JP7693375B2 - Arithmetic method, arithmetic device - Google Patents

Description

The present invention relates to a calculation method and a calculation device.

Calculations using huge amounts of data are used in industry and the field of natural science. Although the processing power of calculation devices has improved due to technological advances, it is desirable for the size of the data to be processed to be small in order to save calculation resources and improve processing speed. Patent Document 1 discloses a method that can be executed by a device that calculates and displays a risk assessment value of an event sequence, which is a partially ordered set that indicates a part of an event group consisting of a finite number of M types of events, and includes the steps of generating an M-dimensional sparse ordering matrix based on the event sequence, interpolating between elements of the generated sparse ordering matrix to calculate a dense ordering matrix, calculating a mapping matrix that maps the similarity relationship between the event sequences into a two-dimensional space or a three-dimensional space using an embedding method based on the calculated dense ordering matrix, and using the calculated mapping matrix to calculate corresponding points in the two-dimensional space or the three-dimensional space of each event sequence, and displaying and outputting the calculated corresponding points in the two-dimensional space or the three-dimensional space.

International Publication No. 2013/084779

The invention described in Patent Document 1 results in large data size.

A calculation method according to a first aspect of the present invention is a calculation method executed by a calculation device, and includes an acquisition step of acquiring independent measurement data including combinations of inconsistent measurement times and measurement values for a plurality of sensors; a preprocessing step of creating, for each measurement time, mask data that is a numerical value indicating the presence or absence of the measurement value and measurement data that is a numerical value based on the measurement value, and creating a preprocessing table including the measurement data and the mask data of the plurality of sensors for all of the measurement times; and a dimensionality reduction step of performing an embedding process on the measurement data in the preprocessing table, integrating the plurality of measurement data based on at least the similarity of the measurement data, and creating a reduced embedding table that integrates the plurality of mask data.
A calculation device according to a second aspect of the present invention includes an acquisition unit that acquires independent measurement data including combinations of inconsistent measurement times and measurement values for multiple sensors, a pre-processing unit that creates, for each measurement time, mask data that is a numerical value indicating the presence or absence of the measurement value and measurement data that is a numerical value based on the measurement value, and creates a pre-processing table including the measurement data and the mask data of the multiple sensors for all of the measurement times, and a dimensionality reduction unit that performs an embedding process on the measurement data in the pre-processing table, and creates a reduced embedding table that integrates the multiple measurement data and the multiple mask data based on at least the similarity of the measurement data.

According to the present invention, similar data is integrated, thereby reducing data size.

1 is a hardware configuration diagram of a computing device according to a first embodiment. Functional configuration diagram of a computing device according to a first embodiment Diagram showing the relationship between independent measurement data and pretreatment table 1 is a flowchart showing an overall optimization process according to the first embodiment. FIG. 13 is a diagram showing an example of a screen displayed by a user interface. Flowchart showing preprocessing by the preprocessing unit A flowchart showing a process of calculating the number of dimensions by a parameter calculation unit. FIG. 8 is a diagram showing a specific example of step S333 in FIG. 7; FIG. 8 is a diagram for explaining the operations in steps S334 to S336 in FIG. 7; Flowchart showing data usage rate calculation processing Flowchart showing time step calculation processing Flowchart showing dimension reduction calculation process 11 is a flowchart showing a second method for creating a pre-processing table in the first modification example. 11 is a flowchart showing a third method for creating a pre-processing table in the second modification example. Flowchart showing the processing of the dimensionality reduction unit in Modification 3 Functional configuration diagram of a computing device according to a second embodiment FIG. 18 is a diagram showing an example of a correspondence table 18. 1 is a flowchart showing an overall optimization process according to a second embodiment. FIG. 13 is a diagram showing an example of a screen display in the second embodiment; Flowchart showing the limitation process FIG. 13 is a diagram showing a dimensionality reduction process according to the second embodiment; 1 is a flowchart showing an analysis process performed by an analysis unit; FIG. 13 is a diagram showing a dimensionality reduction process in a first modified example of the second embodiment;

-First embodiment-
A first embodiment of a computing device will be described below with reference to FIGS.

(Hardware configuration)
1 is a hardware configuration diagram of a computing device according to the present invention. The computing device 1 includes a CPU 2 which is a central processing unit, a ROM 3 which is a read-only storage device, a RAM 4 which is a readable and writable storage device, a storage device 5, a user interface 6, and a communication module 7. The CPU 2 loads a program stored in the ROM 3 into the RAM 4 and executes it to perform computations, which will be described later.

The arithmetic device 1 may be realized by a field programmable gate array (FPGA), which is a rewritable logic circuit, or an application specific integrated circuit (ASIC), which is an integrated circuit for a specific application, instead of the combination of the CPU 2, ROM 3, and RAM 4. The arithmetic device 1 may also be realized by a combination of different configurations, for example, a combination of the CPU 2, ROM 3, RAM 4, and an FPGA, instead of the combination of the CPU 2, ROM 3, and RAM 4.

The storage device 5 is a readable/writable non-volatile storage device, such as a hard disk drive. The user interface 6 is an interface device that exchanges information with the person using the computing device 1, and is, for example, a touch panel that integrates a liquid crystal display and a pointing device. The user interface 6 may be realized by a combination of a liquid crystal display and a pointing device. The communication module 7 communicates with other computing devices and sensors via wireless or wired communication.

(Functional configuration)
2 is a functional configuration diagram of the arithmetic device 1. The arithmetic device 1 includes, as its functions, an acquisition unit 11, a preprocessing unit 12, an interaction unit 13, a parameter calculation unit 14, and a dimension reduction unit 15. In addition, the arithmetic device 1 stores the independent measurement data 16, the preprocessing table 17A, the time normalization table 17B, the embedding table 17C, and the reduced embedding table 17D in either the RAM 4 or the storage device 5. However, in this embodiment, for the sake of brevity, it is assumed that all of the above information is stored in the RAM 4. As will be described in detail later, the preprocessing table 17A, the time normalization table 17B, the embedding table 17C, and the reduced embedding table 17D include information contained in the independent measurement data 16 itself, or information obtained by processing the information contained in the independent measurement data 16.

