Substrate processing apparatus and data variation determination method

Abstract

The present inventive concept provides a substrate processing apparatus. The substrate processing apparatus includes at least one sensor configured to measure a condition of the substrate or the apparatus during the processing of the substrate, a data collection unit configured to collect data measured by the sensor in a time series, and a data processing unit configured to learn the data collected by the data collection unit to detect a change in current data measured by the sensor. The data processing unit includes a data learning unit configured to learn past data collected by the data collection unit using a twin network, and a data checking unit configured to detect whether a problem has occurred in the current data based on the learned data.

Description

Technical Field

[0001] The embodiments of the inventive concept described herein relate to a substrate processing apparatus and a method for determining data changes. More specifically, the inventive concept relates to a method for learning data using a Siamese network and determining whether data has changed based thereon. Background Technology

[0002] Data analysis is a crucial aspect of semiconductor manufacturing facilities, as it can be used for error detection and equipment maintenance. Identifying changed data is a key challenge. A common algorithmic method for determining data differences involves calculating the geometric distance between two data samples and comparing this distance to a predefined threshold. For example, if the calculated distance is less than the threshold (distance < threshold), the two samples are identified as "unchanged (no difference)." In this case, the threshold can be a predefined value or an equation. However, a problem arises when using this method to determine if data changes have occurred: data A1 and data A2 collected at different times in equipment or systems where no issues have arisen are incorrectly identified as changed (different). Therefore, an algorithm and method are needed to accurately determine whether data has changed. Summary of the Invention

[0003] Embodiments of this invention provide an algorithm that can accurately determine whether data has changed when a problem occurs.

[0004] The technical objectives of this invention are not limited to those described above, and other unmentioned technical objectives will become apparent to those skilled in the art from the following description.

[0005] The present invention provides an apparatus for processing a substrate. The apparatus for processing the substrate includes: at least one sensor configured to measure the condition of the substrate or the apparatus during the processing of the substrate; a data collection unit configured to collect data measured by the sensor in a time series; and a data processing unit configured to learn the data collected by the data collection unit to detect changes in the current data measured by the sensor.

[0006] In one embodiment, the data processing unit includes: a data learning unit configured to use a Siamese network to learn past data collected by the data collection unit; and a data checking unit configured to detect whether a problem has occurred in the current data based on the learned data.

[0007] In one embodiment, the data collection unit collects first data prior to the problem and second data subsequent to the problem, and the data learning unit uses the Siamese network to learn the first data and the second data, and to learn whether the data related to the problem is the same and whether it has changed.

[0008] In an embodiment, the data collection unit successively defines data pairs collected in a time series and samples the data pairs.

[0009] In an embodiment, the data learning unit sets any one of the first data as a reference value, and learns by setting the relationship between another first data (excluding the first data) and the reference value to 0 and the relationship between the reference value and the second data to 1.

[0010] In one embodiment, the data inspection unit uses current data measured by the sensor to test the validity of the data learned by the data learning unit.

[0011] In one embodiment, the data checking unit checks the output by inputting two data points identified at the sensor as input values ​​for the Siamese network learned at the data learning unit after completing the validity test.

[0012] In one embodiment, the data inspection unit detects changes in the sensor by inspecting the output.

[0013] In an embodiment, the data inspection unit sets the case where the output is 1 as the fourth data, and sets the previous data as the third data based on this, and checks the time point of the problem by checking the output through continuous sampling.

[0014] In an embodiment, when the output differs from the result learned by the data learning unit, the data checking unit retains the determination (i.e., retains the decision).

[0015] In an embodiment, the data collected from the at least one sensor is digital data related to numbers.

[0016] This invention provides a method for determining whether data generated during a substrate processing process has changed. The data change determination method includes: steps for collecting first data prior to a problem and second data after the problem occurred; steps for learning the first and second data using a twin network; and steps for detecting whether the current data has changed based on the learned twin network.

[0017] In an embodiment, the steps for collecting the first data prior to the occurrence of the problem and the second data after the occurrence of the problem are performed by sequentially sampling the collected time-series data in a defined pair.

[0018] In an embodiment, the step of learning the first data and the second data through the Siamese network sets any one of the first data as a reference value, and learns by setting the relationship between another first data (other than the first data) and the reference value to 0 and setting the relationship between the reference value and the second data to 1.

[0019] In one embodiment, the data change determination method includes the step of performing a validity test on the learned Siamese network.

