JP2010244149A - Data classification method, data classification device, diagnostic method and diagnostic device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To solve the following problem: when there is a difference in variation of data items to be classified, the number of categories for classifying data increases, thereby increasing calculation cost. <P>SOLUTION: A computer obtains standard deviation of each item of sampled input data, preprocesses the input data based on the obtained standard deviation of each item, selects a prototype of a category closest to the preprocessed input data, evaluates whether or not the selected prototype is proper from relation between the prototype and the input data, and classifies a plurality of pieces of the input data constituted of a plurality of items into a plurality of categories. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to a data classification method and apparatus intended for application to an abnormality detection and abnormality diagnosis system.

  As a prior art of the data classification method, there is a data classification method described in Patent Document 1. The prior art described in Patent Document 1 narrows down the abnormal location even when an abnormal event other than the abnormal event assumed in advance occurs by a data classification device using Adaptive Resonance Theory (ART). Diagnostic devices that can be described are described.

JP 05-165375 A

  However, when the conventional technique is applied to abnormality detection, the number of categories may increase under certain conditions, and the calculation cost may increase.

  An object of the present invention is to provide a data classification method for classifying data with fewer categories and reducing calculation costs while maintaining data classification performance equivalent to that of the prior art.

  In the present invention, the computer calculates the standard deviation of each item of the sampled input data, preprocesses the input data based on the calculated standard deviation of each item, and selects the prototype of the category closest to the preprocessed input data The data classification method evaluates whether the selected prototype is valid from the relationship between the prototype and the input data, and classifies a plurality of input data composed of a plurality of items into a plurality of categories.

  According to the present invention, when applied to abnormality detection or abnormality diagnosis, data is preprocessed according to the characteristics of the data, so that the data can be classified with a small number of categories, and the calculation cost can be reduced. It becomes.

The figure which shows an example of measurement data. The figure which shows the measurement data normalized. The figure which shows the outline | summary of the result classified according to the prior art. The figure which shows the structure of the 1st Example of this invention. The figure which shows the processing flow of a data pre-processing apparatus. The figure which shows the learning algorithm by 1st Example of this invention. The figure which shows the outline | summary of the classification result by 1st Example of this invention. The figure which shows the structure of the 2nd Example of this invention. The figure which shows the example of a category correspondence matrix. The figure which shows the learning algorithm by the 2nd Example of this invention. The figure which shows the example of determination mode. The figure which shows the structure of the 3rd Example of this invention. The figure which shows an example of maintenance data. The figure which shows the data conversion algorithm of a data control apparatus.

  The present invention will be described using each example.

  A first embodiment for classifying a plurality of input data into a plurality of categories is proposed.

  A second embodiment is also proposed in which a plurality of input data is classified into a plurality of preset categories B.

  A third embodiment for diagnosing equipment is proposed using the first embodiment for classifying a plurality of input data into a plurality of categories.

  Furthermore, a fourth embodiment for diagnosing equipment using the second embodiment for classifying a plurality of input data into a plurality of preset categories B is proposed.

  According to the embodiments described below, it is possible to provide a data classification method for classifying data in fewer categories and reducing the calculation cost while maintaining data classification performance equivalent to that of the prior art. When the conventional technique is applied to abnormality detection, the number of categories may increase under certain conditions and the calculation cost may increase. An example will be described below with reference to FIGS.

  FIG. 1 is a trend graph of two measurement values (data A and data B) measured when a certain device is operating normally. This example is an example in the case where the vehicle is operated under three operation conditions of operation conditions 1 to 3. Consider the case where this measured value is learned as normal operation data. In general, since the units and sizes of measurement values vary, first, the measurement values are normalized. Here, the values of data A and data B are normalized so as to fall within the range of [0.05, 0.95].

  FIG. 2 shows the relationship between normalized data A and data B. The data consists of three clusters under operating conditions 1 to 3. By normalizing the measurement values, the range that can be taken by the data A and the data B becomes the same, but the variation of the data B is larger than that of the data A, and the data group is elongated in the direction of the data 2. . This is because the range of data and the ratio of data variation are different between data A and data B.

  Next, this data is classified into a plurality of categories, and an area in which the device is in a normal state is defined. FIG. 3 shows the result of classifying data using the prior art. The broken line schematically represents the category area. In this example, the data was classified into eight categories A1 to A8. In the prior art, the size of the category area is variable, but the shape of the area is always circular. Therefore, the number of categories increases when the data group has an elongated shape due to the characteristics of the data.

  The data classification apparatus according to the first embodiment is a data classification apparatus that classifies a plurality of input data into a plurality of categories, and includes a prototype of the category and an area determination parameter that defines the size of the category area. An internal data storage unit to store; data preprocessing means for pre-processing input data so as to obtain the variance of each item of sampled input data, and so that all of the calculated distributions of each item fall within a certain range; and the data pre-processing Prototype selection means that selects the prototype of the category closest to the input data processed by the means, prototype evaluation means that evaluates the similarity between the prototype and the input data based on the distance between the two, and a new Prototype addition means for creating a new prototype, area determination parameters and prototype Internal data correction means capable of correcting at least one of the group, and in the prototype evaluation means, when the prototype is determined to be valid, at least one of the region determination parameter and the prototype is obtained by the internal data correction means. The data classification apparatus is characterized in that a new prototype is created by the prototype addition means when the prototype evaluation means does not determine that the prototype is valid.

