Learning data generation device, learning data generation method, learning data generation program, and classification device

Abstract

To provide a learning data generation device capable of generating sufficient learning data for creating a predictive model through machine learning.SOLUTION: A learning data generation device comprises: raw data creation means 3 for acquiring a feature amount from waveforms already obtained from a device and creating raw data for balancing classes along with the waveform class; minority data increase means 4 for increasing minority data using the raw data to obtain a plurality of sets of new set data composed of waveform feature amounts and class sets; waveform generation means 5 for generating waveforms using the feature amounts set for each set of set data obtained by the minority data increase means 4; and learning data generation means 6 for displaying the generated waveforms on a display device 25, accepting an input of the displayed waveform class, and generating learning data by combining the entered class and the corresponding feature amounts as one set.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] This invention relates to a learning data generation device, a learning data generation method, a learning data generation program, and a classification device.

Background Art

[0002] In semiconductor and various device designs, it is necessary to evaluate signal waveforms using measuring instruments such as oscilloscopes. Currently, to check the signal waveforms acquired by measuring instruments such as oscilloscopes, an operator visually determines pass or fail. On the other hand, when analyzing signals such as voltage acquired from measuring instruments as numerical values, there is a problem that the data cannot be handled as it is because of the accuracy error of the measuring instruments. Also, when creating a prediction model by machine learning, signal waveforms determined as abnormal signals occur very rarely, so there is a problem that it is extremely difficult to create a determination model by machine learning because the samples of abnormal signals for learning are extremely few.

[0003] Patent Document 1 discloses a method, apparatus, and computer program for creating a category determination rule that can create a determination rule with few misjudgments. In these methods, apparatuses, and computer programs for creating category determination rules, the teaching data included in one group of the divided groups is set as verification data, and a verification category determination rule is created based on the teaching data included in the remaining other groups. Then, the operation of performing category determination on the verification data based on the verification category determination rule is mutually performed while sequentially changing the group set as the verification data. The verification data and the category determination for it are presented, prompting for correction or deletion. Upon receiving an instruction, the teaching data is corrected, and a category determination rule is created and output based on the corrected teaching data.

[0004] Patent Document 2 discloses a monitoring method that collects information measured by instruments and monitors a monitoring target. This method includes steps of detecting display information showing the state of the object to be monitored by a non-contact sensor, converting the detected data into digital data, collecting the digital data, and determining the state of the object to be monitored based on the digital data. Specifically, in the detecting step, a measuring instrument is photographed with a camera, and in the converting step, the value of the parameter displayed on the measuring instrument is converted into digital data by analyzing the image obtained by the photographing.

[0005] Patent Document 3 discloses an abnormality detection device capable of performing abnormality detection in consideration of the flow of a process composed of a plurality of operating states. This abnormality detection device includes a feature amount extraction unit that receives time-series data of a target device and extracts a feature amount of the time-series data, a state transition estimation unit that specifies the operating state of the target device from the extracted feature amount and estimates the state transition of the target device from the extracted feature amount with reference to state transition pattern information defining a state transition pattern between a plurality of operating states, and an abnormality detection unit that determines whether the time-series data is abnormal from the specified operating state, the estimated state transition, and the feature amount of the extracted time-series data with reference to the state transition pattern information and normal range information defining a normal range for each operating state. According to this abnormality detection device, since the abnormality determination is made with reference to the state transition pattern information and the normal range information, the abnormality determination can be performed in two stages, namely, the state transition stage and the stage of each operating state. Therefore, it is possible to perform abnormality detection in consideration of the flow of a process composed of a plurality of operating states.

[0006] The above feature amounts include indicators such as the long-term or short-term trend of the waveform of the sensor value, the length of the waveform, the frequency, the update frequency of the sensor value, and the similarity between a plurality of sub-waveforms included in a certain waveform (time-series data).

Prior Art Documents

Patent Documents

[0007] [Patent Document 1] Japanese Patent Application Laid-Open No. 2011-13720 [Patent Document 2] Japanese Patent Application Laid-Open No. 2022-155615 [Patent Document 3] WO2022 / 054256 [Summary of the Invention] [Problems to be Solved by the Invention]

[0008] As described above, although there are known methods for obtaining feature amounts of waveforms and performing determinations and the like, they have not been able to address problems such as not being able to directly use data from measuring instruments with accuracy errors.

