Industrial data inspection method and system based on root cause analysis

Abstract

The invention discloses an industrial data inspection method and system based on root cause analysis, and relates to the technical field of industrial data inspection methods, and the method comprises the steps: determining a plurality of industrial data features based on the recognition of an optimized data combination; and according to the plurality of industrial data characteristics, the corresponding data distribution characteristics and the data discrete characteristics, the corresponding adaptive test mode is determined, so that the accuracy of the adaptive test mode is improved. Therefore, according to the adaptive test mode, the root cause analysis framework and the plurality of pieces of semiconductor data, test effect coefficients of each piece of semiconductor data under different dimensions are determined; and determining a data inspection event of each piece of semiconductor data according to the inspection effect coefficient of each piece of semiconductor data under different dimensions and the data priority corresponding to each piece of semiconductor data, and determining a corresponding data inspection result based on the data inspection event of each piece of semiconductor data, the corresponding data inspection content and the data fusion system. And the accuracy of the data inspection result is improved.

Description

Technical Field

[0001] This invention relates to the technical field of industrial data verification methods, and more particularly to an industrial data verification method and system based on root cause analysis. Background Technology

[0002] For semiconductor technology, semiconductor industry data has been introduced for testing. In existing technologies, multiple industrial data points for semiconductors are collected, and corresponding data combinations are determined based on these data points. However, the impact of missing and outlier values ​​is ignored, which reduces the accuracy of the testing model. This makes it impossible to guarantee the accuracy of the testing effect coefficients of various semiconductor data points under different dimensions, thus affecting the accuracy of the data testing results. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art. This invention provides a method and system for verifying industrial data based on root cause analysis.

[0004] This invention provides a method for verifying industrial data based on root cause analysis, comprising: In the semiconductor database, industrial data sets are collected, and multiple semiconductor data are identified based on the identification of the industrial data sets. Based on each semiconductor data and its corresponding data type, a hybrid industrial data combination is determined. Based on the identification of mixed industrial data combinations, the corresponding missing values ​​and outliers are determined, and the optimized data combinations are determined according to the missing values, outliers and the corresponding data optimization system. Multiple industrial data features are identified based on the optimized data combination; an adaptive verification mode is determined based on the multiple industrial data features, their corresponding data distribution characteristics, and data discrete characteristics. A root cause analysis system is collected, and a corresponding root cause analysis framework is determined based on the identification of the root cause analysis system. The test effect coefficients of each semiconductor data in different dimensions are determined according to the adaptive test mode, the root cause analysis framework and multiple semiconductor data. The data inspection events for each semiconductor data are determined based on the inspection effect coefficients of each semiconductor data under different dimensions and the data priority corresponding to each semiconductor data. The corresponding data inspection results are determined based on the data inspection events of each semiconductor data, the corresponding data inspection content, and the data fusion system.

[0005] This invention provides an apparatus for verifying industrial data based on root cause analysis. The apparatus is applied to the aforementioned method for verifying industrial data based on root cause analysis. The apparatus includes: The hybrid industrial data combination module is used to collect industrial data sets in a semiconductor database, identify multiple semiconductor data based on the identification of industrial data sets, and determine the hybrid industrial data combination based on each semiconductor data and its corresponding data type. The optimized data combination module is used to identify the corresponding missing values ​​and outliers based on the identification of mixed industrial data combinations, and to determine the optimized data combination based on the missing values, outliers and the corresponding data optimization system; The adaptive verification module is used to identify multiple industrial data features based on the optimized data combination; and to determine the corresponding adaptive verification mode according to the multiple industrial data features, the corresponding data distribution characteristics, and the data discrete characteristics. The test effect coefficient module is used to collect the root cause analysis system, determine the corresponding root cause analysis framework based on the identification of the root cause analysis system, and determine the test effect coefficient of each semiconductor data in different dimensions according to the adaptive test mode, the root cause analysis framework and multiple semiconductor data. The data verification result module is used to determine the data verification events for each semiconductor data based on the verification effect coefficients of each semiconductor data in different dimensions and the data priority corresponding to each semiconductor data. Based on the data verification events of each semiconductor data, the corresponding data verification content, and the data fusion system, the corresponding data verification results are determined.

[0006] Compared with the prior art, the beneficial effects of the present invention are: In this embodiment of the invention, the method involves collecting an industrial data set from a semiconductor database, identifying multiple semiconductor data based on the industrial data set, determining a hybrid industrial data combination based on each semiconductor data and its corresponding data type, identifying corresponding missing values ​​and outliers based on the identification of the hybrid industrial data combination, determining an optimized data combination based on the missing values, outliers, and their corresponding data optimization system, identifying multiple industrial data features based on the identification of the optimized data combination, and determining a corresponding adaptive verification mode based on the multiple industrial data features, their corresponding data distribution characteristics, and data discrete characteristics. The introduction of a hybrid industrial data combination further controls the optimized data combination, incorporating a holistic consideration of multiple industrial data features, their corresponding data distribution characteristics, and data discrete characteristics, thereby improving the accuracy of the adaptive verification mode.

[0007] Therefore, a root cause analysis system is collected, and a corresponding root cause analysis framework is determined based on the identification of the root cause analysis system. Based on this adaptive testing mode, the root cause analysis framework, and multiple semiconductor data, the testing effect coefficients of each semiconductor data point under different dimensions are determined. Data testing events for each semiconductor data point are determined based on the testing effect coefficients of each semiconductor data point under different dimensions and the data priority corresponding to each semiconductor data point. Based on the data testing events of each semiconductor data point, the corresponding data testing content, and the data fusion system, the corresponding data testing results are determined. The introduction of testing effect coefficients for each semiconductor data point under different dimensions achieves a holistic consideration of the data testing events, corresponding data testing content, and the data fusion system for each semiconductor data point, thus improving the accuracy of the data testing results. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating the method for verifying industrial data based on root cause analysis in an embodiment of the present invention. Figure 2 This is a flowchart illustrating step S11 in the method for verifying industrial data based on root cause analysis in this embodiment of the invention. Figure 3 This is a flowchart illustrating step S12 in the method for verifying industrial data based on root cause analysis in this embodiment of the invention. Figure 4 This is a flowchart illustrating step S13 in the method for verifying industrial data based on root cause analysis in this embodiment of the invention. Figure 5 This is a flowchart illustrating step S14 in the method for verifying industrial data based on root cause analysis in this embodiment of the invention. Figure 6 This is a flowchart illustrating step S15 in the method for verifying industrial data based on root cause analysis in an embodiment of the present invention. Figure 7 This is a schematic diagram of the structural composition of the industrial data testing device based on root cause analysis in an embodiment of the present invention. Detailed Implementation

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

[0010] Please see Figures 1 to 7 A method for verifying industrial data based on root cause analysis, applied to industrial data verification scenarios; the method for verifying industrial data based on root cause analysis includes: Step S11: In the semiconductor database, collect industrial data sets, identify multiple semiconductor data based on the identification of industrial data sets, and determine a hybrid industrial data combination based on each semiconductor data and its corresponding data type; Step S12: Based on the identification of mixed industrial data combinations, determine the corresponding missing values ​​and outliers, and determine the optimized data combinations according to the missing values, outliers and the corresponding data optimization system; Step S13: Identify multiple industrial data features based on the optimized data combination; determine the corresponding adaptive verification mode according to the multiple industrial data features, the corresponding data distribution characteristics, and the data discrete characteristics; Step S14: Collect the root cause analysis system, determine the corresponding root cause analysis framework based on the identification of the root cause analysis system, and determine the test effect coefficient of each semiconductor data in different dimensions according to the adaptive test mode, the root cause analysis framework and multiple semiconductor data. Step S15: Determine the data inspection events for each semiconductor data based on the inspection effect coefficients of each semiconductor data under different dimensions and the data priority corresponding to each semiconductor data. Determine the corresponding data inspection results based on the data inspection events, corresponding data inspection content, and data fusion system for each semiconductor data.