  A data classification apparatus according to the second embodiment is a data classification apparatus that classifies a plurality of input data into a plurality of preset category B, and includes a category A prototype related to the category B, a category A An internal data storage section for storing area determination parameters that define the size of the area, a category correspondence matrix storage section for storing a category correspondence matrix for converting category A into category B, and each item of sampled input data Data pre-processing means for pre-processing the input data so that the distribution of all the calculated items falls within a certain range, and a category A prototype closest to the input data processed by the data pre-processing means is selected. Prototype selection means to be used and whether the selected prototype is valid Prototype evaluation means for evaluating from the similarity between the input data and the input data, prototype addition means for newly creating a new prototype of category A, at least one of a region determination parameter and a prototype, and internal data for correcting the category correspondence matrix; Matrix correction means and category conversion means for converting the selected category A into category B using the category correspondence matrix, and the internal data and matrix correction means are selected as the category correspondence matrix A In the case of adding a preset category B relationship, if the selected category A already corresponds to another category B, the addition of the corresponding relationship is rejected, and if the addition is rejected, the category A Protota A data classification device, wherein the region determination parameter is corrected so that the region determination parameter is equal to a distance between the category A prototype and the input data. It is.

  The facility diagnosis system according to the third embodiment includes the data classification device according to the first aspect of the facility diagnosis system, and the operation data of the facility is input to the data classification device, and the classified category is obtained. The facility diagnosis system is characterized in that the state of the facility is diagnosed based on the above.

  A facility diagnosis system according to a fourth embodiment includes the data classification device according to the second aspect of the present invention, which gives the state of the facility as a teacher pattern, and the relationship between the state of the facility and the operation data of the facility is the data classification It is an equipment diagnosis system characterized by learning with a device, inputting unlearned operation data into the data classification device after learning, and diagnosing the state of the device based on category B output by the data classification device.

  FIG. 4 shows an outline of the first embodiment of the present invention. The present embodiment is a data classification device that classifies input data into several categories. The category to be classified in advance is not given as teacher data, and is unsupervised classification.

  This embodiment comprises input data preprocessing means 10, data classification means 20, and processing flow control means. The data classification unit 20 includes a data reading unit 21, an internal data initialization unit 22, an internal data storage unit 23, a prototype selection unit 24, a prototype evaluation unit 25, a prototype addition unit 26, and an internal data correction unit 27.

  The input data preprocessing means 10 performs preprocessing so as to correct variations due to each item of input data. Since the input data is generally a physical quantity such as temperature, flow rate, voltage, and signal intensity, the magnitude of the value and the fluctuation amount vary depending on the type and characteristics of the input data to be classified. In order to correct these variations, the input data preprocessing means 10 preprocesses data using the maximum value, minimum value and variance of the input data to be classified. Details of the data processing will be described later.

  Next, the data classification unit 20 will be described. The data reading means 21 reads preprocessed input data in response to a command from the processing flow control means 30. The internal data initialization unit 22 is executed only at the beginning of the data classification process, and initializes the data stored in the internal data storage unit 23 according to the input data.

  The internal data storage unit 23 stores a prototype for each category and a parameter for determining the area of each category (area determination parameter). The prototype is a representative example of each category, and corresponds to the category on a one-to-one basis. If the input data is N-dimensional data, the prototype is also N-dimensional data. The area determination parameter is a parameter that defines the area of each category, and is defined for each category. The prototype selection means 24 is a function that compares the preprocessed input data with the prototype stored in the internal data storage unit 23 and selects the prototype closest to the input data. Which prototype is close to the input data is determined by the distance between the input data and the prototype. That is, the distance between the prototype stored in the internal data storage unit 23 and the preprocessed input data is calculated, and the prototype with the closest distance is selected. Specific processing will be described later.

The prototype evaluation unit 25 evaluates whether the prototype selected by the prototype selection unit 24 is valid. The selected prototype is the closest to the input data among those stored in the internal data storage unit 23, but it is not necessarily a sufficiently close prototype. The prototype evaluation means 25 compares the input data, the distance r _mj between the selected prototype j and the region determination parameter R _j, and if the expression (1) is satisfied, determines that the selected prototype j is valid, and inputs Determine the category into which the data is classified. At the same time, the processing of the internal data correction means 27 is performed.

r _mj ≦ R _j (1)

  If the expression (1) is not satisfied, the processing of the prototype adding means 26 is performed.

  The prototype adding unit 26 adds a prototype when the prototype evaluating unit 25 determines that the selected prototype is not valid. Since the prototype must represent the input data, data in the vicinity of the input data is added as a new prototype as will be described later. Note that the same data as the input data may be added as a prototype.

  The internal data correction means 27 is a means for making the data classification characteristics more appropriate by correcting the prototype and region determination parameters when the selected prototype is valid. Therefore, the internal data correction unit 27 includes an area determination parameter correction unit and a prototype correction unit. The region determination parameter correction means compares the distribution state of the data included in the category region with the region determination parameter, and decreases the region determination parameter when the region determination parameter is larger than the data distribution. The prototype correction means brings the prototype closer to the input data. The degree of approach can be adjusted by a parameter. The prototype is a representative example of a plurality of input data, and the input data is sequentially input. By processing the prototype closer to the input data, it becomes possible to set the prototype more appropriately.

