence modeling system is used to run and evaluate the dataset with multiple models to complete the creation and deployment of an AI model. Finally, when the AI model is deployed on a production line, the automatic tuning system is used to automatically tune the AI model according to the latest data of the production line.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and mainly to a real-time artificial intelligence system at the edge of a production line. Background Technology

[0002] With the continuous development of industrial technology and the deepening of intelligent manufacturing, the global manufacturing industry is accelerating its transformation towards digitalization and intelligence. In a highly dynamic production environment, real-time acquisition and intelligent processing of production line data have become crucial for improving efficiency and ensuring quality. However, traditional production systems mainly rely on static data models and fixed parameter settings, which often struggle to respond in real time and optimize production processes when faced with complex and ever-changing production scenarios. The rise of the Industrial Internet of Things (IIoT) and edge computing technologies has made it possible to perform data preprocessing and intelligent analysis directly on the production floor, providing a new technological path for building adaptive intelligent production lines.

[0003] In precision manufacturing fields such as semiconductor manufacturing, production equipment is complex and generates massive amounts of data, making it difficult for traditional data acquisition and analysis systems to efficiently integrate multi-source heterogeneous data. For example, US Patent 11402826B2 proposes an anomaly detection system based on a self-organizing data collector. While it can enhance data collection capabilities through neural network analysis, it focuses only on the single function of anomaly detection and lacks the ability to optimize the entire production data process in a closed loop. Similarly, Chinese Patent CN112579653B reduces the data space for AI analysis through contextualized analysis, but its reliance on predefined industry knowledge makes it difficult to adapt to complex and ever-changing production scenarios. Chinese Patent CN104254810B attempts to generate virtual sensor outputs by combining physical and empirical models, but the model update mechanism is lagging and cannot meet real-time requirements. While the instantaneous adaptive control method proposed in US Patent 20200166909A1 can train machine learning algorithms based on data from various objects, it does not form a complete solution from data collection to model deployment. In addition, although the defect detection system of Chinese patent CN110659662B improves the detection accuracy through multi-level verification, the lack of linkage between modules makes it difficult to achieve system-level optimization.

[0004] In conclusion, developing a real-time AI system at the production line edge is of critical strategic value for overcoming the technological bottlenecks in the intelligent upgrading of traditional production lines. This system not only enables closed-loop management of the entire process, from data collection to adaptive model optimization, significantly improving production efficiency and product quality, but also significantly reduces the cost and technical barriers for enterprises deploying intelligent systems, providing efficient and adaptive intelligent solutions for high-end manufacturing industries such as semiconductors. The promotion of this system will accelerate the implementation of intelligent manufacturing technologies in discrete manufacturing scenarios, propelling industrial AI from the experimental stage to large-scale commercial applications, and becoming a core driving force for the digital transformation of global manufacturing. Summary of the Invention

[0005] In view of the technical problems existing in the prior art, this application proposes a real-time artificial intelligence system at the production line edge.

[0006] According to one aspect of this application, a real-time artificial intelligence system for the production line edge is proposed, comprising a data collection system, a statistical analysis system, an artificial intelligence modeling system, and an automatic calibration system. The data collection system is used for data preprocessing and cleaning; the statistical analysis system analyzes the preprocessed and cleaned data to generate a usable dataset; the artificial intelligence modeling system runs and evaluates multiple models on the dataset to complete the creation and deployment of the AI ​​model; finally, after the AI ​​model is deployed on the production line, the automatic calibration system automatically calibrates the AI ​​model based on the latest data from the production line. The data collection system includes a data preprocessing module, a data virtualization module, and a data conversion module. The data preprocessing module is used to summarize the data; the data virtualization module is used to integrate and temporarily store multiple heterogeneous data sets for management; and the data conversion module is used to convert the raw data into virtual tables.

[0007] Preferably, the data collection system can collect and filter data for a specific process. The data collection system supports the semiconductor device communication standard SECS / GEM and can be used to collect data in file format, structured and unstructured data, streaming data and other data formats from semiconductor devices.

[0008] Preferably, the statistical analysis system includes a feature extraction module and a wizard function module. The feature extraction module extracts features from data from different equipment or process sequences through feature engineering, and the wizard function module uses wizard functions to assist users in building specific models of the equipment.

