JP7817874B2 - Feature extraction method, program, and device - Google Patents

Description

The present invention relates to a feature extraction method, program, and device for extracting features from multiple unit data.

For example, industrial production facilities are equipped with sensors such as temperature sensors, pressure sensors, and flow rate sensors. The measurement data measured by each sensor is stored in association with a specific physical quantity, such as time. When analyzing the stored measurement data, the difference in unit data, which serves as the unit of processing, becomes an important feature.

Conventionally, multiple unit data sets are extracted from continuously accumulated measurement data, the unit data sets are overlaid at positions derived from the analyst's knowledge and/or insights, graphed, and areas where differences are observed between the unit data sets are extracted by the analyst's visual inspection, generating feature quantities. The generated feature quantities are analyzed using machine learning and statistical analysis methods to identify feature quantities that affect product quality and equipment operation, and are used to identify the causes of quality defects and develop productivity improvement measures.

For example, the following Patent Documents 1 to 4 disclose methods for extracting features from unit data.

Patent No. 4756707 Patent No. 4837520 Patent No. 5348977 Patent No. 5912880

Conventional feature extraction methods such as those described above have had the following issues. For example, unit data for industrial product production equipment may start when the production equipment begins operation. However, even if the starting point of each unit data is used as the origin and the physical quantity positions of each unit data are aligned and compared, it may not be possible to clearly detect differences between the unit data. For example, by overlaying each unit data set with the physical quantity position midway between the unit data as the origin, differences between unit data that would be hidden when aligned with the starting point as the origin may become clear. In other words, unit data is accompanied by extraneous data, and this extra data can hide differences between unit data that should be extracted.

There was also another issue, as described below. The analyst would gradually slide the unit data along the axis of the physical quantity until the curve drawn by the unit data matched most closely, and the remaining differences between the unit data were then used as features. However, the determination that the curve drawn by the unit data best matches depends on the analyst's experience, is not quantified, and can have low reproducibility.

There were also other issues, such as the following. Namely, the generation of features also depended on the analyst's awareness, which meant there was a risk of oversight. In other words, features could not be generated that covered all of the unit data.

The present invention has been made to solve at least one of the above problems, and one of its objectives is to provide a feature extraction method, program, and device that can extract features more appropriately.

A feature extraction method according to one aspect of the present invention is a feature extraction method for extracting feature values from a plurality of unit data, where the plurality of unit data are measurement data measured by a sensor associated with a predetermined physical quantity and serve as processing units. The method includes a change point detection process for detecting change points in the measurement data of the plurality of unit data, an alignment process for aligning the positions of the physical quantities of the plurality of unit data using the detected change points as origins, and a feature extraction process for extracting feature values from the plurality of unit data with the positions of the physical quantities aligned.

A feature extraction method according to another aspect of the present invention is a feature extraction method for extracting feature amounts from multiple unit data, where the multiple unit data are measurement data measured by a sensor associated with a predetermined physical quantity and serve as units of processing. The method includes an alignment step for aligning the physical quantities of the multiple unit data, and a feature extraction step for extracting feature amounts from the multiple unit data with the physical quantities aligned. The alignment step includes a first alignment step for aligning the physical quantities of the multiple unit data with the position of an arbitrary physical quantity as the origin, a second alignment step for calculating a difference area between the multiple unit data after the physical quantities have been aligned, a third alignment step for calculating a difference area sum by adding up the difference areas, a fourth alignment step for repeating the first to third alignment steps while changing the origin, and a fifth alignment step for determining the origin at which the difference area sum is minimized after repeating the first to third alignment steps in the fourth alignment step. In the feature extraction step, feature amounts are extracted from the multiple unit data with the physical quantities aligned at the origin at which the difference area sum is minimized.

A feature extraction method according to yet another aspect of the present invention is a feature extraction method for extracting feature quantities from a plurality of unit data, wherein the plurality of unit data are measurement data measured by a sensor associated with a predetermined physical quantity and serve as units of processing. The method includes an alignment step for aligning the physical quantities of the plurality of unit data, and a feature extraction step for extracting feature quantities from the plurality of unit data with the physical quantities aligned. In the feature extraction step, the plurality of unit data are divided into a plurality of intervals by an arbitrary physical quantity width, and a predetermined statistical quantity is calculated as a feature quantity for each of the plurality of intervals.

A feature extraction program according to one aspect of the present invention is a feature extraction program for extracting feature values from multiple unit data, where the multiple unit data are measurement data measured by a sensor associated with a predetermined physical quantity and serve as units of processing. The program causes a computer to implement a change point detection function that detects change points in the measurement data of the multiple unit data, an alignment function that aligns the positions of the physical quantities of the multiple unit data using the detected change point as the origin, and a feature extraction function that extracts feature values from the multiple unit data with the positions of the physical quantities aligned.

A feature extraction program according to another aspect of the present invention is a feature extraction program for extracting feature amounts from multiple unit data, where the multiple unit data are measurement data measured by a sensor associated with a predetermined physical quantity and serve as units of processing. The feature extraction program causes a computer to implement an alignment function for aligning the physical quantities of the multiple unit data and a feature extraction function for extracting feature amounts from the multiple unit data with the physical quantities aligned. The alignment function includes a first alignment function for aligning the physical quantities of the multiple unit data using the position of an arbitrary physical quantity as the origin, a second alignment function for calculating the difference area between the multiple unit data after the physical quantities have been aligned, a third alignment function for calculating the difference area sum by adding up the difference areas, a fourth alignment function for repeating the first to third alignment functions while changing the origin, and a fifth alignment function for determining the origin at which the difference area sum is minimized after the fourth alignment function has repeated the first to third alignment functions. The feature extraction function extracts feature amounts from the multiple unit data with the physical quantities aligned at the origin at which the difference area sum is minimized.