[0011] refer to Figure 2 In step S11, the specific steps are as follows: S111: Collect semiconductor database and monitor semiconductor database in real time. Based on the detection of semiconductor database, determine industrial data space and determine industrial data set by traversing industrial data space. S112: Based on the detection of industrial data sets, the data content of each industrial data is determined. Multiple semiconductor data are determined according to the data content of each industrial data and the data tags of the semiconductor database. Among the multiple semiconductor data, the corresponding data type is determined based on the identification of each semiconductor data. A hybrid industrial data combination is determined according to each semiconductor data, the corresponding data relationship and the corresponding data type.

[0012] In the embodiments of this application, a semiconductor database is collected and monitored in real time. Based on the detection of the semiconductor database, an industrial data space is determined, and an industrial data set is determined by traversing the industrial data space. This approach incorporates the overall considerations for the detection of the semiconductor database and ensures the accuracy of the industrial data space.

[0013] At this point, the system typically employs techniques such as database log change capture, message queue subscription, or API polling to register "listeners" on change logs of core data sources such as Semiconductor Manufacturing Execution System (MES), Equipment Engineering System (EES), and Statistical Process Control (SPC) databases. The monitored objects are extremely broad, covering process parameters such as exposure energy and focal length of lithography machines, and cavity pressure and gas flow rate of etching machines; measurement data such as critical dimensions (CD), film thickness, and defect density; equipment status information such as equipment ID, uptime, and alarm codes; batch and material data such as wafer batch number and photoresist batch number; and environmental data such as temperature, humidity, and particle size of cleanrooms. The entire monitoring system must possess high throughput and low latency characteristics to ensure the capture of massive amounts of data generated by high-frequency equipment without affecting the normal operation of the production system.

[0014] The "Industrial Data Space" is not a physical database, but a logical, multi-dimensional data framework. It defines the scope and dimensions of data to be collected for effective root cause analysis, serving as a "blueprint" or "metamodel" for data collection. This space consists of multiple dimensions, each representing a perspective of data classification. Typical dimensions include: time dimension (based on time and product generation); product dimension (based on product model and technology node); equipment dimension (based on equipment type and serial number); process step dimension (based on upstream, midstream, downstream, or specific process); and material dimension (based on key raw material batch number). Theoretically, all data collection points can be defined through the Cartesian product of these dimensions. For example, a data point can be uniquely identified as [Time: 2023-Q4, Product: 5nm logic chip, Equipment: Litho-Cell-05, Process: Lithography].

[0015] By systematically "traversing" the industrial data space defined in the previous step, the logical blueprint is transformed into concrete, analyzable data entities. Based on the dimensional definition of the data space, the system generates a series of data query tasks and distributes them to various data sources for data extraction. The traversal process strives for comprehensiveness, ensuring coverage of all defined data points within the space. Since the data comes from different systems and exhibits differences in timestamps and granularity, data aggregation and alignment are crucial tasks in this step. For example, the system associates process parameter data at the wafer level with measurement data at the chip level using a common "wafer ID," forming a complete analysis sample. All extracted and aligned data records are aggregated into a vast "industrial data set," typically existing in the form of a structured data table, where each row represents an analysis unit (such as a wafer), and each column represents a feature variable.

[0016] Furthermore, based on the detection of industrial data sets and the data content of each industrial data set, multiple semiconductor data are determined according to the data content of each industrial data set and the data tags of the semiconductor database. Among the multiple semiconductor data, the corresponding data type is determined based on the identification of each semiconductor data. A hybrid industrial data combination is determined based on each semiconductor data, the corresponding data relationship, and the corresponding data type. This approach takes into account the overall consideration of each semiconductor data, the corresponding data relationship, and the corresponding data type, ensuring the accuracy of the hybrid industrial data combination.

[0017] At this point, the industrial data set output by S111 is subjected to a deep scan and parsing of each record. Each original data record is treated as a structured or semi-structured data entity, and all data fields and their corresponding values ​​are extracted. In terms of technical implementation, the system adopts pattern recognition or metadata-driven methods. For structured data (such as database tables), its schema definition is read directly, while for semi-structured data (such as JSON or XML logs), key-value pairs are extracted through a parser. The parsing object is each field of the data record. The system extracts all field names (such as wafer_id, exposure_energy) and their corresponding values ​​(such as "W5NM-001", 25.5), forming a detailed "data content list".

[0018] The parsed data content is associated with and verified against the metadata (i.e., "data tags") of the semiconductor database, thereby "instantiating" the abstract data fields into "semiconductor data" with clear business meaning. Data tags refer to metadata stored in the database or data dictionary, which describes the business definition, data type, dimension, value range, and relationships of each data field. For example, the tag for tool_id includes: "unique device identifier, associated with the device table, data type is string". The data determination process is that the system matches the parsed field names with the data tags. Through matching, the system not only confirms the validity of the field, but also obtains its rich contextual information. This process "anchors" the raw data to the business model, making it a meaningful object of analysis.

[0019] Based on the established business attributes and inherent characteristics of semiconductor data, it is automatically categorized into different data types, which is crucial for subsequent "hybrid" processing. The judgment logic mainly comes from the data type definitions in the data tags and the system's built-in rule engine. The data is mainly divided into two categories: numerical data, which is data whose field values ​​can be mathematically operated on (such as addition, subtraction, averaging), such as exposure_energy, focus_offset, and cd_measurement (critical dimension measurement values); and categorical data, which is data whose field values ​​are used to identify categories, states, or identities and cannot be mathematically operated on, such as tool_id (equipment model), recipe (recipe version), and operator_id (operator). This step directly corresponds to split_diff_type_of_data in the reference scheme and is a prerequisite for using different processing algorithms for different types of data.

[0020] All identified and classified semiconductor data are reorganized into a unified, structured "hybrid industrial data portfolio" based on their inherent "data relationships." Data relationships refer to the logical connections between data fields, with the core relationship being the "entity-attribute" relationship. For example, exposure_energy and focus_offset are attributes of the wafer with wafer_id "W5NM-001" when it is processed on the equipment with tool_id "LITHO-ASML-05." During the portfolio construction process, the system uses a core business entity (such as wafer ID) as the key to combine all related data fields (whether numerical or categorical) into a complete record. The result of this combination is a structured dataset, typically represented as a wide table that clearly divides the numerical and categorical areas. Each row represents a complete analysis sample, and each column is a feature variable.

[0021] refer to Figure 3 In step S12, the specific steps are as follows: S121: Collect a combination of mixed industrial data, determine numerical and non-numerical industrial data based on the detection of the mixed industrial data combination, and perform synchronous data processing on the numerical and non-numerical industrial data. S122: In the synchronous data processing of numerical industrial data and non-numerical industrial data, multiple independent variables are determined based on the numerical industrial data and non-numerical industrial data, and the corresponding missing values ​​are predicted based on the multiple independent variables and the corresponding neural network model. S123: Collect the data matrix corresponding to the mixed industrial data combination, determine the corresponding low-rank matrix based on the data matrix, numerical industrial data and non-numerical industrial data, and determine the corresponding outliers based on the optimization of the low-rank matrix; determine the optimized data combination based on the missing values, outliers and the corresponding data optimization system.

[0022] In the embodiments of this application, a hybrid industrial data combination is collected, numerical industrial data and non-numerical industrial data are determined based on the detection of the hybrid industrial data combination, and the numerical industrial data and non-numerical industrial data are processed synchronously, which is compatible with the overall consideration of the detection of the hybrid industrial data combination and ensures the accuracy of the numerical industrial data and non-numerical industrial data.

