System for multi-view quantum-tensor partition feature selection and sample elimination and method thereof, and non-transitory computer-readable storage medium

Abstract

This invention proposes a multi-perspective framework for quantum tensor partitioning feature selection (Q-TPFS), which integrates quantum-inspired amplitude collapse, tensor low-rank regularization, and adaptive sample exclusion. It can discard features with low contribution while retaining complementary information across multiple perspectives. It captures potential structures through tensor singular value decomposition and graph Laplace constraints, and adaptively excludes noisy samples to enhance the robustness of clustering and classification. Compared with traditional methods, Q-TPFS solves the problems of feature selection, structure compression, and sample exclusion. This invention guarantees convergence through an augmented Lagrangian framework and quantum-inspired collapse theory. In smart manufacturing, Q-TPFS can accurately grasp the key parameters of multiple process steps under multi-perspective conditions, thereby shortening the fault diagnosis cycle and improving yield.

Claims

1. A multi-view quantum tensor segmentation feature selection and sample exclusion system, comprising: a memory storing a plurality of process data, a plurality of measurement test data, a plurality of factor data, and an outlier threshold, wherein the process data is used or generated when a production machine processes a plurality of workpieces, the process data is mapped one-to-one to the workpieces, each process data includes a plurality of process parameters, the measurement test data includes a plurality of measurement values ​​of the workpieces, and the factor data includes at least two of machine data, a recipe data, and a process step data corresponding to the workpieces; and a processor electrically connected to the memory, the processor receiving the process data, the measurement test data, the factor data, and the outlier threshold, and configured to perform an operation including the following steps: performing a multi-view data segmentation operation, the multi-view data segmentation operation including segmenting the process data and the measurement test data into a plurality of views based on the factor data, the views corresponding to a plurality of features; The process involves: performing a quantum-inspired feature selection operation, which includes confirming whether one of the multiple amplitudes of the features from the given perspectives approaches a preset amplitude value to generate a confirmation result, and retaining or removing one of the features corresponding to the amplitudes of the given perspectives based on the confirmation result to generate multiple important features; performing a tensor low-rank regularization operation, which includes stacking the multiple matrices corresponding to the important features from the given perspectives into a high-order tensor, and performing a low-rank decomposition on the high-order tensor through singular value decomposition to generate multiple samples to preserve a common structure among the given perspectives, wherein each sample has an outlier degree and corresponds to the given perspectives; and performing an adaptive sample exclusion operation, which includes excluding a portion of the samples whose outlier degree is higher than the outlier threshold to converge an objective function corresponding to the given samples.

2. The multi-view quantum tensor partitioning feature selection and sample exclusion system as described in claim 1, wherein the memory stores a critical point, and the processor is configured to perform operations further comprising the following steps: performing an optimization convergence operation, the optimization convergence operation comprising updating the quantum-inspired feature selection operation, the tensor low-rank regularization operation, and the adaptive sample exclusion operation using an alternating direction method of multipliers (ADMM) and a block coordinate descent method, and executing an iterative algorithm to maintain the objective function for the samples within a convergence range and converge to the critical point.

3. The multi-view quantum tensor segmentation feature selection and sample exclusion system as described in claim 2, wherein, In the quantum-inspired feature selection operation, the preset amplitude value is 0; when the confirmation result is yes, the feature corresponding to the amplitude of the corresponding viewpoint is removed; when the confirmation result is no, the feature corresponding to the amplitude of the corresponding viewpoint is retained; in the tensor low-rank regularization operation, each matrix is ​​a Gram matrix, the higher-order tensor is a third-order tensor, the singular value decomposition is a tensor singular value decomposition (t-SVD); and in the optimization convergence operation, the alternating direction multiplier method is an augmented Lagrange method, and the iterative algorithm includes repeatedly performing the quantum-inspired feature selection operation, the tensor low-rank regularization operation, and the adaptive sample exclusion operation according to a number of iterations.