[0009] Preferably, the feature extraction module extracts features from data from different equipment or process sequences through feature engineering. Then, it obtains feature values ​​from the data through multivariate analysis. After integrating knowledge management and statistical analysis content of semiconductor process-specific domain knowledge, it uses Principal Component Analysis (PCA) to achieve variable dimensionality reduction and ranking of relevant variables. After identifying several important variables, it integrates semiconductor process-specific domain knowledge to determine the variables with substantial influence. After identifying the top 50 important variables from 1000 variables using the feature extraction module, it integrates semiconductor process-specific domain knowledge to determine that the variable ranked 5th is not critical, while the variable ranked 2nd or 3rd has a significant influence.

[0010] Preferably, the feature extraction module uses principal component analysis to reduce the dimensionality of variables and rank relevant variables, in order to identify several important variables and integrate semiconductor process-specific domain knowledge. After identifying several important variables, the module integrates semiconductor process-specific domain knowledge to determine the variables with substantial influence. After identifying the top 50 important variables from 1000 variables using the feature extraction module, the module integrates semiconductor process-specific domain knowledge to determine that the variable ranked 5th is not critical, while the variable ranked 2nd or 3rd has a significant influence.

[0011] Preferably, the specific model building of the device includes the following steps: Select Equipment Type: Users can select the type of equipment, reaction chamber, and process to be modeled. Key parameters are retrieved: users are guided to retrieve key equipment parameters based on the selected process or on knowledge of semiconductor process-specific areas. Key parameter weight setting 1: Set the weight of the influence of manufacturing temperature and manufacturing time on the model within the semiconductor process-specific domain knowledge of a specific device or application; Key parameter weight setting two: Set the weight of the influence of manufacturing temperature and manufacturing time on the model within the semiconductor process-specific domain knowledge of another specific device or application; Parameter weight verification and automatic correction: The wizard function automatically verifies whether the parameter weights entered by the user are reasonable and provides automatic adjustment suggestions; Generate a preliminary model: Based on all parameters input by the user, the wizard function generates a preliminary model and displays the simulation results, predicted operating efficiency, energy consumption, and accuracy score. Optimization suggestions: Provide specific optimization suggestions based on the generated preliminary model; and Confirm and complete the model: After the user confirms all parameters and suggestions, the wizard function will complete the final generation of the model.

[0012] Preferably, the artificial intelligence modeling system includes a modeling efficiency optimization module, which uses automatic machine learning technology to train at least two different algorithm types of models in parallel on the input dataset, and selects the best model based on the accuracy data generated by the models.

[0013] More preferably, when the comparison is for data classification, the model performance is evaluated based on the accuracy score.

[0014] More preferably, when the comparison is for data regression, the mean absolute percentage error is used to evaluate the accuracy of the model.

[0015] Furthermore, the modeling efficiency optimization module can provide users with the ability to verify the effectiveness of the model through various statistical analysis charts.

[0016] Preferably, the automatic calibration system includes a model drift detection module, which is responsible for monitoring the performance changes of the AI ​​model and providing a user interface to set the conditions for data drift.

[0017] Preferably, the artificial intelligence modeling system can establish machine tool models through machine learning technology and calculate the prediction of machine tool conditions under specific production parameter settings.

[0018] Preferably, the artificial intelligence modeling system can integrate data including edge artificial intelligence to realize a data exchange mode, and use gateways to protect the data on the data end and the data provider's data, and manage and authorize data through software as a service.

[0019] Preferably, the artificial intelligence modeling system can integrate AI plug-in technology, providing users with a template application function. Users can export their own AI models as preset templates to integrate different AI models.

[0020] More preferably, for specific semiconductor devices, users can export their own AI models as custom templates.

[0021] Preferably, the artificial intelligence modeling system includes an API that can accept query parameters and execute SQL queries to obtain relevant statistical analysis data, so as to effectively integrate generative artificial intelligence into the artificial intelligence modeling system and apply it to the creation of AI models.