[0009] An object of the invention according to an embodiment of the present invention is to address the problem that, due to accuracy errors in measuring instruments, it is not possible to handle the raw data as it is, and to create sufficient learning data for creating a prediction model by learning through machine learning even when signal waveforms determined to be of a certain class occur very rarely. The present invention provides a learning data generation device, a learning data generation method, a learning data generation program, and a classification device. [Means for Solving the Problems]

[0010] A learning data generation device according to an embodiment of the present invention includes raw data creation means for acquiring a feature amount from a waveform already obtained from a device and creating raw data for balancing classes together with the class of the waveform; minor data increase means for increasing minor data using the raw data to obtain a plurality of sets of new set data composed of a set of a feature amount and a class of the waveform; waveform generation means for generating a waveform for each set of the set data obtained by the minor data increase means using the set feature amount; and learning data generation means for displaying the generated waveform on a display device and accepting an input of the class of the displayed waveform, and generating learning data with the input class and the feature amount as one set.

[0011] In the learning data generation device according to an embodiment of the present invention, the original data creation means creates a moving average by dividing the number of waveform data by the number of pixels and further reducing the value, and extracts a feature amount using the value of this moving average.

[0012] In the learning data generation device according to an embodiment of the present invention, the original data creation means extracts the maximum value and the minimum value of the waveform as feature amounts.

[0013] In the learning data generation device according to an embodiment of the present invention, the original data creation means extracts the number of steps at a predetermined height between the maximum value and the minimum value of the waveform as a feature amount.

[0014] In the learning data generation device according to an embodiment of the present invention, the waveform is a waveform of the output voltage of the device, the temperature of the device, the noise of the device, the voice generated from the device, the volume generated from the device, and the vibration generated from the device.

[0015] In the learning data generation device according to an embodiment of the present invention, the minor data increasing means obtains a plurality of sets of new set data using the SMOTE (Synthetic Minority Oversampling TEchnique) method.

[0016] The classification device according to an embodiment of the present invention includes a learning data creation device according to claim 1, prediction model creation means for creating a prediction model that predicts a class from feature amounts by learning using the learning data created by the learning data creation device, and class acquisition means for obtaining a class using the prediction model based on the feature amounts of the waveform obtained from the device to be inspected.

Brief Description of the Drawings

[0017]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Embodiments for Carrying Out the Invention

[0018] Hereinafter, a learning data generation device, a learning data generation method, a learning data generation program, and a classification device according to embodiments of the present invention will be described with reference to the accompanying drawings. In each figure, the same reference numerals are assigned to the same components, and redundant descriptions are omitted.

[0019] As shown in FIG. 1, a learning data generation device according to an embodiment of the present invention can be configured using a computer. That is, the CPU 10 configures the learning data generation device using programs and data in the main memory 11. An external storage interface 13, an input interface 14, a display interface 15, and a data input interface 16 are connected to the CPU 10 via a bus 12.

[0020] An external memory interface 13 is connected to an external memory device 23. The external memory device 23 stores programs and data for the operation of this learning data generation device, and these can be appropriately read out by the CPU 10 to the main memory 11 and used. Therefore, as shown in FIG. 2, the external memory device 23 stores a program for realizing the original data creation means 3, the negative data increase means 4, the waveform generation means 5, and the learning data generation means 6. An input device 24 such as a keyboard or a touch panel and a pointing device 22 such as a mouse are connected to the input interface 14. A display device 25 having a screen such as an LCD is connected to the display interface 15. An oscilloscope 26, which is a measuring instrument, is connected to the data input interface 16. The oscilloscope 26 functions as a measuring device for obtaining the waveform of a measurement target signal from a device to be designed such as a semiconductor. The waveform signal obtained by the oscilloscope 26 is taken in by the CPU 10 through the data input interface 16, and a process of extracting feature amounts is performed.

[0021] In this embodiment, the original data creation means 3, the negative data increase means 4, the waveform generation means 5, and the learning data generation means 6 provided in the external memory device 23 have the following functions. That is, the original data creation means 3 acquires feature amounts from the waveforms already obtained from the device, and creates original data for balancing the classes together with the classes of the waveforms. The negative data increase means 4 uses the original data to increase the negative data and obtains a plurality of sets of new set data composed of a set of waveform feature amounts and classes.