The acquisition unit 11 acquires measurement values from each sensor and stores them in the RAM 4 as independent measurement data 16. The acquisition unit 11 may receive measurement values and time from the sensors and store them in the RAM 4, or the acquisition unit 11 may receive measurement values from the sensors and store the current time at the time of reception and the received measurement values in the RAM 4. The acquisition unit 11 may collect not only the output of sensors that are external to the calculation device 1 shown in FIG. 1, but also the output of a calculation program (not shown) included in the calculation device 1.

The preprocessing unit 12 reads the independent measurement data 16 and creates a preprocessing table 17A. The dialogue unit 13 prompts the user for input using the user interface 6 and presents the calculation results to the user. The parameter calculation unit 14 reads the preprocessing table 17A and two parameters input by the user, and calculates another parameter required to create a reduced embedding table 17D. The parameter calculation unit 14 also creates a time normalization table 17B during the processing. The dimension reduction unit 15 creates a reduced embedding table 17D based on the user's input, and creates an embedding table 17C during the processing.

The independent measurement data 16 stores a combination of the measurement value and measurement time taken by each sensor. The timing of measurement by each sensor is not synchronized and is different. For example, even if the measurement value of sensor S1 is taken at time t1, it is possible that the measurement value of sensor S2 is not recorded in the independent measurement data 16 at time t1.

In the preprocessing table 17A, measurement data based on the measurement values stored in the independent measurement data 16 and mask data indicating the presence or absence of a measurement value are stored for all measurement times. The mask data is created so that it corresponds 1:1 to each of the measurement data. The preprocessing table 17A contains all of the information contained in the independent measurement data 16. For example, if the calculation device 1 receives 10 measurement values from 10 sensors and the measurement times associated with all of the measurement values are different, the size of the preprocessing table 17A will be "20x100". The breakdown of this is that the measurement data and mask data are each "10", so the total is "20", and in time series it is "10" x "10" = "100".

Time normalization table 17B is obtained by integrating multiple data in a time series according to the time step of preprocessing table 17A by parameter calculation unit 14. This time step is a value specified by the user, or a value calculated by parameter calculation unit 14 based on other parameters specified by the user. However, if the time step is very fine, time normalization table 17B will be the same as preprocessing table 17A.

The embedding table 17C is created by the dimensionality reduction unit 15 based on the embedding table 17C. The reduced embedding table 17D is created by the dimensionality reduction unit 15 based on the embedding table 17C. The reduced embedding table 17D is an integration of some measurement data and mask data of the embedding table 17C. For example, when the measurement data of the sensor S1 and the sensor S2 are integrated, the mask data of the sensor S1 and the mask data of the sensor S2 are also integrated. The reduced embedding table 17D is a so-called deliverable in this embodiment, and the amount of data is reduced while retaining the characteristics of the independent measurement data 16. Although not described in this embodiment, the reduced embedding table 17D can be used for various applications, for example, in Neural ODE (Ricky T. Q. Chen et al., Neural Ordinary Differential Equations, 2018).

(Pretreatment table)
Fig. 3 is a diagram showing the relationship between the independent measurement data 16 and the preprocessing table 17A. An example of the independent measurement data 16 is shown in the upper part of Fig. 3, and an example of the preprocessing table 17A is shown in the lower part of Fig. 3. Although Fig. 3 will be described using the outputs of sensors S1 and S2, it is also assumed that the independent measurement data 16 may include outputs from hundreds, thousands, or tens of thousands of sensors.

Sensors S1 and S2 operate independently, measuring at irregular intervals and outputting measured values. However, they are referred to as "sensors" merely for convenience, and each "sensor" does not need to be equipped with sensing hardware, and the information output may not be the measurement value itself but a calculation result or an abnormality detection signal. When an abnormality detection signal is output, it may be designed so that no output is made if there is no abnormality, and a value of "1", for example, is output if an abnormality is detected. Here, the output of sensors S1 and S2 is an integer between +10 and -10.

In the example shown in FIG. 3, the independent measurement data 16 indicates that sensor S1 measured "5" at time t1 and "2" at time t8, and sensor S2 measured "0" at time t3 and "3" at time t8. In the independent measurement data 16, the output of each sensor is expressed as an independent table, but next we will try to integrate these two tables. Since time t8 is common to both tables, the time of the table to be integrated is t1, t2, and t8. If the measured values at each time are listed, it will look like the table in the middle of FIG. 3 (hereinafter referred to as the "intermediate table"). However, in the intermediate table, columns with no measured values, specifically the value of sensor S2 at time t1 and the value of sensor S1 at time t2, are set to "NaN" (Not a Number; missing value), indicating that no numerical value exists.

The intermediate table lists the measured values as they are, and columns with no measured values are clearly distinguished from others by storing "NaN" to avoid any misunderstandings. However, from the perspective of data processing, storing any numerical value allows batch processing regardless of whether there is a measured value or not, making it easier to handle. Therefore, in this embodiment, the concept of "measurement data" that corresponds to the measured value is introduced, and also the concept of "mask data" that indicates whether the measurement data is the measured value itself. In other words, measurement data and mask data exist in a 1:1 ratio, and it can be said that each measurement data has corresponding mask data.