A feature extraction program according to yet another aspect of the present invention is a feature extraction program for extracting feature quantities from multiple unit data, where the multiple unit data are measurement data measured by a sensor associated with a predetermined physical quantity and serve as units of processing. The feature extraction program causes a computer to implement an alignment function that aligns the physical quantities of the multiple unit data and a feature extraction function that extracts feature quantities from the multiple unit data with the physical quantities aligned. The feature extraction function divides the multiple unit data into multiple intervals using an arbitrary physical quantity width and calculates a predetermined statistical quantity as a feature quantity for each of the multiple intervals.

A feature extraction device according to one aspect of the present invention is a feature extraction device for extracting feature values from multiple unit data, where the multiple unit data are measurement data measured by a sensor associated with a predetermined physical quantity and serve as units of processing. The device includes a change point detection unit that detects change points in the measurement data of the multiple unit data, a position alignment unit that aligns the positions of the physical quantities of the multiple unit data using the detected change point as the origin, and a feature extraction unit that extracts feature values from the multiple unit data with the positions of the physical quantities aligned.

A feature extraction device according to another aspect of the present invention is a feature extraction device for extracting feature quantities from multiple unit data, where the multiple unit data are measurement data measured by a sensor associated with a predetermined physical quantity and serve as units of processing. The feature extraction device includes an alignment unit that aligns the physical quantities of the multiple unit data, and a feature extraction unit that extracts feature quantities from the multiple unit data with the physical quantities aligned. The alignment unit includes a first alignment unit that aligns the physical quantities of the multiple unit data using the position of an arbitrary physical quantity as an origin, a second alignment unit that calculates a difference area between the multiple unit data after the physical quantities have been aligned, a third alignment unit that calculates a difference area sum by adding up the difference areas, a fourth alignment unit that repeats the processing of the first to third alignment units while changing the origin, and a fifth alignment unit that determines the origin at which the difference area sum is minimized after the fourth alignment unit has repeated the processing of the first to third alignment units. The feature extraction unit extracts feature quantities from the multiple unit data with the physical quantities aligned at the origin at which the difference area sum is minimized.

A feature extraction device according to yet another aspect of the present invention is a feature extraction device for extracting feature quantities from a plurality of unit data, where the plurality of unit data are measurement data measured by a sensor associated with a predetermined physical quantity and serve as units of processing. The device includes an alignment unit that aligns the physical quantities of the plurality of unit data, and a feature extraction unit that extracts feature quantities from the plurality of unit data with the physical quantities aligned. The feature extraction unit divides the plurality of unit data into a plurality of intervals using an arbitrary physical quantity width, and calculates a predetermined statistical quantity as a feature quantity for each of the plurality of intervals.

In accordance with one aspect of the present invention, the feature extraction method, program, and device extract feature values from multiple unit data sets while aligning the physical quantities of the multiple unit data sets with the detected change point in the measurement data as the origin. This prevents differences between the unit data sets from being hidden by extraneous data accompanying the unit data sets, allowing for more appropriate feature extraction.

Furthermore, according to a feature extraction method, program, and device according to another aspect of the present invention, feature values are extracted from multiple unit data sets with the physical quantities aligned at the origin where the sum of the differential areas is smallest. This allows feature values to be extracted in a state where the curves drawn by the unit data are considered to be most consistent, without relying on the analyst's experience, resulting in more appropriate feature extraction.

Furthermore, according to a feature extraction method, program, and device according to yet another aspect of the present invention, multiple unit data are divided into multiple intervals using an arbitrary physical quantity width, and a predetermined statistical quantity is calculated as a feature for each of the multiple intervals. This makes it possible to generate features that cover all of the unit data, thereby enabling more appropriate feature extraction.

1 is a flowchart showing a feature extraction method according to the first embodiment of the present invention. FIG. 2 is an explanatory diagram showing the unit data extraction process of FIG. 1; FIG. 2 is an explanatory diagram showing the change-point detection process of FIG. 1; 2 is a flowchart showing the alignment process of FIG. 1 in more detail. FIG. 5 is an explanatory diagram showing the first alignment step of FIG. 4 . FIG. 5 is an explanatory diagram showing the second alignment step of FIG. 4 . FIG. 5 is an explanatory diagram showing the fifth alignment step of FIG. 4 . FIG. 2 is an explanatory diagram showing the feature extraction process of FIG. 1; FIG. 2 is a block diagram showing a feature extraction program according to the first embodiment of the present invention. 1 is a block diagram showing a feature extraction device according to a first embodiment of the present invention; 10 is a flowchart showing a feature extraction method according to second and third embodiments of the present invention. FIG. 10 is a block diagram showing a feature extraction program according to a second embodiment of the present invention. FIG. 10 is a block diagram showing a feature extraction device according to a second embodiment of the present invention. FIG. 11 is a block diagram showing a feature extraction program according to a third embodiment of the present invention. FIG. 10 is a block diagram showing a feature extraction device according to a third embodiment of the present invention.

The following describes embodiments of the present invention with reference to the drawings. The present invention is not limited to the embodiments, and the components can be modified and embodied without departing from the spirit of the invention. Furthermore, various inventions can be created by appropriately combining multiple components disclosed in each embodiment. For example, some components may be omitted from all of the components shown in the embodiments. Furthermore, components from different embodiments may be combined as appropriate.

Embodiment 1.
First, a feature extraction method according to a first embodiment of the present invention will be described with reference to FIGS. 1 to 8. FIG. 1 is a flowchart illustrating the feature extraction method according to the first embodiment of the present invention, FIG. 2 is an explanatory diagram illustrating the unit data extraction step of FIG. 1, and FIG. 3 is an explanatory diagram illustrating the change-point detection step of FIG. 1. Also, FIG. 4 is a flowchart illustrating the alignment step of FIG. 1 in more detail, FIG. 5 is an explanatory diagram illustrating the first alignment step of FIG. 4, FIG. 6 is an explanatory diagram illustrating the second alignment step of FIG. 4, and FIG. 7 is an explanatory diagram illustrating the fifth alignment step of FIG. 4. Also, FIG. 8 is an explanatory diagram illustrating the feature extraction step of FIG. 1.

