Analysis device, analysis method, and analysis program

Abstract

To improve prediction accuracy when there is a defect in time series data.SOLUTION: The analysis device includes an input unit that inputs time-series data of observation values of an observation target, a first aggregation unit that aggregates the time-series data of the observation values for each unit time to calculate a statistical value and outputs time-series first one-dimensional array data of the statistical value, and a missing feature amount calculation unit that calculates a missing feature amount by aggregating missing of the observation values for each unit time in the time-series data of the observation values. A second aggregating unit that outputs time-series second one-dimensional array data of the defect feature amount, a dividing unit that divides the first one-dimensional array data based on a specific frequency to generate a plurality of pieces of first two-dimensional array data and divides the second one-dimensional array data based on the specific frequency to generate a plurality of pieces of second two-dimensional array data, a combining unit that combines the plurality of pieces of first two-dimensional array data and the plurality of pieces of second two-dimensional array data, and an identification unit that outputs an identification result by performing a convolution operation on a result of combining to extract a feature amount of the result of combining.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The present invention relates to an analysis device, an analysis method, and an analysis program for analyzing data. [Background technology]

[0002] Patent Document 1 discloses a time-series data feature extraction device. This time-series data feature extraction device includes: a data processing unit that processes a received group of irregularly spaced time-series data into a group of equally spaced time-series data including missing data and a group of missing information indicating the presence or absence of missing data, based on the received input time-series data length and the received minimum observed interval; a model training unit that learns weight vectors for each layer of the model using, as an error, the difference between non-missing elements in the matrix of the group of equally spaced time-series data including missing data and elements of the output result of the output layer of the model, and stores the weight vectors in a storage unit as model parameters; and a feature extraction unit that receives time-series data from which features are to be extracted, inputs the received time-series data from which features are to be extracted into the model, calculates values ​​of intermediate layers of the model using the model parameters stored in the storage unit, and outputs the calculated intermediate layer values ​​as features indicating changes in data over time.

[0003] Non-Patent Document 1 discloses a machine learning network that uses a fast Fourier transform to find characteristic periods in one-dimensional time series data, divides the one-dimensional time series data at the found periods, rearranges the data into a two-dimensional array, and extracts features from the two-dimensional array using a convolutional neural network (CNN). [Prior art documents] [Patent documents]

[0004] [Patent Document 1] International Publication No. WO2018 / 047655 [Non-patent literature]

[0005] [Non-Patent Document 1] Haixu Wu, Tengge Hu, Yong Liu, Hang Zhou, JianminWang, Mingsheng Long, “Timesnet: Temporal 2d-variation modeling for general time series analysis”, 12 Apr 2023 Summary of the Invention [Problem to be solved by the invention]

[0006] However, the technology of Patent Document 1 simply ignores the missing parts and does not take into consideration accurately filling in the missing parts or finding meaning in the missing parts themselves. Furthermore, the technology of Non-Patent Document 1 cannot determine an appropriate analysis period when the time series data contains partial missing parts.

[0007] As described above, the conventional techniques described above are unable to achieve sufficient prediction accuracy when the loss pattern itself can be a meaningful feature for prediction.

[0008] The present invention aims to improve prediction accuracy when there are gaps in time-series data. [Means for solving the problem]

[0009] An analysis device according to one aspect of the invention disclosed in the present application includes an input unit that inputs time-series data of observed values ​​of an observation target; a first counting unit that counts the time-series data of the observed values ​​input by the input unit for each unit time to calculate a statistical value and outputs first one-dimensional array data of the time-series statistical value; a second counting unit that counts missing observed values ​​in the time-series data of the observed values ​​for each unit time to calculate a missing feature amount and outputs second one-dimensional array data of the missing feature amount; and a second counting unit that divides the first one-dimensional array data output by the first counting unit based on a specific frequency to separate elements where the statistical value exists from elements where the statistical value exists. a division unit that generates a plurality of first two-dimensional array data consisting of elements where the missing feature is present and elements where the missing feature is not present, and divides the second one-dimensional array data output by the second aggregation unit based on the specific frequency to generate a plurality of second two-dimensional array data consisting of elements where the missing feature is present and elements where the missing feature is not present; a combination unit that combines the plurality of first two-dimensional array data and the plurality of second two-dimensional array data generated by the division unit; and an identification unit that performs a convolution operation on the combination result by the combination unit to extract the feature of the combination result, and outputs an identification result that identifies the state of the observed object. [Effects of the Invention]