[0023] At this point, the system loads the hybrid industrial data combination generated in stage S11 and performs a comprehensive metadata check on it. During the data loading phase, the system loads the hybrid data combination stored in the data lake or data warehouse into the in-memory computing engine to prepare for subsequent efficient processing. Subsequently, the system does not directly process the data values, but scans the data schema (structure information), reading the name, data type (such as string, float, integer), and constraints (such as value range) of each column. This check process aims to establish a "blueprint" of the data structure, providing a reliable basis for subsequent precise type separation.

[0024] Based on the results of metadata inspection, the system performs an automated type inference and classification process, precisely classifying each column in the data set into either a numeric or non-numeric category. Type inference follows explicit rules: all columns identified as integers, floats, or doubles are classified as numeric industrial data because these data support mathematical operations (such as summation, averaging, and standard deviation); while all columns identified as strings, booleans, or categories are classified as non-numeric industrial data (typically categorized data in industrial scenarios), primarily used to identify categories, states, or identities. The classification results are stored in a configuration object or data structure, explicitly recording which columns are numeric and which are non-numeric. This classification result serves as the routing table for the subsequent "synchronous processing" framework.

[0025] The system constructs a data processing pipeline with parallel processing branches: one dedicated to handling numerical data and the other to handling non-numerical data. These two branches are logically synchronized, achieved primarily through two methods: row-level synchronization and state synchronization. Row-level synchronization uses a unified row index for all operations. When the pipeline decides to operate on a row (e.g., marking it as imputable or identifying it as an anomaly in a subsequent step), the operation is simultaneously applied to both the numerical and non-numerical data within that row, preventing data fragmentation caused by separate processing. State synchronization maintains a global state vector with a length equal to the number of rows in the data set. The state of each row (e.g., "normal," "missing," "anomaly") is recorded in this vector. Both the numerical and non-numerical branches read and update this shared state vector, synchronizing the processing logic. This synchronous processing framework is the best practice for handling mixed-type data. It guarantees the atomicity of data entities (e.g., all information about a wafer), enabling subsequent missing value prediction and outlier detection to be performed on a complete and consistent data view, significantly improving the reliability of data optimization.

[0026] Furthermore, in the synchronous data processing of numerical and non-numerical industrial data, multiple independent variables are determined based on the numerical and non-numerical industrial data. Based on the multiple independent variables and the corresponding neural network model, the corresponding missing values ​​are predicted. This approach takes into account both numerical and non-numerical industrial data and ensures the accuracy of multiple independent variables.

[0027] At this point, the system uses all other variables (including numerical and non-numerical types) in the data combination except for the target variable as the initial candidate feature pool. Subsequently, the system performs fine feature engineering: for categorical independent variables (such as tool_id, recipe), one-hot encoding or embedding techniques are used to convert them into numerical vectors that can be processed by the neural network; for numerical independent variables (such as exposure_energy), standardization or normalization is performed to eliminate the influence of dimensions and accelerate model convergence; the system calculates the correlation (such as Pearson, Spearman, or mutual information) between candidate features and the target variable (using only its non-missing parts), and selects the top-N features with the highest correlation as the final set of independent variables. This can improve model performance and reduce computational complexity.

[0028] The system constructs and trains a neural network model using the selected independent variables to learn the complex nonlinear mapping relationship between the independent and target variables. For tabular data, a multilayer perceptron (MLP) is typically used, consisting of an input layer, several hidden layers, and an output layer, which can effectively learn the interactions between features. The training dataset consists of samples with all target variables that are not missing, with the input being the independent variables selected in the previous step and the output being the true values ​​of the target variables. During model training, the system typically uses mean squared error (MSE) as the loss function and employs optimization algorithms such as Adam or SGD to iteratively update the network weights through backpropagation to minimize the loss. At the same time, to prevent overfitting, techniques such as Dropout and L1 / L2 regularization are used, and the model performance is monitored using a validation set. An early stopping strategy is used to ensure that the model achieves good generalization ability.

[0029] It uses a trained model to predict missing values ​​and accurately fills the results back into the original data set. The system extracts samples with missing target variables (such as wafer W5NM-005) and obtains their corresponding independent variable data. This independent variable data must undergo the same feature engineering processing (encoding, standardization) as during training to ensure the consistency of data distribution. Subsequently, the processed independent variable data is input into the trained neural network model, which performs forward propagation calculations and outputs a predicted value. The system fills the corresponding missing position in the original data set with this predicted value, and this process is traceable. The system will explicitly mark this value as "model prediction filling" in the metadata for differentiation in subsequent analysis.

[0030] Therefore, a data matrix corresponding to the combination of mixed industrial data is collected. Based on this data matrix, numerical industrial data and non-numerical industrial data, a corresponding low-rank matrix is ​​determined, and the corresponding outliers are determined based on the optimization of the low-rank matrix. Based on the missing values, outliers and the corresponding data optimization system, an optimized data combination is determined. This approach takes into account the overall consideration of missing values, outliers and the corresponding data optimization system, ensuring the accuracy of the optimized data combination.

[0031] At this point, the "hybrid industrial data combination" after missing value imputation in S122 is transformed into a numerical data matrix suitable for matrix operations. The system treats the entire data combination as an m x n matrix M, where m is the number of samples (e.g., the number of wafers) and n is the number of features. Since all elements in the matrix must be numerical, all non-numerical (categorical) industrial data, such as tool_id and recipe, must be converted into numerical values ​​through encoding techniques (most commonly one-hot encoding). In addition, in order to eliminate the difference in dimensions between different numerical variables and prevent certain variables with large numerical ranges from dominating the matrix decomposition, the system usually performs Z-score standardization on each column of the matrix (mean is 0, variance is 1).

[0032] It separates the low-rank component representing "normal patterns" and the sparse component representing "abnormalities" from the original data matrix M. This problem is usually modeled as Robust Principal Component Analysis (RPCA), which aims to decompose the original matrix M into the sum of two matrices: M = L + S. Here, L is a low-rank matrix that captures the main, global, linear correlation structure in the data, such as learning a general pattern that "ASML devices typically use high exposure energy and have small CD values." S is a sparse matrix with most of its elements being 0, but its non-zero elements represent outliers in the original data that deviate from the normal patterns. Solving L and S is a non-convex optimization problem. In practice, efficient convex relaxation algorithms, such as the Augmented Lagrange Multiplier Method or the Alternating Direction Multiplier Method (ADMM), are used to iteratively update L and S, eventually converging to an optimal solution. The rank r of the low-rank matrix L is a key hyperparameter, representing the number of main patterns in the data, which can be determined by methods such as cross-validation or the elbow rule.

[0033] Based on the optimization results, outliers are quantified and identified, and all optimization results are ultimately integrated to form the output. The element value S_ij in the sparse matrix S directly measures the "abnormality" of the original data point M_ij. The system sets an outlier threshold (e.g., based on the distribution of the absolute values ​​of the elements in S, such as taking the 99.9th percentile). Any point where |S_ij| exceeds this threshold is judged as an outlier. The identified outliers can be marked according to business needs, replaced with the corresponding predicted value L_ij in the low-rank matrix L (equivalent to "correcting" the outlier), or, in extreme cases, the entire sample can be deleted. The system integrates the missing values ​​filled in S122 and the outliers processed in S123 back into the original data structure, and finally outputs an "optimized data combination", in which missing values ​​are intelligently filled, outliers are identified and processed, and data quality and robustness are greatly improved.

