Semiconductor etching equipment health management method and system based on knowledge graph

Abstract

The application discloses a knowledge graph-based semiconductor etching equipment health management method and system. The method comprises the following steps: preprocessing historical operation data of a semiconductor etching equipment to obtain window-level data; constructing a virtual knowledge graph based on the window-level data; extracting window-level fields from the virtual knowledge graph to obtain a main line training table; performing feature leakage control on the main line training table to obtain a data set; dividing the data set into a training set and a validation set, and dividing the training set into a first training subset and a second training subset; obtaining a fusion weight based on the validation set; obtaining a trained first regression model and a trained second regression model; obtaining real-time operation data of a to-be-tested semiconductor etching equipment, determining to-be-tested features of the real-time operation data, inputting the to-be-tested features into the first regression model and the second regression model respectively, performing weighted fusion on prediction values of the two regression models, and obtaining a remaining life prediction result of the to-be-tested semiconductor etching equipment.

Description

Technical Field

[0001] This application relates to the fields of PHM (Prognostics and Maintenance Management) technology for semiconductor ion beam etching equipment, data semantic modeling and predictive maintenance technology, and in particular to a knowledge graph-based method and system for health management of semiconductor etching equipment. Background Technology

[0002] Semiconductor wafer manufacturing places stringent demands on equipment stability, process consistency, and production continuity. Etching equipment is a core component of wafer manufacturing, and its operational status directly determines wafer processing quality, production line cycle time, equipment maintenance costs, and the overall stability of the production line. During operation, etching equipment continuously generates sensor timing data, process status data, equipment operation data, fault record data, and remaining lifetime tag data. This data comprehensively characterizes the equipment's state changes throughout the entire process, from normal operation and early anomalies to performance degradation and eventual failure. It is the core data source for achieving etching equipment health management and predictive maintenance.

[0003] Currently, predictive maintenance of semiconductor etching equipment generally relies on directly building predictive models from raw tabular data or preprocessed feature tables. While existing solutions can achieve basic fault detection, equipment status identification, and remaining lifespan prediction, they suffer from several technical limitations. First, there is no clear correspondence between raw data fields and equipment operational semantics, failing to characterize the complex relationships between multiple elements such as the equipment itself, time windows, sensor features, health status, and fault modes. Second, training datasets are often composed of concatenated multi-table data, making data field tracing difficult and resulting in poor interpretability of model input semantics. Third, etching equipment is characterized by complex operating conditions, a massive number of sensor dimensions, scarce fault samples, and phased state evolution; conventional tabular modeling methods cannot adapt to actual maintenance scenarios involving multiple devices, multiple states, and multiple fault modes, leading to insufficient model stability. Summary of the Invention

[0004] To address the challenges of expressing semantic relationships in data, tracing the source of training data, lacking semantic interpretation of model inputs, and struggling with risk warning under sparse abnormal sample conditions in the predictive maintenance of existing semiconductor etching equipment, this invention provides a knowledge graph-based method and system for semiconductor etching equipment health management. Instead of directly inputting semiconductor etching equipment operating data into the model, this invention constructs a PHM virtual knowledge graph centered on a TimeWindow, and connects the graph's semantic layer and the predictive model layer through a SPARQL mainline training table, thereby improving data traceability, model interpretation, and predictive maintenance application capabilities.

[0005] On the one hand, a knowledge graph-based method for health management of semiconductor etching equipment is provided, including: Acquire historical operating data of semiconductor etching equipment; preprocess the historical operating data to obtain window-level data; construct a virtual knowledge graph based on the window-level data; Window-level fields are extracted from the virtual knowledge graph to obtain the main training table; feature leakage control is applied to the main training table to obtain the dataset; the dataset is divided into a training set and a validation set, and the training set is further divided into a first training subset and a second training subset; fusion weights are obtained based on the validation set. The first regression model is trained based on the first training subset, and the second regression model is trained based on the second training subset to obtain the trained first regression model and second regression model. The real-time operating data of the semiconductor etching equipment under test is acquired, the test features of the real-time operating data are determined, the test features are input into the first regression model and the second regression model respectively, and the predicted values ​​of the two regression models are weighted and fused to obtain the remaining lifetime prediction result of the semiconductor etching equipment under test.

[0006] On the other hand, a knowledge graph-based health management system for semiconductor etching equipment is provided, including: The acquisition module is configured to: acquire historical operating data of a semiconductor etching device; preprocess the historical operating data to obtain window-level data; and construct a virtual knowledge graph based on the window-level data. The dataset construction module is configured to: extract window-level fields from the virtual knowledge graph to obtain the main training table; perform feature leakage control on the main training table to obtain the dataset; divide the dataset into a training set and a validation set, and divide the training set into a first training subset and a second training subset; and obtain the fusion weights based on the validation set. The model training module is configured to: train the first regression model based on the first training subset, and train the second regression model based on the second training subset, to obtain the trained first regression model and the second regression model. The prediction module is configured to: acquire real-time operating data of the semiconductor etching equipment under test, determine the features to be tested from the real-time operating data, input the features to be tested into a first regression model and a second regression model respectively, and perform weighted fusion of the predicted values ​​of the two regression models to obtain the remaining lifetime prediction result of the semiconductor etching equipment under test.

[0007] Furthermore, an electronic device is also provided, including: Memory, used for non-transitory storage of computer-readable instructions; and Processor, for executing the computer-readable instructions, When the computer-readable instructions are executed by the processor, they perform the method described in the first aspect above.

[0008] In another aspect, a storage medium is also provided for non-transitory storage of computer-readable instructions, wherein when the non-transitory computer-readable instructions are executed by a computer, the method described in the first aspect is performed.

[0009] In another aspect, a computer program product is also provided, including a computer program that, when run on one or more processors, is used to implement the method described in the first aspect above.

[0010] The above technical solution has the following advantages or beneficial effects: 1. This invention constructs a virtual knowledge graph (PHM) for semiconductor etching equipment based on TimeWindow, unifying equipment and sensor statistical characteristics, health status, symptom observation, fault modes, and RUL tags into semantic data assets, thus avoiding the problems of unclear field sources, ambiguous data semantics, and difficult-to-interpret model inputs in existing direct table modeling methods.

[0011] 2. This invention generates a SPARQL mainline training table through RDF / TTL semantic structures and SPARQL queries, enabling window-level fields in the training table to be traced back to objects, attributes, or relationships in a virtual knowledge graph, thus improving the queryability, traceability, and reusability of the training data. In this paper, window-level fields refer to data fields formed after statistical analysis, aggregation, or label alignment at the TimeWindow sample granularity. These differ from the original sensor sampling point fields; instead, they are a set of fields used to describe the device's operating status, sensor statistical characteristics, health status, symptom observations, fault information, and remaining lifespan information within a specific time window.

[0012] 3. This invention adopts a segmented fusion prediction mechanism based on RUL tag quantiles in the training set, models the low-RUL stage and the non-low-RUL stage respectively, and selects fusion weights through the validation set, avoiding target leakage caused by using real RUL segmentation information in the testing stage, thereby improving the deployability of the remaining lifetime prediction scheme.

[0013] 4. This invention sets up a feature leakage control mechanism before model training to exclude fields that may cause information leakage or semantic unfairness, such as RUL labels, health status, risk labels, fault fields, symptom semantic fields, time fields, window numbers, and device numbers. This ensures that the improvement in prediction performance comes from the segmented fusion modeling structure, rather than from the leakage of target information.

[0014] 5. This invention further establishes a multi-tool data normalization and label expansion mechanism. Based on window-level operational data from multiple semiconductor etching equipment tools, it unifies equipment fields, constructs sensor statistical features, constructs strict status labels, and expands risk area labels to generate a multi-tool extended verification table. This multi-tool extended verification table differs from the SPARQL mainline training table; it is not generated by SPARQL queries but by multi-tool data normalization, windowing processing, feature construction, and label expansion. This provides independent data support for strict equipment status identification auxiliary verification and multi-tool event-level auxiliary risk warning.

[0015] 6. This invention further includes a rigorous equipment status identification auxiliary verification module, which performs HEALTHY / NON-HEALTHY binary classification auxiliary verification based on a multi-tool extended verification table to verify the existence of equipment status distinguishing signals in the sensor statistical features. Because the abnormal windows of semiconductor etching equipment are sparse, this module is not used as a window-by-window precise fault diagnosis module, but rather as a pre-verification step for multi-tool event-level auxiliary risk warning.

