A Method for Constructing a Rotating Target Life Prediction Model Based on Incremental Learning

By using incremental learning and Monte Carlo analysis, a multi-source fusion dataset is constructed and the model parameters are dynamically updated. This solves the problems of model statics and uncertainties in target life prediction, and achieves high-precision and adaptive target life prediction, which is applicable to ceramic targets such as IZO and ITO.

CN122838882BActive Publication Date: 2026-11-10东方电气长三角(杭州)创新研究院有限公司 +1
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611343069.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-09-01
Publication Date
2026-11-10
Estimated Expiration
2046-09-01

AI Technical Summary

Technical Problem

Existing methods for predicting the lifespan of target materials suffer from static models, insufficient data utilization, and a lack of physical interpretability, resulting in low prediction accuracy and inadequate uncertainty assessment, making it difficult to meet the requirements of industrial production.

Method used

An incremental learning-based approach is used to construct a multi-source fusion dataset. An initial model is trained using feature engineering and random forest regression algorithms. Uncertainty propagation calculations are performed using Monte Carlo analysis to dynamically update model parameters and output probabilistic lifetime prediction results containing confidence intervals.

Benefits of technology

It achieves high-precision, adaptive target life prediction, adapts to changes in process conditions, provides reliable confidence interval assessment, and improves prediction accuracy and model generalization performance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122838882B_ABST
    Figure CN122838882B_ABST
Patent Text Reader

Abstract

This invention discloses a method for constructing a rotating target life prediction model based on incremental learning, comprising: constructing a multi-dimensional data acquisition system to obtain electrical parameters, process parameters, target state parameters, and time series parameters; forming a standardized multi-source fusion dataset through missing value imputation, outlier identification, and derived feature construction; training an initial prediction model using a random forest regression algorithm; comprehensively evaluating the model using scatter plots of predicted and actual values, residual distribution, feature importance, learning curves, and radar charts; designing an incremental data collection strategy to dynamically update the target consumption rate parameter based on a robust weighted average method and automatically iterating the model version; quantifying measurement uncertainty, model structure uncertainty, and process uncertainty through Monte Carlo analysis, and outputting probabilistic prediction results including confidence intervals. This invention achieves a deep integration of physical constraints and data-driven approaches, significantly improving prediction accuracy, adaptability, and reliability.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and in particular to a method for constructing a rotating target lifetime prediction model based on incremental learning. Background Technology

[0002] In magnetron sputtering, the rotating target is a core consumable component, and its lifespan directly impacts production efficiency and product quality. Ceramic rotating targets, such as IZO (indium zinc oxide) and ITO (indium tin oxide), are widely used in high-end manufacturing fields such as flat panel displays and thin-film solar cells. Accurately predicting the remaining lifespan of the target is crucial for enabling preventative maintenance, reducing production costs, and ensuring process stability and product quality consistency.

[0003] Existing methods for predicting the lifetime of target materials mainly face the following technical challenges: First, there is the problem of model staticization. Traditional prediction methods are mainly based on simple empirical formulas or time-based statistical models, assuming that the sputtering material is consumed uniformly under constant power. However, in actual production, process parameters such as sputtering power, working gas pressure, and gas flow rate are frequently adjusted, and the consumption rate of the sputtering material at different radial and axial positions exhibits significant non-uniformity. Once the trained model is fixed, it cannot absorb data generated by new production batches, causing its predictive performance to drop sharply with process drift or changes in equipment status.

[0004] Second, data utilization is insufficient. Existing methods lack effective fusion processing of multi-source heterogeneous data. The magnetron sputtering process involves various types of data, including electrical parameters, process parameters, target state parameters, and timing parameters. These data have different sampling frequencies, different physical dimensions, and varying noise levels, and directly using them for model training will introduce a large amount of interference. Traditional methods often ignore the timing characteristics of the data and the coupling relationship between process parameters, resulting in prediction accuracy that is difficult to meet the requirements of industrial production.

[0005] Third, there is a lack of physical interpretability and uncertainty assessment. While purely data-driven black-box models can fit certain statistical patterns, their predictions lack physical interpretability, making them difficult for process engineers to trust. More importantly, existing methods cannot quantitatively assess the reliability of their predictions, and predictions lacking uncertainty intervals cannot provide a reliable basis for high-risk decisions such as changing target materials.