[0034] Step S123 provides a powerful, unsupervised anomaly detection method through three technical steps: matrix construction, low-rank decomposition, and anomaly determination. It does not rely on preset rules but learns the "normal" form from the structure of the data itself, thereby accurately identifying hidden anomalies that deviate from the global pattern and are difficult to detect through simple thresholds. It is a key line of defense to ensure the accuracy of data inspection methods in the semiconductor industry.

[0035] refer to Figure 4 In step S13, the specific steps are as follows: S131: Based on the detection of optimized data combinations, multiple industrial data populations are identified. At the same time, target variables of mixed industrial data combinations are collected, and multiple industrial data features are determined according to each industrial data population, target variable and corresponding data application scenario. S132: Determine the corresponding data distribution characteristics based on the first-level data detection of the optimized data combination. These data distribution characteristics include normality and homogeneity of variance. Determine the first adaptive test coefficient based on multiple industrial data features and the corresponding data distribution characteristics. S133: Determine the corresponding data discrete features based on the second-level data detection of the optimized data combination, determine the second adaptive test coefficient based on multiple industrial data features and data discrete features, and determine the corresponding adaptive test mode based on the mapping relationship between the first adaptive test coefficient, the second adaptive test coefficient and the adaptive test mode.

[0036] In the embodiments of this application, multiple industrial data populations are determined based on the detection of optimized data combinations. At the same time, target variables of mixed industrial data combinations are collected. Multiple industrial data features are determined according to each industrial data population, target variable and corresponding data application scenario. This approach takes into account the overall consideration of each industrial data population, target variable and corresponding data application scenario, and ensures the accuracy of multiple industrial data features.

[0037] At this point, all variables (columns) in the data set are logically grouped to form different "populations." This division goes beyond a simple numerical / categorical dichotomy, focusing more on the commonalities of variables in business logic and statistical characteristics. The division is based on multiple dimensions, including the most basic data types (such as continuous numerical and discrete categorical types), variables with similar business meanings (such as all parameters from lithography equipment forming a "lithography parameter population"), and the statistical role of variables in the analysis (such as "input variable population" or "output variable population"). Each population is given a clear definition and label, forming a metadata structure, which is the foundation for the system's deep understanding of data composition.

[0038] The system identifies the target variable y from the data combination. This variable is usually specified by the user when initiating the analysis task, or it is automatically recommended by the system based on preset rules (such as variable names containing keywords such as "yield" or "defect"). Once identified, the system records the data type (numerical or categorical), business meaning (such as "final wafer yield"), and its role in the analysis (dependent variable) of the target variable. This information is crucial for subsequently selecting the correct testing and scoring methods.

[0039] For each data population, the system extracts a series of "industrial data features" that accurately describe its statistical characteristics, taking into account the target variable and the specific application scenario. These features serve as the quantitative basis for subsequent adaptive decision-making. Feature extraction is dynamic and context-dependent: for numerical populations, the system extracts their preliminary correlation with the target variable (such as Pearson or Spearman correlation coefficients), distribution shape (skewness, kurtosis), and dispersion (standard deviation, interquartile range); for categorical populations, it extracts their class cardinality (number of classes), sample distribution (sample size and proportion of each class), and the strength of their association with the target variable (e.g., calculating the mean of the target variable for each class). Furthermore, feature selection is adjusted according to the application scenario; for example, in a "yield improvement" scenario, more attention is paid to features related to yield.

[0040] Specifically, for 5nm logic chip wafers, the system analyzes the optimized data combinations and identifies the following industrial data groups: process parameter group (numerical), including exposure_energy and focus_offset, which are continuous parameters that can be actively adjusted during production; measurement result group (numerical), including cd_mean and film_thickness, which are passively measured values ​​of process results; and equipment and material group (categorical), including tool_id, recipe, and photoresist_batch, which are discrete, identifiable classification information.

[0041] In this analysis, the engineer's goal is to improve wafer yield; therefore, the system collects final_yield (final wafer yield) as the target variable y; the system records it as a numerical variable, representing the yield percentage of each wafer, and sets it as the core object that all subsequent analyses need to interpret.

[0042] The system extracted key features for each population: In the process parameter population, the Pearson correlation coefficient between exposure_energy and final_yield was found to be -0.65, indicating a strong negative correlation; in the measurement results population, the Spearman correlation coefficient between cd_mean and final_yield reached -0.72, showing an even stronger negative correlation; in the equipment and material population, the cardinality of tool_id was 3, and the system calculations showed that the average yield of ASML-05 equipment was 92.5%, while the average yield of NIKON-02 was 89.8%, revealing preliminary inter-group differences and providing important clues for subsequent testing.

[0043] Furthermore, the corresponding data distribution characteristics are determined based on the first layer of data detection of the optimized data combination. These data distribution characteristics include normality and homogeneity of variance. The first adaptive test coefficient is determined based on multiple industrial data features and their corresponding data distribution characteristics, taking into account the overall consideration of multiple industrial data features and their corresponding data distribution characteristics, thus ensuring the accuracy of the first adaptive test coefficient.

[0044] At this point, two core statistical hypothesis tests are performed on the optimized data combination to assess whether the data meets the applicable conditions for parametric tests. The first is the normality test, which aims to test whether a numerical variable follows a normal distribution. This is the theoretical foundation of many parametric tests (such as the t-test and ANOVA). The system applies algorithms such as the Shapiro-Wilk test or D'Agostino's K² test to each numerical industrial data feature (such as exposure_energy and cd_mean), and determines the significance level based on the output p-value and the preset significance level α (usually 0.05). The system performs a comparison to label the data as "normally distributed" or "non-normally distributed." The second step is a homogeneity of variance test, which aims to check whether the variances of a numerical dependent variable are equal across different groups when a categorical variable is used as a grouping factor. This is a crucial prerequisite for comparing inter-group means. For each categorical industrial data feature (such as tool_id), the system uses methods like the Levene test or Bartlett test to check whether the variances of the target variable (or other key numerical variables) under its grouping are homogeneous. Similarly, the p-value is used to label the results as "homogeneous variance" or "non-homogeneous variance."

[0045] The qualitative "data distribution characteristics" obtained in the previous step are transformed into a quantitative, computable "first adaptive test coefficient" (denoted as α1). This coefficient comprehensively evaluates the degree to which the data fits the parametric test and is a comprehensive score between 0 and 1. Its design aims to reflect the "safety" or "confidence" of the data in satisfying the parametric test hypothesis. The system uses a weighted or penalized mechanism to calculate α1. For example, the initial value of α1 is set to 1.0. If the normality test result of a numerical variable is "non-normal", a penalty is applied to the α1 related to that variable (such as multiplying by 0.4). If the homogeneity of variance test result of a categorical variable grouping is "non-homogeneous", another penalty is applied (such as multiplying by 0.5). The closer α1 is to 1, the better the data fits the parametric test. The closer α1 is to 0, the more seriously the parametric test hypothesis is violated, and nonparametric methods must be considered.

[0046] Specifically, the system performs the first level of detection on the optimized data combination. In the normality detection, the Shapiro-Wilk test is performed on exposure_energy, and the p-value is 0.15 (>0.05), indicating that its data distribution characteristics are "normal". The test is performed on cd_mean (mean of critical dimensions), and the p-value is 0.001 (<0.05), indicating that its data distribution characteristics are "non-normal". In the homogeneity of variance detection, the Levene test is performed on final_yield (wafer yield) with tool_id (device ID) as the grouping factor, and the p-value is 0.02 (<0.05), indicating that the yield data distribution characteristics under tool_id grouping are "non-homogeneous".