[0022] Compared with the prior art, this application has the following beneficial effects: The real-time artificial intelligence system at the production line edge of this application has AI calculation and recommendation functions, which provides the ability to use machine learning to infer the optimal production parameter settings after the new machine is installed. It can quickly complete the equipment inspection and put it into the production line earlier, thereby increasing the production line capacity. The real-time artificial intelligence system at the production line edge of this application identifies potential anomalies in advance through the feature capture module, and combines the model drift detection module to set time and numerical thresholds to monitor data drift in real time, thereby achieving early warning and reducing downtime and maintenance costs. The real-time artificial intelligence system at the production line edge of this application can utilize real-time data and key features of the production line to continuously assess the health trend of parts, predict replacement timing in advance, ensure accurate and reliable prediction results, and guarantee the continuous and stable operation of the production line. The feature extraction module of the real-time AI system at the production line edge in this application deeply integrates statistical methods with domain knowledge, accurately selects the most influential variables, lays a highly reliable foundation for subsequent AI models, and ensures that the models are highly matched with the needs of the production line. The automatic calibration system of the real-time artificial intelligence system at the production line edge of this application can automatically trigger model retraining based on equipment maintenance, process changes or environmental fluctuations, and continuously maintain the high accuracy and adaptability of the AI ​​model without human intervention. The real-time artificial intelligence system at the production line edge of this application is not only applicable to the front-end semiconductor process (lithography, etching, thin film, deposition, polishing), but also to the back-end process (assembly, testing, probe testing) and plant systems, and can be extended to related fields such as flat panel displays, materials science, and optoelectronic technology. Attached Figure Description

[0023] The accompanying drawings are included to provide a further understanding of the embodiments and are incorporated in and constitute a part of this specification. The drawings illustrate embodiments and, together with the description, serve to explain the principles of this application. Other embodiments and many anticipated advantages of these embodiments will be readily recognized as they become better understood through reference to the following detailed description. Elements in the drawings are not necessarily to scale. The same reference numerals refer to corresponding similar parts.

[0024] Figure 1 A system architecture diagram of a real-time artificial intelligence system at the production line edge according to an embodiment of this application is shown; Figure 2 A flowchart illustrating the operation of a real-time artificial intelligence system at the production line edge according to an embodiment of this application is shown. Figure 3 The diagram shows the processing effect of a data preprocessing module according to a specific embodiment of this application; Figure 4 The diagram shows the working effect of a data conversion module according to a specific embodiment of this application; Figure 5 A flowchart illustrating a wizard function module of a real-time artificial intelligence system at the production line edge according to an embodiment of this application is shown, guiding a user to set equipment parameters. Figure 6a A partition diagram of a real-time artificial intelligence system at the production line edge according to an embodiment of this application is shown. Figure 6b A Cook distance map of a real-time artificial intelligence system at the production line edge according to an embodiment of this application is shown; Figure 6c A residual graph of a real-time artificial intelligence system at the production line edge according to an embodiment of this application is shown; Figure 6dA prediction confidence interval diagram of a real-time artificial intelligence system at the production line edge according to an embodiment of this application is shown.

[0025] The attached figures are labeled as follows: 100 - Real-time AI system at the production line edge; 200 - Data collection system; 210 - Data preprocessing module; 220 - Data virtualization module; 230 - Data conversion module; 300 - Statistical analysis system; 310 - Feature extraction module; 320 - Wizard function module; 400 - AI modeling system; 410 - Modeling efficiency optimization module; 500 - Automatic calibration system; 510 - Model drift detection module. Detailed Implementation

[0026] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0027] Where there is no conflict, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0028] Figure 1 and Figure 2 The system architecture diagram and operation flowchart of a real-time artificial intelligence system at the production line edge are shown respectively. (Refer to...) Figure 1 and Figure 2 A real-time artificial intelligence system 100 at the production line edge includes a data collection system 200, a statistical analysis system 300, an artificial intelligence modeling system 400, and an automatic calibration system 500. The data collection system 200 is used for data preprocessing and cleaning; the statistical analysis system 300 analyzes the preprocessed and cleaned data to generate a usable dataset; the artificial intelligence modeling system 400 runs and evaluates the dataset using multiple models to create and deploy the AI ​​model; finally, after the AI ​​model is deployed on the production line, the automatic calibration system 500 automatically calibrates the AI ​​model based on the latest data from the production line.

[0029] Specifically, the operation flow and connection relationship of each system in this invention are as follows: Data collection system 200: Used to connect with data acquisition devices such as sensors and programmable controllers to collect data. The collected data undergoes data preprocessing and data cleaning steps to ensure data quality and consistency.

[0030] Statistical Analysis System 300: Utilizes data provided by Data Collection System 200 to perform statistical analysis, gain data insights, and generate usable data sets. The system automatically verifies data parameters to ensure their accuracy and usability.

[0031] Artificial Intelligence Modeling System 400: Automatically imports the dataset generated by Statistical Analysis System 300 into AI modeling. The Artificial Intelligence Modeling System 400 is used for multi-model operation and evaluation, selecting the best model based on accuracy and recommending it to the user to complete the creation and deployment of the AI ​​model.