[0010] According to the exemplary embodiment of the present invention, it is possible to improve the prediction accuracy when there are missing time-series data. Problems, configurations, and effects other than those described above will become clear from the following description of the examples. [Brief explanation of the drawings]

[0011] [Figure 1] FIG. 1 is a block diagram illustrating an example of the hardware configuration of an analysis device. [Figure 2] FIG. 2 is a block diagram illustrating an example of the functional configuration of the analysis device. [Figure 3] FIG. 3 is a block diagram illustrating an example of the functional configuration of the learning unit. [Figure 4] FIG. 4 is an explanatory diagram showing an example of counting by the first counting unit and the second counting unit. [Figure 5] FIG. 5 is an explanatory diagram showing an example of frequency analysis by the frequency analysis unit. [Figure 6] FIG. 6 is an explanatory diagram showing an example of generation of a first two-dimensional array data group by the dividing unit. [Figure 7] FIG. 7 is an explanatory diagram showing an example of generation of a second two-dimensional array data group by the dividing unit. [Figure 8] FIG. 8 is an explanatory diagram showing Example 1 of the combined two-dimensional array data. [Figure 9] FIG. 9 is an explanatory diagram showing Example 2 of the combined two-dimensional array data. [Figure 10] FIG. 10 is a flowchart illustrating an example of a learning process procedure performed by the learning unit. [Figure 11] FIG. 11 is a block diagram illustrating an example of the functional configuration of the prediction unit. [Figure 12] FIG. 12 is a flowchart illustrating an example of a learning process procedure performed by the learning unit. [Figure 13] FIG. 13 is an explanatory diagram showing an example of a display screen. [Figure 14] FIG. 14 is an explanatory diagram showing Example 3 of the combined two-dimensional array data. DETAILED DESCRIPTION OF THE INVENTION

[0012] <Figure 1 Hardware configuration of the analysis device> FIG. 1 is a block diagram showing an example of the hardware configuration of an analysis device. The analysis device 100 includes a processor 101, a storage device 102, an input device 103, an output device 104, and a communication interface (communication IF) 105. The processor 101, the storage device 102, the input device 103, the output device 104, and the communication IF 105 are connected via a bus 106. The processor 101 controls the analysis device 100. The storage device 102 serves as a working area for the processor 101. The storage device 102 is a non-transitory or temporary recording medium that stores various programs and data. Examples of the storage device 102 include a read-only memory (ROM), a random access memory (RAM), a hard disk drive (HDD), and a flash memory. The input device 103 inputs data. Examples of the input device 103 include a keyboard, a mouse, a touch panel, a numeric keypad, a scanner, a microphone, and a sensor. The output device 104 outputs data. The output device 104 includes, for example, a display, a printer, and a speaker. The communication IF 105 connects to a network and transmits and receives data.

[0013] <Figure 2 Functional configuration of analysis device 100> 2 is a block diagram showing an example of the functional configuration of analysis device 100. Analysis device 100 has a learning unit 201 and a prediction unit 202. The learning process involves extracting features of one-dimensional time-series data, learning using the extracted features with teacher labels, and generating a learning model 203. Prediction unit 202 extracts features of the time-series data to be predicted, inputs the extracted features to learning model 203, and identifies the time-series data to be predicted.

[0014] Specifically, learning unit 201 and prediction unit 202 are realized by, for example, causing processor 101 to execute a program stored in storage device 102 shown in Fig. 1. Furthermore, at least one of learning unit 201 and prediction unit 202 may be implemented in another computer that can communicate with analysis device 100.

[0015] The learning model 203 is stored in the storage device 102 shown in FIG. 1 or another computer that can communicate with the analysis device 100.

[0016] <Figure 3: Example of functional configuration of learning unit 201> 3 is a block diagram showing an example of the functional configuration of the learning unit 201. The learning unit 201 includes an input unit 310, a first counting unit 311, a second counting unit 312, a frequency analysis unit 313, a division unit 314, a combination unit 315, a discrimination unit 316, and an adjustment unit 317.