[0047] The system calculates the first adaptive test coefficients for each variable based on the detection results. For `exposure_energy`, it is normally distributed, but when used as the dependent variable to analyze the influence of `tool_id`, it encounters the problem of unequal variances. Assuming the penalty weight for unequal variances is 0.5, then the α1 for analyzing the influence of `tool_id` on `exposure_energy` is 1.0 × 0.5 = 0.5. For `cd_mean`, it is non-normal. Assuming the penalty weight for unequal variances is 0.4, then for any parameter test involving `cd_mean` as the dependent variable, its α1 is 1.0 × 0.4 = 0.4. Similarly, when analyzing the influence of `tool_id` on `final_yield`, due to the unequal variances, the α1 for this test path is also 0.5.

[0048] Therefore, based on the second-level data detection of the optimized data combination, the corresponding data discrete features are determined. The second adaptive test coefficient is determined according to multiple industrial data features and data discrete features. The corresponding adaptive test mode is determined based on the mapping relationship between the first adaptive test coefficient, the second adaptive test coefficient, and the adaptive test mode. This approach takes into account the overall consideration of the mapping relationship between the first adaptive test coefficient, the second adaptive test coefficient, and the adaptive test mode, ensuring the accuracy of the corresponding adaptive test mode. At the same time, a hybrid industrial data combination is introduced to further control the optimized data combination. This approach takes into account the overall consideration of multiple industrial data features, corresponding data distribution characteristics, and data discrete features, thereby improving the accuracy of the adaptive test mode.

[0049] At this point, it is crucial to focus on the robustness and reliability of the data, especially the balance of sample size distribution, as this directly affects the power of statistical tests and the generalizability of the conclusions. The core is to assess the sample distribution of each category of the categorical variable. The system calculates the sample size of each category and quantifies the balance by calculating the ratio of "minimum category sample size / maximum category sample size". The closer the ratio is to 1, the more balanced the categories are; the closer the ratio is to 0, the more severe the sample imbalance is. At the same time, the system also checks for categories with excessively small sample sizes (e.g., less than 30 samples), because small sample groups can lead to unstable variance estimation in statistical tests. These detection results are quantified into specific feature values. For example, the discrete characteristics of one variable are recorded as "highly balanced", while another is recorded as "severely imbalanced".

[0050] The qualitative "discrete characteristics of the data" are transformed into a quantitative "second adaptive test coefficient" (denoted as α2) to assess the robustness of the data in terms of sample distribution. α2 is also a score between 0 and 1, reflecting the robustness of the data to sample imbalance. The system sets a mapping function based on the discrete characteristics to calculate α2. For example: when the balance ratio is >0.8, α2=1.0; when the ratio is between 0.5 and 0.8, α2=0.8; when the ratio is between 0.2 and 0.5, α2=0.5; when the ratio is <0.2, α2=0.2. If there are small sample groups, a further penalty is applied. The closer α2 is to 1, the more ideal the sample distribution and the more reliable the test results. The lower α2 is, the more vigilant one should be about the bias caused by sample imbalance.

[0051] The system combines the α1 (reflecting the distribution hypothesis) and α2 (reflecting sample robustness) obtained in the first two steps and automatically selects the most appropriate statistical test method through a predefined, rule-based mapping engine. This mapping engine is a complex decision tree or rule set that encapsulates the knowledge of statistical experts. Its input is a pair of (α1, α2) coefficients, and its output is a specific test method. For example, the rules are set as follows: if both α1 and α2 are greater than 0.8, then a standard parametric test (such as ANOVA) is selected; if α1 or α2 is less than 0.4, then a nonparametric test (such as Kruskal-Wallis) is selected; if α1 is moderately low but α2 is good, then a robust parametric test (such as Welch ANOVA) is selected. The system determines a clear and adaptive test mode for each "independent variable-dependent variable" analysis pair.

[0052] Specifically, the system performs a second layer of detection on the optimized data combination. When detecting tool_id, it was found that ASML-05 had 400 wafers, ASML-06 had 350, and NIKON-02 had 250. The balance ratio was 250 / 400 = 0.625, and the sample size of all categories was much greater than 30. Therefore, the data dispersion characteristics of tool_id were determined to be "moderately imbalanced, with no small sample groups". When detecting recipe, it was found that recipe A had 800 wafers and recipe B had 200. The balance ratio was 200 / 800 = 0.25, and its data dispersion characteristics were determined to be "severely imbalanced".

[0053] The system calculates α2 based on the discrete characteristics; for tool_id, its discrete characteristic is "moderate imbalance", and according to the above rules, its second adaptive test coefficient α2 is determined to be 0.8; for recipe, its discrete characteristic is "severe imbalance", and its second adaptive test coefficient α2 is determined to be 0.5.

[0054] The system determines the final test mode based on the coefficient pairs. When analyzing the impact of tool_id on final_yield, α1=0.5 (unequal variances) is obtained from S132, and α2=0.8 (moderate imbalance) is obtained from this step. The system queries the mapping engine and matches the rule: α1 is moderately low, α2 is good. The final adaptive test mode is WelchANOVA because it is specifically designed to handle unequal variances. When analyzing the impact of recipe on final_yield, it is assumed that α1=0.9 (satisfies the distribution hypothesis), but α2=0.5 (severe imbalance). The system matches the rule: α2 is low. Although the distribution hypothesis is satisfied, the severe sample imbalance makes standard ANOVA unreliable. The final adaptive test mode is the Kruskal-Wallis test because it is insensitive to sample imbalance.

[0055] refer to Figure 5 In step S14, the specific steps are as follows: S141: Determine the corresponding root cause analysis system based on the matching of semiconductor database and multiple industrial data features; determine multiple key analysis nodes based on the root cause analysis system and multiple industrial data features; and determine the corresponding root cause analysis framework based on the multiple key analysis nodes and the optimized data combination among the multiple key analysis nodes. S142: Determine the first test effectiveness level based on the adaptive testing mode and multiple semiconductor data; determine the second test effectiveness level based on the root cause analysis framework and multiple semiconductor data; S143: Determine the inspection effect coefficient of each semiconductor data in different dimensions based on the first inspection effect level, the second inspection effect level, and the semiconductor processing steps corresponding to multiple semiconductor data.

[0056] In the embodiments of this application, a corresponding root cause analysis system is determined based on the matching of semiconductor database and multiple industrial data features. Multiple key analysis nodes are determined based on the root cause analysis system and multiple industrial data features. Among the multiple key analysis nodes, a corresponding root cause analysis framework is determined based on the multiple key analysis nodes and the optimized data combination. This approach takes into account the overall consideration of multiple key analysis nodes and the optimized data combination, ensuring the accuracy of the corresponding root cause analysis framework.

[0057] At this point, the system finds the most suitable "methodological template" or "analysis paradigm" for the current analysis task. Through matching, the system aligns the specific problem with a general solution framework. Internally, the system maintains a pre-built "root cause analysis system library," a collection of domain knowledge and best practices containing analysis templates for different industrial scenarios, such as "semiconductor yield improvement system," "semiconductor defect diagnosis system," or "equipment drift monitoring system." The matching mechanism employs rule-based or feature-based algorithms, with inputs being the "multiple industrial data features" and "target variables" determined in S131. The system analyzes the type of input features (e.g., if final_yield exists, the yield improvement system is prioritized) and performs strong matching in conjunction with the user-specified application scenario (e.g., "yield analysis"). It can even use natural language processing to understand the business meaning of variable names for more intelligent matching.

[0058] Guided by the selected analytical framework, the system decomposes the macroscopic analytical task into a series of specific and independent "key analytical nodes," each of which is a clearly defined statistical test proposition. The node generation logic is driven by rules defined within the system. The system traverses all input industrial data features (independent variables) and generates corresponding analytical nodes based on their type and relationship with the target variable. For categorical features (such as tool_id), a "between-group difference analysis" node is generated, with the proposition "Test whether there is a significant difference in the mean final_yield of different tool_id groups." For numerical features (such as exposure_energy), a "correlation analysis" node is generated, with the proposition "Test whether there is a significant linear or monotonic relationship between exposure_energy and final_yield." Each node is precisely defined, including the node ID, input variables, output variables, and suggested test method type.

