KR20220028724A - Method and Apparatus for Virtual Measurement for Calculating Predicted Value and Feature Importance Based on Feature Values of Time Series Data - Google Patents

Abstract

Disclosed are a virtual measurement method for calculating a prediction value and feature importance based on feature values of time series data, and an apparatus therefor. A virtual measurement method according to an embodiment of the present invention may include: a time series data obtaining step of obtaining time series data including a time series type variable; a model learning processing step of extracting a feature value for the time series data, calculating a prediction value for learning based on the feature value, and generating a learning model; a virtual measurement processing step of calculating a virtual measurement result by applying new time series data to the learning model; and a virtual measurement result outputting step of providing the virtual measurement result to an external device.

Description

A virtual measurement method for calculating predicted values and feature importance based on feature values of time series data, and an apparatus therefor

The present invention relates to a virtual measurement method for calculating predicted values and feature importance based on feature values of time series data, and an apparatus therefor.

The content described in this section merely provides background information on the embodiments of the present invention and does not constitute the prior art.

The production of products such as semiconductors or flat panel displays (FPDs) consists of numerous manufacturing processes, and sampling and measurement are mostly carried out to shorten production time. Due to this, there is a lack of measurement data, which makes it difficult to analyze the cause of defects and improve product quality.

In order to solve the problem of such metrology-based process management, interest in virtual metrology (VM) has recently been focused. Virtual metrology utilizes sensor data of equipment generated in the manufacturing process to predict measurement values for all wafers without actually performing the measurement process. If the measured value predicted through virtual measurement is linked with the process control system, product quality improvement and manufacturing cost can be reduced.

However, in order to learn the prediction algorithm of virtual measurement processing, the user must process and input appropriate features from raw data. In general, neural networks have the advantage of extracting features from raw data and learning by themselves, but despite many efforts in academia, accurate interpretation of prediction results is impossible due to high nonlinearity, It is not known how much was reflected in the prediction results. Due to this, despite advances in neural network technology, it has been avoided to use in processes and manufacturing sites where cause analysis of results is important. Accordingly, in order to solve the above-mentioned problems, a neural network-based virtual measurement technology capable of extracting and interpreting the features of raw data by itself is required.

The present invention provides a virtual measurement method for extracting and predicting a feature value from time series data, and calculating a predicted value and feature importance based on a feature value of the time series data for calculating the importance for each feature value, and an apparatus therefor. It has a main purpose.

According to an aspect of the present invention, a virtual measurement method for achieving the above object includes: a time series data obtaining step of obtaining time series data including a variable in a time series form; a model learning processing step of extracting a feature value for the time series data, calculating a prediction value for learning based on the feature value, and generating a learning model; a virtual measurement processing step of calculating a virtual measurement result by applying new time series data to the learning model; and outputting a virtual measurement result of providing the virtual measurement result to an external device.

In addition, according to another aspect of the present invention, a virtual measurement device for achieving the above object, at least one or more processors; and a memory for storing one or more programs executed by the processor, wherein when the programs are executed by one or more processors, the one or more processors obtains time series data including a time series variable. acquisition stage; a model learning processing step of extracting a feature value for the time series data, calculating a prediction value for learning based on the feature value, and generating a learning model; a virtual measurement processing step of calculating a virtual measurement result by applying new time series data to the learning model; and outputting a virtual measurement result of providing the virtual measurement result to an external device.

As described above, the present invention has an effect of automatically extracting feature values with high explanatory power for response variables from time series data of explanatory variables through a learning model.

In addition, the present invention has the effect of providing a methodology capable of predicting and analyzing the cause of the process and actual measurement by calculating the importance (contribution) of the automatically extracted feature value.

In addition, the present invention has the effect of improving product quality and reducing manufacturing costs according to process control by automatically extracting feature values from time-series data, calculating their importance, and providing virtual measurement results.