[0006] In recent years, physical information neural networks have demonstrated the potential to integrate physical laws with data-driven approaches in fields such as fluid mechanics and solid mechanics. However, their application in the prediction of industrial equipment lifespan remains unexplored. How to effectively embed the physical mechanisms of target material consumption into a deep learning framework and achieve end-to-end, interpretable, and highly efficient prediction of remaining lifespan from multi-source heterogeneous data remains an unsolved technical challenge in this field. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for constructing a rotating target life prediction model based on incremental learning.

[0008] The objective of this invention is achieved through the following technical solution: a method for constructing a rotating target life prediction model based on incremental learning, comprising the following steps: Step 1: Obtain raw production data containing electrical parameters, process parameters, target material state parameters, and timing parameters; perform data quality checks and preprocessing on the raw production data to construct a standardized multi-source fusion dataset; Step 2: Based on the standardized multi-source fusion dataset, construct a feature engineering system containing basic features and derived features for training the initial lifetime prediction model; the derived features include power-pressure ratio and efficiency index. Step 3: Continuously acquire production data during the new production process; based on this production data, dynamically update the parameters of the initial lifetime prediction model and automatically iterate the model version; Step 4: Identify the sources of uncertainty in the model prediction process; perform uncertainty propagation calculations on the model prediction results using Monte Carlo analysis, and finally output probabilistic lifetime prediction results containing confidence interval information.

[0009] Furthermore, step one is implemented through the following sub-steps: (1.1) Multi-dimensional data acquisition: The acquired parameters include at least the sputtering power, cumulative energy consumption, voltage and current in electrical parameters; working gas pressure, argon flow rate, oxygen ratio and temperature in process parameters; initial thickness, remaining thickness and thin film performance indicators in target material status parameters; and acquisition timestamp and equipment operating status in time parameters.

[0010] (1.2) Data quality checks: including missing value detection and imputation, outlier identification and correction based on physical rationality, data integrity verification and numerical range verification.

[0011] The specific methods for checking physical rationality are as follows: when the thickness reduction is less than 0 or the cumulative energy consumption is less than 0, it is judged as a physical abnormality record and removed; when the power density is less than 0.5 kW / cm² or greater than 10 kW / cm², it is judged as sensor fault data and marked; when the energy density (cumulative energy consumption / thickness reduction) exceeds the normal range determined by the interquartile range method, it is judged as a statistical anomaly and corrected.

[0012] (1.3) Data preprocessing and feature fusion: Time alignment of multi-frequency data (using time-weighted averaging method), Z-score standardization of data, and construction of derived features, including power-pressure ratio (power density / pressure) and efficiency index (thickness reduction / cumulative energy consumption).

[0013] Furthermore, the specific method for checking the physical rationality of the original production data in step (1.2) includes: when the thickness reduction is less than 0 or the cumulative energy consumption is less than 0, it is judged as a physical abnormality record and removed; when the power density is not within the preset range (less than 0.5 kW / cm² or greater than 10 kW / cm²), it is judged as sensor fault data and marked; when the energy density exceeds the normal range determined by the interquartile range method, it is judged as a statistical anomaly and corrected.

[0014] Furthermore, step two is implemented through the following sub-steps: (2.1) Constructing a feature engineering system: Determine the basic feature set as [power density, pressure, argon flow rate, oxygen ratio, temperature, cumulative energy consumption], and combine it with derived features; (2.2) Model training: The random forest regression algorithm is used, and parameters that support incremental learning are set ( The model is trained using the aforementioned feature engineering system as input and the reduction in target material thickness as the prediction target. During training, 5-fold cross-validation is used to evaluate model stability, and the optimal training sample size is determined through learning curve analysis.

[0015] (2.3) Comprehensive Model Evaluation: The mean absolute error (MAE), root mean square error (RMSE), and R² score are calculated using the test set, and a comprehensive evaluation chart is generated. The comprehensive evaluation chart contains six sub-charts: (a) scatter plot of predicted values ​​versus actual values; (b) histogram of residual distribution; (c) bar chart of feature importance; (d) box plot of error analysis; (e) learning curve; and (f) performance radar chart.

[0016] Furthermore, the specific process of generating the performance radar chart in step (2.3) includes: calculating the normalized score of each dimension based on five dimensions: R² score, prediction accuracy, model stability, generalization ability, and feature diversity; distributing the five scores evenly in the polar coordinate system according to angles, and connecting the points to form a closed polygon; calculating the ratio of the polygon area to the area of ​​the regular pentagon as the comprehensive score, and displaying it as a percentage at the center of the radar chart.