[0022] The waveform generation means 5 generates waveforms for each set of the set data obtained by the negative data increase means 4 using the set feature amounts. The learning data generation means 6 displays the generated waveforms on the display device 25 and accepts the input of the classes of the displayed waveforms, and generates learning data with the input class and the feature amount as one set. In this case, the input can be performed from the input device 24 or the pointing device 22.

[0023] The above-mentioned original data creation means 3 creates a moving average by dividing the number of waveform data by the number of pixels and further reducing the value, and extracts feature quantities using the value of this moving average. Here, it is assumed that the waveform data obtained from the oscilloscope 26 is the power supply voltage VDD and is displayed as shown in FIG. 3. The vertical axis represents V (volts), and the horizontal axis represents time in steps of a predetermined time interval.

[0024] As shown in FIG. 3, even at the location indicated by the ○ mark where the potential is monotonically increasing, the magnified waveform always fluctuates up and down as shown in FIG. 4. In FIG. 4, a part of the waveform without magnification is shown as sCH2. In this embodiment, the total number of plots (total number of steps) of the waveform data is N, and in FIGS. 3 and 4, N = 250,000. In this case, the short-term moving average (MA_SHORT) is obtained. MA_SHORT can be obtained by calculating the moving average with a resolution of one-tenth of the value obtained by dividing the number of data by the number of pixels. In FIGS. 3 and 4, when expressing the data of 250,000 rows of plots with 1000 pixels (when divided), the moving average is obtained every 25 steps with a resolution of 1 / 10.

[0025] Also, the long-term moving average (MA_LONG) is obtained. MA_LONG can be obtained by calculating the moving average with the resolution of the value obtained by dividing the number of data by the number of pixels. In FIGS. 3 and 4, when expressing the data of 250,000 rows of plots with 1000 pixels (when divided), the moving average is obtained every 250 steps with that value as the resolution.

[0026] In this embodiment, waveform data is obtained by the above-mentioned moving average, and feature quantities are obtained from this data. Examples of feature quantities include VDD (power supply voltage), the maximum voltage value (VMAX) and the minimum voltage value (VMIN) of the waveform data, Transition (transition rate), and Δpct, and can include EDGE, overshoot, undershoot (at rising), and Drop1 to 4.

[0027] The above VDD (power supply voltage) rounds the voltage data to two decimal places (0.00). If, after obtaining the mode value of this rounded data and removing the mode value from the voltage data, the ratio (40% or more) of the voltage data remaining in the stable section of the screen (GND / VDD state) is left, the extraction of the mode value is repeated again, and the higher value in the list of mode values is to be obtained as VDD. Needless to say, the maximum voltage value (VMAX) and the minimum voltage value (VMIN) of the waveform data are as such. In this way, the original data creation means 3 extracts the maximum value and the minimum value of the waveform as feature quantities. The Transition (mobility) is the value of what percentage of the whole is the window size (= number of steps of index) when crossing (going up and down) 99% of the potential in the waveform of MA_LONG (moving average), and what percentage of the whole is the window size (= number of steps of index) when crossing 1% of the potential in the waveform of MA_LONG (moving average). In this way, the original data creation means 3 extracts the number of steps at a predetermined height between the maximum value and the minimum value of the waveform as a feature quantity.

[0028] Δpct can be obtained by subtracting the index at 1% from the index when crossing 5%, 10%, 30%, 50%, 70%, 90%, 95% of the potential in the waveform of MA_LONG (moving average) and dividing by the total number of data (N = 250,000). Δpct may include those obtained in the same way for (MA_SHORT). For the waveform EDGE of MA_LONG (moving average), if (index at 99%) > (index at 1%), it can be set as "rising", and otherwise as "falling". As overshoot and undershoot (at rising), as overshoot, the maximum potential in the range of twice the window size of Transition from the index of 99% of the potential can be obtained, and as undershoot, the minimum potential in the range going back twice the window size of Transition from the index of 1% of the potential can be obtained. Drop1 to 4 can be obtained by dividing the steps from the index at 99% of the potential to the end into four parts and monitoring the potential values.