1 is a block diagram schematically illustrating a virtual metrology-based process system according to an embodiment of the present invention.
2 is a block diagram schematically illustrating a virtual measurement apparatus according to an embodiment of the present invention.
3 and 4 are block diagrams for explaining a processor operation of a virtual measurement apparatus according to an embodiment of the present invention.
5 is a block diagram schematically showing a model learning processing unit according to an embodiment of the present invention.
6 is a flowchart illustrating a virtual measurement method for model learning processing according to an embodiment of the present invention.
7 is a block diagram schematically illustrating a virtual measurement processing unit according to an embodiment of the present invention.
8 is a flowchart illustrating a virtual measurement method using a learning model according to an embodiment of the present invention.
9A and 9B are exemplary diagrams for explaining a model learning operation of a virtual measurement device according to an embodiment of the present invention.
10A and 10B are exemplary diagrams for explaining an operation of calculating a predicted value and a feature importance of a virtual measurement device according to an embodiment of the present invention.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. In describing the present invention, if it is determined that a detailed description of a related known configuration or function may obscure the gist of the present invention, the detailed description thereof will be omitted. In addition, preferred embodiments of the present invention will be described below, but the technical spirit of the present invention is not limited thereto or may be variously implemented by those skilled in the art without being limited thereto. Hereinafter, a virtual measurement method for calculating predicted values and feature importance based on feature values of time series data proposed by the present invention and an apparatus therefor will be described in detail with reference to the drawings.

1 is a block diagram schematically illustrating a virtual metrology-based process control system according to an embodiment of the present invention.

The virtual metrology-based process control system 10 according to the present embodiment includes a virtual metrology device 100 and a process and measurement device 200 . The process control system 10 of FIG. 1 is according to one embodiment, and not all blocks shown in FIG. 1 are essential components, and in another embodiment, some blocks included in the process control system 10 are added or changed. Or it can be deleted. For example, the process control system 10 may further include an input variable providing device (not shown) that provides input variables, and the process and measuring device 200 is a separate device such as a process device and a measuring device. can be implemented as

The process control system 10 according to the present embodiment is preferably a system that performs at least one process for producing products such as semiconductors or flat panel displays (FPDs), but is not necessarily limited thereto. It may be one of the process systems in various fields to which the learning model for

The virtual measurement device 100 and the process and measurement device 200 of the process control system 10 may operate according to input variables, and the input variables include time series data, sensor data, process condition data at a previous time, and a previous time point. It may be actual measurement data or the like.

The virtual measurement apparatus 100 generates a learning model through learning for virtual measurement, and generates a virtual measurement result including a predicted value and feature importance by using the generated learning model. Specifically, the virtual measurement device 100 extracts a feature value from time series data and calculates the final predicted value and the importance of each feature value by using a learning model learned through the operation of calculating the predicted value for the feature value Performs an operation that generates a measurement result.

The virtual measurement apparatus 100 provides the generated virtual measurement result to the process and measurement apparatus 200 .

The process and measurement device 200 periodically performs actual measurement when the process is in progress, and if the actual measurement value is within a preset reference value, the process proceeds under the current process conditions, and the actual measurement value deviates from the preset reference value or a predetermined value If it is determined that a change in process conditions is necessary according to the standards of

The process and measurement apparatus 200 receives a virtual measurement result from the virtual measurement apparatus 100 , and performs the process by adjusting a reference value or process condition based on the virtual measurement result.

As the process and measurement device 200 receives the virtual measurement result, sampling measurement through actual measurement may be minimized, and product quality may be improved and manufacturing cost may be reduced according to process control.

2 is a block diagram schematically illustrating a virtual measurement apparatus according to an embodiment of the present invention.

The virtual measurement apparatus 100 according to the present embodiment includes an input unit 110 , an output unit 120 , a processor 130 , a memory 140 , and a database 150 . The virtual measurement apparatus 100 of FIG. 2 is according to an embodiment, and not all blocks shown in FIG. 2 are essential components, and in another embodiment, some blocks included in the virtual measurement apparatus 100 are added or changed. Or it can be deleted. Meanwhile, the virtual measurement apparatus 100 may be implemented as a computing device, and each component included in the virtual measurement apparatus 100 may be implemented as a separate software program or as a separate hardware device combined with software. can

The virtual measurement device 100 extracts a feature value from time series data and calculates the final predicted value and the importance for each feature value using the learning model learned through the operation of calculating the predicted value for the feature value, thereby obtaining the virtual measurement result Execute the creation action.

The input unit 110 refers to a means for inputting or acquiring a signal or data for performing a virtual measurement operation in the virtual measurement device 100 . The input unit 110 may interwork with the processor 130 to input various types of signals or data, or may obtain signals or data through interworking with an external device and transmit the signals or data to the processor 130 . Here, the input unit 110 may be implemented as a module for inputting input variables, time series data, learning conditions, virtual measurement performance conditions, etc., but is not necessarily limited thereto.

The output unit 120 may output various information such as a learning result and a virtual measurement result in conjunction with the processor 130 . The output unit 120 may output various information through a display (not shown) provided in the virtual measurement device 100 , but is not limited thereto, and a method of interworking with the process and measurement device 200 or a user terminal Output can be performed in various types of methods, such as

The processor 130 performs a function of executing at least one instruction or program included in the memory 140 .