[0059] The system integrates all independent analysis nodes into an organic, executable "root cause analysis framework." It organizes all key analysis nodes into a structured execution plan, which can be a simple list or a directed acyclic graph (DAG) containing dependencies between nodes. During this process, each node in the framework is bound to a specific data column in the "optimized data combination." For example, the independent variable of node N01 is bound to the `tool_id` column of the data combination, and the dependent variable is bound to the `final_yield` column. The system generates an optimized execution sequence for the framework, allowing independent nodes to be executed in parallel for efficiency, and nodes with dependencies to be executed in the order of dependency. The final output is a complete, self-contained set of analysis instructions that clearly tells the system what to analyze, what data to use, and in what order. This framework is the direct input for the next step, S142, execution verification.

[0060] Furthermore, a first test effectiveness level is determined based on the adaptive testing mode and multiple semiconductor data; a second test effectiveness level is determined based on the root cause analysis framework and multiple semiconductor data, which is compatible with the overall consideration of the root cause analysis framework and multiple semiconductor data, ensuring the accuracy of the second test effectiveness level.

[0061] At this point, the system will traverse each key analysis node in the root cause analysis framework constructed in S141. For each node, it will use the "adaptive test mode" determined by S13 to perform statistical calculations. For example, for the node analyzing the effect of tool_id on final_yield, if the adaptive mode is WelchANOVA, the system will perform the test. For the node analyzing the relationship between exposure_energy and final_yield, if the mode is Pearson correlation analysis, the system will calculate the correlation coefficient and p-value. After the test is completed, the system will map the key statistical outputs such as p-value to a qualitative "first test effect level" according to the preset threshold rules. For example, p < 0.01 is "strongly significant", 0.01 ≤ p < 0.05 is "moderately significant", and p ≥ 0.05 is "not significant". This level purely reflects the "signal strength" at the data level. A "strongly significant" level means that, from a statistical point of view, this variable is a very promising root cause.

[0062] By incorporating business logic and contextual information, the system assesses the "overall quality" of variables as root causes. It examines the results of individual nodes within the overall root cause analysis framework, evaluating their consistency and logical coherence. For example, it checks whether the test results of the variable form a reasonable causal chain with the results of other related variables in the framework. Simultaneously, based on a pre-defined business knowledge base, the system evaluates the variables from dimensions such as business importance (e.g., whether the variable is a core process parameter or an auxiliary parameter, and the scope of its impact) and operability (e.g., the ease and cost of adjusting the variable). The system then integrates these dimensions—consistency, importance, and operability—through a weighted scoring model or rule engine, ultimately outputting a qualitative "second test effect level," such as "high," "medium," or "low."

[0063] Therefore, the inspection effect coefficients of each semiconductor data in different dimensions are determined based on the first inspection effect level, the second inspection effect level, and the semiconductor processing steps corresponding to multiple semiconductor data. This approach takes into account the overall consideration of the first inspection effect level, the second inspection effect level, and the semiconductor processing steps corresponding to multiple semiconductor data, ensuring the accuracy of the inspection effect coefficients of each semiconductor data in different dimensions.

[0064] At this point, the two qualitative levels output by S142 ("strongly significant", "high", etc.) are converted into mathematically computable values ​​to prepare for subsequent weighted calculations. The system internally defines clear numerical mapping rules for each level, which are usually based on expert experience or statistical calibration. For example, for the first test effect level (statistical significance), "strongly significant" can be mapped to f1=0.9, "moderately significant" to f1=0.6, and "not significant" to f1=0.2. For the second test effect level (comprehensive business value), "high" can be mapped to f2=0.9, "moderate" to f2=0.6, and "low" to f2=0.3. After this step, each semiconductor data (such as tool_id, exposure_energy) obtains two quantitative scores, f1 and f2, which represent its performance in the statistical and business dimensions, respectively.

[0065] The system incorporates contextual information about the position of variables within the semiconductor manufacturing process, as variables in different processes have different weights influencing the final result. A built-in "semiconductor processing step weight library" assigns a basic weight to each critical step (such as photolithography, etching, and thin-film deposition) in the 5nm logic chip manufacturing process. The weights are set based on the "criticality" and "value-added" of the process. For example, photolithography, which directly defines the chip circuit pattern, typically has the highest weight, while routine processes like cleaning have relatively lower weights. The system binds each semiconductor data point to its corresponding processing step based on the variable name or metadata and obtains the weight value g for that step.

[0066] The system integrates the quantitative scores (f1, f2) and process weights (g) obtained from the first two steps using a weighted fusion model to output the final test effect coefficient. The system uses a multivariate weighted formula to calculate the core comprehensive dimension coefficient, for example: comprehensive coefficient = (w1×f1+w2×f2+w3×g) / (w1+w2+w3), where w1, w2, and w3 are configurable weights used to adjust the relative importance of statistical significance, business value, and process importance. In addition to the comprehensive coefficient, the system can also output specific coefficients (such as directly using f1 or f2) to provide ranking from different perspectives. The system generates a structured result containing multiple dimension coefficients for each semiconductor data point. These coefficients are standardized and can be directly used for ranking, filtering, and visualization, providing engineers with a clear and quantifiable list of root cause priorities.

[0067] refer to Figure 6 In step S15, the specific steps are as follows: S151: Collect data from each semiconductor, determine the data priority of each semiconductor data based on the preset inspection requirements of the corresponding data application scenarios and industrial data, determine the data inspection events of each semiconductor data based on the inspection effect coefficients of each semiconductor data in different dimensions and the data priority of each semiconductor data, and mark the corresponding data content. S152: Collect the data fusion system corresponding to each semiconductor data, determine multiple sub-data fusion items based on the identification of the data fusion system, and determine the corresponding data verification results according to the data verification events of each semiconductor data, the corresponding data verification content, and the multiple sub-data fusion items.

[0068] In the embodiments of this application, various semiconductor data are collected, and the data priority corresponding to each semiconductor data is determined based on the preset inspection requirements of the industrial data of the corresponding data application scenario. The data inspection events of each semiconductor data are determined based on the inspection effect coefficients of each semiconductor data in different dimensions and the data priority corresponding to each semiconductor data, and the corresponding data content is marked. This approach takes into account the overall consideration of the inspection effect coefficients of each semiconductor data in different dimensions and the data priority corresponding to each semiconductor data, ensuring the accuracy of the data inspection events of each semiconductor data.

[0069] At this point, each potential root cause (semiconductor data) is assigned a priority reflecting its business urgency and importance. This is a weighted calculation process that combines data insights and business rules. The system calls a preset rule base based on the current "data application scenario." This rule base is a digital representation of business knowledge, defining which types of variables need special attention in different scenarios (such as "new equipment online verification" or "mass production yield ramp-up") and assigning them corresponding preset weights. Subsequently, the system uses a weighted model to calculate the final data priority. This model integrates objective evidence from the data and subjective concerns from the business. Its calculation formula can be expressed as: Final data priority = (w_data × inspection effect coefficient) + (w_business × preset inspection requirement weight). Among them, w_data and w_business are configurable weights used to balance the relative importance of data insights and business rules.

[0070] High-priority variables are transformed into specific, structured "data validation events." Each event is a complete analysis story, not just a variable name. The system sets a priority threshold (e.g., 0.85). Only semiconductor data whose final data priority exceeds this threshold will be triggered to generate a data validation event. This effectively filters out secondary information and focuses on core issues. Each data validation event is a structured object containing multiple fields. It encapsulates all the key analytical conclusions and evidence about the variable, including event ID, event title (a concise conclusive description), associated variables, and priority labels (e.g., automatically mapped to "P0-Urgent" or "P1-High" based on the final priority value range).

