JP7419787B2 - Time series data analysis device - Google Patents

Description

The present invention is a method for converting multiple time-series data obtained from observation targets such as production systems, production equipment, and products for PET bottle containers and can containers into image data that can be applied to machine learning and deep learning. The present invention relates to a time series data analysis device.

In recent years, there have been efforts to develop deep learning technology cultivated in image recognition for other applications.
When abstracted, an image can be understood as a three-dimensional array (tensor), so if non-image data can be converted into a tensor, it becomes possible to apply deep learning to images.

Regarding converting non-image data into images and performing machine learning, there is a technique proposed in Patent Document 1, for example.
What is disclosed in Patent Document 1 is to convert audio data (non-image data) into multidimensional data (image data) and generate acoustic features using a neural network, thereby producing smooth and natural speech. This means that it can be synthesized.

Patent No. 6552146

However, all of the techniques for converting non-image data into images and subjecting them to machine learning, including the method disclosed in Patent Document 1, only image data at the same time. It was not possible to deal with time-series data that is constantly generated and output from observation targets such as production systems.
Time-series data cannot be handled simply by imaging data at the same fixed time; appropriate time alignment is required between multiple pieces of time-series data, and delays in response for each time-series data are required. It was also important to carry out visualization with this in mind.
Conventional techniques have not been able to effectively deal with such problems.

The present invention was proposed to solve the above problems, and by quantifying the correlation between data divided in the time direction with respect to observation data consisting of a plurality of time series data, For time-series data consisting of multiple inputs with periodicity, it is possible to extract and visualize features that take into account their spatio-temporal relationships. The purpose of the present invention is to provide a time-series data analysis device that can convert data into data applicable to deep learning.

In order to achieve the above object, the time series data analysis device according to the present invention includes an observation data storage means for storing a plurality of observation data of two or more types consisting of time series data generated by an observation target; , divide one type of selected observation data into predetermined periods, divide all the observation data other than the selected one type of observation data into the predetermined periods, and create a set of partial time series data. A partial time series data storage means for storing the partial time series data, and each array element of a two-dimensional array having the type of the time series data as the first axis and the observation time of the partial time series data as the second axis. out of each element, either the reference element with the latest observation time of the selected one type of observation data, or the partial time series data identified from the first axis and the second axis of the array element. a feature array storage means that calculates a feature amount using a predetermined calculation method based on a combination of two elements with one element and stores it as a feature array; and a feature array output that outputs the feature array. The configuration includes means.

Further, the present invention can be configured as a program executed by the time-series data analysis device according to the present invention as described above.
Furthermore, the present invention can also be implemented as a method that can be implemented by the time-series data analysis device and program according to the present invention as described above.

According to the present invention, it is possible to quantify the correlation between data divided in the time direction for data consisting of a plurality of time series data.
This makes it possible to extract and visualize features of time-series data consisting of multiple inputs with periodicity, taking into account their spatiotemporal relationships. For example, multiple non-image data can be combined into one image data. It becomes possible to compress and generate multiple time-series data, which is non-image data, into data that can be applied to machine learning and deep learning.

FIG. 1 is a functional block diagram showing the configuration of a time-series data analysis device according to an embodiment of the present invention. FIG. 2 is an explanatory diagram schematically showing data to be processed by a time-series data analysis device, in which (a) shows observation data obtained from an observation target, and (b) shows a partial time series obtained by dividing the observation data. Shows a data set. FIG. 3 is an explanatory diagram schematically showing data to be processed by a time series data analysis device, in which (a) shows a partial time series data set, (b) shows correlation data based on the partial time series data, and (c ) indicates data in which correlations based on partial time series data are mapped into a two-dimensional array. FIG. 2 is an explanatory diagram schematically showing data to be processed by a time-series data analysis device, in which (a) shows data in which correlations based on partial time-series data are mapped into a two-dimensional array; This shows the data obtained by converting the values of the mapping array shown in (image) into luminance (image formation). FIG. 2 is an explanatory diagram schematically showing data to be processed by a time-series data analysis device, in which (a) shows a brightness-converted image corresponding to multiple inputs, and (b) shows a plurality of brightness change observed images superimposed. showing something. 2 is a flowchart showing processing operations in the time-series data analysis device, in which (a) shows a case where image output is not performed, and (b) shows a case where image output is performed. FIG. 2 is a block diagram showing the hardware configuration of a time-series data analysis device.