4. The multi-view quantum tensor segmentation feature selection and sample exclusion system as described in claim 1, wherein in the multi-view data segmentation operation, the factor data includes machine data, recipe data, and process step data corresponding to the workpieces; the machine data includes data of a plurality of production machines, which are distinct from each other; the recipe data includes data of a plurality of recipes, which are distinct from each other; the process step data includes data of a plurality of process steps, which are distinct from each other; and the views include a first view, a second view, and a third view, the first view corresponding to the production machines, the second view corresponding to the recipes, and the third view corresponding to the process steps.

5. The multi-view quantum tensor partitioning feature selection and sample exclusion system as described in claim 1, wherein the views correspond to a plurality of factory information systems in a semiconductor manufacturing process, the factory information systems including an Advanced Process Control (APC) system, a Fault Detection and Classification (FDC) system, a Manufacturing Execution System (MES), a Statistical Process Control (SPC) system, and an Equipment Data Acquisition and Tool Log (EDA / Log) system, the factory information systems being electrically connected to the memory and the processor, the semiconductor manufacturing process including a wide bandgap (WBG) epitaxial (Epi) process, the multi-view data partitioning operation, the quantum-inspired feature selection operation, the tensor low-rank regularization operation, and the adaptive sample exclusion operation being performed in the wide bandgap epitaxial process.

6. The multi-view quantum tensor segmentation feature selection and sample exclusion system as described in claim 1, wherein the adaptive sample exclusion operation further comprises: in a first training phase, moving the portion of the samples whose outlier degree is higher than the outlier threshold to a discard region of the memory; and then in a second training phase, determining whether to reinstate the portion of the samples from the discard region based on the convergence of the objective function; wherein, The first training phase preceded the second training phase.

7. The multi-view quantum tensor segmentation feature selection and sample exclusion system as described in claim 6, wherein in the adaptive sample exclusion operation, when the convergence of the objective function meets a stable convergence, the portion of the samples is retrieved from the region to be discarded; and when the convergence of the objective function does not meet the stable convergence, the portion of the samples is maintained in the region to be discarded.

8. A multi-view quantum tensor segmentation-based feature selection and sample exclusion method, comprising the following steps: obtaining, by a processor, a plurality of process data, a plurality of measurement test data, a plurality of factor data, and an outlier threshold in a memory, wherein the process data is used or generated when a production machine processes a plurality of workpieces, the process data is mapped to the workpieces in a one-to-one manner, each process data includes a plurality of process parameters, the measurement test data includes a plurality of measurement values ​​of the workpieces, and the factor data includes at least two of a machine data, a recipe data, and a process step data corresponding to the workpieces; performing a multi-view data segmentation operation by the processor, the multi-view data segmentation operation comprising dividing the process data and the measurement test data into a plurality of views based on the factor data, the views corresponding to a plurality of features; The processor performs a quantum-inspired feature selection operation, which includes confirming whether one of the multiple amplitudes of the features of the views approaches a preset amplitude value to generate a confirmation result, and retaining or removing one of the features corresponding to the amplitudes of the views based on the confirmation result to generate multiple important features; the processor performs a tensor low-rank regularization operation, which includes stacking the multiple matrices corresponding to the important features of the views into a high-order tensor, and performing low-rank decomposition on the high-order tensor through singular value decomposition to generate multiple samples to preserve a common structure among the views, wherein each sample has an outlier degree and corresponds to the views; and the processor performs an adaptive sample exclusion operation, which includes excluding a portion of the samples whose outlier degree is higher than the outlier threshold to converge an objective function corresponding to the samples.

9. The multi-view quantum tensor segmentation feature selection and sample exclusion method as described in claim 8 further comprises: performing an optimization convergence operation by the processor, the optimization convergence operation comprising updating the quantum-inspired feature selection operation, the tensor low-rank regularization operation, and the adaptive sample exclusion operation using an alternating direction method of multipliers (ADMM) and a block coordinate descent method, and executing an iterative algorithm to keep the objective function corresponding to the samples within a convergence range and converge to a critical point.