For example, a "1" in the mask data indicates that the corresponding measurement data is a measured value, and a "0" in the mask data indicates that the corresponding measurement data is not a measured value. In the example shown in FIG. 3, the name of the mask data corresponding to sensor S1 is "M1", and the name of the mask data corresponding to sensor S2 is "M2". In the intermediate table, the column for time t1 of mask M1 is set to "1" because there is a corresponding measured value of "5". Also, the column for time t2 of mask M1 is set to "0" because the corresponding measured value is "NaN", i.e., does not exist.

After clarifying whether a measurement has been performed using this mask vector, as shown in the lower part of Figure 3, in preprocessing table 17A, the "NaN" in the intermediate table is replaced with an arbitrary numerical value, in this case "0". This replacement allows all data to be represented as numerical values, making data processing easier. In this case, it is impossible to tell if the measurement value was "0" or if no measurement was performed at all by looking only at the measurement data in preprocessing table 17A, but this can be determined by looking at the corresponding mask data.

In FIG. 3, for convenience of drawing, the time is written very simply as "t1", but in reality, a specific date and time is written, such as "February 3, 2021 04:05:06". The granularity of the time is arbitrary, and may be, for example, in units of one hour, one minute, one second, or one hundredth of a second. The granularity of the time may also differ for each sensor or measurement. In the following, all the time-series mask data corresponding to a specific sensor will be referred to as "mask information". For example, FIG. 3 includes mask information for mask M1 and mask information for mask M2.

(process)
4 is a flowchart showing the entire optimization process by the arithmetic device 1. In the arithmetic device 1, first, in step S30, the preprocessing unit 12 performs preprocessing and consolidates the outputs of multiple sensors into one table. The preprocessing will be described in detail later. In the following step S31, the dialogue unit 13 uses the user interface 6 to display a screen and waits for input from the user. The user inputs two of the three items: the number of dimensions, the data utilization rate, and the time step.

Figure 5 shows an example of a screen displayed by the user interface 6 in step S31. Figure 5 has an input interface for specifying the three items mentioned above, namely the number of dimensions, data utilization rate, and time step. When the user sets two of these three items, the parameter calculation unit 14 calculates the remaining item as described below and displays it on the screen. Let's return to Figure 4 to continue the explanation.

In the following step S32, the parameter calculation unit 14 identifies the items to be calculated by the calculation device 1 based on the user's input. Specifically, if the parameter calculation unit 14 determines that the number of dimensions should be calculated since the data utilization rate and time step have been set, it proceeds to step S33. Also, if the parameter calculation unit 14 determines that the data utilization rate should be calculated since the time step and number of dimensions have been set, it proceeds to step S34. Furthermore, if the parameter calculation unit 14 determines that the number of dimensions and data utilization rate should be set, it proceeds to step S35.

In step S33, the parameter calculation unit 14 executes a dimension number calculation process described below and proceeds to step S36. In step S34, the parameter calculation unit 14 executes a data utilization calculation process described below and proceeds to step S36. In step S35, the parameter calculation unit 14 executes a time step calculation process described below and proceeds to step S36. In step S36, the dialogue unit 13 further displays the items calculated in any of steps S33 to S35 on the screen and proceeds to step S37. In step S37, the dimension reduction unit 15 executes a dimension reduction process described below and ends the process shown in FIG. 4.

Figure 6 is a flowchart showing the preprocessing by the preprocessing unit 12, and shows details of step S30 in Figure 4. First, in step S301, the computing device 1 reads all data to be processed and sorts all the times listed, for example in ascending order. In this sorting process, if there are identical times, all but one are deleted. For example, in the example shown in Figure C, the data from sensor S1 records times t1 and t8, and the data from sensor S2 records times t2 and t8, but time t8 overlaps between the two, so one report is deleted and the sorted result becomes t1, t2, t8.

In the next step S302, the calculation device 1 creates a table whose width is twice the number of sensors and whose height is the number of unique times. In the example shown in Figure C, there are "2" sensors and the number of unique times is "3" (t1, t2, and t8) as mentioned above, so a table whose width is "4" (twice "2") and whose height is "3" is created.

In the next step S303, the calculation device 1 transcribes the measurement values of each sensor into the left half of the table created in step S302, and writes "NaN" in any column where there is no measurement value. However, the vertical direction of the table indicates the time in ascending order, and the horizontal direction, left-justified, indicates the type of sensor. In the next step S304, the calculation device 1 writes mask data into the right half of the table created in step S302. At this time, the calculation device 1 sets mask data whose corresponding measurement data is "NaN" to "0", and sets mask data whose corresponding measurement data is other than "NaN", i.e., where some measurement value exists, to an arbitrary value. This arbitrary value may be either "0" or "1".

In the next step S305, the calculation device 1 replaces the "NaN" written in step S303 with "0" and ends the process shown in FIG. 6. In other words, when the process of this step is completed, the preprocessing table 17A shown at the bottom in the example shown in FIG. C is completed.

Figure 7 is a flowchart showing the dimensionality calculation process by the parameter calculation unit 14, and shows details of step S33 in Figure 4. First, in step S331, the calculation device 1 reads the time step and data utilization rate value specified by the user. Note that hereinafter, the time step specified by the user is referred to as the "time step DS", and the data utilization rate specified by the user is referred to as the "data utilization rate DU".

In the next step S332, the calculation device 1 corrects the times in the preprocessing table according to the time step specified by the user. Specifically, each time listed in the preprocessing table is divided by the time step DS, and the result, which is rounded up or down, is used as new time information. For example, if the time step DS is 60 seconds, this may be handled by deleting information in seconds from the time information. In this case, "March 4, 2020, 5:45:23" is changed to "March 4, 2020, 5:45." Note that if this process results in overlapping times in the preprocessing table, the next step is performed.