[0016] 7. This invention further incorporates a multi-tool event-level auxiliary risk warning module. Through risk area labeling, category imbalance weighting, short-term historical maximum score smoothing, device-specific adaptive thresholds, single-window alarm triggering, and alarm fragment merging mechanisms, it expands window-level risk identification results into event-level anomaly coverage and early warning. Thus, this invention forms a complete technical closed loop encompassing semiconductor etching equipment operation data cleaning, windowed preprocessing and normalization, PHM virtual knowledge graph construction, RDF / TTL semantic structure expression, SPARQL mainline training table generation, q60 segmented fusion RUL model verification, rigorous equipment status identification auxiliary verification, and multi-tool event-level auxiliary risk warning. This provides a semantic data foundation and engineering model verification support for semiconductor etching equipment health management and predictive maintenance systems. Attached Figure Description

[0017] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.

[0018] Figure 1 This is a flowchart of the method in Example 1. Detailed Implementation

[0019] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0020] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of this application. The terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0021] Semiconductor wafer manufacturing places extremely high demands on equipment stability, process consistency, and production continuity. Etching equipment, as one of the key process devices in wafer manufacturing, directly impacts wafer processing quality, production cycle time, equipment maintenance costs, and the overall stability of the production line. During etching equipment operation, the system continuously generates a large amount of sensor timing data, process status data, equipment operation data, fault record data, and tag data related to remaining lifetime (RUL). This data reflects the equipment's state changes from normal operation, early anomalies, performance degradation, to the point before and after failure, providing a crucial data foundation for equipment health management and predictive maintenance.

[0022] Existing predictive maintenance methods for semiconductor equipment typically model directly based on raw tabular data or preprocessed feature tables. While these methods can accomplish tasks such as fault detection, condition identification, and remaining lifetime prediction to some extent, they still have significant shortcomings. First, there is a lack of clear mapping between raw fields and the semantics of equipment operation, making it difficult to express the complex relationships between equipment, time windows, sensor features, health status, symptom observation, fault modes, and remaining lifetime labels. Second, model training data is usually obtained by concatenating multiple tabular files, making it difficult to trace the source of fields and resulting in a lack of clear semantic interpretation of model inputs. Third, semiconductor etching equipment operates under complex conditions, has many sensor dimensions, relatively sparse fault samples, and phased state changes; simply relying on ordinary tabular modeling cannot stably support predictive maintenance applications in scenarios with multiple devices, multiple states, and multiple fault modes.

[0023] With the development of knowledge graphs, semantic modeling, and machine learning technologies, organizing equipment, time, status, sensors, symptoms, faults, and predictive labels into a queryable semantic structure has become an important direction for improving the health management capabilities of industrial equipment. Knowledge graphs can express the relationships between different objects, and machine learning models can learn the patterns of state changes and remaining life changes from data. The combination of the two is beneficial for improving the data organization capabilities, model interpretation capabilities, and engineering scalability of predictive maintenance systems.

[0024] However, existing technologies still lack a complete technical solution for predictive maintenance scenarios of semiconductor etching equipment, capable of simultaneously achieving windowed processing of raw operating data, construction of a predictive maintenance virtual knowledge graph, generation of a SPARQL mainline training table based on SPARQL queries, remaining lifetime prediction, and event-level risk warning. Therefore, researching a method and system that can semantically organize multi-source operating data of semiconductor etching equipment and further support remaining lifetime prediction and risk warning has significant engineering application value.

[0025] Example 1 This embodiment provides a knowledge graph-based method for health management of semiconductor etching equipment; like Figure 1 As shown, the knowledge graph-based health management method for semiconductor etching equipment includes: S101: Acquire historical operating data of the semiconductor etching equipment; preprocess the historical operating data to obtain window-level data; construct a virtual knowledge graph based on the window-level data; S102: Extract window-level fields from the virtual knowledge graph to obtain the main training table; perform feature leakage control on the main training table to obtain the dataset; divide the dataset into a training set and a validation set, and divide the training set into a first training subset and a second training subset; obtain the fusion weights based on the validation set; S103: Train the first regression model based on the first training subset, and train the second regression model based on the second training subset to obtain the trained first regression model and second regression model; S104: Obtain the real-time operating data of the semiconductor etching equipment under test, determine the characteristics to be tested from the real-time operating data, input the characteristics to be tested into the first regression model and the second regression model respectively, and weight and fuse the predicted values ​​of the two regression models to obtain the remaining lifetime prediction result of the semiconductor etching equipment under test.

[0026] Furthermore, the method also includes: S105: Constructing a multi-tool extended validation table: A multi-tool extended validation table is obtained through multi-tool data normalization, windowing processing, feature construction, and label expansion; S106: Based on the multi-tool extended verification table, perform auxiliary verification for device status identification; S107: Based on the aforementioned multi-tool extended verification table, perform multi-tool event-level assisted risk warning, including: S107-1: Construct risk area labels; S107-2: The short-term historical maximum score smoothing strategy is used to smooth the window-level scores to obtain the smoothed risk scores; S107-3: Based on the smoothed risk score, a device-specific adaptive threshold strategy is adopted to determine the corresponding device adaptive alarm threshold for different semiconductor etching devices; based on the device adaptive alarm threshold, a single-window alarm trigger determination is performed for each smoothed risk score; if the smoothed risk score in the current time window exceeds the corresponding device adaptive alarm threshold, a single-window alarm is triggered. S107-4: Merge alarm points that are adjacent in time, or merge alarm points whose time interval is less than the preset window number to obtain alarm segments; S107-5: Generate a risk warning result table based on single-window alarms and alarm segments.

[0027] Further, S101: acquiring historical operating data of the semiconductor etching equipment specifically includes: Acquire multivariable sensor timing data, process status data, fault record data, and remaining lifetime tag data generated by semiconductor etching equipment during wafer manufacturing; The multivariable sensor time-series data includes one or more of the following: pressure, voltage, current, gas flow rate, rotation angle, usage count, process stage, running recipe, recipe steps, and step duration.

[0028] The fault record data is used to indicate abnormal or fault events that occur during the operation of the equipment; The remaining lifespan tag data is used to represent the remaining time between the current device operating time window and the next failure.

[0029] Further, step S101: preprocessing the historical operation data to obtain window-level data, specifically including: The time-series data of multivariable sensors, process status data, fault record data, and remaining life tag data are cleaned, field normalized, sorted by time, windowed, statistically feature-constructed, and tag-aligned to obtain window-level fault prediction and health management (PHM) data with time window as the sample granularity.

[0030] Furthermore, the cleaning, field normalization, time sorting, windowing, statistical feature construction, and label alignment include: Standardize the naming conventions for device number, window number, window generation time, RUL label, health status, and sensor statistics fields; The window-level samples are sorted according to the order of device operation time. The sensor time series data is windowed, and the sensor statistical characteristics within the window are calculated. The sensor statistical characteristics include one or more of the following: mean, standard deviation, minimum value, maximum value, and last value. Dynamic derived features are constructed based on the statistical characteristics of the sensor. The dynamic derived features include one or more of the following: range of change, coefficient of variation, offset of the final value relative to the mean, position change, and amplitude ratio. Align the fault log data and remaining life tag data with the corresponding TimeWindow to form window-level PHM data.

[0031] The window-level fault prediction and health management data includes: device number, time window number, window generation time, sensor statistical characteristics, health status, symptom observation information, fault mode information, and RUL tag.

[0032] The sensor statistical characteristics include one or more of the following: mean, standard deviation, minimum, maximum and last values ​​of the sensor data within the window.

[0033] Dynamic derived features can also be constructed based on the statistical characteristics of the sensor. These dynamic derived features include one or more of the following: range of variation, coefficient of variation, offset of the final value relative to the mean, positional change, and amplitude ratio.

[0034] It should be understood that the TimeWindow is the core data granularity in this invention that connects raw time-series data, virtual knowledge graphs, and machine learning training tables. By converting raw time-series data into window-level samples, sensor statistical features, health status, symptom observations, fault modes, and RUL labels can be aligned within the same time window.