10. The multi-view quantum tensor segmentation feature selection and sample exclusion method as described in claim 9, wherein, In the quantum-inspired feature selection operation, the preset amplitude value is 0; when the confirmation result is yes, the feature corresponding to the amplitude of the corresponding viewpoint is removed; when the confirmation result is no, the feature corresponding to the amplitude of the corresponding viewpoint is retained; in the tensor low-rank regularization operation, each matrix is ​​a Gram matrix, the higher-order tensor is a third-order tensor, the singular value decomposition is a tensor singular value decomposition (t-SVD); and in the optimization convergence operation, the alternating direction multiplier method is an augmented Lagrange method, and the iterative algorithm includes repeatedly performing the quantum-inspired feature selection operation, the tensor low-rank regularization operation, and the adaptive sample exclusion operation according to a number of iterations.

11. The multi-view quantum tensor segmentation feature selection and sample exclusion method as described in claim 8, wherein in the multi-view data segmentation operation, the factor data includes machine data, recipe data, and process step data corresponding to the workpieces; the machine data includes data of a plurality of production machines, which are distinct from each other; the recipe data includes data of a plurality of recipes, which are distinct from each other; the process step data includes data of a plurality of process steps, which are distinct from each other; and the views include a first view, a second view, and a third view, the first view corresponding to the production machines, the second view corresponding to the recipes, and the third view corresponding to the process steps.

12. The multi-view quantum tensor segmentation feature selection and sample exclusion method as described in claim 8, wherein the views correspond to a plurality of factory information systems in a semiconductor manufacturing process, the factory information systems including an Advanced Process Control (APC) system, a Fault Detection and Classification (FDC) system, a Manufacturing Execution System (MES), a Statistical Process Control (SPC) system, and an Equipment Data Acquisition and Tool Log (EDA / Log) system, the factory information systems being electrically connected to the memory and the processor, the semiconductor manufacturing process including a wide bandgap (WBG) epitaxial (Epi) process, the multi-view data segmentation operation, the quantum-inspired feature selection operation, the tensor low-rank regularization operation, and the adaptive sample exclusion operation being performed in the wide bandgap epitaxial process.

13. The multi-view quantum tensor segmentation feature selection and sample exclusion method as described in claim 8, wherein the adaptive sample exclusion operation further comprises: in a first training phase, moving the portion of the samples whose outlier degree is higher than the outlier threshold to a discard region of the memory; and then in a second training phase, determining whether to reinstate the portion of the samples from the discard region based on the convergence degree of the objective function; wherein, The first training phase preceded the second training phase.

14. The multi-view quantum tensor segmentation feature selection and sample exclusion method as described in claim 13, wherein in the adaptive sample exclusion operation, when the convergence degree of the objective function meets a stable convergence, the portion of the samples is retrieved from the region to be discarded; and when the convergence degree of the objective function does not meet the stable convergence, the portion of the samples is maintained in the region to be discarded.

15. A non-transitory computer-readable storage medium having a plurality of instructions, which, when executed on a processor, cause the processor to perform a multi-view quantum tensor segmentation feature selection and sample exclusion method, the multi-view quantum tensor segmentation feature selection and sample exclusion method comprising the steps of: obtaining, by the processor, a plurality of process data, a plurality of measurement test data, a plurality of factor data, and an outlier threshold in a memory, wherein the process data is used or generated when a production machine processes a plurality of workpieces, the process data is mapped one-to-one to the workpieces, each of the process data includes a plurality of process parameters, the measurement test data includes a plurality of measurement values ​​of the workpieces, and the factor data includes at least two of a machine data, a recipe data, and a process step data corresponding to the workpieces; The processor performs a multi-view data segmentation operation, which includes dividing the process data and measurement test data into a plurality of views based on the factor data, and the views correspond to a plurality of features; the processor performs a quantum-inspired feature selection operation, which includes confirming whether one of the plurality of amplitudes of the features of the views approaches a preset amplitude value to generate a confirmation result, and retaining or removing one of the features corresponding to the amplitudes of the views based on the confirmation result to generate a plurality of important features; The processor performs a tensor low-rank regularization operation, which includes stacking the complex matrices corresponding to the important features of these perspectives into a high-order tensor, and performing a low-rank decomposition on the high-order tensor through a singular value decomposition to generate a complex number of samples to preserve a common structure among these perspectives, wherein each sample has an outlier degree and corresponds to these perspectives; and the processor performs an adaptive sample exclusion operation, which includes excluding a portion of the samples whose outlier degree is higher than the outlier threshold, in order to converge an objective function corresponding to these samples.