  The processing flow control means 30 controls the processing flow in the data classification device. The main functions are to execute the internal data initializing means 22 when the input data is first read and to determine the termination of the repetitive processing from the change of the category output from the prototype evaluating means 25.

  Next, the processing flow of the first embodiment will be described in detail. In this embodiment, the input data is two-dimensional data, but the same processing is performed even if the input data is three-dimensional data or more. Further, it is assumed that there are M sets of input data (X, Y) and teacher categories. The data classification method is a method for appropriately classifying data by repeatedly giving M sets of data several times. In this example, the number of repetitions was K. Hereinafter, the processing flow will be described.

  First, the processing flow of the input data preprocessing means 10 will be described with reference to FIG.

  Step 101: Calculate the maximum value and the minimum value for each of (X, Y). For example, if the data X is the temperature and the data Y is the flow rate, the maximum value of X is 560 and the minimum value is 500, assuming that the temperature fluctuates from 500 ° C. to 560 ° C. and 2.5 t / h to 3.5 t / h, respectively. The maximum value of Y is 3.5 and the minimum value is 2.5.

  Step 102: The input data is normalized with the maximum and minimum values obtained in Step 101. Here, both X and Y are normalized so that the minimum value is 0.1 and the maximum value is 0.9.

  Step 103: Steady state data is extracted from the input data. Here, the steady-state data is data in the case of operating for a while under a fixed condition without changing the operating condition.

  Step 104: Calculate the standard deviation of the extracted data. The extracted data is steady-state data, but the data varies slightly due to sensor errors, process disturbances, etc., so the standard deviation, which is an indicator of the degree of variation, is normalized to X and Y respectively. Calculate on the data.

  Step 105: Using the obtained standard deviation value, the normalized data is corrected so that the variations are approximately the same. Here, assuming that the normalized data is (Xn, Yn), the corrected data is (Xn ′, Yn ′), the standard deviations σx, σy of Xn, Yn, and the standard deviation as a reference is σs, the correction formula is Equations (2) and (3) are obtained.

Xn ′ = (Xn−0.5) * σs / σx + 0.5 (2)
Yn ′ = (Yn−0.5) * σs / σy + 0.5 (3)

  The standard deviation serving as a reference is the smaller of σx and σy.

  Due to the processing of Step 101 to Step 105, the variation of each item of the input data is made uniform. Therefore, when data is classified, it is possible to suppress an increase in the number of categories more than necessary. In addition, as a pre-process for uniformizing the variation in data between each category item, the input data may be pre-processed based on the standard deviation of each item, and methods other than Equation (2) and Equation (3) can be used. good. For example, the input data can be preprocessed based on the standard deviation of each item by equalizing the variation by multiplying the normalized data with the smaller standard deviation by a predetermined value. good.

  Next, the processing flow of the data classification unit 20 will be described.

  Step 201: Initialize prototypes and area determination parameters that are internal data. The prototype initialization process is to create a new prototype. In this example, prototype (Xj, Yj) (j = 1 to 5) was created. Since it is desirable that the initial prototypes are dispersed as much as possible, a combination of x and y in which x and y each fall from 0.1 to 0.9 is created using random numbers. Prototypes are, for example, (0.45, 0.28), (0.78, 0.31), (0.15, 0.17), (0.54, 0.39), (0.78, 0.12).

The region determination parameter R _j was created by the equation (4).

R _j = K_R × √N (4)

  K_R is a constant and N is the number of dimensions of the input data. In this embodiment, since K_R = 0.1 and N = 2, the initial value of R is 0.1414.

  Step 202: Read the data of the input data number m = 1 at the first repetition.

Step 203: The distance r _mj between the read input data (x, y) and the prototype (Xj, Yj) is calculated by the equation (5).

r _mj = √ ((x−Xj) 2+ (y−Yj) 2) (5)

Select j as the prototype candidate for which the distance r _mj is minimum.

Step 204: Evaluate the validity of the prototype. Specifically, the r _mj, compared to reference a region determining parameter R _j, r _mj is less than R _j, evaluating the prototype selected is valid, performs the process of Step 206. If the prototype is not valid, Step 205 is executed.

  Step 205: One new prototype is added here. The prototype to be added is determined by equations (6) and (7).

X′j = kp × xm (6)
Y′j = kp × ym (7)

  Here, kp is a parameter and takes a value in the vicinity of 1. When kp = 1, the prototype is equal to the input data xm, ym.

  Step 206: In the process when the prototype is evaluated to be valid in Step 204, the prototype is corrected according to the input data. More specifically, the correction is performed using Expressions (8) and (9).

X′j = Xj + kw × (xm−Xj) (8)
Y′j = Yj + kw × (ym−Yj) (9)

  Here, kw is a parameter, and takes a value in the range of 0 to 1, and the degree of correction according to the input data is determined by the value of kw. When kw = 1, this means that the prototype is corrected until it completely matches the input data. When kw = 0, the prototype is not corrected.

  The value of kw was changed stepwise as the data classification process progressed. Specifically, in the initial process, the value of kw is increased so that the prototype changes in accordance with the input data, and when the process proceeds to some extent, the value of kw is decreased so that the prototype does not change.