[0035] Further, step S101: Constructing a virtual knowledge graph based on the window-level data specifically includes: Using window-level fault prediction and health management (PHM) data as the core, a virtual knowledge graph of PHM for semiconductor etching equipment is constructed; The virtual knowledge graph includes: device objects, time window objects, sensor statistical feature objects, health status objects, symptom observation objects, fault mode objects, fault occurrence objects, and RUL information objects; wherein, the RUL information object, also known as the RULPrediction object, is used to uniformly carry the remaining lifetime information corresponding to the time window; Establish semantic relationships between the time window and device objects, sensor statistical feature objects, health status objects, symptom observation objects, fault mode objects, and RUL information objects, and express the virtual knowledge graph as an RDF / TTL semantic structure.

[0036] Furthermore, semiconductor etching equipment or etching tools are defined as equipment objects; Define the time window during device operation as a TimeWindow object; Define the sensor statistics within the window as a SensorFeature object; Define the device health stage corresponding to the window as a HealthState object; The remaining lifetime label or prediction result corresponding to the window is defined as a RUL information object, also known as a RULPrediction object; wherein, during the model training phase, the RUL information object is used to carry the RUL label corresponding to the TimeWindow; during the model prediction or deployment phase, the RUL information object is used to carry the RUL prediction result output by the q60 segmented fusion RUL model. Define the symptoms, symptom scores, symptom confidence levels, and number of rules triggered by the window as SymptomObservation objects; Define potential fault types or standardized fault clues as FaultMode objects; Define a fault occurrence record or fault event as a FaultOccurrence object.

[0037] Furthermore, the device object is used to refer to the semiconductor etching equipment or etching tool being monitored.

[0038] The time window object is used to represent a time window during the operation of the device and serves as the core sample object of the virtual knowledge graph.

[0039] The sensor statistical feature object is used to represent the statistical status of each sensor within the current time window.

[0040] The health status object is used to represent the device health stage corresponding to the current time window, and the health stage includes one or more of HEALTHY, WARNING, DEGRADED, and CRITICAL.

[0041] The symptom observation object is used to represent the symptoms triggered in the current time window, symptom score, symptom confidence level, and number of rules.

[0042] The fault mode object is used to represent potential fault types or standardized fault clues.

[0043] The RUL information object, also known as the RULPrediction object, represents the remaining lifetime information associated with the current TimeWindow. During model training, the RUL information object carries the RUL label corresponding to the TimeWindow; during model prediction or deployment, it carries the RUL prediction results output by the q60 segmented fusion RUL model. Therefore, the RUL information object is a unified remaining lifetime information carrier connecting the virtual knowledge graph semantic structure, the SPARQL mainline training table, and the RUL prediction result table.

[0044] Furthermore, the semantic relationships include: The time window between a device object and a time window object has a hasWindow relationship; The device between the time window object and the device object belongs to the Tool relationship; The health status between the time window object and the health status object has a hasHealthState relationship; The sensor features between the time window object and the sensor statistical feature object have a hasSensorFeature relationship; Symptom observations between the time window object and the symptom observation object have a hasSymptomObservation relationship; The prediction results between the time window object and the RUL information object have a hasPrediction relationship; The fault modes between the fault occurrence object and the fault mode object have a hasFaultMode relationship. The relationship between the fault-occurring object and the time window object regarding whether the fault occurred within the time window (occursInWindow); The `supportsFaultMode` relationship exists between the symptom observation object and the fault mode object.

[0045] It should be understood that, through the aforementioned semantic objects and semantic relationships, the present invention can transform the original table fields into a queryable, traceable, and interpretable graph semantic structure, enabling the model training fields to be traced back to specific graph objects or relationships.

[0046] Further, step S102: extracting window-level fields from the virtual knowledge graph to obtain the main training table, specifically includes: Based on SPARQL queries, window-level fields are extracted from the RDF / TTL semantic structure, and SPARQL mainline training tables are generated by aligning RUL labels, window generation time, health status, sensor statistical features, symptom aggregation information, and fault explanation information according to time windows.

[0047] Further, step S102: extracting window-level fields from the virtual knowledge graph to obtain the main training table, specifically includes: Based on RUL tag query, extract the RUL tags corresponding to each time window from the RULPrediction object; Based on the window generation time query, extract the window generation time corresponding to each time window; Based on the health status query, extract the health status, degradation status marker and critical risk status marker from the HealthState object; Based on sensor statistical feature query, window-level sensor statistical features are extracted from SensorFeature objects; Based on symptom aggregation queries, extract and aggregate the number of symptoms, symptom scores, symptom confidence, and number of rules from the SymptomObservation object; The query results are summarized according to the window number of the time window to generate the SPARQL mainline training table.

[0048] Furthermore, the SPARQL query includes RUL label query, window generation time query, health status query, sensor statistical feature query, symptom aggregation query, and fault explanation query.

[0049] RUL tag query is used to extract the RUL tag corresponding to each TimeWindow from the RULPrediction object during the training phase and generate the supervised learning target field in the SPARQL main training table; during the prediction phase, the RULPrediction object is used to carry the RUL prediction results output by the model. Window generation time query is used to extract the generation time corresponding to each time window to support time sorting and time backward backtesting; Health status query is used to extract the health status, degradation status marker and key risk status marker corresponding to each time window; Sensor statistical feature query is used to extract sensor statistical features corresponding to each time window; Symptom aggregation query is used to aggregate multiple symptom observations within a time window into a window-level symptom field; Fault explanation query is used to extract the correlation information between symptom observation and fault mode.

[0050] It should be understood that SPARQL in this invention is not a typical data display method, but rather a crucial bridge from the virtual knowledge graph to the machine learning training table. The SPARQL mainline training table generated through SPARQL queries ensures that training fields have a clear semantic origin, avoiding the problems of unclear data sources and ambiguous field semantics caused by direct table concatenation.

[0051] Further, S102: Performing feature leakage control on the mainline training table to obtain the dataset, specifically including: Feature leakage control is performed on the SPARQL mainline training table, and fields that may cause target leakage or semantic unfairness are deleted: RUL label, health status, risk label, fault field, symptom semantic field, time field, window number and device number, while retaining sensor statistical features and dynamically derived features.

[0052] The SPARQL mainline training table includes time window identifiers, device identifiers, time window generation time, RUL labels, health status, symptom aggregation information, fault explanation information, and sensor statistical features, which are used to support semantic tracing, sample alignment, time sorting, window-level interpretation, and model validation.

[0053] During the training of the RUL prediction model, feature leakage control is performed on the SPARQL mainline training table to exclude fields that may lead to target leakage or semantic unfairness. The excluded fields include one or more of the following: RUL label, health status field, risk label field, fault field, symptom semantic field, time field, window number field, and device number field.

[0054] The sensor's statistical characteristics and their dynamically derived characteristics are retained as inputs to the remaining lifetime prediction model. These dynamically derived characteristics include one or more of the following: range of variation, coefficient of variation, offset of final value relative to the mean, location change, and amplitude ratio.

[0055] Therefore, the baseline RUL model and the q60 segmented fusion RUL model use the same type of input feature set under the same feature selection rules, so that the difference in their prediction performance mainly comes from the difference in modeling structure, rather than the difference in input fields or the leakage of target information.

[0056] The baseline RUL model is implemented using the XGBoost regression algorithm. As a single unified regression model, it directly learns the mapping relationship between the time window and the RUL label based on all training samples. The q60 segmented fusion RUL model is a segmented fusion prediction model, which includes a first regression model, a second regression model, and a fusion weighted output mechanism. The first regression model is trained on samples with low remaining lifetime, and the second regression model is trained on samples with non-low remaining lifetime. Both the first and second regression models are implemented using the XGBoost regression algorithm. Therefore, the baseline RUL model and the q60 segmented fusion RUL model use the same type of basic regression algorithm. The main difference lies in the fact that the q60 segmented fusion RUL model further introduces a sample segmentation training based on RUL quantiles and a fusion weighted prediction mechanism.

[0057] Further, S102: dividing the dataset into a training set and a validation set, and dividing the training set into a first training subset and a second training subset, specifically includes: Based on the quantile threshold of the RUL labels in the training set, the training samples are divided into a first training subset (low-RUL samples) and a second training subset (non-low-RUL samples). The first training subset corresponds to the low-RUL regression model, and the second training subset corresponds to the non-low-RUL regression model.