[0071] The system populates the generated "data verification event" with detailed evidence and explanatory content to make it persuasive and readable. Information is extracted from S14 and the raw data and automatically populated into the various fields of the event structure. This includes: statistical summaries (such as verification methods, statistics, and p-values), quantification of business impact (such as "for every 1 mJ / cm² increase in exposure energy, the average wafer yield decreases by approximately 0.45 percentage points"), visual evidence (the system automatically calls the visualization engine to generate the most suitable chart for the event, such as a scatter plot or box plot), and data slicing (providing interactive slicing functionality for the raw data, allowing engineers to drill down to view the data distribution within a specific device or time range). All populated content is tagged with metadata and indexed for quick retrieval and correlation in the final fusion report.

[0072] Specifically, the current application scenario is "5nm logic chip mass production yield ramp-up". The system activates the corresponding rules, and the preset weight of the variables in the lithography process is 1.2. For exposure_energy, its S14 comprehensive coefficient is 0.90, which belongs to the lithography process. Assuming w_data=0.7 and w_business=0.3, its final data priority = (0.7×0.90)+(0.3×1.2)=0.99. For tool_id, its S14 comprehensive coefficient is 0.78, which also belongs to the lithography process. Its final data priority = (0.7×0.78)+(0.3×1.2)=0.906.

[0073] The priority of exposure_energy is 0.99 and the priority of tool_id is 0.906, both exceeding the threshold of 0.85, thus successfully triggering event generation. The system creates an event for exposure_energy with the title "[P0] Exposure energy has a strong and significant negative correlation with wafer yield"; and creates an event for tool_id with the title "[P1] Different lithography equipment has a significant impact on wafer yield".

[0074] In the P0 event of exposure_energy, the system populated all the Pearson-related details and automatically generated a clear scatter plot with a clear downward trend in the data points, along with a regression line equation. In the P1 event of tool_id, the system populated the WelchANOVA results and generated a box plot, which visually showed that the median yield of the NIKON-02 device was significantly lower than that of the other two devices. After being labeled, these events are now complete information packages containing titles, priorities, statistical evidence, business explanations, and visualizations, ready to be sent to S152 for final report fusion.

[0075] Furthermore, a data fusion system corresponding to each semiconductor data point is collected. Based on the identification of the data fusion system, multiple sub-data fusion projects are determined. The corresponding data verification results are determined according to the data verification events, corresponding data verification content, and multiple sub-data fusion projects of each semiconductor data point. This approach takes into account the overall consideration of the data verification events, corresponding data verification content, and multiple sub-data fusion projects of each semiconductor data point, ensuring the accuracy of the corresponding data verification results. At the same time, the verification effect coefficients of each semiconductor data point under different dimensions are introduced, realizing the overall consideration of the data verification events, corresponding data verification content, and data fusion system of each semiconductor data point, thereby improving the accuracy of the data verification results.

[0076] At this point, based on the purpose and audience of the analysis task, the most suitable report framework and presentation format are selected, as different scenarios require different report styles. The system maintains a pre-built "data fusion system library," where each system is a complete report template that defines the report's structure, layout, interaction methods, and level of detail. For example, there is a "senior executive dashboard system" designed for management, an "engineer in-depth analysis system" designed for process engineers, or a "compliance audit system" designed for quality audits. The system will automatically select the most matching fusion system from the library based on the "application scenario" or "user role" specified when initiating the analysis task, thereby ensuring that the report's presentation format is highly consistent with the user's needs.

[0077] Within the selected fusion framework, the report is broken down into several specific, populateable modules, known as "sub-data fusion projects." Each sub-project is a logical component of the report, corresponding to a chapter, panel, or chart area. These projects are diverse in type and function: summary projects (such as "Execution Summary") automatically extract the core conclusions of all high-priority events; ranking projects (such as "Root Cause Priority Ranking") generate a sortable table listing all events by priority; detail projects (such as "Detailed Event Analysis") provide an independent display space for each event; and metadata projects (such as "Data Quality Report") display statistical information from the data optimization phase to demonstrate the reliability of the analytical basis.

[0078] The system intelligently populates all the "data verification events" generated by S151 into each sub-data fusion project, generating a complete and interactive final report. The system automatically maps and populates event content into corresponding projects: extracting the core conclusions of P0 and P1 events in the "Execution Summary"; sorting all events by priority in the "Ranking Table"; and loading the complete content of each event, including charts and statistical summaries, in the "Details Panel." More advanced fusion systems also perform intelligent association and enhancement. For example, if the system detects a strong correlation between exposure_energy (P0 event) and tool_id (P1 event), it automatically generates a "Relationship Insight" module below the exposure_energy details page, prompting engineers to check the exposure energy settings of different devices. The system generates a unified, accessible URL or file. This report not only presents the findings but, more importantly, guides engineers from phenomena to essence, ultimately pinpointing actionable root causes through structured projects and intelligent associations.

[0079] Specifically, the current analysis task was initiated by a senior process engineer with the goal of deeply identifying the root cause of yield problems. The system identified the scenario and user role, collected and activated the "Engineer Deep Analysis System" from the fusion system library. The system template defines that the report must include modules such as "Execution Summary", "Root Cause Priority Ranking", "Detailed Event Analysis" and "Data Quality Report" to meet the engineer's needs for depth and technical details.

[0080] Under the "Engineer Deep Analysis System", the system identified the following sub-data fusion projects: Project A (Execution Summary), Project B (Root Cause Priority Ranking Table), Project C (Detailed Event Analysis Panel), and Project D (Data Quality and Cleaning Log). These projects together constitute a report structure that can provide both a macro overview and micro-drill-down.

[0081] The system generates a web report; opening the report reveals an "Execution Summary," with the P0 event of exposure_energy highlighted in red; switching to the "Root Cause Ranking" page, exposure_energy is ranked first; clicking on it displays a complete scatter plot and statistical summary in the "Detailed Analysis" panel on the right, along with insights into its association with tool_id at the bottom; based on this highly integrated report, engineers can immediately organize equipment engineers to jointly examine the exposure energy calibration procedures of different lithography equipment, thus quickly entering the problem-solving phase.

[0082] Please see Figure 7 , Figure 7 This is a schematic diagram of the structural composition of the industrial data testing device based on root cause analysis in an embodiment of the present invention; the industrial data testing device based on root cause analysis includes: The hybrid industrial data combination module 21 is used to collect industrial data sets in the semiconductor database, identify multiple semiconductor data based on the identification of industrial data sets, and determine the hybrid industrial data combination based on each semiconductor data and its corresponding data type. The optimized data combination module 22 is used to identify the corresponding missing values ​​and outliers based on the identification of hybrid industrial data combinations, and to determine the optimized data combination based on the missing values, outliers and the corresponding data optimization system; The adaptive verification module 23 is used to identify multiple industrial data features based on the optimized data combination; and to determine the corresponding adaptive verification mode according to the multiple industrial data features, the corresponding data distribution characteristics, and the data discrete characteristics. The test effect coefficient module 24 is used to collect the root cause analysis system, determine the corresponding root cause analysis framework based on the identification of the root cause analysis system, and determine the test effect coefficient of each semiconductor data in different dimensions according to the adaptive test mode, the root cause analysis framework and multiple semiconductor data. The data verification result module 25 is used to determine the data verification events of each semiconductor data according to the verification effect coefficients of each semiconductor data in different dimensions and the data priority of each semiconductor data, and to determine the corresponding data verification results based on the data verification events of each semiconductor data, the corresponding data verification content and the data fusion system.