[0017] Furthermore, step three is implemented through the following sub-steps: (3.1) Incremental data collection: Set data validity verification rules, including a minimum threshold for thickness consumption (0.1 mm) and a reasonable range for consumption rate (50-500 kWh / mm), and only include newly produced data that has passed verification in the incremental learning process. At the same time, add data timestamp markers.

[0018] (3.2) Dynamic update of model parameters: The target consumption rate is updated using a robust weighted average method. The formula for calculating the new consumption rate is: Among them, the weight of new data Old data weights .

[0019] Furthermore, to improve robustness, a dynamic weight correction mechanism based on prediction residuals is introduced. This mechanism works as follows: First, the current model is used to predict each newly collected valid production data, calculating the prediction residual for each sample (actual thickness reduction minus predicted thickness reduction). Then, a weight function is designed based on the residual size, such that samples with larger residuals have lower weights. Finally, when calculating a robust estimate of the new data consumption rate, a weighted median is used instead of the ordinary median; that is, after sorting by consumption rate values, the rate value corresponding to the first time the cumulative weight exceeds half of the total weight is found. This method can effectively suppress the interference of measurement anomalies or transient process fluctuations on parameter updates.

[0020] (3.3) Model version management: After each incremental learning, the model version number is automatically iterated (e.g., from 1.0.0 to 1.1.0), the update time and update content summary are recorded, and the 20 most recent incremental data are retained for subsequent optimization. At the same time, a model backup mechanism is created.

[0021] Furthermore, step four is implemented through the following sub-steps: (4.1) Identification of sources of uncertainty: Identify three main sources of uncertainty: measurement uncertainty (power sensor error ±0.5%, thickness measurement error ±0.05 mm), model structure uncertainty, and process uncertainty (fluctuation of process parameters, changes in equipment status).

[0022] (4.2) Monte Carlo Uncertainty Analysis: Set the number of Monte Carlo simulations N=1000. Independently sample each uncertainty factor according to its probability distribution; substitute the N sets of sampling parameters into the trained prediction model to obtain N predicted remaining lifetime values; calculate the mean, standard deviation, and 2.5% and 97.5% quantiles of these N values ​​as the 95% confidence interval; output the probability density distribution plot and cumulative distribution function plot. If the confidence interval width exceeds the preset threshold (20% of the predicted mean), the system triggers a data supplementation warning. After the warning, the main source of uncertainty is first located through log recording and sensitivity analysis, and then engineers calibrate the sensor, supplement at least 20 valid data points, and check the equipment stability. If the width still exceeds the preset threshold after re-evaluation, a full model retraining is forced until the prediction confidence meets the requirements (width ≤ 20%). This closed-loop mechanism ensures the reliability and adaptability of this invention in complex industrial environments.

[0023] (4.3) Model Adaptive Optimization: Based on the uncertainty analysis results, the model parameters are dynamically adjusted or the prediction results of multiple models are fused. When the width of the predicted confidence interval exceeds the preset threshold, a warning for data supplementation or model retraining is triggered.

[0024] Furthermore, the specific implementation of the Monte Carlo uncertainty analysis in step (4.2) includes: setting the number of simulations N; independently sampling each uncertainty factor according to its probability distribution; substituting the N sets of sampling parameters into the trained prediction model to obtain N predicted remaining lifetime values; calculating the mean, standard deviation, and 2.5% and 97.5% quantiles of these N values ​​as 95% confidence intervals; and outputting the probability density distribution plot and cumulative distribution function plot.

[0025] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method for constructing a rotating target life prediction model based on incremental learning.

[0026] The beneficial effects of this invention are as follows: Through systematic data quality control, advanced machine learning training, intelligent incremental learning mechanisms, and rigorous uncertainty quantification, this invention achieves significantly improved prediction accuracy, strong model adaptability, excellent generalization performance, and high reliability. Under stable process conditions, this invention achieves a prediction accuracy within ±15%, can automatically optimize and adjust based on new data, adapts to different process conditions and target batches, and provides confidence interval assessment through uncertainty quantification, offering an efficient and reliable solution for the intelligent operation and maintenance of magnetron sputtering processes.