[0058] The quantile threshold for dividing the training samples is the q quantile of the RUL label in the training set. Training samples with RUL labels less than or equal to the quantile threshold are classified as low-RUL samples; training samples with RUL labels greater than the quantile threshold are classified as non-low-RUL samples; wherein, quantile q60 represents the 0.60 quantile of the RUL label in the training set.

[0059]

[0060] in, Represents the RUL labels of the training set quantile threshold, Indicates the first training set RUL labels for each time window, express Quantitation function. Preferably, That is, q60.

[0061] Low remaining lifetime samples and non-low remaining lifetime samples can be defined as:

[0062]

[0063] The low-RUL regression model and the non-low-RUL regression model were trained based on the first training subset (low-RUL samples) and the second training subset (non-low-RUL samples).

[0064] Furthermore, the low-RUL regression model and the non-low-RUL regression model can be implemented using a gradient boosting regression model.

[0065] It should be understood that semiconductor etching equipment may exhibit different degradation patterns at different stages of remaining lifetime. A single uniform regression model is insufficient to fully represent the differences between low remaining lifetime stages and non-low remaining lifetime stages. Therefore, this embodiment divides different lifetime stages by remaining lifetime quantiles and trains corresponding regression models for each stage.

[0066] Further, S102: obtaining the fusion weights based on the validation set specifically includes: A candidate fusion weight set A is pre-defined. The candidate fusion weight set A is a finite set obtained by discretizing within the interval [0,1] according to a preset step size. It is used to represent different weight combinations of the prediction results of the first regression model and the prediction results of the second regression model in the final fusion prediction.

[0067] The candidate fusion weight set A is represented as:

[0068] Where M represents the number of candidate weight divisions.

[0069] Preferably, M can be 100, in which case the candidate fusion weight set A includes candidate values ​​such as 0.00, 0.01, 0.02, ..., 0.99, 1.00.

[0070] For each candidate fusion weight in the candidate fusion weight set A compute the corresponding value on the validation set. And select the fusion weights with the lowest validation set error that satisfy robust constraints. .in, This represents the fusion weights of the prediction results from the first regression model. This represents the fusion weights of the prediction results from the second regression model.

[0071] It should be noted that the fusion weights The model parameters are not obtained through gradient descent training, therefore they do not involve random initialization in the sense of model training. As a neutral fusion weight or fixed weight control value, it is used to represent that the prediction results of the first regression model and the second regression model each account for half; however, the fusion weights used in the end are selected from the candidate fusion weight set A on the validation set. .

[0072] It can be represented as:

[0073]

[0074]

[0075] in, Indicates the first The true remaining lifetime of each sample Indicates the first Predicted remaining lifespan for each sample Indicates the sample size. (Using...) Error assessment can reduce the impact of large remaining lifetime numerical scales on assessment results and more stably reflect the relative deviation between predicted and actual values.

[0076] Fusion weights The choice is represented as:

[0077] in, This represents the set of candidate fusion weights after robust constraints. This indicates that within the candidate set, the validation set is... The lowest fusion weight.

[0078] During the testing or deployment phase, both low-RUL regression model predictions and non-low-RUL regression model predictions are obtained simultaneously for the same time window, and the final remaining lifetime prediction is generated:

[0079] in, Indicates the first The final remaining lifetime prediction for each time window. This represents the predicted value output by the low-RUL regression model. This represents the predicted value output by the non-low-RUL regression model. This represents the fusion weights determined based on the validation set, and 0 ≤ ≤ 1.

[0080] During the testing or deployment phase, the system does not use the actual remaining lifetime interval information of the current time window for hard switching, thereby avoiding target leakage caused by using the actual test segment information.

[0081] It should be understood that during the testing or deployment phase, the system cannot know in advance whether the current time window truly belongs to the low-RUL interval or the non-low-RUL interval. Therefore, this invention does not use the actual remaining lifetime interval information for hard switching, but instead adopts a fusion approach to utilize the prediction results of two regression models simultaneously, thereby avoiding target leakage caused by using actual segmentation information during the testing phase.

[0082] A remaining lifetime prediction result table is generated based on the final remaining lifetime prediction value. The remaining lifetime prediction result table includes: time window identifier, equipment identifier, remaining lifetime prediction value, and output fields related to the prediction result, which are used to support subsequent health status display, remaining lifetime prediction result display, and maintenance decision support.

[0083] Further, step S103: Training the first regression model based on the first training subset and training the second regression model based on the second training subset to obtain the trained first and second regression models, specifically includes: The features of the first training subset are used as the input values ​​of the first regression model, and the remaining lifetime of each sample in the first training subset is used as the output value of the first regression model. The parameters of the first regression model are adjusted. When the loss function value of the first regression model no longer decreases, or when the number of iterations exceeds the set number, training is stopped, and the trained first regression model is obtained. The features of the second training subset are used as the input values ​​of the second regression model, and the remaining lifetime of each sample in the second training subset is used as the output value of the second regression model. The parameters of the second regression model are adjusted. When the loss function value of the second regression model no longer decreases, or when the number of iterations exceeds the set number, training is stopped, and the trained second regression model is obtained.

[0084] Furthermore, S104: Acquire real-time operating data of the semiconductor etching equipment under test, specifically including: acquiring multivariable sensor timing data, process status data, and fault record data generated by the semiconductor etching equipment during wafer manufacturing.

[0085] Further, S104: Determining the features to be measured in the real-time running data specifically includes: Multivariate sensor time-series data, process status data, and fault record data are cleaned, field normalized, sorted by time, windowed, and statistically feature-constructed to obtain window-level fault prediction and health management (PHM) data with a time window as the sample granularity. The sensor statistical features include one or more of the following: mean, standard deviation, minimum, maximum, and last value of the sensor data within the window. Based on the sensor statistical features, dynamic derived features are constructed, which include one or more of the following: range of variation, coefficient of variation, offset of the last value relative to the mean, positional change, and amplitude ratio.

[0086] Further, in step S104: the features to be tested are input into the first regression model and the second regression model respectively, and the predicted values ​​of the two regression models are weighted and fused to obtain the health status prediction result of the semiconductor etching equipment under test.

[0087] Further, in S105: Constructing a multi-tool extended verification table: A multi-tool extended verification table is obtained through multi-tool data normalization, windowing processing, feature construction, and label expansion; In this context, "multiple tools" refers to multiple semiconductor etching equipment tools, i.e., multiple semiconductor etching devices. The "tools" in this application are not software tools, but rather equipment tools in a semiconductor manufacturing scenario, specifically semiconductor ion beam etching equipment with different device numbers. Each equipment tool corresponds to an independent operating time series, sensor records, fault records, and remaining lifetime information.

[0088] Multi-tool data normalization refers to organizing window-level operational data generated by multiple semiconductor etching equipment tools into data with a unified field structure, unified naming rules, unified time order, and unified sample granularity. Specifically, this includes: unified equipment identifier field, unified time window identifier field, unified window generation time field, unified sensor statistical feature field, unified health status field, unified fault marker field, unified fault name field, and unified RUL label field, and sorting the window-level samples of each equipment tool according to the equipment operation time order.

[0089] Windowing refers to dividing raw multivariable sensor time-series data into consecutive time windows, with each time window serving as a window-level sample. After windowing, each time window corresponds to a set of sensor statistical characteristics, health status, fault information, and remaining lifetime information within that window. Through windowing, the raw sensor data acquired point-in-time is converted into window-level PHM data with time window granularity.

[0090] Feature construction refers to the calculation of window-level feature fields describing the operating status of equipment based on the raw sensor data within each time window. These window-level feature fields include sensor statistical features and dynamically derived features. Sensor statistical features include the mean, standard deviation, minimum, maximum, and final value of the sensor data within the window; dynamically derived features include the range of variation, coefficient of variation, offset of the final value relative to the mean, positional variation, and amplitude ratio. Through feature construction, the raw multivariate sensor time-series data is transformed into window-level input features that can be directly used by state recognition and risk warning models.

[0091] Label extension refers to constructing strict equipment status labels and risk area labels based on existing health status, fault records, and RUL labels. Strict equipment status labels are used to mark time windows as HEALTHY or NON-HEALTHY, where HEALTHY indicates a normal window and NON-HEALTHY indicates an abnormal window. Risk area labels are used to identify risk windows near actual abnormal events. Specifically, the actual abnormal window itself, along with its five preceding and five following time windows, are marked as risk areas, while the remaining time windows are marked as non-risk areas.