Step 207: It is determined whether or not the region determination parameter R _j is valid from the data included in the region of the selected category. Here, the maximum value r _{mj_max} of the distance r _mj between the prototype of the selected category and the input data included in this category is obtained, and the ratio F_r to R _j is calculated by the equation (10).

F_r = r _{mj_max} / R _j (10)

The threshold value F_r_th set in advance is compared with F_r obtained by Expression (10). If Expression (11) is satisfied, it is determined that the region determination parameter R _j is valid, and the process proceeds to Step 209. If Expression (11) does not hold, it is determined that the region determination parameter R _j is not valid, and the process proceeds to Step 208.

    F_r ≧ F_r_th (11)

  Although the validity is evaluated here based on the distance, another evaluation method for evaluating the validity by, for example, predetermining the relationship between the prototype of the selected category and the input data included in this category may be used. .

Step 208: Change the region determination parameter R _j according to the equation (12).

R _j = r _{mj_max} / F_r_th (12)

  Step 209: It is determined whether or not all M sets of input data have been processed. If there is unprocessed input data, the next input data is processed in Step 203. If all input data has been processed, the process proceeds to Step 210.

  Step 210: The process is terminated when the number of repetitions has been set K times or the data classification process has converged. The convergence determination of the data classification is performed by determining whether the category into which the input data is classified is the same as the previous time (k−1) and the current time (k time). If at least one of the classified categories has changed, it is determined that the process has not converged. If all of the classified categories are the same, it is determined that the process has converged.

  In the first embodiment, the processing illustrated in FIG. 6 is performed from k = 1. However, the processing of Steps 207 to 208 is not performed until k reaches a certain value, and Steps 207 to 208 are performed after k exceeds the certain value. A processing flow for performing the above process may be used.

  FIG. 7 is a conceptual diagram illustrating a classification result according to the first embodiment. The input data is the same as in FIG. 1, but the data B, which has a large variation in data, is pre-processed by the data preprocessing device 10, so that the variation is almost the same as the data A. In this case, the categories were only divided into three categories A1 ′ to A3 ′, and the data could be classified with a small number of categories.

  Therefore, when the classification method according to the present embodiment is applied to abnormality detection and abnormality diagnosis, data can be classified with a small number of categories, so that the calculation cost can be reduced.

  The data classification means 20 is not limited to the classification method described in the present embodiment, and data classification means such as adaptive resonance theory may be used.

  As described above, the computer obtains the standard deviation of each item of the sampled input data, preprocesses the input data based on the obtained standard deviation of each item, and selects the prototype of the category closest to the preprocessed input data. A conventional method for selecting and evaluating whether or not the selected prototype is valid from the relationship between the prototype and the input data, and classifying a plurality of input data composed of a plurality of items into a plurality of categories. While maintaining the same data classification performance, it is possible to classify data with fewer categories and reduce the calculation cost.

  Further, as described above, when the prototype is determined to be valid by the evaluation, at least one of the region determination parameter and the prototype is corrected, and when the prototype is not determined to be valid by the evaluation, a new prototype is selected. The classification characteristics of the data can be made more appropriate by the data classification method to be added.

  Next, a second embodiment of the present invention will be described. FIG. 4 shows an outline of the second embodiment.

  This embodiment is a data classification method for learning a category corresponding to input data. Here, the category that is the classification result in the first embodiment and the category that is learned in advance do not necessarily correspond one-to-one. Therefore, the category which is the classification result of the first embodiment is defined as category A, and the category to be learned in advance is defined as category B, and the second embodiment will be described. Since many of the components of the second embodiment are the same as those of the first embodiment, only the differences from the first embodiment will be described here.

  There are three components newly added by the data classification means 40 of the second embodiment: a category B reading means 41, a category conversion means 42, and a category correspondence matrix 43. Further, the internal data correction means 27 of the first embodiment is replaced with internal data and matrix correction means 28.

  The category B reading means 41 is means for reading category B, which is data (teacher data) to be learned in advance. Similar to the data reading means 21, the processing flow control means 30 controls the timing for reading data.

  The category conversion means 42 is means for converting category A into category B. Here, the category A is the same as the category in which the data is classified in the first embodiment, and some similar data are grouped into one group. Category A is assigned a category number, but the number itself has no meaning. On the other hand, category B is a category that humans have given meaning. When this data classification method is used for abnormality diagnosis, for example, events such as “normal”, “abnormal 1”, and “abnormal 2” are assigned to category B.

  In order to convert category A into category B, a category correspondence matrix that represents the correspondence between category A and category B is used.

  The category correspondence matrix storage unit 43 stores a category correspondence matrix. An example of the category correspondence matrix is shown in FIG. In the example of FIG. 9, there are three types of categories B: “normal”, “abnormal 1”, and “abnormal 2”. Category A has 10 types of categories 1-10. A plurality of categories A correspond to one category B. Each category always corresponds to only one category B, and does not correspond to two or more categories B. By using this category correspondence matrix, if category A is given, it can be converted to category B.

  The internal data and matrix correction means 28 corrects the internal data stored in the internal data storage unit 23 and the category correspondence matrix stored in the category correspondence matrix storage unit 43.