The processor 130 according to this embodiment extracts a feature value for the time series data obtained from the input unit 110 or the database 150, and calculates a prediction value for learning based on the feature value to generate a learning model. carry out

In addition, the processor 130 calculates a feature value for the new time series data using the pre-learned learning model, and calculates a virtual measurement result including a final predicted value and a feature importance for each feature value based on the feature value perform the action A detailed operation of the processor 130 according to the present embodiment will be described with reference to FIGS. 3 to 8 .

The memory 140 includes at least one instruction or program executable by the processor 130 . The memory 140 may include instructions or programs for generating a learning model, generating a virtual measurement result, and the like.

The database 150 means a general data structure implemented in the storage space (hard disk or memory) of a computer system using a database management program (DBMS), and performs data search (extraction), deletion, editing, addition, etc. Relational database management system (RDBMS) such as Oracle, Infomix, Sybase, DB2, Gemston, Orion ), an object-oriented database management system (OODBMS) such as O2, and an XML Native Database such as Excelon, Tamino, Sekaiju, etc. The purpose of an embodiment of the present invention It can be implemented according to the requirements, and has appropriate fields or elements to achieve its function.

The database 150 according to the present embodiment may store data related to input variables, learning model generation, virtual measurement result calculation, and the like, and may provide data related to pre-stored input variables, learning model generation, virtual measurement result calculation, and the like.

The input variable stored in the database 150 may be time series data, sensor data, process condition data of a previous time, actual measurement data of a previous time, and the like. Although the database 150 is described as being implemented in the virtual measurement device 100, it is not necessarily limited thereto, and may be implemented as a separate data storage device.

3 and 4 are block diagrams for explaining a processor operation of a virtual measurement apparatus according to an embodiment of the present invention.

The processor 130 of the virtual measurement apparatus 100 generates a learning model for virtual measurement and performs virtual measurement based on a pre-learned learning model. Referring to FIG. 3 , the processor 130 of the virtual measurement device 100 learns, predicts, and performs virtual measurement using a deep neural network configuration 310 , a white box algorithm configuration 320 , and a white box analysis configuration 330 , etc. Executes an operation such as calculating a result.

The deep neural network configuration 310 operates based on a convolutional neural network (CNN). For example, the deep neural network configuration 310 may input time series data to the convolutional neural network to calculate a feature value and a neural network-based first predicted value.

The white box algorithm configuration 320 and the white box interpretation configuration 330 operate based on the classification prediction model and the regression prediction model based on the white box algorithm. For example, the white box algorithm configuration 320 and the white box analysis configuration 330 may calculate predicted values, feature importance, etc. based on a regression decision tree method.

Referring to FIG. 4 , the processor 130 of the virtual measurement apparatus 100 according to the present embodiment includes a time series data acquisition unit 410 , a model learning processing unit 420 , a virtual measurement processing unit 430 , and a virtual measurement result output unit. 440 may be included. In FIG. 4 , the processor 130 of the virtual measurement device 100 is according to an embodiment, and not all blocks shown in FIG. 4 are essential components, and in another embodiment, some blocks included in the processor 130 are It may be added, changed or deleted. On the other hand, the virtual measurement apparatus 100 may be implemented as a computing device, and each component included in the processor 130 of the virtual measurement apparatus 100 may be implemented as a separate software device, or a separate software device combined with software. It may be implemented as a hardware device.

The time series data obtaining unit 410 obtains time series data including time series type variables.

The time series data obtaining unit 410 may obtain time series data including explanatory variables and response variables in the form of time series.

The time series data may be acquired from an external device (not shown), a separate storage unit (not shown), or acquired by a user's manipulation. The explanatory variables included in the time series data may be univariate or multivariate, and the response variable may be continuous or categorical.

In addition, the time series data may include information on the number of training data, the length of time series data, the number of time series variables, and the like.

The model learning processing unit 420 performs an operation of generating a learning model for virtual measurement. Specifically, the model learning processing unit 420 generates a learning model by extracting a feature value for the time series data, and calculating a prediction value for learning based on the feature value.

Some operations for generating the learning model in the model learning processing unit 420 may be processed through a deep neural network, and other operations may be processed through a white box algorithm.

A detailed description of the model learning processing unit 420 will be described in detail with reference to FIG. 5 .