[0092] The multi-tool extended verification table includes: device identifier field, time window identifier field, window generation time field, sensor statistical feature field, dynamic derived feature field, health status field, strict device status label field, fault mark field, fault name field, RUL label field, and risk area label field.

[0093] Based on window-level PHM data from multiple semiconductor etching equipment tools, a multi-tool extended verification table is generated by unifying equipment fields, unifying sensor statistical features, constructing strict equipment status labels, and expanding risk region labels. This multi-tool extended verification table differs from the SPARQL mainline training table; it is not directly generated from the RDF / TTL semantic structure by the SPARQL query module, but rather generated through multi-tool data normalization, windowing processing, feature construction, and label expansion. The SPARQL mainline training table primarily supports the verification of the q60 segmented fusion RUL model; the multi-tool extended verification table primarily supports strict equipment status identification auxiliary verification and multi-tool event-level auxiliary risk warning.

[0094] Further, step S106: Based on the multi-tool extended verification table, perform auxiliary verification for device status identification, specifically including: Based on the multi-tool extended verification table, a binary classification auxiliary verification of health / unhealth is performed to verify that there is a distinguishing signal between the health status and the unhealthy status of the device in the sensor statistical features.

[0095] S106 differs from the aforementioned q60 segmented fusion RUL model. The q60 segmented fusion RUL model is used to predict the remaining lifetime value corresponding to a time window, which is a regression prediction task. This step, however, is used to determine whether the device status corresponding to the time window is healthy or unhealthy, which is a binary classification auxiliary verification task. S106 does not output RUL prediction values ​​and is not used as the main prediction model of this invention, nor as a window-by-window precise fault diagnosis model. The purpose of this step is to verify that, after excluding the label field, fault field, RUL field, device identification field, and time field, the difference signal between healthy and unhealthy states can still be identified based solely on sensor statistical characteristics, thereby providing a basis for subsequent event-level auxiliary risk warning.

[0096] The device status identification-assisted verification includes the following steps: S106-1: Construction Equipment Status Label: Based on the health status field in the multi-tool extended verification table, a device status label is constructed for each time window. Time windows with a health status of HEALTHY are marked as healthy, and time windows with a health status other than HEALTHY are marked as unhealthy.

[0097] Device status label The definition is as follows:

[0098] in, Indicates the first Health status within a time window Indicates health category, This indicates a non-healthy category.

[0099] S106-2: Determine the input and exclusion fields for device status identification auxiliary verification: In the device status identification assisted verification, the following fields are excluded from the multi-tool extended verification table: health status label, strict status label, risk label, fault mark, fault name, remaining life label, device identifier, time window identifier, window generation time field, and other time fields.

[0100] After excluding the above fields, only sensor statistical features and dynamic derived features are retained as input fields. Among them, sensor statistical features include the mean, standard deviation, minimum, maximum and last values ​​of sensor data within the window; dynamic derived features include the range of variation, coefficient of variation, offset of the last value relative to the mean, positional change and amplitude ratio.

[0101] By using the above field filtering rules, the device status identification auxiliary verification does not use the health status itself, strict status labels, fault results, risk labels, remaining life labels, device number, time window number, or time order fields as model input, thus avoiding leakage of target information and semantic unfairness.

[0102] S106-3: Constructing a sample set for device status identification: Each time window in the multi-tool expanded validation table is treated as a sample. The model input for each time window is denoted as... Its strict equipment status label is recorded as .in, It consists of the sensor statistical features and dynamically derived features corresponding to this time window. The label is divided into two categories: healthy and unhealthy.

[0103] Therefore, the device status identification auxiliary verification sample set is represented as:

[0104] in, This represents the sample set for auxiliary verification of device status identification. Indicates the number of samples in the time window. Indicates the first Sensor feature vectors for each time window, Indicates the first Health / unhealthy status labels for each time window.

[0105] S106-4: Divide the dataset into training, validation, and test sets: The time window samples in the multi-tool extended validation table are sorted according to device identifier and window generation time, and then divided into training, validation, and test sets in chronological order. The training set is used to train the device status recognition auxiliary model, the validation set is used to determine the classification threshold, and the test set is used to evaluate the device status recognition auxiliary validation results. This division method preserves the chronological order of device operation and prevents future window information from entering the training process.

[0106] In this embodiment, the multi-tool extended verification table includes 133,420 time window samples, of which 132,850 are healthy windows and 570 are unhealthy windows, with unhealthy windows accounting for 0.43%. This sample distribution indicates that the number of unhealthy windows is far less than that of healthy windows, and device status recognition assisted verification belongs to a binary classification task with extreme class imbalance.

[0107] S106-5: Training the auxiliary model for device status recognition: A device status recognition auxiliary model is trained based on sensor statistical features and dynamically derived features from the training set. This model employs a gradient boosting classification approach. During training, unhealthy samples are assigned higher class weights than healthy samples, thereby improving the model's ability to identify unhealthy windows.

[0108] The model input is the first Sensor feature vectors within a time window The model output is the probability score of the time window belonging to the unhealthy category. , means as follows:

[0109] in, Indicates the first The predicted probability that a given time window belongs to the unhealthy category. Indicates the first Sensor statistical characteristics and dynamic derived characteristics within a time window.

[0110] S106-6: Determine the classification threshold based on the validation set.

[0111] Set the candidate set for classification threshold And calculate the balanced accuracy corresponding to different classification thresholds on the validation set. Classification threshold Determine using the following formula:

[0112] in, Represents the candidate set for classification thresholds. This indicates that a classification threshold is applied on the validation set. The balance accuracy obtained at that time This represents the classification threshold that maximizes the accuracy of balancing the validation set.

[0113] Based on the classification threshold Device status identification results in the test set The definition is as follows:

[0114] in, Indicates the first The predicted probability that a given time window belongs to the unhealthy category. This represents the classification threshold determined based on the validation set. This indicates that the condition is identified as unhealthy. This indicates that the status is recognized as healthy.

[0115] S106-7: Evaluate the results of auxiliary verification for equipment status identification: Calculate the confusion matrix, accuracy, balanced accuracy, non-healthy class precision, non-healthy class recall, non-healthy class F1 score, and macro-average F1 score on the test set.

[0116] The confusion matrix includes: the number of windows that are actually healthy and predicted to be healthy, the number of windows that are actually healthy and predicted to be unhealthy, the number of windows that are actually unhealthy and predicted to be healthy, and the number of windows that are actually unhealthy and predicted to be unhealthy.

[0117] Among them, the non-healthy recall rate measures the proportion of real non-healthy windows that are identified; the non-healthy precision rate measures the proportion of real non-healthy windows among the samples identified as non-healthy windows; the balanced precision rate measures the ability to identify both healthy and non-healthy classes; and the macro-average F1 score measures the overall recognition performance of healthy and non-healthy classes under class imbalance conditions.

[0118] When the balanced accuracy on the test set is higher than 0.5 and the recall rate for the unhealthy class is greater than 0, it is determined that there is a distinguishing signal between healthy and unhealthy states in the sensor statistical features. This determination condition indicates that, without using health status labels, strict status labels, risk labels, fault fields, remaining life labels, device identifiers, time window identifiers, and time fields as input, the model can still identify some unhealthy windows based on sensor statistical features.

[0119] In this embodiment, the confusion matrix results for the test set are as follows: the number of windows that are actually healthy and predicted to be healthy is 26519, the number of windows that are actually healthy and predicted to be unhealthy is 94, the number of windows that are actually unhealthy and predicted to be healthy is 62, and the number of windows that are actually unhealthy and predicted to be unhealthy is 10. There are a total of 72 truly unhealthy windows in the test set, of which 10 were identified as unhealthy. This result indicates that even when unhealthy windows are extremely sparse, there is still a distinguishing signal between healthy and unhealthy states in the sensor's statistical features.