In the next step S333, the calculation device 1 integrates information from the preprocessing table that has reached the same time as follows. That is, the average value of the sensor measurement values is calculated, and an OR operation, which is a bit operation, is performed on the mask values. However, if the corresponding mask value is "0", no measurement has been performed, so the sensor measurement value can be ignored and not included in the population of the average value. By the processing of this step, the number of time series data in the preprocessing table, i.e. the number of data in the vertical direction of the table, is maintained or decreased. A specific example of the processing of this step will be described with reference to the drawings.

Figure 8 is a diagram showing a specific example of step S333. An excerpt of a portion of the preprocessing table is shown at the top of Figure 8, and record numbers are shown on the left side for ease of explanation. In the example shown in Figure 8, time information is recorded in one-second units for both sensor S1 and sensor S2. Record "52" is a measurement value at 5:45:23 on March 4, 2020, and as can be seen from the mask value, only the measurement value of sensor S2 is recorded. Record "53" is a measurement value at 5:45:48 on March 4, 2020, and as can be seen from the mask value, the measurement values of sensors S1 and S2 are recorded.

Here, if the time step DS is set to 60 seconds, the information in seconds is omitted, and records "52" and "53" are treated as having the same time. In this case, the M1 values of records "52" and "53" are "0" and "1", respectively, so the value of S1, "2", which corresponds to "1", is set as the S1 value of the merged record. The M2 values of records "52" and "53" are both "1", so the S2 value of the merged record is set to "7", which is the average of "8" and "6". Let's return to Figure 7 and continue the explanation.

In the next step S334, the calculation device 1 counts the patterns of mask information, which is each piece of time-series information contained in the mask data, for the preprocessing table 17A processed in step S333. In the next step S335, the calculation device 1 sorts the patterns counted in step S334 in descending order of number. In the next step S336, the calculation device 1 identifies the smallest number of patterns P that exceeds the specified data utilization rate. Here, the processing of steps S334 to S336 will be described in detail with specific examples.

Figure 9 is a diagram explaining the operation of steps S334 to S336. In this specific example, it is assumed that there are 800 pieces of time information in the preprocessing table after processing in step S333, and 1000 target sensors. In this case, if the mask information corresponding to the output of each sensor is treated as vector data, it becomes 800-dimensional vector data, and there are 1000 of these 800-dimensional vector data, the same number as the number of sensors. In the table shown in Figure 9, the specific vector values are arranged from top to bottom in descending order of the number of hits, and a "rank" column is also provided. Furthermore, for convenience in Figure 9, the cumulative number up to that rank is listed at the right end.

As described above, the mask information patterns are tallied by the processing of step S334, and when step S334 is completed, the "vectors" and "counts" shown in FIG. 9 are linked. However, in this state, sorting as shown in FIG. 9 has not yet been performed. When the following step S335 is completed, the data is sorted as shown in FIG. 9. If the data usage rate specified by the user is "85%", the cumulative total up to rank "123" is "847", which is less than 850, which is 85% of 1000. Therefore, since the number of patterns P to be identified in step S336 exceeds the specified 85% by including the next rank "124", the minimum number of patterns P to be identified is "124". Returning to FIG. 7, the explanation will continue.

In step S337, which is executed after step S336, the calculation device 1 calculates the number of dimensions ceil(log2P) and ends the process shown in FIG. 7. Note that hereinafter, the number of dimensions calculated in this step will be referred to as the "number of dimensions D." The function ceil is a function that rounds up real numbers that are arguments, and outputs the result of the rounding up process if the argument is a decimal, and outputs the integer if the argument is an integer. The value calculated in step S337 is the number of dimensions targeted by this flowchart.

Figure 10 is a flowchart showing the data utilization calculation process, and shows details of step S34 in Figure 4. In Figure 10, the same steps as in Figure 7 are given the same step numbers, and the description of these steps will be omitted below. First, in step S341, the calculation device 1 reads the values of the time step and number of dimensions specified by the user. Note that hereinafter, the time step specified by the user will be referred to as the "time step DS", and the number of dimensions specified by the user will be referred to as the "number of dimensions D".

Steps S332 to S335 executed after step S341 are processes using the time step DS read in step S341, and as they have been explained with reference to FIG. 7, explanations will be omitted here. In the following step S346, the calculation device 1 calculates the ratio of the total number of the top two sorted D-power patterns to the total number of patterns as the data utilization rate, and ends the process shown in FIG. 10. The value calculated in step S346 is the data utilization rate targeted by this flowchart.

Here, we will explain in detail the case where the specified number of dimensions D is "3" and the number of sorted patterns is as shown in Figure 9 above. The cube of 2 is "8", and the total number of the top "8" patterns in Figure 9 is "150". And since the total number of items in Figure 9 is "1000", the data utilization rate is calculated to be "15%".

Figure 11 is a flowchart showing the time step calculation process, and shows details of step S35 in Figure 4. First, in step S351, the calculation device 1 reads the data utilization rate and number of dimensions specified by the user. In the following, the data utilization rate specified by the user is referred to as the "data utilization rate DU", and the number of dimensions specified by the user is referred to as the "number of dimensions D". In the following step S352, the calculation device 1 reads the upper limit Tmax and lower limit Tmin of the time step that can be specified. This information may be read from another device, or may be values that have been previously stored in the storage device 5.

In the following step S353, the calculation device 1 sets Ttmp to the average value of Tmin and Tmax. In the following step S354, the calculation device 1 executes step S34, i.e., the process shown in FIG. 10, using a combination of each of Tmin, Tmax, and Ttmp and the specified number of dimensions D to calculate each of Umin, Umax, and Utmp. In other words, in step S354, the process shown in FIG. 10 is executed three times with only the time step parameters changed. Next, the calculation device 1 proceeds to step S355.

