Virtual measurement method using convolutional neural network

Abstract

An embodiment of the present invention provides a virtual measurement method using a convolutional neural network (CNN). In this method, a dynamic time warping (DTW) algorithm is used to delete dissimilar process data and adjust the length of the process data to the same length so that the convolutional neural network can be used for virtual measurement. The virtual measurement model of the embodiment of the present invention includes multiple convolutional neural network models and an estimation model, wherein the multiple inputs of these convolutional neural network models are respectively the time series data of each parameter of the process data, and the multiple outputs of these convolutional neural network models are the inputs of the estimation model.

Description

Technical Field

[0001] The present invention relates to a virtual measurement method using a convolutional neural network (CNN), and more particularly to a virtual measurement method using a convolutional neural network using a dynamic time warping (DTW) algorithm. Background Art

[0002] Virtual measurement has a wide range of applications in various industries, such as the semiconductor industry and the machine tool industry. Virtual measurement can transform offline and delayed quality sampling into online and real-time quality inspection to ensure the process stability of production machines and improve production capacity and yield. For example, when virtual measurement is introduced into the wafer process of the semiconductor industry, process anomalies can be discovered immediately and immediate improvements can be made, thus avoiding the subsequent scrapping of the entire batch of wafers. When virtual measurement is introduced into machine tool processing, the accuracy value of each processed workpiece (such as wheel rims) can be estimated by taking into account both immediacy and accuracy requirements, thereby predicting the processing quality of the machine tool to overcome the known shortcomings of online and offline measurement.

[0003] Although existing virtual measurement is generally suitable for its intended purpose, it has not yet met the needs of all aspects. Summary of the invention

[0004] The object of the present invention is to provide a virtual measurement method using a convolutional neural network and a computer program product thereof, so as to effectively use the convolutional neural network to perform virtual measurement to improve the accuracy of known virtual measurement.

[0005] According to one aspect of the present invention, a virtual measurement method using a convolutional neural network is provided. In this method, first, multiple sets of process data are obtained, wherein these sets of process data are used or generated when a production machine processes multiple workpieces, and these sets of process data correspond to these workpieces in a one-to-one manner, and each set of process data includes the values ​​of multiple parameters, and the value of each parameter is a plurality of sets of timing data corresponding to these workpieces. Then, a data alignment operation is performed on these sets of process data. In this data alignment operation, a frequency distribution (Frequency Distribution) calculation is first performed on a data length of these sets of timing data of each parameter, and a distribution of the number of data occurrences to the data length is obtained, wherein the data length with the largest number of occurrences in these sets of timing data of each parameter is a reference data length. Then, the timing data with this reference data length in these sets of timing data of each parameter is averaged, and a set of reference timing data for each parameter is obtained. Next, a dynamic time warping (DTW) algorithm is used to calculate a distance between each set of timing data of each parameter and its corresponding reference timing data. Then, a distance threshold value is set. When the aforementioned distance is greater than the distance threshold value, the process data corresponding to this distance is deleted. Next, a data length adjustment operation is performed to repeatedly add at least one data point having the value of the last data point after a last data point of each set of timing data of each parameter until the data length of each set of timing data of each parameter is equal to a longest data length of the sets of process data. After obtaining a plurality of actual measurement values ​​of these workpieces, a modeling operation is performed to use these sets of process parameter data and these actual measurement values ​​to establish a virtual measurement model, wherein the virtual measurement model includes at least one convolutional neural network model. Then, after performing the aforementioned data alignment operation on another set of process data of another workpiece, the another set of process data of this another workpiece is input into the virtual measurement model to calculate a virtual measurement value of this another workpiece.

[0006] In some embodiments, the aforementioned virtual measurement method using a convolutional neural network further includes: setting a data length upper limit value before performing the aforementioned data length adjustment operation; and deleting process data in the aforementioned process data whose data length is greater than the data length upper limit value.

[0007] In some embodiments, the aforementioned upper limit value of the data length is Q3+k×IQR, where Q3 is the third quartile of the order of occurrence of data length from small to large in the distribution of the aforementioned data occurrence frequency to data length; IQR is the quartile of the order of occurrence of data length from small to large in the distribution of the data occurrence frequency to data length; k is a constant greater than 0.