[0027] This invention, for the first time, constructs a four-dimensional parameter system covering electrical parameters, process parameters, target state parameters, and time-series parameters, achieving a comprehensive characterization of the sputtering process. Furthermore, it constructs derived features to map the physical essence of sputtering efficiency, providing a solid feature foundation for subsequent incremental learning and Monte Carlo uncertainty quantification. The multi-source parameter system of this application provides sufficient "contextual information" for the incremental learning mechanism, enabling the model to distinguish between consumption rate fluctuations caused by changes in process parameters and consumption rate changes caused by the aging of the target itself, thereby achieving efficient dynamic updates of the model. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a flowchart of the overall process for constructing a rotating target life prediction model based on incremental learning according to the present invention.

[0030] Figure 2 This is a schematic diagram of the data quality inspection and preprocessing process in an embodiment of the present invention.

[0031] Figure 3 This is a schematic diagram illustrating the training and evaluation of machine learning models in an embodiment of the present invention.

[0032] Figure 4 This is a schematic diagram illustrating the working principle of the incremental learning mechanism in this embodiment of the invention.

[0033] Figure 5 This is a flowchart of uncertainty analysis and model optimization in an embodiment of the present invention.

[0034] Figure 6 This is an error distribution analysis and learning curve diagram during the model training process in this embodiment of the invention.

[0035] Figure 7 This is a comparison of predicted and actual values ​​and a residual distribution diagram during the model training process in this embodiment of the invention.

[0036] Figure 8 This is a feature importance analysis and performance radar chart during the model training process in this embodiment of the invention. Detailed Implementation

[0037] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0038] This invention provides a method for constructing a rotating target lifetime prediction model based on incremental learning. See [link to relevant documentation]. Figure 1 This includes the following steps: Step 1: Data Quality Control Stage: Construct a multi-dimensional data acquisition system to obtain raw production data including electrical parameters, process parameters, target material status parameters, and time series parameters; perform data quality checks and preprocessing operations on the raw data to construct a standardized multi-source fusion dataset.

[0039] This embodiment uses IZO rotating target material as the application object, and the data comes from the actual production records of a display panel manufacturer for 6 months.

[0040] First, step (1.1) was performed to construct a multi-dimensional data acquisition system. The data acquisition system (SCADA) connected to the magnetron sputtering equipment was used to collect parameters including: sputtering power and cumulative energy consumption in electrical parameters; working gas pressure, argon flow rate, oxygen ratio, and temperature in process parameters; and initial thickness and thickness reduction in target material status parameters. A total of 1250 valid records were collected.

[0041] Next, proceed to step (1.2) to perform a data quality check. For example... Figure 2 As shown, the data quality inspection process includes missing value detection and imputation, physical plausibility checking (including statistical outlier detection), data integrity verification, and numerical range verification. The DataQualityChecker class is used to scan all feature columns, and linear imputation is used to fill in missing records for some fields. In the physical plausibility check, outlier records with negative thickness reduction or negative cumulative energy consumption are removed. Statistical outlier detection uses the interquartile range (IQR) method; data points exceeding the normal range are corrected or removed based on the degree of deviation. In the data integrity verification, it is checked whether all required parameters in each record (including sputtering power, cumulative energy consumption, working pressure, argon flow rate, oxygen ratio, temperature, initial thickness, and thickness reduction) are non-empty; the continuity of timestamps is verified to ensure that the sampling interval does not exceed a set threshold; if a long period of missing data is found, the data for that period is marked as invalid; the consistency of the number of records in each parameter table within the same batch is checked to prevent dimensional mismatches caused by asynchronous data acquisition. In the numerical range verification, physical upper and lower limits are set for each numerical parameter, and data exceeding the limits are directly marked as invalid. For monotonically increasing parameters such as cumulative energy consumption, it is verified whether the sequence is strictly non-decreasing. If a backsliding occurs, it is determined to be sensor drift or data recording error and is corrected.

[0042] Finally, step (1.3) was performed, involving data preprocessing and feature fusion. A time-weighted average method was used to align all data to the minute level. Z-score standardization was then applied to the feature data. Two derived features, the power-pressure ratio and the efficiency index, were constructed. After preprocessing, the original 1250 data points, after denoising, imputation, and anomaly handling, achieved a 95% retention rate, forming a standardized multi-source fusion dataset.