In step S355, the calculation device 1 determines whether the calculated value of Utmp is greater than the specified data utilization rate Du. If the calculation device 1 determines that the calculated value of Utmp is greater than the specified data utilization rate Du, the calculation device 1 proceeds to step S356, and if the calculation device 1 determines that the calculated value of Utmp is equal to or less than the specified data utilization rate Du, the calculation device 1 proceeds to step S357. In step S356, the calculation device 1 assigns Utmp to Umin, assigns the average value of Ttmp and Tmax to Utmp, and executes the process of FIG. 10 using the number of dimensions D specified in step S351 and the Dtmp updated in this step to update Utmp, and returns to step S355.

In step S357, the calculation device 1 determines whether the calculated value of Utmp is smaller than the specified data utilization rate Du. If the calculation device 1 determines that the calculated value of Utmp is smaller than the specified data utilization rate Du, the calculation device 1 proceeds to step S358, and if the calculation device 1 determines that the calculated value of Utmp is not smaller than the specified data utilization rate Du, that is, equal to the specified data utilization rate Du, the calculation device 1 proceeds to step S359. In step S358, the calculation device 1 assigns Utmp to Umax, assigns the average value of Ttmp and Tmin to Utmp, and executes the process of FIG. 10 using the number of dimensions D specified in step S351 and the Dtmp updated in this step to update Utmp, and returns to step S355.

In step S359, the calculation device 1 executes the processes of steps S332 and S333 shown in FIG. 7 using the previously calculated value of Ttmp as the time step, and ends the process shown in FIG. 11. The preprocessing table is updated by the process of this step. The value used in this step is the time step targeted by this flowchart. In summary, the process shown in FIG. 11 fixes the number of dimensions to a specified value, changes the time step successively, and searches for a time step value that matches the specified data utilization rate.

Figure 12 is a flowchart showing the dimensionality reduction calculation process, and shows details of step S37 in Figure 4. As shown in Figure 4, any one of steps S34 to S35 is executed before step S37 is executed, and regardless of which step is executed, a time normalization table 17B is created based on preprocessing table 17A using the specified time step or the calculated time step.

In step S370, the calculation device 1 performs a known embedding process for the purpose of reducing the dimension of the measurement data of the time normalization table 17B, and creates the embedding table 17C by mapping it to a vector in a low-dimensional space. Known embedding methods include the method of Wang et al. (Jizhe Wang, et.al., Billion-scale Commodity Embedding for E-commerce Recommendation in Alibaba, 2018), the method of Nalmpantis et al. (Christoforos Nalmpantis, et.al., Signal2Vec: Time Series Embedding Representation, 2019), and the method of Kazemi et al. (Seyed Mehran Kazemi, et.al., Time2Vec: Learning a Vector Representation of Time, 2019).

In the next step S371, the calculation device 1 calculates the cosine similarity between the measurement data in the embedding table 17C and proceeds to step S372. In steps S372 and after, the embedding table 17C is processed to create the reduced embedding table 17D. That is, the reduced embedding table 17D is not complete until the processing shown in FIG. 12 is completed, but for convenience, it will be referred to as the reduced embedding table 17D below even if it is still being created. The initial state of the reduced embedding table 17D is the embedding table 17C created in step S370.

In step S372, the calculation device 1 determines whether the number of types of measurement data in the current reduced embedding table 17D is greater than the specified or calculated number of dimensions D. If the calculation device 1 determines that the number of types of measurement data in the current reduced embedding table 17D is greater than the specified or calculated number of dimensions D, the calculation device 1 proceeds to step S373. If the calculation device 1 determines that the number of types of measurement data in the current reduced embedding table 17D is not greater than the specified or calculated number of dimensions D, the calculation device 1 determines that the current reduced embedding table 17D is the final reduced embedding table 17D and ends the processing shown in FIG. 12.

In step S373, the calculation device 1 identifies a plurality of pieces of measurement data having the highest cosine similarity among the non-integrated measurement data calculated in step S371, and proceeds to step S374. In step S374, the calculation device 1 selects the oldest unprocessed time in the measurement data identified in step S373. In the following step S375, the calculation device 1 counts the number of measurement data identified in step S373 for which the mask data value corresponding to the measurement data at the time selected in step S374 is "1". If the calculation device 1 determines that the mask data value of "1" is zero, it proceeds to step S376, if it determines that the mask data value is "1", it proceeds to step S377, and if it determines that the mask data value is "2" or greater, it proceeds to step S378.

In step S376, the calculation device 1 sets the value of the measurement data to be integrated to zero, and proceeds to step S379. In step S377, the calculation device 1 sets the value of the measurement data to be integrated to the value of the measurement data whose corresponding mask data is "1", and proceeds to step S379. In step S378, the calculation device 1 sets the average value or logical sum of the measurement data whose corresponding mask data is "1" to the value of the measurement data to be integrated, and proceeds to step S379. Note that in each of steps S376 to S378, the mask data uses the value of the logical sum of the objects to be integrated.

In step S379, the calculation device 1 determines whether or not all the data at all times has been processed for the multiple measurement data identified in step S373. If it determines that all the data at all times has been processed, the calculation device 1 returns to step S372, and if it determines that processing of all the data at all times has not been completed, the calculation device 1 returns to step S374. This concludes the explanation of FIG. 12.

According to the above-described first embodiment, the following advantageous effects can be obtained.
(1) The calculation method executed by the calculation device 1 includes an acquisition step executed by the acquisition unit 11, which reads independent measurement data including combinations of non-uniform measurement times and measurement values for multiple sensors, a preprocessing step executed by the preprocessing unit 12, which creates mask data, which is a numerical value indicating the presence or absence of a measurement value, and measurement data, which is a numerical value based on the measurement value, for each measurement time, to create a preprocessing table 17A including the measurement data and mask data of multiple sensors for all measurement times, and a dimension reduction step executed by the dimension reduction unit 15, which performs an embedding process on the measurement data in the preprocessing table 17A, and creates a reduced embedding table 17D that integrates multiple measurement data and multiple mask data based on at least the similarity of the measurement data. Therefore, the reduced embedding table 17D is reduced in size compared to the preprocessing table 17A by the dimension reduction step executed by the dimension reduction unit 15, and therefore the data size is reduced, and calculations using the reduced embedding table 17D are expected to reduce required resources and improve processing load.