[0032] Automatic calibration system 500: After the AI ​​model is deployed on the production line, the automatic calibration system 500 automatically calibrates the model based on the latest data collected by the equipment, ensuring the model's continuous accuracy and adaptability. The automatic calibration system 500 includes a model drift detection module 510, which monitors performance changes and retrains the AI ​​model when data drift is detected.

[0033] The present invention provides a real-time artificial intelligence system 100 for the edge of a production line, which can be presented in the form of a website, software, etc., allowing users to import and apply AI models according to their application scenarios, thereby improving the convenience and efficiency of operation.

[0034] The specific operating mechanisms of the above systems will be explained in detail: 1. Data collection system 200: The data collection system 200 connects with data acquisition devices such as sensors and programmable logic controllers (PLCs). After collecting data from these devices, it performs operations such as data preprocessing and data cleaning to ensure data quality and usability. The collected data can then be used for subsequent system and module analysis or applications.

[0035] Taking the semiconductor industry as an example, in a specific embodiment, the data collection system 200 can collect and filter data for a specific process. In addition to supporting the semiconductor device communication standard SECS / GEM, it can also collect nearly 100% of the data from semiconductor devices, including detailed data generated during the operation of semiconductor devices, including different data formats such as file data, structured and unstructured data, and streaming data, to ensure that the system has enough data for subsequent processing.

[0036] Specifically, the data collection system 200 includes a data preprocessing module 210, a data virtualization module 220, and a data conversion module 230. The data preprocessing module 210 is used to summarize the data, the data virtualization module 220 is used to integrate and temporarily store multiple heterogeneous data for management, and the data conversion module 230 is used to convert the raw data into virtual tables.

[0037] As data is continuously collected, the system will collect increasingly detailed data. Therefore, the data preprocessing module 210 will perform preprocessing operations on the raw data, including data preprocessing and data cleaning, to ensure the consistency of subsequent analysis and application of the system, thereby improving the overall operating performance of the system.

[0038] Taking high-frequency sensor data as an example, in a specific embodiment, the data preprocessing section first summarizes the data source (e.g., high-frequency sensor data) to provide more data details, or cleans the data by including line breaks and blank data, thus completing the data preprocessing. The processing effect is as follows: Figure 3 As shown, the upper half of the image shows the data before preprocessing, and the lower half shows the data after preprocessing (wafer-based summary data).

[0039] The raw data will be recorded according to a timeline, containing data from various high-frequency sensors at each point in time (such as...). Figure 3 Data1, Data2, etc., and so on), these data contain many details, such as the same wafer (e.g. Figure 3 The data (Wafer1 and Wafer2 wafers) were read multiple times within a short period. After data preprocessing, this high-frequency data was summarized and categorized based on the unique characteristics of wafer manufacturing. For example, after data preprocessing, only one representative data point was retained for Wafer1 and Wafer2, instead of displaying detailed information for every time point. This processing method effectively reduces the amount of data while retaining key information, thus completing the data preprocessing process.

[0040] Meanwhile, the data preprocessing module 210 enhances the raw data, thereby improving the accuracy of AI modeling in the subsequent AI modeling system 400. The data preprocessing module 210 improves the quality of the raw data through real-time data normalization, data augmentation using statistical methods and machine learning algorithms, adding noise data perturbation to the raw data, rebalancing the labeling of the raw data, and filling in missing data. This enables the modeling efficiency optimization module 410 in the AI ​​modeling system 400 to optimize the training and prediction performance of the AI ​​model. Adding noise data perturbation to the raw data can replenish data lost during the data collection process by the data collection system 200. Real-time data normalization, data augmentation using statistical methods and machine learning algorithms, rebalancing the labeling of the raw data, and filling in missing data further improve the quality of the raw data, enhancing the learning effect of the AI ​​modeling system 400 in AI modeling. Simultaneously, it labels data that remains unchanged and requires multiple readings, thereby improving the AI ​​performance of the AI ​​modeling system 400. Modeling accuracy.

[0041] The data virtualization module 220 integrates heterogeneous data based on data virtualization and includes open memory, which can provide temporary storage and management of data for different devices or process steps. This enables real-time integration and access of various types of data, thereby improving data compatibility and processing efficiency across diverse sources and providing accurate and complete data for subsequent analysis.

[0042] The data conversion module 230 converts raw data (Log) from semiconductor devices and other sources into virtual table data. This conversion process organizes unstructured data into standardized table format data, which facilitates subsequent data processing, analysis and application. It can also convert the format of different data sources according to user needs to ensure that the system operates normally and efficiently.