[0020] In an embodiment, the step of detecting whether the current data has changed based on the learned Siamese network checks the output by inputting two data points identified at the sensor as input values ​​to the Siamese network learned at the data learning unit after completing the validity test.

[0021] In one embodiment, the data change determination method further includes a step of detecting a sensor that has changed by examining the output.

[0022] In an embodiment, the step of detecting whether the current data has changed based on the learned Siamese network sets the case where the output is 1 as the fourth data, and sets the previous data as the third data based on this, and checks the time point of the problem by checking the output via continuous sampling.

[0023] In an embodiment, the output is retained when different results are obtained from the learning from the Siamese network.

[0024] In one embodiment, a computer-readable recording medium has a program for performing a data change determination method.

[0025] According to an embodiment of the present invention, an algorithm is proposed that can accurately determine whether data has changed when a problem occurs.

[0026] According to an embodiment of the present invention, it is determined that there is no variation (difference) between test sample 1 and test sample 2 collected in a device or system where no problem has occurred. If a problem has occurred in the device or system, it can be determined that variation (difference) exists only in the data of the sensor related to the problem.

[0027] The effects of this invention are not limited to those described above, and other effects not mentioned will become apparent to those skilled in the art from the following description. Attached Figure Description

[0028] The above and other objects and features will become apparent from the following description with reference to the accompanying drawings, wherein, unless otherwise stated, the same reference numerals in the various drawings refer to the same parts, and wherein:

[0029] Figure 1 This is a block diagram illustrating the configuration of a substrate processing apparatus according to an embodiment of the present invention.

[0030] Figure 2 This is a block diagram illustrating the configuration of a data processing unit according to an embodiment of the present invention.

[0031] Figures 3A to 3C Learning via twin networks, as conceived according to the present invention, is illustrated.

[0032] Figure 4 A data collection method based on the present invention is illustrated.

[0033] Figure 5 The data collection and variation determination based on the present invention are illustrated.

[0034] Figure 6 A distance learning method in a twin network conceived according to the present invention is illustrated.

[0035] Figure 7 This is a flowchart illustrating a data change determination method based on the present invention.

[0036] Figures 8 to 14 The learning method and determination method according to the present invention are shown in more detail. Detailed Implementation

[0037] The inventive concept can be modified and taken in various forms, and specific embodiments of the inventive concept will be shown and described in detail in the accompanying drawings. However, the embodiments of the inventive concept are not intended to limit the specific forms disclosed, and it should be understood that the inventive concept includes all variations, equivalents, and substitutions included within the spirit and scope of the inventive concept. In the description of the inventive concept, detailed descriptions of relevant prior art may be omitted where it may obscure the essence of the inventive concept.

[0038] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the inventive concept. As used herein, the singular forms “a,” “an,” and “described” are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that, when used in this specification, the terms “comprise” and / or “comprising” specify the presence of stated features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or combinations thereof. As used herein, the term “and / or” includes any and all combinations of one or more associated listed items. Furthermore, the term “exemplary” is intended to refer to an example or illustration.

[0039] In the following, embodiments of the inventive concept will be described in detail with reference to the accompanying drawings.

[0040] Figure 1 This is a block diagram illustrating the configuration of a substrate processing apparatus 1 according to an embodiment of the present invention.

[0041] refer to Figure 1 The substrate processing apparatus 1 according to the present invention may include a sensor 10, a data collection unit 20, and a data processing unit 30. According to an embodiment, the sensor 10 can measure the condition of the substrate or apparatus during substrate processing. Figure 1 In the block diagram, sensor 10 is shown as one, but multiple sensors 10 may be provided. Sensor 10 may be disposed within the substrate processing apparatus 1 or may be disposed externally. According to an embodiment, sensor 10 may be a temperature sensor. According to an embodiment, sensor 10 may be a pressure sensor. According to an embodiment, sensor 10 may be a sensor for measuring the condition of the substrate surface. However, sensor 10 is not limited to these, and may be a sensor capable of measuring parameters related to the substrate processing apparatus or the substrate processing apparatus itself throughout the entire process of the substrate processing apparatus. According to an embodiment, substrate processing apparatus 1 may be a substrate processing apparatus using a non-limiting plasma process. The process performed by substrate processing apparatus 1 may be any process performed by a semiconductor processing apparatus.