(2) The dimension reduction step performed by the dimension reduction unit 15 determines multiple measurement values to be integrated based on the cosine similarity of the measurement data. Therefore, the measurement data can be integrated efficiently.

(3) The calculation method executed by the calculation device 1 includes a time integration step (S332 and S333 in Figs. 7, 10, and 11) that integrates multiple measurement values and multiple mask data combined at different measurement times in the preprocessing table 17A according to a time step, which is a time width. Therefore, the measurement data is arranged in time units of a predetermined time step, making it possible to perform unified processing.

(4) The calculation method executed by the calculation device 1 includes an input step (step S31 in FIG. 4) for accepting the specification of two of the three: the time step, which is the time width to be integrated; the data utilization rate, which is the utilization rate of the measurement data; and the number of dimensions, which is the number of data; a parameter calculation step by the parameter calculation unit 14 for calculating the one of the three that is not specified; and a time integration step for integrating multiple measurement values and multiple mask data that are combined at different measurement times in the preprocessing table 17A according to the time step. Therefore, the three parameters that are related to each other can be determined without contradiction.

(5) The preprocessing step involves setting mask data for each sensor at measurement times when no measurement values exist to zero, and setting measurement data corresponding to nonexistent measurements to zero.

(Variation 1)
In the first embodiment, the mask data was set for the purpose of determining whether a measurement value is present or absent and whether the measurement value is zero. That is, in the first embodiment, when the value of the sensor vector is "0", the mask data was created to determine whether the measurement value is "0" or whether the measurement has not been performed. Therefore, when the measurement data is other than "0", the value of the mask data could be set arbitrarily. However, the design concept of the mask data may be different from that of the first embodiment.

That is, the mask data may be used to determine whether or not a measurement value exists. For example, the mask data may be created so that a "1" in the mask data indicates that a corresponding measurement value exists, i.e., a measurement has been performed, and a "0" in the mask data indicates that a corresponding measurement value does not exist, i.e., a measurement has not been performed. In this case, the value of the sensor data corresponding to a "0" in the mask data can be set arbitrarily.

Figure 13 is a flowchart showing a second method for creating preprocessing table 17A. This flowchart differs from Figure 2 in that steps S304 and S305 are replaced with steps S304A and S305A. This difference is explained below.

In step S304A, the calculation device 1 writes mask data into the right half of the table created in step S302. At this time, the calculation device 1 sets mask data whose corresponding measurement data is "NaN" to "0", and sets mask data whose corresponding measurement data is other than "NaN", i.e., mask data for which some measurement value exists, to "1". In the following step S305A, the calculation device 1 replaces the "NaN" written in step S303 with an arbitrary value, and ends the process shown in FIG. 6. This arbitrary value may be any real number within a specified numerical range, and can take a value at least between the minimum and maximum measurement values.

The following advantageous effects can be obtained in this first modified example.
(6) In the pre-processing step, the mask data at the measurement time when there is no measurement value for each sensor is set to zero, and the mask data at the measurement time when there is a measurement value for each sensor is set to 1. Therefore, the mask data can be created by the second method instead of the first method in accordance with the handling of the mask data in the application that processes the reduced embedding table 17D.

(Variation 2)
Although the first technique described in the first embodiment and the second technique described in the first modified example have different design concepts for the preprocessing table, a preprocessing table may be created to accommodate both of them.

Figure 14 is a flowchart showing a method for creating a preprocessing table that corresponds to both the first method and the second method. This flowchart can be said to have step S304 in Figure 6 changed to step S304A, and step S305A in Figure 13 changed to step S305. By creating a preprocessing table using the method shown in Figure 14, it is possible to create a preprocessing table that corresponds to both the first method and the second method.

(Variation 3)
The processing of the dimension reduction unit 15 may be changed as follows. FIG. 15 is a flowchart showing the processing of the dimension reduction unit 15 in the modified example 3. FIG. 15 is different from FIG. 12 in that steps S370 and S371 are changed to steps S370A and S371A. In step S370A, the dimension reduction unit 15 performs embedding processing not only on the measurement data but also on the mask data. In the following step S371A, the dimension reduction unit 15 concatenates each of the measurement data and the mask data, and calculates the cosine similarity. For example, the measurement data of the sensor S1 and the mask data of the corresponding mask M1 are concatenated into one vector, and the cosine similarity between the vector concatenating the measurement data of the sensor S2 and the mask data of the mask M2 is calculated. The processing from step S372 onwards is the same as that in the first embodiment, so the description will be omitted.

According to the third modification, the following advantageous effects can be obtained.
(7) The dimension reduction step performed by the dimension reduction unit 15 determines multiple measurement values to be integrated based on the cosine similarity between the measurement data and the mask data, which allows for a more precise comparison than when the cosine similarity is calculated only between the measurement data.

(Variation 4)
The arithmetic device 1 does not have to communicate directly with each sensor. For example, the arithmetic device 1 may communicate with the sensor via one or more other devices to receive measurement values, or may read past measurement values of each sensor received by other devices. The arithmetic device 1 may receive past measurement values via the communication module 7, or may read past measurement values stored in a storage medium via a medium reading device (not shown). In other words, the function of the acquisition unit 11 may be to read data from a storage medium.

(Variation 5)
In the first embodiment, the processing executed by the arithmetic device 1 may be realized by a plurality of hardware devices. In other words, the system as a whole may have the functions of the arithmetic device 1, and the number of devices that realize the functions is arbitrary.