The feature extraction method shown in Figure 1 is a method for extracting features from multiple unit data. Unit data is measurement data measured by a sensor associated with a predetermined physical quantity, and serves as a unit of processing. In this embodiment, multiple unit data are assumed to be stored continuously in a database.

Any sensor, such as a temperature sensor, pressure sensor, or flow sensor, can be used as the sensor. Therefore, the measurement data measured by the sensor can correspond to any physical quantity (first physical quantity), such as temperature, pressure, or flow rate. The physical quantity (second physical quantity) to which the measurement data is associated is typically time, but may also be other measurement data measured by another sensor. For example, the temperature at a given location may be associated with the pressure or flow rate at that location, or the temperature at a given location may be associated with the temperature, pressure, or flow rate at another location. In other words, the first physical quantity measured by the sensor may be associated with a second physical quantity of a different type from the first physical quantity, or with the same type of second physical quantity.

Another example of a physical quantity (second physical quantity) with which measurement data is associated is a physical position or distance. For example, in a tunnel kiln for continuously heat-treating industrial products along a predetermined trajectory, the temperature (first physical quantity) measured at each location may be associated with the distance from the entrance of the tunnel kiln (second physical quantity). Another example of a physical quantity (second physical quantity) with which measurement data is associated is a wavelength. For example, in a gas analyzer or frequency analyzer for investigating the properties of a test sample, the wavelength (second physical quantity) may be associated with the strength of the response at that wavelength (first physical quantity).

A database can be understood as a set of data in which measurement data such as those described above is continuously stored. "Continuous" does not necessarily mean uninterrupted in an analog sense, but can also include discrete data such as measurement data spaced at predetermined time intervals. The database can be stored on any hardware, such as a hard disk, memory, computer, or server. The feature extraction method can be performed by any device, such as a computer or dedicated circuit. The device that performs the feature extraction method may include hardware that stores the database, or may be provided separately from that hardware and connected to it.

While the sensor can measure any object, in this embodiment, the sensor's measurement object will be described as a batch production facility for industrial products. More specifically, an example will be described in which a sensor measures the temperature at a predetermined position in a heat treatment furnace that heat treats industrial products for a predetermined period of time, and the sensor's measurement data is associated with time and continuously stored in a database. The temperature of the heat treatment furnace is raised after an industrial product is inserted into the heat treatment furnace, maintained at a high temperature for a predetermined period of time, and then lowered. The temperature of the heat treatment furnace is maintained at a predetermined low temperature until the next industrial product to be processed is inserted into the heat treatment furnace.

Figure 2 shows a database showing such temperature changes in a heat treatment furnace. More specifically, the database shown in Figure 2 includes first unit data D1 relating to the production of a first industrial product, second unit data D2 relating to the production of a second industrial product, and third unit data D3 relating to the production of a third industrial product. The first to third industrial products are the same type of products that are subjected to the same heat treatment.

As shown in FIG. 2, the temperature may be increased and decreased in stages rather than at a uniform rate. In the example shown in FIG. 2, after an industrial product is inserted into the heat treatment furnace, the temperature of the heat treatment furnace is increased from the first temperature T1 to the second temperature T2 at a predetermined rate and maintained at the second temperature T2 for a predetermined period of time. Next, the temperature is increased from the second temperature T2 to the third temperature T3 at a predetermined rate and maintained at the third temperature T3 for a predetermined period of time. Next, the temperature is later decreased from the third temperature T3 to the first temperature T1. When decreasing from the third temperature T3 to the first temperature T1, the temperature is decreased at a predetermined first rate, and then at a second rate that is slower than the first rate.

A processing unit may be a set of measurement data associated with a specific physical quantity, intended to find differences between sets of measurement data. A processing unit may be a set of measurement data during each process when the same type of process is being performed repeatedly. More specifically, a processing unit may be a set of measurement data during each batch process in a batch production facility for industrial products. A processing unit may also be a set of measurement data from the inlet to the outlet in a continuous production facility for industrial products, such as a tunnel kiln. A processing unit may also be a set of measurement data during each measurement process in a test sample measurement device, such as a gas analyzer or frequency analyzer.

As shown in FIG. 1, the feature extraction method of this embodiment 1 includes a unit data extraction process (step S1), a noise removal process (step S2), a change point detection process (step S3), a registration process (step S4), and a feature extraction process (step S5). Each process can be performed in this order.

The unit data extraction process (step S1) is a process of extracting multiple unit data, which serve as processing units, from the database. In this process, first to fourth unit data are extracted from the database shown in FIG. 2. For example, the start point of a unit data is recognized as the timing at which the first temperature T1 is maintained for a predetermined time (the timing at which the first physical quantity is maintained in the first state for a predetermined time), and the end point of the unit data is recognized as the timing at which the first temperature T1 is maintained for a predetermined time from that start point, passing through the second and third temperatures T2 and T3. This allows unit data to be extracted. For example, the first temperature T1 or a temperature between the first temperature T1 and the second temperature T2 can be set as a threshold, and the timing at which the measurement data is maintained at or below that threshold for a predetermined time can be recognized as the timing at which the first temperature T1 is maintained for a predetermined time. There may be data between the first to fourth unit data that is not included in these unit data. Alternatively, the end point of one unit data may be used as the start point of the next unit data.

The noise removal process (step S2) is a process for removing noise from the extracted multiple unit data. Noise is high-frequency fluctuation contained in the measurement data, and can be removed, for example, by obtaining a moving average of the measurement data or by applying a low-pass filter to the measurement data.