16. The non-transitory computer-readable storage medium as described in claim 15, wherein the multi-view quantum tensor partitioned feature selection and sample exclusion method further comprises: performing an optimization convergence operation by the processor, the optimization convergence operation comprising updating the quantum-inspired feature selection operation, the tensor low-rank regularization operation, and the adaptive sample exclusion operation using an alternating direction method of multipliers (ADMM) and a block coordinate descent operation, and executing an iterative algorithm to maintain the objective function for the samples within a convergence range and converge to a critical point.

17. The non-transitory computer-readable storage medium as described in claim 16, wherein, In the quantum-inspired feature selection operation, the preset amplitude value is 0; when the confirmation result is yes, the feature corresponding to the amplitude of the corresponding viewpoint is removed; when the confirmation result is no, the feature corresponding to the amplitude of the corresponding viewpoint is retained; in the tensor low-rank regularization operation, each matrix is ​​a Gram matrix, the higher-order tensor is a third-order tensor, the singular value decomposition is a tensor singular value decomposition (t-SVD); and in the optimization convergence operation, the alternating direction multiplier method is an augmented Lagrange method, and the iterative algorithm includes repeatedly performing the quantum-inspired feature selection operation, the tensor low-rank regularization operation, and the adaptive sample exclusion operation according to a number of iterations.

18. The non-transitory computer-readable storage medium as described in claim 15, wherein in the operation of the multi-view data partitioning, the factor data includes machine data, recipe data, and process step data corresponding to the workpieces, the machine data including data of a plurality of production machines that are different from each other, the recipe data including data of a plurality of recipes that are different from each other, the process step data including data of a plurality of process steps that are different from each other; and the views including a first view, a second view, and a third view, the first view corresponding to the production machines, the second view corresponding to the recipes, and the third view corresponding to the process steps.

19. The non-transitory computer-readable storage medium as described in claim 15, wherein the views correspond to a plurality of factory information systems in a semiconductor manufacturing process, the factory information systems including an Advanced Process Control (APC) system, a Fault Detection and Classification (FDC) system, a Manufacturing Execution System (MES), a Statistical Process Control (SPC) system, and an Equipment Data Acquisition and Tool Log (EDA / Log) system, the factory information systems being electrically connected to the memory and the processor, the semiconductor manufacturing process including a wide bandgap (WBG) epitaxial (Epi) process, the multi-view data partitioning operation, the quantum-inspired feature selection operation, the tensor low-rank regularization operation, and the adaptive sample exclusion operation being performed in the wide bandgap epitaxial process.

20. The non-transitory computer-readable storage medium as described in claim 15, wherein the adaptive sample exclusion operation further comprises: during a first training phase, moving the portion of the samples whose outlier score is higher than the outlier threshold to a discard region of the memory, and then during a second training phase, determining whether to reinstate the portion of the samples from the discard region based on the convergence of one of the objective functions; wherein, The first training phase precedes the second training phase; wherein, when the convergence of the objective function meets a stable convergence, the portion of the samples is retrieved from the region to be discarded; and wherein, when the convergence of the objective function does not meet the stable convergence, the portion of the samples is kept in the region to be discarded.

Images