  The internal data to be corrected is a prototype and a region determination parameter, and is the same as the internal data correction means 27 of the first embodiment. However, the correction method is different from the internal data correction means 27. As shown in FIG. 9, the internal data and matrix correction means 28 corrects the internal data so that the category B does not become two or more when it corresponds to the category A. A specific correction algorithm will be described later.

  The category correspondence matrix is corrected so that the relationship between category A and category B always matches the classification result of the latest input data. Since category B corresponding to each input data is given as teacher data, there is no change. However, if the prototype or region determination parameter changes, category A may change even with the same input data. Therefore, when category A is changed, the correspondence with category B is sequentially corrected.

  Next, the operation of the second embodiment will be described. In the second embodiment, there are two types of operation modes, a learning mode and a determination mode. The learning mode is an operation mode in which input data and category B (teacher data) are given and learning is performed so that the input data is classified into category B. The determination mode is an operation mode in which only input data is given and a category B into which the input data is classified is determined. The input data is two-dimensional data (X, Y) as in the first embodiment.

  Since the processing flow of the input data preprocessing means is exactly the same as in the embodiment, the description is omitted.

  FIG. 10 shows a processing flow of the learning mode of the data classification means 40. A portion surrounded by a broken line in the drawing is a portion different from the processing flow of the first embodiment. The processing flow will be described focusing on the different parts.

  Steps 201 to 205: Normalize input data, initialize internal data such as prototypes, and determine a category A prototype into which the input data is classified.

  Step 206-1: The relationship between the category A into which the input data is classified and the category B given as teacher data is added to the category correspondence matrix. However, in the combination of category A and category B to be added, if the selected category A has already recorded a corresponding relationship with another category B, the addition is rejected. If the correspondence between category A and category B is already recorded in the matrix, no correction is made.

  Step 206-2: The process of the correction of the prototype differs depending on whether or not addition is rejected in Step 206-1. If the addition is not rejected in Step 206-1, the prototype is corrected by Expression (8) and Expression (9). When the addition is rejected, it means that the category A into which the input data is classified is not suitable. Therefore, it is necessary to correct the prototype so that the target input data is not classified into the classified category A. Specifically, the prototype is corrected by the equations (13) and (14).

X′j = Xj−kw × (xm−Xj) (13)
Y′j = Yj−kw × (ym−Yj) (14)

  In equations (13) and (14), if kw> 0, the prototype is corrected away from the input data. The greater the value of kw, the greater the degree. When kw = 0, the prototype is not changed.

Step 207: The validity determination of R _j differs depending on whether or not the expression (11) is established and whether or not additional refusal is present. If Expression (11) is satisfied and there is no additional refusal, it is determined to be valid and the process proceeds to Step 209. If the expression (11) is not satisfied or there is an additional refusal, it is determined that it is not valid, and the process proceeds to Step 208.

  Step 208: If there is an additional refusal, it means that the category A into which the input data is classified is not suitable. Therefore, it is necessary to correct the region determination parameter so that the target input data is not classified into the classified category A. Specifically, the correction is made by the equation (15).

R _j = r _mj (15)

Further, when Expression (11) is not satisfied, R _j is corrected by Expression (12).

  Steps 209 to 210: It is determined whether or not the process is repeated. If it is not necessary, the process proceeds to Step 211. When all processing up to Step 210 is completed, all M input data are classified into a certain category A, and the correspondence between each category A and category B is recorded in the category correspondence matrix.

  Step 211: Convert category A to category B using the category correspondence matrix. The data to be converted is M pieces of input data. When the example of FIG. 9 is used as a matrix, if category A of certain input data is category 3, the corresponding category B is abnormal 1, and category 3 is converted to abnormal 1.

  Through the above processing, M pieces of input data can be classified into category B given as teacher data.

  Note that the processing from Step 206-1 to Step 208 is the same as the processing from Step 206 to Step 208 in the first embodiment except for the processing related to the category correspondence matrix. In the second embodiment, the processing shown in the figure is performed from k = 1, but the processing related to the category correspondence matrix is not performed until k reaches a certain value, and only the processing from Step 206 to Step 208 of the first embodiment is performed. It is good also as a processing flow which implements and adds the process relevant to a category correspondence matrix, after k exceeds a certain value.

  Next, the determination mode will be described with reference to FIG. A case where one input data is determined will be described.

  Step 221: The internal data is brought into a state where the learning mode is finished.

  Step 222: Read input data.

  Step 223: A prototype is selected by the same algorithm as Step 203 in the learning mode.

  Step 224: The validity of the prototype is determined by the same algorithm as Step 204 in the learning mode. If it is valid, it is determined that the category A of the input data is the category A having the selected prototype, and the process proceeds to Step 225. If not valid, the category A of the input data is determined as a new category, and the process proceeds to Step 225.

  Step 225: The category A is converted to the category B by the same algorithm as in the learning mode Step 211. However, when the category A of the input data is a new category, the correspondence relationship is not recorded in the category correspondence matrix, so the category B is undefined.

  Finally, effects of the second embodiment will be described. Also in the second embodiment, the data pre-processing means 10 makes the data variation of each item uniform, so that the category A can be classified with a smaller number of categories compared to the prior art, and the calculation cost can be reduced.