[0029] Including the processing by the original data creation means 3 as described above, the processing of the learning data creation apparatus according to the embodiment of the present invention is realized by the processing by the minor data increasing means 4, the waveform generation means 5, and the learning data generation means 6. That is, as shown in FIG. 5, the CPU 10 captures waveform data from the oscilloscope 26 via the data input interface 16 (S11). Using this waveform data, preprocessing such as moving average acquisition is performed as already described (S12). Further, as already described, feature extraction is performed (S13). At this time, for the waveform data, either a normal or abnormal class is obtained via the data input interface 16 or from the input device 24, and original data for balancing the classes is created.

[0030] Next, using the original data, minor data is increased to obtain a plurality of sets of new set data composed of a set of waveform feature amounts and classes (S14). This step S14 is performed by the minor data increasing means 4. The minor data increasing means 4 obtains a plurality of sets of new set data using the SMOTE (Synthetic Minority Oversampling TEchnique) method, which is known as a method for taking class balance.

[0031] Next, for each set of the obtained set data, a waveform is generated using the feature amounts that are set (S15). The generated waveform is displayed on the display device, and an input of the class of the displayed waveform is received (S16). That is, the generated and displayed waveform is visually inspected by a highly skilled inspector for normality / abnormality, and the normality / abnormality input is obtained from the input device 24.

[0032] In this way, the input class and the feature amount are taken as one set, and learning data for a plurality of sets created by SMOTE is generated (S17). The created data set is displayed and stored in the external storage device 23 and used for creating a classification device.

[0033] Next, a classification device according to an embodiment of the present invention will be described. When the classification device according to the embodiment of the present invention is configured using a computer, it can have the same configuration as shown in FIG. 1, similar to the learning data generation device according to the embodiment of the present invention.

[0034] However, the external storage device 23 stores programs and data for the operation of this classification device, and these can be appropriately read by the CPU 10 into the main memory 11 and used. Therefore, as shown in FIG. 6, the external storage device 23 stores programs for realizing the original data creation means 3, the minor data increasing means 4, the waveform generation means 5, the learning data generation means 6, the prediction model creation means 7, and the class acquisition means 8.

[0035] The prediction model creation means 7 creates a prediction model for predicting a class from feature amounts through learning using the learning data created by the above learning data creation device. The class acquisition means 8 obtains a class using the above prediction model based on the feature amounts of the waveforms obtained from the device to be inspected.

[0036] The classification device of this embodiment operates using a program corresponding to the flowchart shown in FIG. 7, and thus the operation will be described with reference to this flowchart. A prediction model for predicting a class from feature amounts is created through learning using the learning data created by the above learning data creation device (S21). As a result, as shown in FIG. 5, a prediction model (program) is completed in the external storage device 23.

[0037] Thereafter, the inspection of the device to be inspected will be carried out. First, waveform data is read from the new inspection target (S22). Here, the CPU 10 captures the waveform signal obtained by the oscilloscope 26 via the data input interface 16. Similar to what has already been described in the learning data generation device, preprocessing such as moving average acquisition is performed (S23). Further, similar to what has already been described in the learning data generation device, feature extraction is performed (S24). Next, a class is obtained using the prediction model based on the feature amount (S25).

[0038] According to this embodiment, a device for automatically determining a signal waveform by machine learning can be created using a waveform obtained with an inexpensive oscilloscope and even in a situation with little abnormal waveform data.

[0039] The waveform in this embodiment is not limited to the output voltage of the device, and of course, it may be the waveform of the temperature of the device, the noise of the device, the voice generated from the device, the volume generated from the device, or the vibration generated from the device. Also, needless to say, the feature amount is not limited to those described above, and various appropriate feature amounts can be adopted according to the target waveform. Furthermore, although two classes (abnormal and normal) are set, it is not limited to this, and three classes may be used, or four or more classes are also possible.

Explanation of Signs

[0040] 3 Original data creation means 4 Minor data increase means 5 Waveform generation means 6 Learning data generation means 7 Prediction model creation means 8 Class acquisition means 10 CPU 11 Main memory 12 Bus 13 External memory interface 14 Input interface 15 Display interface 16 Data input interface 22 Pointing device 23 External storage device 24 Input device 25 Display device 26 Oscilloscope

Claims

1. Original data creation means for obtaining a feature amount from a waveform already obtained from a device and creating original data for balancing classes together with the class of the waveform; Minor data increasing means for increasing minor data using the original data to obtain a plurality of sets of new set data composed of a set of a feature amount and a class of the waveform; Waveform generation means for generating a waveform using the feature amount set for each set of the set data obtained by the minor data increasing means; Learning data generation means for displaying the generated waveform on a display device, receiving an input of the class of the displayed waveform, and generating learning data with the input class and the feature amount as one set; A learning data creation device comprising the above.