The change point detection process (step S3) is a process for detecting change points in the measurement data in multiple unit data from which noise has been removed. Change points in the measurement data can be detected as timings when the second-order derivatives of multiple unit data exceed a predetermined value or become discontinuous. Figure 3(a) shows unit data from which noise has been removed, (b) shows the first-order derivatives of the unit data, and (c) shows the relationship between the second-order derivatives of the unit data and the pulse P contained in the second-order derivatives. As shown in Figure 3(c), pulse P appears in the second-order derivatives of the unit data at the positions of change points in the measurement data contained in the unit data. By detecting the timings when the second-order derivatives exceed a predetermined value or become discontinuous, the positions of pulse P can be detected and the positions of pulse P can be detected as change points. Alternatively, change points may be detected using other methods, such as detecting points where the measurement data of the noise-removed unit data changes by more than a predetermined value, or detecting the timings when the first-order derivatives of the unit data become discontinuous.

The alignment process (step S4) is a process of aligning the positions of the physical quantities of multiple unit data from which noise has been removed, using the detected change point as the origin. In this embodiment, the position of the time axis of the unit data is adjusted so that the time positions (positions of the physical quantities) detected as change points match. Figure 5 shows an example of overlapping multiple unit data whose time axis positions have been adjusted. When multiple change points are detected in the change point detection process, a specific change point may be selected at least initially according to a predetermined rule, and alignment may be performed using the selected change point as the origin. In the example of Figure 5, the change point at the timing when the temperature reaches the second temperature T2 is set as the origin.

As shown in FIG. 4, the alignment process (step S4) in the feature extraction method of this embodiment 1 includes a first alignment process (step S4-1), a second alignment process (step S4-2), a third alignment process (step S4-3), a fourth alignment process (step S4-4), and a fifth alignment process (step S4-5).

The first alignment process (step S4-1) is a process of aligning the physical quantities of multiple unit data from which noise has been removed, using the change point as the origin. The processing in this process is the same as the processing in the alignment process described above.

The second alignment process (step S4-2) is a process for calculating the difference area between multiple unit data whose physical quantities have been aligned. As shown in Figure 6, the difference area can be understood as the sum of the areas of the differences between unit data, with the measurement data (temperature) on the vertical axis and the physical quantity associated with that measurement data on the horizontal axis. In other words, it can be understood as the sum of the differences between the measurement data of one unit data and the measurement data of another unit data at each physical quantity position. The difference between the measurement data of one unit data and the measurement data of another unit data can be treated as an absolute value before the sum is calculated.

In the second alignment process of this embodiment, after determining the average data (average unit data) of the aligned multiple unit data, the difference area between the average data and the multiple unit data is determined. The average data can be determined by averaging the measurement data (temperature) of each unit data at each physical quantity (time) position. Figure 6 shows the difference area between the average data and the second unit data. Alternatively, one of the multiple unit data may be selected according to a predetermined rule, and the difference area between the selected unit data and the other unit data may be determined.

The third alignment process (step S4-3) is a process of adding up the difference areas to find the sum of the difference areas. In this embodiment, the third alignment process finds the sum of the difference areas by adding up the first difference area between the average data and the first unit data, the second difference area between the average data and the second unit data, and the third difference area between the average data and the third unit data.

The fourth alignment process (step S4-4) is a process in which the first to third alignment processes are repeated while changing the origin. The origin can be changed by shifting it from the current origin by a predetermined physical quantity width (time width). In the initial first alignment process, the unit data is roughly aligned by using the change point as the origin, and in the fourth alignment process, the origin is gradually changed (i.e., the position of the unit data is fine-tuned) to find the smallest difference area sum, and difference area sums can be collected. Alternatively, the origin may be sequentially changed to a change point different from the change point used in the initial first alignment process.

The fifth alignment process (step S4-5) is a process of determining the origin at which the differential area sum is smallest after repeating the first to third alignment processes in the fourth alignment process. Figure 7 shows an example of a graph showing the relationship between the differential area sum collected in the fourth alignment process and the value of the physical quantity (time) used as the origin when calculating the differential area sum. The origin at which the differential area sum is smallest is the origin at which the deviation between each unit of data is smallest, and can be used as the optimal origin for correcting the deviation between each unit of data.

Returning to Figure 1, the feature extraction process (step S5) is a process of extracting feature values from multiple unit data sets while aligning the positions of the physical quantities. In this embodiment, the feature extraction process extracts feature values from multiple unit data sets while aligning the positions of the physical quantities at the origin determined in the fifth alignment process (the origin where the differential area sum is smallest).

The extracted feature values and their extraction method are determined arbitrarily and appropriately depending on the processing purpose of the unit data. In the feature extraction process of this embodiment, as shown in FIG. 8, multiple unit data are divided into multiple intervals S using arbitrary physical quantity widths, and a predetermined statistical value is calculated as a feature value for each of the multiple intervals S. The physical quantity width used to divide into intervals S can be set appropriately depending on the processing purpose of the unit data and/or the data acquisition period, for example, every 5 minutes when the data interval of the unit data is 1 minute. Furthermore, considering that statistical values will be calculated in subsequent processing, the physical quantity width used to divide into intervals S is preferably set so that each interval contains three or more measurement data points. Intervals S can be set for each predetermined physical quantity width, based on the origin. Intervals S can be set adjacent to each other. The entire physical quantity range of the unit data may be divided into intervals S. Alternatively, a partial physical quantity range of the unit data, such as a partial range where measurement data exceeds a predetermined value (i.e., a partial range containing significant measurement data), may be divided into intervals S. Examples of calculated statistics include the average value and/or slope (rate of change) of the measurement data (temperature) for each section S. By comparing and examining the feature amounts extracted from each aligned unit of data, it is possible to more reliably grasp the differences in phenomena between each unit of data.

