JP7702799B2 - DATA EXTRACTION DEVICE, LEARNING MODEL CONSTRUCTION DEVICE, DATA EXTRACTION METHOD, LEARNING MODEL CONSTRUCTION METHOD, AND PROGRAM - Google Patents

Description

The present disclosure relates to a data extraction device, a learning model construction device, a data extraction method, a learning model construction method, and a program.

In recent years, deep learning has been increasingly used. For example, Patent Document 1 discloses a method for improving detection accuracy in target detection in an image using deep learning. In addition, for example, high accuracy may be achieved by using deep learning for multi-value classification problems of time-series data, and the application of CNN (Convolutional Neural Network), for example, is expected. Training and evaluation of CNN requires a significant amount of time. For example, when building a discrimination model for discriminating the operating state of a plant from the plant's process data, building a discrimination model for each plant requires an enormous amount of time for training and evaluation.

Chinese Patent Publication No. 106934346

There is a need for a method to reduce the time required to build a learning model.

The present disclosure provides a data extraction device, a learning model construction device, a data extraction method, a learning model construction method, and a program that can solve the above problems.

The data extraction device disclosed herein is a data extraction device that extracts a portion of learning data from a plurality of learning data cut out for each time window from time series data, and includes a data acquisition unit that acquires the plurality of learning data, a similarity calculation unit that calculates a similarity between the learning data based on features of the learning data excluding temporal differences in the time window, and a data reduction unit that classifies the learning data with high similarity from each other and deletes a portion of the learning data. Furthermore, the similarity calculation unit may calculate the similarity based on the size per unit time of the area where the regions related to the two learning data overlap with respect to a region sandwiched between the waveform data indicated by each learning data and the time axis when the two learning data are shifted in the time axis direction. Furthermore, the similarity calculation unit may calculate the similarity between the two shapes based on the shapes of the respective frequency distributions obtained by frequency analysis of the two learning data. Furthermore, the data extraction device may further include a classification unit that classifies similar pieces of the plurality of learning data into the same group based on a predetermined evaluation scale, and the similarity calculation unit may calculate the similarity between the learning data belonging to the same group .

The learning model construction device of the present disclosure includes the above-mentioned data extraction device and a learning unit that learns the learning data and constructs a learning model, and the learning unit learns the learning data extracted by the data extraction device to construct the learning model, and then adjusts the parameters of the learning model using the learning data before extraction.

The data extraction method disclosed herein is a data extraction method for extracting a portion of learning data from a plurality of learning data cut out for each time window from time series data, and includes the steps of acquiring a plurality of learning data, calculating a similarity between the learning data based on the characteristics of the learning data excluding the temporal difference in the time window, and classifying the learning data having a high similarity from each other and deleting a portion of the learning data. Furthermore, in the step of calculating the similarity, the similarity may be calculated based on the size per unit time of the area where the regions related to the two learning data overlap with respect to the region sandwiched between the waveform data and the time axis indicated by each learning data when the two learning data are shifted in the time axis direction. In addition, in the step of calculating the similarity, the similarity between the two shapes may be calculated based on the shapes of the respective frequency distributions obtained by frequency analysis of the two learning data. In addition, the method further includes the step of classifying similar multiple learning data into the same group based on a predetermined evaluation scale, and in the step of calculating the similarity, the similarity between the learning data belonging to the same group may be calculated .

The learning model construction method disclosed herein constructs a learning model by learning the learning data extracted by the data extraction method described above, and then adjusts the parameters of the learning model using the learning data before extraction.

The program of the present disclosure causes a computer to execute a data extraction process for extracting a portion of learning data from a plurality of learning data cut out for each time window from time series data, the data extraction process including the steps of acquiring a plurality of learning data, calculating a similarity between the learning data based on the characteristics of the learning data excluding the temporal difference in the time window, and classifying the learning data having high similarity from each other and deleting a portion of the learning data. Furthermore, in the step of calculating the similarity, when the two learning data are shifted in the time axis direction, the similarity may be calculated based on the size per unit time of the area where the two learning data overlap with respect to the area sandwiched between the waveform data indicated by each learning data and the time axis. In addition, in the step of calculating the similarity, the similarity between the two shapes may be calculated based on the shapes of the respective frequency distributions obtained by frequency analysis of the two learning data. In addition, the data extraction process executed by the program may further include a step of classifying similar pieces of learning data into the same group based on a predetermined evaluation measure, and the step of calculating the similarity may calculate the similarity between the learning data belonging to the same group .

The program disclosed herein causes a computer to learn the learning data extracted by the data extraction process described above to construct a learning model, and then execute a process of adjusting the parameters of the learning model using the learning data before extraction.

The above-mentioned data extraction device, learning model construction device, data extraction method, learning model construction method, and program can reduce the time required to construct a learning model by reducing the amount of learning data.