[0008] In some embodiments, the aforementioned operation of setting the distance threshold value is to apply the Leave-One-Out (LOO) principle in Cross Validation.

[0009] In some embodiments, the aforementioned virtual measurement model includes multiple convolutional neural network models and an estimation model, the multiple inputs of these convolutional neural network models are respectively the time series data of the aforementioned parameters, and the multiple outputs of these convolutional neural network models are the inputs of this estimation model.

[0010] In some embodiments, the aforementioned estimation model is established based on a neural network-like algorithm, a complex regression algorithm, a partial least squares algorithm, or a support vector machine algorithm.

[0011] According to one aspect of the present invention, a computer program product is provided. When a computer loads and executes the computer program product, the aforementioned virtual measurement method using a convolutional neural network can be completed.

[0012] Therefore, by applying the embodiments of the present invention, the convolutional neural network can be effectively used to perform virtual measurement, thereby improving the accuracy of the conventional virtual measurement. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] For a more complete understanding of the embodiments and their advantages, reference is now made to the following description taken in conjunction with the accompanying drawings, in which:

[0014] Figure 1 The raw process data for illustrating the pressure parameters of the etching process are shown in the figure;

[0015] Figure 2 A block diagram illustrating a virtual measurement application according to some embodiments of the present invention;

[0016] Figure 3A A flowchart of a virtual measurement method based on using an automatic data alignment mechanism (ADAS) 108 according to some embodiments of the present invention is illustrated;

[0017] Figure 3B The data length of the process data of each workpiece in an application example of the present invention is shown;

[0018] Figure 3C Plotting the distribution of the number of process occurrences of each workpiece in this application example versus the data length

[0019] Figure 4 A schematic diagram illustrating an integrated virtual measurement model according to some embodiments of the present invention;

[0020] Figure 5A To illustrate the prediction results of the etching sidewall angle of the application example of the present invention; and

[0021] Figure 5B The figure shows the prediction result of etching depth according to the application example of the present invention.

[0022]

Explanation of symbols

[0023] 10: Curve

[0024] 20: Curve

[0025] 30: Curve

[0026] 42: Curve

[0027] 44: Curve

[0028] 46: Curve

[0029] 52: Curve

[0030] 54: Curve

[0031] 56: Curve

[0032] 102: Process data

[0033] 104: Measurement data

[0034] 106: Process data pre-processing module

[0035] 108:ADAS

[0036] 110: Virtual Measurement Model

[0037] 112: Measurement data pre-processing module

[0038] 120:RI Model

[0039] 130:GSI Model

[0040] 300: Operation

[0041] 302: Operation

[0042] 310: Operation

[0043] 320: Operation

[0044] 330: Operation

[0045] 332: Operation

[0046] 334: Operation

[0047] 340: Operation

[0048] 342: Operation

[0049] 350: Operation

[0050] 360: Operation

[0051] 400: Estimation Model DETAILED DESCRIPTION

[0052] As the process becomes more and more complex, the requirements for virtual measurement accuracy are getting higher and higher. Known virtual measurement uses the Back Propagation Neural Network (BPNN) algorithm. However, even if the number of modeling sample data grows significantly, the performance improvement of known virtual measurement is still limited. The BPNN algorithm also requires a strict and time-consuming feature selection process. On the other hand, when the number of modeling sample data grows, the accuracy of the convolutional neural network (CNN) algorithm will improve accordingly. If there is enough modeling sample data, the performance of the virtual measurement established by the CNN algorithm will be greatly improved. In addition, the CNN algorithm can automatically extract highly informative data features from the sample data. Therefore, when the CNN algorithm is used for virtual measurement, it can not only improve the accuracy of virtual measurement, but also save time and labor costs for feature selection. However, the input values ​​of the CNN algorithm must have the same data length and similar time distribution shape. When using the CNN algorithm for virtual measurement, it is necessary to first overcome the problem of dissimilar time distribution shape and different data length of the input process data.