Next, FIG. 9 is a block diagram showing a feature extraction program 1 according to embodiment 1 of the present invention. The feature extraction program 1 shown in FIG. 9 is for extracting feature quantities of unit data, which serve as processing units, from a database 2 in which measurement data measured by a sensor is associated with a predetermined physical quantity and continuously stored. The program is loaded into a computer 3 and executed to cause the computer 3 to implement functions corresponding to each step of the feature extraction method described above. The database 2 shown in FIG. 9 is a continuous storage of measurement data measured by a sensor, associated with a predetermined physical quantity. Although the database 2 is shown external to the computer 3 in FIG. 9, the database 2 may also be stored in hardware included in the computer 3. Also, although a single computer 3 is shown in FIG. 9, the computer 3 shown in FIG. 9 may be implemented by multiple computers. In this case, some functions may be implemented by one computer and other functions may be implemented by another computer.

When feature extraction program 1 is loaded into computer 3 and executed, the functions realized by computer 3 include a unit data extraction function 31 that extracts multiple unit data from database 2, a noise removal function 32 that removes noise from the extracted multiple unit data, a change point detection function 33 that detects change points in the measurement data in the multiple unit data from which noise has been removed, an alignment function 34 that aligns the positions of the physical quantities of the multiple unit data from which noise has been removed, using the detected change point as the origin, and a feature extraction function 35 that extracts feature quantities from the multiple unit data with the physical quantities aligned.

The change point detection function 33 can detect the timing when the second-order differential values of multiple unit data exceed a predetermined value or become discontinuous as a change point.

The alignment function 34 includes a first alignment function 341 that aligns the physical quantities of multiple unit data from which noise has been removed, using the change point as the origin; a second alignment function 342 that calculates the difference area between multiple unit data whose physical quantities have been aligned; a third alignment function 343 that adds up the difference areas to calculate the difference area sum; a fourth alignment function 344 that repeats the first to third alignment functions 341 to 343 while changing the origin; and a fifth alignment function 345 that determines the origin at which the difference area sum is minimized after the fourth alignment function 344 repeats the first to third alignment functions 341 to 343. The feature extraction function 35 can extract features from multiple unit data with the physical quantities aligned at the origin at which the difference area sum is minimized.

The second alignment function 342 can calculate the average data of the aligned multiple unit data, and then calculate the differential area between the average data and the multiple unit data.

The feature extraction function 35 divides multiple unit data into multiple intervals S with an arbitrary physical quantity width, and can calculate a predetermined statistical quantity as a feature for each of the multiple intervals S.

As mentioned above, each function corresponds to a step in the feature extraction method, and the explanation of each step in the feature extraction method above also applies to each function.

Next, FIG. 10 is a block diagram showing a feature extraction device 4 according to embodiment 1 of the present invention. The feature extraction device 4 shown in FIG. 10 extracts features of unit data, which serve as processing units, from a database 2 in which measurement data measured by a sensor is associated with a predetermined physical quantity and continuously stored, and includes each unit (functional block) for executing each step of the feature extraction method described above. The database 2 shown in FIG. 10 is measurement data measured by a sensor that is associated with a predetermined physical quantity and continuously stored. Although FIG. 10 shows the database 2 outside the feature extraction device 4, the database 2 may also be stored in hardware included in the feature extraction device 4. Furthermore, while FIG. 10 shows a single feature extraction device 4, the feature extraction device 4 shown in FIG. 10 may be implemented by multiple devices. In this case, some functional blocks may be implemented by one device and other functional blocks may be implemented by another device.

The feature extraction device 4 includes a unit data extraction unit 41 that extracts multiple unit data from the database 2, a noise removal unit 42 that removes noise from the extracted multiple unit data, a change point detection unit 43 that detects change points in the measurement data in the multiple unit data from which noise has been removed, a position alignment unit 44 that aligns the positions of the physical quantities of the multiple unit data from which noise has been removed, using the detected change point as the origin, and a feature extraction unit 45 that extracts feature quantities from the multiple unit data with the physical quantities aligned.

The change point detection unit 43 can detect the timing at which the second-order differential values of multiple unit data exceed a predetermined value or become discontinuous as a change point.

The alignment unit 44 includes a first alignment unit 441 that aligns the physical quantities of multiple unit data from which noise has been removed, using the change point as the origin; a second alignment unit 442 that calculates the difference area between multiple unit data whose physical quantities have been aligned; a third alignment unit 443 that calculates the difference area sum by adding up the difference areas; a fourth alignment unit 444 that repeats the first to third alignment units 441 to 443 while changing the origin; and a fifth alignment unit 445 that determines the origin at which the difference area sum is minimized after the fourth alignment unit 444 repeats the first to third alignment units 441 to 443. The feature extraction unit 45 can extract feature quantities from the multiple unit data with the physical quantities aligned at the origin at which the difference area sum is minimized.

After calculating the average data of the multiple unit data that have been aligned, the second alignment unit 442 can calculate the difference area between the average data and the multiple unit data.

The feature extraction unit 45 divides multiple unit data into multiple intervals S with arbitrary physical quantity widths, and can calculate a predetermined statistical quantity as a feature for each of the multiple intervals S.

As described above, each part corresponds to a step in the feature extraction method, and the explanation of each step in the feature extraction method above also applies to each part.

In the feature extraction method, program, and device of this embodiment, feature values are extracted from multiple unit data sets while aligning the physical quantities of the multiple unit data sets, with the change point of the detected measurement data set as the origin. This prevents differences between the unit data sets from being hidden by extraneous data accompanying the unit data sets, allowing for more appropriate feature extraction.

In addition, the timing at which the second-order differential values of multiple unit data exceed a predetermined value or become discontinuous is detected as a change point, making it possible to detect change points more reliably.