[0043] The power-to-pressure ratio is calculated as power density / working pressure. Power density represents the energy input density of the plasma, while working pressure affects the mean free path of particles and sputtering yield. Dividing the two gives a comprehensive representation of the "energy density per unit pressure," indirectly reflecting the balance between sputtering efficiency and gas scattering effects. The efficiency index is calculated as thickness reduction ÷ cumulative energy consumption. The efficiency index represents the "target thickness that can be removed per unit of electrical energy," i.e., energy utilization efficiency. A higher value indicates that more target material can be sputtered with the same amount of electrical energy, indicating high sputtering efficiency and optimized process conditions (such as suitable gas pressure and power). Conversely, a low efficiency index suggests potential energy waste (such as overheating due to excessive power or backsputtering due to unsuitable gas pressure).

[0044] Step 2: Machine Learning Model Training Phase: Based on the standardized multi-source fusion dataset, construct a feature engineering system that includes basic features and derived features; train the initial lifetime prediction model using the random forest regression algorithm, and quantitatively evaluate the model performance using comprehensive evaluation metrics.

[0045] Perform step (2.1) to construct the feature engineering system. Determine the basic feature set as [power density, working pressure, argon flow rate, oxygen ratio, temperature, cumulative energy consumption], and combine it with derived features power-to-pressure ratio and efficiency index. The final feature vector dimension is 8.

[0046] Perform step (2.2) to train the model. Use the Random Forest regression algorithm with the following parameters: n_estimators=100, max_depth=10, random_state=42, n_jobs=-1, and enable warm_start=True. Divide the dataset into training and test sets at an 80% / 20% ratio. Use 5-fold cross-validation during training.

[0047] Step (2.3) is executed for comprehensive model evaluation. Evaluation on the test set yielded the following results: Mean Absolute Error (MAE) of 0.08 mm, Root Mean Square Error (RMSE) of 0.11 mm, R² score of 0.87, and Mean Relative Error of 12.5%. Simultaneously, the `plot_comprehensive_evaluation` function is called to generate a comprehensive evaluation chart, as shown below. Figure 3and Figures 6-8 As shown. Figures 6-8 This is a comprehensive evaluation chart generated during the model training process in this embodiment of the invention, comprising six sub-charts: a scatter plot comparing predicted and actual values, a histogram of residual distribution, a bar chart of feature importance, a box plot of error analysis, a learning curve, and a performance radar chart. Feature importance analysis shows that power density and cumulative energy consumption are the two most important predictors, which are highly consistent with the physical mechanism. The performance radar chart shows a comprehensive score of 87.5%.

[0048] Step 3: Incremental Learning Mechanism Stage: Design an incremental data collection strategy to continuously acquire effective data from the new production process; based on the effective data, use a robust weighted average method to dynamically update the key parameters of the initial lifetime prediction model and automatically iterate the model version.

[0049] Perform step (3.1) and design an incremental data collection strategy. Continuously collect production data in new production batches and set validity verification rules: thickness reduction > 0.1 mm and consumption rate within the range of 50-500 kWh / mm, in order to filter out valid production data.

[0050] Execute step (3.2) to dynamically update the model parameters. When the cumulative amount of newly collected valid production data (hereinafter referred to as new data) reaches a preset threshold (e.g., 83 data points), incremental learning is triggered. The target material consumption rate is updated using a robust weighted average method: the weight of new data w_new = min(0.5, 83×0.1)=0.5, and the weight of old data (referring to the original production data) w_old=0.5. The median consumption rate in the new data, median_new, is 132.1 kWh / mm, and the original consumption rate, rate_old, is 125.3 kWh / mm. Substituting these values ​​into the formula, the new consumption rate is 128.7 kWh / mm.

[0051] In one embodiment, a dynamic weight correction mechanism based on the prediction residual is also introduced, which replaces the ordinary median with the weighted median of the consumption rate in the new data to update the new consumption rate and improve robustness to outliers.

[0052] The ordinary median (also known as the simple median) is the middle value after sorting a set of data by value. If the number of data points is odd, the middle value is used; if the number is even, the average of the two middle values ​​is used. All data points have equal weight in the median calculation. The weighted median, on the other hand, assigns a non-negative weight to each data point, sorts them by value, and finds the value where the cumulative weight reaches half (≥50%) of the total weight. It reflects the differences in importance between different samples more accurately than the ordinary median.