The virtual measurement processing unit 430 performs virtual measurement using the learning model learned by the model learning processing unit 420 . Specifically, the virtual measurement processing unit 430 calculates a virtual measurement result by applying the new time series data to the learning model.

In the virtual metrology processing unit 430 , some operations for virtual measurement may be processed through a deep neural network, and other operations may be processed through a white box algorithm.

A detailed description of the virtual measurement processing unit 430 will be described in detail with reference to FIG. 7 .

The virtual measurement result output unit 440 performs an operation of providing the calculated virtual measurement result to an external device.

The virtual measurement result output unit 440 may provide the virtual measurement result to the process and measurement device 200 , but is not limited thereto, and may be a separate control device or user terminal interworking with the process and measurement device 200 . Virtual measurement results can be provided.

5 is a block diagram schematically showing a model learning processing unit according to an embodiment of the present invention.

The model learning processing unit 420 according to the present embodiment includes a feature value extraction unit 510 , a first prediction value calculation unit 520 , a second prediction value calculation unit 530 , a prediction value calculation unit for learning 540 , and an error value calculation unit 550 and a learning model generator 560 . The model learning processing unit 420 of FIG. 5 is according to an embodiment, and not all blocks shown in FIG. 5 are essential components, and in another embodiment, some blocks included in the model learning processing unit 420 are added or changed. Or it can be deleted. Meanwhile, each component included in the model learning processing unit 420 may be implemented as a separate software program, or may be implemented as a separate hardware device combined with software.

The model learning processing unit 420 generates a learning model by extracting feature values for time series data, and calculating prediction values for learning based on the feature values. Hereinafter, each of the components included in the model learning processing unit 420 will be described.

The feature value extraction unit 510 extracts a feature value for time series data through a neural network.

The feature value extraction unit 510 extracts feature values by passing the time series data through a convolution layer included in the neural network. Here, the feature value is preferably in the form of a vector.

The prediction value calculators 520 , 530 , and 540 calculate prediction values for learning based on the feature values.

The first predicted value calculating unit 520 performs an operation of calculating a first predicted value based on the feature value based on the neural network.

The second predicted value calculator 530 calculates a second predicted value by substituting the feature value into the white box algorithm.

The prediction value calculation unit 540 for learning calculates a similarity by comparing the first predicted value with the second predicted value, and calculates a predicted value for learning by assigning a weight according to the similarity.

The error value calculator 550 calculates an error value for the prediction value for learning.

The learning model generator 560 generates a learning model for virtual measurement based on the predicted value for learning.

The learning model generator 560 may generate a learning model based on the prediction value for learning when the process of calculating an error value for the prediction value for learning is omitted or not performed.

On the other hand, when the process of calculating the error value for the prediction value for learning is performed, the learning model generator 560 compares the error value with a preset threshold to generate the learning model.

The learning model generator 560 compares the error value with a preset threshold, and repeats the operation of updating the neural network for extracting the feature value when the error value is greater than or equal to the threshold.

The learning model generator 560 compares the error value with a preset threshold, and when the error value is less than the threshold, generates a learning model based on the corresponding learning prediction value.

6 is a flowchart illustrating a virtual measurement method for model learning processing according to an embodiment of the present invention.

The virtual measurement device 100 acquires time series data (S610).

The virtual measurement apparatus 100 extracts a feature value of the time series data through a neural network (S620).

The virtual measurement apparatus 100 calculates a first predicted value based on the feature value based on the neural network (S630).

The virtual measurement apparatus 100 calculates a second predicted value by applying the feature value to the white box algorithm (S640).

The virtual measurement device 100 compares the first predicted value and the second predicted value to calculate a learning predicted value ( S650 ). The virtual measurement apparatus 100 calculates a similarity by comparing the first predicted value with the second predicted value, and calculates a predicted value for learning by assigning a weight according to the similarity.

The virtual measurement device 100 calculates an error value for the prediction value for learning ( S660 ).

The virtual measurement apparatus 100 compares the calculated error value with a preset threshold (S670).

In step S670, the virtual measurement apparatus 100 compares the error value with a preset threshold, and repeats the operation of updating the neural network for extracting the feature value when the error value is greater than or equal to the threshold (S672). That is, when the error value is equal to or greater than the threshold value, the virtual measurement apparatus 100 performs backpropagation and returns to step S620 to repeat the learning operation.

In step S670, the virtual measurement device 100 compares the error value with a preset threshold, ends the learning when the error value is less than the threshold, and generates a learning model based on the predicted value for learning (S680).