[0053] The embodiment of the present invention uses multiple sets of modeling samples and establishes a virtual measurement model based on a convolutional neural network algorithm. Each set of modeling samples includes a set of process data and an actual measurement value. The process data is used or generated when the production machine processes the workpiece, and the process data corresponds to the workpiece in a one-to-one manner. Each set of process data includes the values ​​of multiple parameters, and the value of each parameter is a plurality of sets of timing data corresponding to the workpiece respectively. The actual measurement value is obtained after the measurement machine measures the quality items of the workpiece. For wafer processing, the production machine is a wafer processing machine, such as an etching machine, a deposition machine, a sputtering machine, etc.; the actual measurement value is the film thickness, etching depth, etching sidewall angle, etc.; the process data includes temperature. For wafer dicing process, the production machine is the wafer dicing machine; the actual measurement value is the wafer chipping amount; the process data includes blade clogging, coolant flow rate, spindle speed, feed rate, wafer condition (such as thickness, coating, etc.), and / or kerf width. For machine tool processing, the production machine is the machine tool; the actual measurement value includes roughness, straightness, angularity, perpendicularity, parallelism and / or roundness; the process data includes operating current, vibration and / or audio data obtained by a detector installed on the machine tool (for example: a three-axis accelerometer and / or an acoustic emission (AE) sensor).

[0054] Each set of timing data of each parameter of each set of process data is obtained by detecting a workpiece with a sensor at a sampling rate during a process. Figure 1 , Figure 1 The original process data of the pressure parameters of the etching process shown in the figure, wherein the curve 10 marked with a triangle mark represents the time series data of the pressure (parameter) in the reaction chamber when the workpiece sample 1 is etched; the curve 20 marked with a square mark represents the time series data of the pressure (parameter) in the reaction chamber when the workpiece sample 2 is etched; the curve 30 marked with an x ​​mark represents the time series data of the pressure (parameter) in the reaction chamber when the workpiece sample 3 is etched. Figure 1 As shown in the area A1 and the enlarged area A2, even if the etching process has the same recipe, the lengths of the time series data of the pressure (parameter) of the workpiece samples 1, 2, and 3 are still different and cannot be used for modeling. Figure 1As shown in area B, at the beginning of the process, the initial pressure of workpiece sample 3 (curve 30) is not as high as the initial pressure of workpiece samples 1 and 2 (curves 10 and 20). This is because the sensor delay when etching workpiece sample 3 causes the data of the first two seconds to be not reported when collecting workpiece sample 3. Since the time distribution shape of the process data of workpiece sample 3 is not similar to the time distribution shape of the process data of workpiece samples 1 and 2, if the process data of workpiece sample 3 is still used for modeling, it will pollute the model and affect the accuracy of the model.

[0055] Therefore, an embodiment of the present invention provides an automatic data alignment scheme (ADAS) to delete process data with dissimilar time distribution shapes and make the data lengths of the process data the same.

[0056] The virtual measurement, RI (Reliance Index; confidence index) model, GSI (Global Similarity Index; overall similarity index) model, DQI X (Process Data Quality Index; process data quality index), DQI y The Metrology Data Quality Index (MDQI) model and the dual-stage virtual measurement mechanism can refer to the prior US patent No. 8095484B2. That is, the embodiments of the present invention incorporate by reference the relevant provisions of the prior US patent.