[0053] Execute step (3.3), model version management. After completing the parameter update, the model version number automatically iterates from 1.0.0 to 1.1.0, recording the update time and a summary of the update content, and retaining the 20 most recent incremental data points. The incremental learning workflow is as follows: Figure 4 As shown.

[0054] Step 4: Uncertainty Quantification Stage: Identify the sources of uncertainty in the prediction process; perform uncertainty propagation calculations on the prediction results using Monte Carlo analysis, and finally output probabilistic lifetime prediction results containing confidence interval information.

[0055] Step (4.1) is executed to identify sources of uncertainty. This embodiment considers measurement uncertainty (power sensor error ±0.5%, thickness measurement error ±0.05mm), model structure uncertainty, and process parameter uncertainty (process parameter fluctuation ±2%) as a combination of sampling parameters.

[0056] Perform step (4.2) to conduct Monte Carlo uncertainty analysis. Set the number of Monte Carlo simulations N=1000. Independently sample each uncertainty factor according to its probability distribution, and substitute the combined sampling parameters into the prediction model to obtain the distribution sample of remaining lifetime. The calculated mean remaining lifetime for a certain prediction is 45.2 hours, with a 95% confidence interval of [40.5, 49.9] hours. Simultaneously output the probability density distribution plot and the cumulative distribution function plot.

[0057] Preferably, when the confidence interval width obtained from the Monte Carlo analysis exceeds 20% of the predicted mean, the system automatically triggers a data supplementation warning. Upon triggering this warning, the system performs the following operations: (a) generates a warning log and marks the current prediction ID; (b) outputs a ranking of the contribution of the main sources of uncertainty through sensitivity analysis; (c) if the prediction results are used for target material replacement decisions, it automatically suspends execution and issues a manual review request. Upon receiving the warning, the process engineer should prioritize the following measures: calibrate the thickness measurement sensor to an error ≤ ±0.02 mm; collect at least 20 new data points under the same process conditions to trigger incremental learning; and check the operating status of the working gas pressure and argon flow control equipment. After completing the above operations, rerun the Monte Carlo analysis. If the confidence interval width still exceeds 20%, force model retraining and update the model version.

[0058] Execute step (4.3) to establish a model adaptive optimization mechanism. The system automatically adjusts the strategy based on the confidence interval width: if the width is less than 10% of the predicted mean, the result is highly reliable; if the width is between 10% and 20%, it is for reference only; if the width is greater than 20%, a data supplementation warning is triggered. The overall process of uncertainty analysis and model optimization is as follows: Figure 5 As shown.

[0059] In summary, this invention successfully constructs an adaptive, high-precision, and highly reliable rotating sputtering target lifetime prediction model through four stages: data quality control, machine learning training, incremental learning, and uncertainty quantification. The technical solution of this invention is not only applicable to IZO rotating sputtering targets but can also be extended to lifetime prediction of other ceramic sputtering targets such as ITO and IGZO, demonstrating broad industrial application prospects.

[0060] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described method for constructing a rotating target life prediction model based on incremental learning.

[0061] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method for constructing a rotating target life prediction model based on incremental learning.

[0062] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0063] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0064] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0065] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0066] The above embodiments are only used to illustrate the design concept and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, all equivalent changes or modifications made based on the principles and design ideas disclosed in the present invention are within the protection scope of the present invention.

Claims

1. A method for constructing a rotating target life prediction model based on incremental learning, characterized in that, Includes the following steps: Step 1: Obtain raw production data containing electrical parameters, process parameters, target material state parameters, and timing parameters; perform data quality checks and preprocessing on the raw production data to construct a standardized multi-source fusion dataset; Step 2: Based on the standardized multi-source fusion dataset, construct a feature engineering system containing basic features and derived features for training the initial lifetime prediction model; Derivative characteristics include power-to-pressure ratio and efficiency index; Step 3: Continuously acquire production data during the new production process; based on this production data, dynamically update the parameters of the initial lifetime prediction model and automatically iterate the model version; Step three is achieved through the following sub-steps: (3.1) Incremental data collection: Set data validity verification rules, including minimum threshold for thickness consumption and reasonable range of consumption rate, and only include new production data that has passed verification into the incremental learning process; (3.2) Dynamic update of model parameters: The target consumption rate is updated using a robust weighted average method. w_old is the weight of the original production data, rate_old is the original consumption rate, w_new is the weight of the new data, and median_new is the median consumption rate in the new data. (3.3) Model version management: After each incremental learning, the model version number is automatically iterated, the update time is recorded, and the most recent few incremental data are retained; Step 4: Identify the sources of uncertainty in the model prediction process; perform uncertainty propagation calculations on the model prediction results using Monte Carlo analysis, and finally output probabilistic lifetime prediction results containing confidence interval information.