Furthermore, because feature quantities are extracted from multiple unit data sets with the physical quantities aligned at the origin where the differential area sum is minimized, feature quantities can be extracted in a state where the curves drawn by the unit data are considered to be most consistent, without relying on the analyst's experience, resulting in more appropriate feature extraction. This configuration is particularly useful when the range of physical quantities between change points differs between unit data sets. For example, in a batch-type device that uses a vacuum process, there is a program that moves on to the next step, such as increasing the temperature, once a predetermined degree of vacuum is reached. The time it takes to reach a predetermined degree of vacuum can vary depending on the ambient environment, such as the season, meaning that the range of physical quantities between change points can differ between unit data sets. In such cases, it can be difficult to determine which change point is best to align the origin with. The configuration described above can solve this problem.

Furthermore, after calculating the average data of the aligned multiple unit data, the difference area between the average data and the multiple unit data is calculated, which reduces the computational load compared to calculating the difference area between all unit data. Furthermore, compared to calculating the difference area based on specific unit data, the risk of bias in the results obtained can be reduced.

In addition, multiple unit data are divided into multiple intervals S using arbitrary physical quantity widths, and a predetermined statistical quantity is calculated as a feature for each of the multiple intervals S, so that features can be generated that cover the entire unit data, allowing for more appropriate feature extraction.

Embodiment 2.
FIG. 11 is a flowchart showing a feature extraction method according to a second embodiment of the present invention. The feature extraction method according to the second embodiment omits the change-point detection step (step S3) of the first embodiment. The alignment step (step S4) of the second embodiment includes the first alignment step (step S4-1), second alignment step (step S4-2), third alignment step (step S4-3), fourth alignment step (step S4-4), and fifth alignment step (step S4-5) described with reference to FIG. 4. The first alignment step (step S4-1) of the second embodiment aligns the positions of the physical quantities of a plurality of unit data from which noise has been removed, using the position of an arbitrary physical quantity as the origin. That is, while the origin is determined based on the change point in the first embodiment, the origin may be determined according to other methods/criteria, such as the position of a physical quantity specified by an operator. The rest of the method is the same as that of the feature extraction method according to the first embodiment.

Next, FIG. 12 is a block diagram showing a feature extraction program 1 according to a second embodiment of the present invention. The feature extraction program 1 according to the second embodiment is a program in which the change-point detection function 33 is omitted from the functions implemented by a computer described in the first embodiment. The alignment function 34 according to the second embodiment includes a first alignment function 341, a second alignment function 342, a third alignment function 343, a fourth alignment function 344, and a fifth alignment function 345. The first alignment function 341 according to the second embodiment aligns the positions of the physical quantities of multiple unit data from which noise has been removed, using the position of any physical quantity as the origin. That is, while the origin was determined based on the change point in the first embodiment, the origin may also be determined according to other methods/criteria, such as the position of a physical quantity specified by an operator. The rest is the same as the feature extraction program according to the first embodiment.

Next, FIG. 13 is a block diagram showing a feature extraction device 4 according to a second embodiment of the present invention. The feature extraction device of the second embodiment is a device in which the change point detection unit 43 described in the first embodiment is omitted. The alignment unit 44 of the second embodiment includes a first alignment unit 441, a second alignment unit 442, a third alignment unit 443, a fourth alignment unit 444, and a fifth alignment unit 445. The first alignment unit 441 of the second embodiment aligns the positions of the physical quantities of multiple unit data from which noise has been removed, using the position of an arbitrary physical quantity as the origin. That is, while the origin was determined based on the change point in the first embodiment, the origin may also be determined according to other methods/criteria, such as the physical quantity position specified by the operator. Other aspects are the same as those of the feature extraction device of the first embodiment.

In this way, the origin may be determined according to a method/criteria different from that of embodiment 1. In the feature extraction method, program, and device of embodiment 2, feature quantities are extracted from multiple unit data sets while aligning the physical quantities at the origin where the sum of the differential areas is minimized. This allows feature quantities to be extracted in a state where the curves drawn by the unit data are considered to be most consistent, without relying on the analyst's experience, resulting in more appropriate feature quantity extraction. This configuration is particularly useful when the range of physical quantities between change points differs between unit data sets. For example, in a batch-type device involving a vacuum process, there is a program that transitions to the next step, such as increasing the temperature, once a predetermined degree of vacuum is reached. The time required to reach a predetermined degree of vacuum may vary depending on the ambient environment, such as the season, meaning that the range of physical quantities between change points may differ between unit data sets. In such cases, it is difficult to determine which change point is best to align the origin with. The above configuration can solve this problem.

Embodiment 3.
The flowchart illustrating the feature extraction method of the third embodiment is similar to the flowchart illustrating the feature extraction method of the second embodiment, and therefore FIG. 11 is also used. That is, the feature extraction method of the third embodiment, like the second embodiment, omits the change-point detection step (step S3) of the first embodiment. Furthermore, while the alignment step (step S4) of the second embodiment is described as including the first to fifth alignment steps (steps S4-1 to S4-5), the alignment step (step S4) of the third embodiment does not include the first to fifth alignment steps (steps S4-1 to S4-5). In the alignment step (step S4), the positions of the physical quantities of the plurality of unit data from which noise has been removed are aligned by any method/criteria, such as an operator's operation. In the feature extraction step (step S5), the plurality of unit data are divided into a plurality of intervals S by any physical quantity width, and a predetermined statistical quantity is calculated as a feature for each of the plurality of intervals S. The remaining steps are similar to the feature extraction methods of the first and second embodiments.

Next, FIG. 14 is a block diagram showing a feature extraction program 1 according to a third embodiment of the present invention. The feature extraction program of the third embodiment is a program in which the change-point detection function 33 and the first to fifth alignment functions 341 to 345 are omitted from the functions implemented by a computer described in the first embodiment. The alignment function 34 of the second embodiment aligns the positions of the physical quantities of multiple unit data from which noise has been removed, using any method/criteria, such as an operator's operation. The feature extraction function 35 divides multiple unit data into multiple intervals S using any physical quantity width, and calculates a predetermined statistical quantity as a feature for each of the multiple intervals S. The rest is the same as the feature extraction program of the first and second embodiments.