  Next, a third embodiment of the present invention will be described. The third embodiment of the present invention is a remote diagnosis system for a power generation facility using the data classification device according to the second embodiment of the present invention. As will be described in the second embodiment, the data classification apparatus of the first embodiment can also be implemented.

  The configuration of this embodiment is shown in FIG. In this embodiment, the power generation facilities 8 a to 8 n and the monitoring center are connected, and the power generation facilities 8 a to 8 n and the monitoring center are connected by the Internet 7.

  The power generation facilities 8 a to 8 n include a combined cycle power generation system 81, a control device 82, and a data transmission device 83. The combined cycle power generation system 81 includes a compressor, a combustor, a gas turbine, an exhaust heat recovery boiler, a steam turbine, a condenser, and a generator.

  First, the combined cycle power generation system 81 will be described. Compressed air from the compressor is sent to a combustor, mixed with fuel, and burned in the combustor. The gas turbine rotates by the high-pressure gas generated by the combustion to generate electricity. The gas turbine exhaust gas enters the exhaust heat recovery boiler and generates steam from the turbine exhaust gas. The generated steam rotates the steam turbine to generate electricity.

  The control device 82 controls the output of the combined cycle power generation system 81 according to the power demand. In addition, the operation data is measured using a sensor installed in the combined cycle power generation system 81. In the present embodiment, intake air temperature, fuel input amount, gas turbine exhaust gas temperature, gas turbine rotation speed, exhaust heat recovery boiler steam generation amount, steam turbine output, and other state quantities were measured at 1 second intervals as operation data. In the data transmission device, the operation data measured by the control device 82 is transmitted to the monitoring center via the Internet 7.

  The monitoring center receives the operation data of the power generation facilities 8a to 8n received via the Internet 7, and monitors the state of the power generation facilities 8a to 8n. The monitoring center includes a data receiving device 2, an operation database 3, a maintenance database 4, a data control device 5, data classification devices 1 a to 1 n and an operation terminal 6. These components are shown below.

  The data receiving device 2 sends the received data to each database and the data control device 5. The operation data transmitted from the power generation facilities 8 a to 8 n is sent to the operation database 3 at the same time as being sent to the data control device 5. The maintenance information input from the portable terminal by the maintenance staff is sent to the maintenance database 4.

  The operation database 3 stores the operation data sent from the data receiving device 2 for each of the power generation facilities 8a to 8n. The operation data includes state quantities such as intake air temperature, fuel input amount, gas turbine exhaust gas temperature, and the like, and is recorded at intervals of 1 second. In addition to these measurement data, an operation mode is recorded. The operation mode is recorded with numerical values corresponding to “start-up operation”, “rated operation”, and the like. Further, in addition to data transmitted from the power generation facility, classification results by the data classification devices 1a to 1n are also stored. There are two types of classification results. One is a category A number in the data classification device. The other is the state (category B) of the device such as “normal” and “abnormal A”, which are also recorded with numbers corresponding to the state.

  The maintenance database 4 stores maintenance information sent from the data receiving device 2. Maintenance information performed by maintenance personnel is recorded in the maintenance database 4. The accumulated maintenance information is recorded as shown in FIG. 13, for example. In the example of FIG. 13, the actual maintenance work is as follows. At 9:50, an alarm was generated by monitoring and an abnormality was confirmed. Maintenance personnel arrived at the site at 10:30 and maintenance work started. After exchanging the parts at 10:45, the gas turbine generator was commissioned. After confirming that the abnormality was recovered at 11:30, the maintenance work was completed.

  In the data control device 5, the operation data and the maintenance data are classified by site and operation mode, and processed into a format that can be input to the data classification devices 1a to 1n. Details of this data classification and processing method will be described later. The data classification device has an estimation mode, a learning mode, and a preparation mode. The data control device 5 controls input data according to the mode.

  The operation terminal 6 inputs a mode switching command for the data control device 5. In addition, the information recorded in the operation database 3 and the maintenance database 4 or the results classified by the data classification devices 1a to 1n can be output to the monitor screen.

  The data classification devices 1a to 1n have an estimation mode, a learning mode, and a preparation mode. In the estimation mode, the states of the power generation facilities 8a to 8n are diagnosed based on the operation data sent from the data receiving device 20. The diagnosis result is sent to the operation database 3. In the learning mode, the relation between the operation data and the state of the power generation facility is learned using information stored in the operation database 3 and the maintenance database. In addition, the preparation mode does not perform learning or estimation. This mode is a mode when there is no data necessary for learning, such as when a new site is established. These modes can be set individually for each of the diagnostic apparatuses 1a to 1n. The mode is set on the operation terminal 6.

  Next, the operation of the monitoring center will be described. Since there are two modes in which the data classification devices 1a to 1n actually operate, the learning mode and the estimation mode will be described in the order of the learning mode and the estimation mode. As described above, the learning mode and the estimation mode can be individually set for each site, but here, a case where a plurality of sites are all in the learning mode and a case where all the sites are in the estimation mode will be described as an example.

  In the learning mode, information stored in the operation database 3 and the maintenance database 4 is used to learn the relationship between the operation data and the state of the power generation equipment. However, the information stored in both databases cannot be directly input to the data classification devices 1a to 1n. Therefore, the data control device 5 converts the operation data and maintenance information into a data format that is input to the data classification device. The algorithm is shown in FIG. In the following description, measurement items will be described using an example in which three items (DATA1 to DATA3) are selected from the items described above.