[0042] The data collection unit 20 can collect data measured by one or more sensors 10 in a time series manner. When multiple sensors 10 are present, the data collection unit 20 can collect data from each of the multiple sensors 10. The data collection unit 20 can collect time series data at fixed time intervals. The data collection unit 20 can collect first data before a problem occurs and second data after a problem occurs, based on the occurrence of a problem. A detailed data collection method will be described later with reference to the accompanying drawings.

[0043] As used herein, the data-related term "problem" in this invention concept can refer to a significant event that alters the data. According to an embodiment, a problem can be a device malfunction. According to an embodiment, a problem can be a sudden error event.

[0044] The data processing unit 30 can learn from the data collected by the data collection unit 20 to detect whether the current data measured by the sensor 10 has changed. The detailed configuration and learning method of the data processing unit 30 will be described later with reference to the accompanying drawings.

[0045] Figure 2 This is a block diagram illustrating the configuration of a data processing unit 30 according to an embodiment of the present invention.

[0046] The data processing unit 30 according to the present invention may include a data learning unit 31 and a data checking unit 32. The data learning unit 31 may use a twin network to learn from data collected in the past by the data collection unit 30. The data learning unit 31 may use the twin network to learn whether the first data and the second data are the same and whether the data has changed in relation to the problem.

[0047] The data inspection unit 32 can detect whether a problem has occurred in the current data based on the data learned by the data learning unit 31.

[0048] According to the present invention, it is assumed that a problem occurs in substrate processing equipment 1 during substrate processing. In this case, data is collected before and after the problem occurs by sensors 10 installed inside and / or outside the substrate processing equipment 1. Based on the data collected before and after the problem, it is determined which sensors 10 show changes between the two sets of data. Those sensors 10 whose data changed are considered to be sensors associated with the cause of the problem because the data changed after the problem.

[0049] The present invention differs from conventional algorithms in that it can simultaneously utilize the following: According to the present invention, data collected before and after a problem occurs are compared to identify sensors 10 where data changes (differences) exist. According to the present invention, deep learning using Siamese networks determines that there are "changes (differences)" in the data collected from each sensor 10 before and after the current problem occurs. Deep learning using Siamese networks, according to the present invention, determines that the data for the current problem "changes (differences)" exist based on data collected before and after previous problems. Conventionally, determining whether the inputs A and B through the Siamese network are the same is the subject; however, in the case of the present invention, instead, it is assumed that the inputs A and B through the Siamese network are the same and the subject is determining whether a change has occurred. According to the present invention, when determining whether there are changes (differences) in the data for the current problem, the Siamese network only produces a result if the resulting values ​​are consistent. Detailed methods for processing the Siamese network will be described below.

[0050] Figures 3A to 3C This is a view used to explain learning via twin networks according to the present invention.

[0051] Referring to Figure 3, a method for identifying differences in Siamese networks is disclosed. Figure 3A A reference for the difference between data A and data B can be learned through Siamese networks; this is known as the Siamese threshold. Figure 3B Once the Siamese network has completed learning, test1 and test2, which serve as the input values ​​for testing, can be fed into the learned Siamese network. The learned Siamese network then outputs the Siamese distance as the values ​​for test1 and test2. (See reference) Figure 3C The Siamese network conceived according to the present invention can determine whether "a change (difference) exists" based on this Siamese distance, considering a Siamese threshold. In this case, the output value can have values ​​between 0 and 1. In this case, when the output value is 0, it can be determined that there is no change, and when the output value is 1, it can be determined that there is a change. When the output value is between 0 and 1, it can be determined that there is no change when the output value is below the Siamese threshold, and that there is a change when the output value is above the threshold.

[0052] Figure 4 A data collection method based on the present invention is illustrated.

[0053] According to an embodiment, the substrate processing apparatus may include one or more sensors 10. Each sensor 10 included in the substrate processing apparatus may periodically generate data. Each sensor 10 included in the substrate processing apparatus may collect data generated in a time sequence. When each sensor 10 generates time-series data, the normality of the first data and the next data can be defined as a pair. According to the present invention, after sampling the first data pair, subsequent data pairs may be sampled sequentially in time sequence. That is, the data collection unit 20 according to the present invention may sample the normality of each data pair sequentially in time sequence.

[0054] Figure 5 The data collection and defect identification based on the present invention are illustrated.