FIG. 1 is a block diagram showing an example of a learning model construction device according to a first embodiment. FIG. 2 is a diagram illustrating an outline of a data extraction method according to the first embodiment. 5A to 5C are diagrams illustrating a method of calculating a similarity using normalized cross-correlation according to the first embodiment. 5A to 5C are diagrams illustrating a method of calculating a similarity based on a frequency distribution according to the first embodiment. FIG. 2 is a first diagram illustrating the data reduction method according to the first embodiment. FIG. 2 is a second diagram illustrating the data reduction method according to the first embodiment. FIG. 11 is a third diagram illustrating the data reduction method according to the first embodiment. FIG. 4 is a fourth diagram illustrating the data reduction method according to the first embodiment. 11 is a flowchart illustrating an example of a data extraction process according to the first embodiment. 1 is a flowchart illustrating an example of a learning model construction process according to the first embodiment. FIG. 11 is a block diagram showing an example of a learning model construction device according to a second embodiment. 13 is a flowchart illustrating an example of a data extraction process according to the second embodiment. FIG. 2 is a diagram illustrating an example of a hardware configuration of a learning model construction device according to an embodiment.

First Embodiment
Hereinafter, a data extraction method and a learning model construction method according to a first embodiment of the present disclosure will be described with reference to FIGS. 1 to 7.
(composition)
FIG. 1 is a block diagram illustrating an example of a learning model construction device according to an embodiment.
As shown in the figure, the learning model construction device 10 includes a data acquisition unit 11, a data extraction unit 12, a learning unit 13, an input unit 14, an output unit 15, and a memory unit 16.
The data acquisition unit 11 acquires learning data used to construct a learning model. The learning data is, for example, data extracted for each time window of a predetermined time width from time series data such as temperature, pressure, and diversion collected from a monitored device or plant. Here, reference is made to FIG. 2. An example of a process (random sampling) for extracting learning data from time series data is shown in the upper diagram 201 of FIG. 2. The variables A to C are time series data of various parameters such as temperature, pressure, and flow rate. In random sampling, learning data is extracted one by one from the time series data of each parameter with a time window having a common start time and end time, and the time window is arbitrarily shifted to extract multiple learning data. The data acquisition unit 11 acquires data extracted with a common time window for each of the variables A to C. These data are used as learning data for constructing a learning model. Here, the variables A to C are used as an example for explanation, but the number of variables is not limited to three, and may be two or less, or four or more.

The data extraction unit 12 extracts data from the learning data acquired by the data acquisition unit 11 so that the learning accuracy does not decrease. Here, reference is made to the lower diagram 202 of FIG. 2. The data extraction unit 12 extracts only representative data (e.g., W1) from the learning data cut out for each time frame W1 to W9, so as not to include overlapping data (e.g., W1 and W4) that are determined to have the same characteristics in a time series. The data extraction unit 12 includes a similarity calculation unit 121 and a data reduction unit 122. The similarity calculation unit 121 calculates the similarity of data based on (1) normalized cross-correlation, (2) frequency distribution, etc. The method of calculating the similarity will be described later with reference to FIG. 3 to FIG. 4. The data reduction unit 122 reduces data with overlapping characteristics in a time series from the learning data based on the similarity between the data. The reduction method from the learning data will be described later with reference to FIG. 5A to FIG. 5D. Here, data with overlapping features in a time series does not only mean data with similar features when data extracted for each time frame is compared, but also includes data with similar features when the data extracted in the time frame is shifted along the time axis (the data after extraction in a common time frame is shifted, rather than shifting the time frame along the time axis). When constructing a learning model in which time-series information (information before and after time) of the learning data is eliminated, such as in deep learning such as CNN, it is considered that even if data with overlapping features in a time series is reduced, a learning model corresponding to the feature can be constructed by leaving a small amount of learning data having the feature. By excluding data with overlapping features in a time series from the learning data and narrowing down the entire learning data to a small number of learning data having the features necessary for building a learning model, the learning time can be shortened.

The learning unit 13 learns the learning data through deep learning and constructs a learning model.
The input unit 14 is configured using input devices such as a keyboard, a mouse, a touch panel, buttons, etc. The input unit 14 accepts information input using the input devices, and outputs the information to the data extraction unit 12, the learning unit 13, etc.
The output unit 15 outputs various data to an electronic file or a display device. For example, the output unit 15 may output the learning data extracted by the data extracting unit 12 as an electronic file.
The storage unit 16 stores data necessary for extracting learning data and constructing a learning model.

[Similarity calculation]
The similarity calculation unit 121 calculates the similarity between the learning data by using the following normalized cross-correlation or frequency component distribution.