[0057] Please refer to Figure 2 , Figure 2 1 is a block diagram of a virtual measurement application according to some embodiments of the present invention. First, a plurality of sets of process data 102 of historical workpieces as described above are obtained from a production machine. A process data pre-processing module 106 performs a data alignment operation on the process data 102 according to an automatic data alignment mechanism (ADAS) 108 to delete process data 102 with dissimilar time distribution shapes and make the data lengths of the process data 102 the same. Before or after performing the data alignment operation, the process data pre-processing module 106 may X The measurement data pre-processing module 112 uses the DQI model to evaluate the data quality of the process data 102 and organize and standardize the original process data 102 from the production machine. yModel is used to evaluate the data quality of the measurement data 104 (i.e., the multiple actual measurement values ​​described above) of the historical workpiece, and discretize and filter to remove the abnormal values. Then, the measurement data 104 of the historical workpiece and the aligned process data 102 are used as modeling samples, and a virtual measurement model 110, an RI model 120, and a GSI model 130 are established according to the convolutional neural network algorithm, wherein the virtual measurement model 110 includes a CNN model and has a two-stage mechanism. After the virtual measurement model 110, the RI model 120, and the GSI model 130 are modeled, the subsequent workpieces can be virtually measured. After obtaining a set of process data 102 of a workpiece, the process data pre-processing module 106 performs data alignment operations and / or other data pre-processing according to the automatic data alignment mechanism (ADAS) 108. Then, the processed process data 102 is input to the virtual measurement model 110, the RI model 120, and the GSI model 130 to calculate the first stage virtual measurement value (VM) of the workpiece. I ) and its confidence index (RI) and overall similarity index (GSI). After the workpiece is processed by the production machine, if the quality item is measured by the measured machine and its measurement data 104 (i.e., the actual measurement value) is obtained, the process data 102 and the measurement data 104 of the workpiece can be used to retrain or adjust the virtual measurement model 110, the RI model 120, and the GSI model 130 to calculate the second stage virtual measurement value (VM) of the workpiece. II ) and its confidence index value (RI) and the global similarity index value (GSI).

[0058] The operation of data alignment based on the automatic data alignment mechanism (ADAS) 108 is described below.

[0059] Please refer to Figure 3A , Figure 3AThe flowchart of the virtual measurement method based on the use of the automatic data alignment mechanism (ADAS) 108 according to some embodiments of the present invention is illustrated. Before the data alignment operation 300 is performed, the data collection operation 302 is performed to obtain multiple sets of process data, wherein these sets of process data are used or generated when the production machine processes multiple workpieces, and these sets of process data correspond to these workpieces in a one-to-one manner. Each set of process data includes the values ​​of multiple parameters, and the value of each parameter is a plurality of sets of timing data corresponding to these workpieces. In the data alignment operation 300, first, the operation 310 of obtaining the reference data length is performed to perform frequency distribution calculation on the data length of the multiple sets of timing data of each parameter, and obtain the distribution of the number of data occurrences to the data length, wherein the data length with the largest number of occurrences in the multiple sets of timing data of each parameter is a reference data length. Then, the operation 320 of obtaining the reference timing data is performed to perform an average calculation on the timing data with the aforementioned reference data length in the multiple sets of timing data of each parameter, and obtain a set of reference timing data for each parameter.

[0060] The following uses an application example to illustrate the operation 310 of obtaining the reference data length and the operation 320 of obtaining the reference timing data. Figure 3B and 3C , Figure 3B The data length of the process data of each workpiece in an application example of the present invention is shown. Figure 3C The distribution of the number of occurrences of each workpiece in the process of this application example versus the data length is plotted. Figure 3B As shown, this application example provides data of a certain parameter of 100 workpieces. The data of this parameter of each set of process data is as follows Figure 1 The time series data shown in . The data length of the time series data is 444 data points to 462 data points. After performing the frequency distribution calculation, the following can be obtained: Figure 3C The distribution of the number of data occurrences (15 times) to the data length is shown, where the data length with the most occurrences is 453 data points, so the reference data length is 453 data points, and the data length of the process data of 15 workpieces is 453 data points in total. After averaging the timing data of this parameter of these 15 workpieces, a set of reference timing data of this parameter can be obtained.

[0061] like Figure 3AAs shown, after obtaining a set of reference time series data for each parameter, an operation 330 of calculating the data distance is performed to use a dynamic time warping (DTW) algorithm to calculate a distance between each set of time series data for each parameter and its corresponding reference time series data. DTW is a method for measuring the similarity between two time series data (time series), which is widely used in the field of speech recognition to identify whether two speech segments represent the same word. Assume Figure 1 The pressure time series data A of the workpiece sample 1 shown is [a1, ..., a n ], the pressure time series data B of workpiece sample 3 = [b1,...,b m ], where A is the reference timing data, B is the test timing data (i.e., each set of timing data for each parameter), and the data length is n, m∈R. The warping path P between the two timing data contains the index of the alignment component, P=[p1,p2,...,p K ], where p K =(i K ,j K )∈[1:n]×[1:m], k∈[1:K], and satisfy the following restrictions:

[0062] Boundary restrictions: p1 = (1,1), p K =(n,m);

[0063] Step size and direction: (p k+1 -p k )∈{(1,0),(0,1),(1,1)}, where l∈[1:K-1].