[0055] refer to Figure 5 You can use, such as Figure 4 Learning is performed on the sampled data as shown. This can be the processing method of data processing unit 30. In a Siamese network, a pair of data can be input to determine whether the input data pair is identical. (Reference) Figure 5 Normal data A1 and normal data A2 are input as input 1 and input 2 of the Siamese network. Normal data A1 can consist of n I / O data (A11, A12, ..., A1n). Similarly, normal data A2 can consist of n I / O data (A21, A22, ..., A2n). The Siamese network can be learned by matching A1 and A2 with the same (corresponding) I / O data. Subsequently, normality can be learned by each of the same I / O data at input data 1 and input data 2. That is, according to the present invention, a normality test (A1-A2, A2-A3, ..., An-A1) can be performed n times using n normal data segments (A1, A2, ..., An). Then, input B1 (which is data after the problem occurred) can be used as input 2 to perform a normality test. The learned Siamese network can detect abnormal I / O. According to the present invention, anomalous I / O can be defined by the frequency of anomalous I / O detected by the learned Siamese network or by the frequency of anomalous I / O detected by the Siamese network in a number of tests. This results in the effect of quantifying problem correlation using the frequency of anomalous I / O detected in a number of tests. According to an embodiment, when 5 anomalous I / Os are detected in 10 tests, a 50% problem correlation can be determined.

[0056] In other words, according to the present invention, data changes can be determined through the learning of twin networks, the determined results can be used to detect sensors that have caused related problems, and the problem correlation can be determined through multiple inference results.

[0057] Figure 6A learning method in a twin network conceived according to the present invention is illustrated.

[0058] refer to Figure 6 The Siamese network can learn to make the distance between A1 and B1 greater than the distance between A1 and A2. According to an embodiment, the Siamese network can learn distances close to 1 between A1 and B1 and distances close to 0 between A1 and A2. Through this process, the Siamese network can learn data from groups A and B, data before and after a problem occurs, and thereby examine when and whether a problem occurred.

[0059] Figure 7 This is a flowchart illustrating a data change detection method based on the present invention.

[0060] First, according to the concept of the present invention, it is necessary to... Figure 7 The data A, B, C, and D described in the attached figure are defined as follows.

[0061] In the case of data A, the data prior to the occurrence of problem 1 in system 1 is defined as group A. In the case of data B, the data after the occurrence of problem 1 in the same system 1 is defined as group B. In the case of data C, the data before the recurrence of problem 1 in the same system 1 is defined as group C. In the case of data D, the data after the recurrence of problem 1 in the same system 1 is defined as group D. The learning method and the method for verifying the existence of changes according to the present invention will be described in more detail.

[0062] refer to Figure 7The Siamese network can be trained using data from groups A and B, which represent the data from when problem 1 first occurred in the same system 1. The Siamese network learns that data A1 and A2 in group A are unchanged data pairs, and that data B1 in group B is a changed data pair compared to data A1 in group A. Once sufficient learning is complete, validation tests can be performed using data from groups C and D, which represent data before and after problem 1 reoccurs in the current state. In this case, it can be checked whether data C1 in group C is the same as data C2 in group C, and whether data C1 in group C is different from data D1 in group D. This can be checked using data pairs input into the learned Siamese network. If the validity test results are determined to be valid, tests can be performed based on data from groups C and D, which represent data before and after the current problem occurs. Therefore, by measuring the distance between data C1 included in test group C and data D1 included in group D, if the distance is less than a predetermined threshold, C and D are determined to be similar (no change), and if the distance is greater than the predetermined threshold, C and D are determined to be different (change).

[0063] The data learning and determination methods based on the present invention will be described with reference to the following figures.

[0064] Figures 8 to 14 The learning method according to the present invention is shown in more detail.

[0065] according to Figure 8 The criteria for assigning data from group A to group D are disclosed. The procedure for learning problems at the sensors into a Siamese network is described below.

[0066] refer to Figure 9 We can select a sample An from group A and set it as the anchor. In this case, the anchor An can be normal data from the data included in group A. Then, we can select Am instead of An from group A, and the relationship between An and Am is defined as unchanged, i.e., references to the same data. Alternatively, we can select B1, collected immediately after problem 1 occurs, from group B, and the relationship between An and B1 is defined as changed, i.e., references to different data. Afterward, the learning of the Siamese network is divided into two stages, as follows: An and Am are applied to inputs 1 and 2 of the Siamese network to learn a Siamese distance close to zero. In the case of An and B1, a Siamese distance close to 1 is learned. The two convolutional neural networks (CNNs) inside the Siamese network can be set to the same structure. Learning is the process of changing the weight parameters and the biases of CNN1 and CNN2 to the same value while reducing the difference between the Siamese distance and the target value (0 or 1) of the Siamese network to below a certain level. This can be done in... Figure 6 As shown in the image.