Although it is described that each step is sequentially executed in FIG. 6 , the present invention is not limited thereto. In other words, since it may be applicable to changing and executing the steps described in FIG. 6 or executing one or more steps in parallel, FIG. 6 is not limited to a chronological order.

The virtual measurement method for model learning processing according to the present embodiment described in FIG. 6 may be implemented as an application (or program) and recorded in a recording medium readable by a terminal device (or computer). An application (or program) for implementing a virtual measurement method for model learning processing according to this embodiment is recorded and a terminal device (or computer) readable recording medium is all of the data that can be read by the computing system are stored Recording devices or media of any kind.

7 is a block diagram schematically illustrating a virtual measurement processing unit according to an embodiment of the present invention.

The virtual measurement processing unit 430 according to the present embodiment includes a feature value extraction unit 710 , a learning model processing unit 720 , a final predicted value calculation unit 730 , and an importance calculation unit 740 .

The virtual metrology processing unit 430 of FIG. 7 is according to an embodiment, and not all blocks shown in FIG. 7 are essential components, and in another embodiment, some blocks included in the virtual metrology processing unit 430 are added or changed. Or it can be deleted. Meanwhile, each component included in the virtual measurement processing unit 430 may be implemented as a separate software program or as a separate hardware device combined with software.

The virtual measurement processing unit 430 calculates a virtual measurement result by applying the new time series data to the learning model. Hereinafter, each of the components included in the virtual measurement processing unit 430 will be described.

When new time series data is obtained, the feature value extraction unit 710 extracts a feature value for the new time series data.

The virtual measurement result calculation units 720 , 730 , and 740 calculate a virtual measurement result by using a pre-learned learning model.

The learning model processing unit 720 applies the feature value of the new time series data to the learning model based on the white box algorithm to generate a virtual measurement result.

The learning model processing unit 720 stores the learning model 422 learned by the model learning processing unit 420 , and generates an output value by applying a feature value of the new time series data as an input value. Here, the output value may be transmitted to at least one of the final predicted value calculator 730 and the importance calculator 740 .

The final predicted value calculating unit 730 applies the feature value of the new time series data to the white box algorithm-based learning model to calculate the final predicted value, and generates a virtual measurement result including the calculated final predicted value.

The importance calculator 740 applies the feature value for the new time series data to a white box algorithm-based learning model to calculate the feature importance according to the error reduction rate for the feature value, and virtual measurement including the calculated feature importance produce results.

8 is a flowchart illustrating a virtual measurement method using a learning model according to an embodiment of the present invention.

The virtual measurement device 100 acquires time series data (S810). Here, the time series data may be new time series data different from the time series data of the learning process.

The virtual measurement apparatus 100 extracts a feature value for the time series data (S820). Here, the virtual measurement apparatus 100 may calculate a feature value for the new time series data through the neural network included in the pre-trained learning model.

The virtual measurement apparatus 100 calculates a final predicted value through the pre-trained learning model (S830). The virtual measurement device 100 may transmit a virtual measurement result including the calculated final predicted value to an external device.

The virtual measurement apparatus 100 calculates the feature importance of the feature value through the pre-learned learning model (S840). The virtual measurement device 100 may transmit a virtual measurement result including the calculated feature importance to an external device.

Steps S830 and S840 are described as being sequentially performed, but are not necessarily limited thereto, and only one of steps S830 and S840 may be performed to generate a virtual measurement result.

Although it is described that each step is sequentially executed in FIG. 8 , the present invention is not limited thereto. In other words, since it may be applicable to changing and executing the steps described in FIG. 8 or executing one or more steps in parallel, FIG. 8 is not limited to a chronological order.

The virtual measurement method using the learning model according to the present embodiment described in FIG. 8 may be implemented as an application (or program) and recorded in a recording medium readable by a terminal device (or computer). The recording medium in which the application (or program) for implementing the virtual measurement method using the learning model according to this embodiment is recorded and the terminal device (or computer) can read is any type of data that can be read by the computing system is stored of recording devices or media.

9A and 9B are exemplary diagrams for explaining a model learning operation of a virtual measurement device according to an embodiment of the present invention.

9A is a diagram for explaining an operation of generating a learning model by calculating a prediction value for learning in the virtual measurement device 100 .

In step S910 , the virtual measurement device 100 acquires time series data. Step S910 may match the operation of the time series data acquisition unit 410 .

In step S910, the time series data may include raw data (911) for explanatory variables in the form of (n × t × m), and label data for response variables in the form of (n × 1). there is. Here, n is the number of training data, t is the length of time series data, and m is the number of time series variables. For example, the time series data may include (100×64×1) raw data and (100×1) label data.