--Second embodiment--
A second embodiment of the arithmetic device will be described with reference to Figures 16 to 21. In the following description, the same components as in the first embodiment are given the same reference numerals, and differences will be mainly described. Points that are not specifically described are the same as in the first embodiment. This embodiment differs from the first embodiment mainly in that analysis is performed in addition to the processing of the first embodiment.

Figure 16 is a functional configuration diagram of the arithmetic device 1A in the second embodiment. In addition to the configuration of the arithmetic device 1 in the first embodiment, the arithmetic device 1A further includes an analysis unit 19 and a correspondence table 18. The analysis unit 19 reads the reduced embedding table 17D and identifies explanatory variables corresponding to the objective variables. The objective variables may be specified by the user each time the analysis unit 19 operates, or may be specified in advance. The correspondence table 18 stores the correspondence between the preprocessing table 17A and the sensor types in the reduced embedding table 17D. Note that the hardware configuration of the arithmetic device 1A is the same as that of the first embodiment, so a description thereof will be omitted.

Figure 17 is a diagram showing an example of correspondence table 18. Correspondence table 18 shows the correspondence between measurement data before and after reduction by dimensionality reduction processing, in other words, the correspondence between preprocessing table 17A and reduced embedding table 17D. The example shown in Figure 17 shows that sensor 3, sensor 6, and sensor 9 in preprocessing table 17A and embedded table 17C have been integrated into the third sensor from the left in reduced embedding table 17D.

Figure 18 is a flowchart showing the overall optimization process by the computing device 1A in the second embodiment. The differences from the eleventh embodiment are that step S31, which includes screen display, has been changed to S31A, step S41 has been added, and step S37, which is a dimensionality reduction process, has been changed to step S37A. These will be described in detail.

Figure 19 is a diagram showing an example of a screen display in the second embodiment. Figure 19 differs from the first embodiment in that the context period range can be set. That is, this embodiment has four parameters: context period range, time step, number of dimensions, and data utilization rate. And in this embodiment, it is essential for the user to specify the context period range, and like the first embodiment, the user specifies two of the three parameters: time step, number of dimensions, and data utilization rate.

Figure 20 is a flowchart showing the limiting process for limiting the preprocessing table to the context period range. In step S411, the calculation device 1 identifies reference data, which is any measurement data that serves as a reference for the context period. The reference data may be specified by the user each time, or may be specified in advance. In the following step S412, the calculation device 1 refers to the preprocessing table 17A or the independent measurement data 16 to calculate the occurrence interval of the reference data identified in step S411.

In the next step S413, the calculation device 1 calculates the time that satisfies the ratio specified by the user. For example, if the occurrence interval of the reference data is "2 hours" on average and the ratio corresponding to the context period specified by the user is "1σ", then "1 hour 22 minutes", which is 68% of "2 hours", is calculated. In the next step S414, the calculation device 1 deletes unnecessary rows from the preprocessing table, i.e., data other than the period calculated in step S413 from the occurrence time of the reference data, and ends the process shown in FIG. 20.

For example, if the reference data is S1 and the time calculated in step S413 is "1 hour 22 minutes," the rows to be deleted are determined as follows. That is, the calculation device 1 retains data in the preprocessing table 17A at a time when the measurement data of sensor S1 is not zero or the mask data of sensor S1 is not zero, as well as data from that time up to the time "1 hour 22 minutes" before, and deletes the other data.

Figure 21 is a diagram showing the dimension reduction process in the second embodiment. Compared to Figure 12 in the first embodiment, Figure 21 differs in that step S421 is added between step S373 and step S374. In step S421, the calculation device 1 enters the multiple sensor data identified in step S373 in the correspondence table 18 and proceeds to step S374. Specifically, the calculation device 1 adds to the correspondence table 18 the name of the sensor that measured the sensor data identified in step S373 and information indicating the position of the integrated data, for example, the number of values from the left. The process from step S374 onwards is the same as in the first embodiment, so a description will be omitted.

Figure 22 is a flowchart showing the analysis process by the analysis unit 19. First, in step S431, the analysis unit 19 reads the reduced embedding table 17D. In the following step S432, the analysis unit 19 identifies unmeasured measurement data. The unmeasured measurement values are the values in the columns marked "NaN" in Figure 3, and when the preprocessing table 17A is created by the method shown in Figure 6, they are the values in the column of sensor data where the sensor data, which is the objective variable, is "0" and the corresponding mask data is "0". Note that when the objective variable is integrated with other sensor data, the correspondence table 18 is referenced to identify the data string containing the objective variable.

In the next step S433, the analysis unit 19 estimates the measurement value identified in step S432. This estimation can use various known techniques, for example, interpolation in Word2Vec. In the next step S434, the analysis unit 19 estimates explanatory variables for the objective variable. However, if sensor data is integrated in the process of creating the reduced embedded table 17D, this step does not identify explanatory variables, and only identifies, for example, the fourth data column from the left.

In the following step S435, the analysis unit 19 refers to the correspondence table 18 to identify the name of the data included in the data string identified in step S434. However, if the explanatory variables identified in step S434 are not integrated with other sensor data, the processing of this step can be omitted. In the following step S436, the variables identified in step S436 are displayed on the screen, and the processing shown in FIG. 22 is terminated.

In addition, in this embodiment, the dimensionality reduction process shown in FIG. 12 is modified as follows. That is, in step S373, the names of the identified multiple sensor data and the storage location of the integrated data are stored in the correspondence table 18.

According to the above-described second embodiment, the following advantageous effects can be obtained.
(8) An analysis step includes reading reduced embedding table 17D and identifying sensors having other measurement values that are highly correlated with the measurement value of a predetermined sensor. The analysis step identifies the predetermined sensor and sensors having other measurement values by referring to correspondence table 18 that shows the correspondence between preprocessing table 17A and reduced embedding table 17D. Therefore, since calculation device 1A reads reduced embedding table 17D and performs analysis, analysis can be performed with fewer resources.