(1) Similarity Based on Normalized Cross-Correlation Next, a method for calculating similarity based on normalized cross-correlation will be described with reference to Fig. 2 and Fig. 3. The waveform data L1 and waveform data L2 in Fig. 3 are data of variable A in time frame W1 and variable A in time frame W4, respectively, extracted from time series data, normalized so that the average of the data values is 0 and the standard deviation is 1. For example, the similarity calculation unit 121 calculates the area P1 of the overlapping portion between the waveform data L1 and the waveform data L2 while shifting the data in the time axis direction. Then, the similarity calculation unit 121 divides the area P1 of the overlapping portion by the time T1 of the overlapping portion to calculate the unit area of the overlapping portion. In the same manner as for the variable A, the similarity calculation unit 121 calculates normalized waveform data for the data of the time frame W1 and the data of the time frame W4 of the variables B and C, respectively, and calculates the unit area of the portion where the waveform data calculated for the time frame W1 and the waveform data calculated for the time frame W4 overlap when the time frame W4 is shifted in the time direction by the same amount as the variable A. The unit area for the variable B at this time is P2, and the unit area for the variable C is P3. The similarity calculation unit 121 calculates the average value of the unit areas P1 to P3 while adjusting the shift amount in the time direction of the data of the time frame W4, and searches for the shift amount of the time frame W4 that maximizes the average value. Then, the similarity calculation unit 121 determines the average value when the average value of the unit areas P1 to P3 is maximized as the similarity between the data of the time frame W1 and the data of the time frame W4. By such calculation, if there is a period in which the trends are similar even between waveforms with a phase difference, the value of the similarity becomes large. The similarity calculation unit 121 uses the same method to calculate the similarity between time frame W1 and time frames W2 to W9, calculates the similarity between time frame W2 and time frames W3 to W9, and so on for all patterns of combinations of time frames. Note that if the time series values of one learning data are all the same value, normalization is not possible, so the similarity of the calculation target variable is set to 1.

(2) Similarity Based on Frequency Component Distribution Next, a method of calculating similarity based on the distribution of frequency components will be described with reference to FIG. 4. For example, the similarity calculation unit 121 performs frequency analysis by fast Fourier transform on each of data obtained by normalizing data of variable A extracted in time frame W1 and data obtained by normalizing data of variable A extracted in time frame W4. The similarity calculation unit 121 calculates similarity using frequency distribution based on the data after frequency analysis. FIG. 4 shows data Q1 to Q3 obtained by frequency analysis of data of different time frames. The vertical axis of the graph in FIG. 4 indicates the magnitude of the frequency components, and the horizontal axis indicates the frequency. For example, the value of the vertical axis of data Q1 indicates the average value (average in the time axis direction) of the magnitude of each frequency component in time frame W1. When comparing data Q1 to Q3, the shapes of the waveforms are similar even if the magnitudes of the frequency components are different. For example, in any of data Q1 to Q3, in the frequency range of f1, the magnitude of the frequency components increases as the frequency increases, and in the range of f2, the magnitude of the frequency components is roughly constant even if the frequency changes. In this way, when the trends of the frequency distribution are similar, the similarity between the data Q1 to Q3 is high, and when the trends of the frequency distribution are different, the similarity is low. Specifically, the similarity calculation unit 121 calculates the cosine similarity of vectors V1 and V3 composed of each frequency component of the data Q1 and data Q3 in the frequency range f3, for example. The similarity calculation unit 121 also calculates the cosine similarity of vectors (for example, V1a and V3a) composed of each frequency component in the other frequency ranges. For example, the similarity calculation unit 121 sets the average of the cosine similarities calculated for each variable A to C in a predetermined range as the similarity between the data Q1 and data Q3. More specifically, the similarity calculation method using the fast Fourier transform is as follows. (Step 1) A fast Fourier transform is performed on the two data to calculate the real number and imaginary number of each frequency component. (Step 2) A cosine similarity is calculated as the similarity, with the coefficients of the real number and imaginary number calculated above as feature quantities.

The predetermined range is, for example, when the data contains a lot of noise (mainly high-frequency components), a range of frequencies other than the range of frequencies containing a lot of noise (high frequencies). In addition, when the frequency component value for a certain frequency range is 0 or a low value, the cosine similarity may be calculated excluding that range. In addition, for example, the similarity may be calculated using only the frequency range corresponding to each operation mode (starting, rated operation, partial load operation, stopped, etc.) of the target equipment or plant. For example, when in a transient operation state (starting, stopping, changing operation mode), the similarity may be calculated using data in the low frequency range. In this way, the similarity calculation unit 121 calculates the similarity based on the frequency distribution of each of the time frame W1 and the time frames W2 to W9, calculates the similarity of each of the time frame W2 and the time frames W3 to W9, and so on, calculating the similarity for all patterns of combinations of time frames. Note that in the similarity based on the frequency component distribution, the feature amount (frequency component value of each frequency) and similarity are calculated by originally excluding the context in the time axis direction.

Based on the idea that the features of a certain piece of data can be substituted by other data with a high degree of similarity, the data extraction unit 12 uses the data reduction unit 122 to reduce data with overlapping time-series features based on the similarity calculated by the similarity calculation unit 121, and extracts a portion of data with features necessary for constructing a learning model. Next, the data reduction process will be described with reference to Figures 5A to 5D.

[Data reduction process]
Table 500 in FIG. 5A shows the result of sorting the similarity between each data calculated by the similarity calculation method (1) or (2) for data A to I cut out in each time frame. For example, the similarity between data A and data B is 0.9, the similarity between data A and data C is 0.6, and the similarity between data A and data I is 0.2. Here, the data reduction unit 122 classifies each data into similar data and dissimilar data based on a predetermined threshold (for example, 0.9). Groups of similar data after classification are shown in table 501 in FIG. 5B. Also, a distribution based on the feature amount showing the similarity of each data by the distance between data is shown in FIG. 5C. Note that data A to I belong to class 1, and data A' belongs to class 2. For example, class 1 is a set of data of each variable such as pressure and flow rate when the plant state is normal, and class 2 is a set of data when the plant state is abnormal. Alternatively, class 1 is a set of data for each variable when the plant is operating at rated power, and class 2 is a set of data at the time of startup. The data reduction unit 122 deletes learning data within each class. Here, data deletion within class 1 will be described.