[0083] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method of inspection of industrial data based on root cause analysis, characterized in that, The method comprises the following steps: Collecting an industrial data set in a semiconductor database, determining a plurality of semiconductor data according to the identification of the industrial data set, determining a mixed industrial data combination based on each semiconductor data and the corresponding data category; Determining the corresponding missing value and the abnormal value based on the identification of the mixed industrial data combination, and determining the optimized data combination according to the missing value, the abnormal value and the corresponding data optimization system; Determining a plurality of industrial data features based on the identification of the optimized data combination; Determining the corresponding adaptive test mode according to the plurality of industrial data features, the corresponding data distribution characteristics and the data discrete features; Collecting a root cause analysis system, determining the corresponding root cause analysis framework based on the identification of the root cause analysis system, and determining the test effect coefficient of each semiconductor data in different dimensions according to the adaptive test mode, the root cause analysis framework and the plurality of semiconductor data; Determining the data test event of each semiconductor data according to the test effect coefficient of each semiconductor data in different dimensions and the data priority corresponding to each semiconductor data, and determining the corresponding data test result based on the data test event of each semiconductor data, the corresponding data test content and the data fusion system.

2. The method of claim 1, wherein, The method comprises the following steps: Collecting a semiconductor database and monitoring the semiconductor database in real time, determining an industrial data space based on the detection of the semiconductor database, and determining an industrial data set according to the traversal of the industrial data space; Determining a plurality of semiconductor data according to the data content of each industrial data based on the detection of the industrial data set, determining the corresponding data category based on the identification of each semiconductor data in the plurality of semiconductor data, and determining a mixed industrial data combination according to each semiconductor data, the corresponding data relationship and the corresponding data category.

3. The method of claim 1, wherein, The method comprises the following steps: Collecting a mixed industrial data combination, determining numerical industrial data and non-numerical industrial data based on the detection of the mixed industrial data combination, and performing synchronous data processing on the numerical industrial data and the non-numerical industrial data; In the synchronous data processing of the numerical industrial data and the non-numerical industrial data, a plurality of independent variables are determined according to the numerical industrial data and the non-numerical industrial data, and the corresponding missing value is predicted based on the plurality of independent variables and the corresponding neural network model; Collecting a data matrix corresponding to the mixed industrial data combination, determining a corresponding low-rank matrix according to the data matrix, the numerical industrial data and the non-numerical industrial data, and determining the corresponding abnormal value based on the optimization of the low-rank matrix; and determining the optimized data combination based on the missing value, the abnormal value and the corresponding data optimization system.

4. The method of claim 1, wherein, Determine a plurality of industrial data features based on the optimized data combination recognition; Determine a corresponding adaptive test mode according to the plurality of industrial data features, corresponding data distribution characteristics and data dispersion characteristics, including: Determine a plurality of industrial data populations based on the optimized data combination detection, and collect target variables of the mixed industrial data combination. Determine a plurality of industrial data features according to each industrial data population, target variables and corresponding data application scenarios.

5. The method of claim 4, wherein, Determine a plurality of industrial data features based on the optimized data combination recognition; Determine a corresponding adaptive test mode according to the plurality of industrial data features, corresponding data distribution characteristics and data dispersion characteristics, including: Determine corresponding data distribution characteristics based on the first heavy data detection of the optimized data combination, which includes normality and variance homogeneity. Determine a first adaptive test coefficient according to the plurality of industrial data features and corresponding data distribution characteristics; Determine corresponding data dispersion characteristics based on the second heavy data detection of the optimized data combination. Determine a second adaptive test coefficient according to the plurality of industrial data features and data dispersion characteristics. Determine a corresponding adaptive test mode based on the mapping relationship of the first adaptive test coefficient, the second adaptive test coefficient and the adaptive test mode.

6. The method of claim 1, wherein, The collection of the root cause analysis system determines a corresponding root cause analysis framework based on the recognition of the root cause analysis system. Determine the test effect coefficient of each semiconductor data in different dimensions according to the adaptive test mode, the root cause analysis framework and the plurality of semiconductor data, including: Determine a corresponding root cause analysis system based on the matching of the semiconductor database and the plurality of industrial data features. Determine a plurality of key analysis nodes according to the root cause analysis system and the plurality of industrial data features. In the plurality of key analysis nodes, determine a corresponding root cause analysis framework according to the plurality of key analysis nodes and the optimized data combination.

7. The method of claim 6, wherein, The collection of the root cause analysis system determines a corresponding root cause analysis framework based on the recognition of the root cause analysis system. Determine the test effect coefficient of each semiconductor data in different dimensions according to the adaptive test mode, the root cause analysis framework and the plurality of semiconductor data, including: Determine a first test effect level according to the adaptive test mode and the plurality of semiconductor data. Determine a second test effect level according to the root cause analysis framework and the plurality of semiconductor data. Determine the test effect coefficient of each semiconductor data in different dimensions based on the first test effect level, the second test effect level and the semiconductor processing procedure corresponding to the plurality of semiconductor data.

8. The method of claim 1, wherein, Determine the data test event of each semiconductor data according to the test effect coefficient of each semiconductor data in different dimensions and the data priority corresponding to each semiconductor data. Determine the corresponding data test result based on the data test event of each semiconductor data, the corresponding data test content and the data fusion system, including: Collecting each semiconductor data, determining a data priority corresponding to each semiconductor data based on a preset inspection requirement of a corresponding data application scene industrial data of each semiconductor data, determining a data inspection event of each semiconductor data based on an inspection effect coefficient of each semiconductor data under different dimensions and the data priority corresponding to each semiconductor data, and marking corresponding data content.

9. The method of claim 8, wherein, The data inspection event of each semiconductor data is determined based on the inspection effect coefficient of each semiconductor data under different dimensions and the data priority corresponding to each semiconductor data, the corresponding data inspection result is determined based on the data inspection event of each semiconductor data, corresponding data inspection content and a data fusion system, and the method further comprises: Collecting the data fusion system corresponding to each semiconductor data, determining a plurality of sub-data fusion projects based on identification of the data fusion system, and determining the corresponding data inspection result according to the data inspection event of each semiconductor data, the corresponding data inspection content and the plurality of sub-data fusion projects.

10. An apparatus for verifying industrial data based on root cause analysis, characterized by, The industrial data inspection device based on root cause analysis is applied to the industrial data inspection method based on root cause analysis in any one of claims 1-9, and the industrial data inspection device based on root cause analysis comprises: A hybrid industrial data combination module is used to collect an industrial data set in a semiconductor database, determine a plurality of semiconductor data according to identification of the industrial data set, and determine a hybrid industrial data combination based on each semiconductor data and a corresponding data category; An optimized data combination module is used to determine corresponding missing values and abnormal values based on identification of the hybrid industrial data combination, and determine an optimized data combination according to the missing values, abnormal values and a corresponding data optimization system; An adaptive inspection module is used to determine a plurality of industrial data characteristics based on identification of the optimized data combination, and determine a corresponding adaptive inspection mode according to the plurality of industrial data characteristics, a corresponding data distribution characteristic and a data discrete feature; An inspection effect coefficient module is used to collect a root cause analysis system, determine a corresponding root cause analysis framework based on identification of the root cause analysis system, and determine an inspection effect coefficient of each semiconductor data under different dimensions according to the adaptive inspection mode, the root cause analysis framework and the plurality of semiconductor data; A data inspection result module is used to determine a data inspection event of each semiconductor data according to the inspection effect coefficient of each semiconductor data under different dimensions and the data priority corresponding to each semiconductor data, and determine a corresponding data inspection result based on the data inspection event of each semiconductor data, corresponding data inspection content and a data fusion system.

Images