  Step 301: All operation data is grouped by site and operation mode, and the sampling interval of data of each group is determined. An example of grouping is shown below. For example, if the operation mode is “start operation”, “rated operation”, and “stop operation” at the three sites “site 1” to “site 3”, the operation data is grouped into nine groups in total. . The data sampling interval varies depending on the operation mode. In this example, the “rated operation” with little change in operation data was set at 1 minute intervals, and in other modes, the interval was set at 1 second.

  Step 302: Select one group from the groups classified in Step 301.

  Step 303: Steps 101 to 105 of the data preprocessing device 20 shown in FIG. 5 are performed on the data DATA1 to DATA3 using the data corresponding to the group selected in Step 302.

  First, the maximum value and the minimum value are calculated for each measurement item of DATA1 to DATA5. The calculated maximum and minimum values are stored in the data control device 5 for each site.

  Next, the data is normalized using the calculated maximum and minimum values. The method will be described taking DATA1 as an example. The data number of DATA1 is M and the mth measurement value is data1 (m). Further, if the maximum value and the minimum value in the M data are respectively Max_1 and Min_1, normalized data Ndata_1 (m) is calculated by Expression (16).

Ndata_1 (m) = α + (1-α) × (data_1 (m) −Min_1) / (Max_1−Min_1)
... (16)

  Here, it is a constant of α (0 ≦ α <0.5), and the data is normalized to the range of [α, 1−α] by the equation (1). In this embodiment, α = 0.1.

  Next, the processing of Step 103 and Step 104 in FIG. 5 is performed. Here, 200 points of rated operation data were extracted from the operation data, and the standard deviation σi (i = 1 to 3) of DATA1 to DATA3 was calculated. Finally, the normalized data Ndata_i (m) (i = 1 to 3) of DATA1 to DATA3 is corrected by Expression (17).

    Ndata_i ′ (m) = (Ndata_i (m) −0.5) * σs / σi + 0.5 (17)

  Here, σs is the minimum value in σi (i = 1 to 3).

  Step 304: Convert the device status data to create category B data. The maintenance database 4 records the state of the device. For example, if there are three types of states, “normal”, “abnormal A”, and “abnormal B”, each is converted into number data such as “1”, “2”, and “3”. However, the equipment status is not recorded in the maintenance database 4 at fixed time intervals. Therefore, the data is processed according to the time interval of the input data a. By correcting the time interval in this way, data corresponding to the operation data at the same time and the state of the equipment at that time can be obtained.

  Step 305: It is determined whether the processing of Step 303 to Step 305 has been executed for all the groups classified in Step 301. If there are any remaining groups, Steps 303 to 305 are executed for the data corresponding to those groups. If the processing is completed for all groups, the process ends.

  As a result of the above processing, pairs of input data and category B at that time are created for the number classified in Step 301. These are respectively input to different data classification apparatuses 1a to 1n.

  Next, the operation of the data classification devices 1a to 1n will be described. The data classification devices 1a to 1n learn the relationship between input data and category B. There are M input data and category B. In this embodiment, M sets of data are sequentially input to the data classification devices 1a to 1n. Since the algorithm for learning the relationship between the M sets of input data and category B has been described in detail in the second embodiment of the present invention, description thereof will be omitted in this embodiment.

  When the learning is completed, the prototype obtained as a result of the learning is stored in the storage areas in the data classification devices 1a to 1n.

  Next, the operation in the estimation mode will be described. In the estimation mode, the state of the power generation facilities 8a to 8n is diagnosed based on the operation data sent from the data receiving device 2 every minute. Data sent from the data receiving device 2 is processed by the data control device 5 and sent to the data classification devices 1a to 1n.

  The data control device 5 stores a maximum value and a minimum value for normalizing data for each site. Accordingly, the data sent from the data receiving device 2 is instantaneously converted by the processing in step 304 in the learning mode and input to the data classification devices 1a to 1n. The data classification devices 1a to 1n store prototypes obtained in the learning mode. Therefore, the category A and the state of the power generation facility (category B) are estimated by the procedure described in the estimation mode of the second embodiment. The estimated category A and the state of the power generation equipment can be output to the monitor screen of the operation terminal 6. It can also be stored in the operation database 3.

  In the third embodiment, the data classification device according to the second embodiment is used. However, the data classification device according to the first embodiment can be used.

  The apparatus of each embodiment described above can be implemented by a computer or the like having storage means and a CPU, and each processing means as a function of the apparatus is a program module, and the module is read and executed by the computer. Each function can be implemented. Each function can be implemented by causing a computer to read a recording medium on which a program module is recorded.

  In the above-described embodiment, each unit such as the input data preprocessing unit 10, the data classification unit 20, and the processing flow control unit 30 is provided in the data classification device, and the processing is executed. These processes can be executed by separate apparatuses and communicated with each other.