The data reduction unit 122 selects data having many data with high similarity and deletes the selected data. With reference to table 501 in FIG. 5B, for example, the number of data in class 1 similar to data A is 1 (B), the number of data in class 1 similar to data C is 3 (G, H, I), and the number of data in class 1 similar to data H is 5 (C, E, F, G, I). Also, with reference to FIG. 5C, it can be seen that data C, E, F, G, and I with similarity of 0.9 or more are distributed around data H. In such a case, the data reduction unit 122 deletes data H. The relationship between the data after deletion is shown in table 502 in FIG. 5B. With reference to table 502, the number of data similar to data B is 2 (A, C), the number of data similar to data C is 2 (B, I), and the number of similar data for other data is 1 or less. For the data group in table 502, the data reduction unit 122 repeatedly performs a process of deleting data having many data with high similarity. For example, the data reduction unit 122 deletes either data B or data C (either is acceptable). Here, it is assumed that data B is deleted. FIG. 5D shows the relationship between the data after data B is deleted. In FIG. 5D, the number of pieces of data similar to each piece of data is one or less. By repeating such processing, data with overlapping time-series features is deleted. The data reduction unit 122 ends the data reduction when the number of pieces of training data can be sufficiently reduced while preserving the features of the original training data group. For example, the data reduction unit 122 ends the data deletion when the total number of pieces of data becomes a predetermined number or less, or when the number of pieces of data similar to each piece of data becomes a predetermined number or less. The data extraction unit 12 extracts the data remaining after the data with overlapping time-series features is deleted as training data for learning the rough structure of the training model.

(operation)
[Data extraction process]
Next, the flow of the data extraction process of this embodiment will be described with reference to FIG.
FIG. 6 is a flowchart illustrating an example of the data extraction process of the first embodiment.
First, the data acquiring unit 11 acquires a plurality of pieces of learning data cut out for each time window from each of the time series data of a plurality of parameters (e.g., pressure, flow rate, etc.) (step S10). The data acquiring unit 11 may acquire the time series data and cut out the data for each time window to acquire the learning data. The data acquiring unit 11 records the plurality of pieces of learning data cut out for each time frame in the storage unit 16.

Next, the data extraction unit 12 classifies the learning data into classes (step S20). For example, if the user wishes to separate the learning data into data when the plant is in a normal operating state (class 1) and data when the plant is in an abnormal operating state (class 2), the user inputs the time periods when the plant is in a normal operating state and the time periods when the plant is in an abnormal operating state to the learning model construction device 10. The input unit 14 outputs the input information about the time periods to the data extraction unit 12. The data extraction unit 12 classifies the learning data acquired in step S10 into either class 1 or class 2, depending on the time period set by the user.

Then, the data extraction unit 12 performs a process (steps S30 to S60) of excluding data with overlapping time-series features from the multiple learning data sets and extracting only the small amount of learning data necessary to accurately construct the general structure of the learning model. The data extraction unit 12 performs this process for each class.

First, the similarity calculation unit 121 calculates (1) a similarity based on normalized cross-correlation, or (2) a similarity based on frequency distribution (step S30). The similarity calculation unit 121 calculates the similarity by method (1) or (2) for all combination patterns in which two pieces of learning data belonging to the same class classified in step S20 are selected and combined, and records the similarity between each piece of learning data in the memory unit 16. This results in data such as table 500 in FIG. 5A. The similarity calculation unit 121 may calculate the similarity by both methods (1) and (2) and use the weighted average of these as the similarity.

Next, the data reduction unit 122 tabulates combinations of learning data with high similarity (step S40). The data reduction unit 122 associates each learning data with other learning data that is equal to or greater than a predetermined threshold (e.g., 0.9 or greater), and separates the data into those with high similarity. The data reduction unit 122 records the separation results in the storage unit 16. This process results in data such as that shown in table 501 of FIG. 5B.

Next, the data reduction unit 122 deletes the training data that has many data with high similarity (step S50). When the training data as shown in FIG. 5C is obtained, the data reduction unit 122 deletes data H. By deleting data in this way, it is possible to reduce the number of training data while retaining the features of the training data group (while suppressing the loss of features). In the case of the data distribution shown in FIG. 5C, the amount of data can be reduced while leaving the training data scattered around the boundary of class 1 by leaving data C, I, G, F, etc. and deleting data H. As another method of reducing the training data, it is possible to select training data that has many data with high similarity, leave the selected data, and delete other similar training data. For example, in the case of table 501 in FIG. 5B, with respect to data H, data H is left and data C, E, F, G, and I are deleted. As a result, in the case of the example in FIG. 5C, more data can be reduced.