2. The method according to claim 1, characterized in that, Step one is achieved through the following sub-steps: (1.1) Multi-dimensional data acquisition: The acquired parameters include sputtering power and cumulative energy consumption in electrical parameters, working gas pressure, argon flow rate, oxygen ratio and temperature in process parameters, and initial thickness and thickness reduction in target material state parameters. (1.2) Data quality checks: including missing value detection and imputation, outlier identification and correction based on physical rationality, data integrity verification, and numerical range verification; (1.3) Data preprocessing and feature fusion: Time alignment of multi-frequency data, Z-score standardization of data, and construction of derived features.

3. The method according to claim 1, characterized in that, Step two is achieved through the following sub-steps: (2.1) Constructing a feature engineering system: Determine the basic feature set as power density, pressure, argon flow rate, oxygen ratio, temperature, and cumulative energy consumption, and combine it with derived features; (2.2) Model training: Random forest regression algorithm is adopted, parameters that support incremental learning are set, the feature engineering system is used as input, and the reduction of target thickness is used as the prediction target for model training; (2.3) Model comprehensive evaluation: Calculate the mean absolute error, root mean square error and R² score using the test set, and generate scatter plots of predicted values ​​vs. actual values, residual distribution plots, feature importance plots, learning curves and performance radar charts.

4. The method according to claim 1, characterized in that, New data weights The weight of the original production data is w_old = 1 - w_new.

5. The method according to claim 1, characterized in that, Step four is achieved through the following sub-steps: (4.1) Identification of sources of uncertainty: including at least measurement uncertainty, model structure uncertainty, and process parameter fluctuation uncertainty; (4.2) Monte Carlo uncertainty analysis: Random sampling is performed on the sources of uncertainty to simulate multiple combinations of uncertainty, the probability distribution of the prediction results is calculated, and a confidence interval with a 95% confidence level is provided based on the distribution; (4.3) Model adaptive optimization: dynamically adjust model parameters or fuse prediction results from multiple models based on uncertainty analysis results.

6. The method according to claim 2, characterized in that, The specific methods for checking the physical rationality of the original production data in step (1.2) include: when the thickness reduction is less than 0 or the cumulative energy consumption is less than 0, it is judged as a physical abnormal record and removed; when the power density is not within the preset range, it is judged as sensor fault data and marked; when the energy density exceeds the normal range determined by the interquartile range method, it is judged as a statistical abnormal value and corrected.

7. The method according to claim 3, characterized in that, The specific process of generating the performance radar chart in step (2.3) includes: calculating the normalized score of each dimension based on five dimensions: R² score, prediction accuracy, model stability, generalization ability, and feature diversity; distributing the five scores evenly in the polar coordinate system according to the angle, and connecting the points to form a closed polygon; calculating the ratio of the polygon area to the area of ​​the regular pentagon as the comprehensive score, and displaying it as a percentage in the center of the radar chart.

8. The method according to claim 4, characterized in that, In step (3.2), when updating the target consumption rate using the robust weighted average method, a dynamic weight correction mechanism based on the prediction residual is also introduced: the prediction residual of each sample in the incremental dataset is calculated, and the sample weight is assigned according to the size of the residual. The larger the residual, the lower the weight of the sample; the weighted median is used instead of the ordinary median as median_new.

9. The method according to claim 5, characterized in that, The specific implementation of Monte Carlo uncertainty analysis in step (4.2) includes: setting the number of simulations N; independently sampling each uncertainty factor according to its probability distribution; substituting the N sets of sampling parameters into the trained prediction model to obtain N predicted remaining lifetime values; calculating the mean, standard deviation, and 2.5% and 97.5% quantiles of these N values ​​as 95% confidence intervals; and outputting the probability density distribution plot and cumulative distribution function plot.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the method for constructing a rotating target life prediction model based on incremental learning as described in any one of claims 1 to 9.

Citation Information

Patent Citations

  • Magnetron sputtering parameter adjusting method and device, electronic equipment and storage medium

    CN120911252A

  • Target material monitoring method and system based on multi-source data and life prediction

    CN121598797A