DESCRIPTION OF EMBODIMENTS Hereinafter, embodiments of a time-series data analysis device and a program according to the present invention will be described with reference to the drawings.
Here, the time series data analysis apparatus of the present invention described below is realized by processes, means, and functions executed by a computer according to instructions of a program (software). The program can send commands to each component of the computer to cause it to perform predetermined processing, functions, etc. related to the present invention as described below. That is, each process, means, and function in the present invention is realized by specific means in which a program and a computer cooperate.

Note that all or part of the program is provided by, for example, a magnetic disk, optical disk, semiconductor memory, or any other computer-readable recording medium, and the program read from the recording medium is installed on the computer and executed. Ru. Further, the program can also be directly loaded and executed on a computer through a communication line without using a recording medium.
Furthermore, this system can be configured with a single information processing device (for example, one personal computer, etc.) or can be configured with a plurality of information processing devices (for example, a group of multiple server computers, etc.).

[overall structure]
As shown in FIG. 1, a time series data analysis device 10 according to an embodiment of the present invention receives a plurality of time series data generated by an observation target 100 that can be acquired and output by a plurality of sensors 1 to n. It is a device/means that can perform predetermined data analysis processing and input the analysis processing results to any machine learning device 200.
The time-series data analysis device 10, the observation targets (sensors 1 to n), and the machine learning device 200 can be connected online so that they can transmit and receive data via wires or wirelessly, for example. Furthermore, the time-series data analysis device 10, the observation targets (sensors 1 to n), and the machine learning device 200 can each be set offline so that they can send and receive necessary data to and from each other via, for example, a recording medium.

[observation target]
The observation target 100 is a target for generating time series data that is subject to data analysis and visualization (imaging) processing by the time series data analysis device 10, and is, for example, a production system and production of PET bottle containers and can containers. Equipment, products, etc.
Specifically, the "observation target" in the present invention is a "object to be monitored" or a "phenomenon" determined by humans to solve some problem, and is one that generates two or more types of time-series data. If so, it can be an "observation target" to which the present invention can be applied.

For example, the present invention is expected to be applied to abnormality detection and abnormality cause estimation in the production technology field, and therefore, the main "observation targets" include (1) production systems (lines); These include (2) production equipment, (3) products, and (4) their models. Here, the term "model" refers to a simulation model, statistical model, machine learning model, etc. calculated on a computer.
In FIG. 1, for convenience of explanation, a plan view schematically showing a manufacturing line for PET bottle containers is shown as an observation target 100. However, it goes without saying that it is not particularly limited to this.

Furthermore, in the present invention, the expression "generates" in "observation target generates..." is equivalent to the expression "obtained from observation target...", and any observation obtained from observation target This means that the data (time series data) is subject to analysis processing of the present invention.
In addition, "time series data" in the present invention refers to values (data) generated and obtained from a given observation target as described above, which are (1) continuously or constantly It is a set of observed values obtained at each time, or (2) a set of observed values obtained at irregular time intervals and interpolated and estimated into continuous or fixed time values.

Therefore, "observation data/time-series data generated by an observation object" is any value that can be acquired and detected in a fixed/indeterminate period of time, such as temperature, humidity, pressure, vibration, etc. emitted by the object. For example, it may be a value measured by a temperature sensor, a pressure sensor, etc., or a value that can be obtained visually or manually by a person such as a supervisor or worker.
Then, two or more types of time series data (1 to n pieces) generated and output by the observation target 100 as described above are acquired as data/signals via means such as sensors 1 to n, and the data is , are input to the time series data analysis device 10.

In FIG. 1, a plurality of time-series data such as temperature and pressure are acquired over a fixed/indeterminate period of time by a plurality of sensors 1 to n that monitor a production line for PET bottle containers, which is an observation target 100. As an example, FIG. 2A shows three input data (inputs 1 to 3) acquired by three sensors 1 to 3. However, it goes without saying that it is not particularly limited to these.
For convenience of explanation, FIG. 1 shows only one (one set) of the observation target 100 and the plurality of sensors 1 to n that monitor it, and FIG. 2(a) shows three pieces of observation data. (Inputs 1 to 3) are shown, but the invention is not particularly limited to this, and it goes without saying that a plurality of observation targets can be simultaneously targeted by the time-series data analysis device 10.