Next, the data reduction unit 122 determines whether or not the deletion of the learning data is complete (step S60). For example, the data reduction unit 122 determines that the deletion of the learning data is complete when the number of pieces of data similar to each piece of data is equal to or less than a predetermined number. Alternatively, the data reduction unit 122 may determine that the deletion of the learning data is complete when the number of pieces of remaining learning data is equal to or less than a predetermined number, or when the cumulative number of pieces of deleted data reaches a predetermined number. If the deletion of the learning data is not complete (step S60; No), the process from step S40 is repeated. If the deletion of the learning data is complete (step S60; Yes), the data extraction unit 12 records the extracted learning data in the storage unit 16 (step S70). The data extraction unit 12 records the remaining learning data that has not been deleted as extracted learning data in the storage unit 16.

[Learning model construction process]
Next, the process of constructing a learning model will be described.
FIG. 7 is a flowchart illustrating an example of a learning model construction process according to the first embodiment.
First, the data acquiring unit 11 acquires learning data extracted for each time frame from the time-series data of a plurality of parameters (step S1).
Next, the data extraction unit 12 extracts learning data (step S2). The data extraction unit 12 extracts learning data by the process described with reference to FIG.

Next, the learning unit 13 learns the data extracted by the data extraction unit 12 (step S3). The learning unit 13 reads the extracted learning data from the storage unit 16 and constructs a learning model by deep learning such as CNN. For example, the learning unit 13 learns teacher data labeled with information indicating the plant state during the time window (normal, abnormal, event XX has occurred, etc.) for the learning data of multiple parameters of a common time window, and constructs a discrimination model for discriminating events occurring in the plant. This determines approximate values of the hyperparameters of the discrimination model. In this step, since the learning data is narrowed down, learning and evaluation can be performed in a short time compared to the case where learning is performed using all the learning data. In addition, since data that retains the characteristics of the original learning data group is extracted by the process of FIG. 6, a learning model (discrimination model) can be constructed without reducing accuracy.

Next, the learning unit 13 learns all the learning data (step S4). The learning unit 13 proceeds with learning using a large amount of learning data for the learning model constructed in step S3, and adjusts the determined hyperparameter values. Since the hyperparameter values have already been determined in step S3, learning and evaluation can be performed in a relatively short time even if learning is performed using all the learning data. Furthermore, by performing learning using a large amount of data, the accuracy of the learning model constructed in step S3 can be improved. According to the process of FIG. 7, by performing the process of step S3 before constructing the learning data using all the learning data, the learning model can be constructed in a short time compared to the case of constructing the learning model using all the learning data from the beginning.

(effect)
As described above, according to this embodiment, the number of training data can be reduced while maintaining the characteristics of the original training data group. By reducing the training data while maintaining accuracy, the learning time required to build a learning model can be shortened. Assuming that the structure of a deep learning model depends on the variation in the feature amount of the training data, rather than on the amount of training data, a model structure that can obtain the desired accuracy is specified using the training data extraction method of this embodiment, and then a large amount of training data is used to fine-tune each parameter while improving the accuracy of the model, thereby efficiently building a learning model.

The learning data extraction method of this embodiment is effective when it is not possible to find a model that can be learned or a highly accurate model structure, and it is necessary to make several changes to the model structure before evaluating it. In addition, the learning data extraction method of this embodiment is effective for a method (such as deep learning) in which the time sequence of the features of the learning data is ignored (time sequence information is excluded) when learning time series data. In addition, it can be applied to a system that handles a multi-value classification problem of time series data, such as a discrimination model that discriminates abnormal events in a plant. In the embodiment, an example is given in which similarity is calculated using normalized cross-correlation and frequency distribution to determine whether data extracted from a time window has the same characteristics in a time series, but the similarity may be calculated by other methods. In addition, in the above embodiment, supervised deep learning is used as an example, but it can also be applied to unsupervised deep learning.

Second Embodiment
A learning model construction device 10A according to a second embodiment of the present invention will be described below with reference to FIGS.
In the first embodiment, the similarity is calculated for all combinations of learning data. With this method, when the number of learning data becomes huge, the calculation load and calculation time may become huge. Therefore, in the second embodiment, the calculation of the similarity between data with low similarity is omitted to reduce the amount of calculation.

(composition)
FIG. 8 is a block diagram illustrating an example of a learning model construction device according to the second embodiment.
Among the components of the learning model construction device 10A according to the second embodiment of the present invention, the same functional units as those constituting the learning model construction device 10 according to the first embodiment of the present invention are denoted by the same reference numerals, and their description will be omitted. The learning model construction device 10A according to the second embodiment includes a data extraction unit 12A instead of the data extraction unit 12 of the first embodiment. The data extraction unit 12A includes a classification unit 123 in addition to a similarity calculation unit 121 and a data reduction unit 122 similar to those of the first embodiment.

The classification unit 123 classifies the learning data of the same class into groups (clusters) of similar data. For example, the classification unit 123 classifies learning data with similar trends in frequency component distribution into the same cluster by (1) normalized cross-correlation or (2) any clustering method described in the first embodiment. The similarity calculation unit 121 calculates the similarity between combinations of learning data belonging to the same cluster.