In step S920 , the virtual measurement apparatus 100 extracts a feature value for time series data through a neural network. Step S920 may match the operation of the feature value extraction unit 510 .

In step S920, the neural network includes a first convolution layer, a first activation layer, a first pooling layer, a second convolution layer, a second activation layer, a second pooling layer, and Including a flatten layer (Flatten layer).

The output of the first activation layer (block 921) may be defined as [Equation 1].

(n: number of training data, in: input data size, k: kernel size, s: stride size, f: number of filters)

The output of the first pooling layer (block 922) may be defined as [Equation 2].

(n: number of training data, in: input data size, p: pool size, f: number of filters)

The output of the platen layer may be defined as [Equation 3].

값, f: 필터 개수)(n: the number of training data, o: the number of previous nodes

value, f: number of filters)

For example, assuming that the user setting values of the neural network are stride=2, kernel size=2, filter size=8, pool size=2, the output of the first activation layer is a value of the form (100 × 32 × 8) , the output of the first pooling layer is a value of the form (100 × 16 × 8), the output of the second activation layer (block 923) is a value of the form (100 × 4 × 8), and the second pooling layer (block 923) has a value of the form 924) may be a value of the form (100 × 8 × 8). Also, the output of the platen layer (block 925) may be a value (feature value) of the form (100×64).

In operation S930, the virtual measurement apparatus 100 calculates a first predicted value based on the feature value based on the neural network. Step S930 may match the operation of the first prediction value calculator 520 .

In step S930, the neural network further includes a fully connected layer and an output layer. The output layer outputs a first prediction value, and the first prediction value may be a value in the form of (100 × 1).

In operation S940, the virtual measurement apparatus 100 calculates a second predicted value by applying the feature value to the white box algorithm. Step S940 may match the operation of the second prediction value calculator 530 .

In step S940, the virtual metrology device 100 receives a feature value in the form of (100 × 64) that is the output of the platen layer, generates a sample data set through bootstrap, and applies it to the white box algorithm (block 942, block 944). Here, bootstrap refers to a method of creating a data set through random sampling that allows duplication. Here, the data set may be a value in the form of (100 × 64).

In step S940, an output that has passed through the white box algorithm may be different depending on the type of the white box algorithm. For example, an output that has passed through the white box algorithm may include target type data, model information required for prediction, and the like.

In operation S940, a second predicted value may be calculated by substituting a feature value in the form of (100×64), which is the output of the platen layer, into the learned white box algorithm. Here, the second prediction value may be a value of the form (100×1) (block 946).

In steps S950 and S960 , the virtual measurement device 100 calculates a prediction value for learning by comparing the first predicted value and the second predicted value. Specifically, the virtual measurement apparatus 100 calculates a similarity by comparing the first predicted value with the second predicted value, and calculates the predicted value for learning by assigning a weight according to the similarity. Steps S950 and S960 may match the operation of the prediction value calculation unit 540 for learning.

In operation S950 , the virtual measurement apparatus 100 may calculate a similarity by comparing the first predicted value and the second predicted value based on a Pearson correlation coefficient (block 952 ). Here, the output value of step S950 may be a value in the form of (100 × 1).

In step S960, the virtual measurement device 100 assigns a preset weight w according to the degree of similarity (block 962), and calculates a prediction value for learning (block 964). Here, the weight w refers to a scaled value such that the sum of the weights w becomes 1, and the output value may be a value in the form of (100 × 1).

In step S960, the virtual measurement device 100 preferably assigns a weight to the first predicted value, but is not limited thereto, and assigns a weight to the second predicted value, or applies a weight to the average value of the first predicted value and the second predicted value. It is also possible to calculate a prediction value for learning by giving it.

9B is a diagram for explaining an operation of generating a learning model by calculating an error value after calculating a prediction value for learning in the virtual measurement device 100 .

Since operations S910 to S960 of FIG. 9B are the same as those of FIG. 9A , descriptions of overlapping descriptions will be omitted.

In step S970 , the virtual measurement device 100 calculates an error value for the prediction value for learning. Step S970 may match the operation of the error value calculator 550 .

In step S970 , the virtual measurement device 100 calculates an error value by calculating a loss function using the predicted value for learning and the value (actual value) of the label data included in the time series data. In step S970, it is described that a loss function is used for calculating an error value, but the present invention is not limited thereto, and a root mean square error (RMSE) may be used in the regression method.