[Time series data analysis device]
The time series data analysis device 10 is configured to run a predetermined program (software), such as a server system consisting of one or more server computers, personal computers, or one or more virtual servers built on a cloud computing service. The time-series data analysis device of the present invention can be configured by a mounted information processing device.
As shown in FIG. 1, the time series data analysis device 10 according to the present embodiment includes an observation data storage means 11, a partial time series data storage means 12, a feature array storage means 13, a feature array output means 14, It is configured to function and operate as each of image generation means 15, image output means 16, and multi-channel image output means 17.

The observation data storage means 11 stores two or more types of observation data made up of time series data generated by the observation target 100 described above in, for example, a predetermined storage means provided in the time series data analysis device 10.
As described above, the time series data stored and memorized by the observation data storage means 11 is data such as temperature and pressure over a fixed/indeterminate period of time, and for example, a plurality of data indicating detected values such as temperature and detection times. The information is stored for each observation target 100 in chronological order (see FIG. 2(a)).

The partial time series data storage means 12 generates and stores predetermined partial time series data based on the plurality of observation data stored and held by the observation data storage means 11.
Specifically, the partial time series data storage means 12 divides one type of observation data selected from among the plurality of observation data stored and held in the time series data analysis device 10 for each predetermined period. , all observation data other than the selected one type of observation data are divided into predetermined periods to generate partial time series data. Then, they are stored as a set of partial time-series data in, for example, a predetermined storage means provided in the time-series data analysis device 10.

An example of a set of partial time series data generated by the partial time series data storage means 12 is shown in FIG. 2(b).
As shown in the figure, the partial time series data storage means 12 first focuses on one piece of observation data out of a plurality of pieces of observation data, and segments the observation data in the time axis direction by a predetermined period (section width). Create a set of partial time series data.
Next, the partial time series data storage means 12 similarly creates a set of partial time series data for each observation data for other observation data using the above-mentioned predetermined interval width.

Here, as the predetermined period (interval width) for generating partial time series data, set an appropriate time, such as one second interval or several seconds interval, depending on the output characteristics and period of the target observation data. can do.
In particular, it is desirable to set this predetermined period to a length that is equal to or longer than the lowest cycle of all time series data in the target observation data.
If the predetermined period is set to a length shorter than the period of observation data, partial time series data will be generated within each divided period using a flat waveform that does not include the upper and lower limits of the values indicated by the observation data. Therefore, there is a possibility that similar data with little variation may be generated as the feature quantity obtained by the predetermined calculation described later. Therefore, in the partial time-series data storage means 12, it is preferable to set the predetermined period for dividing the time-series data to a length equal to or longer than the minimum period in all the target time-series data (see FIG. 2(b)). reference).

The feature array storage means 13 extracts and transforms the feature values for each of the plurality of partial time series data based on the set of partial time series data generated by the partial time series data storage means 12 as described above. Executes processing and generates and stores converted data.
FIG. 3 shows an example of extraction/conversion processing into a feature array based on partial time series data.
As shown in the figure, the feature array storage means 13 converts and generates partial time series data (see FIG. 3(a)) into predetermined two-dimensional array data (see FIG. 3(c)).

Specifically, as shown in FIG. 3(c), the feature array storage means 13 sets the first axis (for example, the horizontal axis) as the type of time series data, and the second axis (for example, the vertical axis) as the type of time series data. Generate two-dimensional array data with observation times of series data.
To this end, the feature array storage means 13 extracts and selects a reference element and a comparison element to be compared with it as each array element of the two-dimensional array.
First, among the elements of the partial time series data (see Figure 3(a)), the element with the latest (latest) observation time of one type of selected observation data is extracted and determined as the reference element. Ru.
The contrast elements to be compared with the reference element are the elements of the partial time series data identified from the first axis and the second axis of the array element, one by one, among all the elements other than the reference element. Extracted and selected.