[0043] Figure 4 This is a screenshot showing the effect of the data conversion module 230 recording the raw data (HourlyTool ProcessLog) of the equipment's operating status and parameters hourly. Figure 4As shown, the data conversion module 230 filters the raw data, extracts and provides shareable data to the user to protect the confidential information in the raw data. Shareable data can be the raw data required by the user, including raw data that has a critical impact on equipment operation performance and process control, such as pressure, gas flow, RF power, and RF current.

[0044] The data that can be shared retains metadata, including field names, units, and data ranges, thereby ensuring that the data conversion module 230 can selectively share important data that does not involve confidential intellectual property rights with users while protecting intellectual property rights.

[0045] 2. Statistical Analysis System 300: The statistical analysis system 300 performs statistical analysis on the data provided by the data collection system 200 to gain data insights and generate usable data sets. The statistical analysis system 300 has an automated parameter verification function to ensure parameter availability.

[0046] The statistical analysis system 300 automatically filters and verifies parameters based on internal algorithms and indicative data to determine which parameters are valid and usable. This ensures the accuracy and practicality of the data after statistical analysis, enabling its application to data sets such as production lines, and the identification of parameters with reference value.

[0047] Specifically, the statistical analysis system 300 includes a feature extraction module 310 and a wizard function module 320. The feature extraction module 310 extracts features from data from different equipment or process sequences through feature engineering, while the wizard function module 320 uses wizard functions to assist users in building specific models of equipment.

[0048] The feature extraction module 310 extracts features from data from different equipment or process steps through feature engineering. Then, it obtains the feature values ​​of the data through multivariate analysis (MVA). Finally, it integrates domain know-how and statistical analysis to define meaningful key features in the data, thereby filtering out the features with actual influence and their related variables. In a specific embodiment, the feature extraction module 310 can use principal component analysis (PCA) to reduce the dimensionality of variables and rank related variables. After identifying several important variables, it integrates domain knowledge to determine the variables with substantial influence (for example, after identifying the top 50 important variables from 1000 variables, the feature extraction module integrates domain knowledge to determine that the 5th ranked variable is not key, while the 2nd or 3rd ranked variable has significant influence).

[0049] The Wizard function module 320 utilizes a wizard function to assist users in effectively building specific models for various types of devices. The wizard function can generate corresponding model building processes for various devices and verify whether the relevant parameter values ​​set in the model are reasonable.

[0050] Figure 5 The wizard function module provides a flowchart to guide users through setting device parameters, such as... Figure 5 As shown, taking a wafer fabrication etching machine as an example, the specific steps to ensure that the generated final model conforms to the use or application of the wafer fabrication etching machine are as follows: A101. Select Equipment Type: The user selects the equipment type (tool), reaction chamber (chamber), and process (process) to be modeled. For example, the type (tool) can be selected to include wafer manufacturing etching equipment or other specific equipment categories. A102. Extracting Key Parameters: Guides users to extract key equipment parameters based on the selected process or domain knowledge, and determine the variables with substantial impact. After identifying the top 50 important variables from 1000 variables through the feature extraction module, domain knowledge is integrated to determine that the variable ranked 5th is not critical, while the variable ranked 2nd or 3rd has a significant impact. These key equipment parameters include parameters such as special gas valve pressure, chamber gas mass flow, process chamber RF power, and process chamber RF current. A103, Key Parameter Weight Setting 1: Set specific requirements within the domain knowledge of a particular device or application, such as setting the weight of the influence of manufacturing temperature on the model. A104, Key Parameter Weight Setting II: Set specific requirements within the domain knowledge of another specific device or application, such as setting the weight of the impact of manufacturing time on the model; A105. Parameter Weight Verification and Automatic Correction: The wizard function automatically verifies whether the parameter weights entered by the user are reasonable and provides automatic adjustment suggestions; if the RF power and pressure settings are mismatched, the wizard function proposes the best adjustment weight suggestions. A106. Generate a preliminary model: Based on all parameters input by the user, the wizard function generates a preliminary model and displays the simulation results, predicted operating efficiency, energy consumption, and accuracy score. A107. Optimization Suggestions: Provide specific optimization suggestions based on the generated preliminary model; suggestions may include adjusting RF power or gas power parameters to reduce energy consumption based on simulation results. A108. Confirm and complete the model: After the user confirms all parameters and suggestions, the wizard function will complete the final generation of the model.