[0120] The technical significance of this step is as follows: First, it verifies that the sensor statistical features are not invalid noise, but contain information on changes in the health status of the equipment; second, it demonstrates that in a strict window-by-window health / unhealthy identification task, due to the extreme sparseness of the unhealthy windows, relying solely on window-by-window binary classification is insufficient to form stable and accurate fault diagnosis results; third, it provides a basis for subsequent event-level auxiliary risk warning, namely, under the condition that the equipment status differentiation signal exists but the window-by-window abnormal samples are sparse, further risk area label expansion, short-term historical maximum score smoothing, device-specific adaptive thresholds, and alarm segment merging mechanisms are adopted to expand the window-by-window status differentiation signal into event-level abnormal coverage and early warning capabilities.

[0121] Further, in S107: risk area labels are constructed, and risk warning result tables and event-level evaluation result tables are generated by combining a gradient boosting classification model with class imbalance weights, a short-term historical maximum score smoothing strategy, a device-specific adaptive threshold strategy, a single-window alarm triggering mechanism, and an alarm fragment merging mechanism.

[0122] Further, S107-1: Constructing risk area labels specifically includes: The risk region label consists of the risk region labels for the first N time windows and the last N time windows, corresponding to the identifier preN_postN. That is, the actual anomaly window itself, along with its N preceding and N following time windows, are defined as the risk region. The risk region label is used to uniformly mark the actual anomaly window and its adjacent windows as risk regions, and serves as the supervised learning target label for the subsequent multi-tool event-level assisted risk warning model. The risk region label can be represented as:

[0123] in, Indicates the first Does the current time window fall within a risk zone? Indicates the first Whether a time window is a true anomaly window, where N represents the number of preceding and following expansion windows. This indicates that the risk area extension is limited to the same semiconductor etching equipment. Preferably, N = 5, i.e., using the pre5_post5 risk area label.

[0124] The risk area label is constructed by defining the actual anomaly window and a preset number of time windows before and after it as the risk area. In event-level evaluation, "horizon" represents the event-level evaluation range of a continuous time window, not a fixed second, minute, hour, or other fixed physical time length.

[0125] Further, S107-2: Training a window-level risk warning model and smoothing the window-level risk score using a short-term historical maximum score smoothing strategy to obtain a smoothed risk score, specifically including: Based on the sensor statistical features and dynamically derived features in the multi-tool extended validation table, as well as the risk area labels, a window-level risk warning model is trained. This window-level risk warning model is a gradient boosting classification model used to determine whether each time window belongs to a risk area and outputs the probability score of each time window belonging to a risk area.

[0126] During the training process of the window-level risk warning model, the first The sensor statistical characteristics and dynamic derived characteristics of each time window are denoted as follows: , will the The risk zone label for each time window is denoted as .in, Indicates the first This time window falls within a risk zone. Indicates the first The specified time window does not belong to the risk area. Therefore, the training samples for the window-level risk warning model are represented as follows: ; in, This represents the training sample set for the window-level risk warning model. Indicates the number of samples in the time window. Indicates the first Sensor feature vectors for each time window, Indicates the first Risk zone labels for each time window.

[0127] In the model input field selection, only sensor statistical features and dynamically derived features are used as model input fields; risk area labels. Labels are used only as supervised learning target labels and not as input fields to the model. Furthermore, health status labels, strict equipment status labels, fault markers, fault names, remaining lifespan labels, equipment identifiers, time window identifiers, window generation time fields, and other time fields are excluded to avoid leakage of target information and semantic unfairness. Specifically, equipment identifiers are only used for subsequent determination of equipment-specific thresholds and grouping of time windows within the same equipment, and are not used as input features for the window-level risk warning model.

[0128] Because the number of risk area windows is less than the number of non-risk area windows, the window-level risk warning model uses class imbalance weights during training. Let the number of non-risk area samples in the training set be... The number of samples in the risk area is The sample weights for non-risk areas are then... The sample weights for risk areas are:

[0129] in, This represents the class weights of samples in the risk area. This represents the class weights of samples in non-risk areas. By increasing the class weights of samples in risk areas, the model's ability to identify risk area windows is improved during training.

[0130] The window-level risk warning model outputs the first... The probability score of a time window belonging to a risk area , means as follows:

[0131] in, The output of the window-level risk warning model represents the first... Window-level risk score for each time window. Indicates the first Sensor statistical characteristics and dynamic derived characteristics within a time window Indicates the first This time window falls within the risk zone.

[0132] After obtaining the window-level risk score Then, a short-term historical maximum score smoothing strategy is used to smooth the window-level risk score, resulting in a smoothed risk score. The aforementioned short-term historical maximum score smoothing strategy involves taking the maximum window-level risk score of the current time window and its preceding short-term historical time windows within the same semiconductor etching equipment, and using this maximum as the smoothed risk score for the current time window.

[0133] Let the first The set of short-term historical windows for each time window is:

[0134] in, Indicates the first Each time window corresponds to a set of short-term historical windows for the same device. Indicates the first The semiconductor etching equipment corresponding to each time window Indicates the length of the short-term historical window.

[0135] The smoothed risk score is defined as:

[0136] in, Indicates the first Risk score after smoothing over a time window Represents the set of short-term historical windows The Middle Window-level risk score for each time window.

[0137] In this embodiment, the short-term historical window length (k=3) is the maximum value of the window-level risk score of the current time window and its two preceding time windows within the same semiconductor etching equipment. If the number of available historical windows preceding the current time window is less than two, then the maximum value of the risk score of the existing current window and its preceding windows within the same equipment is taken.

[0138] The short-term historical maximum score smoothing strategy is used to retain the highest risk signal within a short-term continuous window, reducing the impact of risk score fluctuations within a single time window on alarm triggering results. The smoothed result... In subsequent S107-3, it serves as the input for the device adaptive threshold determination, used to determine whether a single-window alarm is triggered in the corresponding time window.

[0139] Further, S107-3: Based on the smoothed risk score, an adaptive threshold strategy for each device is used to determine the alarm threshold, specifically including: A device-specific adaptive threshold strategy is adopted to determine a corresponding adaptive alarm threshold for each semiconductor etching device. Since the sensor distribution, operating status, risk score range, and abnormal event distribution vary among different semiconductor etching devices, this step does not use a uniform alarm threshold. Instead, it searches for the alarm threshold that maximizes the event-level evaluation score for each semiconductor etching device on the validation set.

[0140] Set candidate threshold set The candidate threshold set It is in the interval The finite set obtained by discretizing the inner element according to a preset step size is represented as follows:

[0141] in, Represents the set of candidate thresholds. Indicates the first One candidate threshold, This indicates the number of threshold divisions. In this paper, we take... Then the candidate threshold set include .

[0142] For the Taiwan semiconductor etching equipment sequentially traverses the candidate threshold set on the validation set. Each candidate threshold in For those belonging to the first The first semiconductor etching equipment A time window, based on the smoothed risk score. and candidate threshold Generate window-level alarm results for the candidate threshold. , means as follows:

[0143] in, Indicates the first The first semiconductor etching equipment A time window at the candidate threshold The alarm is triggered below; Indicates the first The first semiconductor etching equipment A time window at the candidate threshold No alarm is triggered. Indicates the first Risk score after smoothing over a time window; Indicates the first The semiconductor etching equipment number corresponding to each time window; This indicates that the time window belongs to the first... Taiwan semiconductor etching equipment.

[0144] At each candidate threshold Next, the first The continuous alarm windows in the semiconductor etching equipment verification set are merged into alarm segments, and these alarm segments are matched with actual abnormal events. If an alarm segment is located within a preset event-level evaluation range before the start of an actual abnormal event, or overlaps with a window near the occurrence of an actual abnormal event, then the alarm segment is determined to match the actual abnormal event. The event-level evaluation range is represented by "horizon," which represents the number of continuous time windows and does not represent a fixed second, minute, hour, or other fixed physical time length.

[0145] For the Taiwan semiconductor etching equipment and candidate threshold Define event-level statistics: Indicates the first Taiwanese semiconductor etching equipment at candidate threshold The number of alarm segments generated below; Indicates the first Taiwanese semiconductor etching equipment at candidate threshold The number of alarm segments that can match real abnormal events; Indicates the first The number of real anomalies in the semiconductor etching equipment verification set; Indicates the first Taiwanese semiconductor etching equipment at candidate threshold The number of actual abnormal events detected by the alarm segment.