Note that the extraction and determination of reference elements/comparison elements of partial time series data can be performed not only in the above case but also in other methods (conditions).
For example, the reference element can be determined not only by using the partial time series data with the latest (latest) observation time as described above, but also by using the partial time series data with the oldest (early) observation time as the reference element. You can also.
Furthermore, any one element among all the elements of the partial time series data set can be specified and extracted and determined as the reference element.
Then, based on the combination of the two elements, the reference element/contrast element, extracted as described above, feature quantities are calculated using a predetermined calculation method (see Figure 3(b)), and the features are calculated for all elements. The feature amounts are stored and generated as a two-dimensional array of feature amounts as shown in FIG. 3(c).

In the example shown in Figure 3, among the elements of the partial time series data, the element whose observation time of [Input 1] is the latest [Input 1: -0τ] is used as the reference element, and any other element is used as the reference element. For example, using the element [input 2: -0τ] as a contrast element (see Figure 3(a)), calculate the correlation between the two, and calculate the value of the contrast element with the reference element as the reference (1) (0.1223 ...) is calculated.
Specifically, as shown in Figure 3(b), one axis (for example, the horizontal axis) is set as "input 1" and the other axis (for example, the vertical axis) is set as "input 2", and the values of each partial time series data are By plotting, the correlation (feature amount) between both elements (input 1 and input 2) can be determined.

Then, as shown in FIG. 3(c), the values indicating the feature amount of each element of the partial time series data obtained in this way are set as the type of time series data on the first axis (horizontal axis), and By arranging the partial time series data as two-dimensional array data with two axes (vertical axis) as observation times of the partial time series data, a two-dimensional array mapping showing the correlation of each element is generated.
The feature array represented by such two-dimensional array mapping becomes information equivalent to image data, as shown in Figure 4, and can be applied to machine learning and deep learning based on image recognition. This means that time series data, which was image data, could be converted into data that can be applied to machine learning and deep learning.

Furthermore, in the above two-dimensional array mapping, not only [Input 1] but also [Input 2] and [Input 3], the element with the latest observation time [-0τ] is used as the reference element. , the correlation with each other element can be calculated, and a two-dimensional array mapping can be generated (see FIG. 5).
In this way, by converting and generating two-dimensional array mappings for multiple time series data of different types (inputs 1 to 3), as shown in Figure 5, by overlapping these multiple mapping data, It can be converted into a tensor in the same way as image data, and for example, even if each time series data (inputs 1 to 3) has different detection timing, time lag, response delay, etc., their spatiotemporal relationships can be understood. It becomes possible to take into account feature extraction and visualization (imaging) based on it.

Note that a known calculation method can be used as the predetermined calculation method for determining the correlation between the reference element and the contrast element. For example, at least one of covariance, correlation coefficient, HSIC, and MIC may be used. is preferred.
These known calculation methods can be used selectively depending on the data characteristics. For example, when the distribution of values obtained from time series data is on a straight line (there is an overall trend), they can be used selectively. Dispersion and correlation coefficients are suitable, and HSIC and MIC are suitable when the data values are not on a straight line (distributed on a curve).
Therefore, it is possible to selectively use calculation methods depending on the data, taking into account the type of observation target 100 and the characteristics of time-series data, or to calculate the most preferable calculation method by using all possible calculation methods. It is also possible to use the results, and it is desirable to use at least one of the above-mentioned known calculation methods.

As described above, the feature array storage means 13 calculates one feature corresponding to two elements of the set for a partial time series data set obtained from a plurality of time series data, and calculates the calculated value. By allocating it on a two-dimensional array, it is possible to convert and generate a two-dimensional array mapping equivalent to the image.
The obtained two-dimensional array mapping data (feature array) is output by the feature array output means 14 and inputted to any machine learning device 200, so that it can be used as a target for machine learning/deep learning. .
Further, the feature array outputted by the feature array output means 14 can be inputted into the image generation means 15 to be imaged and visualized into image data that can be visually recognized by humans.

The image generation means 15 converts the value of the corresponding element of each feature array generated and converted by the feature array storage means 13 into luminance by a predetermined conversion method, and converts each value of the image into luminance. By storing it as pixels, a predetermined image is generated.
Specifically, the image generation means 15 converts the value of each element indicated by the feature array data into a luminance value (for example, a value of "0-255") using a predetermined conversion method, and converts the value of each element indicated by the feature array data into a luminance value (for example, a value of "0-255"), and Convert the existing feature array to actual image data.