[0017] The input unit 310 inputs one-dimensional time series data 300. The time series data 300 is one-dimensional data that indicates changes over time in the observed values ​​of an observation target. The observation target is an observation target device such as a facility or machine. The observed values ​​are numerical values ​​observed from the observation target, such as voltage values, current values, acceleration, rotation speed, vibration speed, and temperature (temperature may also be detected from the vicinity of the observation target device) detected by a sensor provided on the observation target device. The time series data 300 is a learning dataset that is paired with a teacher label 370.

[0018] The first counting unit 311 counts observed values ​​from the one-dimensional time series data 300 for each unit time. The unit time is a time span longer than the observation interval between successive observed values ​​in the time series data 300. For example, if the time series data 300 has observed values ​​every second, the unit time is one minute. The first counting unit 311 counts the observed values ​​for each unit time and calculates statistical values. The statistical values ​​are statistical values ​​related to the observed values ​​within the unit time, such as the sum, average, maximum, minimum, median, or mode of the observed values ​​within the unit time. The time series data of the statistical values ​​is referred to as first one-dimensional array data 320. In the first one-dimensional array data 320, successive statistical values ​​may be connected by line segments.

[0019] The second counting unit 312 counts missing observation values ​​for each unit time from the one-dimensional time series data 300. A missing value is an observation time point where an observation should have been made but was not made for some reason, or an observed value that was observed but was outside a predetermined range, such as an outlier, and the time point at which it was made. The second counting unit 312 counts missing values ​​for each unit time and calculates a missing feature. The missing feature is, for example, the number of missing values ​​per unit time (hereinafter referred to as the number of missing values). The time series data of missing features is referred to as second one-dimensional array data 330. In the second one-dimensional array data 330, consecutive missing features may be connected by line segments.

[0020] [Figure 4: Counting by the first counting unit 311 and the second counting unit 312] 4A and 4B are explanatory diagrams showing examples of aggregation by the first aggregation unit 311 and the second aggregation unit 312. (A) shows an example of interpolation when there is a gap in the one-dimensional time series data 300. (B) shows an example of aggregation by the first aggregation unit 311. In FIG. 4, as an example, the observation interval of the time series data 300 is set to 15 seconds, and the unit time is set to 1 minute.

[0021] (A) When there is a missing value in the time series data 300, it is possible to interpolate the missing value using two observation values ​​adjacent to the time when the missing value occurred. However, if the interpolated value is linearly interpolated, it is the average value of the two observation values, and is not the true missing observation value. Therefore, if interpolation is performed, the learning accuracy and prediction accuracy will decrease if you want to find meaning in the missing value.

[0022] (B) The first counting unit 311 counts the observed values ​​for each unit time and calculates statistical values ​​from the counted observed values. For example, in section B1, there are three observed values ​​and one missing value, so the first counting unit 311 calculates the statistical value of the three observed values. In section B2, there are two observed values ​​and two missing values, so the first counting unit 311 calculates the statistical value of the two observed values. The series of statistical values ​​calculated in this way is the first one-dimensional array data.

[0023] (B) The second counting unit 312 counts the number of defects for each unit time and calculates the number of defects by adding up the total number of the counted defects. For example, in section B1, there are three observed values ​​and one defect, so the first counting unit 311 calculates the number of defects as 1. In section B2, there are two observed values ​​and two defects, so the first counting unit 311 calculates the number of defects as 2. The series of numbers of defects calculated in this way constitutes the second one-dimensional array data 330.

[0024] Returning to FIG. 3, the frequency analysis unit 313 performs frequency analysis on the first one-dimensional array data 320 and outputs a specific frequency to the division unit 314.

[0025] [Figure 5 Frequency analysis by frequency analysis unit 313] 5 is an explanatory diagram showing an example of frequency analysis by the frequency analysis unit 313. The frequency analysis unit 313 performs a fast Fourier transform on the first one-dimensional array data 320 to generate a frequency spectrum 500. The frequency spectrum 500 indicates the intensity of each frequency. The intensity of a frequency is a value that depends on changes in the observed value, and the more periodic the change, the higher the intensity of that frequency. In the frequency spectrum 500, the frequencies are listed in descending order as f1, f2, f3, ..., fn (n is an integer equal to or greater than 1). When there is no need to distinguish between f1, f2, f3, ..., fn, they will be referred to as a specific frequency fn.