(operation)
Next, the data extraction process of the second embodiment will be described.
FIG. 9 is a flowchart illustrating an example of the data extraction process according to the second embodiment.
The same processes as those in the first embodiment described with reference to FIG. 6 are denoted by the same reference numerals, and the description thereof will be omitted. First, the data acquisition unit 11 acquires a plurality of pieces of learning data (step S10). Next, the data extraction unit 12 classifies the learning data into classes (step S20). Next, the classification unit 123 performs clustering on the learning data in the same class (step S25). The classification unit 123 classifies similar learning data belonging to class 1 into the same cluster based on a predetermined evaluation measure (for example, similarity of features not considering the chronological relationship between the learning data) using any clustering method (for example, kmeans method). For example, the classification unit 123 calculates similarity based on (1) normalized cross-correlation or (2) frequency component distribution, and classifies those with high similarity into one of a plurality of clusters. The classification unit 123 performs clustering for each class, but if the number of learning data in a certain class is less than a threshold, the processing of step S31 may not be performed for that class. The classification unit 123 records, for each of the learning data that has been subjected to the clustering, in the storage unit 16, identification information of the cluster to which the learning data belongs.

Next, the similarity calculation unit 121 calculates the similarity using methods (1) or (2) for all combination patterns in which two pieces of learning data classified into the same cluster are selected and combined (step S31), and records the similarity between each piece of learning data in the memory unit 16. If the similarity is calculated using a similar method in step S25, the calculation result can be used. By limiting the calculation of the similarity to within the same cluster of the same class, the calculation cost can be reduced compared to step S30 in the first embodiment.

Next, the data reduction unit 122 tabulates combinations of learning data with high similarity (step S40). In step S25, similar learning data is classified based on features that do not take into account chronological context. Therefore, even if the range of similarity calculation is limited in step S31, it is considered that the tabulation result in step S40 is not affected (for example, data with high similarity cannot be distinguished from each other in table 501 in FIG. 5B) and the accuracy of extraction of learning data is not reduced. Next, the data reduction unit 122 deletes learning data that has many data with high similarity (step S50). The data reduction unit 122 repeats data reduction until a predetermined completion condition is satisfied (step S60). The data extraction unit 12 records the remaining learning data that has not been deleted as extracted learning data in the storage unit 16 (step S70). The process of constructing a learning model using the extracted learning data is the same as that in the first embodiment (FIG. 7).

According to the second embodiment, when the number of combinations between learning data is enormous, clustering is performed and the time required for similarity calculation is reduced by omitting similarity calculation with learning data belonging to other clusters that are considered to have low similarity. This makes it possible to reduce the time required for data extraction processing in addition to the effects of the first embodiment.
In addition, in the clustering by the kmeans method, in order to estimate the cluster to which each piece of learning data belongs, it is necessary to construct a kmeans model. When constructing the kmeans model, all learning data may be used, but if it takes time to construct the model, the kmeans model may be constructed using representative learning data.

FIG. 10 is a diagram illustrating an example of the hardware configuration of a learning model construction device according to each embodiment.
The computer 900 includes a CPU 901 , a main memory device 902 , an auxiliary memory device 903 , an input/output interface 904 , and a communication interface 905 .
The above-mentioned learning model construction devices 10 and 10A are implemented in a computer 900. The above-mentioned functions are stored in the auxiliary storage device 903 in the form of a program. The CPU 901 reads the program from the auxiliary storage device 903, expands it in the main storage device 902, and executes the above-mentioned processing according to the program. The CPU 901 also secures a memory area in the main storage device 902 according to the program. The CPU 901 also secures a memory area in the auxiliary storage device 903 for storing data being processed according to the program.

In addition, a program for realizing all or part of the functions of the learning model construction device 10, 10A may be recorded on a computer-readable recording medium, and the program recorded on the recording medium may be read into a computer system and executed to perform processing by each functional unit. The term "computer system" as used herein includes hardware such as an OS and peripheral devices. In addition, if a WWW system is used, the term "computer system" also includes a homepage providing environment (or display environment). In addition, the term "computer-readable recording medium" refers to portable media such as CDs, DVDs, and USBs, and storage devices such as hard disks built into a computer system. In addition, when the program is distributed to a computer 900 via a communication line, the computer 900 that receives the program may expand the program into the main storage device 902 and execute the above processing. In addition, the above program may be for realizing part of the above-mentioned functions, and may further be capable of realizing the above-mentioned functions in combination with a program already recorded in the computer system.
In addition, the learning model construction device 10, 10A may be composed of multiple computers 900.

As described above, several embodiments of the present disclosure have been described, but all of these embodiments are presented as examples and are not intended to limit the scope of the invention. These embodiments can be implemented in various other forms, and various omissions, substitutions, and modifications can be made without departing from the gist of the invention. These embodiments and their modifications are included in the scope of the invention and its equivalents as described in the claims, as well as in the scope and gist of the invention.

<Additional Notes>
The data extraction device, the learning model construction device, the data extraction method, the learning model construction method, and the program described in each embodiment can be understood, for example, as follows.