[0051] By following the steps above, users can complete the model building process.

[0052] 3. Artificial Intelligence Modeling System 400: The AI ​​modeling system 400 imports the data set provided by the statistical analysis system 300 into the AI ​​model, recommends the best value to the user, and completes the deployment of the AI ​​model to the production line.

[0053] The AI ​​modeling system 400 can identify how the model is used, including API connections, extracting datasets for compatibility, and other methods, to provide AI modeling services.

[0054] Specifically, the artificial intelligence modeling system 400 includes a modeling efficiency optimization module 410. The modeling efficiency optimization module 410 applies technologies such as Automated Machine Learning (AutoML), runs multiple models on a dataset, compares the accuracy data generated by each model, selects the best model, and recommends the selected best model to the user to help improve AI modeling efficiency (e.g., for data classification, the accuracy score is used to evaluate model performance; for data regression, the mean absolute percentage error (MAPE) is used to evaluate model accuracy).

[0055] 4. Automatic calibration system 500: The Automatic Calibration System 500 serves as an AI model operation and maintenance platform, connecting in real time with the production line model deployed by the Artificial Intelligence Modeling System 400. By collecting the latest operating data of equipment and machines, it dynamically optimizes model parameters to maintain prediction accuracy.

[0056] Specifically, the automatic calibration system 500 includes a model drift detection module 510. This module monitors the performance changes of the AI ​​model and provides a user interface to set data drift conditions, including anomalies in time and values. It detects data drift by using statistical methods and rule-based methods. For example, it might use a rule based on a five-point consecutive increase or decrease in the AI ​​model's accuracy. If the AI ​​model's accuracy increases or decreases five times consecutively, or if the model is detected as not meeting the original settings, an alert is issued. The model drift detection module 510 automatically initiates model retraining, incorporating the latest data from the equipment and machines on the production line where the model is deployed to retrain the model and improve its performance. In a specific embodiment, for data drift conditions, such as multiple consecutive occurrences of abnormal values ​​or monthly detections of a performance deviation exceeding 10%, the model drift detection module 510 automatically incorporates new data for retraining.

[0057] In a specific embodiment, the real-time artificial intelligence system 100 at the production line edge can also be used for the following applications: AI Calculation and Recommendation: The artificial intelligence modeling system 400 of this invention can establish machine tool models through automatic machine learning technology, and calculate the prediction of machine tool conditions under specific production parameter settings, or recommend the optimal production parameter settings to achieve specific machine tool conditions, which can significantly shorten the time required for machine tool adjustment and reduce resource waste.

[0058] Data integration and protection application: This invention integrates other data, including edge AI data, to achieve a data exchange mode. It uses a gateway to protect the data on the data end and the data of data providers (such as equipment manufacturers), and manages and authorizes data through Software as a Service (SaaS) to ensure the security of the system's data.

[0059] No-Code Technology Integration: The artificial intelligence modeling system 400 of this invention includes No-Code technology integration, allowing users to fill in information through an interface to complete the testing and deployment of AI models.

[0060] Integrating Existing AI Models: The AI ​​modeling system 400 of this invention can integrate AI plug-in technology, providing users with a template application function. Users can export their own AI models as preset templates to integrate different AI models. On the other hand, for specific semiconductor equipment, users can export their own AI models as specific templates, allowing user models to be integrated into a state usable by this system. Therefore, if users have already built models on other platforms, they can also import and apply them to the production line.

[0061] API Integration with Generative Artificial Intelligence (GAI): The AI ​​modeling system 400 of this invention includes an API that can accept query parameters and execute SQL queries to obtain relevant statistical analysis data, so as to effectively integrate generative artificial intelligence (GAI) into the system and apply it to the creation of AI models.

[0062] Model Validity Verification via Statistical Analysis Charts: The modeling efficiency optimization module 410 of the artificial intelligence modeling system 400 of the present invention can provide users with model validity verification through various statistical analysis charts.

[0063] Figures 6a-6d These are the quantile map, Cook distance map, residual map, and prediction confidence interval map of the real-time artificial intelligence system at the production line edge.

[0064] refer to Figure 6aThe vertical axis represents the sample quantiles, and the horizontal axis represents the theoretical quantiles. A quantile plot is used to check whether data conforms to a certain theoretical distribution (usually assumed to be a normal distribution). By comparing the correspondence between the sample quantiles and the theoretical quantiles, if the data points roughly fall on the red diagonal, it means that the sample and the theory have the same distribution.