[0026] The frequency analysis unit 313 selects a specific frequency from the frequency spectrum 500 and outputs it to the division unit 314. The specific frequency is, for example, the maximum frequency f1 in the frequency spectrum 500. The specific frequency is not limited to the maximum frequency f1, and may be, for example, a frequency whose intensity is equal to or greater than a threshold, or any of the frequencies fn in the top n-th order (n is an integer equal to or greater than 1) in intensity.

[0027] Returning to FIG. 3 , the dividing unit 314 divides the first one-dimensional array data 320 based on a specific frequency fn to generate a plurality of first two-dimensional array data. Specifically, for example, the dividing unit 314 divides the first one-dimensional array data 320 at a period 1 / fn, which is the reciprocal of the specific frequency fn. Each of the divided plurality of first two-dimensional array data is matrix data in a coordinate plane formed by a horizontal axis indicating time and a vertical axis indicating statistical values, in which, for example, elements having a statistical value (which may include a line segment if consecutive statistical values ​​are connected by the line segment) have a value of 1 and elements having no statistical value have a value of 0. The values ​​of both elements are merely examples, and may be different values. The divided plurality of first two-dimensional array data is referred to as a first two-dimensional array data group 351.

[0028] Similarly, the dividing unit 314 divides the second one-dimensional array data 330 based on a specific frequency fn to generate multiple pieces of second two-dimensional array data. Specifically, for example, the dividing unit 314 divides the second one-dimensional array data 330 at a period 1 / fn, which is the inverse of the specific frequency fn. Each of the multiple pieces of divided second two-dimensional array data is matrix data in which, on a coordinate plane formed by a horizontal axis indicating time and a vertical axis indicating missing features (number of missing features), the value of an element having a statistical value (which may include a line segment if consecutive missing features are connected by the line segment) is 1 and the value of an element having no statistical value is 0. The values ​​of both elements are merely examples, and may be different values. The multiple pieces of divided second two-dimensional array data are referred to as a second two-dimensional array data group 352.

[0029] The number of divisions of the first one-dimensional sequence data 320 is the same as the number of divisions of the second one-dimensional sequence data 330. Note that the division unit 314 may divide the first one-dimensional sequence data 320 and the second one-dimensional sequence data 330 using a specific frequency fn set in advance, rather than the specific frequency fn from the frequency analysis unit 313.

[0030] FIG. 6: Example of generation of first two-dimensional array data set 351 by division unit 314 6 is an explanatory diagram showing an example of generation of first two-dimensional array data group 351 by dividing unit 314. Dividing unit 314 divides first two-dimensional array data group 351 into first two-dimensional array data 601 to 606 at a period 1 / fn, which is the reciprocal of a specific frequency fn.

[0031] When the first two-dimensional array data 601 to 606 are not distinguished, they are referred to as first two-dimensional array data 600. In the first two-dimensional array data 600, consecutive statistical values ​​may be connected by line segments. The first two-dimensional array data 600 is, for example, matrix data in which elements in the coordinate space where a statistical value (which may include a line segment if consecutive statistical values ​​are connected by a line segment) exists are represented as 1, and elements where no statistical value exists are represented as 0. The first two-dimensional array data 600 may be image data.

[0032] FIG. 7 Example of generation of second two-dimensional array data set 352 by division unit 314 7 is an explanatory diagram showing an example of generation of second two-dimensional array data group 352 by dividing unit 314. Dividing unit 314 divides second two-dimensional array data group 352 into second two-dimensional array data 701 to 706 at a period 1 / fn, which is the reciprocal of a specific frequency fn.

[0033] When the second two-dimensional array data 701 to 706 are not distinguished, they are referred to as second two-dimensional array data 700. In the second two-dimensional array data 700, consecutive statistical values ​​may be connected by a line segment. Like the first two-dimensional array data 600, the second two-dimensional array data 700 is matrix data in which, for example, elements that have a loss count in its coordinate space (which may include a line segment if consecutive loss counts are connected by a line segment) are represented as 1 and elements that do not have a loss count are represented as 0. The second two-dimensional array data 700 may be image data.

[0034] 3, the combining unit 315 combines the first two-dimensional array data group 351 and the second two-dimensional array data group 352, and generates, as a combined result, combined two-dimensional array data 360. Specifically, for example, the combining unit 315 combines the first two-dimensional array data 601-606 and the second two-dimensional array data 701-706 so that they are dispersed.