[0146] No. Taiwanese semiconductor etching equipment at candidate threshold The event-level precision is defined as follows:

[0147] when season:

[0148] No. Taiwanese semiconductor etching equipment at candidate threshold The event-level recall rate is defined as follows:

[0149] when season:

[0150] No. Taiwanese semiconductor etching equipment at candidate threshold The event-level evaluation score is defined as the event-level F1 score, expressed as follows:

[0151] when season ; in, Indicates the first Taiwanese semiconductor etching equipment at candidate threshold The event-level evaluation score is calculated below. This score considers both the ability of the alarm segment to cover real abnormal events and the effectiveness of the alarm segment.

[0152] No. Adaptive alarm thresholds for semiconductor etching equipment Defined as:

[0153] in, Indicates the first The adaptive alarm threshold determined by the semiconductor etching equipment on the verification set. Represents the set of candidate thresholds. Indicates the first Taiwanese semiconductor etching equipment at candidate threshold The event-level evaluation score is as follows.

[0154] If there are multiple candidate thresholds, If the same highest value is obtained, the candidate threshold with the largest value is selected as the highest value. This is to reduce the number of false alarms and improve the stability of alarm triggering.

[0155] The adaptive alarm threshold Used in subsequent single-window alarm triggering steps. For the first... The time window, if its device is the first For semiconductor etching equipment, this time window uses the first... The adaptive alarm threshold corresponding to each device Perform alarm determination. Single-window alarm trigger result. It is expressed as follows:

[0156] in, Indicates the first An alarm is triggered within a specific time window. Indicates the first No alarm is triggered within a certain time window; Indicates the first Risk score after smoothing over a time window; Indicates the first The adaptive alarm thresholds for the semiconductor etching equipment belonging to each time window.

[0157] By employing the aforementioned device-specific adaptive threshold strategy, each semiconductor etching device uses the alarm threshold determined on its own validation set to determine alarm triggering, thereby reducing the impact of differences in risk score distribution among different devices on alarm results and providing window-level alarm results for subsequent alarm segment merging and event-level evaluation.

[0158] Further, in S107-4: alarm points that are adjacent in time are merged, or alarm points with a time interval less than a preset window number are merged to obtain an alarm segment.

[0159] Furthermore, S107-5: Based on single-window alarms and alarm segments, a risk warning result table is generated, specifically including: Based on the event-balanced early warning strategy, the low false alarm priority strategy, or the recall priority strategy, a risk early warning result table and an event-level assessment result table are generated.

[0160]

[0161]

[0162] in, Indicates the actual number of abnormal events. This represents the number of actual abnormal events detected. This represents the number of actual abnormal events that enabled early warning. Indicates the number of alarm events. This represents the number of alarm events that can match real abnormal events; This represents the event-level recall rate, which measures the proportion of real-world abnormal events that are covered by system alerts. This refers to the early warning recall rate, which measures the proportion of real abnormal events that have been covered by the system's early warning before the event occurs. This represents the event-level accuracy, which measures the proportion of alarm events generated by the system that match real abnormal events. This represents the event-level F1 score, used to comprehensively measure event-level recall and event-level precision.

[0163] It should be understood that event-level assisted risk warning is not a window-by-window precise anomaly diagnosis, but rather, under the condition of sparse anomaly windows, it expands the window-by-window identification results into event-level anomaly coverage and early warning through risk area expansion, time smoothing, device-specific adaptive thresholds, and alarm segment merging.

[0164] In this embodiment, publicly available semiconductor ion beam etching equipment PHM data can be used for instantiation. This data includes multivariable sensor timing data, fault log data, and time-to-failure information corresponding to multiple semiconductor etching equipment tools.

[0165] Furthermore, the method also includes: S108: Based on time window identifiers, back-search the equipment objects, sensor statistical feature objects, health status objects, symptom observation objects, fault mode objects, fault occurrence objects, and remaining life information objects in the PHM virtual knowledge graph of semiconductor etching equipment to support health status display, remaining life prediction result display, risk warning display, query traceability, and maintenance decision output. Output the main training table, the multi-tool extended validation table, the prediction result table, the risk warning result table, and the event-level evaluation result table.

[0166] A q60 segmented fusion RUL model was constructed based on the SPARQL mainline training table. Experimental results show that this model can reduce the RUL prediction error compared to the baseline RUL model. Rigorous equipment state recognition-assisted verification and event-level assisted risk warning experiments were conducted based on a multi-tool extended verification table, demonstrating that the proposed virtual knowledge graph semantic foundation can not only support RUL prediction but also be extended to state recognition-assisted verification and event-level risk warning scenarios.

[0167] This invention acquires multivariate sensor time-series data, process status data, fault record data, and remaining lifetime tag data during the operation of semiconductor etching equipment. The data is cleaned, windowed, statistically feature-constructed, and tag-aligned to obtain window-level PHM data with time windows as the sample granularity. A virtual knowledge graph of the semiconductor etching equipment PHM is constructed with the time window as the core and expressed as an RDF / TTL semantic structure. A SPARQL mainline training table is generated based on SPARQL queries. Feature leakage control is applied to the training table, and low-RUL regression models and non-low-RUL regression models are trained based on the q60 quantile threshold of the RUL tags in the training set, where q60 represents the 0.60 quantile of the RUL tags in the training set. The fusion weights are determined through the validation set, and an RUL prediction result table is generated. This invention improves the queryability, traceability, and interpretability of semiconductor etching equipment operation data and supports remaining lifetime prediction and event-level risk warning. Furthermore, based on multi-tool data normalization, windowing processing, feature construction, and label expansion, a multi-tool extended verification table is generated to conduct rigorous equipment status identification auxiliary verification and multi-tool event-level auxiliary risk warning. Thus, without changing the SPARQL mainline training table supporting RUL main model positioning, the system's status differentiation signal verification and event-level risk warning in multi-tool scenarios are extended.

[0168] Example 2 This embodiment provides a knowledge graph-based health management system for semiconductor etching equipment, including: The acquisition module is configured to: acquire historical operating data of a semiconductor etching device; preprocess the historical operating data to obtain window-level data; and construct a virtual knowledge graph based on the window-level data. The dataset construction module is configured to: extract window-level fields from the virtual knowledge graph to obtain the main training table; perform feature leakage control on the main training table to obtain the dataset; divide the dataset into a training set and a validation set, and divide the training set into a first training subset and a second training subset; and obtain the fusion weights based on the validation set. The model training module is configured to: train the first regression model based on the first training subset, and train the second regression model based on the second training subset, to obtain the trained first regression model and the second regression model. The prediction module is configured to: acquire real-time operating data of the semiconductor etching equipment under test, determine the features to be tested from the real-time operating data, input the features to be tested into a first regression model and a second regression model respectively, and perform weighted fusion of the predicted values ​​of the two regression models to obtain the remaining lifetime prediction result of the semiconductor etching equipment under test.

[0169] It should be noted that the acquisition module, dataset construction module, model training module, and prediction module described above correspond to steps S101 to S104 in Embodiment 1. The examples and application scenarios implemented by these modules and their corresponding steps are the same, but they are not limited to the content disclosed in Embodiment 1. It should also be noted that these modules, as part of the system, can be executed in a computer system, such as a set of computer-executable instructions.

[0170] Example 3 This embodiment also provides an electronic device, including: one or more processors, one or more memories, and one or more computer programs; wherein, the processor is connected to the memory, and the one or more computer programs are stored in the memory. When the electronic device is running, the processor executes the one or more computer programs stored in the memory to cause the electronic device to perform the method described in Embodiment 1.

[0171] Example 4 This embodiment also provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the method described in Embodiment 1.

[0172] Example 5 This embodiment also provides a computer program product, including a computer program that, when executed by a processor, implements the method in Embodiment 1.

[0173] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A knowledge graph-based health management method for semiconductor etching equipment, characterized by: include: Acquire historical operating data of semiconductor etching equipment; The historical operation data is preprocessed to obtain window-level data; Based on the aforementioned window-level data, a virtual knowledge graph is constructed; Window-level fields are extracted from the virtual knowledge graph to obtain the main training table; Feature leakage control is applied to the main training table to obtain the dataset; The dataset is divided into a training set and a validation set, and the training set is further divided into a first training subset and a second training subset. The fusion weights are obtained based on the validation set. The first regression model is trained based on the first training subset, and the second regression model is trained based on the second training subset to obtain the trained first regression model and second regression model. The real-time operating data of the semiconductor etching equipment under test is acquired, the test features of the real-time operating data are determined, the test features are input into the first regression model and the second regression model respectively, and the predicted values ​​of the two regression models are weighted and fused to obtain the remaining lifetime prediction result of the semiconductor etching equipment under test.