(1) A data extraction device (learning model construction device 10) according to a first aspect is a data extraction device that extracts a portion of learning data from a plurality of learning data cut out for each time window from time-series data, and includes a data acquisition unit 11 that acquires the plurality of learning data, a similarity calculation unit 121 that calculates a similarity between the learning data based on features of the learning data, excluding temporal differences in the time window, and a data reduction unit 122 that separates the learning data with high similarity from each other and deletes a portion of the learning data.
This allows learning methods (such as deep learning) to ignore the temporal context of the features of the training data (eliminating chronological information), effectively reducing the amount of training data while preserving the features of the original training data group.

(2) A data extraction device (learning model construction device 10) according to a second aspect is the data extraction device of (1), in which the data reduction unit 122 associates, for each of the learning data, other learning data whose similarity is equal to or greater than a threshold, and deletes the learning data with which the other learning data is most frequently associated.
This makes it possible to reduce the training data without compromising the characteristics of the training data set.

(3) A data extraction device (learning model construction device 10) according to a third aspect is a data extraction device according to (1) to (2), wherein the similarity calculation unit 121 calculates the similarity based on the size per unit time of the area of overlap between the areas relating to the two learning data when the two learning data are shifted in the time axis direction, for the area sandwiched between the waveform data indicated by each learning data and the time axis.
This makes it possible to calculate the similarity between training data based on features that exclude temporal differences.

(4) A data extraction device (learning model construction device 10) according to a fourth aspect is a data extraction device according to any one of (1) to (3), wherein the similarity calculation unit 121 calculates the similarity between the two shapes based on the shapes of respective frequency distributions obtained by frequency analysis of the two pieces of learning data.
This makes it possible to calculate the similarity between training data based on time-independent features (features that exclude differences before and after in terms of time).

(5) A data extraction device (learning model construction device 10) according to a fifth aspect is a data extraction device according to any one of (1) to (4), further comprising a classification unit 123 that classifies similar pieces of learning data into the same group (cluster) based on a predetermined evaluation measure (e.g., similarity of features without taking into account the chronological relationship between the learning data), and the similarity calculation unit 121 calculates the similarity between the learning data belonging to the same group (cluster).
This makes it possible to prevent the computational cost of similarity calculation from becoming enormous.

(6) A data extraction device (learning model construction device 10) according to a sixth aspect is a data extraction device according to any one of (1) to (5), in which the learning data is learning data used for deep learning.
This allows the amount of training data for deep learning to be reduced. The reduced training data can be used to determine the rough structure of a deep learning model.

(7) A learning model construction device 10 according to a seventh aspect includes a data extraction device as described in (1) to (6) and a learning unit that learns the learning data and constructs a learning model. The learning unit constructs the learning model using the learning data extracted by the data extraction device, and then adjusts parameters of the learning model using the learning data before extraction.
This reduces the time required to build a learning model. In addition, the accuracy of the learning model can be ensured by adjusting the parameters of the learning model using all data.

(8) The data extraction method according to the eighth aspect is a data extraction method for extracting a portion of learning data from a plurality of learning data cut out for each time window from time series data, and includes the steps of acquiring the plurality of learning data, excluding temporal differences in the time window and calculating the similarity between the learning data based on the characteristics of the learning data, and separating the learning data with high similarity from each other and deleting a portion of the data.

(9) A learning model construction method according to a ninth aspect learns the learning data extracted by the data extraction method of (8) to construct the learning model, and then adjusts the parameters of the learning model using the learning data before extraction.

(10) The program according to the tenth aspect causes a computer 900 to execute a data extraction process for extracting a portion of learning data from a plurality of pieces of learning data cut out from time series data for each time window, the data extraction process having the steps of acquiring the plurality of pieces of learning data, calculating the similarity between the pieces of learning data based on the characteristics of the learning data while excluding the temporal differences in the time window, and separating the pieces of learning data with high similarity from each other and deleting a portion of the pieces of learning data.

(11) The program according to the eleventh aspect causes the computer 900 to execute a process of learning the learning data extracted by the data extraction process of (8) to construct the learning model, and then adjusting the parameters of the learning model using the learning data before extraction.

10, 10A... Learning model construction device 11... Data acquisition unit 12, 12A... Data extraction unit 121... Similarity calculation unit 122... Data reduction unit 123... Classification unit 13... Learning unit 14... Input unit 15... Output unit 16... Memory unit 900... Computer 901... CPU
902: Main memory device 903: Auxiliary memory device 904: Input/output interface 905: Communication interface

Claims (14)