[0035] [Figure 8, Figure 9: Combined two-dimensional array data 360] Fig. 8 is an explanatory diagram showing Example 1 of the combined two-dimensional array data 360. In Fig. 8, the combined two-dimensional array data 360 is an example in which the first two-dimensional array data 600 and the second two-dimensional array data 700 are combined by alternately arranging each piece of data along the vertical axis (statistical value, number of missing pieces).

[0036] Fig. 9 is an explanatory diagram showing Example 2 of the combined two-dimensional array data 360. In Fig. 9, the combined two-dimensional array data 360 is an example in which the first two-dimensional array data 600 and the second two-dimensional array data 700 are combined by alternately arranging two by two in the direction of the vertical axis (statistical value, number of defects).

[0037] The combining performed by the combining unit 315 depends on, for example, the size of the kernel 800. The kernel 800 is a filter used when performing a convolution operation on the combined two-dimensional array data 360 in CNN. The combining unit 315 combines the first two-dimensional array data 601-606 and the second two-dimensional array data 701-706 so that the combined two-dimensional array data 360 includes one or more first two-dimensional array data 600 and one or more second two-dimensional array data 700, regardless of where the kernel is placed in the combined two-dimensional array data 360.

[0038] The size of the kernel 800 is assumed to be m×m (m is an integer equal to or greater than 1). The length of the vertical axis of each of the first two-dimensional array data 600 and the second two-dimensional array data 700 is assumed to be V. If m>V, as shown in FIG. 8, the first two-dimensional array data 600 and the second two-dimensional array data 700 are alternately arranged one by one along the vertical axis. If m>2V, the first two-dimensional array data 600 and the second two-dimensional array data 700 are alternately arranged two by two along the vertical axis. Therefore, if m>kV (k is an integer equal to or greater than 1), the first two-dimensional array data 600 and the second two-dimensional array data 700 are alternately arranged k by k along the vertical axis.

[0039] In Figures 8 and 9, the vertical axis is shorter than the horizontal axis (time), so the combining unit 315 arranges the first two-dimensional array data 600 and the second two-dimensional array data 700 alternately in the vertical axis direction, but if the horizontal axis is shorter than the vertical axis, the combining unit 315 may arrange the first two-dimensional array data 600 and the second two-dimensional array data 700 alternately in the horizontal axis direction.

[0040] Returning to FIG. 3, the classification unit 316 includes a CNN. The classification unit 316 inputs the combined two-dimensional array data 360 into the CNN, extracts feature quantities from the combined two-dimensional array data 360, and outputs a classification result based on the extracted feature quantities. The feature quantities are, for example, an array composed of numerical values ​​obtained by convolving a kernel 800 with a statistical value such as a voltage value detected by a sensor. The classification result is an output value from a fully connected layer included in the CNN, and ranges from 0 to 1.

[0041] The adjustment unit 317 adjusts the weights of the kernels 800 of the CNN based on the classification result from the classification unit 316 and the teacher labels 370 indicating the state of the observation target in the time-series data 300 of the observation values, and sets the learning model 203 that identifies the state of the observation target. Specifically, for example, the adjustment unit 317 calculates the value of the loss function based on the difference between the classification result and the teacher labels 370, and performs backpropagation so as to reduce the value of the loss function, thereby controlling the weights of the kernels 800 used in the convolution operation of the CNN.

[0042] <Fig. 10 Learning process procedure by learning unit 201> FIG. 10 is a flowchart showing an example of the learning process procedure performed by the learning unit 201.

[0043] (Step S1001) The learning unit 201 receives time-series data 300 of observed values ​​via an input unit 310 .

[0044] (Step S1002) The learning unit 201 causes the first counting unit 311 to count the observed values ​​into statistical values ​​per unit time, as shown in FIG.

[0045] (Step S1003) The learning unit 201 uses the second counting unit 312 to count the number of defects per unit time, as shown in FIG.

[0046] (Step S1004) The learning unit 201 performs frequency analysis on the first one-dimensional array data 320 using the frequency analysis unit 313, as shown in FIG.

[0047] (Step S1005) The learning unit 201 selects a specific frequency fn from the frequency spectrum 500 using the dividing unit 314 as shown in FIG.