(Variation 1 of the second embodiment)
In the second embodiment, the processing of the dimension reduction unit 15 may be changed as follows. Fig. 23 is a diagram showing the processing of the dimension reduction unit 15 in the first modified example of the second embodiment. In step S441, the dimension reduction unit 15 calculates the co-occurrence probability between measurement data in the embedding table 17C. In the following step S442, the dimension reduction unit 15 determines, as objects to be integrated, the measurement data whose co-occurrence probability calculated in step S441 is equal to or greater than a predetermined value, and the mask data corresponding thereto.

In the next step S443, the dimension reduction unit 15 calculates the co-occurrence probability between mask data in the embedding table 17C. In the next step S444, the dimension reduction unit 15 determines, as objects to be integrated, the mask data whose co-occurrence probability calculated in step S443 is equal to or greater than a predetermined value, and the measurement data corresponding to these. In the next step S445, the dimension reduction unit 15 determines, for the measurement data and mask data to be integrated, the integrated value to be the average value of the measurement data and the logical sum of the mask data.

(Variation 1 of the second embodiment)
The interpolation process by the analysis unit 19, i.e., the estimation of the measurement value in step S433 of FIG. 22, may utilize the "interpolate" function or the "fillna" function of the "pandas" library in Python.

In each of the above-described embodiments and variations, the functional block configurations are merely examples. Some functional configurations shown as separate functional blocks may be configured together, or a configuration shown in a single functional block diagram may be divided into two or more functions. In addition, some of the functions of each functional block may be provided by other functional blocks.

In each of the above-described embodiments and modifications, the program of the arithmetic device 1 is stored in a ROM (not shown), but the program may be stored in the storage device 5. Furthermore, the arithmetic device 1 may be provided with an input/output interface (not shown), and the program may be read from another device when necessary via the input/output interface and a medium available to the arithmetic device 1. Here, the medium refers to, for example, a storage medium that is detachable from the input/output interface, or a communication medium, i.e., a network such as a wired, wireless, or optical network, or a carrier wave or digital signal that propagates through the network. Furthermore, some or all of the functions realized by the program may be realized by a hardware circuit or an FPGA.

The above-mentioned embodiments and modifications may be combined with each other. Although various embodiments and modifications have been described above, the present invention is not limited to these. Other aspects that are conceivable within the scope of the technical concept of the present invention are also included within the scope of the present invention.

Reference Signs List 1, 1A... Calculation device 11... Acquisition unit 12... Preprocessing unit 13... Dialogue unit 14... Parameter calculation unit 15... Dimension reduction unit 16... Independent measurement data 17A... Preprocessing table 17B... Time normalization table 17C... Embedding table 17D... Reduced embedding table 18... Correspondence table 19... Analysis unit

Claims (9)

A computing method executed by a computing device, comprising:
acquiring independent measurement data including non-uniform measurement time and measurement value combinations for a plurality of sensors;
a pre-processing step of creating, for each of the measurement times, mask data which is a numerical value indicating the presence or absence of the measurement values and measurement data which is a numerical value based on the measurement values, and creating a pre-processing table including the measurement data and the mask data of the plurality of sensors for all of the measurement times;
and a dimensionality reduction step of performing an embedding process on the measurement data in the preprocessing table, integrating a plurality of the measurement data based on at least the similarity of the measurement data, and creating a reduced embedding table that integrates a plurality of the mask data. 2. The method according to claim 1,
The dimension reduction step determines the plurality of measurement data to be integrated based on a cosine similarity of the measurement data. 2. The method according to claim 1,
A computation method, wherein the dimensionality reduction step determines the plurality of measurement data to be integrated based on a cosine similarity between the measurement data and the mask data. 2. The method according to claim 1,
The calculation method further includes a time integration step of integrating the plurality of measurement values and the plurality of mask data combined at different measurement times in the pre-processing table according to a time step that is a time width. 2. The method according to claim 1,
An input step that accepts the specification of two of the following three items: a time step, which is the time width to be integrated; a data utilization rate, which is the utilization rate of the measurement data; and a number of dimensions, which is the number of data.
a parameter calculation step of calculating one of the three parameters that is not specified;
The calculation method further includes a time integration step of integrating the plurality of measurement values and the plurality of mask data combined at different measurement times in the pre-processing table according to the time step. 2. The method according to claim 1,
The pre-processing step sets the mask data at the measurement time when no measurement value exists for each sensor to zero, and sets the measurement data corresponding to the non-existent measurement value to zero. 2. The method according to claim 1,
The pre-processing step sets the mask data for each sensor at the measurement time when no measurement value exists to zero, and sets the mask data for each sensor at the measurement time when a measurement value exists to "1." 2. The method according to claim 1,
further comprising an analyzing step of reading the reduced embedding table and identifying sensors having other measurements that are highly correlated with the measurements of a given sensor;
A calculation method in which the analyzing step identifies the predetermined sensor and the sensors having the other measurement values by referring to a correspondence table indicating the correspondence between the preprocessing table and the reduced embedded table. An acquisition unit that acquires independent measurement data including combinations of non-uniform measurement times and measurement values for a plurality of sensors;
a pre-processing unit that creates, for each of the measurement times, mask data that is a numerical value indicating the presence or absence of the measurement values and measurement data that is a numerical value based on the measurement values, and creates a pre-processing table including the measurement data and the mask data of the plurality of sensors for all of the measurement times;
a dimensionality reduction unit that performs an embedding process on the measurement data in the preprocessing table, integrates a plurality of the measurement data based on at least the similarity of the measurement data, and creates a reduced embedding table that integrates a plurality of the mask data.

Images