In the example shown in FIG. 4, for example, brightness "255" is assigned to the reference element (reference value) "1" as the upper bound (maximum) value, and the minimum value "-0.495" with respect to the reference value is assigned as the lower bound (minimum) value. It is possible to assign a brightness of "0" and convert the value of each element into a corresponding brightness value.
Then, by drawing an image corresponding to the converted luminance, the feature amount array shown in FIG. 4(a) can be converted into the image data shown in FIG. 4(b).

The converted and generated image data is outputted by the image output means 16, displayed as an image on a display means such as a liquid crystal display, and printed by a printing means such as a printer.
Further, the generated image data is outputted by the multi-channel image output means 17 for each of the plural observation data to form an output image set, and the output image set is combined in the channel direction to form a multi-channel can be output as a two-dimensional image.
As a result, as shown in Figure 5, feature arrays are converted into image data for multiple time series data of different types (inputs 1 to 3), and these are also generated and output as multi-channel two-dimensional images. will be done.

By converting feature array data into image data and outputting it in this way, it is possible to convert the feature array data into image data that is equivalent to the image that has been converted into image-recognizable machine learning data, and to create a feature array that can be visually recognized by humans. It can be output and displayed as actual image data.
As a result, time-series data generated by the observation target 100 can be recognized and grasped as an image, and by visually checking the image data, managers and workers of the observation target 100 can, for example, determine whether the production system is normal or not. Abnormalities can be determined.
In addition, by using multi-channel image data obtained by overlapping multiple images and converting them into tensors, it is possible to identify, for example, whether any element has a strong or weak relationship with a particular input sensor (the image becomes "white/black"). This makes it possible to make judgments and determinations based on the characteristics of input data.

Here, as a predetermined conversion method for converting the feature array into an image, the upper and lower bounds (or maximum value It is preferable to perform brightness scaling in a domain corresponding to the bit depth of the image constituting the feature amount array, with a numerical range of (initial value and initial value).
When converting a feature array into an image, it is possible to scale the brightness according to the upper and lower bounds (or maximum and initial values) of the actually obtained feature values, but in that case, the actual If the range of values of the obtained feature quantity is narrow, all values will be converted to almost the same brightness, and the generated image will be a "vague" image with almost the same color (gray) as a whole. Therefore, when viewed with human vision, it may not be possible to make specific judgments or judgments.

Therefore, in image generation/conversion by the image generation means 15, the upper and lower bounds (or maximum value and initial value) of possible values are set according to the numerical range and according to the bit depth of the image. It is desirable to scale the brightness accordingly.
Note that depending on the type of observation target 100 and the characteristics of time-series data, the range of features that can actually be obtained may be relatively wide, or the number of features that can be obtained may be extremely large. Of course, it is also possible to perform brightness scaling based on a range narrower or more limited than the upper and lower bounds (or the maximum value and initial value) of values that can be taken by the feature amount calculation method.

In addition, regarding image generation and output by the image generation means 15, image output means 16, and multi-channel image output means 17 as described above, machine learning/deep learning is performed by the machine learning device 200 based on feature array data. is not necessarily necessary. In other words, the data to be subjected to machine learning/deep learning does not need to be the image itself that humans can visually recognize and grasp, but can be two-dimensional array mapping data in which features equivalent to image data are arranged in two dimensions. , can be used as processing data for machine learning/deep learning.
Therefore, for example, when there is no need for a human such as a manager or worker of the observation target 100 to visually confirm or judge, the image generation means 15, image output means 16, and multi-channel image output means 17 can be used. Image generation/output is not necessary (see FIG. 6(a)), and in that case, the configurations of the image generation means 15, image output means 16, and multichannel image output means 17 in the time series data analysis device 10 are omitted. be able to.

[Machine learning device]
The machine learning device 200 is configured by any information processing device in which a known machine learning program/algorithm is implemented.
Through the analysis process of the time-series data analysis device 10 described above, a plurality of waveforms (time-series data) are outputted as feature-value extracted and tensorized (imaged) data.
By inputting the data to the machine learning device 200 and performing learning processing, it becomes possible to apply deep learning cultivated in image recognition to time series data acquired as non-image data.
Note that in FIG. 1, for convenience of explanation, only one machine learning device 200 is shown, but the invention is not limited to this. Of course, data obtained by the analysis device 10 can be provided and input.