[0048] (Step S1006) The learning unit 201 divides the first one-dimensional array data 320 at a period 1 / fn of a specific frequency fn, as shown in FIG. 6, by the dividing unit 314, to generate a first two-dimensional array data group 351.

[0049] (Step S1007) The learning unit 201 divides the second one-dimensional array data 330 at a period 1 / fn of a specific frequency fn, as shown in FIG. 7, by the dividing unit 314, to generate a second two-dimensional array data group 352.

[0050] (Step S1008) The learning unit 201 uses the combining unit 315 to combine the first two-dimensional array data group 351 and the second two-dimensional array data group 352 as shown in FIGS. 8 and 9 to generate combined two-dimensional array data 360.

[0051] (Step S1009) The learning unit 201 inputs the combined two-dimensional array data 360 to the CNN by the adjustment unit 317, executes a convolution operation, extracts the feature amount of the combined two-dimensional array data 360, and outputs the classification result.

[0052] (Step S1010) The learning unit 201 uses the adjustment unit 317 to learn the relationship between the classification result output in step S1009 and the teacher label 370. That is, the learning unit 201 uses the adjustment unit 317 to calculate the value of the loss function based on the difference between the classification result and the teacher label 370, and controls the weight of the kernel 800 by performing error back propagation so that the value of the loss function becomes smaller.

[0053] The learning unit 201 repeatedly executes steps S1001 to S1010 for each combination of the time-series data 300 and the teacher label 370. As a result, a learning model 203 is generated in which the weight of the finally obtained kernel 800 is set in the CNN.

[0054] In this way, by actively incorporating and learning from defects rather than ignoring them, it is possible to generate a learning model 203 that is capable of finding meaning in the defects themselves.

[0055] <Fig. 11: Example of functional configuration of the prediction unit 202> 11 is a block diagram showing an example of the functional configuration of the prediction unit 202. The prediction unit 202 has an input unit 310, a first counting unit 311, a second counting unit 312, a frequency analysis unit 313, a division unit 314, a combination unit 315, and a discrimination unit 1117. The input unit 310, the first counting unit 311, the second counting unit 312, the frequency analysis unit 313, the division unit 314, and the combination unit 315 have the same configuration as the learning unit 201.

[0056] In prediction unit 202, the one-dimensional data input to input unit 310 is time-series data 1110 to be predicted. Therefore, first counting unit 311 outputs first one-dimensional array data 1120, second counting unit 312 outputs second one-dimensional array data 1130, frequency analysis unit 313 outputs the specific frequency fn selected by learning unit 201, division unit 314 outputs first two-dimensional array data group 1151 and second two-dimensional array data group 1152, and combination unit 315 outputs two-dimensional array data 1160.

[0057] The discrimination unit 1117 inputs the two-dimensional array data 1160 to the learning model 203, and performs a convolution operation with the kernel 800 whose weights are controlled by the learning unit 201, thereby outputting a discrimination result. <Fig. 12 Prediction processing procedure by the prediction unit 202> FIG. 12 is a flowchart showing an example of a learning process procedure performed by the learning unit 201.

[0058] (Step S1201) The prediction unit 202 receives, via the input unit 310, time series data 1110 of the observation value to be predicted.

[0059] (Step S1202) The prediction unit 202 causes the first counting unit 311 to count the observed values ​​into statistical values ​​per unit time, as shown in FIG.

[0060] (Step S1203) The prediction unit 202 uses the second counting unit 312 to count the number of defects per unit time, as shown in FIG.

[0061] (Step S1204) The prediction unit 202 divides the first one-dimensional array data 1120 by the division unit 314 at a period 1 / fn of the specific frequency fn selected by the learning unit 201, as shown in Figure 6, to generate a first two-dimensional array data group 1151.

[0062] (Step S1205) The prediction unit 202 divides the second one-dimensional array data 1130 by the division unit 314 at a period 1 / fn of the specific frequency fn selected by the learning unit 201, as shown in Figure 7, to generate a second two-dimensional array data group 1152.

[0063] (Step S1206) The prediction unit 202 uses the combining unit 315 to combine the first two-dimensional array data set 1151 and the second two-dimensional array data set 1152, as shown in FIGS. 8 and 9, to generate two-dimensional array data 1160.