2. The original data creation means creates a moving average by dividing the number of waveform data by the number of pixels and further reducing the value, and extracts a feature amount using the value of the moving average. The learning data creation device according to claim 1.

3. The original data creation means extracts the maximum value and the minimum value of the waveform as feature amounts. The learning data creation device according to claim 2.

4. The original data creation means extracts the number of steps at a predetermined height between the maximum value and the minimum value of the waveform as a feature amount. The learning data creation device according to claim 2.

5. The waveform is a waveform of the output voltage of the device, the temperature of the device, the noise of the device, the sound generated from the device, the volume generated from the device, and the vibration generated from the device. The learning data creation device according to claim 1.

6. The minor data increasing means obtains a plurality of sets of new set data using the SMOTE (Synthetic Minority Over-sampling Technique) method. The learning data creation device according to claim 1.

7. An original data creation step of obtaining a feature amount from a waveform already obtained from a device and creating original data for balancing classes together with the class of the waveform; A minor data increasing step of increasing minor data using the original data to obtain a plurality of sets of new set data composed of a set of a feature amount and a class of the waveform; For each set of the set data obtained by the negative data increasing step, a waveform generation step of generating a waveform using the feature amount set; A learning data generation step of displaying the generated waveform on a display device and accepting an input of the class of the displayed waveform, and generating learning data with the input class and the feature amount as one set; A learning data creation method characterized by comprising:

8. The original data creation step creates a moving average by dividing the number of waveform data by the number of pixels and further reducing the value, and extracts a feature amount using the value of the moving average. The learning data creation method according to claim 7.

9. The original data creation step extracts the maximum value and the minimum value of the waveform as feature amounts. The learning data creation method according to claim 8.

10. The original data creation step extracts the number of steps at a predetermined height between the maximum value and the minimum value of the waveform as a feature amount. The learning data creation method according to claim 8.

11. The waveform is a waveform of the output voltage of the device, the temperature of the device, the noise of the device, the sound generated from the device, the volume generated from the device, and the vibration generated from the device. The learning data creation method according to claim 7.

12. The negative data increasing step obtains a plurality of sets of new set data using the SMOTE (Synthetic Minority Over-sampling Technique) method. The learning data creation method according to claim 7.

13. A computer, Original data creation means for obtaining a feature amount from a waveform already obtained from a device and creating original data for balancing classes together with the class of the waveform; Negative data increasing means for increasing negative data using the original data to obtain a plurality of sets of new set data composed of a set of waveform feature amounts and classes; Waveform generation means for generating a waveform using the feature amount set for each set of the set data obtained by the negative data increasing means; Learning data generation means for displaying the generated waveform on a display device and accepting an input of the class of the displayed waveform, and generating learning data with the input class and the feature amount as one set; A program for creating learning data, characterized by causing a computer to function as such.

14. The program for creating learning data according to claim 13, wherein the computer is caused to function as the original data creation means, and a moving average is created by dividing the number of waveform data by the number of pixels and further reducing the value, and feature amounts are extracted using the value of this moving average.

15. The program for creating learning data according to claim 14, wherein the computer is caused to function as the original data creation means, and the maximum value and the minimum value of the waveform are extracted as feature amounts.

16. The program for creating learning data according to claim 14, wherein the computer is caused to function as the original data creation means, and the number of steps at a predetermined height between the maximum value and the minimum value of the waveform is extracted as a feature amount.

17. The program for creating learning data according to claim 13, wherein the waveform is a waveform of the output voltage of the device, the temperature of the device, the noise of the device, the voice generated from the device, the volume generated from the device, or the vibration generated from the device.

18. The program for creating learning data according to claim 13, wherein the computer is caused to function as minor data increasing means, and a plurality of sets of new set data are obtained using the SMOTE (Synthetic Minority Over-sampling Technique) method.

19. The learning data creation device according to claim 1, prediction model creation means for creating a prediction model that predicts a class from feature amounts by learning using the learning data created by the learning data creation device, class acquisition means for obtaining a class using the prediction model based on the feature amounts of the waveform obtained from the device to be inspected, A classification device, characterized by comprising the above.

Images