[motion]
Next, specific processing and operation (time series data analysis method) in the time series data analysis device 10 as described above will be explained with reference to FIG. 6.
The time-series data analysis device 10 can analyze and image a plurality of time-series data acquired from the observation target 100 by automatically executing a series of processing operations shown in FIG.
Specifically, as shown in FIG. 6, the time series data analysis device 10 inputs time series data (S01) → generates a partial time series set (S02) → reference partial time series (S03) → generates a partial time series. Selection (S04) → Calculation of feature values (S05) → Store in feature array (S06) → Output of feature array (S07) → Brightness conversion of feature array (S08) → Storing brightness in pixels (S09) → Each process of image output (S10) is executed.
The details of each process will be explained below.

First, with reference to FIG. 6A, a case will be described in which the feature array is not converted and outputted into image data.
In the observation target 100, time series data is generated every moment, and two or more types of time series data are acquired and output by the plurality of sensors 1 to n, and a plurality of time series data as shown in FIG. 2(a) are generated. Data is input to the time series data analysis device 10 (step 01), and is stored and held in the observation data storage means 11 (see FIG. 2(a)).
A plurality of time series data input and stored in the time series data analysis device 10 are generated as a set of partial time series data divided and divided into predetermined periods (section widths), as shown in FIG. 2(b). - Stored (step 02).

The generated partial time series data set is converted and generated into feature array data (steps 03 to 06).
First, a reference partial time series serving as a reference element is selected from among all elements constituting the partial time series data set (step 03). For the partial time series to be the reference element, the element with the latest observation time is extracted and determined as the reference element among all the elements of the partial time series data set of the target observation data.
Next, a partial time series is selected as a comparison element to be compared with the reference element (step 04). For the partial time series to be contrasted elements, each element of the partial time series data set other than the reference element is sequentially extracted and selected one by one.

Then, features are calculated using a predetermined calculation method for the two selected reference/contrast partial time series data (step 05), and the calculated features are generated as a two-dimensional feature array. It is stored (step 06).
In the feature array, the values indicating the feature values of each element of the partial time series data are arranged such that the first axis (horizontal axis) is the type of time series data, and the second axis (vertical axis) is the observation time of the partial time series data. It consists of a two-dimensional array (see Figure 3 (c)), which is equivalent to image data that shows the correlation of each element (Figure 4), and observation data/time series data that was non-image data can be This means that the data has been converted to data that can be applied to training and deep learning.
Thereafter, the feature array data generated and stored by the time series data analysis device 10 is output to any machine learning device 200 (step 07), and can be subjected to machine learning/deep learning.

Next, along FIG. 6(b), a case will be described in which the feature array is converted and output into image data.
Note that when converting and outputting feature array data to image data, the image output shown in Figure 6(a) is performed for the processing operations from inputting time series data to generating and storing feature array data. It is the same as when there is no such information (steps 01 to 06).

The value of each element of the feature array generated and stored in step 06 is subjected to luminance conversion using a predetermined conversion method (step 08), and the converted luminance is stored in the pixel corresponding to the feature array (step 09). , image data is generated.
In the image data, the value of each element shown in the feature array data is replaced with a brightness value (for example, 0-255) using a predetermined conversion method, and the feature array, which was an array of numerical values, is converted into an image. It is something.
The generated image data is output as an image via, for example, a display or a printer (step 10), so that the administrator, operator, etc. of the observation target 100 can view and refer to the image.

Note that, as described above, the machine learning/deep learning performed by the machine learning device 200 based on the feature array data does not require image data that can be visually recognized and grasped by humans.
Therefore, if human visual confirmation of the image is not required, steps 08 to 10 related to image generation and output can be omitted.

As explained above, according to the time series data analysis device 10, for a plurality of time series data generated and acquired by the observation target 100, the correlation between data for partial time series data divided in the time direction is numerically calculated. It becomes possible to convert into
As a result, unlike conventional technology that only visualizes (images) data at the same time, this technology enables multiple time points that are generated and output from observation targets such as production systems (lines) from time to time. With regard to series data, it is possible to extract and visualize feature amounts while taking into account spatiotemporal relationships such as response delays in each time series data while performing appropriate time alignment between data.