In step S980 , the virtual measurement apparatus 100 compares the calculated error value with a preset threshold ε, and generates a learning model based on the comparison result. Step S980 may match the operation of the learning model generator 560 .

In step S980 , the virtual measurement apparatus 100 compares the error value with a preset threshold and repeats the operation of updating the neural network for extracting the feature value when the error value is greater than or equal to the threshold. That is, when the error value is equal to or greater than the threshold value, the virtual measurement apparatus 100 performs backpropagation to update the neural network in the order of steps S930 and S920. Here, the operation of updating the neural network is repeated until it is determined that the error value is smaller than the threshold ε.

In step S980 , the virtual measurement apparatus 100 compares the error value with a preset threshold, ends learning when the error value is less than the threshold, and generates a learning model based on the corresponding learning prediction value.

10A and 10B are exemplary diagrams for explaining an operation of calculating a predicted value and a feature importance of a virtual measurement device according to an embodiment of the present invention.

FIG. 10A is a diagram for explaining an operation of calculating a final predicted value using a pre-trained learning model in the virtual measurement device 100 .

In step S1010 , the virtual measurement device 100 acquires new time series data. Step S1010 may match the operation of the time series data acquisition unit 410 .

In step S1010 , the time series data may include raw data for an explanatory variable in the form of (n × t × m) and label data for a response variable in the form of (n × 1). Here, n is the number of training data, t is the length of time series data, and m is the number of time series variables. For example, the time series data may include raw data in the form of (100 × 64 × 1) and label data in the form of (100 × 1).

In step S1020 , the virtual measurement apparatus 100 extracts a feature value for new time series data through a neural network. Here, the virtual measurement apparatus 100 may calculate a feature value for the new time series data through the neural network included in the pre-trained learning model. Step S1020 may match the operation of the feature value extraction unit 710 .

In steps S1030 and S1040, the virtual measurement device 100 inputs the calculated feature values to the pre-learned learning model, and calculates a final predicted value through the pre-learned learning model. Steps S1030 and S1040 may match operations of the learning model processing unit 720 and the final prediction value calculating unit 730 .

In step S1040, the virtual measurement device 100 calculates and outputs a final predicted value having the same type and length as the label data of the time series data. For example, in step S1040, the virtual measurement device 100 has a (100 × 1) form output the final predicted value of .

FIG. 10B is a diagram for explaining an operation of calculating feature importance by using a pre-learned learning model in the virtual measurement apparatus 100 .

of Figure 10b Operations of S1010 to S1030 are the same as those of FIG. 10A , so description of overlapping descriptions will be omitted.

In steps S1030 and S1050, the virtual measurement device 100 inputs the calculated feature values to the pre-trained learning model, and calculates the feature importance of the feature values through the pre-learned learning model. Steps S1030 and S1050 may match operations of the learning model processing unit 720 and the importance calculating unit 740 .

In operation S1050 , the virtual measurement apparatus 100 may determine the magnitude of the error reduced by the corresponding variable during prediction as the explanatory power of the corresponding variable, that is, the feature value, and calculate the feature importance according to the explanatory power.

For example, in operation S1050 , the virtual measurement apparatus 100 may calculate a feature value and feature importance as shown in [Table 1].

In FIG. 10B , step S1050 is described as being processed after step S1030, but is not limited thereto, and may be additionally performed after step S1040 of FIG. 10A is performed.

The above description is merely illustrative of the technical idea of the embodiment of the present invention, and those of ordinary skill in the art to which the embodiment of the present invention pertains may modify various modifications and transformation will be possible. Accordingly, the embodiments of the present invention are not intended to limit the technical spirit of the embodiment of the present invention, but to explain, and the scope of the technical spirit of the embodiment of the present invention is not limited by these embodiments. The protection scope of the embodiment of the present invention should be construed by the following claims, and all technical ideas within the scope equivalent thereto should be construed as being included in the scope of the embodiment of the present invention.