[0067] The next step illustrates the process of using a Siamese network to perform the analysis for Problem 1. (Reference) Figure 10 Before performing the actual test, perform validation tests for Cn&Cm and Cn&D1. If the results are valid, Cn&D1 and Cn&D1 can be tested to identify sensors with data changes. If the Siamese network output is close to 1 for the inputs (data from group D after the problem recurs and data from group C before the problem recurs), and if the Siamese network output is close to 0 only for the data from group C before the problem recurs, the result can be determined to be valid.

[0068] refer to Figure 11 According to the present invention, when the previous problem 1 occurs again, data D1 collected immediately after the occurrence of the current problem 1 is found. Therefore, the data in System 1 can be divided into group C and group D based on D1. According to an embodiment, Cn (fixed) can be fixed from normal data collected long before the occurrence of event 1 (problem 1). According to an embodiment, Cn can be data collected a month ago. After making Cm and D1 (which are anchor points for unchanged / changed data), the value to the right of Cn can be successively applied to Cm = D1 to determine the result. In this case, when the test result of Ta and the test result of Ta+1 show a large difference, it is considered that problem 1 occurred between Ta and Ta+1. In other words, the time point when the problem occurred can be indicated by this method.

[0069] There are two examples regarding how to analyze the occurrence of a problem.

[0070] exist Figure 12 The first example is shown. According to the embodiment, if D1 (which is the problem data) is found, the cause of the problem is analyzed, while simultaneously, after fixing D1, the values ​​of Cn and Cm are changed towards D1 and Cn and Cm are set to be adjacent to each other, as shown. Figure 12 As shown. In particular, when focusing on comparison Figure 12 In cases 1 and 2 shown, sensors whose data changed (differenced) before and after the problem can be identified.

[0071] exist Figure 13 A cause analysis method according to another embodiment is disclosed. Unlike... Figure 12 In one embodiment, Cn and D1 are fixed, and the value of Cm is moved from Cn to D1 to analyze the cause of the problem. Specifically, in the comparison... Figure 13 In cases 1 and 2 shown, sensors whose data changed (differenced) before and after the problem can be identified.

[0072] By using these methods, data can be used to detect sensors that have malfunctioned and to pinpoint the time when the problem occurred.

[0073] refer to Figure 14 The output values ​​can be used to determine which cause analysis to retain. For example... Figure 14 As shown, if different learning datasets are set up and learned in a Siamese network, and the results of applying the same validation and test data are inconsistent, the Siamese network can retain its judgment on the problem. In other words, problems that have not been experienced before do not show consistent results. For example, for learning dataset 1 and dataset 2, results 1 and 2 indicate "data has changed." For learning dataset 3 and dataset 4, results 3 and 4 indicate "no data change." The Siamese network can retain its judgment on the result when 50% of the test results are the same across a total of 4 tests.

[0074] In other words, the data change determination method according to the present invention can be summarized as follows. According to the present invention, sensors showing data changes (differences) between data A collected from sensors in a normally operating substrate processing device and data B collected after a problem occurs can be identified, and the cause of the problem can be analyzed using sensors associated with the occurrence of the problem. When analyzing causes by comparing data before and after the problem occurs, the present invention differs from conventional techniques in that the criteria used to determine the data changes of each sensor before and after the problem are normal data and problem data from previous problems of the same kind. Furthermore, a twinning threshold can be learned by using a twinning network and normal data and problem data from specific past problems. Additionally, when learning about specific problems on the twinning network, data collected before and after the current problem are input into the learned twinning network regarding the currently recurring problem. The twinning network presents a twinning distance for the current problem data as its output based on past problem data. If the analysis currently focuses on a problem that has not yet been experienced in the past, the twinning network outputs "unknown" without mentioning whether any changes (differences) exist and suspends the analysis of the cause of the problem.

[0075] Furthermore, the data change determination method according to the embodiments of the present invention described above can be implemented in the form of program instructions, which can be executed by various computer means and recorded on a computer-readable recording medium. In this case, the computer-readable recording medium may include program commands, data files, data structures, etc., individually or in combination. Moreover, the program commands recorded on the recording medium may be specifically designed and configured for the present invention, or may be known or usable by a person skilled in computer software.