Therefore, it is possible to generate and compress multiple time-series data (non-image data) obtained from observation targets such as production systems, production equipment, and products for PET bottle containers and can containers into one image data. By applying the feature data and image data generated and output by the time series data analysis device 10 to any machine learning device 200, deep learning technology cultivated in image recognition can be applied to time series data that is non-image data. It can also be applied to data.

An example of the hardware configuration of the time series data analysis device 10 according to the present embodiment as described above is as shown in FIG. It is composed of a processing device. These components are connected by a system bus, and data is exchanged via the system bus. A CPU (Central Processing Unit) 501, also called a central processing unit, is a part that performs central processing of a computer, and controls each device and calculates/processes data. A RAM (Random Access Memory) 502 is a type of memory device in which data can be erased and rewritten, and a ROM (Read Only Memory) 503 is a type of memory device using a semiconductor or the like, and data cannot be written. It can only be used once at the time of manufacture, and only the recorded data can be read out at the time of use. A HDD (Hard Disk Drive) 504 is an auxiliary storage device that records and reads information using the properties of a magnetic material. The input device 505 is used by the user to issue operating instructions to the computer or to input characters, etc., and specifically includes a keyboard, a mouse, and the like. The display device 506 may be configured with, for example, a liquid crystal display, and may have a touch panel function. In addition, it also has a communication function (not shown), and this communication function enables communication with other terminals.

Although the present invention has been described above with reference to preferred embodiments, it goes without saying that the present invention is not limited to the above-described embodiments, and that various modifications can be made within the scope of the present invention.
For example, in the above-mentioned embodiment, a PET bottle container manufacturing line was explained as an example of an observation target to which the present invention is applied, but an observation target to which the present invention can be applied is one in which two or more time series data are generated. - There are no particular limitations as long as the target can be output.

INDUSTRIAL APPLICABILITY The present invention can be used as an apparatus and method suitable for data analysis on an observation target that generates and outputs a plurality of time-series data, such as a manufacturing line for can containers and PET bottle containers.

10 Time series data analysis device 11 Observation data storage means 12 Partial time series data storage means 13 Feature value array storage means 14 Feature value array output means 15 Image generation means 16 Image output means 17 Multi-channel image output means 100 Observation target 200 Machine learning Device

Claims (6)

observation data storage means for storing a plurality of observation data of two or more types consisting of time series data generated by an observation target;
Among the observation data, one type of observation data selected is divided into predetermined periods, and all of the observation data other than the selected one type of observation data is divided into parts according to the predetermined period. Partial time series data storage means for storing as a set of series data;
As each array element of a two-dimensional array in which the first axis is the type of the time series data and the second axis is the observation time of the partial time series data, the selected Two elements: a reference element with the latest observation time of one type of observation data, and any one element of the partial time series data identified from the first axis and the second axis of the array element. a feature amount array storage means for calculating a feature amount using a predetermined calculation method based on the combination and storing it as a feature amount array;
Feature array output means for outputting the feature array;
A time series data analysis device comprising: image generating means for generating a predetermined image by converting the values of corresponding elements of the feature array into luminances using a predetermined conversion method and storing the converted values as each pixel of the image;
image output means for outputting the image;
The time-series data analysis device according to claim 1, comprising: multi-channel image output means for outputting the images for each of the plurality of observation data to form an output image set, and combining the output image sets in the channel direction to output them as a multi-channel two-dimensional image;
The time-series data analysis device according to claim 2, comprising: a time-series data analysis device; As the predetermined period,
The time series data analysis device according to any one of claims 1 to 3, characterized in that the length is set to be longer than the lowest period in all the time series data in the observation data. As the predetermined calculation method,
The time series data analysis device according to any one of claims 1 to 4, characterized in that at least one of covariance, correlation coefficient, HSIC, and MIC is used. As the predetermined conversion method,
The brightness is scaled in a domain corresponding to the bit depth of the image constituting the feature array, in which the upper and lower bounds of values that can be obtained by the predetermined calculation method are numerical ranges. The time series data analysis device according to claim 2.

Images