1 Data classification device 2 Data reception device 3 Operation database 4 Maintenance database 5 Data control device 6 Operation terminal 7 Internet 8 Power generation facility 10 Input data preprocessing means 20, 40 Data classification means 21 Data reading means 22 Internal data initialization means 23 Inside Data storage unit 24 Prototype selection unit 25 Prototype evaluation unit 26 Prototype addition unit 27 Internal data correction unit 28 Internal data and matrix correction unit 30 Processing flow control unit 41 Category B reading unit 42 Category conversion unit 43 Category correspondence matrix storage unit

Claims (12)

Computer
Calculate the standard deviation of each item of sampled input data, pre-process the input data based on the standard deviation of each item found,
Select the prototype of the category closest to the preprocessed input data,
Evaluating whether the selected prototype is valid from the relationship between the prototype and the input data,
A data classification method characterized by classifying a plurality of input data composed of a plurality of items into a plurality of categories. The data classification method according to claim 1,
If the evaluation determines that the prototype is valid, correct at least one of the region determination parameter and the prototype,
A data classification method comprising: adding a new prototype when the prototype is not determined to be valid by the evaluation. A data classification method in which a computer classifies a plurality of natural input data into a plurality of preset categories B,
Storing a prototype of category A related to category B, a region determination parameter that defines the size of the region of category A, and a category correspondence matrix for converting category A to category B;
Calculate the standard deviation of each item of sampled input data, pre-process the input data based on the standard deviation of each item found,
A category A prototype closest to the preprocessed input data is selected, whether the prototype is valid is determined from the relationship between the selected prototype and the preprocessed input data, and the prototype is determined to be valid If not, add a new prototype and repeat selection of the prototype, and if the prototype is determined to be valid, correct the region determination parameter or the prototype and the category correspondence matrix;
A data classification method, wherein the selected category A is converted to category B using a corrected category correspondence matrix. A data classification method according to claim 3, wherein
If the selected category A already corresponds to another category B when adding a relationship between the selected category A and the preset category B to the category correspondence matrix,
The addition of the correspondence relationship is rejected, the category A prototype is corrected so as to be away from the input data, and the region determination parameter is equal to the distance between the category A prototype and the input data. The data classification method characterized by correcting as follows. The computer inputs the operation data of the equipment, executes the data classification method according to claim 1,
A facility diagnosis method, wherein the facility state is diagnosed based on a category classified by the data classification method. The computer inputs the operation data of the equipment, executes the data classification method according to claim 3,
Giving the state of the equipment as a teacher pattern, learning the relationship between the state of the equipment and the operation data of the equipment by the data classification device,
A diagnosis method characterized by inputting unlearned operation data to the data classification device after learning, and diagnosing the state of the facility based on a category B output by the data classification device. A data classification device for classifying a plurality of input data composed of a plurality of items into a plurality of categories,
A data preprocessing means for obtaining a standard deviation of each item of the sampled input data and preprocessing the input data based on the obtained standard deviation of each item;
Prototype selection means for selecting a prototype of the category closest to the input data processed by the data preprocessing means;
A data classification apparatus comprising: prototype evaluation means for evaluating whether the prototype is valid from the relationship between the prototype and the input data. In the data classification device according to claim 7,
The prototype of the category, and an internal data storage unit that stores a region determination parameter that defines the size of the category region;
Prototype addition means to create a new prototype,
Internal data correction means capable of correcting the region determination parameter and the prototype;
In the prototype evaluation unit, when the prototype is determined to be valid, the internal data correction unit corrects at least one of the region determination parameter and the prototype, and if the prototype is not determined to be valid, the prototype evaluation unit A data classification device characterized in that a new prototype is created by an additional means. A data classification device for classifying a plurality of input data composed of a plurality of items into a plurality of preset categories B,
A prototype of category A related to category B, and an internal data storage for storing area determination parameters that define the size of the area of category A;
A category correspondence matrix storage for storing a category correspondence matrix for converting category A to category B;
A data preprocessing means for obtaining a standard deviation of each item of the sampled input data and preprocessing the input data based on the obtained standard deviation of each item;
Prototype selection means for selecting a category A prototype closest to the input data processed by the data preprocessing means;
Prototype evaluation means for evaluating whether the prototype is valid from the relationship between the prototype and the input data;
Prototype adding means for creating a new category A prototype and repeating selection of the prototype when the prototype evaluation means does not determine that the prototype is valid;
Internal data and matrix correction means for correcting the region determination parameter or the prototype, and the category correspondence matrix when the prototype is determined to be valid;
A data classification apparatus comprising: category conversion means for converting the selected category A into a category B using the category correspondence matrix when the prototype is determined to be valid. The data classification device according to claim 9, wherein
In the internal data and matrix correction means, when adding the relationship between the selected category A and the preset category B to the category correspondence matrix, if the selected category A already corresponds to another category B, Refuse to add correspondence,
The category A prototype is corrected away from the input data, and the area determination parameter is corrected so that the area determination parameter is equal to the distance between the category A prototype and the input data. Data classification device. In the equipment diagnostic apparatus, the data classification apparatus according to claim 6 is provided,
A facility diagnosis apparatus, wherein operation data of the facility is input to the data classification device, and the state of the facility is diagnosed based on the classified category. In equipment diagnostic equipment,
A data classification device according to claim 8,
Giving the state of the equipment as a teacher pattern, learning the relationship between the state of the equipment and the operation data of the equipment by the data classification device,
An equipment diagnosis apparatus characterized by inputting unlearned operation data to the data classification apparatus after learning, and diagnosing the state of the equipment based on the category B output by the data classification apparatus.

Images