[0064] The length of the normalized path P satisfies max(n,m)≤K≤n+m, and all normalized paths P between A and B are labeled P m,n The warping cost d(A,B) of the warping path P is defined as follows:

[0065]

[0066] Where d(…) is a distance function, which can be, for example, the Euclidean distance, i.e.

[0067] In order to obtain the best integration path, a cumulative cost matrix D is established, and its formula is as follows:

[0068]

[0069] According to the above formula, the distance (cumulative cost matrix) between the pressure time series data of workpiece sample 3 and the reference time series data (pressure time series data of workpiece sample 1) is 199.79, and the distance (cumulative cost matrix) between the pressure time series data of workpiece sample 2 and the reference time series data (pressure time series data of workpiece sample 1) is 152.29. Since the smaller the distance, the more similar, workpiece sample 2 is more similar to workpiece sample 1 than workpiece sample 3. It is worth mentioning that the distance function can also be, for example, Mahalanobis distance or other algorithms.

[0070] Then, a distance threshold is set (operation 332). When the distance (data distance) between the timing data of the parameter and its corresponding reference timing data is greater than the distance threshold, the process data corresponding to this distance is deleted (operation 334), that is, this set of process data is too dissimilar to other sets of process data and cannot be used for modeling. Setting the distance threshold (DTW T The operation 332 of ) can apply, for example, the Leave-One-Out (LOO) principle in Cross Validation, and the formula is as follows:

[0071]

[0072] in is the average value of the distance between each group of time series data of each parameter and its corresponding reference time series data; is the standard deviation of the distance between each set of timing data of each parameter and its corresponding reference timing data, which is calculated by the leave-one-out (LOO) principle for the process data of all workpieces (i.e., modeling sample data). α is obtained by Chebyshev's theorem, and its formula is as follows:

[0073]

[0074] When using, for example, 95% confidence (λ=0.05) to identify a value less than or equal to DTW T Normal data and above DTW T When the abnormal data

[0075] Next, an upper limit value of the data length is set (operation 340). The upper limit value of the data length may be, for example, Q3+k×IQR, where Q3 is Figure 3C The third quartile of the distribution of the number of data occurrences versus data length in the order of data length from small to large; IQR is the quartile of the distribution of the number of data occurrences versus data length in the order of data length from small to large; k is a constant greater than 0, for example: 1.5. Figure 3C As shown, among the 100 workpieces, the third quartile of the data length appearance order is the data length of the 75th workpiece, that is, Q3 is 455 data points; the first quartile of the data length appearance order is the data length of the 25th workpiece, which is 451 data points; IQR = 455-451 = 4 data points. Therefore, the upper limit of the data length = 455 + 1.5 × 4 = 461 data points. Then, the process data whose data length is greater than the upper limit of the data length (461 data points) is deleted from the process data of all workpieces, that is, the process data of the 99th and 100th workpieces.

[0076] Then, a data length adjustment operation 342 is performed to repeatedly add at least one data point having the value of the last data point after each set of timing data of each parameter, until the data length of each set of timing data of each parameter is equal to the longest data length of all process data.

[0077] Next, a modeling operation 350 is performed to establish a virtual measurement model using the adjusted process parameter data and the corresponding actual measurement values, wherein the virtual measurement model includes at least one convolutional neural network model. Then, a prediction operation 360 is performed. In the prediction operation 360, a set of process data of the next workpiece is first obtained, and then the process data of the workpiece is subjected to a data alignment operation 300. Then, the process data of the workpiece is input into the virtual measurement model to calculate the virtual measurement value of the workpiece.

[0078] The convolutional neural network model is typically composed of a convolutional layer, a pooling layer, a flattening layer, and / or a dropout layer, and a fully connected (FC) layer, wherein the FC layer includes at least one hidden layer and an output layer. Since the convolutional neural network model or algorithm is known to those with ordinary knowledge in the technical field to which the present invention belongs, it will not be described in detail here.