A data extraction device that extracts a portion of learning data from a plurality of learning data extracted for each time window from time-series data, comprising:
A data acquisition unit that acquires a plurality of the learning data;
a similarity calculation unit that calculates a similarity between the learning data based on features of the learning data excluding temporal differences in the time window;
a data reduction unit that classifies the learning data with high similarity from each other and deletes a part of the learning data;
Equipped with
the similarity calculation unit calculates the similarity based on a size per unit time of an overlapping area between the two learning data and waveform data indicated by each learning data and a time axis when the two learning data are shifted in a time axis direction.
Data Extraction Device. A data extraction device that extracts a portion of learning data from a plurality of learning data extracted for each time window from time-series data, comprising:
A data acquisition unit that acquires a plurality of the learning data;
a similarity calculation unit that calculates a similarity between the learning data based on features of the learning data excluding temporal differences in the time window;
a data reduction unit that classifies the learning data with high similarity from each other and deletes a part of the learning data;
Equipped with
the similarity calculation unit calculates a similarity between the two shapes based on respective shapes of frequency distributions obtained by performing a frequency analysis on the two pieces of learning data ;
Data Extraction Device. A data extraction device that extracts a portion of learning data from a plurality of learning data extracted for each time window from time-series data, comprising:
A data acquisition unit that acquires a plurality of the learning data;
A classification unit that classifies the plurality of learning data into the same group so that similar data are classified based on a predetermined evaluation criterion;
a similarity calculation unit that calculates a similarity between the learning data based on features of the learning data excluding temporal differences in the time window;
a data reduction unit that classifies the learning data with high similarity from each other and deletes a part of the learning data;
Equipped with
the similarity calculation unit calculates the similarity between the learning data belonging to the same group.
Data Extraction Device. the data reduction unit associates, for each of the learning data, other learning data whose similarity is equal to or greater than a threshold, and deletes the learning data to which the other learning data is most frequently associated.
The data extraction device according to any one of claims 1 to 3 . The learning data is learning data used for deep learning.
The data extraction device according to any one of claims 1 to 4 . A data extraction device according to any one of claims 1 to 5 ;
A learning unit that learns the learning data and constructs a learning model,
The learning unit learns the learning data extracted by the data extraction device to construct the learning model, and then adjusts parameters of the learning model using the learning data before extraction.
Learning model construction device. A data extraction method for extracting a portion of learning data from a plurality of learning data extracted for each time window from time series data, comprising the steps of:
acquiring a plurality of said learning data;
calculating a similarity between the learning data based on features of the learning data excluding temporal differences in the time window;
A step of classifying the learning data having a high degree of similarity from each other and deleting a part of the learning data;
having
In the step of calculating the similarity, when the two pieces of learning data are shifted in the time axis direction, the similarity is calculated based on a size per unit time of an area where the areas related to the two pieces of learning data overlap with respect to an area sandwiched between the waveform data indicated by each piece of learning data and the time axis.
Data extraction methods. A data extraction method for extracting a portion of learning data from a plurality of learning data extracted for each time window from time series data, comprising the steps of:
acquiring a plurality of said learning data;
calculating a similarity between the learning data based on features of the learning data excluding temporal differences in the time window;
A step of classifying the learning data having a high degree of similarity from each other and deleting a part of the learning data;
having
In the step of calculating the similarity, a similarity between the two shapes is calculated based on shapes of respective frequency distributions obtained by performing a frequency analysis on the two pieces of learning data.
Data extraction methods. A data extraction method for extracting a portion of learning data from a plurality of learning data extracted for each time window from time series data, comprising the steps of:
acquiring a plurality of said learning data;
A step of classifying the plurality of learning data into the same group so that similar data are classified based on a predetermined evaluation criterion;
calculating a similarity between the learning data based on features of the learning data excluding temporal differences in the time window;
A step of classifying the learning data having a high degree of similarity from each other and deleting a part of the learning data;
having
In the step of calculating the similarity, the similarity between the learning data belonging to the same group is calculated.
Data extraction methods. A learning model is constructed by learning the learning data extracted by the data extraction method according to any one of claims 7 to 9 , and then parameters of the learning model are adjusted using the learning data before extraction.
How to build a learning model. On the computer,
A data extraction process for extracting a portion of learning data from a plurality of learning data cut out for each time window from time series data, the data extraction process comprising the steps of:
acquiring a plurality of said learning data;
calculating a similarity between the learning data based on features of the learning data excluding temporal differences in the time window;
A step of classifying the learning data having a high degree of similarity from each other and deleting a part of the learning data;
having
The step of calculating the similarity includes a step of executing a data extraction process to calculate the similarity based on the size per unit time of the area where the regions related to the two learning data overlap, for a region sandwiched between the waveform data represented by each learning data and the time axis when the two learning data are shifted in the time axis direction. On the computer,
A data extraction process for extracting a portion of learning data from a plurality of learning data cut out for each time window from time series data, the data extraction process comprising the steps of:
acquiring a plurality of said learning data;
calculating a similarity between the learning data based on features of the learning data excluding temporal differences in the time window;
A step of classifying the learning data having a high degree of similarity from each other and deleting a part of the learning data;
having
In the step of calculating the similarity, a program is provided for executing a data extraction process that calculates the similarity between the two shapes based on the shapes of respective frequency distributions obtained by frequency analysis of the two pieces of learning data. On the computer,
A data extraction process for extracting a portion of learning data from a plurality of learning data cut out for each time window from time series data, the data extraction process comprising the steps of:
acquiring a plurality of said learning data;
A step of classifying the plurality of learning data into the same group so that similar data are classified based on a predetermined evaluation criterion;
calculating a similarity between the learning data based on features of the learning data excluding temporal differences in the time window;
A step of classifying the learning data having a high degree of similarity from each other and deleting a part of the learning data;
having
The step of calculating the similarity includes the step of: extracting data from the learning data that belong to the same group; On the computer,
A program that executes a process of learning the learning data extracted by the program according to any one of claims 11 to 13 to construct a learning model, and then adjusting parameters of the learning model using the learning data before extraction.

Images