Next, FIG. 15 is a block diagram showing a feature extraction device 4 according to a third embodiment of the present invention. The feature extraction device of the third embodiment is a device in which the change point detection unit 43 and the first to fifth alignment units 441 to 445 described in the first embodiment are omitted. The alignment unit 44 of the second embodiment aligns the positions of the physical quantities of multiple unit data from which noise has been removed, using any method/criteria, such as an operator's operation. The feature extraction unit 45 divides the multiple unit data into multiple intervals S with any physical quantity width, and calculates a predetermined statistical quantity as a feature for each of the multiple intervals S. The rest is the same as the feature extraction devices of the first and second embodiments.

In this way, the positions of the physical quantities of multiple unit data may be aligned according to methods/standards different from those used in embodiments 1 and 2. In the feature extraction method, program, and device of embodiment 3, multiple unit data are divided into multiple intervals S using arbitrary physical quantity widths, and a predetermined statistical quantity is calculated as a feature for each of the multiple intervals S. This makes it possible to generate features that cover all of the unit data, allowing for more appropriate feature extraction.

In the first to third embodiments, it has been described that multiple unit data are stored continuously in a database, but feature amounts may also be extracted from multiple unit data that have been individually prepared in advance. In other words, the unit data extraction process, unit data extraction function, and unit data extraction unit for extracting multiple unit data from a database may be omitted.

Furthermore, in embodiments 1 to 3, the noise removal process, noise removal function, and noise removal unit for removing noise may be omitted, for example, when there is very little noise in the multiple unit data.

1: Feature extraction program 2: Database 3: Computer 31: Unit data extraction function 32: Noise removal function 33: Change point detection function 34: Alignment function 341: First alignment function 342: Second alignment function 343: Third alignment function 344: Fourth alignment function 345: Fifth alignment function 35: Feature extraction function 4: Feature extraction device 41: Unit data extraction unit 42: Noise removal unit 43: Change point detection unit 44: Alignment unit 441: First alignment unit 442: Second alignment unit 443: Third alignment unit 444: Fourth alignment unit 445: Fifth alignment unit 45: Feature extraction unit

Claims (21)