[0079] Ensemble learning can effectively improve accuracy by combining multiple single classifiers in the system. Therefore, the embodiment of the present invention also proposes an ensemble learning method to construct a virtual measurement model. Please refer to Figure 4 , Figure 4 FIG. 1 is a schematic diagram illustrating an integrated virtual measurement model according to some embodiments of the present invention. Figure 4As shown, the virtual measurement model of the embodiment of the present invention includes multiple convolutional neural network (CNN) models (CNN model 1, CNN model 2, ..., CNN model p) and an estimation model 400, wherein the multiple inputs of these convolutional neural network models are the time series data of each parameter of the process data, and the multiple outputs y1-y p is the input of the estimation model 400. The estimation model 400 is established based on, for example, a neural network algorithm, a complex regression algorithm, a partial least squares algorithm, or a support vector machine algorithm. The parameters 1, 2, ..., and p of the process data correspond to the CNN model 1, the CNN model 2, ..., and the CNN model p in a one-to-one manner. This virtual measurement model performs convolution, pooling, flattening, and / or discarding and full connection on each parameter individually to obtain intermediate outputs y1, y2, ..., y p Then, these intermediate outputs y1, y2, ..., y p Integration is performed on an additional fully connected layer (for example, the estimation model 400 is a back propagation neural network (BPNN) model) to generate the final output of the virtual measurement value.

[0080] The following is an example of the present invention using the application of etching depth and etching sidewall angle. Figure 5A , Figure 5A To illustrate the prediction results of the etching sidewall angle of the application example of the present invention, the curve 42 marked with an x ​​mark represents the actual etching sidewall angle; the curve 44 marked with a triangle mark represents the etching sidewall angle predicted by the known BPNN model; the curve 46 marked with a square mark represents the etching sidewall angle predicted by the known BPNN model. Figure 4 The average absolute error (MAE) of curve 44 is 0.383 degrees, and the determination coefficient (R 2 ) is 0.640, the mean absolute error (MAE) of curve 46 is 0.326 degrees, and the determination coefficient (R 2 ) is 0.708. Please refer to Figure 5B , Figure 5B To illustrate the prediction results of the etching depth of the application example of the present invention, the curve 52 marked with an x ​​mark represents the actual etching depth; the curve 54 marked with a triangle mark represents the etching depth predicted by the known BPNN model; the curve 56 marked with a square mark represents the etching depth predicted by the known BPNN model. Figure 4 The average absolute error (MAE) of curve 54 is 22.86 angstroms, and the determination coefficient (R 2 ) is 0.791, the mean absolute error (MAE) of curve 56 is 18.83 angstroms, and the determination coefficient (R 2) is 0.849. The smaller the MAE, the higher the accuracy; the larger the R2, the stronger the statistical relevance (explanatory power) of the model. Therefore, the integrated virtual measurement model provided by the embodiment of the present invention is superior to the known BPNN model.

[0081] It is understood that the virtual measurement method using convolutional neural network (CNN) of the present invention is the implementation steps described above, and the computer program product with built-in storage for measurement sampling of the present invention is used to complete the method for predicting the occurrence of machine processing events and the method for determining whether to start virtual measurement as described above. The order of each implementation step described in the above embodiment can be adjusted, combined or omitted according to actual needs. The above embodiment can be implemented using a computer program product, which can include a machine-readable medium storing multiple instructions, and these instructions can program a computer to perform the steps in the above embodiment. The machine-readable medium can be, but is not limited to, a floppy disk, an optical disk, a read-only optical disk, a magneto-optical disk, a read-only memory, a random access memory, an erasable programmable read-only memory (EPROM), an electronically erasable programmable read-only memory (EEPROM), an optical card (optical card) or a magnetic card, a flash memory, or any machine-readable medium suitable for storing electronic instructions. Furthermore, embodiments of the present invention may also be downloaded as a computer program product, which may be transferred from a remote computer to a requesting computer via data signals using a communications link (eg, a network connection).