[0076] Computer-readable recording media may include hardware devices specifically configured to store and execute program instructions, such as magnetic media (e.g., hard disks, floppy disks, and magnetic tapes), optical media (e.g., CD-ROMs and DVDs), magneto-optical media (e.g., optical disks), and ROM, RAM, flash memory, etc. Furthermore, program instructions include machine language code, such as code created by a compiler, and high-level language code that can be executed by a computer using an interpreter. The aforementioned hardware devices may be configured to operate as one or more software modules to perform the operations of the present invention.

[0077] The effects of this invention are not limited to those described above, and those skilled in the art to which this invention pertains can clearly understand any effects not mentioned from the specification and drawings.

[0078] Although preferred embodiments of the inventive concept have been shown and described up to now, the inventive concept is not limited to the specific embodiments described above, and it should be noted that those skilled in the art to which the inventive concept relates can implement the inventive concept differently without departing from the essence of the inventive concept as claimed in the claims, and should not be interpreted or modified separately from the technical spirit or prospect of the inventive concept.

Claims

1. An apparatus for processing a substrate, comprising: At least one sensor, the at least one sensor being configured to measure the condition of the substrate or the device during the processing of the substrate; A data collection unit configured to collect data measured by the sensor in a time series. as well as A data processing unit configured to learn the data collected by the data collection unit to detect changes in the current data measured by the sensor. The data processing unit includes: A data learning unit, configured to use a Siamese network to learn past data collected by the data collection unit; as well as A data inspection unit is configured to detect whether a problem has occurred in the current data based on the learned data. The data collection unit collects first data prior to the problem and second data subsequent to the problem. The data learning unit uses the Siamese network to learn the first data and the second data, and to learn whether the data related to the problem are the same and whether they have changed. The data learning unit sets any one of the first data as a reference value and learns by setting the relationship between another first data (other than the first data) and the reference value to 0 and setting the relationship between the reference value and the second data to 1.

2. The apparatus for processing a substrate according to claim 1, wherein the data collection unit successively defines data pairs collected in a time sequence and samples the data pairs.

3. The apparatus for processing a substrate according to claim 1, wherein the data inspection unit uses current data measured by the sensor to test the validity of data learned by the data learning unit.

4. The apparatus for processing a substrate according to claim 3, wherein the data inspection unit checks the output by inputting two data points identified at the sensor as input values ​​to the twin network learned at the data learning unit after completing the validity test.

5. The apparatus for processing a substrate according to claim 4, wherein the data inspection unit detects changes in the sensor by inspecting the output.

6. The apparatus for processing a substrate according to claim 5, wherein the data checking unit sets the case where the output is 1 as fourth data, and sets the previous data as third data based thereon, and checks the time point of the problem occurrence by checking the output via continuous sampling.

7. The apparatus for processing a substrate according to any one of claims 4 to 6, wherein the data checking unit retains the determination when the output differs from the result learned by the data learning unit.

8. The apparatus for processing a substrate according to any one of claims 1 to 6, wherein the data collected from the at least one sensor is digital data related to a digital object.

9. A method for determining whether data generated during a substrate processing process has changed, the method comprising: Steps for collecting first data before the problem occurs and second data after the problem occurs; Steps for learning the first data and the second data through a twin network; as well as This step is used to detect whether the current data has changed based on the learned Siamese network. The steps for collecting the first data prior to the occurrence of the problem and the second data after the occurrence of the problem involve sequentially sampling the collected time-series data in a defined pair. The step of learning the first data and the second data through the Siamese network involves setting any one of the first data as a reference value, and learning is performed by setting the relationship between another first data (other than the first data) and the reference value to 0 and by setting the relationship between the reference value and the second data to 1.

10. The method of claim 9, further comprising the step of performing a validity test on the learned Siamese network.

11. The method of claim 10, wherein the step of detecting whether the current data has changed based on the learned Siamese network is performed by checking the output by inputting two data points identified at the sensor as input values ​​to the Siamese network learned at the data learning unit after the validity test is completed.

12. The method of claim 11, further comprising the step of detecting a sensor that has changed by examining the output.

13. The method of claim 12, wherein the step of detecting whether the current data has changed based on the learned Siamese network sets the case when the output is 1 as the fourth data, and sets the previous data as the third data based on this, and checks the time point of the problem by checking the output via continuous sampling.

14. The method of claim 13, wherein the output is retained when different results arise from the results learned from the Siamese network.

15. A computer-readable recording medium having a program for performing the method according to any one of claims 9 to 14.

Images