2. The knowledge graph-based semiconductor etching equipment health management method as described in claim 1, characterized in that, The method further includes: constructing a multi-tool extended verification table: obtaining a multi-tool extended verification table through multi-tool data normalization, windowing processing, feature construction, and label expansion; performing auxiliary verification for device status identification based on the multi-tool extended verification table; performing multi-tool event-level auxiliary risk warning based on the multi-tool extended verification table, including: constructing risk area labels; using a short-term historical maximum score smoothing strategy to smooth window-level scores to obtain smoothed risk scores; using a device-specific adaptive threshold strategy based on the smoothed risk scores to determine corresponding device adaptive alarm thresholds for different semiconductor etching devices; performing single-window alarm trigger determination for each smoothed risk score based on the device adaptive alarm thresholds; triggering a single-window alarm if the smoothed risk score in the current time window exceeds the corresponding device adaptive alarm threshold; merging temporally adjacent alarm points, or merging alarm points with time intervals less than a preset number of windows to obtain alarm segments; generating a risk warning result table based on single-window alarms and alarm segments.

3. The knowledge graph-based semiconductor etching equipment health management method as described in claim 1, characterized in that, Based on the window-level data, a virtual knowledge graph is constructed, which includes: device objects, time window objects, sensor statistical feature objects, health status objects, symptom observation objects, fault mode objects, fault occurrence objects, and RUL information objects; wherein, the RUL information object, also known as the RULPrediction object, is used to uniformly carry the remaining lifetime information corresponding to the time window; Establish semantic relationships between the time window and device objects, sensor statistical feature objects, health status objects, symptom observation objects, fault mode objects, and RUL information objects, and express the virtual knowledge graph as an RDF / TTL semantic structure; Define semiconductor etching equipment or etching tools as device objects; define the time window during device operation as a TimeWindow object; define the sensor statistical status within the window as a SensorFeature object; define the device health stage corresponding to the window as a HealthState object; define the remaining lifetime label or prediction result corresponding to the window as a RUL information object; define the symptoms, symptom scores, symptom confidence, and number of rules triggered by the window as SymptomObservation objects; define potential fault types or standardized fault clues as FaultMode objects; and define fault occurrence records or fault events as FaultOccurrence objects.

4. The knowledge graph-based semiconductor etching equipment health management method as described in claim 3, characterized in that, The semantic relationships include: a hasWindow relationship between a device object and a time window object; a belongsToTool relationship between a time window object and a device object; a hasHealthState relationship between a time window object and a health state object; a hasSensorFeature relationship between a time window object and a sensor statistical feature object; a hasSymptomObservation relationship between a time window object and a symptom observation object; a hasPrediction relationship between a time window object and a RUL information object; a hasFaultMode relationship between a fault occurrence object and a fault mode object; an occurrence-in-Window relationship between a fault occurrence object and a time window object; and supportsFaultMode relationships between a symptom observation object and a fault mode object.

5. The knowledge graph-based semiconductor etching equipment health management method as described in claim 1, characterized in that, The main training table is obtained by extracting window-level fields from the virtual knowledge graph. Specifically, this includes: querying based on RUL tags, extracting the RUL tags corresponding to each time window from the RULPrediction object; querying based on window generation time, extracting the window generation time corresponding to each time window; querying based on health status, extracting health status, deterioration status markers, and key risk status markers from the HealthState object; querying based on sensor statistical features, extracting window-level sensor statistical features from the SensorFeature object; and querying based on symptom aggregation, extracting and aggregating the number of symptoms, symptom scores, symptom confidence, and number of rules from the SymptomObservation object. The query results are then summarized according to the window number of the time window to generate the SPARQL main training table.

6. The knowledge graph-based semiconductor etching equipment health management method as described in claim 1, characterized in that, Extracting window-level fields from the virtual knowledge graph to obtain the main training table specifically includes: extracting window-level fields from the RDF / TTL semantic structure based on SPARQL query, and generating the SPARQL main training table by aligning RUL tags, window generation time, health status, sensor statistical features, symptom aggregation information and fault explanation information according to time window; The SPARQL queries include RUL label query, window generation time query, health status query, sensor statistical feature query, symptom aggregation query, and fault explanation query. Specifically, the RUL label query extracts the RUL label corresponding to each TimeWindow from the RULPrediction object during the training phase and generates the supervised learning target field in the SPARQL mainline training table; during the prediction phase, the RULPrediction object carries the RUL prediction results output by the model. The window generation time query extracts the generation time corresponding to each time window to support time sorting and time-based backtesting. The health status query extracts the health status, degradation status markers, and critical risk status markers corresponding to each time window. The sensor statistical feature query extracts the sensor statistical features corresponding to each time window. The symptom aggregation query aggregates multiple symptom observations within a time window into a window-level symptom field. The fault explanation query extracts the correlation information between symptom observations and fault modes. or, Feature leakage control is performed on the mainline training table to obtain the dataset. Specifically, this includes: performing feature leakage control on the SPARQL mainline training table, deleting fields that may cause target leakage or semantic unfairness: RUL label, health status, risk label, fault field, symptom semantic field, time field, window number and device number, while retaining sensor statistical features and dynamically derived features.

7. The knowledge graph-based semiconductor etching equipment health management method as described in claim 1, characterized in that, The fusion weights are obtained based on the validation set, specifically including: A candidate fusion weight set A is pre-defined. The candidate fusion weight set A is a finite set obtained by discretizing within the interval [0,1] according to a preset step size. It is used to represent different weight combinations of the prediction results of the first regression model and the prediction results of the second regression model in the final fusion prediction. The candidate fusion weight set A is represented as: Where M represents the number of candidate weight partitions; for each candidate fusion weight in the candidate fusion weight set A compute the corresponding value on the validation set. And select the fusion weights with the lowest validation set error that satisfy robust constraints. ; Represented as: in, Indicates the first The true remaining lifetime of each sample Indicates the first Predicted remaining lifespan for each sample Indicates the number of samples; Fusion weights The choice is represented as: ； in, This represents the set of candidate fusion weights after robust constraints. This indicates that within the candidate set, the validation set is... The lowest fusion weight; During the testing or deployment phase, both low-RUL regression model predictions and non-low-RUL regression model predictions are obtained simultaneously for the same time window, and the final remaining lifetime prediction is generated: ； in, Indicates the first The final remaining lifetime prediction for each time window. This represents the predicted value output by the low-RUL regression model. This represents the predicted value output by the non-low-RUL regression model. This represents the fusion weights determined based on the validation set, and 0 ≤ ≤ 1.

8. A knowledge graph-based health management system for semiconductor etching equipment, characterized in that, include: The acquisition module is configured to: acquire historical operating data of the semiconductor etching equipment; preprocess the historical operating data to obtain window-level data; Based on the aforementioned window-level data, a virtual knowledge graph is constructed; The dataset construction module is configured to extract window-level fields from the virtual knowledge graph to obtain the main training table; Feature leakage control is applied to the main training table to obtain the dataset; The dataset is divided into a training set and a validation set, and the training set is further divided into a first training subset and a second training subset. The fusion weights are obtained based on the validation set. The model training module is configured to: train the first regression model based on the first training subset, and train the second regression model based on the second training subset, to obtain the trained first regression model and the second regression model. The prediction module is configured to: acquire real-time operating data of the semiconductor etching equipment under test, determine the features to be tested from the real-time operating data, input the features to be tested into a first regression model and a second regression model respectively, and perform weighted fusion of the predicted values ​​of the two regression models to obtain the remaining lifetime prediction result of the semiconductor etching equipment under test.

9. An electronic device, characterized in that it comprises: Memory is used to store computer-readable instructions in a non-transitory manner. as well as Processor, for executing the computer-readable instructions, When the computer-readable instructions are executed by the processor, they perform the method described in any one of claims 1-7.

10. A storage medium, characterized in that, Non-transitory storage of computer-readable instructions, wherein when the non-transitory computer-readable instructions are executed by a computer, the method of any one of claims 1-7 is performed.

Images