[0064] (Step S1207) The prediction unit 202 inputs the two-dimensional array data 1160 to the CNN using the discrimination unit 1117, executes a convolution operation using the kernel 800 whose weights are controlled by the learning unit 201, extracts features of the two-dimensional array data 1160, and outputs a discrimination result. By simultaneously convolving missing features, it is possible to discriminate that there is a high possibility of a fault when there are many missing features in the time series change of the missing features, even if the time series change of the statistical value is the same. In this way, the learning model 203, which is trained by actively incorporating missing features rather than ignoring them, can obtain a discrimination result in which the missing features themselves are meaningful.

[0065] Therefore, for example, when an industrial machine is close to failing, the vibration frequency deviates from the measurement range and is increasingly recorded as missing data. In this case, the timing of the failure can be predicted from the time series data of the vibration frequency detected by a vibration sensor attached to the industrial machine.

[0066] <Figure 13 Example of display screen> 13 is an explanatory diagram showing an example of a display screen 1300. On a display screen 1300, first one-dimensional array data 320, second one-dimensional array data 330, and frequency spectrum 500 are displayed.

[0067] <Modification> Fig. 14 is an explanatory diagram showing example 3 of two-dimensional array data. In Fig. 8 and Fig. 9, the first two-dimensional array data 600 and the second two-dimensional array data 700 are arranged so that the statistical value and the number of missing data are included in one kernel 800, but the combined two-dimensional array data 360 in Fig. 13 is data arranged so that one kernel 800 includes only one of the first two-dimensional array data 600 and the second two-dimensional array data 700.

[0068] The present invention is not limited to the above-described embodiments, and includes various modifications and equivalent configurations within the spirit and scope of the appended claims. For example, the above-described embodiments have been described in detail to clearly explain the present invention, and the present invention is not necessarily limited to configurations including all of the described configurations. Furthermore, part of the configuration of one embodiment may be replaced with the configuration of another embodiment. Furthermore, the configuration of another embodiment may be added to the configuration of one embodiment. Furthermore, part of the configuration of each embodiment may be added to, deleted from, or replaced with other configurations.

[0069] Furthermore, the aforementioned configurations, functions, processing units, processing means, etc. may be realized in part or in whole in hardware, for example by designing them as integrated circuits, or may be realized in software by having a processor interpret and execute a program that realizes each function.

[0070] Information such as programs, tables, files, etc. that realize each function can be stored in storage devices such as memory, hard disks, SSDs (Solid State Drives), or recording media such as IC (Integrated Circuit) cards, SD cards, and DVDs (Digital Versatile Discs).

[0071] In addition, the control lines and information lines shown are those that are considered necessary for explanation, and do not necessarily represent all the control lines and information lines that are necessary for implementation. In reality, it can be assumed that almost all components are interconnected. [Explanation of symbols]

[0072] 100 Analyzer 201 Learning Department 202 Prediction Department 203 Learning Model 300 Time Series Data 310 Input section 311 First Tabulation Section 312 Second Tally Section 313 Frequency Analysis Unit 314 Split section 315 Joint 316 Identification Unit 317 Generation part 320 One-dimensional array data 330 One-dimensional array data 351 Two-dimensional array data set 352 Two-dimensional array data set 360 Combined Two-Dimensional Array Data 370 Teacher Labels 500 Frequency Spectrum 600 First two-dimensional array data 700 Second two-dimensional array data 800 kernel 1110 Time Series Data 1117 Identification Unit 1120 First one-dimensional array data 1130 Second one-dimensional array data 1151 First two-dimensional array data set 1152 Second two-dimensional array data set 1160 Combined Two-Dimensional Array Data fn specific frequency

Claims

1. an input unit for inputting time series data of observed values ​​of an observation target; a first counting unit that counts the time-series data of the observation values ​​input by the input unit for each unit time to calculate statistical values ​​and outputs first one-dimensional array data of the time-series statistical values; a second counting unit that calculates a missing feature by counting missing observation values ​​in the time-series data of the observation values ​​for each unit time, and outputs second one-dimensional array data of the missing feature values ​​in time series; a division unit that divides first one-dimensional array data output by the first counting unit based on a specific frequency to generate a plurality of first two-dimensional array data configured with elements for which the statistical value exists and elements for which the statistical value does not exist, and divides second one-dimensional array data output by the second counting unit based on the specific frequency to generate a plurality of second two-dimensional array data configured with elements for which the missing feature exists and elements for which the missing feature does not exist; a combining unit that combines the plurality of first two-dimensional array data and the plurality of second two-dimensional array data generated by the dividing unit; a discrimination unit that performs a convolution operation on the combination result by the combination unit to extract a feature value of the combination result, and outputs a discrimination result that discriminates the state of the observation target; An analysis device comprising:

2. The analysis device according to claim 1 , a frequency analysis unit that performs frequency analysis on the first one-dimensional array data and selects the specific frequency from a frequency analysis result; the dividing unit divides the first one-dimensional array data based on the specific frequency selected by the frequency analysis unit to generate the plurality of first two-dimensional array data, and divides the second one-dimensional array data based on the specific frequency selected by the frequency analysis unit to generate the plurality of second two-dimensional array data. An analytical device characterized by:

3. The analysis device according to claim 2, The specific frequency is a frequency having a maximum intensity in the frequency analysis result. An analytical device characterized by:

4. The analysis device according to claim 2, The specific frequency is a frequency whose intensity is equal to or greater than a predetermined threshold value in the frequency analysis result. An analytical device characterized by:

5. The analysis device according to claim 1 , The defect feature amount is the number of defects of the defect. An analytical device characterized by:

6. The analysis device according to claim 1 , The missing point is a point in time when the observed value was not observed. An analytical device characterized by:

7. The analysis device according to claim 1 , The defect is that the observed value is outside a predetermined range. An analytical device characterized by:

8. The analysis device according to claim 1 , the combining unit combines the plurality of first two-dimensional array data and the plurality of second two-dimensional array data so that a kernel used in the convolution operation includes any one of the plurality of first two-dimensional array data and any one of the plurality of second two-dimensional array data; An analytical device characterized by:

9. The analysis device according to claim 1 , an adjustment unit that adjusts a learning model that identifies the state of the observation object based on the identification result and a teacher label that indicates the state of the observation object in the time-series data of the observation value; An analysis device comprising:

10. The analysis device according to claim 1 , a learning model for identifying the state of the observed object; the identification unit inputs the combination result into the learning model, and outputs the identification result from the learning model. An analytical device characterized by:

11. The processor: an input process for inputting time series data of observed values ​​of the observation target; a first counting process for counting the time series data of the observation values ​​input by the input process for each unit time to calculate statistical values ​​and outputting first one-dimensional array data of the time series of the statistical values; a second counting process for counting missing observation values ​​in the time-series data of the observation values ​​for each unit time to calculate missing feature amounts, and outputting second one-dimensional array data of the missing feature amounts in time series; a division process that divides the first one-dimensional array data output by the first counting process based on a specific frequency to generate a plurality of first two-dimensional array data configured with elements for which the statistical value exists and elements for which the statistical value does not exist, and divides the second one-dimensional array data output by the second counting process based on the specific frequency to generate a plurality of second two-dimensional array data configured with elements for which the missing feature exists and elements for which the missing feature does not exist; a combining process for combining the plurality of first two-dimensional array data and the plurality of second two-dimensional array data generated by the dividing process; an identification process for outputting an identification result for identifying the state of the observation target by performing a convolution operation on the combination result of the combination process and extracting a feature amount of the combination result; An analysis method characterized by executing the following.

12. The processor an input process for inputting time series data of observed values ​​of the observation target; a first counting process for counting the time series data of the observation values ​​input by the input process for each unit time to calculate statistical values ​​and outputting first one-dimensional array data of the time series of the statistical values; a second counting process for counting missing observation values ​​in the time-series data of the observation values ​​for each unit time to calculate missing feature amounts, and outputting second one-dimensional array data of the missing feature amounts in time series; a division process that divides the first one-dimensional array data output by the first counting process based on a specific frequency to generate a plurality of first two-dimensional array data configured with elements for which the statistical value exists and elements for which the statistical value does not exist, and divides the second one-dimensional array data output by the second counting process based on the specific frequency to generate a plurality of second two-dimensional array data configured with elements for which the missing feature exists and elements for which the missing feature does not exist; a combining process for combining the plurality of first two-dimensional array data and the plurality of second two-dimensional array data generated by the dividing process; an identification process for outputting an identification result for identifying the state of the observation target by performing a convolution operation on the combination result of the combination process and extracting a feature amount of the combination result; An analysis program characterized by executing the above.

Images