10: Process control system
100: virtual metrology device 200: process and metrology device
110: input unit 120: output unit
130: processor 140: memory
150: database
410: time series data acquisition unit 420: model learning processing unit
430: virtual measurement processing unit 440: virtual measurement result output unit

Claims (13)

A method for performing virtual measurement in a virtual measurement device, the method comprising:
a time series data obtaining step of obtaining time series data including time series type variables;
a model learning processing step of extracting a feature value for the time series data, calculating a prediction value for learning based on the feature value, and generating a learning model;
a virtual measurement processing step of calculating a virtual measurement result by applying new time series data to the learning model; and
Virtual measurement result output step of providing the virtual measurement result to an external device
Virtual measurement method comprising a. The method of claim 1,
The time series data acquisition step includes:
Acquire time series data including explanatory variables and response variables in a time series form, wherein the explanatory variable is a variable for univariate or multivariate, and the response variable is a variable for continuous or categorical type,
The time series data is a virtual measurement method, characterized in that it comprises at least one information of the number of training data, time series data length, and the number of time series variables. The method of claim 1,
The model learning processing step is,
a feature value extraction step of extracting feature values for the time series data through a neural network;
a prediction value calculation step of calculating a prediction value for learning based on the feature value; and
A learning model generation step of generating a learning model for virtual measurement processing based on the prediction value for learning
Virtual measurement method comprising a. 4. The method of claim 3,
The prediction value calculation step is,
a first predicted value calculating step of calculating a first predicted value based on the feature value based on the neural network;
a second predicted value calculating step of calculating a second predicted value by substituting the feature value into a white box algorithm; and
A learning prediction value calculation step of calculating a similarity by comparing the first predicted value with the second predicted value, and calculating a learning predicted value by assigning a weight according to the similarity
Virtual measurement method comprising a. 4. The method of claim 3,
The model learning processing step is,
Further comprising an error value calculation step of calculating an error value for the prediction value for learning,
The learning model creation step is,
Comparing the error value with a preset threshold, repeating the operation of updating the neural network for extracting the feature value when the error value is greater than or equal to the threshold, and generating the learning model when the error value is less than the threshold Virtual measurement method characterized. The method of claim 1,
The virtual measurement processing step includes:
a feature value extraction step of extracting a feature value for the new time series data when new time series data is obtained;
A virtual measurement result calculation step of calculating a virtual measurement result using the learning model
Virtual measurement method comprising a. 7. The method of claim 6,
The virtual measurement result calculation step includes:
A virtual measurement method, characterized in that by applying the feature value of the new time series data to the learning model based on a white box algorithm, calculating a final predicted value, and generating the virtual measurement result including the calculated final predicted value. 7. The method of claim 6,
The step of calculating the virtual measurement result,
By applying the feature value for the new time series data to the learning model based on the white box algorithm, calculating the feature importance according to the error reduction rate for the feature value, and generating the virtual measurement result including the calculated feature importance Virtual measurement method, characterized in that. A device for performing virtual metrology, comprising:
at least one processor; and a memory storing one or more programs executed by the processor, wherein the programs, when executed by the one or more processors, in the one or more processors;
a time series data obtaining step of obtaining time series data including time series type variables;
a model learning processing step of extracting a feature value for the time series data, calculating a prediction value for learning based on the feature value, and generating a learning model;
a virtual measurement processing step of calculating a virtual measurement result by applying new time series data to the learning model; and
Virtual measurement result output step of providing the virtual measurement result to an external device
A virtual metrology device, characterized in that it performs operations comprising: 10. The method of claim 9,
The model learning processing step is,
a feature value extraction step of extracting feature values for the time series data through a neural network;
a prediction value calculation step of calculating a prediction value for learning based on the feature value;
an error value calculation step of calculating an error value for the learning prediction value; and
Comparing the error value with a preset threshold, and repeating the operation of updating the neural network for extracting the feature value when the error value is greater than or equal to the threshold, and when the error value is less than the threshold, a learning model for virtual measurement processing Steps to create a learning model that creates
Virtual measurement device comprising a. 11. The method of claim 10,
The predicted value calculation step is,
a first predicted value calculating step of calculating a first predicted value based on the feature value based on the neural network;
a second predicted value calculating step of calculating a second predicted value by substituting the feature value into a white box algorithm; and
A learning prediction value calculation step of calculating a similarity by comparing the first predicted value with the second predicted value, and calculating a learning predicted value by assigning a weight according to the similarity
Virtual measurement device comprising a. 10. The method of claim 9,
The virtual measurement processing step includes:
a feature value extraction step of extracting a feature value for the new time series data when new time series data is obtained;
A virtual measurement result calculation step of calculating a virtual measurement result using the learning model
Virtual measurement device comprising a. 13. The method of claim 12,
The step of calculating the virtual measurement result,
An operation of calculating a final predicted value by applying the feature value of the new time series data to the learning model based on the white box algorithm, and applying the feature value of the new time series data to the learning model based on the white box algorithm, the feature value and generating the virtual measurement result including at least one of the final predicted value and the feature importance by performing at least one of the operations of calculating the feature importance according to the error reduction rate.

Images