[0082] It is also noted that the present invention may also be described in the context of a manufacturing system. Although the present invention may be implemented in semiconductor manufacturing, the present invention is not limited to semiconductor manufacturing and may also be applied to other manufacturing industries. A manufacturing system is configured to manufacture workpieces or products, including but not limited to microprocessors, memory devices, digital signal processors, application specific circuits (ASICs) or other similar devices. The present invention may also be applied to other workpieces or products other than semiconductor devices, such as vehicle rims, screws. The manufacturing system includes one or more processing tools that can be used to form one or more products or parts of products on or in a workpiece (e.g., a wafer, a glass substrate). It should be known to those skilled in the art of the invention that the production machine may be of any number and any type, including lithography machines, deposition machines, etching machines, grinding machines, annealing machines, machine tools and the like. In an embodiment, the measurement machine includes a scatterometer, an ellipsometer, a scanning electron microscope and the like.

[0083] In summary, the embodiments of the present invention can effectively use convolutional neural networks to perform virtual measurement, thereby improving the accuracy of conventional virtual measurement.

[0084] Although the present invention has been disclosed as above by way of embodiments, it is not intended to limit the present invention. Any person having ordinary knowledge in the technical field may make some changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention shall be determined by the scope defined in the attached claims.

Claims

1. A virtual measurement method using a convolutional neural network, characterized in that: Include: Obtaining a plurality of sets of process data, wherein the sets of process data are used or generated when a production machine processes a plurality of workpieces, the sets of process data correspond to the workpieces in a one-to-one manner, each of the sets of process data includes a plurality of parameter values, and each of the parameter values ​​is a plurality of sets of timing data respectively corresponding to the workpieces; Performing a data alignment operation on the sets of process data, the data alignment operation comprising: Performing a number distribution calculation on a data length of each of the sets of time series data of each of the parameters to obtain a distribution of the number of data occurrences versus the data length, wherein the data length with the largest number of occurrences in the sets of time series data of each of the parameters is a reference data length; Performing average calculation on the time series data having the reference data length in the time series data sets of each of the parameters to obtain a set of reference time series data for each of the parameters; Using a dynamic time warping algorithm to calculate a distance between each of the sets of time series data of each of the parameters and its corresponding reference time series data; Set a distance threshold; When the distance is greater than the distance threshold, deleting the process data corresponding to the distance; and Performing a data length adjustment operation to repeatedly add at least one data point having a value of the last data point after a last data point of each of the sets of timing data of each of the parameters until the data length of each of the sets of timing data of each of the parameters is equal to a longest data length of the sets of process data; Obtaining a plurality of actual measurement values ​​of the workpieces; Performing a modeling operation to establish a virtual metrology model using the sets of process parameter data and the actual measurement values, wherein the virtual metrology model includes at least one convolutional neural network model; and After performing the data alignment operation on another set of process data of another workpiece, the another set of process data of the another workpiece is input into the virtual measurement model to calculate a virtual measurement value of the another workpiece.

2. The virtual measurement method using a convolutional neural network according to claim 1, characterized in that: Also includes: Before performing the data length adjustment operation, setting a data length upper limit value; and The process data whose data length is greater than the upper limit of the data length is deleted from the groups of process data.

3. The virtual measurement method using a convolutional neural network according to claim 2, characterized in that: The upper limit of the data length is Q3+k×IQR, where Q3 is the third quartile of the order of data length from small to large in the distribution of the number of data occurrences versus data length; IQR is the quartile of the order of data length from small to large in the distribution of the number of data occurrences versus data length; and k is a constant greater than 0.

4. The virtual measurement method using a convolutional neural network according to claim 1, characterized in that: The operation of setting the distance threshold value is to apply the leave-one-out principle in interactive validation.

5. The virtual measurement method using a convolutional neural network according to claim 1, characterized in that: The virtual measurement model includes multiple convolutional neural network models and an estimation model, wherein the multiple inputs of the convolutional neural network models are respectively the time series data of the parameters, and the multiple outputs of the convolutional neural network models are the inputs of the estimation model.

6. The virtual measurement method using a convolutional neural network according to claim 5, characterized in that: The estimation model is established based on a neural network algorithm, a complex regression algorithm, a partial least squares algorithm or a support vector machine algorithm.

Images