[0065] refer to Figure 6b The vertical axis represents the Cook distance, and the horizontal axis represents the index of the data points. The Cook distance plot measures the impact of each data point on the regression model. Generally, a larger Cook distance indicates that the data point has a greater impact on the model.

[0066] refer to Figure 6c The vertical axis represents the residuals, and the horizontal axis represents the predicted values. The residual plot shows the relationship between the predicted values ​​and the residuals, used to evaluate the model's fit. Ideally, the residuals should be randomly distributed around 0 without a clear pattern. The presence of a trend or structural distribution may indicate that the model needs further improvement.

[0067] Continue to refer to Figure 6d The vertical axis represents the predicted value, and the horizontal axis represents the actual value. The prediction confidence interval plot shows the correspondence between the predicted and actual values, and includes a 95% confidence interval. The narrower the interval, the higher the model's prediction accuracy; if most actual values ​​fall within the interval, the model's reliability is high.

[0068] This application constructs a highly flexible and adaptive intelligent system that can instantly perceive changes in process and equipment status based on real-time production line data, and automatically update and optimize the AI ​​model. As a result, production efficiency and product quality are improved simultaneously, downtime and delays caused by unexpected anomalies are significantly reduced, and ultimately, effective cost reduction and continuous capacity release are achieved.

[0069] The specific embodiments of this application have been described above, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0070] In the description of this application, it should be understood that the terms "upper," "lower," "inner," "outer," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are used only for the convenience of describing this application and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application. The terms "comprising" or "including" do not exclude the presence of elements or steps not listed in the claims. The terms "a" or "an" preceding an element do not exclude the presence of a plurality of such elements. The simple fact that certain measures are recited in mutually different dependent claims does not indicate that combinations of these measures cannot be used for improvement. Any reference signs in the claims should not be construed as limiting the scope.

Claims

1. A real-time artificial intelligence system at the production line edge, comprising: a data collection system for preprocessing and cleaning data; a statistical analysis system for analyzing the preprocessed and cleaned data to generate a usable dataset; an artificial intelligence modeling system for running and evaluating multiple models on the dataset to create and deploy an AI model; and an automatic calibration system for automatically calibrating the AI ​​model based on the latest data from the production line after the AI ​​model is deployed on the production line; characterized in that: The data collection system includes a data preprocessing module, a data virtualization module, and a data conversion module. The data preprocessing module summarizes the data; the data virtualization module integrates and temporarily stores multiple heterogeneous data sets for management; and the data conversion module converts raw data into a virtual table. This conversion module also filters the raw data, extracting and providing shareable data to users to protect confidential information within the raw data. Specifically, shareable data refers to raw data that users require, including pressure, gas flow, and radio frequency power. The system collects raw data such as power and radio frequency current (RFCurrent), which have a critical impact on equipment operation performance and process control. It includes a feature extraction module and a wizard module. The feature extraction module extracts features from data from different equipment or process sequences through feature engineering. The wizard module assists users in building a specific model for a particular equipment. The feature extraction module extracts features from data from different equipment or process sequences through feature engineering, then obtains feature values ​​through multivariate analysis. After integrating knowledge management and statistical analysis content from semiconductor process-specific domains, it uses Principal Component Analysis (PCA) to reduce variable dimensionality and rank relevant variables. After identifying several important variables, it integrates semiconductor process-specific domain knowledge to determine variables with substantial impact. From 1000 variables, the feature extraction module identifies the top 50 important variables. Then, integrating semiconductor process-specific domain knowledge, it determines that the 5th ranked variable is not critical, while the 2nd or 3rd ranked variable has a significant impact. This feature extraction module uses principal component analysis to reduce the dimensionality of variables and rank related variables in order to identify several important variables and integrate semiconductor process-specific knowledge to determine the variables with substantial impact.

2. The real-time artificial intelligence system at the production line edge according to claim 1, characterized in that, This data collection system can collect and filter data for specific processes. It supports the semiconductor device communication standard SECS / GEM and can be used to collect data in file format, structured and unstructured data, streaming data and other data formats from semiconductor devices.

3. The real-time artificial intelligence system at the production line edge according to claim 1, characterized in that, This feature extraction module utilizes principal component analysis (PCA) to reduce the dimensionality of variables and rank relevant variables. It identifies several important variables and integrates semiconductor process-specific knowledge to determine the variables with substantial impact. From 1000 variables, the feature extraction module identifies the top 50 most important variables. Then, by integrating semiconductor process-specific knowledge, it determines that the variable ranked 5th is not critical, while the 2nd or 3rd variable has a significant impact.