A feature extraction method for extracting feature values from a plurality of unit data, wherein the plurality of unit data are measurement data measured by a sensor associated with a predetermined physical quantity and serve as a unit of processing;
a change point detection step of detecting a change point of the measurement data in the plurality of unit data;
an alignment step of aligning positions of the physical quantities of the plurality of unit data with the detected change point as an origin;
a feature extraction step of extracting the feature from the plurality of unit data while aligning the positions of the physical quantities,
The alignment step includes:
a first alignment step of aligning positions of the physical quantities of the plurality of unit data using the change point as an origin;
a second alignment step of calculating a difference area between the plurality of unit data after aligning the physical quantities;
a third alignment step of adding up the difference areas to obtain a sum of the difference areas;
a fourth alignment step of repeating the first to third alignment steps while changing the origin;
a fifth alignment step of determining an origin at which the difference area sum is minimum after repeating the first to third alignment steps in the fourth alignment step;
Including,
In the feature extraction step, the feature is extracted from the plurality of unit data in a state where the physical quantity is aligned with an origin at which the difference area sum is minimum.
Feature extraction methods. In the change point detection step, timings at which second-order differential values of the plurality of unit data items exceed a predetermined value or become discontinuous are detected as the change points.
The feature extraction method according to claim 1 . In the second alignment step, average data of the aligned plurality of unit data is calculated, and then the difference area between the average data and the plurality of unit data is calculated.
The feature extraction method according to claim 1 or 2 . In the feature extraction step, the plurality of unit data are divided into a plurality of sections by an arbitrary physical quantity width, and a predetermined statistical quantity is calculated as the feature for each of the plurality of sections.
The feature extraction method according to any one of claims 1 to 3 . the plurality of unit data are continuously stored as a database,
a unit data extraction step of extracting the plurality of unit data from the database,
the change point detection step detects the change points in the plurality of unit data extracted in the unit data extraction step;
The feature extraction method according to any one of claims 1 to 4 . a noise removal step of removing noise from the extracted plurality of unit data,
In the change point detection step, a change point of the measurement data in the plurality of unit data from which the noise has been removed is detected.
The feature extraction method according to any one of claims 1 to 5 . A feature extraction method for extracting feature values from a plurality of unit data, wherein the plurality of unit data are measurement data measured by a sensor associated with a predetermined physical quantity and serve as a unit of processing;
an alignment step of aligning positions of the physical quantities of the plurality of unit data;
a feature extraction step of extracting the feature from the plurality of unit data while aligning the positions of the physical quantities,
The alignment step includes:
a first alignment step of aligning positions of the physical quantities of the plurality of unit data with a position of any one of the physical quantities as an origin;
a second alignment step of calculating a difference area between the plurality of unit data after aligning the physical quantities;
a third alignment step of adding up the difference areas to obtain a sum of the difference areas;
a fourth alignment step of repeating the first to third alignment steps while changing the origin;
a fifth alignment step of determining an origin at which the difference area sum is minimum after repeating the first to third alignment steps in the fourth alignment step;
In the feature extraction step, the feature is extracted from the plurality of unit data in a state where the physical quantity is aligned with an origin at which the difference area sum is minimum.
Feature extraction methods. A feature extraction program for extracting feature values from a plurality of unit data, the plurality of unit data being measurement data measured by a sensor associated with a predetermined physical quantity and serving as a unit of processing;
a change point detection function for detecting a change point of the measurement data in the plurality of unit data;
a positioning function for aligning positions of the physical quantities of the plurality of unit data with the detected change point as an origin;
a feature extraction function of extracting the feature from the plurality of unit data while aligning the positions of the physical quantities ,
The alignment function includes:
a first alignment function that aligns positions of the physical quantities of the plurality of unit data using the change point as an origin;
a second alignment function for calculating a difference area between the plurality of unit data items after aligning the physical quantities;
a third alignment function for calculating a sum of the difference areas by adding up the difference areas;
a fourth alignment function that repeats the first to third alignment functions while changing the origin;
a fifth alignment function that determines an origin at which the difference area sum is minimum after the fourth alignment function repeats the first to third alignment functions; and
Including,
the feature extraction function extracts the feature from the plurality of unit data in a state where the physical quantity is aligned at an origin where the difference area sum is minimum;
Feature extraction program . the change point detection function detects, as the change point, a timing at which second-order differential values of the plurality of unit data exceed a predetermined value or become discontinuous;
The feature extraction program according to claim 8 . the second alignment function calculates average data of the aligned plurality of unit data, and then calculates the difference area between the average data and the plurality of unit data.
10. The feature extraction program according to claim 8 or 9 . the feature extraction function divides the plurality of unit data into a plurality of sections by an arbitrary physical quantity width, and calculates a predetermined statistical quantity as the feature for each of the plurality of sections;
11. The feature extraction program according to claim 8 . the plurality of unit data are continuously stored as a database,
further realizing in the computer a unit data extraction function for extracting the plurality of unit data from the database;
the change point detection function detects the change points in the plurality of unit data extracted by the unit data extraction function;
12. The feature extraction program according to claim 8 . further realizing a noise removal function in the computer for removing noise from the extracted plurality of unit data;
the change point detection function detects change points of the measurement data in the plurality of unit data from which the noise has been removed;
13. The feature extraction program according to claim 8 . A feature extraction program for extracting feature values from a plurality of unit data, the plurality of unit data being measurement data measured by a sensor associated with a predetermined physical quantity and serving as a unit of processing;
a position alignment function for aligning positions of the physical quantities of the plurality of unit data;
a feature extraction function of extracting the feature from the plurality of unit data while aligning the positions of the physical quantities,
The alignment function includes:
a first alignment function that aligns positions of the physical quantities of the plurality of unit data with a position of any one of the physical quantities as an origin;
a second alignment function for calculating a difference area between the plurality of unit data items after aligning the physical quantities;
a third alignment function for calculating a sum of the difference areas by adding up the difference areas;
a fourth alignment function that repeats the first to third alignment functions while changing the origin;
a fifth alignment function that determines an origin at which the difference area sum is minimum after the fourth alignment function repeats the first to third alignment functions;
the feature extraction function extracts the feature from the plurality of unit data in a state where the physical quantity is aligned at an origin where the difference area sum is minimum;
Feature extraction program. A feature extraction device for extracting feature values from a plurality of unit data, wherein the plurality of unit data are measurement data measured by a sensor associated with a predetermined physical quantity and serve as a unit of processing;
a change point detection unit that detects change points of the measurement data in the plurality of unit data;
a positioning unit that aligns positions of the physical quantities of the plurality of unit data with the detected change point as an origin;
a feature extraction unit that extracts the feature from the plurality of unit data while aligning the positions of the physical quantities,
The alignment unit is
a first alignment unit that aligns positions of the physical quantities of the plurality of unit data using the change point as an origin;
a second alignment unit that calculates a difference area between the plurality of unit data items after aligning the physical quantities;
a third alignment unit that calculates a sum of the difference areas by adding up the difference areas;
a fourth alignment unit that repeats the processes of the first to third alignment units while changing the origin;
a fifth alignment unit that determines an origin at which the difference area sum is minimum after the fourth alignment unit has repeated the processes of the first to third alignment units; and
Including,
the feature extraction unit extracts the feature from the plurality of unit data in a state where the physical quantities are aligned at an origin where the difference area sum is minimum;
Feature extraction device. the change point detection unit detects, as the change point, a timing at which second-order differential values of the plurality of unit data exceed a predetermined value or become discontinuous.
The feature extraction device according to claim 15 . the second alignment unit calculates average data of the aligned plurality of unit data, and then calculates the difference area between the average data and the plurality of unit data.
17. The feature extraction device according to claim 15 or 16 . the feature extraction unit divides the plurality of unit data into a plurality of sections by an arbitrary physical quantity width, and calculates a predetermined statistical quantity as the feature for each of the plurality of sections;
18. The feature extraction device according to claim 15 . the plurality of unit data are continuously stored as a database,
a unit data extraction unit that extracts the plurality of unit data from the database,
the change point detection unit detects the change points in the plurality of unit data extracted by the unit data extraction unit.
19. The feature extraction device according to claim 15 . a noise removal unit that removes noise from the extracted plurality of unit data,
the change point detection unit detects change points of the measurement data in the plurality of unit data from which the noise has been removed.
20. The feature extraction device according to claim 15 . A feature extraction device for extracting feature values from a plurality of unit data, wherein the plurality of unit data are measurement data measured by a sensor associated with a predetermined physical quantity and serve as a unit of processing;
a positioning unit that aligns positions of the physical quantities of the plurality of unit data;
a feature extraction unit that extracts the feature from the plurality of unit data while aligning the positions of the physical quantities,
The alignment unit is
a first alignment unit that aligns positions of the physical quantities of the plurality of unit data with a position of any one of the physical quantities as an origin;
a second alignment unit that calculates a difference area between the plurality of unit data items after aligning the physical quantities;
a third alignment unit that calculates a sum of the difference areas by adding up the difference areas;
a fourth alignment unit that repeats the processes of the first to third alignment units while changing the origin;
a fifth alignment unit that determines an origin at which the difference area sum is minimum after the fourth alignment unit has repeated the processes of the first to third alignment units,
the feature extraction unit extracts the feature from the plurality of unit data in a state where the physical quantities are aligned at an origin where the difference area sum is minimum;
Feature extraction device.