4. The real-time artificial intelligence system at the production line edge according to claim 1, characterized in that, The creation of this specific model involves the following steps: Select Equipment Type: The user selects the type of equipment, reaction chamber, and process to be modeled. Key parameter extraction: Guide the user to extract a key equipment parameter based on the selected process or semiconductor process-specific knowledge, and determine the variable with substantial impact. After the feature extraction module identifies the top 50 important variables from 1000 variables, it integrates semiconductor process-specific knowledge to determine that the variable ranked 5th is not critical, while the variable ranked 2nd or 3rd has a significant impact. The key equipment parameter includes parameters such as special gas valve pressure, specific reaction chamber pipeline gas massflow, process chamber RF power, and process chamber RF current. Key parameter weight setting 1: Set the weight of the influence of manufacturing temperature and manufacturing time on the model within the proprietary knowledge of semiconductor manufacturing processes for a specific device or application; Key parameter weight setting 2: Set the weight of the influence of manufacturing temperature and manufacturing time on the model within the knowledge of semiconductor process expertise for another specific device or application; Parameter weight verification and automatic correction: This wizard function automatically verifies whether the parameter weights entered by the user are reasonable and provides automatic adjustment suggestions; Generate a preliminary model: Based on all the parameters input by the user, the wizard function generates a preliminary model and displays the simulation results, predicted operating efficiency, energy consumption, and accuracy score. Optimization suggestions: Provide specific optimization suggestions based on the generated preliminary model; and Confirm and complete the model: After the user confirms all parameters and suggestions, the wizard function will complete the final generation of the model.

5. The real-time artificial intelligence system at the production line edge according to claim 1, characterized in that, The artificial intelligence modeling system includes a modeling efficiency optimization module. This module uses automated machine learning technology to run multiple models on the dataset and compares the accuracy data generated by each model to select the best model.

6. The real-time artificial intelligence system at the production line edge according to claim 5, characterized in that, When classifying data, this comparison evaluates model performance based on accuracy scores.

7. The real-time artificial intelligence system at the production line edge according to claim 5, characterized in that, When performing regression on data, the mean absolute percentage error is used to evaluate the accuracy of the model.

8. The real-time artificial intelligence system at the production line edge according to claim 5, characterized in that, This modeling efficiency optimization module can provide users with a way to verify the effectiveness of the model through various statistical analysis charts.

9. The real-time artificial intelligence system at the production line edge according to claim 1, characterized in that, The automatic calibration system includes a model drift detection module, which is responsible for monitoring the performance changes of the AI ​​model and provides a user interface to set the conditions for data drift.

10. The real-time artificial intelligence system at the production line edge according to claim 9, characterized in that, Data drift is determined using methods including statistical approaches and rule-based methods.

11. The real-time artificial intelligence system at the production line edge according to claim 1, characterized in that, This artificial intelligence modeling system can build a machine model using machine learning technology and calculate the prediction of the machine's condition under specific production parameter settings.

12. The real-time artificial intelligence system at the production line edge according to claim 1, characterized in that, This artificial intelligence modeling system can integrate data, including edge AI, to realize data exchange mode, and uses gateways to protect data from both the data source and the data provider, and manages and authorizes data through Software as a Service.

13. The real-time artificial intelligence system at the production line edge according to claim 1, characterized in that, This artificial intelligence modeling system can integrate AIPlug-in technology, providing users with a set of template functions. Users can export their own AI models as a preset template to integrate different AI models.

14. The real-time artificial intelligence system at the production line edge according to claim 13, characterized in that, For a specific semiconductor device, the user can export their own AI model as a custom template.

15. The real-time artificial intelligence system at the production line edge according to claim 1, characterized in that, This artificial intelligence modeling system includes an API that can accept query parameters and execute SQL queries to obtain relevant statistical analysis data, effectively integrating generative artificial intelligence into the artificial intelligence modeling system and applying it to the creation of AI models.

16. The real-time artificial intelligence system at the production line edge according to claim 1, wherein the data preprocessing module enhances the quality of the original data by including real-time data normalization, data augmentation using statistical methods and machine learning algorithms, adding noise data perturbation to the original data, rebalancing the labeling of the original data and filling in missing data, thereby achieving optimal performance of AI model training and prediction.

Images